diff --git a/.gitignore b/.gitignore
index f81086b..b0b248c 100644
--- a/.gitignore
+++ b/.gitignore
@@ -14,3 +14,6 @@ _build/
 
 # coverage.py default output file
 .coverage
+
+# IDE cruft
+.idea/
diff --git a/docs/performance.rst b/docs/performance.rst
index 2d12b9c..d22b06a 100644
--- a/docs/performance.rst
+++ b/docs/performance.rst
@@ -36,7 +36,7 @@ Both of these approaches will be much faster than using `load/s()`, since
 they avoid loading the parts of the document we didn't care about.
 
 Both `Object` and `Array` have a `mini` property that returns their entire
-content as a minified Python `str`. A message router for example would only
+content as a minified Python `bytes`. A message router for example would only
 parse the document and retrieve a single property, the destination, and forward
 the payload without ever turning it into a Python object. Here's a (bad)
 example:
diff --git a/simdjson/__init__.pyi b/simdjson/__init__.pyi
index d42b480..606e804 100644
--- a/simdjson/__init__.pyi
+++ b/simdjson/__init__.pyi
@@ -52,7 +52,7 @@ class Object(Mapping[str, SimValue]):
         ...
 
     @property
-    def mini(self) -> str:
+    def mini(self) -> bytes:
         ...
 
 
@@ -73,7 +73,7 @@ class Array(Sequence[SimValue]):
         ...
 
     @property
-    def mini(self) -> str:
+    def mini(self) -> bytes:
         ...
 
 
diff --git a/simdjson/csimdjson.cpp b/simdjson/csimdjson.cpp
index 9145f61..6336efc 100644
--- a/simdjson/csimdjson.cpp
+++ b/simdjson/csimdjson.cpp
@@ -1,4 +1,4 @@
-/* Generated by Cython 0.29.30 */
+/* Generated by Cython 0.29.32 */
 
 /* BEGIN: Cython Metadata
 {
@@ -32,8 +32,8 @@ END: Cython Metadata */
 #elif PY_VERSION_HEX < 0x02060000 || (0x03000000 <= PY_VERSION_HEX && PY_VERSION_HEX < 0x03030000)
     #error Cython requires Python 2.6+ or Python 3.3+.
 #else
-#define CYTHON_ABI "0_29_30"
-#define CYTHON_HEX_VERSION 0x001D1EF0
+#define CYTHON_ABI "0_29_32"
+#define CYTHON_HEX_VERSION 0x001D20F0
 #define CYTHON_FUTURE_DIVISION 1
 #include <stddef.h>
 #ifndef offsetof
@@ -72,6 +72,7 @@ END: Cython Metadata */
   #define CYTHON_COMPILING_IN_PYPY 1
   #define CYTHON_COMPILING_IN_PYSTON 0
   #define CYTHON_COMPILING_IN_CPYTHON 0
+  #define CYTHON_COMPILING_IN_NOGIL 0
   #undef CYTHON_USE_TYPE_SLOTS
   #define CYTHON_USE_TYPE_SLOTS 0
   #undef CYTHON_USE_PYTYPE_LOOKUP
@@ -115,6 +116,7 @@ END: Cython Metadata */
   #define CYTHON_COMPILING_IN_PYPY 0
   #define CYTHON_COMPILING_IN_PYSTON 1
   #define CYTHON_COMPILING_IN_CPYTHON 0
+  #define CYTHON_COMPILING_IN_NOGIL 0
   #ifndef CYTHON_USE_TYPE_SLOTS
     #define CYTHON_USE_TYPE_SLOTS 1
   #endif
@@ -155,10 +157,56 @@ END: Cython Metadata */
   #ifndef CYTHON_UPDATE_DESCRIPTOR_DOC
     #define CYTHON_UPDATE_DESCRIPTOR_DOC 0
   #endif
+#elif defined(PY_NOGIL)
+  #define CYTHON_COMPILING_IN_PYPY 0
+  #define CYTHON_COMPILING_IN_PYSTON 0
+  #define CYTHON_COMPILING_IN_CPYTHON 0
+  #define CYTHON_COMPILING_IN_NOGIL 1
+  #ifndef CYTHON_USE_TYPE_SLOTS
+    #define CYTHON_USE_TYPE_SLOTS 1
+  #endif
+  #undef CYTHON_USE_PYTYPE_LOOKUP
+  #define CYTHON_USE_PYTYPE_LOOKUP 0
+  #ifndef CYTHON_USE_ASYNC_SLOTS
+    #define CYTHON_USE_ASYNC_SLOTS 1
+  #endif
+  #undef CYTHON_USE_PYLIST_INTERNALS
+  #define CYTHON_USE_PYLIST_INTERNALS 0
+  #ifndef CYTHON_USE_UNICODE_INTERNALS
+    #define CYTHON_USE_UNICODE_INTERNALS 1
+  #endif
+  #undef CYTHON_USE_UNICODE_WRITER
+  #define CYTHON_USE_UNICODE_WRITER 0
+  #undef CYTHON_USE_PYLONG_INTERNALS
+  #define CYTHON_USE_PYLONG_INTERNALS 0
+  #ifndef CYTHON_AVOID_BORROWED_REFS
+    #define CYTHON_AVOID_BORROWED_REFS 0
+  #endif
+  #ifndef CYTHON_ASSUME_SAFE_MACROS
+    #define CYTHON_ASSUME_SAFE_MACROS 1
+  #endif
+  #ifndef CYTHON_UNPACK_METHODS
+    #define CYTHON_UNPACK_METHODS 1
+  #endif
+  #undef CYTHON_FAST_THREAD_STATE
+  #define CYTHON_FAST_THREAD_STATE 0
+  #undef CYTHON_FAST_PYCALL
+  #define CYTHON_FAST_PYCALL 0
+  #ifndef CYTHON_PEP489_MULTI_PHASE_INIT
+    #define CYTHON_PEP489_MULTI_PHASE_INIT 1
+  #endif
+  #ifndef CYTHON_USE_TP_FINALIZE
+    #define CYTHON_USE_TP_FINALIZE 1
+  #endif
+  #undef CYTHON_USE_DICT_VERSIONS
+  #define CYTHON_USE_DICT_VERSIONS 0
+  #undef CYTHON_USE_EXC_INFO_STACK
+  #define CYTHON_USE_EXC_INFO_STACK 0
 #else
   #define CYTHON_COMPILING_IN_PYPY 0
   #define CYTHON_COMPILING_IN_PYSTON 0
   #define CYTHON_COMPILING_IN_CPYTHON 1
+  #define CYTHON_COMPILING_IN_NOGIL 0
   #ifndef CYTHON_USE_TYPE_SLOTS
     #define CYTHON_USE_TYPE_SLOTS 1
   #endif
@@ -955,7 +1003,7 @@ static const char *__pyx_filename;
 static const char *__pyx_f[] = {
   "simdjson/csimdjson.pyx",
   "stringsource",
-  "venv/lib/python3.10/site-packages/Cython/Includes/cpython/type.pxd",
+  "backend/venv/lib/python3.9/site-packages/Cython/Includes/cpython/type.pxd",
 };
 /* MemviewSliceStruct.proto */
 struct __pyx_memoryview_obj;
@@ -973,30 +1021,26 @@ typedef struct {
 #ifndef CYTHON_ATOMICS
     #define CYTHON_ATOMICS 1
 #endif
+#define __PYX_CYTHON_ATOMICS_ENABLED() CYTHON_ATOMICS
 #define __pyx_atomic_int_type int
-#if CYTHON_ATOMICS && __GNUC__ >= 4 && (__GNUC_MINOR__ > 1 ||\
-                    (__GNUC_MINOR__ == 1 && __GNUC_PATCHLEVEL >= 2)) &&\
-                    !defined(__i386__)
-    #define __pyx_atomic_incr_aligned(value, lock) __sync_fetch_and_add(value, 1)
-    #define __pyx_atomic_decr_aligned(value, lock) __sync_fetch_and_sub(value, 1)
+#if CYTHON_ATOMICS && (__GNUC__ >= 5 || (__GNUC__ == 4 &&\
+                    (__GNUC_MINOR__ > 1 ||\
+                    (__GNUC_MINOR__ == 1 && __GNUC_PATCHLEVEL__ >= 2))))
+    #define __pyx_atomic_incr_aligned(value) __sync_fetch_and_add(value, 1)
+    #define __pyx_atomic_decr_aligned(value) __sync_fetch_and_sub(value, 1)
     #ifdef __PYX_DEBUG_ATOMICS
         #warning "Using GNU atomics"
     #endif
-#elif CYTHON_ATOMICS && defined(_MSC_VER) && 0
-    #include <Windows.h>
+#elif CYTHON_ATOMICS && defined(_MSC_VER) && CYTHON_COMPILING_IN_NOGIL
+    #include <intrin.h>
     #undef __pyx_atomic_int_type
-    #define __pyx_atomic_int_type LONG
-    #define __pyx_atomic_incr_aligned(value, lock) InterlockedIncrement(value)
-    #define __pyx_atomic_decr_aligned(value, lock) InterlockedDecrement(value)
+    #define __pyx_atomic_int_type long
+    #pragma intrinsic (_InterlockedExchangeAdd)
+    #define __pyx_atomic_incr_aligned(value) _InterlockedExchangeAdd(value, 1)
+    #define __pyx_atomic_decr_aligned(value) _InterlockedExchangeAdd(value, -1)
     #ifdef __PYX_DEBUG_ATOMICS
         #pragma message ("Using MSVC atomics")
     #endif
-#elif CYTHON_ATOMICS && (defined(__ICC) || defined(__INTEL_COMPILER)) && 0
-    #define __pyx_atomic_incr_aligned(value, lock) _InterlockedIncrement(value)
-    #define __pyx_atomic_decr_aligned(value, lock) _InterlockedDecrement(value)
-    #ifdef __PYX_DEBUG_ATOMICS
-        #warning "Using Intel atomics"
-    #endif
 #else
     #undef CYTHON_ATOMICS
     #define CYTHON_ATOMICS 0
@@ -1007,9 +1051,9 @@ typedef struct {
 typedef volatile __pyx_atomic_int_type __pyx_atomic_int;
 #if CYTHON_ATOMICS
     #define __pyx_add_acquisition_count(memview)\
-             __pyx_atomic_incr_aligned(__pyx_get_slice_count_pointer(memview), memview->lock)
+             __pyx_atomic_incr_aligned(__pyx_get_slice_count_pointer(memview))
     #define __pyx_sub_acquisition_count(memview)\
-            __pyx_atomic_decr_aligned(__pyx_get_slice_count_pointer(memview), memview->lock)
+            __pyx_atomic_decr_aligned(__pyx_get_slice_count_pointer(memview))
 #else
     #define __pyx_add_acquisition_count(memview)\
             __pyx_add_acquisition_count_locked(__pyx_get_slice_count_pointer(memview), memview->lock)
@@ -1231,7 +1275,7 @@ struct __pyx_obj_9csimdjson___pyx_scope_struct_4_get_implementations {
 };
 
 
-/* "View.MemoryView":105
+/* "View.MemoryView":106
  * 
  * @cname("__pyx_array")
  * cdef class array:             # <<<<<<<<<<<<<<
@@ -1256,7 +1300,7 @@ struct __pyx_array_obj {
 };
 
 
-/* "View.MemoryView":279
+/* "View.MemoryView":280
  * 
  * @cname('__pyx_MemviewEnum')
  * cdef class Enum(object):             # <<<<<<<<<<<<<<
@@ -1269,7 +1313,7 @@ struct __pyx_MemviewEnum_obj {
 };
 
 
-/* "View.MemoryView":330
+/* "View.MemoryView":331
  * 
  * @cname('__pyx_memoryview')
  * cdef class memoryview(object):             # <<<<<<<<<<<<<<
@@ -1292,7 +1336,7 @@ struct __pyx_memoryview_obj {
 };
 
 
-/* "View.MemoryView":965
+/* "View.MemoryView":967
  * 
  * @cname('__pyx_memoryviewslice')
  * cdef class _memoryviewslice(memoryview):             # <<<<<<<<<<<<<<
@@ -1354,7 +1398,7 @@ static struct __pyx_vtabstruct_9csimdjson_Object *__pyx_vtabptr_9csimdjson_Objec
 static CYTHON_INLINE PyObject *__pyx_f_9csimdjson_6Object_from_element(struct __pyx_obj_9csimdjson_Parser *, simdjson::dom::element);
 
 
-/* "View.MemoryView":105
+/* "View.MemoryView":106
  * 
  * @cname("__pyx_array")
  * cdef class array:             # <<<<<<<<<<<<<<
@@ -1368,7 +1412,7 @@ struct __pyx_vtabstruct_array {
 static struct __pyx_vtabstruct_array *__pyx_vtabptr_array;
 
 
-/* "View.MemoryView":330
+/* "View.MemoryView":331
  * 
  * @cname('__pyx_memoryview')
  * cdef class memoryview(object):             # <<<<<<<<<<<<<<
@@ -1388,7 +1432,7 @@ struct __pyx_vtabstruct_memoryview {
 static struct __pyx_vtabstruct_memoryview *__pyx_vtabptr_memoryview;
 
 
-/* "View.MemoryView":965
+/* "View.MemoryView":967
  * 
  * @cname('__pyx_memoryviewslice')
  * cdef class _memoryviewslice(memoryview):             # <<<<<<<<<<<<<<
@@ -8676,7 +8720,7 @@ static CYTHON_INLINE PyObject *__pyx_convert_PyByteArray_string_to_py_std__in_st
   return __pyx_r;
 }
 
-/* "View.MemoryView":122
+/* "View.MemoryView":123
  *         cdef bint dtype_is_object
  * 
  *     def __cinit__(array self, tuple shape, Py_ssize_t itemsize, format not None,             # <<<<<<<<<<<<<<
@@ -8728,13 +8772,13 @@ static int __pyx_array___cinit__(PyObject *__pyx_v_self, PyObject *__pyx_args, P
         case  1:
         if (likely((values[1] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_itemsize)) != 0)) kw_args--;
         else {
-          __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 3, 5, 1); __PYX_ERR(1, 122, __pyx_L3_error)
+          __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 3, 5, 1); __PYX_ERR(1, 123, __pyx_L3_error)
         }
         CYTHON_FALLTHROUGH;
         case  2:
         if (likely((values[2] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_format)) != 0)) kw_args--;
         else {
-          __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 3, 5, 2); __PYX_ERR(1, 122, __pyx_L3_error)
+          __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 3, 5, 2); __PYX_ERR(1, 123, __pyx_L3_error)
         }
         CYTHON_FALLTHROUGH;
         case  3:
@@ -8750,7 +8794,7 @@ static int __pyx_array___cinit__(PyObject *__pyx_v_self, PyObject *__pyx_args, P
         }
       }
       if (unlikely(kw_args > 0)) {
-        if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_pyargnames, 0, values, pos_args, "__cinit__") < 0)) __PYX_ERR(1, 122, __pyx_L3_error)
+        if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_pyargnames, 0, values, pos_args, "__cinit__") < 0)) __PYX_ERR(1, 123, __pyx_L3_error)
       }
     } else {
       switch (PyTuple_GET_SIZE(__pyx_args)) {
@@ -8766,14 +8810,14 @@ static int __pyx_array___cinit__(PyObject *__pyx_v_self, PyObject *__pyx_args, P
       }
     }
     __pyx_v_shape = ((PyObject*)values[0]);
-    __pyx_v_itemsize = __Pyx_PyIndex_AsSsize_t(values[1]); if (unlikely((__pyx_v_itemsize == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 122, __pyx_L3_error)
+    __pyx_v_itemsize = __Pyx_PyIndex_AsSsize_t(values[1]); if (unlikely((__pyx_v_itemsize == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 123, __pyx_L3_error)
     __pyx_v_format = values[2];
     __pyx_v_mode = values[3];
     if (values[4]) {
-      __pyx_v_allocate_buffer = __Pyx_PyObject_IsTrue(values[4]); if (unlikely((__pyx_v_allocate_buffer == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 123, __pyx_L3_error)
+      __pyx_v_allocate_buffer = __Pyx_PyObject_IsTrue(values[4]); if (unlikely((__pyx_v_allocate_buffer == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 124, __pyx_L3_error)
     } else {
 
-      /* "View.MemoryView":123
+      /* "View.MemoryView":124
  * 
  *     def __cinit__(array self, tuple shape, Py_ssize_t itemsize, format not None,
  *                   mode="c", bint allocate_buffer=True):             # <<<<<<<<<<<<<<
@@ -8785,19 +8829,19 @@ static int __pyx_array___cinit__(PyObject *__pyx_v_self, PyObject *__pyx_args, P
   }
   goto __pyx_L4_argument_unpacking_done;
   __pyx_L5_argtuple_error:;
-  __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 3, 5, PyTuple_GET_SIZE(__pyx_args)); __PYX_ERR(1, 122, __pyx_L3_error)
+  __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 3, 5, PyTuple_GET_SIZE(__pyx_args)); __PYX_ERR(1, 123, __pyx_L3_error)
   __pyx_L3_error:;
   __Pyx_AddTraceback("View.MemoryView.array.__cinit__", __pyx_clineno, __pyx_lineno, __pyx_filename);
   __Pyx_RefNannyFinishContext();
   return -1;
   __pyx_L4_argument_unpacking_done:;
-  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_shape), (&PyTuple_Type), 1, "shape", 1))) __PYX_ERR(1, 122, __pyx_L1_error)
+  if (unlikely(!__Pyx_ArgTypeTest(((PyObject *)__pyx_v_shape), (&PyTuple_Type), 1, "shape", 1))) __PYX_ERR(1, 123, __pyx_L1_error)
   if (unlikely(((PyObject *)__pyx_v_format) == Py_None)) {
-    PyErr_Format(PyExc_TypeError, "Argument '%.200s' must not be None", "format"); __PYX_ERR(1, 122, __pyx_L1_error)
+    PyErr_Format(PyExc_TypeError, "Argument '%.200s' must not be None", "format"); __PYX_ERR(1, 123, __pyx_L1_error)
   }
   __pyx_r = __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(((struct __pyx_array_obj *)__pyx_v_self), __pyx_v_shape, __pyx_v_itemsize, __pyx_v_format, __pyx_v_mode, __pyx_v_allocate_buffer);
 
-  /* "View.MemoryView":122
+  /* "View.MemoryView":123
  *         cdef bint dtype_is_object
  * 
  *     def __cinit__(array self, tuple shape, Py_ssize_t itemsize, format not None,             # <<<<<<<<<<<<<<
@@ -8839,7 +8883,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
   __Pyx_RefNannySetupContext("__cinit__", 0);
   __Pyx_INCREF(__pyx_v_format);
 
-  /* "View.MemoryView":129
+  /* "View.MemoryView":130
  *         cdef PyObject **p
  * 
  *         self.ndim = <int> len(shape)             # <<<<<<<<<<<<<<
@@ -8848,12 +8892,12 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
   if (unlikely(__pyx_v_shape == Py_None)) {
     PyErr_SetString(PyExc_TypeError, "object of type 'NoneType' has no len()");
-    __PYX_ERR(1, 129, __pyx_L1_error)
+    __PYX_ERR(1, 130, __pyx_L1_error)
   }
-  __pyx_t_1 = PyTuple_GET_SIZE(__pyx_v_shape); if (unlikely(__pyx_t_1 == ((Py_ssize_t)-1))) __PYX_ERR(1, 129, __pyx_L1_error)
+  __pyx_t_1 = PyTuple_GET_SIZE(__pyx_v_shape); if (unlikely(__pyx_t_1 == ((Py_ssize_t)-1))) __PYX_ERR(1, 130, __pyx_L1_error)
   __pyx_v_self->ndim = ((int)__pyx_t_1);
 
-  /* "View.MemoryView":130
+  /* "View.MemoryView":131
  * 
  *         self.ndim = <int> len(shape)
  *         self.itemsize = itemsize             # <<<<<<<<<<<<<<
@@ -8862,7 +8906,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
   __pyx_v_self->itemsize = __pyx_v_itemsize;
 
-  /* "View.MemoryView":132
+  /* "View.MemoryView":133
  *         self.itemsize = itemsize
  * 
  *         if not self.ndim:             # <<<<<<<<<<<<<<
@@ -8872,20 +8916,20 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
   __pyx_t_2 = ((!(__pyx_v_self->ndim != 0)) != 0);
   if (unlikely(__pyx_t_2)) {
 
-    /* "View.MemoryView":133
+    /* "View.MemoryView":134
  * 
  *         if not self.ndim:
  *             raise ValueError("Empty shape tuple for cython.array")             # <<<<<<<<<<<<<<
  * 
  *         if itemsize <= 0:
  */
-    __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__16, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 133, __pyx_L1_error)
+    __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__16, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 134, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
     __Pyx_Raise(__pyx_t_3, 0, 0, 0);
     __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
-    __PYX_ERR(1, 133, __pyx_L1_error)
+    __PYX_ERR(1, 134, __pyx_L1_error)
 
-    /* "View.MemoryView":132
+    /* "View.MemoryView":133
  *         self.itemsize = itemsize
  * 
  *         if not self.ndim:             # <<<<<<<<<<<<<<
@@ -8894,7 +8938,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
   }
 
-  /* "View.MemoryView":135
+  /* "View.MemoryView":136
  *             raise ValueError("Empty shape tuple for cython.array")
  * 
  *         if itemsize <= 0:             # <<<<<<<<<<<<<<
@@ -8904,20 +8948,20 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
   __pyx_t_2 = ((__pyx_v_itemsize <= 0) != 0);
   if (unlikely(__pyx_t_2)) {
 
-    /* "View.MemoryView":136
+    /* "View.MemoryView":137
  * 
  *         if itemsize <= 0:
  *             raise ValueError("itemsize <= 0 for cython.array")             # <<<<<<<<<<<<<<
  * 
  *         if not isinstance(format, bytes):
  */
-    __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__17, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 136, __pyx_L1_error)
+    __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__17, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 137, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
     __Pyx_Raise(__pyx_t_3, 0, 0, 0);
     __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
-    __PYX_ERR(1, 136, __pyx_L1_error)
+    __PYX_ERR(1, 137, __pyx_L1_error)
 
-    /* "View.MemoryView":135
+    /* "View.MemoryView":136
  *             raise ValueError("Empty shape tuple for cython.array")
  * 
  *         if itemsize <= 0:             # <<<<<<<<<<<<<<
@@ -8926,7 +8970,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
   }
 
-  /* "View.MemoryView":138
+  /* "View.MemoryView":139
  *             raise ValueError("itemsize <= 0 for cython.array")
  * 
  *         if not isinstance(format, bytes):             # <<<<<<<<<<<<<<
@@ -8937,14 +8981,14 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
   __pyx_t_4 = ((!(__pyx_t_2 != 0)) != 0);
   if (__pyx_t_4) {
 
-    /* "View.MemoryView":139
+    /* "View.MemoryView":140
  * 
  *         if not isinstance(format, bytes):
  *             format = format.encode('ASCII')             # <<<<<<<<<<<<<<
  *         self._format = format  # keep a reference to the byte string
  *         self.format = self._format
  */
-    __pyx_t_5 = __Pyx_PyObject_GetAttrStr(__pyx_v_format, __pyx_n_s_encode); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 139, __pyx_L1_error)
+    __pyx_t_5 = __Pyx_PyObject_GetAttrStr(__pyx_v_format, __pyx_n_s_encode); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 140, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_5);
     __pyx_t_6 = NULL;
     if (CYTHON_UNPACK_METHODS && likely(PyMethod_Check(__pyx_t_5))) {
@@ -8958,13 +9002,13 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
     }
     __pyx_t_3 = (__pyx_t_6) ? __Pyx_PyObject_Call2Args(__pyx_t_5, __pyx_t_6, __pyx_n_s_ASCII) : __Pyx_PyObject_CallOneArg(__pyx_t_5, __pyx_n_s_ASCII);
     __Pyx_XDECREF(__pyx_t_6); __pyx_t_6 = 0;
-    if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 139, __pyx_L1_error)
+    if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 140, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
     __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
     __Pyx_DECREF_SET(__pyx_v_format, __pyx_t_3);
     __pyx_t_3 = 0;
 
-    /* "View.MemoryView":138
+    /* "View.MemoryView":139
  *             raise ValueError("itemsize <= 0 for cython.array")
  * 
  *         if not isinstance(format, bytes):             # <<<<<<<<<<<<<<
@@ -8973,14 +9017,14 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
   }
 
-  /* "View.MemoryView":140
+  /* "View.MemoryView":141
  *         if not isinstance(format, bytes):
  *             format = format.encode('ASCII')
  *         self._format = format  # keep a reference to the byte string             # <<<<<<<<<<<<<<
  *         self.format = self._format
  * 
  */
-  if (!(likely(PyBytes_CheckExact(__pyx_v_format))||((__pyx_v_format) == Py_None)||(PyErr_Format(PyExc_TypeError, "Expected %.16s, got %.200s", "bytes", Py_TYPE(__pyx_v_format)->tp_name), 0))) __PYX_ERR(1, 140, __pyx_L1_error)
+  if (!(likely(PyBytes_CheckExact(__pyx_v_format))||((__pyx_v_format) == Py_None)||(PyErr_Format(PyExc_TypeError, "Expected %.16s, got %.200s", "bytes", Py_TYPE(__pyx_v_format)->tp_name), 0))) __PYX_ERR(1, 141, __pyx_L1_error)
   __pyx_t_3 = __pyx_v_format;
   __Pyx_INCREF(__pyx_t_3);
   __Pyx_GIVEREF(__pyx_t_3);
@@ -8989,7 +9033,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
   __pyx_v_self->_format = ((PyObject*)__pyx_t_3);
   __pyx_t_3 = 0;
 
-  /* "View.MemoryView":141
+  /* "View.MemoryView":142
  *             format = format.encode('ASCII')
  *         self._format = format  # keep a reference to the byte string
  *         self.format = self._format             # <<<<<<<<<<<<<<
@@ -8998,12 +9042,12 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
   if (unlikely(__pyx_v_self->_format == Py_None)) {
     PyErr_SetString(PyExc_TypeError, "expected bytes, NoneType found");
-    __PYX_ERR(1, 141, __pyx_L1_error)
+    __PYX_ERR(1, 142, __pyx_L1_error)
   }
-  __pyx_t_7 = __Pyx_PyBytes_AsWritableString(__pyx_v_self->_format); if (unlikely((!__pyx_t_7) && PyErr_Occurred())) __PYX_ERR(1, 141, __pyx_L1_error)
+  __pyx_t_7 = __Pyx_PyBytes_AsWritableString(__pyx_v_self->_format); if (unlikely((!__pyx_t_7) && PyErr_Occurred())) __PYX_ERR(1, 142, __pyx_L1_error)
   __pyx_v_self->format = __pyx_t_7;
 
-  /* "View.MemoryView":144
+  /* "View.MemoryView":145
  * 
  * 
  *         self._shape = <Py_ssize_t *> PyObject_Malloc(sizeof(Py_ssize_t)*self.ndim*2)             # <<<<<<<<<<<<<<
@@ -9012,7 +9056,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
   __pyx_v_self->_shape = ((Py_ssize_t *)PyObject_Malloc((((sizeof(Py_ssize_t)) * __pyx_v_self->ndim) * 2)));
 
-  /* "View.MemoryView":145
+  /* "View.MemoryView":146
  * 
  *         self._shape = <Py_ssize_t *> PyObject_Malloc(sizeof(Py_ssize_t)*self.ndim*2)
  *         self._strides = self._shape + self.ndim             # <<<<<<<<<<<<<<
@@ -9021,7 +9065,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
   __pyx_v_self->_strides = (__pyx_v_self->_shape + __pyx_v_self->ndim);
 
-  /* "View.MemoryView":147
+  /* "View.MemoryView":148
  *         self._strides = self._shape + self.ndim
  * 
  *         if not self._shape:             # <<<<<<<<<<<<<<
@@ -9031,20 +9075,20 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
   __pyx_t_4 = ((!(__pyx_v_self->_shape != 0)) != 0);
   if (unlikely(__pyx_t_4)) {
 
-    /* "View.MemoryView":148
+    /* "View.MemoryView":149
  * 
  *         if not self._shape:
  *             raise MemoryError("unable to allocate shape and strides.")             # <<<<<<<<<<<<<<
  * 
  * 
  */
-    __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_MemoryError, __pyx_tuple__18, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 148, __pyx_L1_error)
+    __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_MemoryError, __pyx_tuple__18, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 149, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
     __Pyx_Raise(__pyx_t_3, 0, 0, 0);
     __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
-    __PYX_ERR(1, 148, __pyx_L1_error)
+    __PYX_ERR(1, 149, __pyx_L1_error)
 
-    /* "View.MemoryView":147
+    /* "View.MemoryView":148
  *         self._strides = self._shape + self.ndim
  * 
  *         if not self._shape:             # <<<<<<<<<<<<<<
@@ -9053,7 +9097,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
   }
 
-  /* "View.MemoryView":151
+  /* "View.MemoryView":152
  * 
  * 
  *         for idx, dim in enumerate(shape):             # <<<<<<<<<<<<<<
@@ -9065,18 +9109,18 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
   for (;;) {
     if (__pyx_t_1 >= PyTuple_GET_SIZE(__pyx_t_3)) break;
     #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS
-    __pyx_t_5 = PyTuple_GET_ITEM(__pyx_t_3, __pyx_t_1); __Pyx_INCREF(__pyx_t_5); __pyx_t_1++; if (unlikely(0 < 0)) __PYX_ERR(1, 151, __pyx_L1_error)
+    __pyx_t_5 = PyTuple_GET_ITEM(__pyx_t_3, __pyx_t_1); __Pyx_INCREF(__pyx_t_5); __pyx_t_1++; if (unlikely(0 < 0)) __PYX_ERR(1, 152, __pyx_L1_error)
     #else
-    __pyx_t_5 = PySequence_ITEM(__pyx_t_3, __pyx_t_1); __pyx_t_1++; if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 151, __pyx_L1_error)
+    __pyx_t_5 = PySequence_ITEM(__pyx_t_3, __pyx_t_1); __pyx_t_1++; if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 152, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_5);
     #endif
-    __pyx_t_9 = __Pyx_PyIndex_AsSsize_t(__pyx_t_5); if (unlikely((__pyx_t_9 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 151, __pyx_L1_error)
+    __pyx_t_9 = __Pyx_PyIndex_AsSsize_t(__pyx_t_5); if (unlikely((__pyx_t_9 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 152, __pyx_L1_error)
     __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
     __pyx_v_dim = __pyx_t_9;
     __pyx_v_idx = __pyx_t_8;
     __pyx_t_8 = (__pyx_t_8 + 1);
 
-    /* "View.MemoryView":152
+    /* "View.MemoryView":153
  * 
  *         for idx, dim in enumerate(shape):
  *             if dim <= 0:             # <<<<<<<<<<<<<<
@@ -9086,18 +9130,18 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
     __pyx_t_4 = ((__pyx_v_dim <= 0) != 0);
     if (unlikely(__pyx_t_4)) {
 
-      /* "View.MemoryView":153
+      /* "View.MemoryView":154
  *         for idx, dim in enumerate(shape):
  *             if dim <= 0:
  *                 raise ValueError("Invalid shape in axis %d: %d." % (idx, dim))             # <<<<<<<<<<<<<<
  *             self._shape[idx] = dim
  * 
  */
-      __pyx_t_5 = __Pyx_PyInt_From_int(__pyx_v_idx); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 153, __pyx_L1_error)
+      __pyx_t_5 = __Pyx_PyInt_From_int(__pyx_v_idx); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 154, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_5);
-      __pyx_t_6 = PyInt_FromSsize_t(__pyx_v_dim); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 153, __pyx_L1_error)
+      __pyx_t_6 = PyInt_FromSsize_t(__pyx_v_dim); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 154, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_6);
-      __pyx_t_10 = PyTuple_New(2); if (unlikely(!__pyx_t_10)) __PYX_ERR(1, 153, __pyx_L1_error)
+      __pyx_t_10 = PyTuple_New(2); if (unlikely(!__pyx_t_10)) __PYX_ERR(1, 154, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_10);
       __Pyx_GIVEREF(__pyx_t_5);
       PyTuple_SET_ITEM(__pyx_t_10, 0, __pyx_t_5);
@@ -9105,17 +9149,17 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
       PyTuple_SET_ITEM(__pyx_t_10, 1, __pyx_t_6);
       __pyx_t_5 = 0;
       __pyx_t_6 = 0;
-      __pyx_t_6 = __Pyx_PyString_Format(__pyx_kp_s_Invalid_shape_in_axis_d_d, __pyx_t_10); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 153, __pyx_L1_error)
+      __pyx_t_6 = __Pyx_PyString_Format(__pyx_kp_s_Invalid_shape_in_axis_d_d, __pyx_t_10); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 154, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_6);
       __Pyx_DECREF(__pyx_t_10); __pyx_t_10 = 0;
-      __pyx_t_10 = __Pyx_PyObject_CallOneArg(__pyx_builtin_ValueError, __pyx_t_6); if (unlikely(!__pyx_t_10)) __PYX_ERR(1, 153, __pyx_L1_error)
+      __pyx_t_10 = __Pyx_PyObject_CallOneArg(__pyx_builtin_ValueError, __pyx_t_6); if (unlikely(!__pyx_t_10)) __PYX_ERR(1, 154, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_10);
       __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0;
       __Pyx_Raise(__pyx_t_10, 0, 0, 0);
       __Pyx_DECREF(__pyx_t_10); __pyx_t_10 = 0;
-      __PYX_ERR(1, 153, __pyx_L1_error)
+      __PYX_ERR(1, 154, __pyx_L1_error)
 
-      /* "View.MemoryView":152
+      /* "View.MemoryView":153
  * 
  *         for idx, dim in enumerate(shape):
  *             if dim <= 0:             # <<<<<<<<<<<<<<
@@ -9124,7 +9168,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
     }
 
-    /* "View.MemoryView":154
+    /* "View.MemoryView":155
  *             if dim <= 0:
  *                 raise ValueError("Invalid shape in axis %d: %d." % (idx, dim))
  *             self._shape[idx] = dim             # <<<<<<<<<<<<<<
@@ -9133,7 +9177,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
     (__pyx_v_self->_shape[__pyx_v_idx]) = __pyx_v_dim;
 
-    /* "View.MemoryView":151
+    /* "View.MemoryView":152
  * 
  * 
  *         for idx, dim in enumerate(shape):             # <<<<<<<<<<<<<<
@@ -9143,17 +9187,17 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
   }
   __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
 
-  /* "View.MemoryView":157
+  /* "View.MemoryView":158
  * 
  *         cdef char order
  *         if mode == 'fortran':             # <<<<<<<<<<<<<<
  *             order = b'F'
  *             self.mode = u'fortran'
  */
-  __pyx_t_4 = (__Pyx_PyString_Equals(__pyx_v_mode, __pyx_n_s_fortran, Py_EQ)); if (unlikely(__pyx_t_4 < 0)) __PYX_ERR(1, 157, __pyx_L1_error)
+  __pyx_t_4 = (__Pyx_PyString_Equals(__pyx_v_mode, __pyx_n_s_fortran, Py_EQ)); if (unlikely(__pyx_t_4 < 0)) __PYX_ERR(1, 158, __pyx_L1_error)
   if (__pyx_t_4) {
 
-    /* "View.MemoryView":158
+    /* "View.MemoryView":159
  *         cdef char order
  *         if mode == 'fortran':
  *             order = b'F'             # <<<<<<<<<<<<<<
@@ -9162,7 +9206,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
     __pyx_v_order = 'F';
 
-    /* "View.MemoryView":159
+    /* "View.MemoryView":160
  *         if mode == 'fortran':
  *             order = b'F'
  *             self.mode = u'fortran'             # <<<<<<<<<<<<<<
@@ -9175,7 +9219,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
     __Pyx_DECREF(__pyx_v_self->mode);
     __pyx_v_self->mode = __pyx_n_u_fortran;
 
-    /* "View.MemoryView":157
+    /* "View.MemoryView":158
  * 
  *         cdef char order
  *         if mode == 'fortran':             # <<<<<<<<<<<<<<
@@ -9185,17 +9229,17 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
     goto __pyx_L10;
   }
 
-  /* "View.MemoryView":160
+  /* "View.MemoryView":161
  *             order = b'F'
  *             self.mode = u'fortran'
  *         elif mode == 'c':             # <<<<<<<<<<<<<<
  *             order = b'C'
  *             self.mode = u'c'
  */
-  __pyx_t_4 = (__Pyx_PyString_Equals(__pyx_v_mode, __pyx_n_s_c, Py_EQ)); if (unlikely(__pyx_t_4 < 0)) __PYX_ERR(1, 160, __pyx_L1_error)
+  __pyx_t_4 = (__Pyx_PyString_Equals(__pyx_v_mode, __pyx_n_s_c, Py_EQ)); if (unlikely(__pyx_t_4 < 0)) __PYX_ERR(1, 161, __pyx_L1_error)
   if (likely(__pyx_t_4)) {
 
-    /* "View.MemoryView":161
+    /* "View.MemoryView":162
  *             self.mode = u'fortran'
  *         elif mode == 'c':
  *             order = b'C'             # <<<<<<<<<<<<<<
@@ -9204,7 +9248,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
     __pyx_v_order = 'C';
 
-    /* "View.MemoryView":162
+    /* "View.MemoryView":163
  *         elif mode == 'c':
  *             order = b'C'
  *             self.mode = u'c'             # <<<<<<<<<<<<<<
@@ -9217,7 +9261,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
     __Pyx_DECREF(__pyx_v_self->mode);
     __pyx_v_self->mode = __pyx_n_u_c;
 
-    /* "View.MemoryView":160
+    /* "View.MemoryView":161
  *             order = b'F'
  *             self.mode = u'fortran'
  *         elif mode == 'c':             # <<<<<<<<<<<<<<
@@ -9227,7 +9271,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
     goto __pyx_L10;
   }
 
-  /* "View.MemoryView":164
+  /* "View.MemoryView":165
  *             self.mode = u'c'
  *         else:
  *             raise ValueError("Invalid mode, expected 'c' or 'fortran', got %s" % mode)             # <<<<<<<<<<<<<<
@@ -9235,18 +9279,18 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  *         self.len = fill_contig_strides_array(self._shape, self._strides,
  */
   /*else*/ {
-    __pyx_t_3 = __Pyx_PyString_FormatSafe(__pyx_kp_s_Invalid_mode_expected_c_or_fortr, __pyx_v_mode); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 164, __pyx_L1_error)
+    __pyx_t_3 = __Pyx_PyString_FormatSafe(__pyx_kp_s_Invalid_mode_expected_c_or_fortr, __pyx_v_mode); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 165, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
-    __pyx_t_10 = __Pyx_PyObject_CallOneArg(__pyx_builtin_ValueError, __pyx_t_3); if (unlikely(!__pyx_t_10)) __PYX_ERR(1, 164, __pyx_L1_error)
+    __pyx_t_10 = __Pyx_PyObject_CallOneArg(__pyx_builtin_ValueError, __pyx_t_3); if (unlikely(!__pyx_t_10)) __PYX_ERR(1, 165, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_10);
     __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
     __Pyx_Raise(__pyx_t_10, 0, 0, 0);
     __Pyx_DECREF(__pyx_t_10); __pyx_t_10 = 0;
-    __PYX_ERR(1, 164, __pyx_L1_error)
+    __PYX_ERR(1, 165, __pyx_L1_error)
   }
   __pyx_L10:;
 
-  /* "View.MemoryView":166
+  /* "View.MemoryView":167
  *             raise ValueError("Invalid mode, expected 'c' or 'fortran', got %s" % mode)
  * 
  *         self.len = fill_contig_strides_array(self._shape, self._strides,             # <<<<<<<<<<<<<<
@@ -9255,7 +9299,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
   __pyx_v_self->len = __pyx_fill_contig_strides_array(__pyx_v_self->_shape, __pyx_v_self->_strides, __pyx_v_itemsize, __pyx_v_self->ndim, __pyx_v_order);
 
-  /* "View.MemoryView":169
+  /* "View.MemoryView":170
  *                                              itemsize, self.ndim, order)
  * 
  *         self.free_data = allocate_buffer             # <<<<<<<<<<<<<<
@@ -9264,19 +9308,19 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
   __pyx_v_self->free_data = __pyx_v_allocate_buffer;
 
-  /* "View.MemoryView":170
+  /* "View.MemoryView":171
  * 
  *         self.free_data = allocate_buffer
  *         self.dtype_is_object = format == b'O'             # <<<<<<<<<<<<<<
  *         if allocate_buffer:
  * 
  */
-  __pyx_t_10 = PyObject_RichCompare(__pyx_v_format, __pyx_n_b_O, Py_EQ); __Pyx_XGOTREF(__pyx_t_10); if (unlikely(!__pyx_t_10)) __PYX_ERR(1, 170, __pyx_L1_error)
-  __pyx_t_4 = __Pyx_PyObject_IsTrue(__pyx_t_10); if (unlikely((__pyx_t_4 == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 170, __pyx_L1_error)
+  __pyx_t_10 = PyObject_RichCompare(__pyx_v_format, __pyx_n_b_O, Py_EQ); __Pyx_XGOTREF(__pyx_t_10); if (unlikely(!__pyx_t_10)) __PYX_ERR(1, 171, __pyx_L1_error)
+  __pyx_t_4 = __Pyx_PyObject_IsTrue(__pyx_t_10); if (unlikely((__pyx_t_4 == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 171, __pyx_L1_error)
   __Pyx_DECREF(__pyx_t_10); __pyx_t_10 = 0;
   __pyx_v_self->dtype_is_object = __pyx_t_4;
 
-  /* "View.MemoryView":171
+  /* "View.MemoryView":172
  *         self.free_data = allocate_buffer
  *         self.dtype_is_object = format == b'O'
  *         if allocate_buffer:             # <<<<<<<<<<<<<<
@@ -9286,7 +9330,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
   __pyx_t_4 = (__pyx_v_allocate_buffer != 0);
   if (__pyx_t_4) {
 
-    /* "View.MemoryView":174
+    /* "View.MemoryView":175
  * 
  * 
  *             self.data = <char *>malloc(self.len)             # <<<<<<<<<<<<<<
@@ -9295,7 +9339,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
     __pyx_v_self->data = ((char *)malloc(__pyx_v_self->len));
 
-    /* "View.MemoryView":175
+    /* "View.MemoryView":176
  * 
  *             self.data = <char *>malloc(self.len)
  *             if not self.data:             # <<<<<<<<<<<<<<
@@ -9305,20 +9349,20 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
     __pyx_t_4 = ((!(__pyx_v_self->data != 0)) != 0);
     if (unlikely(__pyx_t_4)) {
 
-      /* "View.MemoryView":176
+      /* "View.MemoryView":177
  *             self.data = <char *>malloc(self.len)
  *             if not self.data:
  *                 raise MemoryError("unable to allocate array data.")             # <<<<<<<<<<<<<<
  * 
  *             if self.dtype_is_object:
  */
-      __pyx_t_10 = __Pyx_PyObject_Call(__pyx_builtin_MemoryError, __pyx_tuple__19, NULL); if (unlikely(!__pyx_t_10)) __PYX_ERR(1, 176, __pyx_L1_error)
+      __pyx_t_10 = __Pyx_PyObject_Call(__pyx_builtin_MemoryError, __pyx_tuple__19, NULL); if (unlikely(!__pyx_t_10)) __PYX_ERR(1, 177, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_10);
       __Pyx_Raise(__pyx_t_10, 0, 0, 0);
       __Pyx_DECREF(__pyx_t_10); __pyx_t_10 = 0;
-      __PYX_ERR(1, 176, __pyx_L1_error)
+      __PYX_ERR(1, 177, __pyx_L1_error)
 
-      /* "View.MemoryView":175
+      /* "View.MemoryView":176
  * 
  *             self.data = <char *>malloc(self.len)
  *             if not self.data:             # <<<<<<<<<<<<<<
@@ -9327,7 +9371,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
     }
 
-    /* "View.MemoryView":178
+    /* "View.MemoryView":179
  *                 raise MemoryError("unable to allocate array data.")
  * 
  *             if self.dtype_is_object:             # <<<<<<<<<<<<<<
@@ -9337,7 +9381,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
     __pyx_t_4 = (__pyx_v_self->dtype_is_object != 0);
     if (__pyx_t_4) {
 
-      /* "View.MemoryView":179
+      /* "View.MemoryView":180
  * 
  *             if self.dtype_is_object:
  *                 p = <PyObject **> self.data             # <<<<<<<<<<<<<<
@@ -9346,7 +9390,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
       __pyx_v_p = ((PyObject **)__pyx_v_self->data);
 
-      /* "View.MemoryView":180
+      /* "View.MemoryView":181
  *             if self.dtype_is_object:
  *                 p = <PyObject **> self.data
  *                 for i in range(self.len / itemsize):             # <<<<<<<<<<<<<<
@@ -9355,18 +9399,18 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
       if (unlikely(__pyx_v_itemsize == 0)) {
         PyErr_SetString(PyExc_ZeroDivisionError, "integer division or modulo by zero");
-        __PYX_ERR(1, 180, __pyx_L1_error)
+        __PYX_ERR(1, 181, __pyx_L1_error)
       }
       else if (sizeof(Py_ssize_t) == sizeof(long) && (!(((Py_ssize_t)-1) > 0)) && unlikely(__pyx_v_itemsize == (Py_ssize_t)-1)  && unlikely(UNARY_NEG_WOULD_OVERFLOW(__pyx_v_self->len))) {
         PyErr_SetString(PyExc_OverflowError, "value too large to perform division");
-        __PYX_ERR(1, 180, __pyx_L1_error)
+        __PYX_ERR(1, 181, __pyx_L1_error)
       }
       __pyx_t_1 = __Pyx_div_Py_ssize_t(__pyx_v_self->len, __pyx_v_itemsize);
       __pyx_t_9 = __pyx_t_1;
       for (__pyx_t_11 = 0; __pyx_t_11 < __pyx_t_9; __pyx_t_11+=1) {
         __pyx_v_i = __pyx_t_11;
 
-        /* "View.MemoryView":181
+        /* "View.MemoryView":182
  *                 p = <PyObject **> self.data
  *                 for i in range(self.len / itemsize):
  *                     p[i] = Py_None             # <<<<<<<<<<<<<<
@@ -9375,7 +9419,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
         (__pyx_v_p[__pyx_v_i]) = Py_None;
 
-        /* "View.MemoryView":182
+        /* "View.MemoryView":183
  *                 for i in range(self.len / itemsize):
  *                     p[i] = Py_None
  *                     Py_INCREF(Py_None)             # <<<<<<<<<<<<<<
@@ -9385,7 +9429,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
         Py_INCREF(Py_None);
       }
 
-      /* "View.MemoryView":178
+      /* "View.MemoryView":179
  *                 raise MemoryError("unable to allocate array data.")
  * 
  *             if self.dtype_is_object:             # <<<<<<<<<<<<<<
@@ -9394,7 +9438,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
     }
 
-    /* "View.MemoryView":171
+    /* "View.MemoryView":172
  *         self.free_data = allocate_buffer
  *         self.dtype_is_object = format == b'O'
  *         if allocate_buffer:             # <<<<<<<<<<<<<<
@@ -9403,7 +9447,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
  */
   }
 
-  /* "View.MemoryView":122
+  /* "View.MemoryView":123
  *         cdef bint dtype_is_object
  * 
  *     def __cinit__(array self, tuple shape, Py_ssize_t itemsize, format not None,             # <<<<<<<<<<<<<<
@@ -9427,7 +9471,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array___cinit__(struct __
   return __pyx_r;
 }
 
-/* "View.MemoryView":185
+/* "View.MemoryView":186
  * 
  *     @cname('getbuffer')
  *     def __getbuffer__(self, Py_buffer *info, int flags):             # <<<<<<<<<<<<<<
@@ -9470,7 +9514,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
   __pyx_v_info->obj = Py_None; __Pyx_INCREF(Py_None);
   __Pyx_GIVEREF(__pyx_v_info->obj);
 
-  /* "View.MemoryView":186
+  /* "View.MemoryView":187
  *     @cname('getbuffer')
  *     def __getbuffer__(self, Py_buffer *info, int flags):
  *         cdef int bufmode = -1             # <<<<<<<<<<<<<<
@@ -9479,18 +9523,18 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
  */
   __pyx_v_bufmode = -1;
 
-  /* "View.MemoryView":187
+  /* "View.MemoryView":188
  *     def __getbuffer__(self, Py_buffer *info, int flags):
  *         cdef int bufmode = -1
  *         if self.mode == u"c":             # <<<<<<<<<<<<<<
  *             bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS
  *         elif self.mode == u"fortran":
  */
-  __pyx_t_1 = (__Pyx_PyUnicode_Equals(__pyx_v_self->mode, __pyx_n_u_c, Py_EQ)); if (unlikely(__pyx_t_1 < 0)) __PYX_ERR(1, 187, __pyx_L1_error)
+  __pyx_t_1 = (__Pyx_PyUnicode_Equals(__pyx_v_self->mode, __pyx_n_u_c, Py_EQ)); if (unlikely(__pyx_t_1 < 0)) __PYX_ERR(1, 188, __pyx_L1_error)
   __pyx_t_2 = (__pyx_t_1 != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":188
+    /* "View.MemoryView":189
  *         cdef int bufmode = -1
  *         if self.mode == u"c":
  *             bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS             # <<<<<<<<<<<<<<
@@ -9499,7 +9543,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
  */
     __pyx_v_bufmode = (PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS);
 
-    /* "View.MemoryView":187
+    /* "View.MemoryView":188
  *     def __getbuffer__(self, Py_buffer *info, int flags):
  *         cdef int bufmode = -1
  *         if self.mode == u"c":             # <<<<<<<<<<<<<<
@@ -9509,18 +9553,18 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
     goto __pyx_L3;
   }
 
-  /* "View.MemoryView":189
+  /* "View.MemoryView":190
  *         if self.mode == u"c":
  *             bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS
  *         elif self.mode == u"fortran":             # <<<<<<<<<<<<<<
  *             bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS
  *         if not (flags & bufmode):
  */
-  __pyx_t_2 = (__Pyx_PyUnicode_Equals(__pyx_v_self->mode, __pyx_n_u_fortran, Py_EQ)); if (unlikely(__pyx_t_2 < 0)) __PYX_ERR(1, 189, __pyx_L1_error)
+  __pyx_t_2 = (__Pyx_PyUnicode_Equals(__pyx_v_self->mode, __pyx_n_u_fortran, Py_EQ)); if (unlikely(__pyx_t_2 < 0)) __PYX_ERR(1, 190, __pyx_L1_error)
   __pyx_t_1 = (__pyx_t_2 != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":190
+    /* "View.MemoryView":191
  *             bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS
  *         elif self.mode == u"fortran":
  *             bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS             # <<<<<<<<<<<<<<
@@ -9529,7 +9573,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
  */
     __pyx_v_bufmode = (PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS);
 
-    /* "View.MemoryView":189
+    /* "View.MemoryView":190
  *         if self.mode == u"c":
  *             bufmode = PyBUF_C_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS
  *         elif self.mode == u"fortran":             # <<<<<<<<<<<<<<
@@ -9539,7 +9583,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
   }
   __pyx_L3:;
 
-  /* "View.MemoryView":191
+  /* "View.MemoryView":192
  *         elif self.mode == u"fortran":
  *             bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS
  *         if not (flags & bufmode):             # <<<<<<<<<<<<<<
@@ -9549,20 +9593,20 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
   __pyx_t_1 = ((!((__pyx_v_flags & __pyx_v_bufmode) != 0)) != 0);
   if (unlikely(__pyx_t_1)) {
 
-    /* "View.MemoryView":192
+    /* "View.MemoryView":193
  *             bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS
  *         if not (flags & bufmode):
  *             raise ValueError("Can only create a buffer that is contiguous in memory.")             # <<<<<<<<<<<<<<
  *         info.buf = self.data
  *         info.len = self.len
  */
-    __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__20, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 192, __pyx_L1_error)
+    __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__20, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 193, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
     __Pyx_Raise(__pyx_t_3, 0, 0, 0);
     __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
-    __PYX_ERR(1, 192, __pyx_L1_error)
+    __PYX_ERR(1, 193, __pyx_L1_error)
 
-    /* "View.MemoryView":191
+    /* "View.MemoryView":192
  *         elif self.mode == u"fortran":
  *             bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS
  *         if not (flags & bufmode):             # <<<<<<<<<<<<<<
@@ -9571,7 +9615,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
  */
   }
 
-  /* "View.MemoryView":193
+  /* "View.MemoryView":194
  *         if not (flags & bufmode):
  *             raise ValueError("Can only create a buffer that is contiguous in memory.")
  *         info.buf = self.data             # <<<<<<<<<<<<<<
@@ -9581,7 +9625,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
   __pyx_t_4 = __pyx_v_self->data;
   __pyx_v_info->buf = __pyx_t_4;
 
-  /* "View.MemoryView":194
+  /* "View.MemoryView":195
  *             raise ValueError("Can only create a buffer that is contiguous in memory.")
  *         info.buf = self.data
  *         info.len = self.len             # <<<<<<<<<<<<<<
@@ -9591,7 +9635,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
   __pyx_t_5 = __pyx_v_self->len;
   __pyx_v_info->len = __pyx_t_5;
 
-  /* "View.MemoryView":195
+  /* "View.MemoryView":196
  *         info.buf = self.data
  *         info.len = self.len
  *         info.ndim = self.ndim             # <<<<<<<<<<<<<<
@@ -9601,7 +9645,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
   __pyx_t_6 = __pyx_v_self->ndim;
   __pyx_v_info->ndim = __pyx_t_6;
 
-  /* "View.MemoryView":196
+  /* "View.MemoryView":197
  *         info.len = self.len
  *         info.ndim = self.ndim
  *         info.shape = self._shape             # <<<<<<<<<<<<<<
@@ -9611,7 +9655,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
   __pyx_t_7 = __pyx_v_self->_shape;
   __pyx_v_info->shape = __pyx_t_7;
 
-  /* "View.MemoryView":197
+  /* "View.MemoryView":198
  *         info.ndim = self.ndim
  *         info.shape = self._shape
  *         info.strides = self._strides             # <<<<<<<<<<<<<<
@@ -9621,7 +9665,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
   __pyx_t_7 = __pyx_v_self->_strides;
   __pyx_v_info->strides = __pyx_t_7;
 
-  /* "View.MemoryView":198
+  /* "View.MemoryView":199
  *         info.shape = self._shape
  *         info.strides = self._strides
  *         info.suboffsets = NULL             # <<<<<<<<<<<<<<
@@ -9630,7 +9674,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
  */
   __pyx_v_info->suboffsets = NULL;
 
-  /* "View.MemoryView":199
+  /* "View.MemoryView":200
  *         info.strides = self._strides
  *         info.suboffsets = NULL
  *         info.itemsize = self.itemsize             # <<<<<<<<<<<<<<
@@ -9640,7 +9684,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
   __pyx_t_5 = __pyx_v_self->itemsize;
   __pyx_v_info->itemsize = __pyx_t_5;
 
-  /* "View.MemoryView":200
+  /* "View.MemoryView":201
  *         info.suboffsets = NULL
  *         info.itemsize = self.itemsize
  *         info.readonly = 0             # <<<<<<<<<<<<<<
@@ -9649,7 +9693,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
  */
   __pyx_v_info->readonly = 0;
 
-  /* "View.MemoryView":202
+  /* "View.MemoryView":203
  *         info.readonly = 0
  * 
  *         if flags & PyBUF_FORMAT:             # <<<<<<<<<<<<<<
@@ -9659,7 +9703,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
   __pyx_t_1 = ((__pyx_v_flags & PyBUF_FORMAT) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":203
+    /* "View.MemoryView":204
  * 
  *         if flags & PyBUF_FORMAT:
  *             info.format = self.format             # <<<<<<<<<<<<<<
@@ -9669,7 +9713,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
     __pyx_t_4 = __pyx_v_self->format;
     __pyx_v_info->format = __pyx_t_4;
 
-    /* "View.MemoryView":202
+    /* "View.MemoryView":203
  *         info.readonly = 0
  * 
  *         if flags & PyBUF_FORMAT:             # <<<<<<<<<<<<<<
@@ -9679,7 +9723,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
     goto __pyx_L5;
   }
 
-  /* "View.MemoryView":205
+  /* "View.MemoryView":206
  *             info.format = self.format
  *         else:
  *             info.format = NULL             # <<<<<<<<<<<<<<
@@ -9691,7 +9735,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
   }
   __pyx_L5:;
 
-  /* "View.MemoryView":207
+  /* "View.MemoryView":208
  *             info.format = NULL
  * 
  *         info.obj = self             # <<<<<<<<<<<<<<
@@ -9704,7 +9748,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
   __Pyx_DECREF(__pyx_v_info->obj);
   __pyx_v_info->obj = ((PyObject *)__pyx_v_self);
 
-  /* "View.MemoryView":185
+  /* "View.MemoryView":186
  * 
  *     @cname('getbuffer')
  *     def __getbuffer__(self, Py_buffer *info, int flags):             # <<<<<<<<<<<<<<
@@ -9734,7 +9778,7 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_2__getbuffer__(stru
   return __pyx_r;
 }
 
-/* "View.MemoryView":211
+/* "View.MemoryView":212
  *     __pyx_getbuffer = capsule(<void *> &__pyx_array_getbuffer, "getbuffer(obj, view, flags)")
  * 
  *     def __dealloc__(array self):             # <<<<<<<<<<<<<<
@@ -9758,7 +9802,7 @@ static void __pyx_array___pyx_pf_15View_dot_MemoryView_5array_4__dealloc__(struc
   int __pyx_t_1;
   __Pyx_RefNannySetupContext("__dealloc__", 0);
 
-  /* "View.MemoryView":212
+  /* "View.MemoryView":213
  * 
  *     def __dealloc__(array self):
  *         if self.callback_free_data != NULL:             # <<<<<<<<<<<<<<
@@ -9768,7 +9812,7 @@ static void __pyx_array___pyx_pf_15View_dot_MemoryView_5array_4__dealloc__(struc
   __pyx_t_1 = ((__pyx_v_self->callback_free_data != NULL) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":213
+    /* "View.MemoryView":214
  *     def __dealloc__(array self):
  *         if self.callback_free_data != NULL:
  *             self.callback_free_data(self.data)             # <<<<<<<<<<<<<<
@@ -9777,7 +9821,7 @@ static void __pyx_array___pyx_pf_15View_dot_MemoryView_5array_4__dealloc__(struc
  */
     __pyx_v_self->callback_free_data(__pyx_v_self->data);
 
-    /* "View.MemoryView":212
+    /* "View.MemoryView":213
  * 
  *     def __dealloc__(array self):
  *         if self.callback_free_data != NULL:             # <<<<<<<<<<<<<<
@@ -9787,7 +9831,7 @@ static void __pyx_array___pyx_pf_15View_dot_MemoryView_5array_4__dealloc__(struc
     goto __pyx_L3;
   }
 
-  /* "View.MemoryView":214
+  /* "View.MemoryView":215
  *         if self.callback_free_data != NULL:
  *             self.callback_free_data(self.data)
  *         elif self.free_data:             # <<<<<<<<<<<<<<
@@ -9797,7 +9841,7 @@ static void __pyx_array___pyx_pf_15View_dot_MemoryView_5array_4__dealloc__(struc
   __pyx_t_1 = (__pyx_v_self->free_data != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":215
+    /* "View.MemoryView":216
  *             self.callback_free_data(self.data)
  *         elif self.free_data:
  *             if self.dtype_is_object:             # <<<<<<<<<<<<<<
@@ -9807,7 +9851,7 @@ static void __pyx_array___pyx_pf_15View_dot_MemoryView_5array_4__dealloc__(struc
     __pyx_t_1 = (__pyx_v_self->dtype_is_object != 0);
     if (__pyx_t_1) {
 
-      /* "View.MemoryView":216
+      /* "View.MemoryView":217
  *         elif self.free_data:
  *             if self.dtype_is_object:
  *                 refcount_objects_in_slice(self.data, self._shape,             # <<<<<<<<<<<<<<
@@ -9816,7 +9860,7 @@ static void __pyx_array___pyx_pf_15View_dot_MemoryView_5array_4__dealloc__(struc
  */
       __pyx_memoryview_refcount_objects_in_slice(__pyx_v_self->data, __pyx_v_self->_shape, __pyx_v_self->_strides, __pyx_v_self->ndim, 0);
 
-      /* "View.MemoryView":215
+      /* "View.MemoryView":216
  *             self.callback_free_data(self.data)
  *         elif self.free_data:
  *             if self.dtype_is_object:             # <<<<<<<<<<<<<<
@@ -9825,7 +9869,7 @@ static void __pyx_array___pyx_pf_15View_dot_MemoryView_5array_4__dealloc__(struc
  */
     }
 
-    /* "View.MemoryView":218
+    /* "View.MemoryView":219
  *                 refcount_objects_in_slice(self.data, self._shape,
  *                                           self._strides, self.ndim, False)
  *             free(self.data)             # <<<<<<<<<<<<<<
@@ -9834,7 +9878,7 @@ static void __pyx_array___pyx_pf_15View_dot_MemoryView_5array_4__dealloc__(struc
  */
     free(__pyx_v_self->data);
 
-    /* "View.MemoryView":214
+    /* "View.MemoryView":215
  *         if self.callback_free_data != NULL:
  *             self.callback_free_data(self.data)
  *         elif self.free_data:             # <<<<<<<<<<<<<<
@@ -9844,7 +9888,7 @@ static void __pyx_array___pyx_pf_15View_dot_MemoryView_5array_4__dealloc__(struc
   }
   __pyx_L3:;
 
-  /* "View.MemoryView":219
+  /* "View.MemoryView":220
  *                                           self._strides, self.ndim, False)
  *             free(self.data)
  *         PyObject_Free(self._shape)             # <<<<<<<<<<<<<<
@@ -9853,7 +9897,7 @@ static void __pyx_array___pyx_pf_15View_dot_MemoryView_5array_4__dealloc__(struc
  */
   PyObject_Free(__pyx_v_self->_shape);
 
-  /* "View.MemoryView":211
+  /* "View.MemoryView":212
  *     __pyx_getbuffer = capsule(<void *> &__pyx_array_getbuffer, "getbuffer(obj, view, flags)")
  * 
  *     def __dealloc__(array self):             # <<<<<<<<<<<<<<
@@ -9865,7 +9909,7 @@ static void __pyx_array___pyx_pf_15View_dot_MemoryView_5array_4__dealloc__(struc
   __Pyx_RefNannyFinishContext();
 }
 
-/* "View.MemoryView":222
+/* "View.MemoryView":223
  * 
  *     @property
  *     def memview(self):             # <<<<<<<<<<<<<<
@@ -9895,7 +9939,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_5array_7memview___get__(struct _
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("__get__", 0);
 
-  /* "View.MemoryView":223
+  /* "View.MemoryView":224
  *     @property
  *     def memview(self):
  *         return self.get_memview()             # <<<<<<<<<<<<<<
@@ -9903,13 +9947,13 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_5array_7memview___get__(struct _
  *     @cname('get_memview')
  */
   __Pyx_XDECREF(__pyx_r);
-  __pyx_t_1 = ((struct __pyx_vtabstruct_array *)__pyx_v_self->__pyx_vtab)->get_memview(__pyx_v_self); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 223, __pyx_L1_error)
+  __pyx_t_1 = ((struct __pyx_vtabstruct_array *)__pyx_v_self->__pyx_vtab)->get_memview(__pyx_v_self); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 224, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
   __pyx_r = __pyx_t_1;
   __pyx_t_1 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":222
+  /* "View.MemoryView":223
  * 
  *     @property
  *     def memview(self):             # <<<<<<<<<<<<<<
@@ -9928,7 +9972,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_5array_7memview___get__(struct _
   return __pyx_r;
 }
 
-/* "View.MemoryView":226
+/* "View.MemoryView":227
  * 
  *     @cname('get_memview')
  *     cdef get_memview(self):             # <<<<<<<<<<<<<<
@@ -9948,7 +9992,7 @@ static PyObject *__pyx_array_get_memview(struct __pyx_array_obj *__pyx_v_self) {
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("get_memview", 0);
 
-  /* "View.MemoryView":227
+  /* "View.MemoryView":228
  *     @cname('get_memview')
  *     cdef get_memview(self):
  *         flags =  PyBUF_ANY_CONTIGUOUS|PyBUF_FORMAT|PyBUF_WRITABLE             # <<<<<<<<<<<<<<
@@ -9957,7 +10001,7 @@ static PyObject *__pyx_array_get_memview(struct __pyx_array_obj *__pyx_v_self) {
  */
   __pyx_v_flags = ((PyBUF_ANY_CONTIGUOUS | PyBUF_FORMAT) | PyBUF_WRITABLE);
 
-  /* "View.MemoryView":228
+  /* "View.MemoryView":229
  *     cdef get_memview(self):
  *         flags =  PyBUF_ANY_CONTIGUOUS|PyBUF_FORMAT|PyBUF_WRITABLE
  *         return  memoryview(self, flags, self.dtype_is_object)             # <<<<<<<<<<<<<<
@@ -9965,11 +10009,11 @@ static PyObject *__pyx_array_get_memview(struct __pyx_array_obj *__pyx_v_self) {
  *     def __len__(self):
  */
   __Pyx_XDECREF(__pyx_r);
-  __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_flags); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 228, __pyx_L1_error)
+  __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_flags); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 229, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
-  __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_v_self->dtype_is_object); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 228, __pyx_L1_error)
+  __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_v_self->dtype_is_object); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 229, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
-  __pyx_t_3 = PyTuple_New(3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 228, __pyx_L1_error)
+  __pyx_t_3 = PyTuple_New(3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 229, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_3);
   __Pyx_INCREF(((PyObject *)__pyx_v_self));
   __Pyx_GIVEREF(((PyObject *)__pyx_v_self));
@@ -9980,14 +10024,14 @@ static PyObject *__pyx_array_get_memview(struct __pyx_array_obj *__pyx_v_self) {
   PyTuple_SET_ITEM(__pyx_t_3, 2, __pyx_t_2);
   __pyx_t_1 = 0;
   __pyx_t_2 = 0;
-  __pyx_t_2 = __Pyx_PyObject_Call(((PyObject *)__pyx_memoryview_type), __pyx_t_3, NULL); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 228, __pyx_L1_error)
+  __pyx_t_2 = __Pyx_PyObject_Call(((PyObject *)__pyx_memoryview_type), __pyx_t_3, NULL); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 229, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
   __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
   __pyx_r = __pyx_t_2;
   __pyx_t_2 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":226
+  /* "View.MemoryView":227
  * 
  *     @cname('get_memview')
  *     cdef get_memview(self):             # <<<<<<<<<<<<<<
@@ -10008,7 +10052,7 @@ static PyObject *__pyx_array_get_memview(struct __pyx_array_obj *__pyx_v_self) {
   return __pyx_r;
 }
 
-/* "View.MemoryView":230
+/* "View.MemoryView":231
  *         return  memoryview(self, flags, self.dtype_is_object)
  * 
  *     def __len__(self):             # <<<<<<<<<<<<<<
@@ -10034,7 +10078,7 @@ static Py_ssize_t __pyx_array___pyx_pf_15View_dot_MemoryView_5array_6__len__(str
   __Pyx_RefNannyDeclarations
   __Pyx_RefNannySetupContext("__len__", 0);
 
-  /* "View.MemoryView":231
+  /* "View.MemoryView":232
  * 
  *     def __len__(self):
  *         return self._shape[0]             # <<<<<<<<<<<<<<
@@ -10044,7 +10088,7 @@ static Py_ssize_t __pyx_array___pyx_pf_15View_dot_MemoryView_5array_6__len__(str
   __pyx_r = (__pyx_v_self->_shape[0]);
   goto __pyx_L0;
 
-  /* "View.MemoryView":230
+  /* "View.MemoryView":231
  *         return  memoryview(self, flags, self.dtype_is_object)
  * 
  *     def __len__(self):             # <<<<<<<<<<<<<<
@@ -10058,7 +10102,7 @@ static Py_ssize_t __pyx_array___pyx_pf_15View_dot_MemoryView_5array_6__len__(str
   return __pyx_r;
 }
 
-/* "View.MemoryView":233
+/* "View.MemoryView":234
  *         return self._shape[0]
  * 
  *     def __getattr__(self, attr):             # <<<<<<<<<<<<<<
@@ -10089,7 +10133,7 @@ static PyObject *__pyx_array___pyx_pf_15View_dot_MemoryView_5array_8__getattr__(
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("__getattr__", 0);
 
-  /* "View.MemoryView":234
+  /* "View.MemoryView":235
  * 
  *     def __getattr__(self, attr):
  *         return getattr(self.memview, attr)             # <<<<<<<<<<<<<<
@@ -10097,16 +10141,16 @@ static PyObject *__pyx_array___pyx_pf_15View_dot_MemoryView_5array_8__getattr__(
  *     def __getitem__(self, item):
  */
   __Pyx_XDECREF(__pyx_r);
-  __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_memview); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 234, __pyx_L1_error)
+  __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_memview); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 235, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
-  __pyx_t_2 = __Pyx_GetAttr(__pyx_t_1, __pyx_v_attr); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 234, __pyx_L1_error)
+  __pyx_t_2 = __Pyx_GetAttr(__pyx_t_1, __pyx_v_attr); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 235, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
   __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
   __pyx_r = __pyx_t_2;
   __pyx_t_2 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":233
+  /* "View.MemoryView":234
  *         return self._shape[0]
  * 
  *     def __getattr__(self, attr):             # <<<<<<<<<<<<<<
@@ -10126,7 +10170,7 @@ static PyObject *__pyx_array___pyx_pf_15View_dot_MemoryView_5array_8__getattr__(
   return __pyx_r;
 }
 
-/* "View.MemoryView":236
+/* "View.MemoryView":237
  *         return getattr(self.memview, attr)
  * 
  *     def __getitem__(self, item):             # <<<<<<<<<<<<<<
@@ -10157,7 +10201,7 @@ static PyObject *__pyx_array___pyx_pf_15View_dot_MemoryView_5array_10__getitem__
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("__getitem__", 0);
 
-  /* "View.MemoryView":237
+  /* "View.MemoryView":238
  * 
  *     def __getitem__(self, item):
  *         return self.memview[item]             # <<<<<<<<<<<<<<
@@ -10165,16 +10209,16 @@ static PyObject *__pyx_array___pyx_pf_15View_dot_MemoryView_5array_10__getitem__
  *     def __setitem__(self, item, value):
  */
   __Pyx_XDECREF(__pyx_r);
-  __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_memview); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 237, __pyx_L1_error)
+  __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_memview); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 238, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
-  __pyx_t_2 = __Pyx_PyObject_GetItem(__pyx_t_1, __pyx_v_item); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 237, __pyx_L1_error)
+  __pyx_t_2 = __Pyx_PyObject_GetItem(__pyx_t_1, __pyx_v_item); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 238, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
   __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
   __pyx_r = __pyx_t_2;
   __pyx_t_2 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":236
+  /* "View.MemoryView":237
  *         return getattr(self.memview, attr)
  * 
  *     def __getitem__(self, item):             # <<<<<<<<<<<<<<
@@ -10194,7 +10238,7 @@ static PyObject *__pyx_array___pyx_pf_15View_dot_MemoryView_5array_10__getitem__
   return __pyx_r;
 }
 
-/* "View.MemoryView":239
+/* "View.MemoryView":240
  *         return self.memview[item]
  * 
  *     def __setitem__(self, item, value):             # <<<<<<<<<<<<<<
@@ -10224,19 +10268,19 @@ static int __pyx_array___pyx_pf_15View_dot_MemoryView_5array_12__setitem__(struc
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("__setitem__", 0);
 
-  /* "View.MemoryView":240
+  /* "View.MemoryView":241
  * 
  *     def __setitem__(self, item, value):
  *         self.memview[item] = value             # <<<<<<<<<<<<<<
  * 
  * 
  */
-  __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_memview); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 240, __pyx_L1_error)
+  __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_memview); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 241, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
-  if (unlikely(PyObject_SetItem(__pyx_t_1, __pyx_v_item, __pyx_v_value) < 0)) __PYX_ERR(1, 240, __pyx_L1_error)
+  if (unlikely(PyObject_SetItem(__pyx_t_1, __pyx_v_item, __pyx_v_value) < 0)) __PYX_ERR(1, 241, __pyx_L1_error)
   __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
 
-  /* "View.MemoryView":239
+  /* "View.MemoryView":240
  *         return self.memview[item]
  * 
  *     def __setitem__(self, item, value):             # <<<<<<<<<<<<<<
@@ -10369,7 +10413,7 @@ static PyObject *__pyx_pf___pyx_array_2__setstate_cython__(CYTHON_UNUSED struct
   return __pyx_r;
 }
 
-/* "View.MemoryView":244
+/* "View.MemoryView":245
  * 
  * @cname("__pyx_array_new")
  * cdef array array_cwrapper(tuple shape, Py_ssize_t itemsize, char *format,             # <<<<<<<<<<<<<<
@@ -10391,7 +10435,7 @@ static struct __pyx_array_obj *__pyx_array_new(PyObject *__pyx_v_shape, Py_ssize
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("array_cwrapper", 0);
 
-  /* "View.MemoryView":248
+  /* "View.MemoryView":249
  *     cdef array result
  * 
  *     if buf == NULL:             # <<<<<<<<<<<<<<
@@ -10401,20 +10445,20 @@ static struct __pyx_array_obj *__pyx_array_new(PyObject *__pyx_v_shape, Py_ssize
   __pyx_t_1 = ((__pyx_v_buf == NULL) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":249
+    /* "View.MemoryView":250
  * 
  *     if buf == NULL:
  *         result = array(shape, itemsize, format, mode.decode('ASCII'))             # <<<<<<<<<<<<<<
  *     else:
  *         result = array(shape, itemsize, format, mode.decode('ASCII'),
  */
-    __pyx_t_2 = PyInt_FromSsize_t(__pyx_v_itemsize); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 249, __pyx_L1_error)
+    __pyx_t_2 = PyInt_FromSsize_t(__pyx_v_itemsize); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 250, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_2);
-    __pyx_t_3 = __Pyx_PyBytes_FromString(__pyx_v_format); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 249, __pyx_L1_error)
+    __pyx_t_3 = __Pyx_PyBytes_FromString(__pyx_v_format); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 250, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
-    __pyx_t_4 = __Pyx_decode_c_string(__pyx_v_mode, 0, strlen(__pyx_v_mode), NULL, NULL, PyUnicode_DecodeASCII); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 249, __pyx_L1_error)
+    __pyx_t_4 = __Pyx_decode_c_string(__pyx_v_mode, 0, strlen(__pyx_v_mode), NULL, NULL, PyUnicode_DecodeASCII); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 250, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_4);
-    __pyx_t_5 = PyTuple_New(4); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 249, __pyx_L1_error)
+    __pyx_t_5 = PyTuple_New(4); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 250, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_5);
     __Pyx_INCREF(__pyx_v_shape);
     __Pyx_GIVEREF(__pyx_v_shape);
@@ -10428,13 +10472,13 @@ static struct __pyx_array_obj *__pyx_array_new(PyObject *__pyx_v_shape, Py_ssize
     __pyx_t_2 = 0;
     __pyx_t_3 = 0;
     __pyx_t_4 = 0;
-    __pyx_t_4 = __Pyx_PyObject_Call(((PyObject *)__pyx_array_type), __pyx_t_5, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 249, __pyx_L1_error)
+    __pyx_t_4 = __Pyx_PyObject_Call(((PyObject *)__pyx_array_type), __pyx_t_5, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 250, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_4);
     __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
     __pyx_v_result = ((struct __pyx_array_obj *)__pyx_t_4);
     __pyx_t_4 = 0;
 
-    /* "View.MemoryView":248
+    /* "View.MemoryView":249
  *     cdef array result
  * 
  *     if buf == NULL:             # <<<<<<<<<<<<<<
@@ -10444,7 +10488,7 @@ static struct __pyx_array_obj *__pyx_array_new(PyObject *__pyx_v_shape, Py_ssize
     goto __pyx_L3;
   }
 
-  /* "View.MemoryView":251
+  /* "View.MemoryView":252
  *         result = array(shape, itemsize, format, mode.decode('ASCII'))
  *     else:
  *         result = array(shape, itemsize, format, mode.decode('ASCII'),             # <<<<<<<<<<<<<<
@@ -10452,13 +10496,13 @@ static struct __pyx_array_obj *__pyx_array_new(PyObject *__pyx_v_shape, Py_ssize
  *         result.data = buf
  */
   /*else*/ {
-    __pyx_t_4 = PyInt_FromSsize_t(__pyx_v_itemsize); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 251, __pyx_L1_error)
+    __pyx_t_4 = PyInt_FromSsize_t(__pyx_v_itemsize); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 252, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_4);
-    __pyx_t_5 = __Pyx_PyBytes_FromString(__pyx_v_format); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 251, __pyx_L1_error)
+    __pyx_t_5 = __Pyx_PyBytes_FromString(__pyx_v_format); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 252, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_5);
-    __pyx_t_3 = __Pyx_decode_c_string(__pyx_v_mode, 0, strlen(__pyx_v_mode), NULL, NULL, PyUnicode_DecodeASCII); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 251, __pyx_L1_error)
+    __pyx_t_3 = __Pyx_decode_c_string(__pyx_v_mode, 0, strlen(__pyx_v_mode), NULL, NULL, PyUnicode_DecodeASCII); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 252, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
-    __pyx_t_2 = PyTuple_New(4); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 251, __pyx_L1_error)
+    __pyx_t_2 = PyTuple_New(4); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 252, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_2);
     __Pyx_INCREF(__pyx_v_shape);
     __Pyx_GIVEREF(__pyx_v_shape);
@@ -10473,32 +10517,32 @@ static struct __pyx_array_obj *__pyx_array_new(PyObject *__pyx_v_shape, Py_ssize
     __pyx_t_5 = 0;
     __pyx_t_3 = 0;
 
-    /* "View.MemoryView":252
+    /* "View.MemoryView":253
  *     else:
  *         result = array(shape, itemsize, format, mode.decode('ASCII'),
  *                        allocate_buffer=False)             # <<<<<<<<<<<<<<
  *         result.data = buf
  * 
  */
-    __pyx_t_3 = __Pyx_PyDict_NewPresized(1); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 252, __pyx_L1_error)
+    __pyx_t_3 = __Pyx_PyDict_NewPresized(1); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 253, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
-    if (PyDict_SetItem(__pyx_t_3, __pyx_n_s_allocate_buffer, Py_False) < 0) __PYX_ERR(1, 252, __pyx_L1_error)
+    if (PyDict_SetItem(__pyx_t_3, __pyx_n_s_allocate_buffer, Py_False) < 0) __PYX_ERR(1, 253, __pyx_L1_error)
 
-    /* "View.MemoryView":251
+    /* "View.MemoryView":252
  *         result = array(shape, itemsize, format, mode.decode('ASCII'))
  *     else:
  *         result = array(shape, itemsize, format, mode.decode('ASCII'),             # <<<<<<<<<<<<<<
  *                        allocate_buffer=False)
  *         result.data = buf
  */
-    __pyx_t_5 = __Pyx_PyObject_Call(((PyObject *)__pyx_array_type), __pyx_t_2, __pyx_t_3); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 251, __pyx_L1_error)
+    __pyx_t_5 = __Pyx_PyObject_Call(((PyObject *)__pyx_array_type), __pyx_t_2, __pyx_t_3); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 252, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_5);
     __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
     __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
     __pyx_v_result = ((struct __pyx_array_obj *)__pyx_t_5);
     __pyx_t_5 = 0;
 
-    /* "View.MemoryView":253
+    /* "View.MemoryView":254
  *         result = array(shape, itemsize, format, mode.decode('ASCII'),
  *                        allocate_buffer=False)
  *         result.data = buf             # <<<<<<<<<<<<<<
@@ -10509,7 +10553,7 @@ static struct __pyx_array_obj *__pyx_array_new(PyObject *__pyx_v_shape, Py_ssize
   }
   __pyx_L3:;
 
-  /* "View.MemoryView":255
+  /* "View.MemoryView":256
  *         result.data = buf
  * 
  *     return result             # <<<<<<<<<<<<<<
@@ -10521,7 +10565,7 @@ static struct __pyx_array_obj *__pyx_array_new(PyObject *__pyx_v_shape, Py_ssize
   __pyx_r = __pyx_v_result;
   goto __pyx_L0;
 
-  /* "View.MemoryView":244
+  /* "View.MemoryView":245
  * 
  * @cname("__pyx_array_new")
  * cdef array array_cwrapper(tuple shape, Py_ssize_t itemsize, char *format,             # <<<<<<<<<<<<<<
@@ -10544,7 +10588,7 @@ static struct __pyx_array_obj *__pyx_array_new(PyObject *__pyx_v_shape, Py_ssize
   return __pyx_r;
 }
 
-/* "View.MemoryView":281
+/* "View.MemoryView":282
  * cdef class Enum(object):
  *     cdef object name
  *     def __init__(self, name):             # <<<<<<<<<<<<<<
@@ -10581,7 +10625,7 @@ static int __pyx_MemviewEnum___init__(PyObject *__pyx_v_self, PyObject *__pyx_ar
         else goto __pyx_L5_argtuple_error;
       }
       if (unlikely(kw_args > 0)) {
-        if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_pyargnames, 0, values, pos_args, "__init__") < 0)) __PYX_ERR(1, 281, __pyx_L3_error)
+        if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_pyargnames, 0, values, pos_args, "__init__") < 0)) __PYX_ERR(1, 282, __pyx_L3_error)
       }
     } else if (PyTuple_GET_SIZE(__pyx_args) != 1) {
       goto __pyx_L5_argtuple_error;
@@ -10592,7 +10636,7 @@ static int __pyx_MemviewEnum___init__(PyObject *__pyx_v_self, PyObject *__pyx_ar
   }
   goto __pyx_L4_argument_unpacking_done;
   __pyx_L5_argtuple_error:;
-  __Pyx_RaiseArgtupleInvalid("__init__", 1, 1, 1, PyTuple_GET_SIZE(__pyx_args)); __PYX_ERR(1, 281, __pyx_L3_error)
+  __Pyx_RaiseArgtupleInvalid("__init__", 1, 1, 1, PyTuple_GET_SIZE(__pyx_args)); __PYX_ERR(1, 282, __pyx_L3_error)
   __pyx_L3_error:;
   __Pyx_AddTraceback("View.MemoryView.Enum.__init__", __pyx_clineno, __pyx_lineno, __pyx_filename);
   __Pyx_RefNannyFinishContext();
@@ -10610,7 +10654,7 @@ static int __pyx_MemviewEnum___pyx_pf_15View_dot_MemoryView_4Enum___init__(struc
   __Pyx_RefNannyDeclarations
   __Pyx_RefNannySetupContext("__init__", 0);
 
-  /* "View.MemoryView":282
+  /* "View.MemoryView":283
  *     cdef object name
  *     def __init__(self, name):
  *         self.name = name             # <<<<<<<<<<<<<<
@@ -10623,7 +10667,7 @@ static int __pyx_MemviewEnum___pyx_pf_15View_dot_MemoryView_4Enum___init__(struc
   __Pyx_DECREF(__pyx_v_self->name);
   __pyx_v_self->name = __pyx_v_name;
 
-  /* "View.MemoryView":281
+  /* "View.MemoryView":282
  * cdef class Enum(object):
  *     cdef object name
  *     def __init__(self, name):             # <<<<<<<<<<<<<<
@@ -10637,7 +10681,7 @@ static int __pyx_MemviewEnum___pyx_pf_15View_dot_MemoryView_4Enum___init__(struc
   return __pyx_r;
 }
 
-/* "View.MemoryView":283
+/* "View.MemoryView":284
  *     def __init__(self, name):
  *         self.name = name
  *     def __repr__(self):             # <<<<<<<<<<<<<<
@@ -10663,7 +10707,7 @@ static PyObject *__pyx_MemviewEnum___pyx_pf_15View_dot_MemoryView_4Enum_2__repr_
   __Pyx_RefNannyDeclarations
   __Pyx_RefNannySetupContext("__repr__", 0);
 
-  /* "View.MemoryView":284
+  /* "View.MemoryView":285
  *         self.name = name
  *     def __repr__(self):
  *         return self.name             # <<<<<<<<<<<<<<
@@ -10675,7 +10719,7 @@ static PyObject *__pyx_MemviewEnum___pyx_pf_15View_dot_MemoryView_4Enum_2__repr_
   __pyx_r = __pyx_v_self->name;
   goto __pyx_L0;
 
-  /* "View.MemoryView":283
+  /* "View.MemoryView":284
  *     def __init__(self, name):
  *         self.name = name
  *     def __repr__(self):             # <<<<<<<<<<<<<<
@@ -10983,7 +11027,7 @@ static PyObject *__pyx_pf___pyx_MemviewEnum_2__setstate_cython__(struct __pyx_Me
   return __pyx_r;
 }
 
-/* "View.MemoryView":298
+/* "View.MemoryView":299
  * 
  * @cname('__pyx_align_pointer')
  * cdef void *align_pointer(void *memory, size_t alignment) nogil:             # <<<<<<<<<<<<<<
@@ -10997,7 +11041,7 @@ static void *__pyx_align_pointer(void *__pyx_v_memory, size_t __pyx_v_alignment)
   void *__pyx_r;
   int __pyx_t_1;
 
-  /* "View.MemoryView":300
+  /* "View.MemoryView":301
  * cdef void *align_pointer(void *memory, size_t alignment) nogil:
  *     "Align pointer memory on a given boundary"
  *     cdef Py_intptr_t aligned_p = <Py_intptr_t> memory             # <<<<<<<<<<<<<<
@@ -11006,7 +11050,7 @@ static void *__pyx_align_pointer(void *__pyx_v_memory, size_t __pyx_v_alignment)
  */
   __pyx_v_aligned_p = ((Py_intptr_t)__pyx_v_memory);
 
-  /* "View.MemoryView":304
+  /* "View.MemoryView":305
  * 
  *     with cython.cdivision(True):
  *         offset = aligned_p % alignment             # <<<<<<<<<<<<<<
@@ -11015,7 +11059,7 @@ static void *__pyx_align_pointer(void *__pyx_v_memory, size_t __pyx_v_alignment)
  */
   __pyx_v_offset = (__pyx_v_aligned_p % __pyx_v_alignment);
 
-  /* "View.MemoryView":306
+  /* "View.MemoryView":307
  *         offset = aligned_p % alignment
  * 
  *     if offset > 0:             # <<<<<<<<<<<<<<
@@ -11025,7 +11069,7 @@ static void *__pyx_align_pointer(void *__pyx_v_memory, size_t __pyx_v_alignment)
   __pyx_t_1 = ((__pyx_v_offset > 0) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":307
+    /* "View.MemoryView":308
  * 
  *     if offset > 0:
  *         aligned_p += alignment - offset             # <<<<<<<<<<<<<<
@@ -11034,7 +11078,7 @@ static void *__pyx_align_pointer(void *__pyx_v_memory, size_t __pyx_v_alignment)
  */
     __pyx_v_aligned_p = (__pyx_v_aligned_p + (__pyx_v_alignment - __pyx_v_offset));
 
-    /* "View.MemoryView":306
+    /* "View.MemoryView":307
  *         offset = aligned_p % alignment
  * 
  *     if offset > 0:             # <<<<<<<<<<<<<<
@@ -11043,7 +11087,7 @@ static void *__pyx_align_pointer(void *__pyx_v_memory, size_t __pyx_v_alignment)
  */
   }
 
-  /* "View.MemoryView":309
+  /* "View.MemoryView":310
  *         aligned_p += alignment - offset
  * 
  *     return <void *> aligned_p             # <<<<<<<<<<<<<<
@@ -11053,7 +11097,7 @@ static void *__pyx_align_pointer(void *__pyx_v_memory, size_t __pyx_v_alignment)
   __pyx_r = ((void *)__pyx_v_aligned_p);
   goto __pyx_L0;
 
-  /* "View.MemoryView":298
+  /* "View.MemoryView":299
  * 
  * @cname('__pyx_align_pointer')
  * cdef void *align_pointer(void *memory, size_t alignment) nogil:             # <<<<<<<<<<<<<<
@@ -11066,7 +11110,7 @@ static void *__pyx_align_pointer(void *__pyx_v_memory, size_t __pyx_v_alignment)
   return __pyx_r;
 }
 
-/* "View.MemoryView":345
+/* "View.MemoryView":346
  *     cdef __Pyx_TypeInfo *typeinfo
  * 
  *     def __cinit__(memoryview self, object obj, int flags, bint dtype_is_object=False):             # <<<<<<<<<<<<<<
@@ -11111,7 +11155,7 @@ static int __pyx_memoryview___cinit__(PyObject *__pyx_v_self, PyObject *__pyx_ar
         case  1:
         if (likely((values[1] = __Pyx_PyDict_GetItemStr(__pyx_kwds, __pyx_n_s_flags)) != 0)) kw_args--;
         else {
-          __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 2, 3, 1); __PYX_ERR(1, 345, __pyx_L3_error)
+          __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 2, 3, 1); __PYX_ERR(1, 346, __pyx_L3_error)
         }
         CYTHON_FALLTHROUGH;
         case  2:
@@ -11121,7 +11165,7 @@ static int __pyx_memoryview___cinit__(PyObject *__pyx_v_self, PyObject *__pyx_ar
         }
       }
       if (unlikely(kw_args > 0)) {
-        if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_pyargnames, 0, values, pos_args, "__cinit__") < 0)) __PYX_ERR(1, 345, __pyx_L3_error)
+        if (unlikely(__Pyx_ParseOptionalKeywords(__pyx_kwds, __pyx_pyargnames, 0, values, pos_args, "__cinit__") < 0)) __PYX_ERR(1, 346, __pyx_L3_error)
       }
     } else {
       switch (PyTuple_GET_SIZE(__pyx_args)) {
@@ -11134,16 +11178,16 @@ static int __pyx_memoryview___cinit__(PyObject *__pyx_v_self, PyObject *__pyx_ar
       }
     }
     __pyx_v_obj = values[0];
-    __pyx_v_flags = __Pyx_PyInt_As_int(values[1]); if (unlikely((__pyx_v_flags == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 345, __pyx_L3_error)
+    __pyx_v_flags = __Pyx_PyInt_As_int(values[1]); if (unlikely((__pyx_v_flags == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 346, __pyx_L3_error)
     if (values[2]) {
-      __pyx_v_dtype_is_object = __Pyx_PyObject_IsTrue(values[2]); if (unlikely((__pyx_v_dtype_is_object == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 345, __pyx_L3_error)
+      __pyx_v_dtype_is_object = __Pyx_PyObject_IsTrue(values[2]); if (unlikely((__pyx_v_dtype_is_object == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 346, __pyx_L3_error)
     } else {
       __pyx_v_dtype_is_object = ((int)0);
     }
   }
   goto __pyx_L4_argument_unpacking_done;
   __pyx_L5_argtuple_error:;
-  __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 2, 3, PyTuple_GET_SIZE(__pyx_args)); __PYX_ERR(1, 345, __pyx_L3_error)
+  __Pyx_RaiseArgtupleInvalid("__cinit__", 0, 2, 3, PyTuple_GET_SIZE(__pyx_args)); __PYX_ERR(1, 346, __pyx_L3_error)
   __pyx_L3_error:;
   __Pyx_AddTraceback("View.MemoryView.memoryview.__cinit__", __pyx_clineno, __pyx_lineno, __pyx_filename);
   __Pyx_RefNannyFinishContext();
@@ -11168,7 +11212,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit_
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("__cinit__", 0);
 
-  /* "View.MemoryView":346
+  /* "View.MemoryView":347
  * 
  *     def __cinit__(memoryview self, object obj, int flags, bint dtype_is_object=False):
  *         self.obj = obj             # <<<<<<<<<<<<<<
@@ -11181,7 +11225,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit_
   __Pyx_DECREF(__pyx_v_self->obj);
   __pyx_v_self->obj = __pyx_v_obj;
 
-  /* "View.MemoryView":347
+  /* "View.MemoryView":348
  *     def __cinit__(memoryview self, object obj, int flags, bint dtype_is_object=False):
  *         self.obj = obj
  *         self.flags = flags             # <<<<<<<<<<<<<<
@@ -11190,7 +11234,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit_
  */
   __pyx_v_self->flags = __pyx_v_flags;
 
-  /* "View.MemoryView":348
+  /* "View.MemoryView":349
  *         self.obj = obj
  *         self.flags = flags
  *         if type(self) is memoryview or obj is not None:             # <<<<<<<<<<<<<<
@@ -11210,16 +11254,16 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit_
   __pyx_L4_bool_binop_done:;
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":349
+    /* "View.MemoryView":350
  *         self.flags = flags
  *         if type(self) is memoryview or obj is not None:
  *             __Pyx_GetBuffer(obj, &self.view, flags)             # <<<<<<<<<<<<<<
  *             if <PyObject *> self.view.obj == NULL:
  *                 (<__pyx_buffer *> &self.view).obj = Py_None
  */
-    __pyx_t_4 = __Pyx_GetBuffer(__pyx_v_obj, (&__pyx_v_self->view), __pyx_v_flags); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(1, 349, __pyx_L1_error)
+    __pyx_t_4 = __Pyx_GetBuffer(__pyx_v_obj, (&__pyx_v_self->view), __pyx_v_flags); if (unlikely(__pyx_t_4 == ((int)-1))) __PYX_ERR(1, 350, __pyx_L1_error)
 
-    /* "View.MemoryView":350
+    /* "View.MemoryView":351
  *         if type(self) is memoryview or obj is not None:
  *             __Pyx_GetBuffer(obj, &self.view, flags)
  *             if <PyObject *> self.view.obj == NULL:             # <<<<<<<<<<<<<<
@@ -11229,7 +11273,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit_
     __pyx_t_1 = ((((PyObject *)__pyx_v_self->view.obj) == NULL) != 0);
     if (__pyx_t_1) {
 
-      /* "View.MemoryView":351
+      /* "View.MemoryView":352
  *             __Pyx_GetBuffer(obj, &self.view, flags)
  *             if <PyObject *> self.view.obj == NULL:
  *                 (<__pyx_buffer *> &self.view).obj = Py_None             # <<<<<<<<<<<<<<
@@ -11238,16 +11282,16 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit_
  */
       ((Py_buffer *)(&__pyx_v_self->view))->obj = Py_None;
 
-      /* "View.MemoryView":352
+      /* "View.MemoryView":353
  *             if <PyObject *> self.view.obj == NULL:
  *                 (<__pyx_buffer *> &self.view).obj = Py_None
  *                 Py_INCREF(Py_None)             # <<<<<<<<<<<<<<
  * 
- *         global __pyx_memoryview_thread_locks_used
+ *         if not __PYX_CYTHON_ATOMICS_ENABLED():
  */
       Py_INCREF(Py_None);
 
-      /* "View.MemoryView":350
+      /* "View.MemoryView":351
  *         if type(self) is memoryview or obj is not None:
  *             __Pyx_GetBuffer(obj, &self.view, flags)
  *             if <PyObject *> self.view.obj == NULL:             # <<<<<<<<<<<<<<
@@ -11256,7 +11300,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit_
  */
     }
 
-    /* "View.MemoryView":348
+    /* "View.MemoryView":349
  *         self.obj = obj
  *         self.flags = flags
  *         if type(self) is memoryview or obj is not None:             # <<<<<<<<<<<<<<
@@ -11266,100 +11310,119 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit_
   }
 
   /* "View.MemoryView":355
+ *                 Py_INCREF(Py_None)
  * 
- *         global __pyx_memoryview_thread_locks_used
- *         if __pyx_memoryview_thread_locks_used < THREAD_LOCKS_PREALLOCATED:             # <<<<<<<<<<<<<<
- *             self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]
- *             __pyx_memoryview_thread_locks_used += 1
+ *         if not __PYX_CYTHON_ATOMICS_ENABLED():             # <<<<<<<<<<<<<<
+ *             global __pyx_memoryview_thread_locks_used
+ *             if __pyx_memoryview_thread_locks_used < THREAD_LOCKS_PREALLOCATED:
  */
-  __pyx_t_1 = ((__pyx_memoryview_thread_locks_used < 8) != 0);
+  __pyx_t_1 = ((!(__PYX_CYTHON_ATOMICS_ENABLED() != 0)) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":356
- *         global __pyx_memoryview_thread_locks_used
- *         if __pyx_memoryview_thread_locks_used < THREAD_LOCKS_PREALLOCATED:
- *             self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]             # <<<<<<<<<<<<<<
- *             __pyx_memoryview_thread_locks_used += 1
- *         if self.lock is NULL:
- */
-    __pyx_v_self->lock = (__pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]);
-
     /* "View.MemoryView":357
- *         if __pyx_memoryview_thread_locks_used < THREAD_LOCKS_PREALLOCATED:
- *             self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]
- *             __pyx_memoryview_thread_locks_used += 1             # <<<<<<<<<<<<<<
- *         if self.lock is NULL:
- *             self.lock = PyThread_allocate_lock()
+ *         if not __PYX_CYTHON_ATOMICS_ENABLED():
+ *             global __pyx_memoryview_thread_locks_used
+ *             if __pyx_memoryview_thread_locks_used < THREAD_LOCKS_PREALLOCATED:             # <<<<<<<<<<<<<<
+ *                 self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]
+ *                 __pyx_memoryview_thread_locks_used += 1
  */
-    __pyx_memoryview_thread_locks_used = (__pyx_memoryview_thread_locks_used + 1);
+    __pyx_t_1 = ((__pyx_memoryview_thread_locks_used < 8) != 0);
+    if (__pyx_t_1) {
 
-    /* "View.MemoryView":355
- * 
- *         global __pyx_memoryview_thread_locks_used
- *         if __pyx_memoryview_thread_locks_used < THREAD_LOCKS_PREALLOCATED:             # <<<<<<<<<<<<<<
- *             self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]
- *             __pyx_memoryview_thread_locks_used += 1
+      /* "View.MemoryView":358
+ *             global __pyx_memoryview_thread_locks_used
+ *             if __pyx_memoryview_thread_locks_used < THREAD_LOCKS_PREALLOCATED:
+ *                 self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]             # <<<<<<<<<<<<<<
+ *                 __pyx_memoryview_thread_locks_used += 1
+ *             if self.lock is NULL:
  */
-  }
+      __pyx_v_self->lock = (__pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]);
 
-  /* "View.MemoryView":358
- *             self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]
- *             __pyx_memoryview_thread_locks_used += 1
- *         if self.lock is NULL:             # <<<<<<<<<<<<<<
- *             self.lock = PyThread_allocate_lock()
+      /* "View.MemoryView":359
+ *             if __pyx_memoryview_thread_locks_used < THREAD_LOCKS_PREALLOCATED:
+ *                 self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]
+ *                 __pyx_memoryview_thread_locks_used += 1             # <<<<<<<<<<<<<<
  *             if self.lock is NULL:
+ *                 self.lock = PyThread_allocate_lock()
  */
-  __pyx_t_1 = ((__pyx_v_self->lock == NULL) != 0);
-  if (__pyx_t_1) {
+      __pyx_memoryview_thread_locks_used = (__pyx_memoryview_thread_locks_used + 1);
 
-    /* "View.MemoryView":359
- *             __pyx_memoryview_thread_locks_used += 1
- *         if self.lock is NULL:
- *             self.lock = PyThread_allocate_lock()             # <<<<<<<<<<<<<<
- *             if self.lock is NULL:
- *                 raise MemoryError
+      /* "View.MemoryView":357
+ *         if not __PYX_CYTHON_ATOMICS_ENABLED():
+ *             global __pyx_memoryview_thread_locks_used
+ *             if __pyx_memoryview_thread_locks_used < THREAD_LOCKS_PREALLOCATED:             # <<<<<<<<<<<<<<
+ *                 self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]
+ *                 __pyx_memoryview_thread_locks_used += 1
  */
-    __pyx_v_self->lock = PyThread_allocate_lock();
+    }
 
     /* "View.MemoryView":360
- *         if self.lock is NULL:
- *             self.lock = PyThread_allocate_lock()
+ *                 self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]
+ *                 __pyx_memoryview_thread_locks_used += 1
  *             if self.lock is NULL:             # <<<<<<<<<<<<<<
- *                 raise MemoryError
- * 
+ *                 self.lock = PyThread_allocate_lock()
+ *                 if self.lock is NULL:
  */
     __pyx_t_1 = ((__pyx_v_self->lock == NULL) != 0);
-    if (unlikely(__pyx_t_1)) {
+    if (__pyx_t_1) {
 
       /* "View.MemoryView":361
- *             self.lock = PyThread_allocate_lock()
+ *                 __pyx_memoryview_thread_locks_used += 1
  *             if self.lock is NULL:
- *                 raise MemoryError             # <<<<<<<<<<<<<<
+ *                 self.lock = PyThread_allocate_lock()             # <<<<<<<<<<<<<<
+ *                 if self.lock is NULL:
+ *                     raise MemoryError
+ */
+      __pyx_v_self->lock = PyThread_allocate_lock();
+
+      /* "View.MemoryView":362
+ *             if self.lock is NULL:
+ *                 self.lock = PyThread_allocate_lock()
+ *                 if self.lock is NULL:             # <<<<<<<<<<<<<<
+ *                     raise MemoryError
+ * 
+ */
+      __pyx_t_1 = ((__pyx_v_self->lock == NULL) != 0);
+      if (unlikely(__pyx_t_1)) {
+
+        /* "View.MemoryView":363
+ *                 self.lock = PyThread_allocate_lock()
+ *                 if self.lock is NULL:
+ *                     raise MemoryError             # <<<<<<<<<<<<<<
  * 
  *         if flags & PyBUF_FORMAT:
  */
-      PyErr_NoMemory(); __PYX_ERR(1, 361, __pyx_L1_error)
+        PyErr_NoMemory(); __PYX_ERR(1, 363, __pyx_L1_error)
+
+        /* "View.MemoryView":362
+ *             if self.lock is NULL:
+ *                 self.lock = PyThread_allocate_lock()
+ *                 if self.lock is NULL:             # <<<<<<<<<<<<<<
+ *                     raise MemoryError
+ * 
+ */
+      }
 
       /* "View.MemoryView":360
- *         if self.lock is NULL:
- *             self.lock = PyThread_allocate_lock()
+ *                 self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]
+ *                 __pyx_memoryview_thread_locks_used += 1
  *             if self.lock is NULL:             # <<<<<<<<<<<<<<
- *                 raise MemoryError
- * 
+ *                 self.lock = PyThread_allocate_lock()
+ *                 if self.lock is NULL:
  */
     }
 
-    /* "View.MemoryView":358
- *             self.lock = __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]
- *             __pyx_memoryview_thread_locks_used += 1
- *         if self.lock is NULL:             # <<<<<<<<<<<<<<
- *             self.lock = PyThread_allocate_lock()
- *             if self.lock is NULL:
+    /* "View.MemoryView":355
+ *                 Py_INCREF(Py_None)
+ * 
+ *         if not __PYX_CYTHON_ATOMICS_ENABLED():             # <<<<<<<<<<<<<<
+ *             global __pyx_memoryview_thread_locks_used
+ *             if __pyx_memoryview_thread_locks_used < THREAD_LOCKS_PREALLOCATED:
  */
   }
 
-  /* "View.MemoryView":363
- *                 raise MemoryError
+  /* "View.MemoryView":365
+ *                     raise MemoryError
  * 
  *         if flags & PyBUF_FORMAT:             # <<<<<<<<<<<<<<
  *             self.dtype_is_object = (self.view.format[0] == b'O' and self.view.format[1] == b'\0')
@@ -11368,7 +11431,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit_
   __pyx_t_1 = ((__pyx_v_flags & PyBUF_FORMAT) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":364
+    /* "View.MemoryView":366
  * 
  *         if flags & PyBUF_FORMAT:
  *             self.dtype_is_object = (self.view.format[0] == b'O' and self.view.format[1] == b'\0')             # <<<<<<<<<<<<<<
@@ -11379,24 +11442,24 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit_
     if (__pyx_t_2) {
     } else {
       __pyx_t_1 = __pyx_t_2;
-      goto __pyx_L11_bool_binop_done;
+      goto __pyx_L12_bool_binop_done;
     }
     __pyx_t_2 = (((__pyx_v_self->view.format[1]) == '\x00') != 0);
     __pyx_t_1 = __pyx_t_2;
-    __pyx_L11_bool_binop_done:;
+    __pyx_L12_bool_binop_done:;
     __pyx_v_self->dtype_is_object = __pyx_t_1;
 
-    /* "View.MemoryView":363
- *                 raise MemoryError
+    /* "View.MemoryView":365
+ *                     raise MemoryError
  * 
  *         if flags & PyBUF_FORMAT:             # <<<<<<<<<<<<<<
  *             self.dtype_is_object = (self.view.format[0] == b'O' and self.view.format[1] == b'\0')
  *         else:
  */
-    goto __pyx_L10;
+    goto __pyx_L11;
   }
 
-  /* "View.MemoryView":366
+  /* "View.MemoryView":368
  *             self.dtype_is_object = (self.view.format[0] == b'O' and self.view.format[1] == b'\0')
  *         else:
  *             self.dtype_is_object = dtype_is_object             # <<<<<<<<<<<<<<
@@ -11406,9 +11469,9 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit_
   /*else*/ {
     __pyx_v_self->dtype_is_object = __pyx_v_dtype_is_object;
   }
-  __pyx_L10:;
+  __pyx_L11:;
 
-  /* "View.MemoryView":368
+  /* "View.MemoryView":370
  *             self.dtype_is_object = dtype_is_object
  * 
  *         self.acquisition_count_aligned_p = <__pyx_atomic_int *> align_pointer(             # <<<<<<<<<<<<<<
@@ -11417,7 +11480,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit_
  */
   __pyx_v_self->acquisition_count_aligned_p = ((__pyx_atomic_int *)__pyx_align_pointer(((void *)(&(__pyx_v_self->acquisition_count[0]))), (sizeof(__pyx_atomic_int))));
 
-  /* "View.MemoryView":370
+  /* "View.MemoryView":372
  *         self.acquisition_count_aligned_p = <__pyx_atomic_int *> align_pointer(
  *                   <void *> &self.acquisition_count[0], sizeof(__pyx_atomic_int))
  *         self.typeinfo = NULL             # <<<<<<<<<<<<<<
@@ -11426,7 +11489,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit_
  */
   __pyx_v_self->typeinfo = NULL;
 
-  /* "View.MemoryView":345
+  /* "View.MemoryView":346
  *     cdef __Pyx_TypeInfo *typeinfo
  * 
  *     def __cinit__(memoryview self, object obj, int flags, bint dtype_is_object=False):             # <<<<<<<<<<<<<<
@@ -11445,7 +11508,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview___cinit_
   return __pyx_r;
 }
 
-/* "View.MemoryView":372
+/* "View.MemoryView":374
  *         self.typeinfo = NULL
  * 
  *     def __dealloc__(memoryview self):             # <<<<<<<<<<<<<<
@@ -11476,7 +11539,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
   PyThread_type_lock __pyx_t_7;
   __Pyx_RefNannySetupContext("__dealloc__", 0);
 
-  /* "View.MemoryView":373
+  /* "View.MemoryView":375
  * 
  *     def __dealloc__(memoryview self):
  *         if self.obj is not None:             # <<<<<<<<<<<<<<
@@ -11487,7 +11550,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
   __pyx_t_2 = (__pyx_t_1 != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":374
+    /* "View.MemoryView":376
  *     def __dealloc__(memoryview self):
  *         if self.obj is not None:
  *             __Pyx_ReleaseBuffer(&self.view)             # <<<<<<<<<<<<<<
@@ -11496,7 +11559,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
  */
     __Pyx_ReleaseBuffer((&__pyx_v_self->view));
 
-    /* "View.MemoryView":373
+    /* "View.MemoryView":375
  * 
  *     def __dealloc__(memoryview self):
  *         if self.obj is not None:             # <<<<<<<<<<<<<<
@@ -11506,7 +11569,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
     goto __pyx_L3;
   }
 
-  /* "View.MemoryView":375
+  /* "View.MemoryView":377
  *         if self.obj is not None:
  *             __Pyx_ReleaseBuffer(&self.view)
  *         elif (<__pyx_buffer *> &self.view).obj == Py_None:             # <<<<<<<<<<<<<<
@@ -11516,7 +11579,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
   __pyx_t_2 = ((((Py_buffer *)(&__pyx_v_self->view))->obj == Py_None) != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":377
+    /* "View.MemoryView":379
  *         elif (<__pyx_buffer *> &self.view).obj == Py_None:
  * 
  *             (<__pyx_buffer *> &self.view).obj = NULL             # <<<<<<<<<<<<<<
@@ -11525,7 +11588,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
  */
     ((Py_buffer *)(&__pyx_v_self->view))->obj = NULL;
 
-    /* "View.MemoryView":378
+    /* "View.MemoryView":380
  * 
  *             (<__pyx_buffer *> &self.view).obj = NULL
  *             Py_DECREF(Py_None)             # <<<<<<<<<<<<<<
@@ -11534,7 +11597,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
  */
     Py_DECREF(Py_None);
 
-    /* "View.MemoryView":375
+    /* "View.MemoryView":377
  *         if self.obj is not None:
  *             __Pyx_ReleaseBuffer(&self.view)
  *         elif (<__pyx_buffer *> &self.view).obj == Py_None:             # <<<<<<<<<<<<<<
@@ -11544,7 +11607,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
   }
   __pyx_L3:;
 
-  /* "View.MemoryView":382
+  /* "View.MemoryView":384
  *         cdef int i
  *         global __pyx_memoryview_thread_locks_used
  *         if self.lock != NULL:             # <<<<<<<<<<<<<<
@@ -11554,7 +11617,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
   __pyx_t_2 = ((__pyx_v_self->lock != NULL) != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":383
+    /* "View.MemoryView":385
  *         global __pyx_memoryview_thread_locks_used
  *         if self.lock != NULL:
  *             for i in range(__pyx_memoryview_thread_locks_used):             # <<<<<<<<<<<<<<
@@ -11566,7 +11629,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
     for (__pyx_t_5 = 0; __pyx_t_5 < __pyx_t_4; __pyx_t_5+=1) {
       __pyx_v_i = __pyx_t_5;
 
-      /* "View.MemoryView":384
+      /* "View.MemoryView":386
  *         if self.lock != NULL:
  *             for i in range(__pyx_memoryview_thread_locks_used):
  *                 if __pyx_memoryview_thread_locks[i] is self.lock:             # <<<<<<<<<<<<<<
@@ -11576,7 +11639,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
       __pyx_t_2 = (((__pyx_memoryview_thread_locks[__pyx_v_i]) == __pyx_v_self->lock) != 0);
       if (__pyx_t_2) {
 
-        /* "View.MemoryView":385
+        /* "View.MemoryView":387
  *             for i in range(__pyx_memoryview_thread_locks_used):
  *                 if __pyx_memoryview_thread_locks[i] is self.lock:
  *                     __pyx_memoryview_thread_locks_used -= 1             # <<<<<<<<<<<<<<
@@ -11585,7 +11648,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
  */
         __pyx_memoryview_thread_locks_used = (__pyx_memoryview_thread_locks_used - 1);
 
-        /* "View.MemoryView":386
+        /* "View.MemoryView":388
  *                 if __pyx_memoryview_thread_locks[i] is self.lock:
  *                     __pyx_memoryview_thread_locks_used -= 1
  *                     if i != __pyx_memoryview_thread_locks_used:             # <<<<<<<<<<<<<<
@@ -11595,7 +11658,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
         __pyx_t_2 = ((__pyx_v_i != __pyx_memoryview_thread_locks_used) != 0);
         if (__pyx_t_2) {
 
-          /* "View.MemoryView":388
+          /* "View.MemoryView":390
  *                     if i != __pyx_memoryview_thread_locks_used:
  *                         __pyx_memoryview_thread_locks[i], __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] = (
  *                             __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used], __pyx_memoryview_thread_locks[i])             # <<<<<<<<<<<<<<
@@ -11605,7 +11668,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
           __pyx_t_6 = (__pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]);
           __pyx_t_7 = (__pyx_memoryview_thread_locks[__pyx_v_i]);
 
-          /* "View.MemoryView":387
+          /* "View.MemoryView":389
  *                     __pyx_memoryview_thread_locks_used -= 1
  *                     if i != __pyx_memoryview_thread_locks_used:
  *                         __pyx_memoryview_thread_locks[i], __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] = (             # <<<<<<<<<<<<<<
@@ -11615,7 +11678,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
           (__pyx_memoryview_thread_locks[__pyx_v_i]) = __pyx_t_6;
           (__pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used]) = __pyx_t_7;
 
-          /* "View.MemoryView":386
+          /* "View.MemoryView":388
  *                 if __pyx_memoryview_thread_locks[i] is self.lock:
  *                     __pyx_memoryview_thread_locks_used -= 1
  *                     if i != __pyx_memoryview_thread_locks_used:             # <<<<<<<<<<<<<<
@@ -11624,7 +11687,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
  */
         }
 
-        /* "View.MemoryView":389
+        /* "View.MemoryView":391
  *                         __pyx_memoryview_thread_locks[i], __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used] = (
  *                             __pyx_memoryview_thread_locks[__pyx_memoryview_thread_locks_used], __pyx_memoryview_thread_locks[i])
  *                     break             # <<<<<<<<<<<<<<
@@ -11633,7 +11696,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
  */
         goto __pyx_L6_break;
 
-        /* "View.MemoryView":384
+        /* "View.MemoryView":386
  *         if self.lock != NULL:
  *             for i in range(__pyx_memoryview_thread_locks_used):
  *                 if __pyx_memoryview_thread_locks[i] is self.lock:             # <<<<<<<<<<<<<<
@@ -11644,7 +11707,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
     }
     /*else*/ {
 
-      /* "View.MemoryView":391
+      /* "View.MemoryView":393
  *                     break
  *             else:
  *                 PyThread_free_lock(self.lock)             # <<<<<<<<<<<<<<
@@ -11655,7 +11718,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
     }
     __pyx_L6_break:;
 
-    /* "View.MemoryView":382
+    /* "View.MemoryView":384
  *         cdef int i
  *         global __pyx_memoryview_thread_locks_used
  *         if self.lock != NULL:             # <<<<<<<<<<<<<<
@@ -11664,7 +11727,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
  */
   }
 
-  /* "View.MemoryView":372
+  /* "View.MemoryView":374
  *         self.typeinfo = NULL
  * 
  *     def __dealloc__(memoryview self):             # <<<<<<<<<<<<<<
@@ -11676,7 +11739,7 @@ static void __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_2__deal
   __Pyx_RefNannyFinishContext();
 }
 
-/* "View.MemoryView":393
+/* "View.MemoryView":395
  *                 PyThread_free_lock(self.lock)
  * 
  *     cdef char *get_item_pointer(memoryview self, object index) except NULL:             # <<<<<<<<<<<<<<
@@ -11702,7 +11765,7 @@ static char *__pyx_memoryview_get_item_pointer(struct __pyx_memoryview_obj *__py
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("get_item_pointer", 0);
 
-  /* "View.MemoryView":395
+  /* "View.MemoryView":397
  *     cdef char *get_item_pointer(memoryview self, object index) except NULL:
  *         cdef Py_ssize_t dim
  *         cdef char *itemp = <char *> self.view.buf             # <<<<<<<<<<<<<<
@@ -11711,7 +11774,7 @@ static char *__pyx_memoryview_get_item_pointer(struct __pyx_memoryview_obj *__py
  */
   __pyx_v_itemp = ((char *)__pyx_v_self->view.buf);
 
-  /* "View.MemoryView":397
+  /* "View.MemoryView":399
  *         cdef char *itemp = <char *> self.view.buf
  * 
  *         for dim, idx in enumerate(index):             # <<<<<<<<<<<<<<
@@ -11723,26 +11786,26 @@ static char *__pyx_memoryview_get_item_pointer(struct __pyx_memoryview_obj *__py
     __pyx_t_2 = __pyx_v_index; __Pyx_INCREF(__pyx_t_2); __pyx_t_3 = 0;
     __pyx_t_4 = NULL;
   } else {
-    __pyx_t_3 = -1; __pyx_t_2 = PyObject_GetIter(__pyx_v_index); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 397, __pyx_L1_error)
+    __pyx_t_3 = -1; __pyx_t_2 = PyObject_GetIter(__pyx_v_index); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 399, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_2);
-    __pyx_t_4 = Py_TYPE(__pyx_t_2)->tp_iternext; if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 397, __pyx_L1_error)
+    __pyx_t_4 = Py_TYPE(__pyx_t_2)->tp_iternext; if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 399, __pyx_L1_error)
   }
   for (;;) {
     if (likely(!__pyx_t_4)) {
       if (likely(PyList_CheckExact(__pyx_t_2))) {
         if (__pyx_t_3 >= PyList_GET_SIZE(__pyx_t_2)) break;
         #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS
-        __pyx_t_5 = PyList_GET_ITEM(__pyx_t_2, __pyx_t_3); __Pyx_INCREF(__pyx_t_5); __pyx_t_3++; if (unlikely(0 < 0)) __PYX_ERR(1, 397, __pyx_L1_error)
+        __pyx_t_5 = PyList_GET_ITEM(__pyx_t_2, __pyx_t_3); __Pyx_INCREF(__pyx_t_5); __pyx_t_3++; if (unlikely(0 < 0)) __PYX_ERR(1, 399, __pyx_L1_error)
         #else
-        __pyx_t_5 = PySequence_ITEM(__pyx_t_2, __pyx_t_3); __pyx_t_3++; if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 397, __pyx_L1_error)
+        __pyx_t_5 = PySequence_ITEM(__pyx_t_2, __pyx_t_3); __pyx_t_3++; if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 399, __pyx_L1_error)
         __Pyx_GOTREF(__pyx_t_5);
         #endif
       } else {
         if (__pyx_t_3 >= PyTuple_GET_SIZE(__pyx_t_2)) break;
         #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS
-        __pyx_t_5 = PyTuple_GET_ITEM(__pyx_t_2, __pyx_t_3); __Pyx_INCREF(__pyx_t_5); __pyx_t_3++; if (unlikely(0 < 0)) __PYX_ERR(1, 397, __pyx_L1_error)
+        __pyx_t_5 = PyTuple_GET_ITEM(__pyx_t_2, __pyx_t_3); __Pyx_INCREF(__pyx_t_5); __pyx_t_3++; if (unlikely(0 < 0)) __PYX_ERR(1, 399, __pyx_L1_error)
         #else
-        __pyx_t_5 = PySequence_ITEM(__pyx_t_2, __pyx_t_3); __pyx_t_3++; if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 397, __pyx_L1_error)
+        __pyx_t_5 = PySequence_ITEM(__pyx_t_2, __pyx_t_3); __pyx_t_3++; if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 399, __pyx_L1_error)
         __Pyx_GOTREF(__pyx_t_5);
         #endif
       }
@@ -11752,7 +11815,7 @@ static char *__pyx_memoryview_get_item_pointer(struct __pyx_memoryview_obj *__py
         PyObject* exc_type = PyErr_Occurred();
         if (exc_type) {
           if (likely(__Pyx_PyErr_GivenExceptionMatches(exc_type, PyExc_StopIteration))) PyErr_Clear();
-          else __PYX_ERR(1, 397, __pyx_L1_error)
+          else __PYX_ERR(1, 399, __pyx_L1_error)
         }
         break;
       }
@@ -11763,18 +11826,18 @@ static char *__pyx_memoryview_get_item_pointer(struct __pyx_memoryview_obj *__py
     __pyx_v_dim = __pyx_t_1;
     __pyx_t_1 = (__pyx_t_1 + 1);
 
-    /* "View.MemoryView":398
+    /* "View.MemoryView":400
  * 
  *         for dim, idx in enumerate(index):
  *             itemp = pybuffer_index(&self.view, itemp, idx, dim)             # <<<<<<<<<<<<<<
  * 
  *         return itemp
  */
-    __pyx_t_6 = __Pyx_PyIndex_AsSsize_t(__pyx_v_idx); if (unlikely((__pyx_t_6 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 398, __pyx_L1_error)
-    __pyx_t_7 = __pyx_pybuffer_index((&__pyx_v_self->view), __pyx_v_itemp, __pyx_t_6, __pyx_v_dim); if (unlikely(__pyx_t_7 == ((char *)NULL))) __PYX_ERR(1, 398, __pyx_L1_error)
+    __pyx_t_6 = __Pyx_PyIndex_AsSsize_t(__pyx_v_idx); if (unlikely((__pyx_t_6 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 400, __pyx_L1_error)
+    __pyx_t_7 = __pyx_pybuffer_index((&__pyx_v_self->view), __pyx_v_itemp, __pyx_t_6, __pyx_v_dim); if (unlikely(__pyx_t_7 == ((char *)NULL))) __PYX_ERR(1, 400, __pyx_L1_error)
     __pyx_v_itemp = __pyx_t_7;
 
-    /* "View.MemoryView":397
+    /* "View.MemoryView":399
  *         cdef char *itemp = <char *> self.view.buf
  * 
  *         for dim, idx in enumerate(index):             # <<<<<<<<<<<<<<
@@ -11784,7 +11847,7 @@ static char *__pyx_memoryview_get_item_pointer(struct __pyx_memoryview_obj *__py
   }
   __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
 
-  /* "View.MemoryView":400
+  /* "View.MemoryView":402
  *             itemp = pybuffer_index(&self.view, itemp, idx, dim)
  * 
  *         return itemp             # <<<<<<<<<<<<<<
@@ -11794,7 +11857,7 @@ static char *__pyx_memoryview_get_item_pointer(struct __pyx_memoryview_obj *__py
   __pyx_r = __pyx_v_itemp;
   goto __pyx_L0;
 
-  /* "View.MemoryView":393
+  /* "View.MemoryView":395
  *                 PyThread_free_lock(self.lock)
  * 
  *     cdef char *get_item_pointer(memoryview self, object index) except NULL:             # <<<<<<<<<<<<<<
@@ -11814,7 +11877,7 @@ static char *__pyx_memoryview_get_item_pointer(struct __pyx_memoryview_obj *__py
   return __pyx_r;
 }
 
-/* "View.MemoryView":403
+/* "View.MemoryView":405
  * 
  * 
  *     def __getitem__(memoryview self, object index):             # <<<<<<<<<<<<<<
@@ -11852,7 +11915,7 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_4_
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("__getitem__", 0);
 
-  /* "View.MemoryView":404
+  /* "View.MemoryView":406
  * 
  *     def __getitem__(memoryview self, object index):
  *         if index is Ellipsis:             # <<<<<<<<<<<<<<
@@ -11863,7 +11926,7 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_4_
   __pyx_t_2 = (__pyx_t_1 != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":405
+    /* "View.MemoryView":407
  *     def __getitem__(memoryview self, object index):
  *         if index is Ellipsis:
  *             return self             # <<<<<<<<<<<<<<
@@ -11875,7 +11938,7 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_4_
     __pyx_r = ((PyObject *)__pyx_v_self);
     goto __pyx_L0;
 
-    /* "View.MemoryView":404
+    /* "View.MemoryView":406
  * 
  *     def __getitem__(memoryview self, object index):
  *         if index is Ellipsis:             # <<<<<<<<<<<<<<
@@ -11884,14 +11947,14 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_4_
  */
   }
 
-  /* "View.MemoryView":407
+  /* "View.MemoryView":409
  *             return self
  * 
  *         have_slices, indices = _unellipsify(index, self.view.ndim)             # <<<<<<<<<<<<<<
  * 
  *         cdef char *itemp
  */
-  __pyx_t_3 = _unellipsify(__pyx_v_index, __pyx_v_self->view.ndim); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 407, __pyx_L1_error)
+  __pyx_t_3 = _unellipsify(__pyx_v_index, __pyx_v_self->view.ndim); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 409, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_3);
   if (likely(__pyx_t_3 != Py_None)) {
     PyObject* sequence = __pyx_t_3;
@@ -11899,7 +11962,7 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_4_
     if (unlikely(size != 2)) {
       if (size > 2) __Pyx_RaiseTooManyValuesError(2);
       else if (size >= 0) __Pyx_RaiseNeedMoreValuesError(size);
-      __PYX_ERR(1, 407, __pyx_L1_error)
+      __PYX_ERR(1, 409, __pyx_L1_error)
     }
     #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS
     __pyx_t_4 = PyTuple_GET_ITEM(sequence, 0); 
@@ -11907,31 +11970,31 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_4_
     __Pyx_INCREF(__pyx_t_4);
     __Pyx_INCREF(__pyx_t_5);
     #else
-    __pyx_t_4 = PySequence_ITEM(sequence, 0); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 407, __pyx_L1_error)
+    __pyx_t_4 = PySequence_ITEM(sequence, 0); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 409, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_4);
-    __pyx_t_5 = PySequence_ITEM(sequence, 1); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 407, __pyx_L1_error)
+    __pyx_t_5 = PySequence_ITEM(sequence, 1); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 409, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_5);
     #endif
     __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
   } else {
-    __Pyx_RaiseNoneNotIterableError(); __PYX_ERR(1, 407, __pyx_L1_error)
+    __Pyx_RaiseNoneNotIterableError(); __PYX_ERR(1, 409, __pyx_L1_error)
   }
   __pyx_v_have_slices = __pyx_t_4;
   __pyx_t_4 = 0;
   __pyx_v_indices = __pyx_t_5;
   __pyx_t_5 = 0;
 
-  /* "View.MemoryView":410
+  /* "View.MemoryView":412
  * 
  *         cdef char *itemp
  *         if have_slices:             # <<<<<<<<<<<<<<
  *             return memview_slice(self, indices)
  *         else:
  */
-  __pyx_t_2 = __Pyx_PyObject_IsTrue(__pyx_v_have_slices); if (unlikely(__pyx_t_2 < 0)) __PYX_ERR(1, 410, __pyx_L1_error)
+  __pyx_t_2 = __Pyx_PyObject_IsTrue(__pyx_v_have_slices); if (unlikely(__pyx_t_2 < 0)) __PYX_ERR(1, 412, __pyx_L1_error)
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":411
+    /* "View.MemoryView":413
  *         cdef char *itemp
  *         if have_slices:
  *             return memview_slice(self, indices)             # <<<<<<<<<<<<<<
@@ -11939,13 +12002,13 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_4_
  *             itemp = self.get_item_pointer(indices)
  */
     __Pyx_XDECREF(__pyx_r);
-    __pyx_t_3 = ((PyObject *)__pyx_memview_slice(__pyx_v_self, __pyx_v_indices)); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 411, __pyx_L1_error)
+    __pyx_t_3 = ((PyObject *)__pyx_memview_slice(__pyx_v_self, __pyx_v_indices)); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 413, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
     __pyx_r = __pyx_t_3;
     __pyx_t_3 = 0;
     goto __pyx_L0;
 
-    /* "View.MemoryView":410
+    /* "View.MemoryView":412
  * 
  *         cdef char *itemp
  *         if have_slices:             # <<<<<<<<<<<<<<
@@ -11954,7 +12017,7 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_4_
  */
   }
 
-  /* "View.MemoryView":413
+  /* "View.MemoryView":415
  *             return memview_slice(self, indices)
  *         else:
  *             itemp = self.get_item_pointer(indices)             # <<<<<<<<<<<<<<
@@ -11962,10 +12025,10 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_4_
  * 
  */
   /*else*/ {
-    __pyx_t_6 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->get_item_pointer(__pyx_v_self, __pyx_v_indices); if (unlikely(__pyx_t_6 == ((char *)NULL))) __PYX_ERR(1, 413, __pyx_L1_error)
+    __pyx_t_6 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->get_item_pointer(__pyx_v_self, __pyx_v_indices); if (unlikely(__pyx_t_6 == ((char *)NULL))) __PYX_ERR(1, 415, __pyx_L1_error)
     __pyx_v_itemp = __pyx_t_6;
 
-    /* "View.MemoryView":414
+    /* "View.MemoryView":416
  *         else:
  *             itemp = self.get_item_pointer(indices)
  *             return self.convert_item_to_object(itemp)             # <<<<<<<<<<<<<<
@@ -11973,14 +12036,14 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_4_
  *     def __setitem__(memoryview self, object index, object value):
  */
     __Pyx_XDECREF(__pyx_r);
-    __pyx_t_3 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->convert_item_to_object(__pyx_v_self, __pyx_v_itemp); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 414, __pyx_L1_error)
+    __pyx_t_3 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->convert_item_to_object(__pyx_v_self, __pyx_v_itemp); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 416, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
     __pyx_r = __pyx_t_3;
     __pyx_t_3 = 0;
     goto __pyx_L0;
   }
 
-  /* "View.MemoryView":403
+  /* "View.MemoryView":405
  * 
  * 
  *     def __getitem__(memoryview self, object index):             # <<<<<<<<<<<<<<
@@ -12003,7 +12066,7 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_4_
   return __pyx_r;
 }
 
-/* "View.MemoryView":416
+/* "View.MemoryView":418
  *             return self.convert_item_to_object(itemp)
  * 
  *     def __setitem__(memoryview self, object index, object value):             # <<<<<<<<<<<<<<
@@ -12039,7 +12102,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_6__setit
   __Pyx_RefNannySetupContext("__setitem__", 0);
   __Pyx_INCREF(__pyx_v_index);
 
-  /* "View.MemoryView":417
+  /* "View.MemoryView":419
  * 
  *     def __setitem__(memoryview self, object index, object value):
  *         if self.view.readonly:             # <<<<<<<<<<<<<<
@@ -12049,20 +12112,20 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_6__setit
   __pyx_t_1 = (__pyx_v_self->view.readonly != 0);
   if (unlikely(__pyx_t_1)) {
 
-    /* "View.MemoryView":418
+    /* "View.MemoryView":420
  *     def __setitem__(memoryview self, object index, object value):
  *         if self.view.readonly:
  *             raise TypeError("Cannot assign to read-only memoryview")             # <<<<<<<<<<<<<<
  * 
  *         have_slices, index = _unellipsify(index, self.view.ndim)
  */
-    __pyx_t_2 = __Pyx_PyObject_Call(__pyx_builtin_TypeError, __pyx_tuple__23, NULL); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 418, __pyx_L1_error)
+    __pyx_t_2 = __Pyx_PyObject_Call(__pyx_builtin_TypeError, __pyx_tuple__23, NULL); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 420, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_2);
     __Pyx_Raise(__pyx_t_2, 0, 0, 0);
     __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
-    __PYX_ERR(1, 418, __pyx_L1_error)
+    __PYX_ERR(1, 420, __pyx_L1_error)
 
-    /* "View.MemoryView":417
+    /* "View.MemoryView":419
  * 
  *     def __setitem__(memoryview self, object index, object value):
  *         if self.view.readonly:             # <<<<<<<<<<<<<<
@@ -12071,14 +12134,14 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_6__setit
  */
   }
 
-  /* "View.MemoryView":420
+  /* "View.MemoryView":422
  *             raise TypeError("Cannot assign to read-only memoryview")
  * 
  *         have_slices, index = _unellipsify(index, self.view.ndim)             # <<<<<<<<<<<<<<
  * 
  *         if have_slices:
  */
-  __pyx_t_2 = _unellipsify(__pyx_v_index, __pyx_v_self->view.ndim); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 420, __pyx_L1_error)
+  __pyx_t_2 = _unellipsify(__pyx_v_index, __pyx_v_self->view.ndim); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 422, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
   if (likely(__pyx_t_2 != Py_None)) {
     PyObject* sequence = __pyx_t_2;
@@ -12086,7 +12149,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_6__setit
     if (unlikely(size != 2)) {
       if (size > 2) __Pyx_RaiseTooManyValuesError(2);
       else if (size >= 0) __Pyx_RaiseNeedMoreValuesError(size);
-      __PYX_ERR(1, 420, __pyx_L1_error)
+      __PYX_ERR(1, 422, __pyx_L1_error)
     }
     #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS
     __pyx_t_3 = PyTuple_GET_ITEM(sequence, 0); 
@@ -12094,67 +12157,67 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_6__setit
     __Pyx_INCREF(__pyx_t_3);
     __Pyx_INCREF(__pyx_t_4);
     #else
-    __pyx_t_3 = PySequence_ITEM(sequence, 0); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 420, __pyx_L1_error)
+    __pyx_t_3 = PySequence_ITEM(sequence, 0); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 422, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
-    __pyx_t_4 = PySequence_ITEM(sequence, 1); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 420, __pyx_L1_error)
+    __pyx_t_4 = PySequence_ITEM(sequence, 1); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 422, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_4);
     #endif
     __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
   } else {
-    __Pyx_RaiseNoneNotIterableError(); __PYX_ERR(1, 420, __pyx_L1_error)
+    __Pyx_RaiseNoneNotIterableError(); __PYX_ERR(1, 422, __pyx_L1_error)
   }
   __pyx_v_have_slices = __pyx_t_3;
   __pyx_t_3 = 0;
   __Pyx_DECREF_SET(__pyx_v_index, __pyx_t_4);
   __pyx_t_4 = 0;
 
-  /* "View.MemoryView":422
+  /* "View.MemoryView":424
  *         have_slices, index = _unellipsify(index, self.view.ndim)
  * 
  *         if have_slices:             # <<<<<<<<<<<<<<
  *             obj = self.is_slice(value)
  *             if obj:
  */
-  __pyx_t_1 = __Pyx_PyObject_IsTrue(__pyx_v_have_slices); if (unlikely(__pyx_t_1 < 0)) __PYX_ERR(1, 422, __pyx_L1_error)
+  __pyx_t_1 = __Pyx_PyObject_IsTrue(__pyx_v_have_slices); if (unlikely(__pyx_t_1 < 0)) __PYX_ERR(1, 424, __pyx_L1_error)
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":423
+    /* "View.MemoryView":425
  * 
  *         if have_slices:
  *             obj = self.is_slice(value)             # <<<<<<<<<<<<<<
  *             if obj:
  *                 self.setitem_slice_assignment(self[index], obj)
  */
-    __pyx_t_2 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->is_slice(__pyx_v_self, __pyx_v_value); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 423, __pyx_L1_error)
+    __pyx_t_2 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->is_slice(__pyx_v_self, __pyx_v_value); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 425, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_2);
     __pyx_v_obj = __pyx_t_2;
     __pyx_t_2 = 0;
 
-    /* "View.MemoryView":424
+    /* "View.MemoryView":426
  *         if have_slices:
  *             obj = self.is_slice(value)
  *             if obj:             # <<<<<<<<<<<<<<
  *                 self.setitem_slice_assignment(self[index], obj)
  *             else:
  */
-    __pyx_t_1 = __Pyx_PyObject_IsTrue(__pyx_v_obj); if (unlikely(__pyx_t_1 < 0)) __PYX_ERR(1, 424, __pyx_L1_error)
+    __pyx_t_1 = __Pyx_PyObject_IsTrue(__pyx_v_obj); if (unlikely(__pyx_t_1 < 0)) __PYX_ERR(1, 426, __pyx_L1_error)
     if (__pyx_t_1) {
 
-      /* "View.MemoryView":425
+      /* "View.MemoryView":427
  *             obj = self.is_slice(value)
  *             if obj:
  *                 self.setitem_slice_assignment(self[index], obj)             # <<<<<<<<<<<<<<
  *             else:
  *                 self.setitem_slice_assign_scalar(self[index], value)
  */
-      __pyx_t_2 = __Pyx_PyObject_GetItem(((PyObject *)__pyx_v_self), __pyx_v_index); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 425, __pyx_L1_error)
+      __pyx_t_2 = __Pyx_PyObject_GetItem(((PyObject *)__pyx_v_self), __pyx_v_index); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 427, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_2);
-      __pyx_t_4 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->setitem_slice_assignment(__pyx_v_self, __pyx_t_2, __pyx_v_obj); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 425, __pyx_L1_error)
+      __pyx_t_4 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->setitem_slice_assignment(__pyx_v_self, __pyx_t_2, __pyx_v_obj); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 427, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_4);
       __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
       __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
 
-      /* "View.MemoryView":424
+      /* "View.MemoryView":426
  *         if have_slices:
  *             obj = self.is_slice(value)
  *             if obj:             # <<<<<<<<<<<<<<
@@ -12164,7 +12227,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_6__setit
       goto __pyx_L5;
     }
 
-    /* "View.MemoryView":427
+    /* "View.MemoryView":429
  *                 self.setitem_slice_assignment(self[index], obj)
  *             else:
  *                 self.setitem_slice_assign_scalar(self[index], value)             # <<<<<<<<<<<<<<
@@ -12172,17 +12235,17 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_6__setit
  *             self.setitem_indexed(index, value)
  */
     /*else*/ {
-      __pyx_t_4 = __Pyx_PyObject_GetItem(((PyObject *)__pyx_v_self), __pyx_v_index); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 427, __pyx_L1_error)
+      __pyx_t_4 = __Pyx_PyObject_GetItem(((PyObject *)__pyx_v_self), __pyx_v_index); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 429, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_4);
-      if (!(likely(((__pyx_t_4) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_4, __pyx_memoryview_type))))) __PYX_ERR(1, 427, __pyx_L1_error)
-      __pyx_t_2 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->setitem_slice_assign_scalar(__pyx_v_self, ((struct __pyx_memoryview_obj *)__pyx_t_4), __pyx_v_value); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 427, __pyx_L1_error)
+      if (!(likely(((__pyx_t_4) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_4, __pyx_memoryview_type))))) __PYX_ERR(1, 429, __pyx_L1_error)
+      __pyx_t_2 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->setitem_slice_assign_scalar(__pyx_v_self, ((struct __pyx_memoryview_obj *)__pyx_t_4), __pyx_v_value); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 429, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_2);
       __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
       __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
     }
     __pyx_L5:;
 
-    /* "View.MemoryView":422
+    /* "View.MemoryView":424
  *         have_slices, index = _unellipsify(index, self.view.ndim)
  * 
  *         if have_slices:             # <<<<<<<<<<<<<<
@@ -12192,7 +12255,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_6__setit
     goto __pyx_L4;
   }
 
-  /* "View.MemoryView":429
+  /* "View.MemoryView":431
  *                 self.setitem_slice_assign_scalar(self[index], value)
  *         else:
  *             self.setitem_indexed(index, value)             # <<<<<<<<<<<<<<
@@ -12200,13 +12263,13 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_6__setit
  *     cdef is_slice(self, obj):
  */
   /*else*/ {
-    __pyx_t_2 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->setitem_indexed(__pyx_v_self, __pyx_v_index, __pyx_v_value); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 429, __pyx_L1_error)
+    __pyx_t_2 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->setitem_indexed(__pyx_v_self, __pyx_v_index, __pyx_v_value); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 431, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_2);
     __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
   }
   __pyx_L4:;
 
-  /* "View.MemoryView":416
+  /* "View.MemoryView":418
  *             return self.convert_item_to_object(itemp)
  * 
  *     def __setitem__(memoryview self, object index, object value):             # <<<<<<<<<<<<<<
@@ -12231,7 +12294,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_6__setit
   return __pyx_r;
 }
 
-/* "View.MemoryView":431
+/* "View.MemoryView":433
  *             self.setitem_indexed(index, value)
  * 
  *     cdef is_slice(self, obj):             # <<<<<<<<<<<<<<
@@ -12257,7 +12320,7 @@ static PyObject *__pyx_memoryview_is_slice(struct __pyx_memoryview_obj *__pyx_v_
   __Pyx_RefNannySetupContext("is_slice", 0);
   __Pyx_INCREF(__pyx_v_obj);
 
-  /* "View.MemoryView":432
+  /* "View.MemoryView":434
  * 
  *     cdef is_slice(self, obj):
  *         if not isinstance(obj, memoryview):             # <<<<<<<<<<<<<<
@@ -12268,7 +12331,7 @@ static PyObject *__pyx_memoryview_is_slice(struct __pyx_memoryview_obj *__pyx_v_
   __pyx_t_2 = ((!(__pyx_t_1 != 0)) != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":433
+    /* "View.MemoryView":435
  *     cdef is_slice(self, obj):
  *         if not isinstance(obj, memoryview):
  *             try:             # <<<<<<<<<<<<<<
@@ -12284,34 +12347,34 @@ static PyObject *__pyx_memoryview_is_slice(struct __pyx_memoryview_obj *__pyx_v_
       __Pyx_XGOTREF(__pyx_t_5);
       /*try:*/ {
 
-        /* "View.MemoryView":434
+        /* "View.MemoryView":436
  *         if not isinstance(obj, memoryview):
  *             try:
  *                 obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS,             # <<<<<<<<<<<<<<
  *                                  self.dtype_is_object)
  *             except TypeError:
  */
-        __pyx_t_6 = __Pyx_PyInt_From_int(((__pyx_v_self->flags & (~PyBUF_WRITABLE)) | PyBUF_ANY_CONTIGUOUS)); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 434, __pyx_L4_error)
+        __pyx_t_6 = __Pyx_PyInt_From_int(((__pyx_v_self->flags & (~PyBUF_WRITABLE)) | PyBUF_ANY_CONTIGUOUS)); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 436, __pyx_L4_error)
         __Pyx_GOTREF(__pyx_t_6);
 
-        /* "View.MemoryView":435
+        /* "View.MemoryView":437
  *             try:
  *                 obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS,
  *                                  self.dtype_is_object)             # <<<<<<<<<<<<<<
  *             except TypeError:
  *                 return None
  */
-        __pyx_t_7 = __Pyx_PyBool_FromLong(__pyx_v_self->dtype_is_object); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 435, __pyx_L4_error)
+        __pyx_t_7 = __Pyx_PyBool_FromLong(__pyx_v_self->dtype_is_object); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 437, __pyx_L4_error)
         __Pyx_GOTREF(__pyx_t_7);
 
-        /* "View.MemoryView":434
+        /* "View.MemoryView":436
  *         if not isinstance(obj, memoryview):
  *             try:
  *                 obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS,             # <<<<<<<<<<<<<<
  *                                  self.dtype_is_object)
  *             except TypeError:
  */
-        __pyx_t_8 = PyTuple_New(3); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 434, __pyx_L4_error)
+        __pyx_t_8 = PyTuple_New(3); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 436, __pyx_L4_error)
         __Pyx_GOTREF(__pyx_t_8);
         __Pyx_INCREF(__pyx_v_obj);
         __Pyx_GIVEREF(__pyx_v_obj);
@@ -12322,13 +12385,13 @@ static PyObject *__pyx_memoryview_is_slice(struct __pyx_memoryview_obj *__pyx_v_
         PyTuple_SET_ITEM(__pyx_t_8, 2, __pyx_t_7);
         __pyx_t_6 = 0;
         __pyx_t_7 = 0;
-        __pyx_t_7 = __Pyx_PyObject_Call(((PyObject *)__pyx_memoryview_type), __pyx_t_8, NULL); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 434, __pyx_L4_error)
+        __pyx_t_7 = __Pyx_PyObject_Call(((PyObject *)__pyx_memoryview_type), __pyx_t_8, NULL); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 436, __pyx_L4_error)
         __Pyx_GOTREF(__pyx_t_7);
         __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0;
         __Pyx_DECREF_SET(__pyx_v_obj, __pyx_t_7);
         __pyx_t_7 = 0;
 
-        /* "View.MemoryView":433
+        /* "View.MemoryView":435
  *     cdef is_slice(self, obj):
  *         if not isinstance(obj, memoryview):
  *             try:             # <<<<<<<<<<<<<<
@@ -12345,7 +12408,7 @@ static PyObject *__pyx_memoryview_is_slice(struct __pyx_memoryview_obj *__pyx_v_
       __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0;
       __Pyx_XDECREF(__pyx_t_8); __pyx_t_8 = 0;
 
-      /* "View.MemoryView":436
+      /* "View.MemoryView":438
  *                 obj = memoryview(obj, self.flags & ~PyBUF_WRITABLE | PyBUF_ANY_CONTIGUOUS,
  *                                  self.dtype_is_object)
  *             except TypeError:             # <<<<<<<<<<<<<<
@@ -12355,12 +12418,12 @@ static PyObject *__pyx_memoryview_is_slice(struct __pyx_memoryview_obj *__pyx_v_
       __pyx_t_9 = __Pyx_PyErr_ExceptionMatches(__pyx_builtin_TypeError);
       if (__pyx_t_9) {
         __Pyx_AddTraceback("View.MemoryView.memoryview.is_slice", __pyx_clineno, __pyx_lineno, __pyx_filename);
-        if (__Pyx_GetException(&__pyx_t_7, &__pyx_t_8, &__pyx_t_6) < 0) __PYX_ERR(1, 436, __pyx_L6_except_error)
+        if (__Pyx_GetException(&__pyx_t_7, &__pyx_t_8, &__pyx_t_6) < 0) __PYX_ERR(1, 438, __pyx_L6_except_error)
         __Pyx_GOTREF(__pyx_t_7);
         __Pyx_GOTREF(__pyx_t_8);
         __Pyx_GOTREF(__pyx_t_6);
 
-        /* "View.MemoryView":437
+        /* "View.MemoryView":439
  *                                  self.dtype_is_object)
  *             except TypeError:
  *                 return None             # <<<<<<<<<<<<<<
@@ -12377,7 +12440,7 @@ static PyObject *__pyx_memoryview_is_slice(struct __pyx_memoryview_obj *__pyx_v_
       goto __pyx_L6_except_error;
       __pyx_L6_except_error:;
 
-      /* "View.MemoryView":433
+      /* "View.MemoryView":435
  *     cdef is_slice(self, obj):
  *         if not isinstance(obj, memoryview):
  *             try:             # <<<<<<<<<<<<<<
@@ -12398,7 +12461,7 @@ static PyObject *__pyx_memoryview_is_slice(struct __pyx_memoryview_obj *__pyx_v_
       __pyx_L9_try_end:;
     }
 
-    /* "View.MemoryView":432
+    /* "View.MemoryView":434
  * 
  *     cdef is_slice(self, obj):
  *         if not isinstance(obj, memoryview):             # <<<<<<<<<<<<<<
@@ -12407,7 +12470,7 @@ static PyObject *__pyx_memoryview_is_slice(struct __pyx_memoryview_obj *__pyx_v_
  */
   }
 
-  /* "View.MemoryView":439
+  /* "View.MemoryView":441
  *                 return None
  * 
  *         return obj             # <<<<<<<<<<<<<<
@@ -12419,7 +12482,7 @@ static PyObject *__pyx_memoryview_is_slice(struct __pyx_memoryview_obj *__pyx_v_
   __pyx_r = __pyx_v_obj;
   goto __pyx_L0;
 
-  /* "View.MemoryView":431
+  /* "View.MemoryView":433
  *             self.setitem_indexed(index, value)
  * 
  *     cdef is_slice(self, obj):             # <<<<<<<<<<<<<<
@@ -12441,7 +12504,7 @@ static PyObject *__pyx_memoryview_is_slice(struct __pyx_memoryview_obj *__pyx_v_
   return __pyx_r;
 }
 
-/* "View.MemoryView":441
+/* "View.MemoryView":443
  *         return obj
  * 
  *     cdef setitem_slice_assignment(self, dst, src):             # <<<<<<<<<<<<<<
@@ -12465,52 +12528,52 @@ static PyObject *__pyx_memoryview_setitem_slice_assignment(struct __pyx_memoryvi
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("setitem_slice_assignment", 0);
 
-  /* "View.MemoryView":445
+  /* "View.MemoryView":447
  *         cdef __Pyx_memviewslice src_slice
  * 
  *         memoryview_copy_contents(get_slice_from_memview(src, &src_slice)[0],             # <<<<<<<<<<<<<<
  *                                  get_slice_from_memview(dst, &dst_slice)[0],
  *                                  src.ndim, dst.ndim, self.dtype_is_object)
  */
-  if (!(likely(((__pyx_v_src) == Py_None) || likely(__Pyx_TypeTest(__pyx_v_src, __pyx_memoryview_type))))) __PYX_ERR(1, 445, __pyx_L1_error)
-  __pyx_t_1 = __pyx_memoryview_get_slice_from_memoryview(((struct __pyx_memoryview_obj *)__pyx_v_src), (&__pyx_v_src_slice)); if (unlikely(__pyx_t_1 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 445, __pyx_L1_error)
+  if (!(likely(((__pyx_v_src) == Py_None) || likely(__Pyx_TypeTest(__pyx_v_src, __pyx_memoryview_type))))) __PYX_ERR(1, 447, __pyx_L1_error)
+  __pyx_t_1 = __pyx_memoryview_get_slice_from_memoryview(((struct __pyx_memoryview_obj *)__pyx_v_src), (&__pyx_v_src_slice)); if (unlikely(__pyx_t_1 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 447, __pyx_L1_error)
 
-  /* "View.MemoryView":446
+  /* "View.MemoryView":448
  * 
  *         memoryview_copy_contents(get_slice_from_memview(src, &src_slice)[0],
  *                                  get_slice_from_memview(dst, &dst_slice)[0],             # <<<<<<<<<<<<<<
  *                                  src.ndim, dst.ndim, self.dtype_is_object)
  * 
  */
-  if (!(likely(((__pyx_v_dst) == Py_None) || likely(__Pyx_TypeTest(__pyx_v_dst, __pyx_memoryview_type))))) __PYX_ERR(1, 446, __pyx_L1_error)
-  __pyx_t_2 = __pyx_memoryview_get_slice_from_memoryview(((struct __pyx_memoryview_obj *)__pyx_v_dst), (&__pyx_v_dst_slice)); if (unlikely(__pyx_t_2 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 446, __pyx_L1_error)
+  if (!(likely(((__pyx_v_dst) == Py_None) || likely(__Pyx_TypeTest(__pyx_v_dst, __pyx_memoryview_type))))) __PYX_ERR(1, 448, __pyx_L1_error)
+  __pyx_t_2 = __pyx_memoryview_get_slice_from_memoryview(((struct __pyx_memoryview_obj *)__pyx_v_dst), (&__pyx_v_dst_slice)); if (unlikely(__pyx_t_2 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 448, __pyx_L1_error)
 
-  /* "View.MemoryView":447
+  /* "View.MemoryView":449
  *         memoryview_copy_contents(get_slice_from_memview(src, &src_slice)[0],
  *                                  get_slice_from_memview(dst, &dst_slice)[0],
  *                                  src.ndim, dst.ndim, self.dtype_is_object)             # <<<<<<<<<<<<<<
  * 
  *     cdef setitem_slice_assign_scalar(self, memoryview dst, value):
  */
-  __pyx_t_3 = __Pyx_PyObject_GetAttrStr(__pyx_v_src, __pyx_n_s_ndim); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 447, __pyx_L1_error)
+  __pyx_t_3 = __Pyx_PyObject_GetAttrStr(__pyx_v_src, __pyx_n_s_ndim); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 449, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_3);
-  __pyx_t_4 = __Pyx_PyInt_As_int(__pyx_t_3); if (unlikely((__pyx_t_4 == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 447, __pyx_L1_error)
+  __pyx_t_4 = __Pyx_PyInt_As_int(__pyx_t_3); if (unlikely((__pyx_t_4 == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 449, __pyx_L1_error)
   __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
-  __pyx_t_3 = __Pyx_PyObject_GetAttrStr(__pyx_v_dst, __pyx_n_s_ndim); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 447, __pyx_L1_error)
+  __pyx_t_3 = __Pyx_PyObject_GetAttrStr(__pyx_v_dst, __pyx_n_s_ndim); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 449, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_3);
-  __pyx_t_5 = __Pyx_PyInt_As_int(__pyx_t_3); if (unlikely((__pyx_t_5 == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 447, __pyx_L1_error)
+  __pyx_t_5 = __Pyx_PyInt_As_int(__pyx_t_3); if (unlikely((__pyx_t_5 == (int)-1) && PyErr_Occurred())) __PYX_ERR(1, 449, __pyx_L1_error)
   __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
 
-  /* "View.MemoryView":445
+  /* "View.MemoryView":447
  *         cdef __Pyx_memviewslice src_slice
  * 
  *         memoryview_copy_contents(get_slice_from_memview(src, &src_slice)[0],             # <<<<<<<<<<<<<<
  *                                  get_slice_from_memview(dst, &dst_slice)[0],
  *                                  src.ndim, dst.ndim, self.dtype_is_object)
  */
-  __pyx_t_6 = __pyx_memoryview_copy_contents((__pyx_t_1[0]), (__pyx_t_2[0]), __pyx_t_4, __pyx_t_5, __pyx_v_self->dtype_is_object); if (unlikely(__pyx_t_6 == ((int)-1))) __PYX_ERR(1, 445, __pyx_L1_error)
+  __pyx_t_6 = __pyx_memoryview_copy_contents((__pyx_t_1[0]), (__pyx_t_2[0]), __pyx_t_4, __pyx_t_5, __pyx_v_self->dtype_is_object); if (unlikely(__pyx_t_6 == ((int)-1))) __PYX_ERR(1, 447, __pyx_L1_error)
 
-  /* "View.MemoryView":441
+  /* "View.MemoryView":443
  *         return obj
  * 
  *     cdef setitem_slice_assignment(self, dst, src):             # <<<<<<<<<<<<<<
@@ -12531,7 +12594,7 @@ static PyObject *__pyx_memoryview_setitem_slice_assignment(struct __pyx_memoryvi
   return __pyx_r;
 }
 
-/* "View.MemoryView":449
+/* "View.MemoryView":451
  *                                  src.ndim, dst.ndim, self.dtype_is_object)
  * 
  *     cdef setitem_slice_assign_scalar(self, memoryview dst, value):             # <<<<<<<<<<<<<<
@@ -12564,7 +12627,7 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("setitem_slice_assign_scalar", 0);
 
-  /* "View.MemoryView":451
+  /* "View.MemoryView":453
  *     cdef setitem_slice_assign_scalar(self, memoryview dst, value):
  *         cdef int array[128]
  *         cdef void *tmp = NULL             # <<<<<<<<<<<<<<
@@ -12573,17 +12636,17 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
  */
   __pyx_v_tmp = NULL;
 
-  /* "View.MemoryView":456
+  /* "View.MemoryView":458
  *         cdef __Pyx_memviewslice *dst_slice
  *         cdef __Pyx_memviewslice tmp_slice
  *         dst_slice = get_slice_from_memview(dst, &tmp_slice)             # <<<<<<<<<<<<<<
  * 
  *         if <size_t>self.view.itemsize > sizeof(array):
  */
-  __pyx_t_1 = __pyx_memoryview_get_slice_from_memoryview(__pyx_v_dst, (&__pyx_v_tmp_slice)); if (unlikely(__pyx_t_1 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 456, __pyx_L1_error)
+  __pyx_t_1 = __pyx_memoryview_get_slice_from_memoryview(__pyx_v_dst, (&__pyx_v_tmp_slice)); if (unlikely(__pyx_t_1 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 458, __pyx_L1_error)
   __pyx_v_dst_slice = __pyx_t_1;
 
-  /* "View.MemoryView":458
+  /* "View.MemoryView":460
  *         dst_slice = get_slice_from_memview(dst, &tmp_slice)
  * 
  *         if <size_t>self.view.itemsize > sizeof(array):             # <<<<<<<<<<<<<<
@@ -12593,7 +12656,7 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
   __pyx_t_2 = ((((size_t)__pyx_v_self->view.itemsize) > (sizeof(__pyx_v_array))) != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":459
+    /* "View.MemoryView":461
  * 
  *         if <size_t>self.view.itemsize > sizeof(array):
  *             tmp = PyMem_Malloc(self.view.itemsize)             # <<<<<<<<<<<<<<
@@ -12602,7 +12665,7 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
  */
     __pyx_v_tmp = PyMem_Malloc(__pyx_v_self->view.itemsize);
 
-    /* "View.MemoryView":460
+    /* "View.MemoryView":462
  *         if <size_t>self.view.itemsize > sizeof(array):
  *             tmp = PyMem_Malloc(self.view.itemsize)
  *             if tmp == NULL:             # <<<<<<<<<<<<<<
@@ -12612,16 +12675,16 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
     __pyx_t_2 = ((__pyx_v_tmp == NULL) != 0);
     if (unlikely(__pyx_t_2)) {
 
-      /* "View.MemoryView":461
+      /* "View.MemoryView":463
  *             tmp = PyMem_Malloc(self.view.itemsize)
  *             if tmp == NULL:
  *                 raise MemoryError             # <<<<<<<<<<<<<<
  *             item = tmp
  *         else:
  */
-      PyErr_NoMemory(); __PYX_ERR(1, 461, __pyx_L1_error)
+      PyErr_NoMemory(); __PYX_ERR(1, 463, __pyx_L1_error)
 
-      /* "View.MemoryView":460
+      /* "View.MemoryView":462
  *         if <size_t>self.view.itemsize > sizeof(array):
  *             tmp = PyMem_Malloc(self.view.itemsize)
  *             if tmp == NULL:             # <<<<<<<<<<<<<<
@@ -12630,7 +12693,7 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
  */
     }
 
-    /* "View.MemoryView":462
+    /* "View.MemoryView":464
  *             if tmp == NULL:
  *                 raise MemoryError
  *             item = tmp             # <<<<<<<<<<<<<<
@@ -12639,7 +12702,7 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
  */
     __pyx_v_item = __pyx_v_tmp;
 
-    /* "View.MemoryView":458
+    /* "View.MemoryView":460
  *         dst_slice = get_slice_from_memview(dst, &tmp_slice)
  * 
  *         if <size_t>self.view.itemsize > sizeof(array):             # <<<<<<<<<<<<<<
@@ -12649,7 +12712,7 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
     goto __pyx_L3;
   }
 
-  /* "View.MemoryView":464
+  /* "View.MemoryView":466
  *             item = tmp
  *         else:
  *             item = <void *> array             # <<<<<<<<<<<<<<
@@ -12661,7 +12724,7 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
   }
   __pyx_L3:;
 
-  /* "View.MemoryView":466
+  /* "View.MemoryView":468
  *             item = <void *> array
  * 
  *         try:             # <<<<<<<<<<<<<<
@@ -12670,7 +12733,7 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
  */
   /*try:*/ {
 
-    /* "View.MemoryView":467
+    /* "View.MemoryView":469
  * 
  *         try:
  *             if self.dtype_is_object:             # <<<<<<<<<<<<<<
@@ -12680,7 +12743,7 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
     __pyx_t_2 = (__pyx_v_self->dtype_is_object != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":468
+      /* "View.MemoryView":470
  *         try:
  *             if self.dtype_is_object:
  *                 (<PyObject **> item)[0] = <PyObject *> value             # <<<<<<<<<<<<<<
@@ -12689,7 +12752,7 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
  */
       (((PyObject **)__pyx_v_item)[0]) = ((PyObject *)__pyx_v_value);
 
-      /* "View.MemoryView":467
+      /* "View.MemoryView":469
  * 
  *         try:
  *             if self.dtype_is_object:             # <<<<<<<<<<<<<<
@@ -12699,7 +12762,7 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
       goto __pyx_L8;
     }
 
-    /* "View.MemoryView":470
+    /* "View.MemoryView":472
  *                 (<PyObject **> item)[0] = <PyObject *> value
  *             else:
  *                 self.assign_item_from_object(<char *> item, value)             # <<<<<<<<<<<<<<
@@ -12707,13 +12770,13 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
  * 
  */
     /*else*/ {
-      __pyx_t_3 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->assign_item_from_object(__pyx_v_self, ((char *)__pyx_v_item), __pyx_v_value); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 470, __pyx_L6_error)
+      __pyx_t_3 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->assign_item_from_object(__pyx_v_self, ((char *)__pyx_v_item), __pyx_v_value); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 472, __pyx_L6_error)
       __Pyx_GOTREF(__pyx_t_3);
       __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
     }
     __pyx_L8:;
 
-    /* "View.MemoryView":474
+    /* "View.MemoryView":476
  * 
  * 
  *             if self.view.suboffsets != NULL:             # <<<<<<<<<<<<<<
@@ -12723,18 +12786,18 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
     __pyx_t_2 = ((__pyx_v_self->view.suboffsets != NULL) != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":475
+      /* "View.MemoryView":477
  * 
  *             if self.view.suboffsets != NULL:
  *                 assert_direct_dimensions(self.view.suboffsets, self.view.ndim)             # <<<<<<<<<<<<<<
  *             slice_assign_scalar(dst_slice, dst.view.ndim, self.view.itemsize,
  *                                 item, self.dtype_is_object)
  */
-      __pyx_t_3 = assert_direct_dimensions(__pyx_v_self->view.suboffsets, __pyx_v_self->view.ndim); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 475, __pyx_L6_error)
+      __pyx_t_3 = assert_direct_dimensions(__pyx_v_self->view.suboffsets, __pyx_v_self->view.ndim); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 477, __pyx_L6_error)
       __Pyx_GOTREF(__pyx_t_3);
       __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
 
-      /* "View.MemoryView":474
+      /* "View.MemoryView":476
  * 
  * 
  *             if self.view.suboffsets != NULL:             # <<<<<<<<<<<<<<
@@ -12743,7 +12806,7 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
  */
     }
 
-    /* "View.MemoryView":476
+    /* "View.MemoryView":478
  *             if self.view.suboffsets != NULL:
  *                 assert_direct_dimensions(self.view.suboffsets, self.view.ndim)
  *             slice_assign_scalar(dst_slice, dst.view.ndim, self.view.itemsize,             # <<<<<<<<<<<<<<
@@ -12753,7 +12816,7 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
     __pyx_memoryview_slice_assign_scalar(__pyx_v_dst_slice, __pyx_v_dst->view.ndim, __pyx_v_self->view.itemsize, __pyx_v_item, __pyx_v_self->dtype_is_object);
   }
 
-  /* "View.MemoryView":479
+  /* "View.MemoryView":481
  *                                 item, self.dtype_is_object)
  *         finally:
  *             PyMem_Free(tmp)             # <<<<<<<<<<<<<<
@@ -12800,7 +12863,7 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
     __pyx_L7:;
   }
 
-  /* "View.MemoryView":449
+  /* "View.MemoryView":451
  *                                  src.ndim, dst.ndim, self.dtype_is_object)
  * 
  *     cdef setitem_slice_assign_scalar(self, memoryview dst, value):             # <<<<<<<<<<<<<<
@@ -12821,7 +12884,7 @@ static PyObject *__pyx_memoryview_setitem_slice_assign_scalar(struct __pyx_memor
   return __pyx_r;
 }
 
-/* "View.MemoryView":481
+/* "View.MemoryView":483
  *             PyMem_Free(tmp)
  * 
  *     cdef setitem_indexed(self, index, value):             # <<<<<<<<<<<<<<
@@ -12840,28 +12903,28 @@ static PyObject *__pyx_memoryview_setitem_indexed(struct __pyx_memoryview_obj *_
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("setitem_indexed", 0);
 
-  /* "View.MemoryView":482
+  /* "View.MemoryView":484
  * 
  *     cdef setitem_indexed(self, index, value):
  *         cdef char *itemp = self.get_item_pointer(index)             # <<<<<<<<<<<<<<
  *         self.assign_item_from_object(itemp, value)
  * 
  */
-  __pyx_t_1 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->get_item_pointer(__pyx_v_self, __pyx_v_index); if (unlikely(__pyx_t_1 == ((char *)NULL))) __PYX_ERR(1, 482, __pyx_L1_error)
+  __pyx_t_1 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->get_item_pointer(__pyx_v_self, __pyx_v_index); if (unlikely(__pyx_t_1 == ((char *)NULL))) __PYX_ERR(1, 484, __pyx_L1_error)
   __pyx_v_itemp = __pyx_t_1;
 
-  /* "View.MemoryView":483
+  /* "View.MemoryView":485
  *     cdef setitem_indexed(self, index, value):
  *         cdef char *itemp = self.get_item_pointer(index)
  *         self.assign_item_from_object(itemp, value)             # <<<<<<<<<<<<<<
  * 
  *     cdef convert_item_to_object(self, char *itemp):
  */
-  __pyx_t_2 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->assign_item_from_object(__pyx_v_self, __pyx_v_itemp, __pyx_v_value); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 483, __pyx_L1_error)
+  __pyx_t_2 = ((struct __pyx_vtabstruct_memoryview *)__pyx_v_self->__pyx_vtab)->assign_item_from_object(__pyx_v_self, __pyx_v_itemp, __pyx_v_value); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 485, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
   __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
 
-  /* "View.MemoryView":481
+  /* "View.MemoryView":483
  *             PyMem_Free(tmp)
  * 
  *     cdef setitem_indexed(self, index, value):             # <<<<<<<<<<<<<<
@@ -12882,7 +12945,7 @@ static PyObject *__pyx_memoryview_setitem_indexed(struct __pyx_memoryview_obj *_
   return __pyx_r;
 }
 
-/* "View.MemoryView":485
+/* "View.MemoryView":487
  *         self.assign_item_from_object(itemp, value)
  * 
  *     cdef convert_item_to_object(self, char *itemp):             # <<<<<<<<<<<<<<
@@ -12912,31 +12975,31 @@ static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("convert_item_to_object", 0);
 
-  /* "View.MemoryView":488
+  /* "View.MemoryView":490
  *         """Only used if instantiated manually by the user, or if Cython doesn't
  *         know how to convert the type"""
  *         import struct             # <<<<<<<<<<<<<<
  *         cdef bytes bytesitem
  * 
  */
-  __pyx_t_1 = __Pyx_Import(__pyx_n_s_struct, 0, 0); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 488, __pyx_L1_error)
+  __pyx_t_1 = __Pyx_Import(__pyx_n_s_struct, 0, 0); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 490, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
   __pyx_v_struct = __pyx_t_1;
   __pyx_t_1 = 0;
 
-  /* "View.MemoryView":491
+  /* "View.MemoryView":493
  *         cdef bytes bytesitem
  * 
  *         bytesitem = itemp[:self.view.itemsize]             # <<<<<<<<<<<<<<
  *         try:
  *             result = struct.unpack(self.view.format, bytesitem)
  */
-  __pyx_t_1 = __Pyx_PyBytes_FromStringAndSize(__pyx_v_itemp + 0, __pyx_v_self->view.itemsize - 0); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 491, __pyx_L1_error)
+  __pyx_t_1 = __Pyx_PyBytes_FromStringAndSize(__pyx_v_itemp + 0, __pyx_v_self->view.itemsize - 0); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 493, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
   __pyx_v_bytesitem = ((PyObject*)__pyx_t_1);
   __pyx_t_1 = 0;
 
-  /* "View.MemoryView":492
+  /* "View.MemoryView":494
  * 
  *         bytesitem = itemp[:self.view.itemsize]
  *         try:             # <<<<<<<<<<<<<<
@@ -12952,16 +13015,16 @@ static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview
     __Pyx_XGOTREF(__pyx_t_4);
     /*try:*/ {
 
-      /* "View.MemoryView":493
+      /* "View.MemoryView":495
  *         bytesitem = itemp[:self.view.itemsize]
  *         try:
  *             result = struct.unpack(self.view.format, bytesitem)             # <<<<<<<<<<<<<<
  *         except struct.error:
  *             raise ValueError("Unable to convert item to object")
  */
-      __pyx_t_5 = __Pyx_PyObject_GetAttrStr(__pyx_v_struct, __pyx_n_s_unpack); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 493, __pyx_L3_error)
+      __pyx_t_5 = __Pyx_PyObject_GetAttrStr(__pyx_v_struct, __pyx_n_s_unpack); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 495, __pyx_L3_error)
       __Pyx_GOTREF(__pyx_t_5);
-      __pyx_t_6 = __Pyx_PyBytes_FromString(__pyx_v_self->view.format); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 493, __pyx_L3_error)
+      __pyx_t_6 = __Pyx_PyBytes_FromString(__pyx_v_self->view.format); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 495, __pyx_L3_error)
       __Pyx_GOTREF(__pyx_t_6);
       __pyx_t_7 = NULL;
       __pyx_t_8 = 0;
@@ -12978,7 +13041,7 @@ static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview
       #if CYTHON_FAST_PYCALL
       if (PyFunction_Check(__pyx_t_5)) {
         PyObject *__pyx_temp[3] = {__pyx_t_7, __pyx_t_6, __pyx_v_bytesitem};
-        __pyx_t_1 = __Pyx_PyFunction_FastCall(__pyx_t_5, __pyx_temp+1-__pyx_t_8, 2+__pyx_t_8); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 493, __pyx_L3_error)
+        __pyx_t_1 = __Pyx_PyFunction_FastCall(__pyx_t_5, __pyx_temp+1-__pyx_t_8, 2+__pyx_t_8); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 495, __pyx_L3_error)
         __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0;
         __Pyx_GOTREF(__pyx_t_1);
         __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0;
@@ -12987,14 +13050,14 @@ static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview
       #if CYTHON_FAST_PYCCALL
       if (__Pyx_PyFastCFunction_Check(__pyx_t_5)) {
         PyObject *__pyx_temp[3] = {__pyx_t_7, __pyx_t_6, __pyx_v_bytesitem};
-        __pyx_t_1 = __Pyx_PyCFunction_FastCall(__pyx_t_5, __pyx_temp+1-__pyx_t_8, 2+__pyx_t_8); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 493, __pyx_L3_error)
+        __pyx_t_1 = __Pyx_PyCFunction_FastCall(__pyx_t_5, __pyx_temp+1-__pyx_t_8, 2+__pyx_t_8); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 495, __pyx_L3_error)
         __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0;
         __Pyx_GOTREF(__pyx_t_1);
         __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0;
       } else
       #endif
       {
-        __pyx_t_9 = PyTuple_New(2+__pyx_t_8); if (unlikely(!__pyx_t_9)) __PYX_ERR(1, 493, __pyx_L3_error)
+        __pyx_t_9 = PyTuple_New(2+__pyx_t_8); if (unlikely(!__pyx_t_9)) __PYX_ERR(1, 495, __pyx_L3_error)
         __Pyx_GOTREF(__pyx_t_9);
         if (__pyx_t_7) {
           __Pyx_GIVEREF(__pyx_t_7); PyTuple_SET_ITEM(__pyx_t_9, 0, __pyx_t_7); __pyx_t_7 = NULL;
@@ -13005,7 +13068,7 @@ static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview
         __Pyx_GIVEREF(__pyx_v_bytesitem);
         PyTuple_SET_ITEM(__pyx_t_9, 1+__pyx_t_8, __pyx_v_bytesitem);
         __pyx_t_6 = 0;
-        __pyx_t_1 = __Pyx_PyObject_Call(__pyx_t_5, __pyx_t_9, NULL); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 493, __pyx_L3_error)
+        __pyx_t_1 = __Pyx_PyObject_Call(__pyx_t_5, __pyx_t_9, NULL); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 495, __pyx_L3_error)
         __Pyx_GOTREF(__pyx_t_1);
         __Pyx_DECREF(__pyx_t_9); __pyx_t_9 = 0;
       }
@@ -13013,7 +13076,7 @@ static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview
       __pyx_v_result = __pyx_t_1;
       __pyx_t_1 = 0;
 
-      /* "View.MemoryView":492
+      /* "View.MemoryView":494
  * 
  *         bytesitem = itemp[:self.view.itemsize]
  *         try:             # <<<<<<<<<<<<<<
@@ -13022,7 +13085,7 @@ static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview
  */
     }
 
-    /* "View.MemoryView":497
+    /* "View.MemoryView":499
  *             raise ValueError("Unable to convert item to object")
  *         else:
  *             if len(self.view.format) == 1:             # <<<<<<<<<<<<<<
@@ -13034,7 +13097,7 @@ static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview
       __pyx_t_11 = ((__pyx_t_10 == 1) != 0);
       if (__pyx_t_11) {
 
-        /* "View.MemoryView":498
+        /* "View.MemoryView":500
  *         else:
  *             if len(self.view.format) == 1:
  *                 return result[0]             # <<<<<<<<<<<<<<
@@ -13042,13 +13105,13 @@ static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview
  * 
  */
         __Pyx_XDECREF(__pyx_r);
-        __pyx_t_1 = __Pyx_GetItemInt(__pyx_v_result, 0, long, 1, __Pyx_PyInt_From_long, 0, 0, 1); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 498, __pyx_L5_except_error)
+        __pyx_t_1 = __Pyx_GetItemInt(__pyx_v_result, 0, long, 1, __Pyx_PyInt_From_long, 0, 0, 1); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 500, __pyx_L5_except_error)
         __Pyx_GOTREF(__pyx_t_1);
         __pyx_r = __pyx_t_1;
         __pyx_t_1 = 0;
         goto __pyx_L6_except_return;
 
-        /* "View.MemoryView":497
+        /* "View.MemoryView":499
  *             raise ValueError("Unable to convert item to object")
  *         else:
  *             if len(self.view.format) == 1:             # <<<<<<<<<<<<<<
@@ -13057,7 +13120,7 @@ static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview
  */
       }
 
-      /* "View.MemoryView":499
+      /* "View.MemoryView":501
  *             if len(self.view.format) == 1:
  *                 return result[0]
  *             return result             # <<<<<<<<<<<<<<
@@ -13076,7 +13139,7 @@ static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview
     __Pyx_XDECREF(__pyx_t_7); __pyx_t_7 = 0;
     __Pyx_XDECREF(__pyx_t_9); __pyx_t_9 = 0;
 
-    /* "View.MemoryView":494
+    /* "View.MemoryView":496
  *         try:
  *             result = struct.unpack(self.view.format, bytesitem)
  *         except struct.error:             # <<<<<<<<<<<<<<
@@ -13084,7 +13147,7 @@ static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview
  *         else:
  */
     __Pyx_ErrFetch(&__pyx_t_1, &__pyx_t_5, &__pyx_t_9);
-    __pyx_t_6 = __Pyx_PyObject_GetAttrStr(__pyx_v_struct, __pyx_n_s_error); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 494, __pyx_L5_except_error)
+    __pyx_t_6 = __Pyx_PyObject_GetAttrStr(__pyx_v_struct, __pyx_n_s_error); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 496, __pyx_L5_except_error)
     __Pyx_GOTREF(__pyx_t_6);
     __pyx_t_8 = __Pyx_PyErr_GivenExceptionMatches(__pyx_t_1, __pyx_t_6);
     __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0;
@@ -13092,28 +13155,28 @@ static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview
     __pyx_t_1 = 0; __pyx_t_5 = 0; __pyx_t_9 = 0;
     if (__pyx_t_8) {
       __Pyx_AddTraceback("View.MemoryView.memoryview.convert_item_to_object", __pyx_clineno, __pyx_lineno, __pyx_filename);
-      if (__Pyx_GetException(&__pyx_t_9, &__pyx_t_5, &__pyx_t_1) < 0) __PYX_ERR(1, 494, __pyx_L5_except_error)
+      if (__Pyx_GetException(&__pyx_t_9, &__pyx_t_5, &__pyx_t_1) < 0) __PYX_ERR(1, 496, __pyx_L5_except_error)
       __Pyx_GOTREF(__pyx_t_9);
       __Pyx_GOTREF(__pyx_t_5);
       __Pyx_GOTREF(__pyx_t_1);
 
-      /* "View.MemoryView":495
+      /* "View.MemoryView":497
  *             result = struct.unpack(self.view.format, bytesitem)
  *         except struct.error:
  *             raise ValueError("Unable to convert item to object")             # <<<<<<<<<<<<<<
  *         else:
  *             if len(self.view.format) == 1:
  */
-      __pyx_t_6 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__24, NULL); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 495, __pyx_L5_except_error)
+      __pyx_t_6 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__24, NULL); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 497, __pyx_L5_except_error)
       __Pyx_GOTREF(__pyx_t_6);
       __Pyx_Raise(__pyx_t_6, 0, 0, 0);
       __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0;
-      __PYX_ERR(1, 495, __pyx_L5_except_error)
+      __PYX_ERR(1, 497, __pyx_L5_except_error)
     }
     goto __pyx_L5_except_error;
     __pyx_L5_except_error:;
 
-    /* "View.MemoryView":492
+    /* "View.MemoryView":494
  * 
  *         bytesitem = itemp[:self.view.itemsize]
  *         try:             # <<<<<<<<<<<<<<
@@ -13133,7 +13196,7 @@ static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview
     goto __pyx_L0;
   }
 
-  /* "View.MemoryView":485
+  /* "View.MemoryView":487
  *         self.assign_item_from_object(itemp, value)
  * 
  *     cdef convert_item_to_object(self, char *itemp):             # <<<<<<<<<<<<<<
@@ -13159,7 +13222,7 @@ static PyObject *__pyx_memoryview_convert_item_to_object(struct __pyx_memoryview
   return __pyx_r;
 }
 
-/* "View.MemoryView":501
+/* "View.MemoryView":503
  *             return result
  * 
  *     cdef assign_item_from_object(self, char *itemp, object value):             # <<<<<<<<<<<<<<
@@ -13193,19 +13256,19 @@ static PyObject *__pyx_memoryview_assign_item_from_object(struct __pyx_memoryvie
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("assign_item_from_object", 0);
 
-  /* "View.MemoryView":504
+  /* "View.MemoryView":506
  *         """Only used if instantiated manually by the user, or if Cython doesn't
  *         know how to convert the type"""
  *         import struct             # <<<<<<<<<<<<<<
  *         cdef char c
  *         cdef bytes bytesvalue
  */
-  __pyx_t_1 = __Pyx_Import(__pyx_n_s_struct, 0, 0); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 504, __pyx_L1_error)
+  __pyx_t_1 = __Pyx_Import(__pyx_n_s_struct, 0, 0); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 506, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
   __pyx_v_struct = __pyx_t_1;
   __pyx_t_1 = 0;
 
-  /* "View.MemoryView":509
+  /* "View.MemoryView":511
  *         cdef Py_ssize_t i
  * 
  *         if isinstance(value, tuple):             # <<<<<<<<<<<<<<
@@ -13216,37 +13279,37 @@ static PyObject *__pyx_memoryview_assign_item_from_object(struct __pyx_memoryvie
   __pyx_t_3 = (__pyx_t_2 != 0);
   if (__pyx_t_3) {
 
-    /* "View.MemoryView":510
+    /* "View.MemoryView":512
  * 
  *         if isinstance(value, tuple):
  *             bytesvalue = struct.pack(self.view.format, *value)             # <<<<<<<<<<<<<<
  *         else:
  *             bytesvalue = struct.pack(self.view.format, value)
  */
-    __pyx_t_1 = __Pyx_PyObject_GetAttrStr(__pyx_v_struct, __pyx_n_s_pack); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 510, __pyx_L1_error)
+    __pyx_t_1 = __Pyx_PyObject_GetAttrStr(__pyx_v_struct, __pyx_n_s_pack); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 512, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_1);
-    __pyx_t_4 = __Pyx_PyBytes_FromString(__pyx_v_self->view.format); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 510, __pyx_L1_error)
+    __pyx_t_4 = __Pyx_PyBytes_FromString(__pyx_v_self->view.format); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 512, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_4);
-    __pyx_t_5 = PyTuple_New(1); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 510, __pyx_L1_error)
+    __pyx_t_5 = PyTuple_New(1); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 512, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_5);
     __Pyx_GIVEREF(__pyx_t_4);
     PyTuple_SET_ITEM(__pyx_t_5, 0, __pyx_t_4);
     __pyx_t_4 = 0;
-    __pyx_t_4 = __Pyx_PySequence_Tuple(__pyx_v_value); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 510, __pyx_L1_error)
+    __pyx_t_4 = __Pyx_PySequence_Tuple(__pyx_v_value); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 512, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_4);
-    __pyx_t_6 = PyNumber_Add(__pyx_t_5, __pyx_t_4); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 510, __pyx_L1_error)
+    __pyx_t_6 = PyNumber_Add(__pyx_t_5, __pyx_t_4); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 512, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_6);
     __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
     __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
-    __pyx_t_4 = __Pyx_PyObject_Call(__pyx_t_1, __pyx_t_6, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 510, __pyx_L1_error)
+    __pyx_t_4 = __Pyx_PyObject_Call(__pyx_t_1, __pyx_t_6, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 512, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_4);
     __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
     __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0;
-    if (!(likely(PyBytes_CheckExact(__pyx_t_4))||((__pyx_t_4) == Py_None)||(PyErr_Format(PyExc_TypeError, "Expected %.16s, got %.200s", "bytes", Py_TYPE(__pyx_t_4)->tp_name), 0))) __PYX_ERR(1, 510, __pyx_L1_error)
+    if (!(likely(PyBytes_CheckExact(__pyx_t_4))||((__pyx_t_4) == Py_None)||(PyErr_Format(PyExc_TypeError, "Expected %.16s, got %.200s", "bytes", Py_TYPE(__pyx_t_4)->tp_name), 0))) __PYX_ERR(1, 512, __pyx_L1_error)
     __pyx_v_bytesvalue = ((PyObject*)__pyx_t_4);
     __pyx_t_4 = 0;
 
-    /* "View.MemoryView":509
+    /* "View.MemoryView":511
  *         cdef Py_ssize_t i
  * 
  *         if isinstance(value, tuple):             # <<<<<<<<<<<<<<
@@ -13256,7 +13319,7 @@ static PyObject *__pyx_memoryview_assign_item_from_object(struct __pyx_memoryvie
     goto __pyx_L3;
   }
 
-  /* "View.MemoryView":512
+  /* "View.MemoryView":514
  *             bytesvalue = struct.pack(self.view.format, *value)
  *         else:
  *             bytesvalue = struct.pack(self.view.format, value)             # <<<<<<<<<<<<<<
@@ -13264,9 +13327,9 @@ static PyObject *__pyx_memoryview_assign_item_from_object(struct __pyx_memoryvie
  *         for i, c in enumerate(bytesvalue):
  */
   /*else*/ {
-    __pyx_t_6 = __Pyx_PyObject_GetAttrStr(__pyx_v_struct, __pyx_n_s_pack); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 512, __pyx_L1_error)
+    __pyx_t_6 = __Pyx_PyObject_GetAttrStr(__pyx_v_struct, __pyx_n_s_pack); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 514, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_6);
-    __pyx_t_1 = __Pyx_PyBytes_FromString(__pyx_v_self->view.format); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 512, __pyx_L1_error)
+    __pyx_t_1 = __Pyx_PyBytes_FromString(__pyx_v_self->view.format); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 514, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_1);
     __pyx_t_5 = NULL;
     __pyx_t_7 = 0;
@@ -13283,7 +13346,7 @@ static PyObject *__pyx_memoryview_assign_item_from_object(struct __pyx_memoryvie
     #if CYTHON_FAST_PYCALL
     if (PyFunction_Check(__pyx_t_6)) {
       PyObject *__pyx_temp[3] = {__pyx_t_5, __pyx_t_1, __pyx_v_value};
-      __pyx_t_4 = __Pyx_PyFunction_FastCall(__pyx_t_6, __pyx_temp+1-__pyx_t_7, 2+__pyx_t_7); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 512, __pyx_L1_error)
+      __pyx_t_4 = __Pyx_PyFunction_FastCall(__pyx_t_6, __pyx_temp+1-__pyx_t_7, 2+__pyx_t_7); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 514, __pyx_L1_error)
       __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0;
       __Pyx_GOTREF(__pyx_t_4);
       __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
@@ -13292,14 +13355,14 @@ static PyObject *__pyx_memoryview_assign_item_from_object(struct __pyx_memoryvie
     #if CYTHON_FAST_PYCCALL
     if (__Pyx_PyFastCFunction_Check(__pyx_t_6)) {
       PyObject *__pyx_temp[3] = {__pyx_t_5, __pyx_t_1, __pyx_v_value};
-      __pyx_t_4 = __Pyx_PyCFunction_FastCall(__pyx_t_6, __pyx_temp+1-__pyx_t_7, 2+__pyx_t_7); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 512, __pyx_L1_error)
+      __pyx_t_4 = __Pyx_PyCFunction_FastCall(__pyx_t_6, __pyx_temp+1-__pyx_t_7, 2+__pyx_t_7); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 514, __pyx_L1_error)
       __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0;
       __Pyx_GOTREF(__pyx_t_4);
       __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
     } else
     #endif
     {
-      __pyx_t_8 = PyTuple_New(2+__pyx_t_7); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 512, __pyx_L1_error)
+      __pyx_t_8 = PyTuple_New(2+__pyx_t_7); if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 514, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_8);
       if (__pyx_t_5) {
         __Pyx_GIVEREF(__pyx_t_5); PyTuple_SET_ITEM(__pyx_t_8, 0, __pyx_t_5); __pyx_t_5 = NULL;
@@ -13310,18 +13373,18 @@ static PyObject *__pyx_memoryview_assign_item_from_object(struct __pyx_memoryvie
       __Pyx_GIVEREF(__pyx_v_value);
       PyTuple_SET_ITEM(__pyx_t_8, 1+__pyx_t_7, __pyx_v_value);
       __pyx_t_1 = 0;
-      __pyx_t_4 = __Pyx_PyObject_Call(__pyx_t_6, __pyx_t_8, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 512, __pyx_L1_error)
+      __pyx_t_4 = __Pyx_PyObject_Call(__pyx_t_6, __pyx_t_8, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 514, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_4);
       __Pyx_DECREF(__pyx_t_8); __pyx_t_8 = 0;
     }
     __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0;
-    if (!(likely(PyBytes_CheckExact(__pyx_t_4))||((__pyx_t_4) == Py_None)||(PyErr_Format(PyExc_TypeError, "Expected %.16s, got %.200s", "bytes", Py_TYPE(__pyx_t_4)->tp_name), 0))) __PYX_ERR(1, 512, __pyx_L1_error)
+    if (!(likely(PyBytes_CheckExact(__pyx_t_4))||((__pyx_t_4) == Py_None)||(PyErr_Format(PyExc_TypeError, "Expected %.16s, got %.200s", "bytes", Py_TYPE(__pyx_t_4)->tp_name), 0))) __PYX_ERR(1, 514, __pyx_L1_error)
     __pyx_v_bytesvalue = ((PyObject*)__pyx_t_4);
     __pyx_t_4 = 0;
   }
   __pyx_L3:;
 
-  /* "View.MemoryView":514
+  /* "View.MemoryView":516
  *             bytesvalue = struct.pack(self.view.format, value)
  * 
  *         for i, c in enumerate(bytesvalue):             # <<<<<<<<<<<<<<
@@ -13331,7 +13394,7 @@ static PyObject *__pyx_memoryview_assign_item_from_object(struct __pyx_memoryvie
   __pyx_t_9 = 0;
   if (unlikely(__pyx_v_bytesvalue == Py_None)) {
     PyErr_SetString(PyExc_TypeError, "'NoneType' is not iterable");
-    __PYX_ERR(1, 514, __pyx_L1_error)
+    __PYX_ERR(1, 516, __pyx_L1_error)
   }
   __Pyx_INCREF(__pyx_v_bytesvalue);
   __pyx_t_10 = __pyx_v_bytesvalue;
@@ -13341,7 +13404,7 @@ static PyObject *__pyx_memoryview_assign_item_from_object(struct __pyx_memoryvie
     __pyx_t_11 = __pyx_t_14;
     __pyx_v_c = (__pyx_t_11[0]);
 
-    /* "View.MemoryView":515
+    /* "View.MemoryView":517
  * 
  *         for i, c in enumerate(bytesvalue):
  *             itemp[i] = c             # <<<<<<<<<<<<<<
@@ -13350,7 +13413,7 @@ static PyObject *__pyx_memoryview_assign_item_from_object(struct __pyx_memoryvie
  */
     __pyx_v_i = __pyx_t_9;
 
-    /* "View.MemoryView":514
+    /* "View.MemoryView":516
  *             bytesvalue = struct.pack(self.view.format, value)
  * 
  *         for i, c in enumerate(bytesvalue):             # <<<<<<<<<<<<<<
@@ -13359,7 +13422,7 @@ static PyObject *__pyx_memoryview_assign_item_from_object(struct __pyx_memoryvie
  */
     __pyx_t_9 = (__pyx_t_9 + 1);
 
-    /* "View.MemoryView":515
+    /* "View.MemoryView":517
  * 
  *         for i, c in enumerate(bytesvalue):
  *             itemp[i] = c             # <<<<<<<<<<<<<<
@@ -13370,7 +13433,7 @@ static PyObject *__pyx_memoryview_assign_item_from_object(struct __pyx_memoryvie
   }
   __Pyx_DECREF(__pyx_t_10); __pyx_t_10 = 0;
 
-  /* "View.MemoryView":501
+  /* "View.MemoryView":503
  *             return result
  * 
  *     cdef assign_item_from_object(self, char *itemp, object value):             # <<<<<<<<<<<<<<
@@ -13398,7 +13461,7 @@ static PyObject *__pyx_memoryview_assign_item_from_object(struct __pyx_memoryvie
   return __pyx_r;
 }
 
-/* "View.MemoryView":518
+/* "View.MemoryView":520
  * 
  *     @cname('getbuffer')
  *     def __getbuffer__(self, Py_buffer *info, int flags):             # <<<<<<<<<<<<<<
@@ -13441,7 +13504,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
   __pyx_v_info->obj = Py_None; __Pyx_INCREF(Py_None);
   __Pyx_GIVEREF(__pyx_v_info->obj);
 
-  /* "View.MemoryView":519
+  /* "View.MemoryView":521
  *     @cname('getbuffer')
  *     def __getbuffer__(self, Py_buffer *info, int flags):
  *         if flags & PyBUF_WRITABLE and self.view.readonly:             # <<<<<<<<<<<<<<
@@ -13459,20 +13522,20 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
   __pyx_L4_bool_binop_done:;
   if (unlikely(__pyx_t_1)) {
 
-    /* "View.MemoryView":520
+    /* "View.MemoryView":522
  *     def __getbuffer__(self, Py_buffer *info, int flags):
  *         if flags & PyBUF_WRITABLE and self.view.readonly:
  *             raise ValueError("Cannot create writable memory view from read-only memoryview")             # <<<<<<<<<<<<<<
  * 
  *         if flags & PyBUF_ND:
  */
-    __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__25, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 520, __pyx_L1_error)
+    __pyx_t_3 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__25, NULL); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 522, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
     __Pyx_Raise(__pyx_t_3, 0, 0, 0);
     __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
-    __PYX_ERR(1, 520, __pyx_L1_error)
+    __PYX_ERR(1, 522, __pyx_L1_error)
 
-    /* "View.MemoryView":519
+    /* "View.MemoryView":521
  *     @cname('getbuffer')
  *     def __getbuffer__(self, Py_buffer *info, int flags):
  *         if flags & PyBUF_WRITABLE and self.view.readonly:             # <<<<<<<<<<<<<<
@@ -13481,7 +13544,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
  */
   }
 
-  /* "View.MemoryView":522
+  /* "View.MemoryView":524
  *             raise ValueError("Cannot create writable memory view from read-only memoryview")
  * 
  *         if flags & PyBUF_ND:             # <<<<<<<<<<<<<<
@@ -13491,7 +13554,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
   __pyx_t_1 = ((__pyx_v_flags & PyBUF_ND) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":523
+    /* "View.MemoryView":525
  * 
  *         if flags & PyBUF_ND:
  *             info.shape = self.view.shape             # <<<<<<<<<<<<<<
@@ -13501,7 +13564,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
     __pyx_t_4 = __pyx_v_self->view.shape;
     __pyx_v_info->shape = __pyx_t_4;
 
-    /* "View.MemoryView":522
+    /* "View.MemoryView":524
  *             raise ValueError("Cannot create writable memory view from read-only memoryview")
  * 
  *         if flags & PyBUF_ND:             # <<<<<<<<<<<<<<
@@ -13511,7 +13574,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
     goto __pyx_L6;
   }
 
-  /* "View.MemoryView":525
+  /* "View.MemoryView":527
  *             info.shape = self.view.shape
  *         else:
  *             info.shape = NULL             # <<<<<<<<<<<<<<
@@ -13523,7 +13586,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
   }
   __pyx_L6:;
 
-  /* "View.MemoryView":527
+  /* "View.MemoryView":529
  *             info.shape = NULL
  * 
  *         if flags & PyBUF_STRIDES:             # <<<<<<<<<<<<<<
@@ -13533,7 +13596,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
   __pyx_t_1 = ((__pyx_v_flags & PyBUF_STRIDES) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":528
+    /* "View.MemoryView":530
  * 
  *         if flags & PyBUF_STRIDES:
  *             info.strides = self.view.strides             # <<<<<<<<<<<<<<
@@ -13543,7 +13606,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
     __pyx_t_4 = __pyx_v_self->view.strides;
     __pyx_v_info->strides = __pyx_t_4;
 
-    /* "View.MemoryView":527
+    /* "View.MemoryView":529
  *             info.shape = NULL
  * 
  *         if flags & PyBUF_STRIDES:             # <<<<<<<<<<<<<<
@@ -13553,7 +13616,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
     goto __pyx_L7;
   }
 
-  /* "View.MemoryView":530
+  /* "View.MemoryView":532
  *             info.strides = self.view.strides
  *         else:
  *             info.strides = NULL             # <<<<<<<<<<<<<<
@@ -13565,7 +13628,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
   }
   __pyx_L7:;
 
-  /* "View.MemoryView":532
+  /* "View.MemoryView":534
  *             info.strides = NULL
  * 
  *         if flags & PyBUF_INDIRECT:             # <<<<<<<<<<<<<<
@@ -13575,7 +13638,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
   __pyx_t_1 = ((__pyx_v_flags & PyBUF_INDIRECT) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":533
+    /* "View.MemoryView":535
  * 
  *         if flags & PyBUF_INDIRECT:
  *             info.suboffsets = self.view.suboffsets             # <<<<<<<<<<<<<<
@@ -13585,7 +13648,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
     __pyx_t_4 = __pyx_v_self->view.suboffsets;
     __pyx_v_info->suboffsets = __pyx_t_4;
 
-    /* "View.MemoryView":532
+    /* "View.MemoryView":534
  *             info.strides = NULL
  * 
  *         if flags & PyBUF_INDIRECT:             # <<<<<<<<<<<<<<
@@ -13595,7 +13658,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
     goto __pyx_L8;
   }
 
-  /* "View.MemoryView":535
+  /* "View.MemoryView":537
  *             info.suboffsets = self.view.suboffsets
  *         else:
  *             info.suboffsets = NULL             # <<<<<<<<<<<<<<
@@ -13607,7 +13670,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
   }
   __pyx_L8:;
 
-  /* "View.MemoryView":537
+  /* "View.MemoryView":539
  *             info.suboffsets = NULL
  * 
  *         if flags & PyBUF_FORMAT:             # <<<<<<<<<<<<<<
@@ -13617,7 +13680,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
   __pyx_t_1 = ((__pyx_v_flags & PyBUF_FORMAT) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":538
+    /* "View.MemoryView":540
  * 
  *         if flags & PyBUF_FORMAT:
  *             info.format = self.view.format             # <<<<<<<<<<<<<<
@@ -13627,7 +13690,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
     __pyx_t_5 = __pyx_v_self->view.format;
     __pyx_v_info->format = __pyx_t_5;
 
-    /* "View.MemoryView":537
+    /* "View.MemoryView":539
  *             info.suboffsets = NULL
  * 
  *         if flags & PyBUF_FORMAT:             # <<<<<<<<<<<<<<
@@ -13637,7 +13700,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
     goto __pyx_L9;
   }
 
-  /* "View.MemoryView":540
+  /* "View.MemoryView":542
  *             info.format = self.view.format
  *         else:
  *             info.format = NULL             # <<<<<<<<<<<<<<
@@ -13649,7 +13712,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
   }
   __pyx_L9:;
 
-  /* "View.MemoryView":542
+  /* "View.MemoryView":544
  *             info.format = NULL
  * 
  *         info.buf = self.view.buf             # <<<<<<<<<<<<<<
@@ -13659,7 +13722,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
   __pyx_t_6 = __pyx_v_self->view.buf;
   __pyx_v_info->buf = __pyx_t_6;
 
-  /* "View.MemoryView":543
+  /* "View.MemoryView":545
  * 
  *         info.buf = self.view.buf
  *         info.ndim = self.view.ndim             # <<<<<<<<<<<<<<
@@ -13669,7 +13732,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
   __pyx_t_7 = __pyx_v_self->view.ndim;
   __pyx_v_info->ndim = __pyx_t_7;
 
-  /* "View.MemoryView":544
+  /* "View.MemoryView":546
  *         info.buf = self.view.buf
  *         info.ndim = self.view.ndim
  *         info.itemsize = self.view.itemsize             # <<<<<<<<<<<<<<
@@ -13679,7 +13742,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
   __pyx_t_8 = __pyx_v_self->view.itemsize;
   __pyx_v_info->itemsize = __pyx_t_8;
 
-  /* "View.MemoryView":545
+  /* "View.MemoryView":547
  *         info.ndim = self.view.ndim
  *         info.itemsize = self.view.itemsize
  *         info.len = self.view.len             # <<<<<<<<<<<<<<
@@ -13689,7 +13752,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
   __pyx_t_8 = __pyx_v_self->view.len;
   __pyx_v_info->len = __pyx_t_8;
 
-  /* "View.MemoryView":546
+  /* "View.MemoryView":548
  *         info.itemsize = self.view.itemsize
  *         info.len = self.view.len
  *         info.readonly = self.view.readonly             # <<<<<<<<<<<<<<
@@ -13699,7 +13762,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
   __pyx_t_1 = __pyx_v_self->view.readonly;
   __pyx_v_info->readonly = __pyx_t_1;
 
-  /* "View.MemoryView":547
+  /* "View.MemoryView":549
  *         info.len = self.view.len
  *         info.readonly = self.view.readonly
  *         info.obj = self             # <<<<<<<<<<<<<<
@@ -13712,7 +13775,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
   __Pyx_DECREF(__pyx_v_info->obj);
   __pyx_v_info->obj = ((PyObject *)__pyx_v_self);
 
-  /* "View.MemoryView":518
+  /* "View.MemoryView":520
  * 
  *     @cname('getbuffer')
  *     def __getbuffer__(self, Py_buffer *info, int flags):             # <<<<<<<<<<<<<<
@@ -13742,7 +13805,7 @@ static int __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_8__getbu
   return __pyx_r;
 }
 
-/* "View.MemoryView":553
+/* "View.MemoryView":555
  * 
  *     @property
  *     def T(self):             # <<<<<<<<<<<<<<
@@ -13774,29 +13837,29 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_1T___get__(struct _
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("__get__", 0);
 
-  /* "View.MemoryView":554
+  /* "View.MemoryView":556
  *     @property
  *     def T(self):
  *         cdef _memoryviewslice result = memoryview_copy(self)             # <<<<<<<<<<<<<<
  *         transpose_memslice(&result.from_slice)
  *         return result
  */
-  __pyx_t_1 = __pyx_memoryview_copy_object(__pyx_v_self); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 554, __pyx_L1_error)
+  __pyx_t_1 = __pyx_memoryview_copy_object(__pyx_v_self); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 556, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
-  if (!(likely(((__pyx_t_1) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_1, __pyx_memoryviewslice_type))))) __PYX_ERR(1, 554, __pyx_L1_error)
+  if (!(likely(((__pyx_t_1) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_1, __pyx_memoryviewslice_type))))) __PYX_ERR(1, 556, __pyx_L1_error)
   __pyx_v_result = ((struct __pyx_memoryviewslice_obj *)__pyx_t_1);
   __pyx_t_1 = 0;
 
-  /* "View.MemoryView":555
+  /* "View.MemoryView":557
  *     def T(self):
  *         cdef _memoryviewslice result = memoryview_copy(self)
  *         transpose_memslice(&result.from_slice)             # <<<<<<<<<<<<<<
  *         return result
  * 
  */
-  __pyx_t_2 = __pyx_memslice_transpose((&__pyx_v_result->from_slice)); if (unlikely(__pyx_t_2 == ((int)0))) __PYX_ERR(1, 555, __pyx_L1_error)
+  __pyx_t_2 = __pyx_memslice_transpose((&__pyx_v_result->from_slice)); if (unlikely(__pyx_t_2 == ((int)0))) __PYX_ERR(1, 557, __pyx_L1_error)
 
-  /* "View.MemoryView":556
+  /* "View.MemoryView":558
  *         cdef _memoryviewslice result = memoryview_copy(self)
  *         transpose_memslice(&result.from_slice)
  *         return result             # <<<<<<<<<<<<<<
@@ -13808,7 +13871,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_1T___get__(struct _
   __pyx_r = ((PyObject *)__pyx_v_result);
   goto __pyx_L0;
 
-  /* "View.MemoryView":553
+  /* "View.MemoryView":555
  * 
  *     @property
  *     def T(self):             # <<<<<<<<<<<<<<
@@ -13828,7 +13891,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_1T___get__(struct _
   return __pyx_r;
 }
 
-/* "View.MemoryView":559
+/* "View.MemoryView":561
  * 
  *     @property
  *     def base(self):             # <<<<<<<<<<<<<<
@@ -13854,7 +13917,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4base___get__(struc
   __Pyx_RefNannyDeclarations
   __Pyx_RefNannySetupContext("__get__", 0);
 
-  /* "View.MemoryView":560
+  /* "View.MemoryView":562
  *     @property
  *     def base(self):
  *         return self.obj             # <<<<<<<<<<<<<<
@@ -13866,7 +13929,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4base___get__(struc
   __pyx_r = __pyx_v_self->obj;
   goto __pyx_L0;
 
-  /* "View.MemoryView":559
+  /* "View.MemoryView":561
  * 
  *     @property
  *     def base(self):             # <<<<<<<<<<<<<<
@@ -13881,7 +13944,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4base___get__(struc
   return __pyx_r;
 }
 
-/* "View.MemoryView":563
+/* "View.MemoryView":565
  * 
  *     @property
  *     def shape(self):             # <<<<<<<<<<<<<<
@@ -13916,7 +13979,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_5shape___get__(stru
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("__get__", 0);
 
-  /* "View.MemoryView":564
+  /* "View.MemoryView":566
  *     @property
  *     def shape(self):
  *         return tuple([length for length in self.view.shape[:self.view.ndim]])             # <<<<<<<<<<<<<<
@@ -13924,25 +13987,25 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_5shape___get__(stru
  *     @property
  */
   __Pyx_XDECREF(__pyx_r);
-  __pyx_t_1 = PyList_New(0); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 564, __pyx_L1_error)
+  __pyx_t_1 = PyList_New(0); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 566, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
   __pyx_t_3 = (__pyx_v_self->view.shape + __pyx_v_self->view.ndim);
   for (__pyx_t_4 = __pyx_v_self->view.shape; __pyx_t_4 < __pyx_t_3; __pyx_t_4++) {
     __pyx_t_2 = __pyx_t_4;
     __pyx_v_length = (__pyx_t_2[0]);
-    __pyx_t_5 = PyInt_FromSsize_t(__pyx_v_length); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 564, __pyx_L1_error)
+    __pyx_t_5 = PyInt_FromSsize_t(__pyx_v_length); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 566, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_5);
-    if (unlikely(__Pyx_ListComp_Append(__pyx_t_1, (PyObject*)__pyx_t_5))) __PYX_ERR(1, 564, __pyx_L1_error)
+    if (unlikely(__Pyx_ListComp_Append(__pyx_t_1, (PyObject*)__pyx_t_5))) __PYX_ERR(1, 566, __pyx_L1_error)
     __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
   }
-  __pyx_t_5 = PyList_AsTuple(((PyObject*)__pyx_t_1)); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 564, __pyx_L1_error)
+  __pyx_t_5 = PyList_AsTuple(((PyObject*)__pyx_t_1)); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 566, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_5);
   __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
   __pyx_r = __pyx_t_5;
   __pyx_t_5 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":563
+  /* "View.MemoryView":565
  * 
  *     @property
  *     def shape(self):             # <<<<<<<<<<<<<<
@@ -13962,7 +14025,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_5shape___get__(stru
   return __pyx_r;
 }
 
-/* "View.MemoryView":567
+/* "View.MemoryView":569
  * 
  *     @property
  *     def strides(self):             # <<<<<<<<<<<<<<
@@ -13998,7 +14061,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_7strides___get__(st
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("__get__", 0);
 
-  /* "View.MemoryView":568
+  /* "View.MemoryView":570
  *     @property
  *     def strides(self):
  *         if self.view.strides == NULL:             # <<<<<<<<<<<<<<
@@ -14008,20 +14071,20 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_7strides___get__(st
   __pyx_t_1 = ((__pyx_v_self->view.strides == NULL) != 0);
   if (unlikely(__pyx_t_1)) {
 
-    /* "View.MemoryView":570
+    /* "View.MemoryView":572
  *         if self.view.strides == NULL:
  * 
  *             raise ValueError("Buffer view does not expose strides")             # <<<<<<<<<<<<<<
  * 
  *         return tuple([stride for stride in self.view.strides[:self.view.ndim]])
  */
-    __pyx_t_2 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__26, NULL); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 570, __pyx_L1_error)
+    __pyx_t_2 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__26, NULL); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 572, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_2);
     __Pyx_Raise(__pyx_t_2, 0, 0, 0);
     __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
-    __PYX_ERR(1, 570, __pyx_L1_error)
+    __PYX_ERR(1, 572, __pyx_L1_error)
 
-    /* "View.MemoryView":568
+    /* "View.MemoryView":570
  *     @property
  *     def strides(self):
  *         if self.view.strides == NULL:             # <<<<<<<<<<<<<<
@@ -14030,7 +14093,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_7strides___get__(st
  */
   }
 
-  /* "View.MemoryView":572
+  /* "View.MemoryView":574
  *             raise ValueError("Buffer view does not expose strides")
  * 
  *         return tuple([stride for stride in self.view.strides[:self.view.ndim]])             # <<<<<<<<<<<<<<
@@ -14038,25 +14101,25 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_7strides___get__(st
  *     @property
  */
   __Pyx_XDECREF(__pyx_r);
-  __pyx_t_2 = PyList_New(0); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 572, __pyx_L1_error)
+  __pyx_t_2 = PyList_New(0); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 574, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
   __pyx_t_4 = (__pyx_v_self->view.strides + __pyx_v_self->view.ndim);
   for (__pyx_t_5 = __pyx_v_self->view.strides; __pyx_t_5 < __pyx_t_4; __pyx_t_5++) {
     __pyx_t_3 = __pyx_t_5;
     __pyx_v_stride = (__pyx_t_3[0]);
-    __pyx_t_6 = PyInt_FromSsize_t(__pyx_v_stride); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 572, __pyx_L1_error)
+    __pyx_t_6 = PyInt_FromSsize_t(__pyx_v_stride); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 574, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_6);
-    if (unlikely(__Pyx_ListComp_Append(__pyx_t_2, (PyObject*)__pyx_t_6))) __PYX_ERR(1, 572, __pyx_L1_error)
+    if (unlikely(__Pyx_ListComp_Append(__pyx_t_2, (PyObject*)__pyx_t_6))) __PYX_ERR(1, 574, __pyx_L1_error)
     __Pyx_DECREF(__pyx_t_6); __pyx_t_6 = 0;
   }
-  __pyx_t_6 = PyList_AsTuple(((PyObject*)__pyx_t_2)); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 572, __pyx_L1_error)
+  __pyx_t_6 = PyList_AsTuple(((PyObject*)__pyx_t_2)); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 574, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_6);
   __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
   __pyx_r = __pyx_t_6;
   __pyx_t_6 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":567
+  /* "View.MemoryView":569
  * 
  *     @property
  *     def strides(self):             # <<<<<<<<<<<<<<
@@ -14076,7 +14139,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_7strides___get__(st
   return __pyx_r;
 }
 
-/* "View.MemoryView":575
+/* "View.MemoryView":577
  * 
  *     @property
  *     def suboffsets(self):             # <<<<<<<<<<<<<<
@@ -14112,7 +14175,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_10suboffsets___get_
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("__get__", 0);
 
-  /* "View.MemoryView":576
+  /* "View.MemoryView":578
  *     @property
  *     def suboffsets(self):
  *         if self.view.suboffsets == NULL:             # <<<<<<<<<<<<<<
@@ -14122,7 +14185,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_10suboffsets___get_
   __pyx_t_1 = ((__pyx_v_self->view.suboffsets == NULL) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":577
+    /* "View.MemoryView":579
  *     def suboffsets(self):
  *         if self.view.suboffsets == NULL:
  *             return (-1,) * self.view.ndim             # <<<<<<<<<<<<<<
@@ -14130,16 +14193,16 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_10suboffsets___get_
  *         return tuple([suboffset for suboffset in self.view.suboffsets[:self.view.ndim]])
  */
     __Pyx_XDECREF(__pyx_r);
-    __pyx_t_2 = __Pyx_PyInt_From_int(__pyx_v_self->view.ndim); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 577, __pyx_L1_error)
+    __pyx_t_2 = __Pyx_PyInt_From_int(__pyx_v_self->view.ndim); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 579, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_2);
-    __pyx_t_3 = PyNumber_Multiply(__pyx_tuple__27, __pyx_t_2); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 577, __pyx_L1_error)
+    __pyx_t_3 = PyNumber_Multiply(__pyx_tuple__27, __pyx_t_2); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 579, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
     __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
     __pyx_r = __pyx_t_3;
     __pyx_t_3 = 0;
     goto __pyx_L0;
 
-    /* "View.MemoryView":576
+    /* "View.MemoryView":578
  *     @property
  *     def suboffsets(self):
  *         if self.view.suboffsets == NULL:             # <<<<<<<<<<<<<<
@@ -14148,7 +14211,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_10suboffsets___get_
  */
   }
 
-  /* "View.MemoryView":579
+  /* "View.MemoryView":581
  *             return (-1,) * self.view.ndim
  * 
  *         return tuple([suboffset for suboffset in self.view.suboffsets[:self.view.ndim]])             # <<<<<<<<<<<<<<
@@ -14156,25 +14219,25 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_10suboffsets___get_
  *     @property
  */
   __Pyx_XDECREF(__pyx_r);
-  __pyx_t_3 = PyList_New(0); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 579, __pyx_L1_error)
+  __pyx_t_3 = PyList_New(0); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 581, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_3);
   __pyx_t_5 = (__pyx_v_self->view.suboffsets + __pyx_v_self->view.ndim);
   for (__pyx_t_6 = __pyx_v_self->view.suboffsets; __pyx_t_6 < __pyx_t_5; __pyx_t_6++) {
     __pyx_t_4 = __pyx_t_6;
     __pyx_v_suboffset = (__pyx_t_4[0]);
-    __pyx_t_2 = PyInt_FromSsize_t(__pyx_v_suboffset); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 579, __pyx_L1_error)
+    __pyx_t_2 = PyInt_FromSsize_t(__pyx_v_suboffset); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 581, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_2);
-    if (unlikely(__Pyx_ListComp_Append(__pyx_t_3, (PyObject*)__pyx_t_2))) __PYX_ERR(1, 579, __pyx_L1_error)
+    if (unlikely(__Pyx_ListComp_Append(__pyx_t_3, (PyObject*)__pyx_t_2))) __PYX_ERR(1, 581, __pyx_L1_error)
     __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
   }
-  __pyx_t_2 = PyList_AsTuple(((PyObject*)__pyx_t_3)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 579, __pyx_L1_error)
+  __pyx_t_2 = PyList_AsTuple(((PyObject*)__pyx_t_3)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 581, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
   __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
   __pyx_r = __pyx_t_2;
   __pyx_t_2 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":575
+  /* "View.MemoryView":577
  * 
  *     @property
  *     def suboffsets(self):             # <<<<<<<<<<<<<<
@@ -14194,7 +14257,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_10suboffsets___get_
   return __pyx_r;
 }
 
-/* "View.MemoryView":582
+/* "View.MemoryView":584
  * 
  *     @property
  *     def ndim(self):             # <<<<<<<<<<<<<<
@@ -14224,7 +14287,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4ndim___get__(struc
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("__get__", 0);
 
-  /* "View.MemoryView":583
+  /* "View.MemoryView":585
  *     @property
  *     def ndim(self):
  *         return self.view.ndim             # <<<<<<<<<<<<<<
@@ -14232,13 +14295,13 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4ndim___get__(struc
  *     @property
  */
   __Pyx_XDECREF(__pyx_r);
-  __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_self->view.ndim); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 583, __pyx_L1_error)
+  __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_self->view.ndim); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 585, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
   __pyx_r = __pyx_t_1;
   __pyx_t_1 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":582
+  /* "View.MemoryView":584
  * 
  *     @property
  *     def ndim(self):             # <<<<<<<<<<<<<<
@@ -14257,7 +14320,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4ndim___get__(struc
   return __pyx_r;
 }
 
-/* "View.MemoryView":586
+/* "View.MemoryView":588
  * 
  *     @property
  *     def itemsize(self):             # <<<<<<<<<<<<<<
@@ -14287,7 +14350,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_8itemsize___get__(s
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("__get__", 0);
 
-  /* "View.MemoryView":587
+  /* "View.MemoryView":589
  *     @property
  *     def itemsize(self):
  *         return self.view.itemsize             # <<<<<<<<<<<<<<
@@ -14295,13 +14358,13 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_8itemsize___get__(s
  *     @property
  */
   __Pyx_XDECREF(__pyx_r);
-  __pyx_t_1 = PyInt_FromSsize_t(__pyx_v_self->view.itemsize); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 587, __pyx_L1_error)
+  __pyx_t_1 = PyInt_FromSsize_t(__pyx_v_self->view.itemsize); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 589, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
   __pyx_r = __pyx_t_1;
   __pyx_t_1 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":586
+  /* "View.MemoryView":588
  * 
  *     @property
  *     def itemsize(self):             # <<<<<<<<<<<<<<
@@ -14320,7 +14383,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_8itemsize___get__(s
   return __pyx_r;
 }
 
-/* "View.MemoryView":590
+/* "View.MemoryView":592
  * 
  *     @property
  *     def nbytes(self):             # <<<<<<<<<<<<<<
@@ -14352,7 +14415,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_6nbytes___get__(str
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("__get__", 0);
 
-  /* "View.MemoryView":591
+  /* "View.MemoryView":593
  *     @property
  *     def nbytes(self):
  *         return self.size * self.view.itemsize             # <<<<<<<<<<<<<<
@@ -14360,11 +14423,11 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_6nbytes___get__(str
  *     @property
  */
   __Pyx_XDECREF(__pyx_r);
-  __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_size); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 591, __pyx_L1_error)
+  __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_size); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 593, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
-  __pyx_t_2 = PyInt_FromSsize_t(__pyx_v_self->view.itemsize); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 591, __pyx_L1_error)
+  __pyx_t_2 = PyInt_FromSsize_t(__pyx_v_self->view.itemsize); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 593, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
-  __pyx_t_3 = PyNumber_Multiply(__pyx_t_1, __pyx_t_2); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 591, __pyx_L1_error)
+  __pyx_t_3 = PyNumber_Multiply(__pyx_t_1, __pyx_t_2); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 593, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_3);
   __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
   __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
@@ -14372,7 +14435,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_6nbytes___get__(str
   __pyx_t_3 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":590
+  /* "View.MemoryView":592
  * 
  *     @property
  *     def nbytes(self):             # <<<<<<<<<<<<<<
@@ -14393,7 +14456,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_6nbytes___get__(str
   return __pyx_r;
 }
 
-/* "View.MemoryView":594
+/* "View.MemoryView":596
  * 
  *     @property
  *     def size(self):             # <<<<<<<<<<<<<<
@@ -14430,7 +14493,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4size___get__(struc
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("__get__", 0);
 
-  /* "View.MemoryView":595
+  /* "View.MemoryView":597
  *     @property
  *     def size(self):
  *         if self._size is None:             # <<<<<<<<<<<<<<
@@ -14441,7 +14504,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4size___get__(struc
   __pyx_t_2 = (__pyx_t_1 != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":596
+    /* "View.MemoryView":598
  *     def size(self):
  *         if self._size is None:
  *             result = 1             # <<<<<<<<<<<<<<
@@ -14451,7 +14514,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4size___get__(struc
     __Pyx_INCREF(__pyx_int_1);
     __pyx_v_result = __pyx_int_1;
 
-    /* "View.MemoryView":598
+    /* "View.MemoryView":600
  *             result = 1
  * 
  *             for length in self.view.shape[:self.view.ndim]:             # <<<<<<<<<<<<<<
@@ -14461,25 +14524,25 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4size___get__(struc
     __pyx_t_4 = (__pyx_v_self->view.shape + __pyx_v_self->view.ndim);
     for (__pyx_t_5 = __pyx_v_self->view.shape; __pyx_t_5 < __pyx_t_4; __pyx_t_5++) {
       __pyx_t_3 = __pyx_t_5;
-      __pyx_t_6 = PyInt_FromSsize_t((__pyx_t_3[0])); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 598, __pyx_L1_error)
+      __pyx_t_6 = PyInt_FromSsize_t((__pyx_t_3[0])); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 600, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_6);
       __Pyx_XDECREF_SET(__pyx_v_length, __pyx_t_6);
       __pyx_t_6 = 0;
 
-      /* "View.MemoryView":599
+      /* "View.MemoryView":601
  * 
  *             for length in self.view.shape[:self.view.ndim]:
  *                 result *= length             # <<<<<<<<<<<<<<
  * 
  *             self._size = result
  */
-      __pyx_t_6 = PyNumber_InPlaceMultiply(__pyx_v_result, __pyx_v_length); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 599, __pyx_L1_error)
+      __pyx_t_6 = PyNumber_InPlaceMultiply(__pyx_v_result, __pyx_v_length); if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 601, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_6);
       __Pyx_DECREF_SET(__pyx_v_result, __pyx_t_6);
       __pyx_t_6 = 0;
     }
 
-    /* "View.MemoryView":601
+    /* "View.MemoryView":603
  *                 result *= length
  * 
  *             self._size = result             # <<<<<<<<<<<<<<
@@ -14492,7 +14555,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4size___get__(struc
     __Pyx_DECREF(__pyx_v_self->_size);
     __pyx_v_self->_size = __pyx_v_result;
 
-    /* "View.MemoryView":595
+    /* "View.MemoryView":597
  *     @property
  *     def size(self):
  *         if self._size is None:             # <<<<<<<<<<<<<<
@@ -14501,7 +14564,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4size___get__(struc
  */
   }
 
-  /* "View.MemoryView":603
+  /* "View.MemoryView":605
  *             self._size = result
  * 
  *         return self._size             # <<<<<<<<<<<<<<
@@ -14513,7 +14576,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4size___get__(struc
   __pyx_r = __pyx_v_self->_size;
   goto __pyx_L0;
 
-  /* "View.MemoryView":594
+  /* "View.MemoryView":596
  * 
  *     @property
  *     def size(self):             # <<<<<<<<<<<<<<
@@ -14534,7 +14597,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_10memoryview_4size___get__(struc
   return __pyx_r;
 }
 
-/* "View.MemoryView":605
+/* "View.MemoryView":607
  *         return self._size
  * 
  *     def __len__(self):             # <<<<<<<<<<<<<<
@@ -14561,7 +14624,7 @@ static Py_ssize_t __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_1
   int __pyx_t_1;
   __Pyx_RefNannySetupContext("__len__", 0);
 
-  /* "View.MemoryView":606
+  /* "View.MemoryView":608
  * 
  *     def __len__(self):
  *         if self.view.ndim >= 1:             # <<<<<<<<<<<<<<
@@ -14571,7 +14634,7 @@ static Py_ssize_t __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_1
   __pyx_t_1 = ((__pyx_v_self->view.ndim >= 1) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":607
+    /* "View.MemoryView":609
  *     def __len__(self):
  *         if self.view.ndim >= 1:
  *             return self.view.shape[0]             # <<<<<<<<<<<<<<
@@ -14581,7 +14644,7 @@ static Py_ssize_t __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_1
     __pyx_r = (__pyx_v_self->view.shape[0]);
     goto __pyx_L0;
 
-    /* "View.MemoryView":606
+    /* "View.MemoryView":608
  * 
  *     def __len__(self):
  *         if self.view.ndim >= 1:             # <<<<<<<<<<<<<<
@@ -14590,7 +14653,7 @@ static Py_ssize_t __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_1
  */
   }
 
-  /* "View.MemoryView":609
+  /* "View.MemoryView":611
  *             return self.view.shape[0]
  * 
  *         return 0             # <<<<<<<<<<<<<<
@@ -14600,7 +14663,7 @@ static Py_ssize_t __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_1
   __pyx_r = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":605
+  /* "View.MemoryView":607
  *         return self._size
  * 
  *     def __len__(self):             # <<<<<<<<<<<<<<
@@ -14614,7 +14677,7 @@ static Py_ssize_t __pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_1
   return __pyx_r;
 }
 
-/* "View.MemoryView":611
+/* "View.MemoryView":613
  *         return 0
  * 
  *     def __repr__(self):             # <<<<<<<<<<<<<<
@@ -14646,7 +14709,7 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_12
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("__repr__", 0);
 
-  /* "View.MemoryView":612
+  /* "View.MemoryView":614
  * 
  *     def __repr__(self):
  *         return "<MemoryView of %r at 0x%x>" % (self.base.__class__.__name__,             # <<<<<<<<<<<<<<
@@ -14654,33 +14717,33 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_12
  * 
  */
   __Pyx_XDECREF(__pyx_r);
-  __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_base); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 612, __pyx_L1_error)
+  __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_base); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 614, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
-  __pyx_t_2 = __Pyx_PyObject_GetAttrStr(__pyx_t_1, __pyx_n_s_class); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 612, __pyx_L1_error)
+  __pyx_t_2 = __Pyx_PyObject_GetAttrStr(__pyx_t_1, __pyx_n_s_class); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 614, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
   __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
-  __pyx_t_1 = __Pyx_PyObject_GetAttrStr(__pyx_t_2, __pyx_n_s_name_2); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 612, __pyx_L1_error)
+  __pyx_t_1 = __Pyx_PyObject_GetAttrStr(__pyx_t_2, __pyx_n_s_name_2); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 614, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
   __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
 
-  /* "View.MemoryView":613
+  /* "View.MemoryView":615
  *     def __repr__(self):
  *         return "<MemoryView of %r at 0x%x>" % (self.base.__class__.__name__,
  *                                                id(self))             # <<<<<<<<<<<<<<
  * 
  *     def __str__(self):
  */
-  __pyx_t_2 = __Pyx_PyObject_CallOneArg(__pyx_builtin_id, ((PyObject *)__pyx_v_self)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 613, __pyx_L1_error)
+  __pyx_t_2 = __Pyx_PyObject_CallOneArg(__pyx_builtin_id, ((PyObject *)__pyx_v_self)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 615, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
 
-  /* "View.MemoryView":612
+  /* "View.MemoryView":614
  * 
  *     def __repr__(self):
  *         return "<MemoryView of %r at 0x%x>" % (self.base.__class__.__name__,             # <<<<<<<<<<<<<<
  *                                                id(self))
  * 
  */
-  __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 612, __pyx_L1_error)
+  __pyx_t_3 = PyTuple_New(2); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 614, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_3);
   __Pyx_GIVEREF(__pyx_t_1);
   PyTuple_SET_ITEM(__pyx_t_3, 0, __pyx_t_1);
@@ -14688,14 +14751,14 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_12
   PyTuple_SET_ITEM(__pyx_t_3, 1, __pyx_t_2);
   __pyx_t_1 = 0;
   __pyx_t_2 = 0;
-  __pyx_t_2 = __Pyx_PyString_Format(__pyx_kp_s_MemoryView_of_r_at_0x_x, __pyx_t_3); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 612, __pyx_L1_error)
+  __pyx_t_2 = __Pyx_PyString_Format(__pyx_kp_s_MemoryView_of_r_at_0x_x, __pyx_t_3); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 614, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
   __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
   __pyx_r = __pyx_t_2;
   __pyx_t_2 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":611
+  /* "View.MemoryView":613
  *         return 0
  * 
  *     def __repr__(self):             # <<<<<<<<<<<<<<
@@ -14716,7 +14779,7 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_12
   return __pyx_r;
 }
 
-/* "View.MemoryView":615
+/* "View.MemoryView":617
  *                                                id(self))
  * 
  *     def __str__(self):             # <<<<<<<<<<<<<<
@@ -14747,7 +14810,7 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_14
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("__str__", 0);
 
-  /* "View.MemoryView":616
+  /* "View.MemoryView":618
  * 
  *     def __str__(self):
  *         return "<MemoryView of %r object>" % (self.base.__class__.__name__,)             # <<<<<<<<<<<<<<
@@ -14755,27 +14818,27 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_14
  * 
  */
   __Pyx_XDECREF(__pyx_r);
-  __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_base); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 616, __pyx_L1_error)
+  __pyx_t_1 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_self), __pyx_n_s_base); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 618, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
-  __pyx_t_2 = __Pyx_PyObject_GetAttrStr(__pyx_t_1, __pyx_n_s_class); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 616, __pyx_L1_error)
+  __pyx_t_2 = __Pyx_PyObject_GetAttrStr(__pyx_t_1, __pyx_n_s_class); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 618, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
   __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
-  __pyx_t_1 = __Pyx_PyObject_GetAttrStr(__pyx_t_2, __pyx_n_s_name_2); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 616, __pyx_L1_error)
+  __pyx_t_1 = __Pyx_PyObject_GetAttrStr(__pyx_t_2, __pyx_n_s_name_2); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 618, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
   __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
-  __pyx_t_2 = PyTuple_New(1); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 616, __pyx_L1_error)
+  __pyx_t_2 = PyTuple_New(1); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 618, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
   __Pyx_GIVEREF(__pyx_t_1);
   PyTuple_SET_ITEM(__pyx_t_2, 0, __pyx_t_1);
   __pyx_t_1 = 0;
-  __pyx_t_1 = __Pyx_PyString_Format(__pyx_kp_s_MemoryView_of_r_object, __pyx_t_2); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 616, __pyx_L1_error)
+  __pyx_t_1 = __Pyx_PyString_Format(__pyx_kp_s_MemoryView_of_r_object, __pyx_t_2); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 618, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
   __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
   __pyx_r = __pyx_t_1;
   __pyx_t_1 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":615
+  /* "View.MemoryView":617
  *                                                id(self))
  * 
  *     def __str__(self):             # <<<<<<<<<<<<<<
@@ -14795,7 +14858,7 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_14
   return __pyx_r;
 }
 
-/* "View.MemoryView":619
+/* "View.MemoryView":621
  * 
  * 
  *     def is_c_contig(self):             # <<<<<<<<<<<<<<
@@ -14828,17 +14891,17 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_16
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("is_c_contig", 0);
 
-  /* "View.MemoryView":622
+  /* "View.MemoryView":624
  *         cdef __Pyx_memviewslice *mslice
  *         cdef __Pyx_memviewslice tmp
  *         mslice = get_slice_from_memview(self, &tmp)             # <<<<<<<<<<<<<<
  *         return slice_is_contig(mslice[0], 'C', self.view.ndim)
  * 
  */
-  __pyx_t_1 = __pyx_memoryview_get_slice_from_memoryview(__pyx_v_self, (&__pyx_v_tmp)); if (unlikely(__pyx_t_1 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 622, __pyx_L1_error)
+  __pyx_t_1 = __pyx_memoryview_get_slice_from_memoryview(__pyx_v_self, (&__pyx_v_tmp)); if (unlikely(__pyx_t_1 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 624, __pyx_L1_error)
   __pyx_v_mslice = __pyx_t_1;
 
-  /* "View.MemoryView":623
+  /* "View.MemoryView":625
  *         cdef __Pyx_memviewslice tmp
  *         mslice = get_slice_from_memview(self, &tmp)
  *         return slice_is_contig(mslice[0], 'C', self.view.ndim)             # <<<<<<<<<<<<<<
@@ -14846,13 +14909,13 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_16
  *     def is_f_contig(self):
  */
   __Pyx_XDECREF(__pyx_r);
-  __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_memviewslice_is_contig((__pyx_v_mslice[0]), 'C', __pyx_v_self->view.ndim)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 623, __pyx_L1_error)
+  __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_memviewslice_is_contig((__pyx_v_mslice[0]), 'C', __pyx_v_self->view.ndim)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 625, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
   __pyx_r = __pyx_t_2;
   __pyx_t_2 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":619
+  /* "View.MemoryView":621
  * 
  * 
  *     def is_c_contig(self):             # <<<<<<<<<<<<<<
@@ -14871,7 +14934,7 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_16
   return __pyx_r;
 }
 
-/* "View.MemoryView":625
+/* "View.MemoryView":627
  *         return slice_is_contig(mslice[0], 'C', self.view.ndim)
  * 
  *     def is_f_contig(self):             # <<<<<<<<<<<<<<
@@ -14904,17 +14967,17 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_18
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("is_f_contig", 0);
 
-  /* "View.MemoryView":628
+  /* "View.MemoryView":630
  *         cdef __Pyx_memviewslice *mslice
  *         cdef __Pyx_memviewslice tmp
  *         mslice = get_slice_from_memview(self, &tmp)             # <<<<<<<<<<<<<<
  *         return slice_is_contig(mslice[0], 'F', self.view.ndim)
  * 
  */
-  __pyx_t_1 = __pyx_memoryview_get_slice_from_memoryview(__pyx_v_self, (&__pyx_v_tmp)); if (unlikely(__pyx_t_1 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 628, __pyx_L1_error)
+  __pyx_t_1 = __pyx_memoryview_get_slice_from_memoryview(__pyx_v_self, (&__pyx_v_tmp)); if (unlikely(__pyx_t_1 == ((__Pyx_memviewslice *)NULL))) __PYX_ERR(1, 630, __pyx_L1_error)
   __pyx_v_mslice = __pyx_t_1;
 
-  /* "View.MemoryView":629
+  /* "View.MemoryView":631
  *         cdef __Pyx_memviewslice tmp
  *         mslice = get_slice_from_memview(self, &tmp)
  *         return slice_is_contig(mslice[0], 'F', self.view.ndim)             # <<<<<<<<<<<<<<
@@ -14922,13 +14985,13 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_18
  *     def copy(self):
  */
   __Pyx_XDECREF(__pyx_r);
-  __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_memviewslice_is_contig((__pyx_v_mslice[0]), 'F', __pyx_v_self->view.ndim)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 629, __pyx_L1_error)
+  __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_memviewslice_is_contig((__pyx_v_mslice[0]), 'F', __pyx_v_self->view.ndim)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 631, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
   __pyx_r = __pyx_t_2;
   __pyx_t_2 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":625
+  /* "View.MemoryView":627
  *         return slice_is_contig(mslice[0], 'C', self.view.ndim)
  * 
  *     def is_f_contig(self):             # <<<<<<<<<<<<<<
@@ -14947,7 +15010,7 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_18
   return __pyx_r;
 }
 
-/* "View.MemoryView":631
+/* "View.MemoryView":633
  *         return slice_is_contig(mslice[0], 'F', self.view.ndim)
  * 
  *     def copy(self):             # <<<<<<<<<<<<<<
@@ -14980,7 +15043,7 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_20
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("copy", 0);
 
-  /* "View.MemoryView":633
+  /* "View.MemoryView":635
  *     def copy(self):
  *         cdef __Pyx_memviewslice mslice
  *         cdef int flags = self.flags & ~PyBUF_F_CONTIGUOUS             # <<<<<<<<<<<<<<
@@ -14989,7 +15052,7 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_20
  */
   __pyx_v_flags = (__pyx_v_self->flags & (~PyBUF_F_CONTIGUOUS));
 
-  /* "View.MemoryView":635
+  /* "View.MemoryView":637
  *         cdef int flags = self.flags & ~PyBUF_F_CONTIGUOUS
  * 
  *         slice_copy(self, &mslice)             # <<<<<<<<<<<<<<
@@ -14998,17 +15061,17 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_20
  */
   __pyx_memoryview_slice_copy(__pyx_v_self, (&__pyx_v_mslice));
 
-  /* "View.MemoryView":636
+  /* "View.MemoryView":638
  * 
  *         slice_copy(self, &mslice)
  *         mslice = slice_copy_contig(&mslice, "c", self.view.ndim,             # <<<<<<<<<<<<<<
  *                                    self.view.itemsize,
  *                                    flags|PyBUF_C_CONTIGUOUS,
  */
-  __pyx_t_1 = __pyx_memoryview_copy_new_contig((&__pyx_v_mslice), ((char *)"c"), __pyx_v_self->view.ndim, __pyx_v_self->view.itemsize, (__pyx_v_flags | PyBUF_C_CONTIGUOUS), __pyx_v_self->dtype_is_object); if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 636, __pyx_L1_error)
+  __pyx_t_1 = __pyx_memoryview_copy_new_contig((&__pyx_v_mslice), ((char *)"c"), __pyx_v_self->view.ndim, __pyx_v_self->view.itemsize, (__pyx_v_flags | PyBUF_C_CONTIGUOUS), __pyx_v_self->dtype_is_object); if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 638, __pyx_L1_error)
   __pyx_v_mslice = __pyx_t_1;
 
-  /* "View.MemoryView":641
+  /* "View.MemoryView":643
  *                                    self.dtype_is_object)
  * 
  *         return memoryview_copy_from_slice(self, &mslice)             # <<<<<<<<<<<<<<
@@ -15016,13 +15079,13 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_20
  *     def copy_fortran(self):
  */
   __Pyx_XDECREF(__pyx_r);
-  __pyx_t_2 = __pyx_memoryview_copy_object_from_slice(__pyx_v_self, (&__pyx_v_mslice)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 641, __pyx_L1_error)
+  __pyx_t_2 = __pyx_memoryview_copy_object_from_slice(__pyx_v_self, (&__pyx_v_mslice)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 643, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
   __pyx_r = __pyx_t_2;
   __pyx_t_2 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":631
+  /* "View.MemoryView":633
  *         return slice_is_contig(mslice[0], 'F', self.view.ndim)
  * 
  *     def copy(self):             # <<<<<<<<<<<<<<
@@ -15041,7 +15104,7 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_20
   return __pyx_r;
 }
 
-/* "View.MemoryView":643
+/* "View.MemoryView":645
  *         return memoryview_copy_from_slice(self, &mslice)
  * 
  *     def copy_fortran(self):             # <<<<<<<<<<<<<<
@@ -15075,7 +15138,7 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_22
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("copy_fortran", 0);
 
-  /* "View.MemoryView":645
+  /* "View.MemoryView":647
  *     def copy_fortran(self):
  *         cdef __Pyx_memviewslice src, dst
  *         cdef int flags = self.flags & ~PyBUF_C_CONTIGUOUS             # <<<<<<<<<<<<<<
@@ -15084,7 +15147,7 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_22
  */
   __pyx_v_flags = (__pyx_v_self->flags & (~PyBUF_C_CONTIGUOUS));
 
-  /* "View.MemoryView":647
+  /* "View.MemoryView":649
  *         cdef int flags = self.flags & ~PyBUF_C_CONTIGUOUS
  * 
  *         slice_copy(self, &src)             # <<<<<<<<<<<<<<
@@ -15093,17 +15156,17 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_22
  */
   __pyx_memoryview_slice_copy(__pyx_v_self, (&__pyx_v_src));
 
-  /* "View.MemoryView":648
+  /* "View.MemoryView":650
  * 
  *         slice_copy(self, &src)
  *         dst = slice_copy_contig(&src, "fortran", self.view.ndim,             # <<<<<<<<<<<<<<
  *                                 self.view.itemsize,
  *                                 flags|PyBUF_F_CONTIGUOUS,
  */
-  __pyx_t_1 = __pyx_memoryview_copy_new_contig((&__pyx_v_src), ((char *)"fortran"), __pyx_v_self->view.ndim, __pyx_v_self->view.itemsize, (__pyx_v_flags | PyBUF_F_CONTIGUOUS), __pyx_v_self->dtype_is_object); if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 648, __pyx_L1_error)
+  __pyx_t_1 = __pyx_memoryview_copy_new_contig((&__pyx_v_src), ((char *)"fortran"), __pyx_v_self->view.ndim, __pyx_v_self->view.itemsize, (__pyx_v_flags | PyBUF_F_CONTIGUOUS), __pyx_v_self->dtype_is_object); if (unlikely(PyErr_Occurred())) __PYX_ERR(1, 650, __pyx_L1_error)
   __pyx_v_dst = __pyx_t_1;
 
-  /* "View.MemoryView":653
+  /* "View.MemoryView":655
  *                                 self.dtype_is_object)
  * 
  *         return memoryview_copy_from_slice(self, &dst)             # <<<<<<<<<<<<<<
@@ -15111,13 +15174,13 @@ static PyObject *__pyx_memoryview___pyx_pf_15View_dot_MemoryView_10memoryview_22
  * 
  */
   __Pyx_XDECREF(__pyx_r);
-  __pyx_t_2 = __pyx_memoryview_copy_object_from_slice(__pyx_v_self, (&__pyx_v_dst)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 653, __pyx_L1_error)
+  __pyx_t_2 = __pyx_memoryview_copy_object_from_slice(__pyx_v_self, (&__pyx_v_dst)); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 655, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
   __pyx_r = __pyx_t_2;
   __pyx_t_2 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":643
+  /* "View.MemoryView":645
  *         return memoryview_copy_from_slice(self, &mslice)
  * 
  *     def copy_fortran(self):             # <<<<<<<<<<<<<<
@@ -15249,7 +15312,7 @@ static PyObject *__pyx_pf___pyx_memoryview_2__setstate_cython__(CYTHON_UNUSED st
   return __pyx_r;
 }
 
-/* "View.MemoryView":657
+/* "View.MemoryView":659
  * 
  * @cname('__pyx_memoryview_new')
  * cdef memoryview_cwrapper(object o, int flags, bint dtype_is_object, __Pyx_TypeInfo *typeinfo):             # <<<<<<<<<<<<<<
@@ -15269,18 +15332,18 @@ static PyObject *__pyx_memoryview_new(PyObject *__pyx_v_o, int __pyx_v_flags, in
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("memoryview_cwrapper", 0);
 
-  /* "View.MemoryView":658
+  /* "View.MemoryView":660
  * @cname('__pyx_memoryview_new')
  * cdef memoryview_cwrapper(object o, int flags, bint dtype_is_object, __Pyx_TypeInfo *typeinfo):
  *     cdef memoryview result = memoryview(o, flags, dtype_is_object)             # <<<<<<<<<<<<<<
  *     result.typeinfo = typeinfo
  *     return result
  */
-  __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_flags); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 658, __pyx_L1_error)
+  __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_flags); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 660, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
-  __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_v_dtype_is_object); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 658, __pyx_L1_error)
+  __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_v_dtype_is_object); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 660, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
-  __pyx_t_3 = PyTuple_New(3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 658, __pyx_L1_error)
+  __pyx_t_3 = PyTuple_New(3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 660, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_3);
   __Pyx_INCREF(__pyx_v_o);
   __Pyx_GIVEREF(__pyx_v_o);
@@ -15291,13 +15354,13 @@ static PyObject *__pyx_memoryview_new(PyObject *__pyx_v_o, int __pyx_v_flags, in
   PyTuple_SET_ITEM(__pyx_t_3, 2, __pyx_t_2);
   __pyx_t_1 = 0;
   __pyx_t_2 = 0;
-  __pyx_t_2 = __Pyx_PyObject_Call(((PyObject *)__pyx_memoryview_type), __pyx_t_3, NULL); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 658, __pyx_L1_error)
+  __pyx_t_2 = __Pyx_PyObject_Call(((PyObject *)__pyx_memoryview_type), __pyx_t_3, NULL); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 660, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
   __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
   __pyx_v_result = ((struct __pyx_memoryview_obj *)__pyx_t_2);
   __pyx_t_2 = 0;
 
-  /* "View.MemoryView":659
+  /* "View.MemoryView":661
  * cdef memoryview_cwrapper(object o, int flags, bint dtype_is_object, __Pyx_TypeInfo *typeinfo):
  *     cdef memoryview result = memoryview(o, flags, dtype_is_object)
  *     result.typeinfo = typeinfo             # <<<<<<<<<<<<<<
@@ -15306,7 +15369,7 @@ static PyObject *__pyx_memoryview_new(PyObject *__pyx_v_o, int __pyx_v_flags, in
  */
   __pyx_v_result->typeinfo = __pyx_v_typeinfo;
 
-  /* "View.MemoryView":660
+  /* "View.MemoryView":662
  *     cdef memoryview result = memoryview(o, flags, dtype_is_object)
  *     result.typeinfo = typeinfo
  *     return result             # <<<<<<<<<<<<<<
@@ -15318,7 +15381,7 @@ static PyObject *__pyx_memoryview_new(PyObject *__pyx_v_o, int __pyx_v_flags, in
   __pyx_r = ((PyObject *)__pyx_v_result);
   goto __pyx_L0;
 
-  /* "View.MemoryView":657
+  /* "View.MemoryView":659
  * 
  * @cname('__pyx_memoryview_new')
  * cdef memoryview_cwrapper(object o, int flags, bint dtype_is_object, __Pyx_TypeInfo *typeinfo):             # <<<<<<<<<<<<<<
@@ -15340,7 +15403,7 @@ static PyObject *__pyx_memoryview_new(PyObject *__pyx_v_o, int __pyx_v_flags, in
   return __pyx_r;
 }
 
-/* "View.MemoryView":663
+/* "View.MemoryView":665
  * 
  * @cname('__pyx_memoryview_check')
  * cdef inline bint memoryview_check(object o):             # <<<<<<<<<<<<<<
@@ -15354,7 +15417,7 @@ static CYTHON_INLINE int __pyx_memoryview_check(PyObject *__pyx_v_o) {
   int __pyx_t_1;
   __Pyx_RefNannySetupContext("memoryview_check", 0);
 
-  /* "View.MemoryView":664
+  /* "View.MemoryView":666
  * @cname('__pyx_memoryview_check')
  * cdef inline bint memoryview_check(object o):
  *     return isinstance(o, memoryview)             # <<<<<<<<<<<<<<
@@ -15365,7 +15428,7 @@ static CYTHON_INLINE int __pyx_memoryview_check(PyObject *__pyx_v_o) {
   __pyx_r = __pyx_t_1;
   goto __pyx_L0;
 
-  /* "View.MemoryView":663
+  /* "View.MemoryView":665
  * 
  * @cname('__pyx_memoryview_check')
  * cdef inline bint memoryview_check(object o):             # <<<<<<<<<<<<<<
@@ -15379,7 +15442,7 @@ static CYTHON_INLINE int __pyx_memoryview_check(PyObject *__pyx_v_o) {
   return __pyx_r;
 }
 
-/* "View.MemoryView":666
+/* "View.MemoryView":668
  *     return isinstance(o, memoryview)
  * 
  * cdef tuple _unellipsify(object index, int ndim):             # <<<<<<<<<<<<<<
@@ -15413,7 +15476,7 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("_unellipsify", 0);
 
-  /* "View.MemoryView":671
+  /* "View.MemoryView":673
  *     full slices.
  *     """
  *     if not isinstance(index, tuple):             # <<<<<<<<<<<<<<
@@ -15424,14 +15487,14 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
   __pyx_t_2 = ((!(__pyx_t_1 != 0)) != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":672
+    /* "View.MemoryView":674
  *     """
  *     if not isinstance(index, tuple):
  *         tup = (index,)             # <<<<<<<<<<<<<<
  *     else:
  *         tup = index
  */
-    __pyx_t_3 = PyTuple_New(1); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 672, __pyx_L1_error)
+    __pyx_t_3 = PyTuple_New(1); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 674, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
     __Pyx_INCREF(__pyx_v_index);
     __Pyx_GIVEREF(__pyx_v_index);
@@ -15439,7 +15502,7 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
     __pyx_v_tup = __pyx_t_3;
     __pyx_t_3 = 0;
 
-    /* "View.MemoryView":671
+    /* "View.MemoryView":673
  *     full slices.
  *     """
  *     if not isinstance(index, tuple):             # <<<<<<<<<<<<<<
@@ -15449,7 +15512,7 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
     goto __pyx_L3;
   }
 
-  /* "View.MemoryView":674
+  /* "View.MemoryView":676
  *         tup = (index,)
  *     else:
  *         tup = index             # <<<<<<<<<<<<<<
@@ -15462,19 +15525,19 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
   }
   __pyx_L3:;
 
-  /* "View.MemoryView":676
+  /* "View.MemoryView":678
  *         tup = index
  * 
  *     result = []             # <<<<<<<<<<<<<<
  *     have_slices = False
  *     seen_ellipsis = False
  */
-  __pyx_t_3 = PyList_New(0); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 676, __pyx_L1_error)
+  __pyx_t_3 = PyList_New(0); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 678, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_3);
   __pyx_v_result = ((PyObject*)__pyx_t_3);
   __pyx_t_3 = 0;
 
-  /* "View.MemoryView":677
+  /* "View.MemoryView":679
  * 
  *     result = []
  *     have_slices = False             # <<<<<<<<<<<<<<
@@ -15483,7 +15546,7 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
  */
   __pyx_v_have_slices = 0;
 
-  /* "View.MemoryView":678
+  /* "View.MemoryView":680
  *     result = []
  *     have_slices = False
  *     seen_ellipsis = False             # <<<<<<<<<<<<<<
@@ -15492,7 +15555,7 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
  */
   __pyx_v_seen_ellipsis = 0;
 
-  /* "View.MemoryView":679
+  /* "View.MemoryView":681
  *     have_slices = False
  *     seen_ellipsis = False
  *     for idx, item in enumerate(tup):             # <<<<<<<<<<<<<<
@@ -15505,26 +15568,26 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
     __pyx_t_4 = __pyx_v_tup; __Pyx_INCREF(__pyx_t_4); __pyx_t_5 = 0;
     __pyx_t_6 = NULL;
   } else {
-    __pyx_t_5 = -1; __pyx_t_4 = PyObject_GetIter(__pyx_v_tup); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 679, __pyx_L1_error)
+    __pyx_t_5 = -1; __pyx_t_4 = PyObject_GetIter(__pyx_v_tup); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 681, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_4);
-    __pyx_t_6 = Py_TYPE(__pyx_t_4)->tp_iternext; if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 679, __pyx_L1_error)
+    __pyx_t_6 = Py_TYPE(__pyx_t_4)->tp_iternext; if (unlikely(!__pyx_t_6)) __PYX_ERR(1, 681, __pyx_L1_error)
   }
   for (;;) {
     if (likely(!__pyx_t_6)) {
       if (likely(PyList_CheckExact(__pyx_t_4))) {
         if (__pyx_t_5 >= PyList_GET_SIZE(__pyx_t_4)) break;
         #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS
-        __pyx_t_7 = PyList_GET_ITEM(__pyx_t_4, __pyx_t_5); __Pyx_INCREF(__pyx_t_7); __pyx_t_5++; if (unlikely(0 < 0)) __PYX_ERR(1, 679, __pyx_L1_error)
+        __pyx_t_7 = PyList_GET_ITEM(__pyx_t_4, __pyx_t_5); __Pyx_INCREF(__pyx_t_7); __pyx_t_5++; if (unlikely(0 < 0)) __PYX_ERR(1, 681, __pyx_L1_error)
         #else
-        __pyx_t_7 = PySequence_ITEM(__pyx_t_4, __pyx_t_5); __pyx_t_5++; if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 679, __pyx_L1_error)
+        __pyx_t_7 = PySequence_ITEM(__pyx_t_4, __pyx_t_5); __pyx_t_5++; if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 681, __pyx_L1_error)
         __Pyx_GOTREF(__pyx_t_7);
         #endif
       } else {
         if (__pyx_t_5 >= PyTuple_GET_SIZE(__pyx_t_4)) break;
         #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS
-        __pyx_t_7 = PyTuple_GET_ITEM(__pyx_t_4, __pyx_t_5); __Pyx_INCREF(__pyx_t_7); __pyx_t_5++; if (unlikely(0 < 0)) __PYX_ERR(1, 679, __pyx_L1_error)
+        __pyx_t_7 = PyTuple_GET_ITEM(__pyx_t_4, __pyx_t_5); __Pyx_INCREF(__pyx_t_7); __pyx_t_5++; if (unlikely(0 < 0)) __PYX_ERR(1, 681, __pyx_L1_error)
         #else
-        __pyx_t_7 = PySequence_ITEM(__pyx_t_4, __pyx_t_5); __pyx_t_5++; if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 679, __pyx_L1_error)
+        __pyx_t_7 = PySequence_ITEM(__pyx_t_4, __pyx_t_5); __pyx_t_5++; if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 681, __pyx_L1_error)
         __Pyx_GOTREF(__pyx_t_7);
         #endif
       }
@@ -15534,7 +15597,7 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
         PyObject* exc_type = PyErr_Occurred();
         if (exc_type) {
           if (likely(__Pyx_PyErr_GivenExceptionMatches(exc_type, PyExc_StopIteration))) PyErr_Clear();
-          else __PYX_ERR(1, 679, __pyx_L1_error)
+          else __PYX_ERR(1, 681, __pyx_L1_error)
         }
         break;
       }
@@ -15544,13 +15607,13 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
     __pyx_t_7 = 0;
     __Pyx_INCREF(__pyx_t_3);
     __Pyx_XDECREF_SET(__pyx_v_idx, __pyx_t_3);
-    __pyx_t_7 = __Pyx_PyInt_AddObjC(__pyx_t_3, __pyx_int_1, 1, 0, 0); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 679, __pyx_L1_error)
+    __pyx_t_7 = __Pyx_PyInt_AddObjC(__pyx_t_3, __pyx_int_1, 1, 0, 0); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 681, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_7);
     __Pyx_DECREF(__pyx_t_3);
     __pyx_t_3 = __pyx_t_7;
     __pyx_t_7 = 0;
 
-    /* "View.MemoryView":680
+    /* "View.MemoryView":682
  *     seen_ellipsis = False
  *     for idx, item in enumerate(tup):
  *         if item is Ellipsis:             # <<<<<<<<<<<<<<
@@ -15561,7 +15624,7 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
     __pyx_t_1 = (__pyx_t_2 != 0);
     if (__pyx_t_1) {
 
-      /* "View.MemoryView":681
+      /* "View.MemoryView":683
  *     for idx, item in enumerate(tup):
  *         if item is Ellipsis:
  *             if not seen_ellipsis:             # <<<<<<<<<<<<<<
@@ -15571,15 +15634,15 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
       __pyx_t_1 = ((!(__pyx_v_seen_ellipsis != 0)) != 0);
       if (__pyx_t_1) {
 
-        /* "View.MemoryView":682
+        /* "View.MemoryView":684
  *         if item is Ellipsis:
  *             if not seen_ellipsis:
  *                 result.extend([slice(None)] * (ndim - len(tup) + 1))             # <<<<<<<<<<<<<<
  *                 seen_ellipsis = True
  *             else:
  */
-        __pyx_t_8 = PyObject_Length(__pyx_v_tup); if (unlikely(__pyx_t_8 == ((Py_ssize_t)-1))) __PYX_ERR(1, 682, __pyx_L1_error)
-        __pyx_t_7 = PyList_New(1 * ((((__pyx_v_ndim - __pyx_t_8) + 1)<0) ? 0:((__pyx_v_ndim - __pyx_t_8) + 1))); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 682, __pyx_L1_error)
+        __pyx_t_8 = PyObject_Length(__pyx_v_tup); if (unlikely(__pyx_t_8 == ((Py_ssize_t)-1))) __PYX_ERR(1, 684, __pyx_L1_error)
+        __pyx_t_7 = PyList_New(1 * ((((__pyx_v_ndim - __pyx_t_8) + 1)<0) ? 0:((__pyx_v_ndim - __pyx_t_8) + 1))); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 684, __pyx_L1_error)
         __Pyx_GOTREF(__pyx_t_7);
         { Py_ssize_t __pyx_temp;
           for (__pyx_temp=0; __pyx_temp < ((__pyx_v_ndim - __pyx_t_8) + 1); __pyx_temp++) {
@@ -15588,10 +15651,10 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
             PyList_SET_ITEM(__pyx_t_7, __pyx_temp, __pyx_slice__30);
           }
         }
-        __pyx_t_9 = __Pyx_PyList_Extend(__pyx_v_result, __pyx_t_7); if (unlikely(__pyx_t_9 == ((int)-1))) __PYX_ERR(1, 682, __pyx_L1_error)
+        __pyx_t_9 = __Pyx_PyList_Extend(__pyx_v_result, __pyx_t_7); if (unlikely(__pyx_t_9 == ((int)-1))) __PYX_ERR(1, 684, __pyx_L1_error)
         __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0;
 
-        /* "View.MemoryView":683
+        /* "View.MemoryView":685
  *             if not seen_ellipsis:
  *                 result.extend([slice(None)] * (ndim - len(tup) + 1))
  *                 seen_ellipsis = True             # <<<<<<<<<<<<<<
@@ -15600,7 +15663,7 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
  */
         __pyx_v_seen_ellipsis = 1;
 
-        /* "View.MemoryView":681
+        /* "View.MemoryView":683
  *     for idx, item in enumerate(tup):
  *         if item is Ellipsis:
  *             if not seen_ellipsis:             # <<<<<<<<<<<<<<
@@ -15610,7 +15673,7 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
         goto __pyx_L7;
       }
 
-      /* "View.MemoryView":685
+      /* "View.MemoryView":687
  *                 seen_ellipsis = True
  *             else:
  *                 result.append(slice(None))             # <<<<<<<<<<<<<<
@@ -15618,11 +15681,11 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
  *         else:
  */
       /*else*/ {
-        __pyx_t_9 = __Pyx_PyList_Append(__pyx_v_result, __pyx_slice__30); if (unlikely(__pyx_t_9 == ((int)-1))) __PYX_ERR(1, 685, __pyx_L1_error)
+        __pyx_t_9 = __Pyx_PyList_Append(__pyx_v_result, __pyx_slice__30); if (unlikely(__pyx_t_9 == ((int)-1))) __PYX_ERR(1, 687, __pyx_L1_error)
       }
       __pyx_L7:;
 
-      /* "View.MemoryView":686
+      /* "View.MemoryView":688
  *             else:
  *                 result.append(slice(None))
  *             have_slices = True             # <<<<<<<<<<<<<<
@@ -15631,7 +15694,7 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
  */
       __pyx_v_have_slices = 1;
 
-      /* "View.MemoryView":680
+      /* "View.MemoryView":682
  *     seen_ellipsis = False
  *     for idx, item in enumerate(tup):
  *         if item is Ellipsis:             # <<<<<<<<<<<<<<
@@ -15641,7 +15704,7 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
       goto __pyx_L6;
     }
 
-    /* "View.MemoryView":688
+    /* "View.MemoryView":690
  *             have_slices = True
  *         else:
  *             if not isinstance(item, slice) and not PyIndex_Check(item):             # <<<<<<<<<<<<<<
@@ -15661,23 +15724,23 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
       __pyx_L9_bool_binop_done:;
       if (unlikely(__pyx_t_1)) {
 
-        /* "View.MemoryView":689
+        /* "View.MemoryView":691
  *         else:
  *             if not isinstance(item, slice) and not PyIndex_Check(item):
  *                 raise TypeError("Cannot index with type '%s'" % type(item))             # <<<<<<<<<<<<<<
  * 
  *             have_slices = have_slices or isinstance(item, slice)
  */
-        __pyx_t_7 = __Pyx_PyString_FormatSafe(__pyx_kp_s_Cannot_index_with_type_s, ((PyObject *)Py_TYPE(__pyx_v_item))); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 689, __pyx_L1_error)
+        __pyx_t_7 = __Pyx_PyString_FormatSafe(__pyx_kp_s_Cannot_index_with_type_s, ((PyObject *)Py_TYPE(__pyx_v_item))); if (unlikely(!__pyx_t_7)) __PYX_ERR(1, 691, __pyx_L1_error)
         __Pyx_GOTREF(__pyx_t_7);
-        __pyx_t_11 = __Pyx_PyObject_CallOneArg(__pyx_builtin_TypeError, __pyx_t_7); if (unlikely(!__pyx_t_11)) __PYX_ERR(1, 689, __pyx_L1_error)
+        __pyx_t_11 = __Pyx_PyObject_CallOneArg(__pyx_builtin_TypeError, __pyx_t_7); if (unlikely(!__pyx_t_11)) __PYX_ERR(1, 691, __pyx_L1_error)
         __Pyx_GOTREF(__pyx_t_11);
         __Pyx_DECREF(__pyx_t_7); __pyx_t_7 = 0;
         __Pyx_Raise(__pyx_t_11, 0, 0, 0);
         __Pyx_DECREF(__pyx_t_11); __pyx_t_11 = 0;
-        __PYX_ERR(1, 689, __pyx_L1_error)
+        __PYX_ERR(1, 691, __pyx_L1_error)
 
-        /* "View.MemoryView":688
+        /* "View.MemoryView":690
  *             have_slices = True
  *         else:
  *             if not isinstance(item, slice) and not PyIndex_Check(item):             # <<<<<<<<<<<<<<
@@ -15686,7 +15749,7 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
  */
       }
 
-      /* "View.MemoryView":691
+      /* "View.MemoryView":693
  *                 raise TypeError("Cannot index with type '%s'" % type(item))
  * 
  *             have_slices = have_slices or isinstance(item, slice)             # <<<<<<<<<<<<<<
@@ -15705,18 +15768,18 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
       __pyx_L11_bool_binop_done:;
       __pyx_v_have_slices = __pyx_t_1;
 
-      /* "View.MemoryView":692
+      /* "View.MemoryView":694
  * 
  *             have_slices = have_slices or isinstance(item, slice)
  *             result.append(item)             # <<<<<<<<<<<<<<
  * 
  *     nslices = ndim - len(result)
  */
-      __pyx_t_9 = __Pyx_PyList_Append(__pyx_v_result, __pyx_v_item); if (unlikely(__pyx_t_9 == ((int)-1))) __PYX_ERR(1, 692, __pyx_L1_error)
+      __pyx_t_9 = __Pyx_PyList_Append(__pyx_v_result, __pyx_v_item); if (unlikely(__pyx_t_9 == ((int)-1))) __PYX_ERR(1, 694, __pyx_L1_error)
     }
     __pyx_L6:;
 
-    /* "View.MemoryView":679
+    /* "View.MemoryView":681
  *     have_slices = False
  *     seen_ellipsis = False
  *     for idx, item in enumerate(tup):             # <<<<<<<<<<<<<<
@@ -15727,17 +15790,17 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
   __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
   __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
 
-  /* "View.MemoryView":694
+  /* "View.MemoryView":696
  *             result.append(item)
  * 
  *     nslices = ndim - len(result)             # <<<<<<<<<<<<<<
  *     if nslices:
  *         result.extend([slice(None)] * nslices)
  */
-  __pyx_t_5 = PyList_GET_SIZE(__pyx_v_result); if (unlikely(__pyx_t_5 == ((Py_ssize_t)-1))) __PYX_ERR(1, 694, __pyx_L1_error)
+  __pyx_t_5 = PyList_GET_SIZE(__pyx_v_result); if (unlikely(__pyx_t_5 == ((Py_ssize_t)-1))) __PYX_ERR(1, 696, __pyx_L1_error)
   __pyx_v_nslices = (__pyx_v_ndim - __pyx_t_5);
 
-  /* "View.MemoryView":695
+  /* "View.MemoryView":697
  * 
  *     nslices = ndim - len(result)
  *     if nslices:             # <<<<<<<<<<<<<<
@@ -15747,14 +15810,14 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
   __pyx_t_1 = (__pyx_v_nslices != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":696
+    /* "View.MemoryView":698
  *     nslices = ndim - len(result)
  *     if nslices:
  *         result.extend([slice(None)] * nslices)             # <<<<<<<<<<<<<<
  * 
  *     return have_slices or nslices, tuple(result)
  */
-    __pyx_t_3 = PyList_New(1 * ((__pyx_v_nslices<0) ? 0:__pyx_v_nslices)); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 696, __pyx_L1_error)
+    __pyx_t_3 = PyList_New(1 * ((__pyx_v_nslices<0) ? 0:__pyx_v_nslices)); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 698, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
     { Py_ssize_t __pyx_temp;
       for (__pyx_temp=0; __pyx_temp < __pyx_v_nslices; __pyx_temp++) {
@@ -15763,10 +15826,10 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
         PyList_SET_ITEM(__pyx_t_3, __pyx_temp, __pyx_slice__30);
       }
     }
-    __pyx_t_9 = __Pyx_PyList_Extend(__pyx_v_result, __pyx_t_3); if (unlikely(__pyx_t_9 == ((int)-1))) __PYX_ERR(1, 696, __pyx_L1_error)
+    __pyx_t_9 = __Pyx_PyList_Extend(__pyx_v_result, __pyx_t_3); if (unlikely(__pyx_t_9 == ((int)-1))) __PYX_ERR(1, 698, __pyx_L1_error)
     __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
 
-    /* "View.MemoryView":695
+    /* "View.MemoryView":697
  * 
  *     nslices = ndim - len(result)
  *     if nslices:             # <<<<<<<<<<<<<<
@@ -15775,7 +15838,7 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
  */
   }
 
-  /* "View.MemoryView":698
+  /* "View.MemoryView":700
  *         result.extend([slice(None)] * nslices)
  * 
  *     return have_slices or nslices, tuple(result)             # <<<<<<<<<<<<<<
@@ -15785,20 +15848,20 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
   __Pyx_XDECREF(__pyx_r);
   if (!__pyx_v_have_slices) {
   } else {
-    __pyx_t_4 = __Pyx_PyBool_FromLong(__pyx_v_have_slices); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 698, __pyx_L1_error)
+    __pyx_t_4 = __Pyx_PyBool_FromLong(__pyx_v_have_slices); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 700, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_4);
     __pyx_t_3 = __pyx_t_4;
     __pyx_t_4 = 0;
     goto __pyx_L14_bool_binop_done;
   }
-  __pyx_t_4 = PyInt_FromSsize_t(__pyx_v_nslices); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 698, __pyx_L1_error)
+  __pyx_t_4 = PyInt_FromSsize_t(__pyx_v_nslices); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 700, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_4);
   __pyx_t_3 = __pyx_t_4;
   __pyx_t_4 = 0;
   __pyx_L14_bool_binop_done:;
-  __pyx_t_4 = PyList_AsTuple(__pyx_v_result); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 698, __pyx_L1_error)
+  __pyx_t_4 = PyList_AsTuple(__pyx_v_result); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 700, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_4);
-  __pyx_t_11 = PyTuple_New(2); if (unlikely(!__pyx_t_11)) __PYX_ERR(1, 698, __pyx_L1_error)
+  __pyx_t_11 = PyTuple_New(2); if (unlikely(!__pyx_t_11)) __PYX_ERR(1, 700, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_11);
   __Pyx_GIVEREF(__pyx_t_3);
   PyTuple_SET_ITEM(__pyx_t_11, 0, __pyx_t_3);
@@ -15810,7 +15873,7 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
   __pyx_t_11 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":666
+  /* "View.MemoryView":668
  *     return isinstance(o, memoryview)
  * 
  * cdef tuple _unellipsify(object index, int ndim):             # <<<<<<<<<<<<<<
@@ -15836,7 +15899,7 @@ static PyObject *_unellipsify(PyObject *__pyx_v_index, int __pyx_v_ndim) {
   return __pyx_r;
 }
 
-/* "View.MemoryView":700
+/* "View.MemoryView":702
  *     return have_slices or nslices, tuple(result)
  * 
  * cdef assert_direct_dimensions(Py_ssize_t *suboffsets, int ndim):             # <<<<<<<<<<<<<<
@@ -15858,7 +15921,7 @@ static PyObject *assert_direct_dimensions(Py_ssize_t *__pyx_v_suboffsets, int __
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("assert_direct_dimensions", 0);
 
-  /* "View.MemoryView":701
+  /* "View.MemoryView":703
  * 
  * cdef assert_direct_dimensions(Py_ssize_t *suboffsets, int ndim):
  *     for suboffset in suboffsets[:ndim]:             # <<<<<<<<<<<<<<
@@ -15870,7 +15933,7 @@ static PyObject *assert_direct_dimensions(Py_ssize_t *__pyx_v_suboffsets, int __
     __pyx_t_1 = __pyx_t_3;
     __pyx_v_suboffset = (__pyx_t_1[0]);
 
-    /* "View.MemoryView":702
+    /* "View.MemoryView":704
  * cdef assert_direct_dimensions(Py_ssize_t *suboffsets, int ndim):
  *     for suboffset in suboffsets[:ndim]:
  *         if suboffset >= 0:             # <<<<<<<<<<<<<<
@@ -15880,20 +15943,20 @@ static PyObject *assert_direct_dimensions(Py_ssize_t *__pyx_v_suboffsets, int __
     __pyx_t_4 = ((__pyx_v_suboffset >= 0) != 0);
     if (unlikely(__pyx_t_4)) {
 
-      /* "View.MemoryView":703
+      /* "View.MemoryView":705
  *     for suboffset in suboffsets[:ndim]:
  *         if suboffset >= 0:
  *             raise ValueError("Indirect dimensions not supported")             # <<<<<<<<<<<<<<
  * 
  * 
  */
-      __pyx_t_5 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__31, NULL); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 703, __pyx_L1_error)
+      __pyx_t_5 = __Pyx_PyObject_Call(__pyx_builtin_ValueError, __pyx_tuple__31, NULL); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 705, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_5);
       __Pyx_Raise(__pyx_t_5, 0, 0, 0);
       __Pyx_DECREF(__pyx_t_5); __pyx_t_5 = 0;
-      __PYX_ERR(1, 703, __pyx_L1_error)
+      __PYX_ERR(1, 705, __pyx_L1_error)
 
-      /* "View.MemoryView":702
+      /* "View.MemoryView":704
  * cdef assert_direct_dimensions(Py_ssize_t *suboffsets, int ndim):
  *     for suboffset in suboffsets[:ndim]:
  *         if suboffset >= 0:             # <<<<<<<<<<<<<<
@@ -15903,7 +15966,7 @@ static PyObject *assert_direct_dimensions(Py_ssize_t *__pyx_v_suboffsets, int __
     }
   }
 
-  /* "View.MemoryView":700
+  /* "View.MemoryView":702
  *     return have_slices or nslices, tuple(result)
  * 
  * cdef assert_direct_dimensions(Py_ssize_t *suboffsets, int ndim):             # <<<<<<<<<<<<<<
@@ -15924,7 +15987,7 @@ static PyObject *assert_direct_dimensions(Py_ssize_t *__pyx_v_suboffsets, int __
   return __pyx_r;
 }
 
-/* "View.MemoryView":710
+/* "View.MemoryView":712
  * 
  * @cname('__pyx_memview_slice')
  * cdef memoryview memview_slice(memoryview memview, object indices):             # <<<<<<<<<<<<<<
@@ -15968,7 +16031,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("memview_slice", 0);
 
-  /* "View.MemoryView":711
+  /* "View.MemoryView":713
  * @cname('__pyx_memview_slice')
  * cdef memoryview memview_slice(memoryview memview, object indices):
  *     cdef int new_ndim = 0, suboffset_dim = -1, dim             # <<<<<<<<<<<<<<
@@ -15978,7 +16041,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
   __pyx_v_new_ndim = 0;
   __pyx_v_suboffset_dim = -1;
 
-  /* "View.MemoryView":718
+  /* "View.MemoryView":720
  * 
  * 
  *     memset(&dst, 0, sizeof(dst))             # <<<<<<<<<<<<<<
@@ -15987,7 +16050,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
  */
   (void)(memset((&__pyx_v_dst), 0, (sizeof(__pyx_v_dst))));
 
-  /* "View.MemoryView":722
+  /* "View.MemoryView":724
  *     cdef _memoryviewslice memviewsliceobj
  * 
  *     assert memview.view.ndim > 0             # <<<<<<<<<<<<<<
@@ -15998,12 +16061,12 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
   if (unlikely(!Py_OptimizeFlag)) {
     if (unlikely(!((__pyx_v_memview->view.ndim > 0) != 0))) {
       PyErr_SetNone(PyExc_AssertionError);
-      __PYX_ERR(1, 722, __pyx_L1_error)
+      __PYX_ERR(1, 724, __pyx_L1_error)
     }
   }
   #endif
 
-  /* "View.MemoryView":724
+  /* "View.MemoryView":726
  *     assert memview.view.ndim > 0
  * 
  *     if isinstance(memview, _memoryviewslice):             # <<<<<<<<<<<<<<
@@ -16014,20 +16077,20 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
   __pyx_t_2 = (__pyx_t_1 != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":725
+    /* "View.MemoryView":727
  * 
  *     if isinstance(memview, _memoryviewslice):
  *         memviewsliceobj = memview             # <<<<<<<<<<<<<<
  *         p_src = &memviewsliceobj.from_slice
  *     else:
  */
-    if (!(likely(((((PyObject *)__pyx_v_memview)) == Py_None) || likely(__Pyx_TypeTest(((PyObject *)__pyx_v_memview), __pyx_memoryviewslice_type))))) __PYX_ERR(1, 725, __pyx_L1_error)
+    if (!(likely(((((PyObject *)__pyx_v_memview)) == Py_None) || likely(__Pyx_TypeTest(((PyObject *)__pyx_v_memview), __pyx_memoryviewslice_type))))) __PYX_ERR(1, 727, __pyx_L1_error)
     __pyx_t_3 = ((PyObject *)__pyx_v_memview);
     __Pyx_INCREF(__pyx_t_3);
     __pyx_v_memviewsliceobj = ((struct __pyx_memoryviewslice_obj *)__pyx_t_3);
     __pyx_t_3 = 0;
 
-    /* "View.MemoryView":726
+    /* "View.MemoryView":728
  *     if isinstance(memview, _memoryviewslice):
  *         memviewsliceobj = memview
  *         p_src = &memviewsliceobj.from_slice             # <<<<<<<<<<<<<<
@@ -16036,7 +16099,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
  */
     __pyx_v_p_src = (&__pyx_v_memviewsliceobj->from_slice);
 
-    /* "View.MemoryView":724
+    /* "View.MemoryView":726
  *     assert memview.view.ndim > 0
  * 
  *     if isinstance(memview, _memoryviewslice):             # <<<<<<<<<<<<<<
@@ -16046,7 +16109,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
     goto __pyx_L3;
   }
 
-  /* "View.MemoryView":728
+  /* "View.MemoryView":730
  *         p_src = &memviewsliceobj.from_slice
  *     else:
  *         slice_copy(memview, &src)             # <<<<<<<<<<<<<<
@@ -16056,7 +16119,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
   /*else*/ {
     __pyx_memoryview_slice_copy(__pyx_v_memview, (&__pyx_v_src));
 
-    /* "View.MemoryView":729
+    /* "View.MemoryView":731
  *     else:
  *         slice_copy(memview, &src)
  *         p_src = &src             # <<<<<<<<<<<<<<
@@ -16067,7 +16130,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
   }
   __pyx_L3:;
 
-  /* "View.MemoryView":735
+  /* "View.MemoryView":737
  * 
  * 
  *     dst.memview = p_src.memview             # <<<<<<<<<<<<<<
@@ -16077,7 +16140,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
   __pyx_t_4 = __pyx_v_p_src->memview;
   __pyx_v_dst.memview = __pyx_t_4;
 
-  /* "View.MemoryView":736
+  /* "View.MemoryView":738
  * 
  *     dst.memview = p_src.memview
  *     dst.data = p_src.data             # <<<<<<<<<<<<<<
@@ -16087,7 +16150,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
   __pyx_t_5 = __pyx_v_p_src->data;
   __pyx_v_dst.data = __pyx_t_5;
 
-  /* "View.MemoryView":741
+  /* "View.MemoryView":743
  * 
  * 
  *     cdef __Pyx_memviewslice *p_dst = &dst             # <<<<<<<<<<<<<<
@@ -16096,7 +16159,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
  */
   __pyx_v_p_dst = (&__pyx_v_dst);
 
-  /* "View.MemoryView":742
+  /* "View.MemoryView":744
  * 
  *     cdef __Pyx_memviewslice *p_dst = &dst
  *     cdef int *p_suboffset_dim = &suboffset_dim             # <<<<<<<<<<<<<<
@@ -16105,7 +16168,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
  */
   __pyx_v_p_suboffset_dim = (&__pyx_v_suboffset_dim);
 
-  /* "View.MemoryView":746
+  /* "View.MemoryView":748
  *     cdef bint have_start, have_stop, have_step
  * 
  *     for dim, index in enumerate(indices):             # <<<<<<<<<<<<<<
@@ -16117,26 +16180,26 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
     __pyx_t_3 = __pyx_v_indices; __Pyx_INCREF(__pyx_t_3); __pyx_t_7 = 0;
     __pyx_t_8 = NULL;
   } else {
-    __pyx_t_7 = -1; __pyx_t_3 = PyObject_GetIter(__pyx_v_indices); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 746, __pyx_L1_error)
+    __pyx_t_7 = -1; __pyx_t_3 = PyObject_GetIter(__pyx_v_indices); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 748, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
-    __pyx_t_8 = Py_TYPE(__pyx_t_3)->tp_iternext; if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 746, __pyx_L1_error)
+    __pyx_t_8 = Py_TYPE(__pyx_t_3)->tp_iternext; if (unlikely(!__pyx_t_8)) __PYX_ERR(1, 748, __pyx_L1_error)
   }
   for (;;) {
     if (likely(!__pyx_t_8)) {
       if (likely(PyList_CheckExact(__pyx_t_3))) {
         if (__pyx_t_7 >= PyList_GET_SIZE(__pyx_t_3)) break;
         #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS
-        __pyx_t_9 = PyList_GET_ITEM(__pyx_t_3, __pyx_t_7); __Pyx_INCREF(__pyx_t_9); __pyx_t_7++; if (unlikely(0 < 0)) __PYX_ERR(1, 746, __pyx_L1_error)
+        __pyx_t_9 = PyList_GET_ITEM(__pyx_t_3, __pyx_t_7); __Pyx_INCREF(__pyx_t_9); __pyx_t_7++; if (unlikely(0 < 0)) __PYX_ERR(1, 748, __pyx_L1_error)
         #else
-        __pyx_t_9 = PySequence_ITEM(__pyx_t_3, __pyx_t_7); __pyx_t_7++; if (unlikely(!__pyx_t_9)) __PYX_ERR(1, 746, __pyx_L1_error)
+        __pyx_t_9 = PySequence_ITEM(__pyx_t_3, __pyx_t_7); __pyx_t_7++; if (unlikely(!__pyx_t_9)) __PYX_ERR(1, 748, __pyx_L1_error)
         __Pyx_GOTREF(__pyx_t_9);
         #endif
       } else {
         if (__pyx_t_7 >= PyTuple_GET_SIZE(__pyx_t_3)) break;
         #if CYTHON_ASSUME_SAFE_MACROS && !CYTHON_AVOID_BORROWED_REFS
-        __pyx_t_9 = PyTuple_GET_ITEM(__pyx_t_3, __pyx_t_7); __Pyx_INCREF(__pyx_t_9); __pyx_t_7++; if (unlikely(0 < 0)) __PYX_ERR(1, 746, __pyx_L1_error)
+        __pyx_t_9 = PyTuple_GET_ITEM(__pyx_t_3, __pyx_t_7); __Pyx_INCREF(__pyx_t_9); __pyx_t_7++; if (unlikely(0 < 0)) __PYX_ERR(1, 748, __pyx_L1_error)
         #else
-        __pyx_t_9 = PySequence_ITEM(__pyx_t_3, __pyx_t_7); __pyx_t_7++; if (unlikely(!__pyx_t_9)) __PYX_ERR(1, 746, __pyx_L1_error)
+        __pyx_t_9 = PySequence_ITEM(__pyx_t_3, __pyx_t_7); __pyx_t_7++; if (unlikely(!__pyx_t_9)) __PYX_ERR(1, 748, __pyx_L1_error)
         __Pyx_GOTREF(__pyx_t_9);
         #endif
       }
@@ -16146,7 +16209,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
         PyObject* exc_type = PyErr_Occurred();
         if (exc_type) {
           if (likely(__Pyx_PyErr_GivenExceptionMatches(exc_type, PyExc_StopIteration))) PyErr_Clear();
-          else __PYX_ERR(1, 746, __pyx_L1_error)
+          else __PYX_ERR(1, 748, __pyx_L1_error)
         }
         break;
       }
@@ -16157,7 +16220,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
     __pyx_v_dim = __pyx_t_6;
     __pyx_t_6 = (__pyx_t_6 + 1);
 
-    /* "View.MemoryView":747
+    /* "View.MemoryView":749
  * 
  *     for dim, index in enumerate(indices):
  *         if PyIndex_Check(index):             # <<<<<<<<<<<<<<
@@ -16167,25 +16230,25 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
     __pyx_t_2 = (PyIndex_Check(__pyx_v_index) != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":751
+      /* "View.MemoryView":753
  *                 p_dst, p_src.shape[dim], p_src.strides[dim], p_src.suboffsets[dim],
  *                 dim, new_ndim, p_suboffset_dim,
  *                 index, 0, 0, # start, stop, step             # <<<<<<<<<<<<<<
  *                 0, 0, 0, # have_{start,stop,step}
  *                 False)
  */
-      __pyx_t_10 = __Pyx_PyIndex_AsSsize_t(__pyx_v_index); if (unlikely((__pyx_t_10 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 751, __pyx_L1_error)
+      __pyx_t_10 = __Pyx_PyIndex_AsSsize_t(__pyx_v_index); if (unlikely((__pyx_t_10 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 753, __pyx_L1_error)
 
-      /* "View.MemoryView":748
+      /* "View.MemoryView":750
  *     for dim, index in enumerate(indices):
  *         if PyIndex_Check(index):
  *             slice_memviewslice(             # <<<<<<<<<<<<<<
  *                 p_dst, p_src.shape[dim], p_src.strides[dim], p_src.suboffsets[dim],
  *                 dim, new_ndim, p_suboffset_dim,
  */
-      __pyx_t_11 = __pyx_memoryview_slice_memviewslice(__pyx_v_p_dst, (__pyx_v_p_src->shape[__pyx_v_dim]), (__pyx_v_p_src->strides[__pyx_v_dim]), (__pyx_v_p_src->suboffsets[__pyx_v_dim]), __pyx_v_dim, __pyx_v_new_ndim, __pyx_v_p_suboffset_dim, __pyx_t_10, 0, 0, 0, 0, 0, 0); if (unlikely(__pyx_t_11 == ((int)-1))) __PYX_ERR(1, 748, __pyx_L1_error)
+      __pyx_t_11 = __pyx_memoryview_slice_memviewslice(__pyx_v_p_dst, (__pyx_v_p_src->shape[__pyx_v_dim]), (__pyx_v_p_src->strides[__pyx_v_dim]), (__pyx_v_p_src->suboffsets[__pyx_v_dim]), __pyx_v_dim, __pyx_v_new_ndim, __pyx_v_p_suboffset_dim, __pyx_t_10, 0, 0, 0, 0, 0, 0); if (unlikely(__pyx_t_11 == ((int)-1))) __PYX_ERR(1, 750, __pyx_L1_error)
 
-      /* "View.MemoryView":747
+      /* "View.MemoryView":749
  * 
  *     for dim, index in enumerate(indices):
  *         if PyIndex_Check(index):             # <<<<<<<<<<<<<<
@@ -16195,7 +16258,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
       goto __pyx_L6;
     }
 
-    /* "View.MemoryView":754
+    /* "View.MemoryView":756
  *                 0, 0, 0, # have_{start,stop,step}
  *                 False)
  *         elif index is None:             # <<<<<<<<<<<<<<
@@ -16206,7 +16269,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
     __pyx_t_1 = (__pyx_t_2 != 0);
     if (__pyx_t_1) {
 
-      /* "View.MemoryView":755
+      /* "View.MemoryView":757
  *                 False)
  *         elif index is None:
  *             p_dst.shape[new_ndim] = 1             # <<<<<<<<<<<<<<
@@ -16215,7 +16278,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
  */
       (__pyx_v_p_dst->shape[__pyx_v_new_ndim]) = 1;
 
-      /* "View.MemoryView":756
+      /* "View.MemoryView":758
  *         elif index is None:
  *             p_dst.shape[new_ndim] = 1
  *             p_dst.strides[new_ndim] = 0             # <<<<<<<<<<<<<<
@@ -16224,7 +16287,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
  */
       (__pyx_v_p_dst->strides[__pyx_v_new_ndim]) = 0;
 
-      /* "View.MemoryView":757
+      /* "View.MemoryView":759
  *             p_dst.shape[new_ndim] = 1
  *             p_dst.strides[new_ndim] = 0
  *             p_dst.suboffsets[new_ndim] = -1             # <<<<<<<<<<<<<<
@@ -16233,7 +16296,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
  */
       (__pyx_v_p_dst->suboffsets[__pyx_v_new_ndim]) = -1L;
 
-      /* "View.MemoryView":758
+      /* "View.MemoryView":760
  *             p_dst.strides[new_ndim] = 0
  *             p_dst.suboffsets[new_ndim] = -1
  *             new_ndim += 1             # <<<<<<<<<<<<<<
@@ -16242,7 +16305,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
  */
       __pyx_v_new_ndim = (__pyx_v_new_ndim + 1);
 
-      /* "View.MemoryView":754
+      /* "View.MemoryView":756
  *                 0, 0, 0, # have_{start,stop,step}
  *                 False)
  *         elif index is None:             # <<<<<<<<<<<<<<
@@ -16252,7 +16315,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
       goto __pyx_L6;
     }
 
-    /* "View.MemoryView":760
+    /* "View.MemoryView":762
  *             new_ndim += 1
  *         else:
  *             start = index.start or 0             # <<<<<<<<<<<<<<
@@ -16260,13 +16323,13 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
  *             step = index.step or 0
  */
     /*else*/ {
-      __pyx_t_9 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_start); if (unlikely(!__pyx_t_9)) __PYX_ERR(1, 760, __pyx_L1_error)
+      __pyx_t_9 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_start); if (unlikely(!__pyx_t_9)) __PYX_ERR(1, 762, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_9);
-      __pyx_t_1 = __Pyx_PyObject_IsTrue(__pyx_t_9); if (unlikely(__pyx_t_1 < 0)) __PYX_ERR(1, 760, __pyx_L1_error)
+      __pyx_t_1 = __Pyx_PyObject_IsTrue(__pyx_t_9); if (unlikely(__pyx_t_1 < 0)) __PYX_ERR(1, 762, __pyx_L1_error)
       if (!__pyx_t_1) {
         __Pyx_DECREF(__pyx_t_9); __pyx_t_9 = 0;
       } else {
-        __pyx_t_12 = __Pyx_PyIndex_AsSsize_t(__pyx_t_9); if (unlikely((__pyx_t_12 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 760, __pyx_L1_error)
+        __pyx_t_12 = __Pyx_PyIndex_AsSsize_t(__pyx_t_9); if (unlikely((__pyx_t_12 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 762, __pyx_L1_error)
         __pyx_t_10 = __pyx_t_12;
         __Pyx_DECREF(__pyx_t_9); __pyx_t_9 = 0;
         goto __pyx_L7_bool_binop_done;
@@ -16275,20 +16338,20 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
       __pyx_L7_bool_binop_done:;
       __pyx_v_start = __pyx_t_10;
 
-      /* "View.MemoryView":761
+      /* "View.MemoryView":763
  *         else:
  *             start = index.start or 0
  *             stop = index.stop or 0             # <<<<<<<<<<<<<<
  *             step = index.step or 0
  * 
  */
-      __pyx_t_9 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_stop); if (unlikely(!__pyx_t_9)) __PYX_ERR(1, 761, __pyx_L1_error)
+      __pyx_t_9 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_stop); if (unlikely(!__pyx_t_9)) __PYX_ERR(1, 763, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_9);
-      __pyx_t_1 = __Pyx_PyObject_IsTrue(__pyx_t_9); if (unlikely(__pyx_t_1 < 0)) __PYX_ERR(1, 761, __pyx_L1_error)
+      __pyx_t_1 = __Pyx_PyObject_IsTrue(__pyx_t_9); if (unlikely(__pyx_t_1 < 0)) __PYX_ERR(1, 763, __pyx_L1_error)
       if (!__pyx_t_1) {
         __Pyx_DECREF(__pyx_t_9); __pyx_t_9 = 0;
       } else {
-        __pyx_t_12 = __Pyx_PyIndex_AsSsize_t(__pyx_t_9); if (unlikely((__pyx_t_12 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 761, __pyx_L1_error)
+        __pyx_t_12 = __Pyx_PyIndex_AsSsize_t(__pyx_t_9); if (unlikely((__pyx_t_12 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 763, __pyx_L1_error)
         __pyx_t_10 = __pyx_t_12;
         __Pyx_DECREF(__pyx_t_9); __pyx_t_9 = 0;
         goto __pyx_L9_bool_binop_done;
@@ -16297,20 +16360,20 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
       __pyx_L9_bool_binop_done:;
       __pyx_v_stop = __pyx_t_10;
 
-      /* "View.MemoryView":762
+      /* "View.MemoryView":764
  *             start = index.start or 0
  *             stop = index.stop or 0
  *             step = index.step or 0             # <<<<<<<<<<<<<<
  * 
  *             have_start = index.start is not None
  */
-      __pyx_t_9 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_step); if (unlikely(!__pyx_t_9)) __PYX_ERR(1, 762, __pyx_L1_error)
+      __pyx_t_9 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_step); if (unlikely(!__pyx_t_9)) __PYX_ERR(1, 764, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_9);
-      __pyx_t_1 = __Pyx_PyObject_IsTrue(__pyx_t_9); if (unlikely(__pyx_t_1 < 0)) __PYX_ERR(1, 762, __pyx_L1_error)
+      __pyx_t_1 = __Pyx_PyObject_IsTrue(__pyx_t_9); if (unlikely(__pyx_t_1 < 0)) __PYX_ERR(1, 764, __pyx_L1_error)
       if (!__pyx_t_1) {
         __Pyx_DECREF(__pyx_t_9); __pyx_t_9 = 0;
       } else {
-        __pyx_t_12 = __Pyx_PyIndex_AsSsize_t(__pyx_t_9); if (unlikely((__pyx_t_12 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 762, __pyx_L1_error)
+        __pyx_t_12 = __Pyx_PyIndex_AsSsize_t(__pyx_t_9); if (unlikely((__pyx_t_12 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 764, __pyx_L1_error)
         __pyx_t_10 = __pyx_t_12;
         __Pyx_DECREF(__pyx_t_9); __pyx_t_9 = 0;
         goto __pyx_L11_bool_binop_done;
@@ -16319,55 +16382,55 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
       __pyx_L11_bool_binop_done:;
       __pyx_v_step = __pyx_t_10;
 
-      /* "View.MemoryView":764
+      /* "View.MemoryView":766
  *             step = index.step or 0
  * 
  *             have_start = index.start is not None             # <<<<<<<<<<<<<<
  *             have_stop = index.stop is not None
  *             have_step = index.step is not None
  */
-      __pyx_t_9 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_start); if (unlikely(!__pyx_t_9)) __PYX_ERR(1, 764, __pyx_L1_error)
+      __pyx_t_9 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_start); if (unlikely(!__pyx_t_9)) __PYX_ERR(1, 766, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_9);
       __pyx_t_1 = (__pyx_t_9 != Py_None);
       __Pyx_DECREF(__pyx_t_9); __pyx_t_9 = 0;
       __pyx_v_have_start = __pyx_t_1;
 
-      /* "View.MemoryView":765
+      /* "View.MemoryView":767
  * 
  *             have_start = index.start is not None
  *             have_stop = index.stop is not None             # <<<<<<<<<<<<<<
  *             have_step = index.step is not None
  * 
  */
-      __pyx_t_9 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_stop); if (unlikely(!__pyx_t_9)) __PYX_ERR(1, 765, __pyx_L1_error)
+      __pyx_t_9 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_stop); if (unlikely(!__pyx_t_9)) __PYX_ERR(1, 767, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_9);
       __pyx_t_1 = (__pyx_t_9 != Py_None);
       __Pyx_DECREF(__pyx_t_9); __pyx_t_9 = 0;
       __pyx_v_have_stop = __pyx_t_1;
 
-      /* "View.MemoryView":766
+      /* "View.MemoryView":768
  *             have_start = index.start is not None
  *             have_stop = index.stop is not None
  *             have_step = index.step is not None             # <<<<<<<<<<<<<<
  * 
  *             slice_memviewslice(
  */
-      __pyx_t_9 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_step); if (unlikely(!__pyx_t_9)) __PYX_ERR(1, 766, __pyx_L1_error)
+      __pyx_t_9 = __Pyx_PyObject_GetAttrStr(__pyx_v_index, __pyx_n_s_step); if (unlikely(!__pyx_t_9)) __PYX_ERR(1, 768, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_9);
       __pyx_t_1 = (__pyx_t_9 != Py_None);
       __Pyx_DECREF(__pyx_t_9); __pyx_t_9 = 0;
       __pyx_v_have_step = __pyx_t_1;
 
-      /* "View.MemoryView":768
+      /* "View.MemoryView":770
  *             have_step = index.step is not None
  * 
  *             slice_memviewslice(             # <<<<<<<<<<<<<<
  *                 p_dst, p_src.shape[dim], p_src.strides[dim], p_src.suboffsets[dim],
  *                 dim, new_ndim, p_suboffset_dim,
  */
-      __pyx_t_11 = __pyx_memoryview_slice_memviewslice(__pyx_v_p_dst, (__pyx_v_p_src->shape[__pyx_v_dim]), (__pyx_v_p_src->strides[__pyx_v_dim]), (__pyx_v_p_src->suboffsets[__pyx_v_dim]), __pyx_v_dim, __pyx_v_new_ndim, __pyx_v_p_suboffset_dim, __pyx_v_start, __pyx_v_stop, __pyx_v_step, __pyx_v_have_start, __pyx_v_have_stop, __pyx_v_have_step, 1); if (unlikely(__pyx_t_11 == ((int)-1))) __PYX_ERR(1, 768, __pyx_L1_error)
+      __pyx_t_11 = __pyx_memoryview_slice_memviewslice(__pyx_v_p_dst, (__pyx_v_p_src->shape[__pyx_v_dim]), (__pyx_v_p_src->strides[__pyx_v_dim]), (__pyx_v_p_src->suboffsets[__pyx_v_dim]), __pyx_v_dim, __pyx_v_new_ndim, __pyx_v_p_suboffset_dim, __pyx_v_start, __pyx_v_stop, __pyx_v_step, __pyx_v_have_start, __pyx_v_have_stop, __pyx_v_have_step, 1); if (unlikely(__pyx_t_11 == ((int)-1))) __PYX_ERR(1, 770, __pyx_L1_error)
 
-      /* "View.MemoryView":774
+      /* "View.MemoryView":776
  *                 have_start, have_stop, have_step,
  *                 True)
  *             new_ndim += 1             # <<<<<<<<<<<<<<
@@ -16378,7 +16441,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
     }
     __pyx_L6:;
 
-    /* "View.MemoryView":746
+    /* "View.MemoryView":748
  *     cdef bint have_start, have_stop, have_step
  * 
  *     for dim, index in enumerate(indices):             # <<<<<<<<<<<<<<
@@ -16388,7 +16451,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
   }
   __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
 
-  /* "View.MemoryView":776
+  /* "View.MemoryView":778
  *             new_ndim += 1
  * 
  *     if isinstance(memview, _memoryviewslice):             # <<<<<<<<<<<<<<
@@ -16399,7 +16462,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
   __pyx_t_2 = (__pyx_t_1 != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":777
+    /* "View.MemoryView":779
  * 
  *     if isinstance(memview, _memoryviewslice):
  *         return memoryview_fromslice(dst, new_ndim,             # <<<<<<<<<<<<<<
@@ -16408,39 +16471,39 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
  */
     __Pyx_XDECREF(((PyObject *)__pyx_r));
 
-    /* "View.MemoryView":778
+    /* "View.MemoryView":780
  *     if isinstance(memview, _memoryviewslice):
  *         return memoryview_fromslice(dst, new_ndim,
  *                                     memviewsliceobj.to_object_func,             # <<<<<<<<<<<<<<
  *                                     memviewsliceobj.to_dtype_func,
  *                                     memview.dtype_is_object)
  */
-    if (unlikely(!__pyx_v_memviewsliceobj)) { __Pyx_RaiseUnboundLocalError("memviewsliceobj"); __PYX_ERR(1, 778, __pyx_L1_error) }
+    if (unlikely(!__pyx_v_memviewsliceobj)) { __Pyx_RaiseUnboundLocalError("memviewsliceobj"); __PYX_ERR(1, 780, __pyx_L1_error) }
 
-    /* "View.MemoryView":779
+    /* "View.MemoryView":781
  *         return memoryview_fromslice(dst, new_ndim,
  *                                     memviewsliceobj.to_object_func,
  *                                     memviewsliceobj.to_dtype_func,             # <<<<<<<<<<<<<<
  *                                     memview.dtype_is_object)
  *     else:
  */
-    if (unlikely(!__pyx_v_memviewsliceobj)) { __Pyx_RaiseUnboundLocalError("memviewsliceobj"); __PYX_ERR(1, 779, __pyx_L1_error) }
+    if (unlikely(!__pyx_v_memviewsliceobj)) { __Pyx_RaiseUnboundLocalError("memviewsliceobj"); __PYX_ERR(1, 781, __pyx_L1_error) }
 
-    /* "View.MemoryView":777
+    /* "View.MemoryView":779
  * 
  *     if isinstance(memview, _memoryviewslice):
  *         return memoryview_fromslice(dst, new_ndim,             # <<<<<<<<<<<<<<
  *                                     memviewsliceobj.to_object_func,
  *                                     memviewsliceobj.to_dtype_func,
  */
-    __pyx_t_3 = __pyx_memoryview_fromslice(__pyx_v_dst, __pyx_v_new_ndim, __pyx_v_memviewsliceobj->to_object_func, __pyx_v_memviewsliceobj->to_dtype_func, __pyx_v_memview->dtype_is_object); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 777, __pyx_L1_error)
+    __pyx_t_3 = __pyx_memoryview_fromslice(__pyx_v_dst, __pyx_v_new_ndim, __pyx_v_memviewsliceobj->to_object_func, __pyx_v_memviewsliceobj->to_dtype_func, __pyx_v_memview->dtype_is_object); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 779, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
-    if (!(likely(((__pyx_t_3) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_3, __pyx_memoryview_type))))) __PYX_ERR(1, 777, __pyx_L1_error)
+    if (!(likely(((__pyx_t_3) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_3, __pyx_memoryview_type))))) __PYX_ERR(1, 779, __pyx_L1_error)
     __pyx_r = ((struct __pyx_memoryview_obj *)__pyx_t_3);
     __pyx_t_3 = 0;
     goto __pyx_L0;
 
-    /* "View.MemoryView":776
+    /* "View.MemoryView":778
  *             new_ndim += 1
  * 
  *     if isinstance(memview, _memoryviewslice):             # <<<<<<<<<<<<<<
@@ -16449,7 +16512,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
  */
   }
 
-  /* "View.MemoryView":782
+  /* "View.MemoryView":784
  *                                     memview.dtype_is_object)
  *     else:
  *         return memoryview_fromslice(dst, new_ndim, NULL, NULL,             # <<<<<<<<<<<<<<
@@ -16459,30 +16522,30 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
   /*else*/ {
     __Pyx_XDECREF(((PyObject *)__pyx_r));
 
-    /* "View.MemoryView":783
+    /* "View.MemoryView":785
  *     else:
  *         return memoryview_fromslice(dst, new_ndim, NULL, NULL,
  *                                     memview.dtype_is_object)             # <<<<<<<<<<<<<<
  * 
  * 
  */
-    __pyx_t_3 = __pyx_memoryview_fromslice(__pyx_v_dst, __pyx_v_new_ndim, NULL, NULL, __pyx_v_memview->dtype_is_object); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 782, __pyx_L1_error)
+    __pyx_t_3 = __pyx_memoryview_fromslice(__pyx_v_dst, __pyx_v_new_ndim, NULL, NULL, __pyx_v_memview->dtype_is_object); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 784, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
 
-    /* "View.MemoryView":782
+    /* "View.MemoryView":784
  *                                     memview.dtype_is_object)
  *     else:
  *         return memoryview_fromslice(dst, new_ndim, NULL, NULL,             # <<<<<<<<<<<<<<
  *                                     memview.dtype_is_object)
  * 
  */
-    if (!(likely(((__pyx_t_3) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_3, __pyx_memoryview_type))))) __PYX_ERR(1, 782, __pyx_L1_error)
+    if (!(likely(((__pyx_t_3) == Py_None) || likely(__Pyx_TypeTest(__pyx_t_3, __pyx_memoryview_type))))) __PYX_ERR(1, 784, __pyx_L1_error)
     __pyx_r = ((struct __pyx_memoryview_obj *)__pyx_t_3);
     __pyx_t_3 = 0;
     goto __pyx_L0;
   }
 
-  /* "View.MemoryView":710
+  /* "View.MemoryView":712
  * 
  * @cname('__pyx_memview_slice')
  * cdef memoryview memview_slice(memoryview memview, object indices):             # <<<<<<<<<<<<<<
@@ -16504,7 +16567,7 @@ static struct __pyx_memoryview_obj *__pyx_memview_slice(struct __pyx_memoryview_
   return __pyx_r;
 }
 
-/* "View.MemoryView":807
+/* "View.MemoryView":809
  * 
  * @cname('__pyx_memoryview_slice_memviewslice')
  * cdef int slice_memviewslice(             # <<<<<<<<<<<<<<
@@ -16523,7 +16586,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
   const char *__pyx_filename = NULL;
   int __pyx_clineno = 0;
 
-  /* "View.MemoryView":827
+  /* "View.MemoryView":829
  *     cdef bint negative_step
  * 
  *     if not is_slice:             # <<<<<<<<<<<<<<
@@ -16533,7 +16596,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
   __pyx_t_1 = ((!(__pyx_v_is_slice != 0)) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":829
+    /* "View.MemoryView":831
  *     if not is_slice:
  * 
  *         if start < 0:             # <<<<<<<<<<<<<<
@@ -16543,7 +16606,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
     __pyx_t_1 = ((__pyx_v_start < 0) != 0);
     if (__pyx_t_1) {
 
-      /* "View.MemoryView":830
+      /* "View.MemoryView":832
  * 
  *         if start < 0:
  *             start += shape             # <<<<<<<<<<<<<<
@@ -16552,7 +16615,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
       __pyx_v_start = (__pyx_v_start + __pyx_v_shape);
 
-      /* "View.MemoryView":829
+      /* "View.MemoryView":831
  *     if not is_slice:
  * 
  *         if start < 0:             # <<<<<<<<<<<<<<
@@ -16561,7 +16624,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
     }
 
-    /* "View.MemoryView":831
+    /* "View.MemoryView":833
  *         if start < 0:
  *             start += shape
  *         if not 0 <= start < shape:             # <<<<<<<<<<<<<<
@@ -16575,16 +16638,16 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
     __pyx_t_2 = ((!(__pyx_t_1 != 0)) != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":832
+      /* "View.MemoryView":834
  *             start += shape
  *         if not 0 <= start < shape:
  *             _err_dim(IndexError, "Index out of bounds (axis %d)", dim)             # <<<<<<<<<<<<<<
  *     else:
  * 
  */
-      __pyx_t_3 = __pyx_memoryview_err_dim(__pyx_builtin_IndexError, ((char *)"Index out of bounds (axis %d)"), __pyx_v_dim); if (unlikely(__pyx_t_3 == ((int)-1))) __PYX_ERR(1, 832, __pyx_L1_error)
+      __pyx_t_3 = __pyx_memoryview_err_dim(__pyx_builtin_IndexError, ((char *)"Index out of bounds (axis %d)"), __pyx_v_dim); if (unlikely(__pyx_t_3 == ((int)-1))) __PYX_ERR(1, 834, __pyx_L1_error)
 
-      /* "View.MemoryView":831
+      /* "View.MemoryView":833
  *         if start < 0:
  *             start += shape
  *         if not 0 <= start < shape:             # <<<<<<<<<<<<<<
@@ -16593,7 +16656,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
     }
 
-    /* "View.MemoryView":827
+    /* "View.MemoryView":829
  *     cdef bint negative_step
  * 
  *     if not is_slice:             # <<<<<<<<<<<<<<
@@ -16603,7 +16666,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
     goto __pyx_L3;
   }
 
-  /* "View.MemoryView":835
+  /* "View.MemoryView":837
  *     else:
  * 
  *         negative_step = have_step != 0 and step < 0             # <<<<<<<<<<<<<<
@@ -16622,7 +16685,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
     __pyx_L6_bool_binop_done:;
     __pyx_v_negative_step = __pyx_t_2;
 
-    /* "View.MemoryView":837
+    /* "View.MemoryView":839
  *         negative_step = have_step != 0 and step < 0
  * 
  *         if have_step and step == 0:             # <<<<<<<<<<<<<<
@@ -16640,16 +16703,16 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
     __pyx_L9_bool_binop_done:;
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":838
+      /* "View.MemoryView":840
  * 
  *         if have_step and step == 0:
  *             _err_dim(ValueError, "Step may not be zero (axis %d)", dim)             # <<<<<<<<<<<<<<
  * 
  * 
  */
-      __pyx_t_3 = __pyx_memoryview_err_dim(__pyx_builtin_ValueError, ((char *)"Step may not be zero (axis %d)"), __pyx_v_dim); if (unlikely(__pyx_t_3 == ((int)-1))) __PYX_ERR(1, 838, __pyx_L1_error)
+      __pyx_t_3 = __pyx_memoryview_err_dim(__pyx_builtin_ValueError, ((char *)"Step may not be zero (axis %d)"), __pyx_v_dim); if (unlikely(__pyx_t_3 == ((int)-1))) __PYX_ERR(1, 840, __pyx_L1_error)
 
-      /* "View.MemoryView":837
+      /* "View.MemoryView":839
  *         negative_step = have_step != 0 and step < 0
  * 
  *         if have_step and step == 0:             # <<<<<<<<<<<<<<
@@ -16658,7 +16721,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
     }
 
-    /* "View.MemoryView":841
+    /* "View.MemoryView":843
  * 
  * 
  *         if have_start:             # <<<<<<<<<<<<<<
@@ -16668,7 +16731,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
     __pyx_t_2 = (__pyx_v_have_start != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":842
+      /* "View.MemoryView":844
  * 
  *         if have_start:
  *             if start < 0:             # <<<<<<<<<<<<<<
@@ -16678,7 +16741,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
       __pyx_t_2 = ((__pyx_v_start < 0) != 0);
       if (__pyx_t_2) {
 
-        /* "View.MemoryView":843
+        /* "View.MemoryView":845
  *         if have_start:
  *             if start < 0:
  *                 start += shape             # <<<<<<<<<<<<<<
@@ -16687,7 +16750,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
         __pyx_v_start = (__pyx_v_start + __pyx_v_shape);
 
-        /* "View.MemoryView":844
+        /* "View.MemoryView":846
  *             if start < 0:
  *                 start += shape
  *                 if start < 0:             # <<<<<<<<<<<<<<
@@ -16697,7 +16760,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
         __pyx_t_2 = ((__pyx_v_start < 0) != 0);
         if (__pyx_t_2) {
 
-          /* "View.MemoryView":845
+          /* "View.MemoryView":847
  *                 start += shape
  *                 if start < 0:
  *                     start = 0             # <<<<<<<<<<<<<<
@@ -16706,7 +16769,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
           __pyx_v_start = 0;
 
-          /* "View.MemoryView":844
+          /* "View.MemoryView":846
  *             if start < 0:
  *                 start += shape
  *                 if start < 0:             # <<<<<<<<<<<<<<
@@ -16715,7 +16778,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
         }
 
-        /* "View.MemoryView":842
+        /* "View.MemoryView":844
  * 
  *         if have_start:
  *             if start < 0:             # <<<<<<<<<<<<<<
@@ -16725,7 +16788,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
         goto __pyx_L12;
       }
 
-      /* "View.MemoryView":846
+      /* "View.MemoryView":848
  *                 if start < 0:
  *                     start = 0
  *             elif start >= shape:             # <<<<<<<<<<<<<<
@@ -16735,7 +16798,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
       __pyx_t_2 = ((__pyx_v_start >= __pyx_v_shape) != 0);
       if (__pyx_t_2) {
 
-        /* "View.MemoryView":847
+        /* "View.MemoryView":849
  *                     start = 0
  *             elif start >= shape:
  *                 if negative_step:             # <<<<<<<<<<<<<<
@@ -16745,7 +16808,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
         __pyx_t_2 = (__pyx_v_negative_step != 0);
         if (__pyx_t_2) {
 
-          /* "View.MemoryView":848
+          /* "View.MemoryView":850
  *             elif start >= shape:
  *                 if negative_step:
  *                     start = shape - 1             # <<<<<<<<<<<<<<
@@ -16754,7 +16817,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
           __pyx_v_start = (__pyx_v_shape - 1);
 
-          /* "View.MemoryView":847
+          /* "View.MemoryView":849
  *                     start = 0
  *             elif start >= shape:
  *                 if negative_step:             # <<<<<<<<<<<<<<
@@ -16764,7 +16827,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
           goto __pyx_L14;
         }
 
-        /* "View.MemoryView":850
+        /* "View.MemoryView":852
  *                     start = shape - 1
  *                 else:
  *                     start = shape             # <<<<<<<<<<<<<<
@@ -16776,7 +16839,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
         }
         __pyx_L14:;
 
-        /* "View.MemoryView":846
+        /* "View.MemoryView":848
  *                 if start < 0:
  *                     start = 0
  *             elif start >= shape:             # <<<<<<<<<<<<<<
@@ -16786,7 +16849,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
       }
       __pyx_L12:;
 
-      /* "View.MemoryView":841
+      /* "View.MemoryView":843
  * 
  * 
  *         if have_start:             # <<<<<<<<<<<<<<
@@ -16796,7 +16859,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
       goto __pyx_L11;
     }
 
-    /* "View.MemoryView":852
+    /* "View.MemoryView":854
  *                     start = shape
  *         else:
  *             if negative_step:             # <<<<<<<<<<<<<<
@@ -16807,7 +16870,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
       __pyx_t_2 = (__pyx_v_negative_step != 0);
       if (__pyx_t_2) {
 
-        /* "View.MemoryView":853
+        /* "View.MemoryView":855
  *         else:
  *             if negative_step:
  *                 start = shape - 1             # <<<<<<<<<<<<<<
@@ -16816,7 +16879,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
         __pyx_v_start = (__pyx_v_shape - 1);
 
-        /* "View.MemoryView":852
+        /* "View.MemoryView":854
  *                     start = shape
  *         else:
  *             if negative_step:             # <<<<<<<<<<<<<<
@@ -16826,7 +16889,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
         goto __pyx_L15;
       }
 
-      /* "View.MemoryView":855
+      /* "View.MemoryView":857
  *                 start = shape - 1
  *             else:
  *                 start = 0             # <<<<<<<<<<<<<<
@@ -16840,7 +16903,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
     }
     __pyx_L11:;
 
-    /* "View.MemoryView":857
+    /* "View.MemoryView":859
  *                 start = 0
  * 
  *         if have_stop:             # <<<<<<<<<<<<<<
@@ -16850,7 +16913,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
     __pyx_t_2 = (__pyx_v_have_stop != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":858
+      /* "View.MemoryView":860
  * 
  *         if have_stop:
  *             if stop < 0:             # <<<<<<<<<<<<<<
@@ -16860,7 +16923,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
       __pyx_t_2 = ((__pyx_v_stop < 0) != 0);
       if (__pyx_t_2) {
 
-        /* "View.MemoryView":859
+        /* "View.MemoryView":861
  *         if have_stop:
  *             if stop < 0:
  *                 stop += shape             # <<<<<<<<<<<<<<
@@ -16869,7 +16932,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
         __pyx_v_stop = (__pyx_v_stop + __pyx_v_shape);
 
-        /* "View.MemoryView":860
+        /* "View.MemoryView":862
  *             if stop < 0:
  *                 stop += shape
  *                 if stop < 0:             # <<<<<<<<<<<<<<
@@ -16879,7 +16942,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
         __pyx_t_2 = ((__pyx_v_stop < 0) != 0);
         if (__pyx_t_2) {
 
-          /* "View.MemoryView":861
+          /* "View.MemoryView":863
  *                 stop += shape
  *                 if stop < 0:
  *                     stop = 0             # <<<<<<<<<<<<<<
@@ -16888,7 +16951,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
           __pyx_v_stop = 0;
 
-          /* "View.MemoryView":860
+          /* "View.MemoryView":862
  *             if stop < 0:
  *                 stop += shape
  *                 if stop < 0:             # <<<<<<<<<<<<<<
@@ -16897,7 +16960,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
         }
 
-        /* "View.MemoryView":858
+        /* "View.MemoryView":860
  * 
  *         if have_stop:
  *             if stop < 0:             # <<<<<<<<<<<<<<
@@ -16907,7 +16970,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
         goto __pyx_L17;
       }
 
-      /* "View.MemoryView":862
+      /* "View.MemoryView":864
  *                 if stop < 0:
  *                     stop = 0
  *             elif stop > shape:             # <<<<<<<<<<<<<<
@@ -16917,7 +16980,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
       __pyx_t_2 = ((__pyx_v_stop > __pyx_v_shape) != 0);
       if (__pyx_t_2) {
 
-        /* "View.MemoryView":863
+        /* "View.MemoryView":865
  *                     stop = 0
  *             elif stop > shape:
  *                 stop = shape             # <<<<<<<<<<<<<<
@@ -16926,7 +16989,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
         __pyx_v_stop = __pyx_v_shape;
 
-        /* "View.MemoryView":862
+        /* "View.MemoryView":864
  *                 if stop < 0:
  *                     stop = 0
  *             elif stop > shape:             # <<<<<<<<<<<<<<
@@ -16936,7 +16999,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
       }
       __pyx_L17:;
 
-      /* "View.MemoryView":857
+      /* "View.MemoryView":859
  *                 start = 0
  * 
  *         if have_stop:             # <<<<<<<<<<<<<<
@@ -16946,7 +17009,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
       goto __pyx_L16;
     }
 
-    /* "View.MemoryView":865
+    /* "View.MemoryView":867
  *                 stop = shape
  *         else:
  *             if negative_step:             # <<<<<<<<<<<<<<
@@ -16957,7 +17020,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
       __pyx_t_2 = (__pyx_v_negative_step != 0);
       if (__pyx_t_2) {
 
-        /* "View.MemoryView":866
+        /* "View.MemoryView":868
  *         else:
  *             if negative_step:
  *                 stop = -1             # <<<<<<<<<<<<<<
@@ -16966,7 +17029,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
         __pyx_v_stop = -1L;
 
-        /* "View.MemoryView":865
+        /* "View.MemoryView":867
  *                 stop = shape
  *         else:
  *             if negative_step:             # <<<<<<<<<<<<<<
@@ -16976,7 +17039,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
         goto __pyx_L19;
       }
 
-      /* "View.MemoryView":868
+      /* "View.MemoryView":870
  *                 stop = -1
  *             else:
  *                 stop = shape             # <<<<<<<<<<<<<<
@@ -16990,7 +17053,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
     }
     __pyx_L16:;
 
-    /* "View.MemoryView":870
+    /* "View.MemoryView":872
  *                 stop = shape
  * 
  *         if not have_step:             # <<<<<<<<<<<<<<
@@ -17000,7 +17063,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
     __pyx_t_2 = ((!(__pyx_v_have_step != 0)) != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":871
+      /* "View.MemoryView":873
  * 
  *         if not have_step:
  *             step = 1             # <<<<<<<<<<<<<<
@@ -17009,7 +17072,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
       __pyx_v_step = 1;
 
-      /* "View.MemoryView":870
+      /* "View.MemoryView":872
  *                 stop = shape
  * 
  *         if not have_step:             # <<<<<<<<<<<<<<
@@ -17018,7 +17081,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
     }
 
-    /* "View.MemoryView":875
+    /* "View.MemoryView":877
  * 
  *         with cython.cdivision(True):
  *             new_shape = (stop - start) // step             # <<<<<<<<<<<<<<
@@ -17027,7 +17090,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
     __pyx_v_new_shape = ((__pyx_v_stop - __pyx_v_start) / __pyx_v_step);
 
-    /* "View.MemoryView":877
+    /* "View.MemoryView":879
  *             new_shape = (stop - start) // step
  * 
  *             if (stop - start) - step * new_shape:             # <<<<<<<<<<<<<<
@@ -17037,7 +17100,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
     __pyx_t_2 = (((__pyx_v_stop - __pyx_v_start) - (__pyx_v_step * __pyx_v_new_shape)) != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":878
+      /* "View.MemoryView":880
  * 
  *             if (stop - start) - step * new_shape:
  *                 new_shape += 1             # <<<<<<<<<<<<<<
@@ -17046,7 +17109,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
       __pyx_v_new_shape = (__pyx_v_new_shape + 1);
 
-      /* "View.MemoryView":877
+      /* "View.MemoryView":879
  *             new_shape = (stop - start) // step
  * 
  *             if (stop - start) - step * new_shape:             # <<<<<<<<<<<<<<
@@ -17055,7 +17118,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
     }
 
-    /* "View.MemoryView":880
+    /* "View.MemoryView":882
  *                 new_shape += 1
  * 
  *         if new_shape < 0:             # <<<<<<<<<<<<<<
@@ -17065,7 +17128,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
     __pyx_t_2 = ((__pyx_v_new_shape < 0) != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":881
+      /* "View.MemoryView":883
  * 
  *         if new_shape < 0:
  *             new_shape = 0             # <<<<<<<<<<<<<<
@@ -17074,7 +17137,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
       __pyx_v_new_shape = 0;
 
-      /* "View.MemoryView":880
+      /* "View.MemoryView":882
  *                 new_shape += 1
  * 
  *         if new_shape < 0:             # <<<<<<<<<<<<<<
@@ -17083,7 +17146,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
     }
 
-    /* "View.MemoryView":884
+    /* "View.MemoryView":886
  * 
  * 
  *         dst.strides[new_ndim] = stride * step             # <<<<<<<<<<<<<<
@@ -17092,7 +17155,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
     (__pyx_v_dst->strides[__pyx_v_new_ndim]) = (__pyx_v_stride * __pyx_v_step);
 
-    /* "View.MemoryView":885
+    /* "View.MemoryView":887
  * 
  *         dst.strides[new_ndim] = stride * step
  *         dst.shape[new_ndim] = new_shape             # <<<<<<<<<<<<<<
@@ -17101,7 +17164,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
     (__pyx_v_dst->shape[__pyx_v_new_ndim]) = __pyx_v_new_shape;
 
-    /* "View.MemoryView":886
+    /* "View.MemoryView":888
  *         dst.strides[new_ndim] = stride * step
  *         dst.shape[new_ndim] = new_shape
  *         dst.suboffsets[new_ndim] = suboffset             # <<<<<<<<<<<<<<
@@ -17112,7 +17175,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
   }
   __pyx_L3:;
 
-  /* "View.MemoryView":889
+  /* "View.MemoryView":891
  * 
  * 
  *     if suboffset_dim[0] < 0:             # <<<<<<<<<<<<<<
@@ -17122,7 +17185,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
   __pyx_t_2 = (((__pyx_v_suboffset_dim[0]) < 0) != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":890
+    /* "View.MemoryView":892
  * 
  *     if suboffset_dim[0] < 0:
  *         dst.data += start * stride             # <<<<<<<<<<<<<<
@@ -17131,7 +17194,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
     __pyx_v_dst->data = (__pyx_v_dst->data + (__pyx_v_start * __pyx_v_stride));
 
-    /* "View.MemoryView":889
+    /* "View.MemoryView":891
  * 
  * 
  *     if suboffset_dim[0] < 0:             # <<<<<<<<<<<<<<
@@ -17141,7 +17204,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
     goto __pyx_L23;
   }
 
-  /* "View.MemoryView":892
+  /* "View.MemoryView":894
  *         dst.data += start * stride
  *     else:
  *         dst.suboffsets[suboffset_dim[0]] += start * stride             # <<<<<<<<<<<<<<
@@ -17154,7 +17217,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
   }
   __pyx_L23:;
 
-  /* "View.MemoryView":894
+  /* "View.MemoryView":896
  *         dst.suboffsets[suboffset_dim[0]] += start * stride
  * 
  *     if suboffset >= 0:             # <<<<<<<<<<<<<<
@@ -17164,7 +17227,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
   __pyx_t_2 = ((__pyx_v_suboffset >= 0) != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":895
+    /* "View.MemoryView":897
  * 
  *     if suboffset >= 0:
  *         if not is_slice:             # <<<<<<<<<<<<<<
@@ -17174,7 +17237,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
     __pyx_t_2 = ((!(__pyx_v_is_slice != 0)) != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":896
+      /* "View.MemoryView":898
  *     if suboffset >= 0:
  *         if not is_slice:
  *             if new_ndim == 0:             # <<<<<<<<<<<<<<
@@ -17184,7 +17247,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
       __pyx_t_2 = ((__pyx_v_new_ndim == 0) != 0);
       if (__pyx_t_2) {
 
-        /* "View.MemoryView":897
+        /* "View.MemoryView":899
  *         if not is_slice:
  *             if new_ndim == 0:
  *                 dst.data = (<char **> dst.data)[0] + suboffset             # <<<<<<<<<<<<<<
@@ -17193,7 +17256,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
         __pyx_v_dst->data = ((((char **)__pyx_v_dst->data)[0]) + __pyx_v_suboffset);
 
-        /* "View.MemoryView":896
+        /* "View.MemoryView":898
  *     if suboffset >= 0:
  *         if not is_slice:
  *             if new_ndim == 0:             # <<<<<<<<<<<<<<
@@ -17203,7 +17266,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
         goto __pyx_L26;
       }
 
-      /* "View.MemoryView":899
+      /* "View.MemoryView":901
  *                 dst.data = (<char **> dst.data)[0] + suboffset
  *             else:
  *                 _err_dim(IndexError, "All dimensions preceding dimension %d "             # <<<<<<<<<<<<<<
@@ -17212,18 +17275,18 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
       /*else*/ {
 
-        /* "View.MemoryView":900
+        /* "View.MemoryView":902
  *             else:
  *                 _err_dim(IndexError, "All dimensions preceding dimension %d "
  *                                      "must be indexed and not sliced", dim)             # <<<<<<<<<<<<<<
  *         else:
  *             suboffset_dim[0] = new_ndim
  */
-        __pyx_t_3 = __pyx_memoryview_err_dim(__pyx_builtin_IndexError, ((char *)"All dimensions preceding dimension %d must be indexed and not sliced"), __pyx_v_dim); if (unlikely(__pyx_t_3 == ((int)-1))) __PYX_ERR(1, 899, __pyx_L1_error)
+        __pyx_t_3 = __pyx_memoryview_err_dim(__pyx_builtin_IndexError, ((char *)"All dimensions preceding dimension %d must be indexed and not sliced"), __pyx_v_dim); if (unlikely(__pyx_t_3 == ((int)-1))) __PYX_ERR(1, 901, __pyx_L1_error)
       }
       __pyx_L26:;
 
-      /* "View.MemoryView":895
+      /* "View.MemoryView":897
  * 
  *     if suboffset >= 0:
  *         if not is_slice:             # <<<<<<<<<<<<<<
@@ -17233,7 +17296,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
       goto __pyx_L25;
     }
 
-    /* "View.MemoryView":902
+    /* "View.MemoryView":904
  *                                      "must be indexed and not sliced", dim)
  *         else:
  *             suboffset_dim[0] = new_ndim             # <<<<<<<<<<<<<<
@@ -17245,7 +17308,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
     }
     __pyx_L25:;
 
-    /* "View.MemoryView":894
+    /* "View.MemoryView":896
  *         dst.suboffsets[suboffset_dim[0]] += start * stride
  * 
  *     if suboffset >= 0:             # <<<<<<<<<<<<<<
@@ -17254,7 +17317,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
  */
   }
 
-  /* "View.MemoryView":904
+  /* "View.MemoryView":906
  *             suboffset_dim[0] = new_ndim
  * 
  *     return 0             # <<<<<<<<<<<<<<
@@ -17264,7 +17327,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
   __pyx_r = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":807
+  /* "View.MemoryView":809
  * 
  * @cname('__pyx_memoryview_slice_memviewslice')
  * cdef int slice_memviewslice(             # <<<<<<<<<<<<<<
@@ -17288,7 +17351,7 @@ static int __pyx_memoryview_slice_memviewslice(__Pyx_memviewslice *__pyx_v_dst,
   return __pyx_r;
 }
 
-/* "View.MemoryView":910
+/* "View.MemoryView":912
  * 
  * @cname('__pyx_pybuffer_index')
  * cdef char *pybuffer_index(Py_buffer *view, char *bufp, Py_ssize_t index,             # <<<<<<<<<<<<<<
@@ -17313,7 +17376,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("pybuffer_index", 0);
 
-  /* "View.MemoryView":912
+  /* "View.MemoryView":914
  * cdef char *pybuffer_index(Py_buffer *view, char *bufp, Py_ssize_t index,
  *                           Py_ssize_t dim) except NULL:
  *     cdef Py_ssize_t shape, stride, suboffset = -1             # <<<<<<<<<<<<<<
@@ -17322,7 +17385,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
  */
   __pyx_v_suboffset = -1L;
 
-  /* "View.MemoryView":913
+  /* "View.MemoryView":915
  *                           Py_ssize_t dim) except NULL:
  *     cdef Py_ssize_t shape, stride, suboffset = -1
  *     cdef Py_ssize_t itemsize = view.itemsize             # <<<<<<<<<<<<<<
@@ -17332,7 +17395,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
   __pyx_t_1 = __pyx_v_view->itemsize;
   __pyx_v_itemsize = __pyx_t_1;
 
-  /* "View.MemoryView":916
+  /* "View.MemoryView":918
  *     cdef char *resultp
  * 
  *     if view.ndim == 0:             # <<<<<<<<<<<<<<
@@ -17342,7 +17405,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
   __pyx_t_2 = ((__pyx_v_view->ndim == 0) != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":917
+    /* "View.MemoryView":919
  * 
  *     if view.ndim == 0:
  *         shape = view.len / itemsize             # <<<<<<<<<<<<<<
@@ -17351,15 +17414,15 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
  */
     if (unlikely(__pyx_v_itemsize == 0)) {
       PyErr_SetString(PyExc_ZeroDivisionError, "integer division or modulo by zero");
-      __PYX_ERR(1, 917, __pyx_L1_error)
+      __PYX_ERR(1, 919, __pyx_L1_error)
     }
     else if (sizeof(Py_ssize_t) == sizeof(long) && (!(((Py_ssize_t)-1) > 0)) && unlikely(__pyx_v_itemsize == (Py_ssize_t)-1)  && unlikely(UNARY_NEG_WOULD_OVERFLOW(__pyx_v_view->len))) {
       PyErr_SetString(PyExc_OverflowError, "value too large to perform division");
-      __PYX_ERR(1, 917, __pyx_L1_error)
+      __PYX_ERR(1, 919, __pyx_L1_error)
     }
     __pyx_v_shape = __Pyx_div_Py_ssize_t(__pyx_v_view->len, __pyx_v_itemsize);
 
-    /* "View.MemoryView":918
+    /* "View.MemoryView":920
  *     if view.ndim == 0:
  *         shape = view.len / itemsize
  *         stride = itemsize             # <<<<<<<<<<<<<<
@@ -17368,7 +17431,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
  */
     __pyx_v_stride = __pyx_v_itemsize;
 
-    /* "View.MemoryView":916
+    /* "View.MemoryView":918
  *     cdef char *resultp
  * 
  *     if view.ndim == 0:             # <<<<<<<<<<<<<<
@@ -17378,7 +17441,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
     goto __pyx_L3;
   }
 
-  /* "View.MemoryView":920
+  /* "View.MemoryView":922
  *         stride = itemsize
  *     else:
  *         shape = view.shape[dim]             # <<<<<<<<<<<<<<
@@ -17388,7 +17451,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
   /*else*/ {
     __pyx_v_shape = (__pyx_v_view->shape[__pyx_v_dim]);
 
-    /* "View.MemoryView":921
+    /* "View.MemoryView":923
  *     else:
  *         shape = view.shape[dim]
  *         stride = view.strides[dim]             # <<<<<<<<<<<<<<
@@ -17397,7 +17460,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
  */
     __pyx_v_stride = (__pyx_v_view->strides[__pyx_v_dim]);
 
-    /* "View.MemoryView":922
+    /* "View.MemoryView":924
  *         shape = view.shape[dim]
  *         stride = view.strides[dim]
  *         if view.suboffsets != NULL:             # <<<<<<<<<<<<<<
@@ -17407,7 +17470,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
     __pyx_t_2 = ((__pyx_v_view->suboffsets != NULL) != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":923
+      /* "View.MemoryView":925
  *         stride = view.strides[dim]
  *         if view.suboffsets != NULL:
  *             suboffset = view.suboffsets[dim]             # <<<<<<<<<<<<<<
@@ -17416,7 +17479,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
  */
       __pyx_v_suboffset = (__pyx_v_view->suboffsets[__pyx_v_dim]);
 
-      /* "View.MemoryView":922
+      /* "View.MemoryView":924
  *         shape = view.shape[dim]
  *         stride = view.strides[dim]
  *         if view.suboffsets != NULL:             # <<<<<<<<<<<<<<
@@ -17427,7 +17490,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
   }
   __pyx_L3:;
 
-  /* "View.MemoryView":925
+  /* "View.MemoryView":927
  *             suboffset = view.suboffsets[dim]
  * 
  *     if index < 0:             # <<<<<<<<<<<<<<
@@ -17437,7 +17500,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
   __pyx_t_2 = ((__pyx_v_index < 0) != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":926
+    /* "View.MemoryView":928
  * 
  *     if index < 0:
  *         index += view.shape[dim]             # <<<<<<<<<<<<<<
@@ -17446,7 +17509,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
  */
     __pyx_v_index = (__pyx_v_index + (__pyx_v_view->shape[__pyx_v_dim]));
 
-    /* "View.MemoryView":927
+    /* "View.MemoryView":929
  *     if index < 0:
  *         index += view.shape[dim]
  *         if index < 0:             # <<<<<<<<<<<<<<
@@ -17456,26 +17519,26 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
     __pyx_t_2 = ((__pyx_v_index < 0) != 0);
     if (unlikely(__pyx_t_2)) {
 
-      /* "View.MemoryView":928
+      /* "View.MemoryView":930
  *         index += view.shape[dim]
  *         if index < 0:
  *             raise IndexError("Out of bounds on buffer access (axis %d)" % dim)             # <<<<<<<<<<<<<<
  * 
  *     if index >= shape:
  */
-      __pyx_t_3 = PyInt_FromSsize_t(__pyx_v_dim); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 928, __pyx_L1_error)
+      __pyx_t_3 = PyInt_FromSsize_t(__pyx_v_dim); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 930, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_3);
-      __pyx_t_4 = __Pyx_PyString_Format(__pyx_kp_s_Out_of_bounds_on_buffer_access_a, __pyx_t_3); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 928, __pyx_L1_error)
+      __pyx_t_4 = __Pyx_PyString_Format(__pyx_kp_s_Out_of_bounds_on_buffer_access_a, __pyx_t_3); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 930, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_4);
       __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
-      __pyx_t_3 = __Pyx_PyObject_CallOneArg(__pyx_builtin_IndexError, __pyx_t_4); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 928, __pyx_L1_error)
+      __pyx_t_3 = __Pyx_PyObject_CallOneArg(__pyx_builtin_IndexError, __pyx_t_4); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 930, __pyx_L1_error)
       __Pyx_GOTREF(__pyx_t_3);
       __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
       __Pyx_Raise(__pyx_t_3, 0, 0, 0);
       __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
-      __PYX_ERR(1, 928, __pyx_L1_error)
+      __PYX_ERR(1, 930, __pyx_L1_error)
 
-      /* "View.MemoryView":927
+      /* "View.MemoryView":929
  *     if index < 0:
  *         index += view.shape[dim]
  *         if index < 0:             # <<<<<<<<<<<<<<
@@ -17484,7 +17547,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
  */
     }
 
-    /* "View.MemoryView":925
+    /* "View.MemoryView":927
  *             suboffset = view.suboffsets[dim]
  * 
  *     if index < 0:             # <<<<<<<<<<<<<<
@@ -17493,7 +17556,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
  */
   }
 
-  /* "View.MemoryView":930
+  /* "View.MemoryView":932
  *             raise IndexError("Out of bounds on buffer access (axis %d)" % dim)
  * 
  *     if index >= shape:             # <<<<<<<<<<<<<<
@@ -17503,26 +17566,26 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
   __pyx_t_2 = ((__pyx_v_index >= __pyx_v_shape) != 0);
   if (unlikely(__pyx_t_2)) {
 
-    /* "View.MemoryView":931
+    /* "View.MemoryView":933
  * 
  *     if index >= shape:
  *         raise IndexError("Out of bounds on buffer access (axis %d)" % dim)             # <<<<<<<<<<<<<<
  * 
  *     resultp = bufp + index * stride
  */
-    __pyx_t_3 = PyInt_FromSsize_t(__pyx_v_dim); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 931, __pyx_L1_error)
+    __pyx_t_3 = PyInt_FromSsize_t(__pyx_v_dim); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 933, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
-    __pyx_t_4 = __Pyx_PyString_Format(__pyx_kp_s_Out_of_bounds_on_buffer_access_a, __pyx_t_3); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 931, __pyx_L1_error)
+    __pyx_t_4 = __Pyx_PyString_Format(__pyx_kp_s_Out_of_bounds_on_buffer_access_a, __pyx_t_3); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 933, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_4);
     __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
-    __pyx_t_3 = __Pyx_PyObject_CallOneArg(__pyx_builtin_IndexError, __pyx_t_4); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 931, __pyx_L1_error)
+    __pyx_t_3 = __Pyx_PyObject_CallOneArg(__pyx_builtin_IndexError, __pyx_t_4); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 933, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
     __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
     __Pyx_Raise(__pyx_t_3, 0, 0, 0);
     __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
-    __PYX_ERR(1, 931, __pyx_L1_error)
+    __PYX_ERR(1, 933, __pyx_L1_error)
 
-    /* "View.MemoryView":930
+    /* "View.MemoryView":932
  *             raise IndexError("Out of bounds on buffer access (axis %d)" % dim)
  * 
  *     if index >= shape:             # <<<<<<<<<<<<<<
@@ -17531,7 +17594,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
  */
   }
 
-  /* "View.MemoryView":933
+  /* "View.MemoryView":935
  *         raise IndexError("Out of bounds on buffer access (axis %d)" % dim)
  * 
  *     resultp = bufp + index * stride             # <<<<<<<<<<<<<<
@@ -17540,7 +17603,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
  */
   __pyx_v_resultp = (__pyx_v_bufp + (__pyx_v_index * __pyx_v_stride));
 
-  /* "View.MemoryView":934
+  /* "View.MemoryView":936
  * 
  *     resultp = bufp + index * stride
  *     if suboffset >= 0:             # <<<<<<<<<<<<<<
@@ -17550,7 +17613,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
   __pyx_t_2 = ((__pyx_v_suboffset >= 0) != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":935
+    /* "View.MemoryView":937
  *     resultp = bufp + index * stride
  *     if suboffset >= 0:
  *         resultp = (<char **> resultp)[0] + suboffset             # <<<<<<<<<<<<<<
@@ -17559,7 +17622,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
  */
     __pyx_v_resultp = ((((char **)__pyx_v_resultp)[0]) + __pyx_v_suboffset);
 
-    /* "View.MemoryView":934
+    /* "View.MemoryView":936
  * 
  *     resultp = bufp + index * stride
  *     if suboffset >= 0:             # <<<<<<<<<<<<<<
@@ -17568,7 +17631,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
  */
   }
 
-  /* "View.MemoryView":937
+  /* "View.MemoryView":939
  *         resultp = (<char **> resultp)[0] + suboffset
  * 
  *     return resultp             # <<<<<<<<<<<<<<
@@ -17578,7 +17641,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
   __pyx_r = __pyx_v_resultp;
   goto __pyx_L0;
 
-  /* "View.MemoryView":910
+  /* "View.MemoryView":912
  * 
  * @cname('__pyx_pybuffer_index')
  * cdef char *pybuffer_index(Py_buffer *view, char *bufp, Py_ssize_t index,             # <<<<<<<<<<<<<<
@@ -17597,7 +17660,7 @@ static char *__pyx_pybuffer_index(Py_buffer *__pyx_v_view, char *__pyx_v_bufp, P
   return __pyx_r;
 }
 
-/* "View.MemoryView":943
+/* "View.MemoryView":945
  * 
  * @cname('__pyx_memslice_transpose')
  * cdef int transpose_memslice(__Pyx_memviewslice *memslice) nogil except 0:             # <<<<<<<<<<<<<<
@@ -17625,7 +17688,7 @@ static int __pyx_memslice_transpose(__Pyx_memviewslice *__pyx_v_memslice) {
   const char *__pyx_filename = NULL;
   int __pyx_clineno = 0;
 
-  /* "View.MemoryView":944
+  /* "View.MemoryView":946
  * @cname('__pyx_memslice_transpose')
  * cdef int transpose_memslice(__Pyx_memviewslice *memslice) nogil except 0:
  *     cdef int ndim = memslice.memview.view.ndim             # <<<<<<<<<<<<<<
@@ -17635,7 +17698,7 @@ static int __pyx_memslice_transpose(__Pyx_memviewslice *__pyx_v_memslice) {
   __pyx_t_1 = __pyx_v_memslice->memview->view.ndim;
   __pyx_v_ndim = __pyx_t_1;
 
-  /* "View.MemoryView":946
+  /* "View.MemoryView":948
  *     cdef int ndim = memslice.memview.view.ndim
  * 
  *     cdef Py_ssize_t *shape = memslice.shape             # <<<<<<<<<<<<<<
@@ -17645,7 +17708,7 @@ static int __pyx_memslice_transpose(__Pyx_memviewslice *__pyx_v_memslice) {
   __pyx_t_2 = __pyx_v_memslice->shape;
   __pyx_v_shape = __pyx_t_2;
 
-  /* "View.MemoryView":947
+  /* "View.MemoryView":949
  * 
  *     cdef Py_ssize_t *shape = memslice.shape
  *     cdef Py_ssize_t *strides = memslice.strides             # <<<<<<<<<<<<<<
@@ -17655,7 +17718,7 @@ static int __pyx_memslice_transpose(__Pyx_memviewslice *__pyx_v_memslice) {
   __pyx_t_2 = __pyx_v_memslice->strides;
   __pyx_v_strides = __pyx_t_2;
 
-  /* "View.MemoryView":951
+  /* "View.MemoryView":953
  * 
  *     cdef int i, j
  *     for i in range(ndim / 2):             # <<<<<<<<<<<<<<
@@ -17667,7 +17730,7 @@ static int __pyx_memslice_transpose(__Pyx_memviewslice *__pyx_v_memslice) {
   for (__pyx_t_1 = 0; __pyx_t_1 < __pyx_t_4; __pyx_t_1+=1) {
     __pyx_v_i = __pyx_t_1;
 
-    /* "View.MemoryView":952
+    /* "View.MemoryView":954
  *     cdef int i, j
  *     for i in range(ndim / 2):
  *         j = ndim - 1 - i             # <<<<<<<<<<<<<<
@@ -17676,7 +17739,7 @@ static int __pyx_memslice_transpose(__Pyx_memviewslice *__pyx_v_memslice) {
  */
     __pyx_v_j = ((__pyx_v_ndim - 1) - __pyx_v_i);
 
-    /* "View.MemoryView":953
+    /* "View.MemoryView":955
  *     for i in range(ndim / 2):
  *         j = ndim - 1 - i
  *         strides[i], strides[j] = strides[j], strides[i]             # <<<<<<<<<<<<<<
@@ -17688,7 +17751,7 @@ static int __pyx_memslice_transpose(__Pyx_memviewslice *__pyx_v_memslice) {
     (__pyx_v_strides[__pyx_v_i]) = __pyx_t_5;
     (__pyx_v_strides[__pyx_v_j]) = __pyx_t_6;
 
-    /* "View.MemoryView":954
+    /* "View.MemoryView":956
  *         j = ndim - 1 - i
  *         strides[i], strides[j] = strides[j], strides[i]
  *         shape[i], shape[j] = shape[j], shape[i]             # <<<<<<<<<<<<<<
@@ -17700,7 +17763,7 @@ static int __pyx_memslice_transpose(__Pyx_memviewslice *__pyx_v_memslice) {
     (__pyx_v_shape[__pyx_v_i]) = __pyx_t_6;
     (__pyx_v_shape[__pyx_v_j]) = __pyx_t_5;
 
-    /* "View.MemoryView":956
+    /* "View.MemoryView":958
  *         shape[i], shape[j] = shape[j], shape[i]
  * 
  *         if memslice.suboffsets[i] >= 0 or memslice.suboffsets[j] >= 0:             # <<<<<<<<<<<<<<
@@ -17718,16 +17781,16 @@ static int __pyx_memslice_transpose(__Pyx_memviewslice *__pyx_v_memslice) {
     __pyx_L6_bool_binop_done:;
     if (__pyx_t_7) {
 
-      /* "View.MemoryView":957
+      /* "View.MemoryView":959
  * 
  *         if memslice.suboffsets[i] >= 0 or memslice.suboffsets[j] >= 0:
  *             _err(ValueError, "Cannot transpose memoryview with indirect dimensions")             # <<<<<<<<<<<<<<
  * 
  *     return 1
  */
-      __pyx_t_9 = __pyx_memoryview_err(__pyx_builtin_ValueError, ((char *)"Cannot transpose memoryview with indirect dimensions")); if (unlikely(__pyx_t_9 == ((int)-1))) __PYX_ERR(1, 957, __pyx_L1_error)
+      __pyx_t_9 = __pyx_memoryview_err(__pyx_builtin_ValueError, ((char *)"Cannot transpose memoryview with indirect dimensions")); if (unlikely(__pyx_t_9 == ((int)-1))) __PYX_ERR(1, 959, __pyx_L1_error)
 
-      /* "View.MemoryView":956
+      /* "View.MemoryView":958
  *         shape[i], shape[j] = shape[j], shape[i]
  * 
  *         if memslice.suboffsets[i] >= 0 or memslice.suboffsets[j] >= 0:             # <<<<<<<<<<<<<<
@@ -17737,7 +17800,7 @@ static int __pyx_memslice_transpose(__Pyx_memviewslice *__pyx_v_memslice) {
     }
   }
 
-  /* "View.MemoryView":959
+  /* "View.MemoryView":961
  *             _err(ValueError, "Cannot transpose memoryview with indirect dimensions")
  * 
  *     return 1             # <<<<<<<<<<<<<<
@@ -17747,7 +17810,7 @@ static int __pyx_memslice_transpose(__Pyx_memviewslice *__pyx_v_memslice) {
   __pyx_r = 1;
   goto __pyx_L0;
 
-  /* "View.MemoryView":943
+  /* "View.MemoryView":945
  * 
  * @cname('__pyx_memslice_transpose')
  * cdef int transpose_memslice(__Pyx_memviewslice *memslice) nogil except 0:             # <<<<<<<<<<<<<<
@@ -17771,7 +17834,7 @@ static int __pyx_memslice_transpose(__Pyx_memviewslice *__pyx_v_memslice) {
   return __pyx_r;
 }
 
-/* "View.MemoryView":976
+/* "View.MemoryView":978
  *     cdef int (*to_dtype_func)(char *, object) except 0
  * 
  *     def __dealloc__(self):             # <<<<<<<<<<<<<<
@@ -17794,7 +17857,7 @@ static void __pyx_memoryviewslice___pyx_pf_15View_dot_MemoryView_16_memoryviewsl
   __Pyx_RefNannyDeclarations
   __Pyx_RefNannySetupContext("__dealloc__", 0);
 
-  /* "View.MemoryView":977
+  /* "View.MemoryView":979
  * 
  *     def __dealloc__(self):
  *         __PYX_XDEC_MEMVIEW(&self.from_slice, 1)             # <<<<<<<<<<<<<<
@@ -17803,7 +17866,7 @@ static void __pyx_memoryviewslice___pyx_pf_15View_dot_MemoryView_16_memoryviewsl
  */
   __PYX_XDEC_MEMVIEW((&__pyx_v_self->from_slice), 1);
 
-  /* "View.MemoryView":976
+  /* "View.MemoryView":978
  *     cdef int (*to_dtype_func)(char *, object) except 0
  * 
  *     def __dealloc__(self):             # <<<<<<<<<<<<<<
@@ -17815,7 +17878,7 @@ static void __pyx_memoryviewslice___pyx_pf_15View_dot_MemoryView_16_memoryviewsl
   __Pyx_RefNannyFinishContext();
 }
 
-/* "View.MemoryView":979
+/* "View.MemoryView":981
  *         __PYX_XDEC_MEMVIEW(&self.from_slice, 1)
  * 
  *     cdef convert_item_to_object(self, char *itemp):             # <<<<<<<<<<<<<<
@@ -17833,7 +17896,7 @@ static PyObject *__pyx_memoryviewslice_convert_item_to_object(struct __pyx_memor
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("convert_item_to_object", 0);
 
-  /* "View.MemoryView":980
+  /* "View.MemoryView":982
  * 
  *     cdef convert_item_to_object(self, char *itemp):
  *         if self.to_object_func != NULL:             # <<<<<<<<<<<<<<
@@ -17843,7 +17906,7 @@ static PyObject *__pyx_memoryviewslice_convert_item_to_object(struct __pyx_memor
   __pyx_t_1 = ((__pyx_v_self->to_object_func != NULL) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":981
+    /* "View.MemoryView":983
  *     cdef convert_item_to_object(self, char *itemp):
  *         if self.to_object_func != NULL:
  *             return self.to_object_func(itemp)             # <<<<<<<<<<<<<<
@@ -17851,13 +17914,13 @@ static PyObject *__pyx_memoryviewslice_convert_item_to_object(struct __pyx_memor
  *             return memoryview.convert_item_to_object(self, itemp)
  */
     __Pyx_XDECREF(__pyx_r);
-    __pyx_t_2 = __pyx_v_self->to_object_func(__pyx_v_itemp); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 981, __pyx_L1_error)
+    __pyx_t_2 = __pyx_v_self->to_object_func(__pyx_v_itemp); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 983, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_2);
     __pyx_r = __pyx_t_2;
     __pyx_t_2 = 0;
     goto __pyx_L0;
 
-    /* "View.MemoryView":980
+    /* "View.MemoryView":982
  * 
  *     cdef convert_item_to_object(self, char *itemp):
  *         if self.to_object_func != NULL:             # <<<<<<<<<<<<<<
@@ -17866,7 +17929,7 @@ static PyObject *__pyx_memoryviewslice_convert_item_to_object(struct __pyx_memor
  */
   }
 
-  /* "View.MemoryView":983
+  /* "View.MemoryView":985
  *             return self.to_object_func(itemp)
  *         else:
  *             return memoryview.convert_item_to_object(self, itemp)             # <<<<<<<<<<<<<<
@@ -17875,14 +17938,14 @@ static PyObject *__pyx_memoryviewslice_convert_item_to_object(struct __pyx_memor
  */
   /*else*/ {
     __Pyx_XDECREF(__pyx_r);
-    __pyx_t_2 = __pyx_memoryview_convert_item_to_object(((struct __pyx_memoryview_obj *)__pyx_v_self), __pyx_v_itemp); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 983, __pyx_L1_error)
+    __pyx_t_2 = __pyx_memoryview_convert_item_to_object(((struct __pyx_memoryview_obj *)__pyx_v_self), __pyx_v_itemp); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 985, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_2);
     __pyx_r = __pyx_t_2;
     __pyx_t_2 = 0;
     goto __pyx_L0;
   }
 
-  /* "View.MemoryView":979
+  /* "View.MemoryView":981
  *         __PYX_XDEC_MEMVIEW(&self.from_slice, 1)
  * 
  *     cdef convert_item_to_object(self, char *itemp):             # <<<<<<<<<<<<<<
@@ -17901,7 +17964,7 @@ static PyObject *__pyx_memoryviewslice_convert_item_to_object(struct __pyx_memor
   return __pyx_r;
 }
 
-/* "View.MemoryView":985
+/* "View.MemoryView":987
  *             return memoryview.convert_item_to_object(self, itemp)
  * 
  *     cdef assign_item_from_object(self, char *itemp, object value):             # <<<<<<<<<<<<<<
@@ -17920,7 +17983,7 @@ static PyObject *__pyx_memoryviewslice_assign_item_from_object(struct __pyx_memo
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("assign_item_from_object", 0);
 
-  /* "View.MemoryView":986
+  /* "View.MemoryView":988
  * 
  *     cdef assign_item_from_object(self, char *itemp, object value):
  *         if self.to_dtype_func != NULL:             # <<<<<<<<<<<<<<
@@ -17930,16 +17993,16 @@ static PyObject *__pyx_memoryviewslice_assign_item_from_object(struct __pyx_memo
   __pyx_t_1 = ((__pyx_v_self->to_dtype_func != NULL) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":987
+    /* "View.MemoryView":989
  *     cdef assign_item_from_object(self, char *itemp, object value):
  *         if self.to_dtype_func != NULL:
  *             self.to_dtype_func(itemp, value)             # <<<<<<<<<<<<<<
  *         else:
  *             memoryview.assign_item_from_object(self, itemp, value)
  */
-    __pyx_t_2 = __pyx_v_self->to_dtype_func(__pyx_v_itemp, __pyx_v_value); if (unlikely(__pyx_t_2 == ((int)0))) __PYX_ERR(1, 987, __pyx_L1_error)
+    __pyx_t_2 = __pyx_v_self->to_dtype_func(__pyx_v_itemp, __pyx_v_value); if (unlikely(__pyx_t_2 == ((int)0))) __PYX_ERR(1, 989, __pyx_L1_error)
 
-    /* "View.MemoryView":986
+    /* "View.MemoryView":988
  * 
  *     cdef assign_item_from_object(self, char *itemp, object value):
  *         if self.to_dtype_func != NULL:             # <<<<<<<<<<<<<<
@@ -17949,7 +18012,7 @@ static PyObject *__pyx_memoryviewslice_assign_item_from_object(struct __pyx_memo
     goto __pyx_L3;
   }
 
-  /* "View.MemoryView":989
+  /* "View.MemoryView":991
  *             self.to_dtype_func(itemp, value)
  *         else:
  *             memoryview.assign_item_from_object(self, itemp, value)             # <<<<<<<<<<<<<<
@@ -17957,13 +18020,13 @@ static PyObject *__pyx_memoryviewslice_assign_item_from_object(struct __pyx_memo
  *     @property
  */
   /*else*/ {
-    __pyx_t_3 = __pyx_memoryview_assign_item_from_object(((struct __pyx_memoryview_obj *)__pyx_v_self), __pyx_v_itemp, __pyx_v_value); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 989, __pyx_L1_error)
+    __pyx_t_3 = __pyx_memoryview_assign_item_from_object(((struct __pyx_memoryview_obj *)__pyx_v_self), __pyx_v_itemp, __pyx_v_value); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 991, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
     __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
   }
   __pyx_L3:;
 
-  /* "View.MemoryView":985
+  /* "View.MemoryView":987
  *             return memoryview.convert_item_to_object(self, itemp)
  * 
  *     cdef assign_item_from_object(self, char *itemp, object value):             # <<<<<<<<<<<<<<
@@ -17984,7 +18047,7 @@ static PyObject *__pyx_memoryviewslice_assign_item_from_object(struct __pyx_memo
   return __pyx_r;
 }
 
-/* "View.MemoryView":992
+/* "View.MemoryView":994
  * 
  *     @property
  *     def base(self):             # <<<<<<<<<<<<<<
@@ -18010,7 +18073,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_16_memoryviewslice_4base___get__
   __Pyx_RefNannyDeclarations
   __Pyx_RefNannySetupContext("__get__", 0);
 
-  /* "View.MemoryView":993
+  /* "View.MemoryView":995
  *     @property
  *     def base(self):
  *         return self.from_object             # <<<<<<<<<<<<<<
@@ -18022,7 +18085,7 @@ static PyObject *__pyx_pf_15View_dot_MemoryView_16_memoryviewslice_4base___get__
   __pyx_r = __pyx_v_self->from_object;
   goto __pyx_L0;
 
-  /* "View.MemoryView":992
+  /* "View.MemoryView":994
  * 
  *     @property
  *     def base(self):             # <<<<<<<<<<<<<<
@@ -18150,7 +18213,7 @@ static PyObject *__pyx_pf___pyx_memoryviewslice_2__setstate_cython__(CYTHON_UNUS
   return __pyx_r;
 }
 
-/* "View.MemoryView":999
+/* "View.MemoryView":1001
  * 
  * @cname('__pyx_memoryview_fromslice')
  * cdef memoryview_fromslice(__Pyx_memviewslice memviewslice,             # <<<<<<<<<<<<<<
@@ -18178,7 +18241,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("memoryview_fromslice", 0);
 
-  /* "View.MemoryView":1007
+  /* "View.MemoryView":1009
  *     cdef _memoryviewslice result
  * 
  *     if <PyObject *> memviewslice.memview == Py_None:             # <<<<<<<<<<<<<<
@@ -18188,7 +18251,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
   __pyx_t_1 = ((((PyObject *)__pyx_v_memviewslice.memview) == Py_None) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":1008
+    /* "View.MemoryView":1010
  * 
  *     if <PyObject *> memviewslice.memview == Py_None:
  *         return None             # <<<<<<<<<<<<<<
@@ -18199,7 +18262,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
     __pyx_r = Py_None; __Pyx_INCREF(Py_None);
     goto __pyx_L0;
 
-    /* "View.MemoryView":1007
+    /* "View.MemoryView":1009
  *     cdef _memoryviewslice result
  * 
  *     if <PyObject *> memviewslice.memview == Py_None:             # <<<<<<<<<<<<<<
@@ -18208,16 +18271,16 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
  */
   }
 
-  /* "View.MemoryView":1013
+  /* "View.MemoryView":1015
  * 
  * 
  *     result = _memoryviewslice(None, 0, dtype_is_object)             # <<<<<<<<<<<<<<
  * 
  *     result.from_slice = memviewslice
  */
-  __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_v_dtype_is_object); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1013, __pyx_L1_error)
+  __pyx_t_2 = __Pyx_PyBool_FromLong(__pyx_v_dtype_is_object); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1015, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
-  __pyx_t_3 = PyTuple_New(3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 1013, __pyx_L1_error)
+  __pyx_t_3 = PyTuple_New(3); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 1015, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_3);
   __Pyx_INCREF(Py_None);
   __Pyx_GIVEREF(Py_None);
@@ -18228,13 +18291,13 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
   __Pyx_GIVEREF(__pyx_t_2);
   PyTuple_SET_ITEM(__pyx_t_3, 2, __pyx_t_2);
   __pyx_t_2 = 0;
-  __pyx_t_2 = __Pyx_PyObject_Call(((PyObject *)__pyx_memoryviewslice_type), __pyx_t_3, NULL); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1013, __pyx_L1_error)
+  __pyx_t_2 = __Pyx_PyObject_Call(((PyObject *)__pyx_memoryviewslice_type), __pyx_t_3, NULL); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1015, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
   __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
   __pyx_v_result = ((struct __pyx_memoryviewslice_obj *)__pyx_t_2);
   __pyx_t_2 = 0;
 
-  /* "View.MemoryView":1015
+  /* "View.MemoryView":1017
  *     result = _memoryviewslice(None, 0, dtype_is_object)
  * 
  *     result.from_slice = memviewslice             # <<<<<<<<<<<<<<
@@ -18243,7 +18306,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
  */
   __pyx_v_result->from_slice = __pyx_v_memviewslice;
 
-  /* "View.MemoryView":1016
+  /* "View.MemoryView":1018
  * 
  *     result.from_slice = memviewslice
  *     __PYX_INC_MEMVIEW(&memviewslice, 1)             # <<<<<<<<<<<<<<
@@ -18252,14 +18315,14 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
  */
   __PYX_INC_MEMVIEW((&__pyx_v_memviewslice), 1);
 
-  /* "View.MemoryView":1018
+  /* "View.MemoryView":1020
  *     __PYX_INC_MEMVIEW(&memviewslice, 1)
  * 
  *     result.from_object = (<memoryview> memviewslice.memview).base             # <<<<<<<<<<<<<<
  *     result.typeinfo = memviewslice.memview.typeinfo
  * 
  */
-  __pyx_t_2 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_memviewslice.memview), __pyx_n_s_base); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1018, __pyx_L1_error)
+  __pyx_t_2 = __Pyx_PyObject_GetAttrStr(((PyObject *)__pyx_v_memviewslice.memview), __pyx_n_s_base); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1020, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
   __Pyx_GIVEREF(__pyx_t_2);
   __Pyx_GOTREF(__pyx_v_result->from_object);
@@ -18267,7 +18330,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
   __pyx_v_result->from_object = __pyx_t_2;
   __pyx_t_2 = 0;
 
-  /* "View.MemoryView":1019
+  /* "View.MemoryView":1021
  * 
  *     result.from_object = (<memoryview> memviewslice.memview).base
  *     result.typeinfo = memviewslice.memview.typeinfo             # <<<<<<<<<<<<<<
@@ -18277,7 +18340,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
   __pyx_t_4 = __pyx_v_memviewslice.memview->typeinfo;
   __pyx_v_result->__pyx_base.typeinfo = __pyx_t_4;
 
-  /* "View.MemoryView":1021
+  /* "View.MemoryView":1023
  *     result.typeinfo = memviewslice.memview.typeinfo
  * 
  *     result.view = memviewslice.memview.view             # <<<<<<<<<<<<<<
@@ -18287,7 +18350,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
   __pyx_t_5 = __pyx_v_memviewslice.memview->view;
   __pyx_v_result->__pyx_base.view = __pyx_t_5;
 
-  /* "View.MemoryView":1022
+  /* "View.MemoryView":1024
  * 
  *     result.view = memviewslice.memview.view
  *     result.view.buf = <void *> memviewslice.data             # <<<<<<<<<<<<<<
@@ -18296,7 +18359,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
  */
   __pyx_v_result->__pyx_base.view.buf = ((void *)__pyx_v_memviewslice.data);
 
-  /* "View.MemoryView":1023
+  /* "View.MemoryView":1025
  *     result.view = memviewslice.memview.view
  *     result.view.buf = <void *> memviewslice.data
  *     result.view.ndim = ndim             # <<<<<<<<<<<<<<
@@ -18305,7 +18368,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
  */
   __pyx_v_result->__pyx_base.view.ndim = __pyx_v_ndim;
 
-  /* "View.MemoryView":1024
+  /* "View.MemoryView":1026
  *     result.view.buf = <void *> memviewslice.data
  *     result.view.ndim = ndim
  *     (<__pyx_buffer *> &result.view).obj = Py_None             # <<<<<<<<<<<<<<
@@ -18314,7 +18377,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
  */
   ((Py_buffer *)(&__pyx_v_result->__pyx_base.view))->obj = Py_None;
 
-  /* "View.MemoryView":1025
+  /* "View.MemoryView":1027
  *     result.view.ndim = ndim
  *     (<__pyx_buffer *> &result.view).obj = Py_None
  *     Py_INCREF(Py_None)             # <<<<<<<<<<<<<<
@@ -18323,7 +18386,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
  */
   Py_INCREF(Py_None);
 
-  /* "View.MemoryView":1027
+  /* "View.MemoryView":1029
  *     Py_INCREF(Py_None)
  * 
  *     if (<memoryview>memviewslice.memview).flags & PyBUF_WRITABLE:             # <<<<<<<<<<<<<<
@@ -18333,7 +18396,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
   __pyx_t_1 = ((((struct __pyx_memoryview_obj *)__pyx_v_memviewslice.memview)->flags & PyBUF_WRITABLE) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":1028
+    /* "View.MemoryView":1030
  * 
  *     if (<memoryview>memviewslice.memview).flags & PyBUF_WRITABLE:
  *         result.flags = PyBUF_RECORDS             # <<<<<<<<<<<<<<
@@ -18342,7 +18405,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
  */
     __pyx_v_result->__pyx_base.flags = PyBUF_RECORDS;
 
-    /* "View.MemoryView":1027
+    /* "View.MemoryView":1029
  *     Py_INCREF(Py_None)
  * 
  *     if (<memoryview>memviewslice.memview).flags & PyBUF_WRITABLE:             # <<<<<<<<<<<<<<
@@ -18352,7 +18415,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
     goto __pyx_L4;
   }
 
-  /* "View.MemoryView":1030
+  /* "View.MemoryView":1032
  *         result.flags = PyBUF_RECORDS
  *     else:
  *         result.flags = PyBUF_RECORDS_RO             # <<<<<<<<<<<<<<
@@ -18364,7 +18427,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
   }
   __pyx_L4:;
 
-  /* "View.MemoryView":1032
+  /* "View.MemoryView":1034
  *         result.flags = PyBUF_RECORDS_RO
  * 
  *     result.view.shape = <Py_ssize_t *> result.from_slice.shape             # <<<<<<<<<<<<<<
@@ -18373,7 +18436,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
  */
   __pyx_v_result->__pyx_base.view.shape = ((Py_ssize_t *)__pyx_v_result->from_slice.shape);
 
-  /* "View.MemoryView":1033
+  /* "View.MemoryView":1035
  * 
  *     result.view.shape = <Py_ssize_t *> result.from_slice.shape
  *     result.view.strides = <Py_ssize_t *> result.from_slice.strides             # <<<<<<<<<<<<<<
@@ -18382,7 +18445,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
  */
   __pyx_v_result->__pyx_base.view.strides = ((Py_ssize_t *)__pyx_v_result->from_slice.strides);
 
-  /* "View.MemoryView":1036
+  /* "View.MemoryView":1038
  * 
  * 
  *     result.view.suboffsets = NULL             # <<<<<<<<<<<<<<
@@ -18391,7 +18454,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
  */
   __pyx_v_result->__pyx_base.view.suboffsets = NULL;
 
-  /* "View.MemoryView":1037
+  /* "View.MemoryView":1039
  * 
  *     result.view.suboffsets = NULL
  *     for suboffset in result.from_slice.suboffsets[:ndim]:             # <<<<<<<<<<<<<<
@@ -18403,7 +18466,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
     __pyx_t_6 = __pyx_t_8;
     __pyx_v_suboffset = (__pyx_t_6[0]);
 
-    /* "View.MemoryView":1038
+    /* "View.MemoryView":1040
  *     result.view.suboffsets = NULL
  *     for suboffset in result.from_slice.suboffsets[:ndim]:
  *         if suboffset >= 0:             # <<<<<<<<<<<<<<
@@ -18413,7 +18476,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
     __pyx_t_1 = ((__pyx_v_suboffset >= 0) != 0);
     if (__pyx_t_1) {
 
-      /* "View.MemoryView":1039
+      /* "View.MemoryView":1041
  *     for suboffset in result.from_slice.suboffsets[:ndim]:
  *         if suboffset >= 0:
  *             result.view.suboffsets = <Py_ssize_t *> result.from_slice.suboffsets             # <<<<<<<<<<<<<<
@@ -18422,7 +18485,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
  */
       __pyx_v_result->__pyx_base.view.suboffsets = ((Py_ssize_t *)__pyx_v_result->from_slice.suboffsets);
 
-      /* "View.MemoryView":1040
+      /* "View.MemoryView":1042
  *         if suboffset >= 0:
  *             result.view.suboffsets = <Py_ssize_t *> result.from_slice.suboffsets
  *             break             # <<<<<<<<<<<<<<
@@ -18431,7 +18494,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
  */
       goto __pyx_L6_break;
 
-      /* "View.MemoryView":1038
+      /* "View.MemoryView":1040
  *     result.view.suboffsets = NULL
  *     for suboffset in result.from_slice.suboffsets[:ndim]:
  *         if suboffset >= 0:             # <<<<<<<<<<<<<<
@@ -18442,7 +18505,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
   }
   __pyx_L6_break:;
 
-  /* "View.MemoryView":1042
+  /* "View.MemoryView":1044
  *             break
  * 
  *     result.view.len = result.view.itemsize             # <<<<<<<<<<<<<<
@@ -18452,7 +18515,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
   __pyx_t_9 = __pyx_v_result->__pyx_base.view.itemsize;
   __pyx_v_result->__pyx_base.view.len = __pyx_t_9;
 
-  /* "View.MemoryView":1043
+  /* "View.MemoryView":1045
  * 
  *     result.view.len = result.view.itemsize
  *     for length in result.view.shape[:ndim]:             # <<<<<<<<<<<<<<
@@ -18462,29 +18525,29 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
   __pyx_t_7 = (__pyx_v_result->__pyx_base.view.shape + __pyx_v_ndim);
   for (__pyx_t_8 = __pyx_v_result->__pyx_base.view.shape; __pyx_t_8 < __pyx_t_7; __pyx_t_8++) {
     __pyx_t_6 = __pyx_t_8;
-    __pyx_t_2 = PyInt_FromSsize_t((__pyx_t_6[0])); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1043, __pyx_L1_error)
+    __pyx_t_2 = PyInt_FromSsize_t((__pyx_t_6[0])); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1045, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_2);
     __Pyx_XDECREF_SET(__pyx_v_length, __pyx_t_2);
     __pyx_t_2 = 0;
 
-    /* "View.MemoryView":1044
+    /* "View.MemoryView":1046
  *     result.view.len = result.view.itemsize
  *     for length in result.view.shape[:ndim]:
  *         result.view.len *= length             # <<<<<<<<<<<<<<
  * 
  *     result.to_object_func = to_object_func
  */
-    __pyx_t_2 = PyInt_FromSsize_t(__pyx_v_result->__pyx_base.view.len); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1044, __pyx_L1_error)
+    __pyx_t_2 = PyInt_FromSsize_t(__pyx_v_result->__pyx_base.view.len); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1046, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_2);
-    __pyx_t_3 = PyNumber_InPlaceMultiply(__pyx_t_2, __pyx_v_length); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 1044, __pyx_L1_error)
+    __pyx_t_3 = PyNumber_InPlaceMultiply(__pyx_t_2, __pyx_v_length); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 1046, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
     __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
-    __pyx_t_9 = __Pyx_PyIndex_AsSsize_t(__pyx_t_3); if (unlikely((__pyx_t_9 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 1044, __pyx_L1_error)
+    __pyx_t_9 = __Pyx_PyIndex_AsSsize_t(__pyx_t_3); if (unlikely((__pyx_t_9 == (Py_ssize_t)-1) && PyErr_Occurred())) __PYX_ERR(1, 1046, __pyx_L1_error)
     __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
     __pyx_v_result->__pyx_base.view.len = __pyx_t_9;
   }
 
-  /* "View.MemoryView":1046
+  /* "View.MemoryView":1048
  *         result.view.len *= length
  * 
  *     result.to_object_func = to_object_func             # <<<<<<<<<<<<<<
@@ -18493,7 +18556,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
  */
   __pyx_v_result->to_object_func = __pyx_v_to_object_func;
 
-  /* "View.MemoryView":1047
+  /* "View.MemoryView":1049
  * 
  *     result.to_object_func = to_object_func
  *     result.to_dtype_func = to_dtype_func             # <<<<<<<<<<<<<<
@@ -18502,7 +18565,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
  */
   __pyx_v_result->to_dtype_func = __pyx_v_to_dtype_func;
 
-  /* "View.MemoryView":1049
+  /* "View.MemoryView":1051
  *     result.to_dtype_func = to_dtype_func
  * 
  *     return result             # <<<<<<<<<<<<<<
@@ -18514,7 +18577,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
   __pyx_r = ((PyObject *)__pyx_v_result);
   goto __pyx_L0;
 
-  /* "View.MemoryView":999
+  /* "View.MemoryView":1001
  * 
  * @cname('__pyx_memoryview_fromslice')
  * cdef memoryview_fromslice(__Pyx_memviewslice memviewslice,             # <<<<<<<<<<<<<<
@@ -18536,7 +18599,7 @@ static PyObject *__pyx_memoryview_fromslice(__Pyx_memviewslice __pyx_v_memviewsl
   return __pyx_r;
 }
 
-/* "View.MemoryView":1052
+/* "View.MemoryView":1054
  * 
  * @cname('__pyx_memoryview_get_slice_from_memoryview')
  * cdef __Pyx_memviewslice *get_slice_from_memview(memoryview memview,             # <<<<<<<<<<<<<<
@@ -18556,7 +18619,7 @@ static __Pyx_memviewslice *__pyx_memoryview_get_slice_from_memoryview(struct __p
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("get_slice_from_memview", 0);
 
-  /* "View.MemoryView":1055
+  /* "View.MemoryView":1057
  *                                                    __Pyx_memviewslice *mslice) except NULL:
  *     cdef _memoryviewslice obj
  *     if isinstance(memview, _memoryviewslice):             # <<<<<<<<<<<<<<
@@ -18567,20 +18630,20 @@ static __Pyx_memviewslice *__pyx_memoryview_get_slice_from_memoryview(struct __p
   __pyx_t_2 = (__pyx_t_1 != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":1056
+    /* "View.MemoryView":1058
  *     cdef _memoryviewslice obj
  *     if isinstance(memview, _memoryviewslice):
  *         obj = memview             # <<<<<<<<<<<<<<
  *         return &obj.from_slice
  *     else:
  */
-    if (!(likely(((((PyObject *)__pyx_v_memview)) == Py_None) || likely(__Pyx_TypeTest(((PyObject *)__pyx_v_memview), __pyx_memoryviewslice_type))))) __PYX_ERR(1, 1056, __pyx_L1_error)
+    if (!(likely(((((PyObject *)__pyx_v_memview)) == Py_None) || likely(__Pyx_TypeTest(((PyObject *)__pyx_v_memview), __pyx_memoryviewslice_type))))) __PYX_ERR(1, 1058, __pyx_L1_error)
     __pyx_t_3 = ((PyObject *)__pyx_v_memview);
     __Pyx_INCREF(__pyx_t_3);
     __pyx_v_obj = ((struct __pyx_memoryviewslice_obj *)__pyx_t_3);
     __pyx_t_3 = 0;
 
-    /* "View.MemoryView":1057
+    /* "View.MemoryView":1059
  *     if isinstance(memview, _memoryviewslice):
  *         obj = memview
  *         return &obj.from_slice             # <<<<<<<<<<<<<<
@@ -18590,7 +18653,7 @@ static __Pyx_memviewslice *__pyx_memoryview_get_slice_from_memoryview(struct __p
     __pyx_r = (&__pyx_v_obj->from_slice);
     goto __pyx_L0;
 
-    /* "View.MemoryView":1055
+    /* "View.MemoryView":1057
  *                                                    __Pyx_memviewslice *mslice) except NULL:
  *     cdef _memoryviewslice obj
  *     if isinstance(memview, _memoryviewslice):             # <<<<<<<<<<<<<<
@@ -18599,7 +18662,7 @@ static __Pyx_memviewslice *__pyx_memoryview_get_slice_from_memoryview(struct __p
  */
   }
 
-  /* "View.MemoryView":1059
+  /* "View.MemoryView":1061
  *         return &obj.from_slice
  *     else:
  *         slice_copy(memview, mslice)             # <<<<<<<<<<<<<<
@@ -18609,7 +18672,7 @@ static __Pyx_memviewslice *__pyx_memoryview_get_slice_from_memoryview(struct __p
   /*else*/ {
     __pyx_memoryview_slice_copy(__pyx_v_memview, __pyx_v_mslice);
 
-    /* "View.MemoryView":1060
+    /* "View.MemoryView":1062
  *     else:
  *         slice_copy(memview, mslice)
  *         return mslice             # <<<<<<<<<<<<<<
@@ -18620,7 +18683,7 @@ static __Pyx_memviewslice *__pyx_memoryview_get_slice_from_memoryview(struct __p
     goto __pyx_L0;
   }
 
-  /* "View.MemoryView":1052
+  /* "View.MemoryView":1054
  * 
  * @cname('__pyx_memoryview_get_slice_from_memoryview')
  * cdef __Pyx_memviewslice *get_slice_from_memview(memoryview memview,             # <<<<<<<<<<<<<<
@@ -18639,7 +18702,7 @@ static __Pyx_memviewslice *__pyx_memoryview_get_slice_from_memoryview(struct __p
   return __pyx_r;
 }
 
-/* "View.MemoryView":1063
+/* "View.MemoryView":1065
  * 
  * @cname('__pyx_memoryview_slice_copy')
  * cdef void slice_copy(memoryview memview, __Pyx_memviewslice *dst):             # <<<<<<<<<<<<<<
@@ -18660,7 +18723,7 @@ static void __pyx_memoryview_slice_copy(struct __pyx_memoryview_obj *__pyx_v_mem
   Py_ssize_t __pyx_t_5;
   __Pyx_RefNannySetupContext("slice_copy", 0);
 
-  /* "View.MemoryView":1067
+  /* "View.MemoryView":1069
  *     cdef (Py_ssize_t*) shape, strides, suboffsets
  * 
  *     shape = memview.view.shape             # <<<<<<<<<<<<<<
@@ -18670,7 +18733,7 @@ static void __pyx_memoryview_slice_copy(struct __pyx_memoryview_obj *__pyx_v_mem
   __pyx_t_1 = __pyx_v_memview->view.shape;
   __pyx_v_shape = __pyx_t_1;
 
-  /* "View.MemoryView":1068
+  /* "View.MemoryView":1070
  * 
  *     shape = memview.view.shape
  *     strides = memview.view.strides             # <<<<<<<<<<<<<<
@@ -18680,7 +18743,7 @@ static void __pyx_memoryview_slice_copy(struct __pyx_memoryview_obj *__pyx_v_mem
   __pyx_t_1 = __pyx_v_memview->view.strides;
   __pyx_v_strides = __pyx_t_1;
 
-  /* "View.MemoryView":1069
+  /* "View.MemoryView":1071
  *     shape = memview.view.shape
  *     strides = memview.view.strides
  *     suboffsets = memview.view.suboffsets             # <<<<<<<<<<<<<<
@@ -18690,7 +18753,7 @@ static void __pyx_memoryview_slice_copy(struct __pyx_memoryview_obj *__pyx_v_mem
   __pyx_t_1 = __pyx_v_memview->view.suboffsets;
   __pyx_v_suboffsets = __pyx_t_1;
 
-  /* "View.MemoryView":1071
+  /* "View.MemoryView":1073
  *     suboffsets = memview.view.suboffsets
  * 
  *     dst.memview = <__pyx_memoryview *> memview             # <<<<<<<<<<<<<<
@@ -18699,7 +18762,7 @@ static void __pyx_memoryview_slice_copy(struct __pyx_memoryview_obj *__pyx_v_mem
  */
   __pyx_v_dst->memview = ((struct __pyx_memoryview_obj *)__pyx_v_memview);
 
-  /* "View.MemoryView":1072
+  /* "View.MemoryView":1074
  * 
  *     dst.memview = <__pyx_memoryview *> memview
  *     dst.data = <char *> memview.view.buf             # <<<<<<<<<<<<<<
@@ -18708,7 +18771,7 @@ static void __pyx_memoryview_slice_copy(struct __pyx_memoryview_obj *__pyx_v_mem
  */
   __pyx_v_dst->data = ((char *)__pyx_v_memview->view.buf);
 
-  /* "View.MemoryView":1074
+  /* "View.MemoryView":1076
  *     dst.data = <char *> memview.view.buf
  * 
  *     for dim in range(memview.view.ndim):             # <<<<<<<<<<<<<<
@@ -18720,7 +18783,7 @@ static void __pyx_memoryview_slice_copy(struct __pyx_memoryview_obj *__pyx_v_mem
   for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) {
     __pyx_v_dim = __pyx_t_4;
 
-    /* "View.MemoryView":1075
+    /* "View.MemoryView":1077
  * 
  *     for dim in range(memview.view.ndim):
  *         dst.shape[dim] = shape[dim]             # <<<<<<<<<<<<<<
@@ -18729,7 +18792,7 @@ static void __pyx_memoryview_slice_copy(struct __pyx_memoryview_obj *__pyx_v_mem
  */
     (__pyx_v_dst->shape[__pyx_v_dim]) = (__pyx_v_shape[__pyx_v_dim]);
 
-    /* "View.MemoryView":1076
+    /* "View.MemoryView":1078
  *     for dim in range(memview.view.ndim):
  *         dst.shape[dim] = shape[dim]
  *         dst.strides[dim] = strides[dim]             # <<<<<<<<<<<<<<
@@ -18738,7 +18801,7 @@ static void __pyx_memoryview_slice_copy(struct __pyx_memoryview_obj *__pyx_v_mem
  */
     (__pyx_v_dst->strides[__pyx_v_dim]) = (__pyx_v_strides[__pyx_v_dim]);
 
-    /* "View.MemoryView":1077
+    /* "View.MemoryView":1079
  *         dst.shape[dim] = shape[dim]
  *         dst.strides[dim] = strides[dim]
  *         dst.suboffsets[dim] = suboffsets[dim] if suboffsets else -1             # <<<<<<<<<<<<<<
@@ -18753,7 +18816,7 @@ static void __pyx_memoryview_slice_copy(struct __pyx_memoryview_obj *__pyx_v_mem
     (__pyx_v_dst->suboffsets[__pyx_v_dim]) = __pyx_t_5;
   }
 
-  /* "View.MemoryView":1063
+  /* "View.MemoryView":1065
  * 
  * @cname('__pyx_memoryview_slice_copy')
  * cdef void slice_copy(memoryview memview, __Pyx_memviewslice *dst):             # <<<<<<<<<<<<<<
@@ -18765,7 +18828,7 @@ static void __pyx_memoryview_slice_copy(struct __pyx_memoryview_obj *__pyx_v_mem
   __Pyx_RefNannyFinishContext();
 }
 
-/* "View.MemoryView":1080
+/* "View.MemoryView":1082
  * 
  * @cname('__pyx_memoryview_copy_object')
  * cdef memoryview_copy(memoryview memview):             # <<<<<<<<<<<<<<
@@ -18783,7 +18846,7 @@ static PyObject *__pyx_memoryview_copy_object(struct __pyx_memoryview_obj *__pyx
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("memoryview_copy", 0);
 
-  /* "View.MemoryView":1083
+  /* "View.MemoryView":1085
  *     "Create a new memoryview object"
  *     cdef __Pyx_memviewslice memviewslice
  *     slice_copy(memview, &memviewslice)             # <<<<<<<<<<<<<<
@@ -18792,7 +18855,7 @@ static PyObject *__pyx_memoryview_copy_object(struct __pyx_memoryview_obj *__pyx
  */
   __pyx_memoryview_slice_copy(__pyx_v_memview, (&__pyx_v_memviewslice));
 
-  /* "View.MemoryView":1084
+  /* "View.MemoryView":1086
  *     cdef __Pyx_memviewslice memviewslice
  *     slice_copy(memview, &memviewslice)
  *     return memoryview_copy_from_slice(memview, &memviewslice)             # <<<<<<<<<<<<<<
@@ -18800,13 +18863,13 @@ static PyObject *__pyx_memoryview_copy_object(struct __pyx_memoryview_obj *__pyx
  * @cname('__pyx_memoryview_copy_object_from_slice')
  */
   __Pyx_XDECREF(__pyx_r);
-  __pyx_t_1 = __pyx_memoryview_copy_object_from_slice(__pyx_v_memview, (&__pyx_v_memviewslice)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 1084, __pyx_L1_error)
+  __pyx_t_1 = __pyx_memoryview_copy_object_from_slice(__pyx_v_memview, (&__pyx_v_memviewslice)); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 1086, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
   __pyx_r = __pyx_t_1;
   __pyx_t_1 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":1080
+  /* "View.MemoryView":1082
  * 
  * @cname('__pyx_memoryview_copy_object')
  * cdef memoryview_copy(memoryview memview):             # <<<<<<<<<<<<<<
@@ -18825,7 +18888,7 @@ static PyObject *__pyx_memoryview_copy_object(struct __pyx_memoryview_obj *__pyx
   return __pyx_r;
 }
 
-/* "View.MemoryView":1087
+/* "View.MemoryView":1089
  * 
  * @cname('__pyx_memoryview_copy_object_from_slice')
  * cdef memoryview_copy_from_slice(memoryview memview, __Pyx_memviewslice *memviewslice):             # <<<<<<<<<<<<<<
@@ -18848,7 +18911,7 @@ static PyObject *__pyx_memoryview_copy_object_from_slice(struct __pyx_memoryview
   int __pyx_clineno = 0;
   __Pyx_RefNannySetupContext("memoryview_copy_from_slice", 0);
 
-  /* "View.MemoryView":1094
+  /* "View.MemoryView":1096
  *     cdef int (*to_dtype_func)(char *, object) except 0
  * 
  *     if isinstance(memview, _memoryviewslice):             # <<<<<<<<<<<<<<
@@ -18859,7 +18922,7 @@ static PyObject *__pyx_memoryview_copy_object_from_slice(struct __pyx_memoryview
   __pyx_t_2 = (__pyx_t_1 != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":1095
+    /* "View.MemoryView":1097
  * 
  *     if isinstance(memview, _memoryviewslice):
  *         to_object_func = (<_memoryviewslice> memview).to_object_func             # <<<<<<<<<<<<<<
@@ -18869,7 +18932,7 @@ static PyObject *__pyx_memoryview_copy_object_from_slice(struct __pyx_memoryview
     __pyx_t_3 = ((struct __pyx_memoryviewslice_obj *)__pyx_v_memview)->to_object_func;
     __pyx_v_to_object_func = __pyx_t_3;
 
-    /* "View.MemoryView":1096
+    /* "View.MemoryView":1098
  *     if isinstance(memview, _memoryviewslice):
  *         to_object_func = (<_memoryviewslice> memview).to_object_func
  *         to_dtype_func = (<_memoryviewslice> memview).to_dtype_func             # <<<<<<<<<<<<<<
@@ -18879,7 +18942,7 @@ static PyObject *__pyx_memoryview_copy_object_from_slice(struct __pyx_memoryview
     __pyx_t_4 = ((struct __pyx_memoryviewslice_obj *)__pyx_v_memview)->to_dtype_func;
     __pyx_v_to_dtype_func = __pyx_t_4;
 
-    /* "View.MemoryView":1094
+    /* "View.MemoryView":1096
  *     cdef int (*to_dtype_func)(char *, object) except 0
  * 
  *     if isinstance(memview, _memoryviewslice):             # <<<<<<<<<<<<<<
@@ -18889,7 +18952,7 @@ static PyObject *__pyx_memoryview_copy_object_from_slice(struct __pyx_memoryview
     goto __pyx_L3;
   }
 
-  /* "View.MemoryView":1098
+  /* "View.MemoryView":1100
  *         to_dtype_func = (<_memoryviewslice> memview).to_dtype_func
  *     else:
  *         to_object_func = NULL             # <<<<<<<<<<<<<<
@@ -18899,7 +18962,7 @@ static PyObject *__pyx_memoryview_copy_object_from_slice(struct __pyx_memoryview
   /*else*/ {
     __pyx_v_to_object_func = NULL;
 
-    /* "View.MemoryView":1099
+    /* "View.MemoryView":1101
  *     else:
  *         to_object_func = NULL
  *         to_dtype_func = NULL             # <<<<<<<<<<<<<<
@@ -18910,7 +18973,7 @@ static PyObject *__pyx_memoryview_copy_object_from_slice(struct __pyx_memoryview
   }
   __pyx_L3:;
 
-  /* "View.MemoryView":1101
+  /* "View.MemoryView":1103
  *         to_dtype_func = NULL
  * 
  *     return memoryview_fromslice(memviewslice[0], memview.view.ndim,             # <<<<<<<<<<<<<<
@@ -18919,20 +18982,20 @@ static PyObject *__pyx_memoryview_copy_object_from_slice(struct __pyx_memoryview
  */
   __Pyx_XDECREF(__pyx_r);
 
-  /* "View.MemoryView":1103
+  /* "View.MemoryView":1105
  *     return memoryview_fromslice(memviewslice[0], memview.view.ndim,
  *                                 to_object_func, to_dtype_func,
  *                                 memview.dtype_is_object)             # <<<<<<<<<<<<<<
  * 
  * 
  */
-  __pyx_t_5 = __pyx_memoryview_fromslice((__pyx_v_memviewslice[0]), __pyx_v_memview->view.ndim, __pyx_v_to_object_func, __pyx_v_to_dtype_func, __pyx_v_memview->dtype_is_object); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 1101, __pyx_L1_error)
+  __pyx_t_5 = __pyx_memoryview_fromslice((__pyx_v_memviewslice[0]), __pyx_v_memview->view.ndim, __pyx_v_to_object_func, __pyx_v_to_dtype_func, __pyx_v_memview->dtype_is_object); if (unlikely(!__pyx_t_5)) __PYX_ERR(1, 1103, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_5);
   __pyx_r = __pyx_t_5;
   __pyx_t_5 = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":1087
+  /* "View.MemoryView":1089
  * 
  * @cname('__pyx_memoryview_copy_object_from_slice')
  * cdef memoryview_copy_from_slice(memoryview memview, __Pyx_memviewslice *memviewslice):             # <<<<<<<<<<<<<<
@@ -18951,7 +19014,7 @@ static PyObject *__pyx_memoryview_copy_object_from_slice(struct __pyx_memoryview
   return __pyx_r;
 }
 
-/* "View.MemoryView":1109
+/* "View.MemoryView":1111
  * 
  * 
  * cdef Py_ssize_t abs_py_ssize_t(Py_ssize_t arg) nogil:             # <<<<<<<<<<<<<<
@@ -18963,7 +19026,7 @@ static Py_ssize_t abs_py_ssize_t(Py_ssize_t __pyx_v_arg) {
   Py_ssize_t __pyx_r;
   int __pyx_t_1;
 
-  /* "View.MemoryView":1110
+  /* "View.MemoryView":1112
  * 
  * cdef Py_ssize_t abs_py_ssize_t(Py_ssize_t arg) nogil:
  *     if arg < 0:             # <<<<<<<<<<<<<<
@@ -18973,7 +19036,7 @@ static Py_ssize_t abs_py_ssize_t(Py_ssize_t __pyx_v_arg) {
   __pyx_t_1 = ((__pyx_v_arg < 0) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":1111
+    /* "View.MemoryView":1113
  * cdef Py_ssize_t abs_py_ssize_t(Py_ssize_t arg) nogil:
  *     if arg < 0:
  *         return -arg             # <<<<<<<<<<<<<<
@@ -18983,7 +19046,7 @@ static Py_ssize_t abs_py_ssize_t(Py_ssize_t __pyx_v_arg) {
     __pyx_r = (-__pyx_v_arg);
     goto __pyx_L0;
 
-    /* "View.MemoryView":1110
+    /* "View.MemoryView":1112
  * 
  * cdef Py_ssize_t abs_py_ssize_t(Py_ssize_t arg) nogil:
  *     if arg < 0:             # <<<<<<<<<<<<<<
@@ -18992,7 +19055,7 @@ static Py_ssize_t abs_py_ssize_t(Py_ssize_t __pyx_v_arg) {
  */
   }
 
-  /* "View.MemoryView":1113
+  /* "View.MemoryView":1115
  *         return -arg
  *     else:
  *         return arg             # <<<<<<<<<<<<<<
@@ -19004,7 +19067,7 @@ static Py_ssize_t abs_py_ssize_t(Py_ssize_t __pyx_v_arg) {
     goto __pyx_L0;
   }
 
-  /* "View.MemoryView":1109
+  /* "View.MemoryView":1111
  * 
  * 
  * cdef Py_ssize_t abs_py_ssize_t(Py_ssize_t arg) nogil:             # <<<<<<<<<<<<<<
@@ -19017,7 +19080,7 @@ static Py_ssize_t abs_py_ssize_t(Py_ssize_t __pyx_v_arg) {
   return __pyx_r;
 }
 
-/* "View.MemoryView":1116
+/* "View.MemoryView":1118
  * 
  * @cname('__pyx_get_best_slice_order')
  * cdef char get_best_order(__Pyx_memviewslice *mslice, int ndim) nogil:             # <<<<<<<<<<<<<<
@@ -19035,7 +19098,7 @@ static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int _
   int __pyx_t_3;
   int __pyx_t_4;
 
-  /* "View.MemoryView":1121
+  /* "View.MemoryView":1123
  *     """
  *     cdef int i
  *     cdef Py_ssize_t c_stride = 0             # <<<<<<<<<<<<<<
@@ -19044,7 +19107,7 @@ static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int _
  */
   __pyx_v_c_stride = 0;
 
-  /* "View.MemoryView":1122
+  /* "View.MemoryView":1124
  *     cdef int i
  *     cdef Py_ssize_t c_stride = 0
  *     cdef Py_ssize_t f_stride = 0             # <<<<<<<<<<<<<<
@@ -19053,7 +19116,7 @@ static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int _
  */
   __pyx_v_f_stride = 0;
 
-  /* "View.MemoryView":1124
+  /* "View.MemoryView":1126
  *     cdef Py_ssize_t f_stride = 0
  * 
  *     for i in range(ndim - 1, -1, -1):             # <<<<<<<<<<<<<<
@@ -19063,7 +19126,7 @@ static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int _
   for (__pyx_t_1 = (__pyx_v_ndim - 1); __pyx_t_1 > -1; __pyx_t_1-=1) {
     __pyx_v_i = __pyx_t_1;
 
-    /* "View.MemoryView":1125
+    /* "View.MemoryView":1127
  * 
  *     for i in range(ndim - 1, -1, -1):
  *         if mslice.shape[i] > 1:             # <<<<<<<<<<<<<<
@@ -19073,7 +19136,7 @@ static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int _
     __pyx_t_2 = (((__pyx_v_mslice->shape[__pyx_v_i]) > 1) != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":1126
+      /* "View.MemoryView":1128
  *     for i in range(ndim - 1, -1, -1):
  *         if mslice.shape[i] > 1:
  *             c_stride = mslice.strides[i]             # <<<<<<<<<<<<<<
@@ -19082,7 +19145,7 @@ static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int _
  */
       __pyx_v_c_stride = (__pyx_v_mslice->strides[__pyx_v_i]);
 
-      /* "View.MemoryView":1127
+      /* "View.MemoryView":1129
  *         if mslice.shape[i] > 1:
  *             c_stride = mslice.strides[i]
  *             break             # <<<<<<<<<<<<<<
@@ -19091,7 +19154,7 @@ static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int _
  */
       goto __pyx_L4_break;
 
-      /* "View.MemoryView":1125
+      /* "View.MemoryView":1127
  * 
  *     for i in range(ndim - 1, -1, -1):
  *         if mslice.shape[i] > 1:             # <<<<<<<<<<<<<<
@@ -19102,7 +19165,7 @@ static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int _
   }
   __pyx_L4_break:;
 
-  /* "View.MemoryView":1129
+  /* "View.MemoryView":1131
  *             break
  * 
  *     for i in range(ndim):             # <<<<<<<<<<<<<<
@@ -19114,7 +19177,7 @@ static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int _
   for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) {
     __pyx_v_i = __pyx_t_4;
 
-    /* "View.MemoryView":1130
+    /* "View.MemoryView":1132
  * 
  *     for i in range(ndim):
  *         if mslice.shape[i] > 1:             # <<<<<<<<<<<<<<
@@ -19124,7 +19187,7 @@ static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int _
     __pyx_t_2 = (((__pyx_v_mslice->shape[__pyx_v_i]) > 1) != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":1131
+      /* "View.MemoryView":1133
  *     for i in range(ndim):
  *         if mslice.shape[i] > 1:
  *             f_stride = mslice.strides[i]             # <<<<<<<<<<<<<<
@@ -19133,7 +19196,7 @@ static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int _
  */
       __pyx_v_f_stride = (__pyx_v_mslice->strides[__pyx_v_i]);
 
-      /* "View.MemoryView":1132
+      /* "View.MemoryView":1134
  *         if mslice.shape[i] > 1:
  *             f_stride = mslice.strides[i]
  *             break             # <<<<<<<<<<<<<<
@@ -19142,7 +19205,7 @@ static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int _
  */
       goto __pyx_L7_break;
 
-      /* "View.MemoryView":1130
+      /* "View.MemoryView":1132
  * 
  *     for i in range(ndim):
  *         if mslice.shape[i] > 1:             # <<<<<<<<<<<<<<
@@ -19153,7 +19216,7 @@ static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int _
   }
   __pyx_L7_break:;
 
-  /* "View.MemoryView":1134
+  /* "View.MemoryView":1136
  *             break
  * 
  *     if abs_py_ssize_t(c_stride) <= abs_py_ssize_t(f_stride):             # <<<<<<<<<<<<<<
@@ -19163,7 +19226,7 @@ static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int _
   __pyx_t_2 = ((abs_py_ssize_t(__pyx_v_c_stride) <= abs_py_ssize_t(__pyx_v_f_stride)) != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":1135
+    /* "View.MemoryView":1137
  * 
  *     if abs_py_ssize_t(c_stride) <= abs_py_ssize_t(f_stride):
  *         return 'C'             # <<<<<<<<<<<<<<
@@ -19173,7 +19236,7 @@ static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int _
     __pyx_r = 'C';
     goto __pyx_L0;
 
-    /* "View.MemoryView":1134
+    /* "View.MemoryView":1136
  *             break
  * 
  *     if abs_py_ssize_t(c_stride) <= abs_py_ssize_t(f_stride):             # <<<<<<<<<<<<<<
@@ -19182,7 +19245,7 @@ static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int _
  */
   }
 
-  /* "View.MemoryView":1137
+  /* "View.MemoryView":1139
  *         return 'C'
  *     else:
  *         return 'F'             # <<<<<<<<<<<<<<
@@ -19194,7 +19257,7 @@ static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int _
     goto __pyx_L0;
   }
 
-  /* "View.MemoryView":1116
+  /* "View.MemoryView":1118
  * 
  * @cname('__pyx_get_best_slice_order')
  * cdef char get_best_order(__Pyx_memviewslice *mslice, int ndim) nogil:             # <<<<<<<<<<<<<<
@@ -19207,7 +19270,7 @@ static char __pyx_get_best_slice_order(__Pyx_memviewslice *__pyx_v_mslice, int _
   return __pyx_r;
 }
 
-/* "View.MemoryView":1140
+/* "View.MemoryView":1142
  * 
  * @cython.cdivision(True)
  * cdef void _copy_strided_to_strided(char *src_data, Py_ssize_t *src_strides,             # <<<<<<<<<<<<<<
@@ -19228,7 +19291,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
   Py_ssize_t __pyx_t_5;
   Py_ssize_t __pyx_t_6;
 
-  /* "View.MemoryView":1147
+  /* "View.MemoryView":1149
  * 
  *     cdef Py_ssize_t i
  *     cdef Py_ssize_t src_extent = src_shape[0]             # <<<<<<<<<<<<<<
@@ -19237,7 +19300,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
  */
   __pyx_v_src_extent = (__pyx_v_src_shape[0]);
 
-  /* "View.MemoryView":1148
+  /* "View.MemoryView":1150
  *     cdef Py_ssize_t i
  *     cdef Py_ssize_t src_extent = src_shape[0]
  *     cdef Py_ssize_t dst_extent = dst_shape[0]             # <<<<<<<<<<<<<<
@@ -19246,7 +19309,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
  */
   __pyx_v_dst_extent = (__pyx_v_dst_shape[0]);
 
-  /* "View.MemoryView":1149
+  /* "View.MemoryView":1151
  *     cdef Py_ssize_t src_extent = src_shape[0]
  *     cdef Py_ssize_t dst_extent = dst_shape[0]
  *     cdef Py_ssize_t src_stride = src_strides[0]             # <<<<<<<<<<<<<<
@@ -19255,7 +19318,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
  */
   __pyx_v_src_stride = (__pyx_v_src_strides[0]);
 
-  /* "View.MemoryView":1150
+  /* "View.MemoryView":1152
  *     cdef Py_ssize_t dst_extent = dst_shape[0]
  *     cdef Py_ssize_t src_stride = src_strides[0]
  *     cdef Py_ssize_t dst_stride = dst_strides[0]             # <<<<<<<<<<<<<<
@@ -19264,7 +19327,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
  */
   __pyx_v_dst_stride = (__pyx_v_dst_strides[0]);
 
-  /* "View.MemoryView":1152
+  /* "View.MemoryView":1154
  *     cdef Py_ssize_t dst_stride = dst_strides[0]
  * 
  *     if ndim == 1:             # <<<<<<<<<<<<<<
@@ -19274,7 +19337,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
   __pyx_t_1 = ((__pyx_v_ndim == 1) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":1153
+    /* "View.MemoryView":1155
  * 
  *     if ndim == 1:
  *        if (src_stride > 0 and dst_stride > 0 and             # <<<<<<<<<<<<<<
@@ -19294,7 +19357,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
       goto __pyx_L5_bool_binop_done;
     }
 
-    /* "View.MemoryView":1154
+    /* "View.MemoryView":1156
  *     if ndim == 1:
  *        if (src_stride > 0 and dst_stride > 0 and
  *            <size_t> src_stride == itemsize == <size_t> dst_stride):             # <<<<<<<<<<<<<<
@@ -19309,7 +19372,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
     __pyx_t_1 = __pyx_t_3;
     __pyx_L5_bool_binop_done:;
 
-    /* "View.MemoryView":1153
+    /* "View.MemoryView":1155
  * 
  *     if ndim == 1:
  *        if (src_stride > 0 and dst_stride > 0 and             # <<<<<<<<<<<<<<
@@ -19318,7 +19381,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
  */
     if (__pyx_t_1) {
 
-      /* "View.MemoryView":1155
+      /* "View.MemoryView":1157
  *        if (src_stride > 0 and dst_stride > 0 and
  *            <size_t> src_stride == itemsize == <size_t> dst_stride):
  *            memcpy(dst_data, src_data, itemsize * dst_extent)             # <<<<<<<<<<<<<<
@@ -19327,7 +19390,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
  */
       (void)(memcpy(__pyx_v_dst_data, __pyx_v_src_data, (__pyx_v_itemsize * __pyx_v_dst_extent)));
 
-      /* "View.MemoryView":1153
+      /* "View.MemoryView":1155
  * 
  *     if ndim == 1:
  *        if (src_stride > 0 and dst_stride > 0 and             # <<<<<<<<<<<<<<
@@ -19337,7 +19400,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
       goto __pyx_L4;
     }
 
-    /* "View.MemoryView":1157
+    /* "View.MemoryView":1159
  *            memcpy(dst_data, src_data, itemsize * dst_extent)
  *        else:
  *            for i in range(dst_extent):             # <<<<<<<<<<<<<<
@@ -19350,7 +19413,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
       for (__pyx_t_6 = 0; __pyx_t_6 < __pyx_t_5; __pyx_t_6+=1) {
         __pyx_v_i = __pyx_t_6;
 
-        /* "View.MemoryView":1158
+        /* "View.MemoryView":1160
  *        else:
  *            for i in range(dst_extent):
  *                memcpy(dst_data, src_data, itemsize)             # <<<<<<<<<<<<<<
@@ -19359,7 +19422,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
  */
         (void)(memcpy(__pyx_v_dst_data, __pyx_v_src_data, __pyx_v_itemsize));
 
-        /* "View.MemoryView":1159
+        /* "View.MemoryView":1161
  *            for i in range(dst_extent):
  *                memcpy(dst_data, src_data, itemsize)
  *                src_data += src_stride             # <<<<<<<<<<<<<<
@@ -19368,7 +19431,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
  */
         __pyx_v_src_data = (__pyx_v_src_data + __pyx_v_src_stride);
 
-        /* "View.MemoryView":1160
+        /* "View.MemoryView":1162
  *                memcpy(dst_data, src_data, itemsize)
  *                src_data += src_stride
  *                dst_data += dst_stride             # <<<<<<<<<<<<<<
@@ -19380,7 +19443,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
     }
     __pyx_L4:;
 
-    /* "View.MemoryView":1152
+    /* "View.MemoryView":1154
  *     cdef Py_ssize_t dst_stride = dst_strides[0]
  * 
  *     if ndim == 1:             # <<<<<<<<<<<<<<
@@ -19390,7 +19453,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
     goto __pyx_L3;
   }
 
-  /* "View.MemoryView":1162
+  /* "View.MemoryView":1164
  *                dst_data += dst_stride
  *     else:
  *         for i in range(dst_extent):             # <<<<<<<<<<<<<<
@@ -19403,7 +19466,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
     for (__pyx_t_6 = 0; __pyx_t_6 < __pyx_t_5; __pyx_t_6+=1) {
       __pyx_v_i = __pyx_t_6;
 
-      /* "View.MemoryView":1163
+      /* "View.MemoryView":1165
  *     else:
  *         for i in range(dst_extent):
  *             _copy_strided_to_strided(src_data, src_strides + 1,             # <<<<<<<<<<<<<<
@@ -19412,7 +19475,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
  */
       _copy_strided_to_strided(__pyx_v_src_data, (__pyx_v_src_strides + 1), __pyx_v_dst_data, (__pyx_v_dst_strides + 1), (__pyx_v_src_shape + 1), (__pyx_v_dst_shape + 1), (__pyx_v_ndim - 1), __pyx_v_itemsize);
 
-      /* "View.MemoryView":1167
+      /* "View.MemoryView":1169
  *                                      src_shape + 1, dst_shape + 1,
  *                                      ndim - 1, itemsize)
  *             src_data += src_stride             # <<<<<<<<<<<<<<
@@ -19421,7 +19484,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
  */
       __pyx_v_src_data = (__pyx_v_src_data + __pyx_v_src_stride);
 
-      /* "View.MemoryView":1168
+      /* "View.MemoryView":1170
  *                                      ndim - 1, itemsize)
  *             src_data += src_stride
  *             dst_data += dst_stride             # <<<<<<<<<<<<<<
@@ -19433,7 +19496,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
   }
   __pyx_L3:;
 
-  /* "View.MemoryView":1140
+  /* "View.MemoryView":1142
  * 
  * @cython.cdivision(True)
  * cdef void _copy_strided_to_strided(char *src_data, Py_ssize_t *src_strides,             # <<<<<<<<<<<<<<
@@ -19444,7 +19507,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
   /* function exit code */
 }
 
-/* "View.MemoryView":1170
+/* "View.MemoryView":1172
  *             dst_data += dst_stride
  * 
  * cdef void copy_strided_to_strided(__Pyx_memviewslice *src,             # <<<<<<<<<<<<<<
@@ -19454,7 +19517,7 @@ static void _copy_strided_to_strided(char *__pyx_v_src_data, Py_ssize_t *__pyx_v
 
 static void copy_strided_to_strided(__Pyx_memviewslice *__pyx_v_src, __Pyx_memviewslice *__pyx_v_dst, int __pyx_v_ndim, size_t __pyx_v_itemsize) {
 
-  /* "View.MemoryView":1173
+  /* "View.MemoryView":1175
  *                                   __Pyx_memviewslice *dst,
  *                                   int ndim, size_t itemsize) nogil:
  *     _copy_strided_to_strided(src.data, src.strides, dst.data, dst.strides,             # <<<<<<<<<<<<<<
@@ -19463,7 +19526,7 @@ static void copy_strided_to_strided(__Pyx_memviewslice *__pyx_v_src, __Pyx_memvi
  */
   _copy_strided_to_strided(__pyx_v_src->data, __pyx_v_src->strides, __pyx_v_dst->data, __pyx_v_dst->strides, __pyx_v_src->shape, __pyx_v_dst->shape, __pyx_v_ndim, __pyx_v_itemsize);
 
-  /* "View.MemoryView":1170
+  /* "View.MemoryView":1172
  *             dst_data += dst_stride
  * 
  * cdef void copy_strided_to_strided(__Pyx_memviewslice *src,             # <<<<<<<<<<<<<<
@@ -19474,7 +19537,7 @@ static void copy_strided_to_strided(__Pyx_memviewslice *__pyx_v_src, __Pyx_memvi
   /* function exit code */
 }
 
-/* "View.MemoryView":1177
+/* "View.MemoryView":1179
  * 
  * @cname('__pyx_memoryview_slice_get_size')
  * cdef Py_ssize_t slice_get_size(__Pyx_memviewslice *src, int ndim) nogil:             # <<<<<<<<<<<<<<
@@ -19491,7 +19554,7 @@ static Py_ssize_t __pyx_memoryview_slice_get_size(__Pyx_memviewslice *__pyx_v_sr
   Py_ssize_t *__pyx_t_3;
   Py_ssize_t *__pyx_t_4;
 
-  /* "View.MemoryView":1179
+  /* "View.MemoryView":1181
  * cdef Py_ssize_t slice_get_size(__Pyx_memviewslice *src, int ndim) nogil:
  *     "Return the size of the memory occupied by the slice in number of bytes"
  *     cdef Py_ssize_t shape, size = src.memview.view.itemsize             # <<<<<<<<<<<<<<
@@ -19501,7 +19564,7 @@ static Py_ssize_t __pyx_memoryview_slice_get_size(__Pyx_memviewslice *__pyx_v_sr
   __pyx_t_1 = __pyx_v_src->memview->view.itemsize;
   __pyx_v_size = __pyx_t_1;
 
-  /* "View.MemoryView":1181
+  /* "View.MemoryView":1183
  *     cdef Py_ssize_t shape, size = src.memview.view.itemsize
  * 
  *     for shape in src.shape[:ndim]:             # <<<<<<<<<<<<<<
@@ -19513,7 +19576,7 @@ static Py_ssize_t __pyx_memoryview_slice_get_size(__Pyx_memviewslice *__pyx_v_sr
     __pyx_t_2 = __pyx_t_4;
     __pyx_v_shape = (__pyx_t_2[0]);
 
-    /* "View.MemoryView":1182
+    /* "View.MemoryView":1184
  * 
  *     for shape in src.shape[:ndim]:
  *         size *= shape             # <<<<<<<<<<<<<<
@@ -19523,7 +19586,7 @@ static Py_ssize_t __pyx_memoryview_slice_get_size(__Pyx_memviewslice *__pyx_v_sr
     __pyx_v_size = (__pyx_v_size * __pyx_v_shape);
   }
 
-  /* "View.MemoryView":1184
+  /* "View.MemoryView":1186
  *         size *= shape
  * 
  *     return size             # <<<<<<<<<<<<<<
@@ -19533,7 +19596,7 @@ static Py_ssize_t __pyx_memoryview_slice_get_size(__Pyx_memviewslice *__pyx_v_sr
   __pyx_r = __pyx_v_size;
   goto __pyx_L0;
 
-  /* "View.MemoryView":1177
+  /* "View.MemoryView":1179
  * 
  * @cname('__pyx_memoryview_slice_get_size')
  * cdef Py_ssize_t slice_get_size(__Pyx_memviewslice *src, int ndim) nogil:             # <<<<<<<<<<<<<<
@@ -19546,7 +19609,7 @@ static Py_ssize_t __pyx_memoryview_slice_get_size(__Pyx_memviewslice *__pyx_v_sr
   return __pyx_r;
 }
 
-/* "View.MemoryView":1187
+/* "View.MemoryView":1189
  * 
  * @cname('__pyx_fill_contig_strides_array')
  * cdef Py_ssize_t fill_contig_strides_array(             # <<<<<<<<<<<<<<
@@ -19562,7 +19625,7 @@ static Py_ssize_t __pyx_fill_contig_strides_array(Py_ssize_t *__pyx_v_shape, Py_
   int __pyx_t_3;
   int __pyx_t_4;
 
-  /* "View.MemoryView":1196
+  /* "View.MemoryView":1198
  *     cdef int idx
  * 
  *     if order == 'F':             # <<<<<<<<<<<<<<
@@ -19572,7 +19635,7 @@ static Py_ssize_t __pyx_fill_contig_strides_array(Py_ssize_t *__pyx_v_shape, Py_
   __pyx_t_1 = ((__pyx_v_order == 'F') != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":1197
+    /* "View.MemoryView":1199
  * 
  *     if order == 'F':
  *         for idx in range(ndim):             # <<<<<<<<<<<<<<
@@ -19584,7 +19647,7 @@ static Py_ssize_t __pyx_fill_contig_strides_array(Py_ssize_t *__pyx_v_shape, Py_
     for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) {
       __pyx_v_idx = __pyx_t_4;
 
-      /* "View.MemoryView":1198
+      /* "View.MemoryView":1200
  *     if order == 'F':
  *         for idx in range(ndim):
  *             strides[idx] = stride             # <<<<<<<<<<<<<<
@@ -19593,7 +19656,7 @@ static Py_ssize_t __pyx_fill_contig_strides_array(Py_ssize_t *__pyx_v_shape, Py_
  */
       (__pyx_v_strides[__pyx_v_idx]) = __pyx_v_stride;
 
-      /* "View.MemoryView":1199
+      /* "View.MemoryView":1201
  *         for idx in range(ndim):
  *             strides[idx] = stride
  *             stride *= shape[idx]             # <<<<<<<<<<<<<<
@@ -19603,7 +19666,7 @@ static Py_ssize_t __pyx_fill_contig_strides_array(Py_ssize_t *__pyx_v_shape, Py_
       __pyx_v_stride = (__pyx_v_stride * (__pyx_v_shape[__pyx_v_idx]));
     }
 
-    /* "View.MemoryView":1196
+    /* "View.MemoryView":1198
  *     cdef int idx
  * 
  *     if order == 'F':             # <<<<<<<<<<<<<<
@@ -19613,7 +19676,7 @@ static Py_ssize_t __pyx_fill_contig_strides_array(Py_ssize_t *__pyx_v_shape, Py_
     goto __pyx_L3;
   }
 
-  /* "View.MemoryView":1201
+  /* "View.MemoryView":1203
  *             stride *= shape[idx]
  *     else:
  *         for idx in range(ndim - 1, -1, -1):             # <<<<<<<<<<<<<<
@@ -19624,7 +19687,7 @@ static Py_ssize_t __pyx_fill_contig_strides_array(Py_ssize_t *__pyx_v_shape, Py_
     for (__pyx_t_2 = (__pyx_v_ndim - 1); __pyx_t_2 > -1; __pyx_t_2-=1) {
       __pyx_v_idx = __pyx_t_2;
 
-      /* "View.MemoryView":1202
+      /* "View.MemoryView":1204
  *     else:
  *         for idx in range(ndim - 1, -1, -1):
  *             strides[idx] = stride             # <<<<<<<<<<<<<<
@@ -19633,7 +19696,7 @@ static Py_ssize_t __pyx_fill_contig_strides_array(Py_ssize_t *__pyx_v_shape, Py_
  */
       (__pyx_v_strides[__pyx_v_idx]) = __pyx_v_stride;
 
-      /* "View.MemoryView":1203
+      /* "View.MemoryView":1205
  *         for idx in range(ndim - 1, -1, -1):
  *             strides[idx] = stride
  *             stride *= shape[idx]             # <<<<<<<<<<<<<<
@@ -19645,7 +19708,7 @@ static Py_ssize_t __pyx_fill_contig_strides_array(Py_ssize_t *__pyx_v_shape, Py_
   }
   __pyx_L3:;
 
-  /* "View.MemoryView":1205
+  /* "View.MemoryView":1207
  *             stride *= shape[idx]
  * 
  *     return stride             # <<<<<<<<<<<<<<
@@ -19655,7 +19718,7 @@ static Py_ssize_t __pyx_fill_contig_strides_array(Py_ssize_t *__pyx_v_shape, Py_
   __pyx_r = __pyx_v_stride;
   goto __pyx_L0;
 
-  /* "View.MemoryView":1187
+  /* "View.MemoryView":1189
  * 
  * @cname('__pyx_fill_contig_strides_array')
  * cdef Py_ssize_t fill_contig_strides_array(             # <<<<<<<<<<<<<<
@@ -19668,7 +19731,7 @@ static Py_ssize_t __pyx_fill_contig_strides_array(Py_ssize_t *__pyx_v_shape, Py_
   return __pyx_r;
 }
 
-/* "View.MemoryView":1208
+/* "View.MemoryView":1210
  * 
  * @cname('__pyx_memoryview_copy_data_to_temp')
  * cdef void *copy_data_to_temp(__Pyx_memviewslice *src,             # <<<<<<<<<<<<<<
@@ -19692,7 +19755,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
   const char *__pyx_filename = NULL;
   int __pyx_clineno = 0;
 
-  /* "View.MemoryView":1219
+  /* "View.MemoryView":1221
  *     cdef void *result
  * 
  *     cdef size_t itemsize = src.memview.view.itemsize             # <<<<<<<<<<<<<<
@@ -19702,7 +19765,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
   __pyx_t_1 = __pyx_v_src->memview->view.itemsize;
   __pyx_v_itemsize = __pyx_t_1;
 
-  /* "View.MemoryView":1220
+  /* "View.MemoryView":1222
  * 
  *     cdef size_t itemsize = src.memview.view.itemsize
  *     cdef size_t size = slice_get_size(src, ndim)             # <<<<<<<<<<<<<<
@@ -19711,7 +19774,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
  */
   __pyx_v_size = __pyx_memoryview_slice_get_size(__pyx_v_src, __pyx_v_ndim);
 
-  /* "View.MemoryView":1222
+  /* "View.MemoryView":1224
  *     cdef size_t size = slice_get_size(src, ndim)
  * 
  *     result = malloc(size)             # <<<<<<<<<<<<<<
@@ -19720,7 +19783,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
  */
   __pyx_v_result = malloc(__pyx_v_size);
 
-  /* "View.MemoryView":1223
+  /* "View.MemoryView":1225
  * 
  *     result = malloc(size)
  *     if not result:             # <<<<<<<<<<<<<<
@@ -19730,16 +19793,16 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
   __pyx_t_2 = ((!(__pyx_v_result != 0)) != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":1224
+    /* "View.MemoryView":1226
  *     result = malloc(size)
  *     if not result:
  *         _err(MemoryError, NULL)             # <<<<<<<<<<<<<<
  * 
  * 
  */
-    __pyx_t_3 = __pyx_memoryview_err(__pyx_builtin_MemoryError, NULL); if (unlikely(__pyx_t_3 == ((int)-1))) __PYX_ERR(1, 1224, __pyx_L1_error)
+    __pyx_t_3 = __pyx_memoryview_err(__pyx_builtin_MemoryError, NULL); if (unlikely(__pyx_t_3 == ((int)-1))) __PYX_ERR(1, 1226, __pyx_L1_error)
 
-    /* "View.MemoryView":1223
+    /* "View.MemoryView":1225
  * 
  *     result = malloc(size)
  *     if not result:             # <<<<<<<<<<<<<<
@@ -19748,7 +19811,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
  */
   }
 
-  /* "View.MemoryView":1227
+  /* "View.MemoryView":1229
  * 
  * 
  *     tmpslice.data = <char *> result             # <<<<<<<<<<<<<<
@@ -19757,7 +19820,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
  */
   __pyx_v_tmpslice->data = ((char *)__pyx_v_result);
 
-  /* "View.MemoryView":1228
+  /* "View.MemoryView":1230
  * 
  *     tmpslice.data = <char *> result
  *     tmpslice.memview = src.memview             # <<<<<<<<<<<<<<
@@ -19767,7 +19830,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
   __pyx_t_4 = __pyx_v_src->memview;
   __pyx_v_tmpslice->memview = __pyx_t_4;
 
-  /* "View.MemoryView":1229
+  /* "View.MemoryView":1231
  *     tmpslice.data = <char *> result
  *     tmpslice.memview = src.memview
  *     for i in range(ndim):             # <<<<<<<<<<<<<<
@@ -19779,7 +19842,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
   for (__pyx_t_6 = 0; __pyx_t_6 < __pyx_t_5; __pyx_t_6+=1) {
     __pyx_v_i = __pyx_t_6;
 
-    /* "View.MemoryView":1230
+    /* "View.MemoryView":1232
  *     tmpslice.memview = src.memview
  *     for i in range(ndim):
  *         tmpslice.shape[i] = src.shape[i]             # <<<<<<<<<<<<<<
@@ -19788,7 +19851,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
  */
     (__pyx_v_tmpslice->shape[__pyx_v_i]) = (__pyx_v_src->shape[__pyx_v_i]);
 
-    /* "View.MemoryView":1231
+    /* "View.MemoryView":1233
  *     for i in range(ndim):
  *         tmpslice.shape[i] = src.shape[i]
  *         tmpslice.suboffsets[i] = -1             # <<<<<<<<<<<<<<
@@ -19798,7 +19861,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
     (__pyx_v_tmpslice->suboffsets[__pyx_v_i]) = -1L;
   }
 
-  /* "View.MemoryView":1233
+  /* "View.MemoryView":1235
  *         tmpslice.suboffsets[i] = -1
  * 
  *     fill_contig_strides_array(&tmpslice.shape[0], &tmpslice.strides[0], itemsize,             # <<<<<<<<<<<<<<
@@ -19807,7 +19870,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
  */
   (void)(__pyx_fill_contig_strides_array((&(__pyx_v_tmpslice->shape[0])), (&(__pyx_v_tmpslice->strides[0])), __pyx_v_itemsize, __pyx_v_ndim, __pyx_v_order));
 
-  /* "View.MemoryView":1237
+  /* "View.MemoryView":1239
  * 
  * 
  *     for i in range(ndim):             # <<<<<<<<<<<<<<
@@ -19819,7 +19882,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
   for (__pyx_t_6 = 0; __pyx_t_6 < __pyx_t_5; __pyx_t_6+=1) {
     __pyx_v_i = __pyx_t_6;
 
-    /* "View.MemoryView":1238
+    /* "View.MemoryView":1240
  * 
  *     for i in range(ndim):
  *         if tmpslice.shape[i] == 1:             # <<<<<<<<<<<<<<
@@ -19829,7 +19892,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
     __pyx_t_2 = (((__pyx_v_tmpslice->shape[__pyx_v_i]) == 1) != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":1239
+      /* "View.MemoryView":1241
  *     for i in range(ndim):
  *         if tmpslice.shape[i] == 1:
  *             tmpslice.strides[i] = 0             # <<<<<<<<<<<<<<
@@ -19838,7 +19901,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
  */
       (__pyx_v_tmpslice->strides[__pyx_v_i]) = 0;
 
-      /* "View.MemoryView":1238
+      /* "View.MemoryView":1240
  * 
  *     for i in range(ndim):
  *         if tmpslice.shape[i] == 1:             # <<<<<<<<<<<<<<
@@ -19848,7 +19911,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
     }
   }
 
-  /* "View.MemoryView":1241
+  /* "View.MemoryView":1243
  *             tmpslice.strides[i] = 0
  * 
  *     if slice_is_contig(src[0], order, ndim):             # <<<<<<<<<<<<<<
@@ -19858,7 +19921,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
   __pyx_t_2 = (__pyx_memviewslice_is_contig((__pyx_v_src[0]), __pyx_v_order, __pyx_v_ndim) != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":1242
+    /* "View.MemoryView":1244
  * 
  *     if slice_is_contig(src[0], order, ndim):
  *         memcpy(result, src.data, size)             # <<<<<<<<<<<<<<
@@ -19867,7 +19930,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
  */
     (void)(memcpy(__pyx_v_result, __pyx_v_src->data, __pyx_v_size));
 
-    /* "View.MemoryView":1241
+    /* "View.MemoryView":1243
  *             tmpslice.strides[i] = 0
  * 
  *     if slice_is_contig(src[0], order, ndim):             # <<<<<<<<<<<<<<
@@ -19877,7 +19940,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
     goto __pyx_L9;
   }
 
-  /* "View.MemoryView":1244
+  /* "View.MemoryView":1246
  *         memcpy(result, src.data, size)
  *     else:
  *         copy_strided_to_strided(src, tmpslice, ndim, itemsize)             # <<<<<<<<<<<<<<
@@ -19889,7 +19952,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
   }
   __pyx_L9:;
 
-  /* "View.MemoryView":1246
+  /* "View.MemoryView":1248
  *         copy_strided_to_strided(src, tmpslice, ndim, itemsize)
  * 
  *     return result             # <<<<<<<<<<<<<<
@@ -19899,7 +19962,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
   __pyx_r = __pyx_v_result;
   goto __pyx_L0;
 
-  /* "View.MemoryView":1208
+  /* "View.MemoryView":1210
  * 
  * @cname('__pyx_memoryview_copy_data_to_temp')
  * cdef void *copy_data_to_temp(__Pyx_memviewslice *src,             # <<<<<<<<<<<<<<
@@ -19923,7 +19986,7 @@ static void *__pyx_memoryview_copy_data_to_temp(__Pyx_memviewslice *__pyx_v_src,
   return __pyx_r;
 }
 
-/* "View.MemoryView":1251
+/* "View.MemoryView":1253
  * 
  * @cname('__pyx_memoryview_err_extents')
  * cdef int _err_extents(int i, Py_ssize_t extent1,             # <<<<<<<<<<<<<<
@@ -19946,20 +20009,20 @@ static int __pyx_memoryview_err_extents(int __pyx_v_i, Py_ssize_t __pyx_v_extent
   #endif
   __Pyx_RefNannySetupContext("_err_extents", 0);
 
-  /* "View.MemoryView":1254
+  /* "View.MemoryView":1256
  *                              Py_ssize_t extent2) except -1 with gil:
  *     raise ValueError("got differing extents in dimension %d (got %d and %d)" %
  *                                                         (i, extent1, extent2))             # <<<<<<<<<<<<<<
  * 
  * @cname('__pyx_memoryview_err_dim')
  */
-  __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_i); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 1254, __pyx_L1_error)
+  __pyx_t_1 = __Pyx_PyInt_From_int(__pyx_v_i); if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 1256, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
-  __pyx_t_2 = PyInt_FromSsize_t(__pyx_v_extent1); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1254, __pyx_L1_error)
+  __pyx_t_2 = PyInt_FromSsize_t(__pyx_v_extent1); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1256, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
-  __pyx_t_3 = PyInt_FromSsize_t(__pyx_v_extent2); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 1254, __pyx_L1_error)
+  __pyx_t_3 = PyInt_FromSsize_t(__pyx_v_extent2); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 1256, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_3);
-  __pyx_t_4 = PyTuple_New(3); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 1254, __pyx_L1_error)
+  __pyx_t_4 = PyTuple_New(3); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 1256, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_4);
   __Pyx_GIVEREF(__pyx_t_1);
   PyTuple_SET_ITEM(__pyx_t_4, 0, __pyx_t_1);
@@ -19971,24 +20034,24 @@ static int __pyx_memoryview_err_extents(int __pyx_v_i, Py_ssize_t __pyx_v_extent
   __pyx_t_2 = 0;
   __pyx_t_3 = 0;
 
-  /* "View.MemoryView":1253
+  /* "View.MemoryView":1255
  * cdef int _err_extents(int i, Py_ssize_t extent1,
  *                              Py_ssize_t extent2) except -1 with gil:
  *     raise ValueError("got differing extents in dimension %d (got %d and %d)" %             # <<<<<<<<<<<<<<
  *                                                         (i, extent1, extent2))
  * 
  */
-  __pyx_t_3 = __Pyx_PyString_Format(__pyx_kp_s_got_differing_extents_in_dimensi, __pyx_t_4); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 1253, __pyx_L1_error)
+  __pyx_t_3 = __Pyx_PyString_Format(__pyx_kp_s_got_differing_extents_in_dimensi, __pyx_t_4); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 1255, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_3);
   __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
-  __pyx_t_4 = __Pyx_PyObject_CallOneArg(__pyx_builtin_ValueError, __pyx_t_3); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 1253, __pyx_L1_error)
+  __pyx_t_4 = __Pyx_PyObject_CallOneArg(__pyx_builtin_ValueError, __pyx_t_3); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 1255, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_4);
   __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
   __Pyx_Raise(__pyx_t_4, 0, 0, 0);
   __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
-  __PYX_ERR(1, 1253, __pyx_L1_error)
+  __PYX_ERR(1, 1255, __pyx_L1_error)
 
-  /* "View.MemoryView":1251
+  /* "View.MemoryView":1253
  * 
  * @cname('__pyx_memoryview_err_extents')
  * cdef int _err_extents(int i, Py_ssize_t extent1,             # <<<<<<<<<<<<<<
@@ -20011,7 +20074,7 @@ static int __pyx_memoryview_err_extents(int __pyx_v_i, Py_ssize_t __pyx_v_extent
   return __pyx_r;
 }
 
-/* "View.MemoryView":1257
+/* "View.MemoryView":1259
  * 
  * @cname('__pyx_memoryview_err_dim')
  * cdef int _err_dim(object error, char *msg, int dim) except -1 with gil:             # <<<<<<<<<<<<<<
@@ -20035,18 +20098,18 @@ static int __pyx_memoryview_err_dim(PyObject *__pyx_v_error, char *__pyx_v_msg,
   __Pyx_RefNannySetupContext("_err_dim", 0);
   __Pyx_INCREF(__pyx_v_error);
 
-  /* "View.MemoryView":1258
+  /* "View.MemoryView":1260
  * @cname('__pyx_memoryview_err_dim')
  * cdef int _err_dim(object error, char *msg, int dim) except -1 with gil:
  *     raise error(msg.decode('ascii') % dim)             # <<<<<<<<<<<<<<
  * 
  * @cname('__pyx_memoryview_err')
  */
-  __pyx_t_2 = __Pyx_decode_c_string(__pyx_v_msg, 0, strlen(__pyx_v_msg), NULL, NULL, PyUnicode_DecodeASCII); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1258, __pyx_L1_error)
+  __pyx_t_2 = __Pyx_decode_c_string(__pyx_v_msg, 0, strlen(__pyx_v_msg), NULL, NULL, PyUnicode_DecodeASCII); if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1260, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_2);
-  __pyx_t_3 = __Pyx_PyInt_From_int(__pyx_v_dim); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 1258, __pyx_L1_error)
+  __pyx_t_3 = __Pyx_PyInt_From_int(__pyx_v_dim); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 1260, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_3);
-  __pyx_t_4 = PyUnicode_Format(__pyx_t_2, __pyx_t_3); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 1258, __pyx_L1_error)
+  __pyx_t_4 = PyUnicode_Format(__pyx_t_2, __pyx_t_3); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 1260, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_4);
   __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
   __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
@@ -20064,14 +20127,14 @@ static int __pyx_memoryview_err_dim(PyObject *__pyx_v_error, char *__pyx_v_msg,
   __pyx_t_1 = (__pyx_t_2) ? __Pyx_PyObject_Call2Args(__pyx_t_3, __pyx_t_2, __pyx_t_4) : __Pyx_PyObject_CallOneArg(__pyx_t_3, __pyx_t_4);
   __Pyx_XDECREF(__pyx_t_2); __pyx_t_2 = 0;
   __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
-  if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 1258, __pyx_L1_error)
+  if (unlikely(!__pyx_t_1)) __PYX_ERR(1, 1260, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_1);
   __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
   __Pyx_Raise(__pyx_t_1, 0, 0, 0);
   __Pyx_DECREF(__pyx_t_1); __pyx_t_1 = 0;
-  __PYX_ERR(1, 1258, __pyx_L1_error)
+  __PYX_ERR(1, 1260, __pyx_L1_error)
 
-  /* "View.MemoryView":1257
+  /* "View.MemoryView":1259
  * 
  * @cname('__pyx_memoryview_err_dim')
  * cdef int _err_dim(object error, char *msg, int dim) except -1 with gil:             # <<<<<<<<<<<<<<
@@ -20095,7 +20158,7 @@ static int __pyx_memoryview_err_dim(PyObject *__pyx_v_error, char *__pyx_v_msg,
   return __pyx_r;
 }
 
-/* "View.MemoryView":1261
+/* "View.MemoryView":1263
  * 
  * @cname('__pyx_memoryview_err')
  * cdef int _err(object error, char *msg) except -1 with gil:             # <<<<<<<<<<<<<<
@@ -20120,7 +20183,7 @@ static int __pyx_memoryview_err(PyObject *__pyx_v_error, char *__pyx_v_msg) {
   __Pyx_RefNannySetupContext("_err", 0);
   __Pyx_INCREF(__pyx_v_error);
 
-  /* "View.MemoryView":1262
+  /* "View.MemoryView":1264
  * @cname('__pyx_memoryview_err')
  * cdef int _err(object error, char *msg) except -1 with gil:
  *     if msg != NULL:             # <<<<<<<<<<<<<<
@@ -20130,14 +20193,14 @@ static int __pyx_memoryview_err(PyObject *__pyx_v_error, char *__pyx_v_msg) {
   __pyx_t_1 = ((__pyx_v_msg != NULL) != 0);
   if (unlikely(__pyx_t_1)) {
 
-    /* "View.MemoryView":1263
+    /* "View.MemoryView":1265
  * cdef int _err(object error, char *msg) except -1 with gil:
  *     if msg != NULL:
  *         raise error(msg.decode('ascii'))             # <<<<<<<<<<<<<<
  *     else:
  *         raise error
  */
-    __pyx_t_3 = __Pyx_decode_c_string(__pyx_v_msg, 0, strlen(__pyx_v_msg), NULL, NULL, PyUnicode_DecodeASCII); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 1263, __pyx_L1_error)
+    __pyx_t_3 = __Pyx_decode_c_string(__pyx_v_msg, 0, strlen(__pyx_v_msg), NULL, NULL, PyUnicode_DecodeASCII); if (unlikely(!__pyx_t_3)) __PYX_ERR(1, 1265, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_3);
     __Pyx_INCREF(__pyx_v_error);
     __pyx_t_4 = __pyx_v_error; __pyx_t_5 = NULL;
@@ -20153,14 +20216,14 @@ static int __pyx_memoryview_err(PyObject *__pyx_v_error, char *__pyx_v_msg) {
     __pyx_t_2 = (__pyx_t_5) ? __Pyx_PyObject_Call2Args(__pyx_t_4, __pyx_t_5, __pyx_t_3) : __Pyx_PyObject_CallOneArg(__pyx_t_4, __pyx_t_3);
     __Pyx_XDECREF(__pyx_t_5); __pyx_t_5 = 0;
     __Pyx_DECREF(__pyx_t_3); __pyx_t_3 = 0;
-    if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1263, __pyx_L1_error)
+    if (unlikely(!__pyx_t_2)) __PYX_ERR(1, 1265, __pyx_L1_error)
     __Pyx_GOTREF(__pyx_t_2);
     __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
     __Pyx_Raise(__pyx_t_2, 0, 0, 0);
     __Pyx_DECREF(__pyx_t_2); __pyx_t_2 = 0;
-    __PYX_ERR(1, 1263, __pyx_L1_error)
+    __PYX_ERR(1, 1265, __pyx_L1_error)
 
-    /* "View.MemoryView":1262
+    /* "View.MemoryView":1264
  * @cname('__pyx_memoryview_err')
  * cdef int _err(object error, char *msg) except -1 with gil:
  *     if msg != NULL:             # <<<<<<<<<<<<<<
@@ -20169,7 +20232,7 @@ static int __pyx_memoryview_err(PyObject *__pyx_v_error, char *__pyx_v_msg) {
  */
   }
 
-  /* "View.MemoryView":1265
+  /* "View.MemoryView":1267
  *         raise error(msg.decode('ascii'))
  *     else:
  *         raise error             # <<<<<<<<<<<<<<
@@ -20178,10 +20241,10 @@ static int __pyx_memoryview_err(PyObject *__pyx_v_error, char *__pyx_v_msg) {
  */
   /*else*/ {
     __Pyx_Raise(__pyx_v_error, 0, 0, 0);
-    __PYX_ERR(1, 1265, __pyx_L1_error)
+    __PYX_ERR(1, 1267, __pyx_L1_error)
   }
 
-  /* "View.MemoryView":1261
+  /* "View.MemoryView":1263
  * 
  * @cname('__pyx_memoryview_err')
  * cdef int _err(object error, char *msg) except -1 with gil:             # <<<<<<<<<<<<<<
@@ -20205,7 +20268,7 @@ static int __pyx_memoryview_err(PyObject *__pyx_v_error, char *__pyx_v_msg) {
   return __pyx_r;
 }
 
-/* "View.MemoryView":1268
+/* "View.MemoryView":1270
  * 
  * @cname('__pyx_memoryview_copy_contents')
  * cdef int memoryview_copy_contents(__Pyx_memviewslice src,             # <<<<<<<<<<<<<<
@@ -20235,7 +20298,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
   const char *__pyx_filename = NULL;
   int __pyx_clineno = 0;
 
-  /* "View.MemoryView":1276
+  /* "View.MemoryView":1278
  *     Check for overlapping memory and verify the shapes.
  *     """
  *     cdef void *tmpdata = NULL             # <<<<<<<<<<<<<<
@@ -20244,7 +20307,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
   __pyx_v_tmpdata = NULL;
 
-  /* "View.MemoryView":1277
+  /* "View.MemoryView":1279
  *     """
  *     cdef void *tmpdata = NULL
  *     cdef size_t itemsize = src.memview.view.itemsize             # <<<<<<<<<<<<<<
@@ -20254,7 +20317,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
   __pyx_t_1 = __pyx_v_src.memview->view.itemsize;
   __pyx_v_itemsize = __pyx_t_1;
 
-  /* "View.MemoryView":1279
+  /* "View.MemoryView":1281
  *     cdef size_t itemsize = src.memview.view.itemsize
  *     cdef int i
  *     cdef char order = get_best_order(&src, src_ndim)             # <<<<<<<<<<<<<<
@@ -20263,7 +20326,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
   __pyx_v_order = __pyx_get_best_slice_order((&__pyx_v_src), __pyx_v_src_ndim);
 
-  /* "View.MemoryView":1280
+  /* "View.MemoryView":1282
  *     cdef int i
  *     cdef char order = get_best_order(&src, src_ndim)
  *     cdef bint broadcasting = False             # <<<<<<<<<<<<<<
@@ -20272,7 +20335,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
   __pyx_v_broadcasting = 0;
 
-  /* "View.MemoryView":1281
+  /* "View.MemoryView":1283
  *     cdef char order = get_best_order(&src, src_ndim)
  *     cdef bint broadcasting = False
  *     cdef bint direct_copy = False             # <<<<<<<<<<<<<<
@@ -20281,7 +20344,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
   __pyx_v_direct_copy = 0;
 
-  /* "View.MemoryView":1284
+  /* "View.MemoryView":1286
  *     cdef __Pyx_memviewslice tmp
  * 
  *     if src_ndim < dst_ndim:             # <<<<<<<<<<<<<<
@@ -20291,7 +20354,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
   __pyx_t_2 = ((__pyx_v_src_ndim < __pyx_v_dst_ndim) != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":1285
+    /* "View.MemoryView":1287
  * 
  *     if src_ndim < dst_ndim:
  *         broadcast_leading(&src, src_ndim, dst_ndim)             # <<<<<<<<<<<<<<
@@ -20300,7 +20363,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
     __pyx_memoryview_broadcast_leading((&__pyx_v_src), __pyx_v_src_ndim, __pyx_v_dst_ndim);
 
-    /* "View.MemoryView":1284
+    /* "View.MemoryView":1286
  *     cdef __Pyx_memviewslice tmp
  * 
  *     if src_ndim < dst_ndim:             # <<<<<<<<<<<<<<
@@ -20310,7 +20373,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
     goto __pyx_L3;
   }
 
-  /* "View.MemoryView":1286
+  /* "View.MemoryView":1288
  *     if src_ndim < dst_ndim:
  *         broadcast_leading(&src, src_ndim, dst_ndim)
  *     elif dst_ndim < src_ndim:             # <<<<<<<<<<<<<<
@@ -20320,7 +20383,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
   __pyx_t_2 = ((__pyx_v_dst_ndim < __pyx_v_src_ndim) != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":1287
+    /* "View.MemoryView":1289
  *         broadcast_leading(&src, src_ndim, dst_ndim)
  *     elif dst_ndim < src_ndim:
  *         broadcast_leading(&dst, dst_ndim, src_ndim)             # <<<<<<<<<<<<<<
@@ -20329,7 +20392,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
     __pyx_memoryview_broadcast_leading((&__pyx_v_dst), __pyx_v_dst_ndim, __pyx_v_src_ndim);
 
-    /* "View.MemoryView":1286
+    /* "View.MemoryView":1288
  *     if src_ndim < dst_ndim:
  *         broadcast_leading(&src, src_ndim, dst_ndim)
  *     elif dst_ndim < src_ndim:             # <<<<<<<<<<<<<<
@@ -20339,7 +20402,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
   }
   __pyx_L3:;
 
-  /* "View.MemoryView":1289
+  /* "View.MemoryView":1291
  *         broadcast_leading(&dst, dst_ndim, src_ndim)
  * 
  *     cdef int ndim = max(src_ndim, dst_ndim)             # <<<<<<<<<<<<<<
@@ -20355,7 +20418,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
   }
   __pyx_v_ndim = __pyx_t_5;
 
-  /* "View.MemoryView":1291
+  /* "View.MemoryView":1293
  *     cdef int ndim = max(src_ndim, dst_ndim)
  * 
  *     for i in range(ndim):             # <<<<<<<<<<<<<<
@@ -20367,7 +20430,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
   for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) {
     __pyx_v_i = __pyx_t_4;
 
-    /* "View.MemoryView":1292
+    /* "View.MemoryView":1294
  * 
  *     for i in range(ndim):
  *         if src.shape[i] != dst.shape[i]:             # <<<<<<<<<<<<<<
@@ -20377,7 +20440,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
     __pyx_t_2 = (((__pyx_v_src.shape[__pyx_v_i]) != (__pyx_v_dst.shape[__pyx_v_i])) != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":1293
+      /* "View.MemoryView":1295
  *     for i in range(ndim):
  *         if src.shape[i] != dst.shape[i]:
  *             if src.shape[i] == 1:             # <<<<<<<<<<<<<<
@@ -20387,7 +20450,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
       __pyx_t_2 = (((__pyx_v_src.shape[__pyx_v_i]) == 1) != 0);
       if (__pyx_t_2) {
 
-        /* "View.MemoryView":1294
+        /* "View.MemoryView":1296
  *         if src.shape[i] != dst.shape[i]:
  *             if src.shape[i] == 1:
  *                 broadcasting = True             # <<<<<<<<<<<<<<
@@ -20396,7 +20459,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
         __pyx_v_broadcasting = 1;
 
-        /* "View.MemoryView":1295
+        /* "View.MemoryView":1297
  *             if src.shape[i] == 1:
  *                 broadcasting = True
  *                 src.strides[i] = 0             # <<<<<<<<<<<<<<
@@ -20405,7 +20468,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
         (__pyx_v_src.strides[__pyx_v_i]) = 0;
 
-        /* "View.MemoryView":1293
+        /* "View.MemoryView":1295
  *     for i in range(ndim):
  *         if src.shape[i] != dst.shape[i]:
  *             if src.shape[i] == 1:             # <<<<<<<<<<<<<<
@@ -20415,7 +20478,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
         goto __pyx_L7;
       }
 
-      /* "View.MemoryView":1297
+      /* "View.MemoryView":1299
  *                 src.strides[i] = 0
  *             else:
  *                 _err_extents(i, dst.shape[i], src.shape[i])             # <<<<<<<<<<<<<<
@@ -20423,11 +20486,11 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  *         if src.suboffsets[i] >= 0:
  */
       /*else*/ {
-        __pyx_t_6 = __pyx_memoryview_err_extents(__pyx_v_i, (__pyx_v_dst.shape[__pyx_v_i]), (__pyx_v_src.shape[__pyx_v_i])); if (unlikely(__pyx_t_6 == ((int)-1))) __PYX_ERR(1, 1297, __pyx_L1_error)
+        __pyx_t_6 = __pyx_memoryview_err_extents(__pyx_v_i, (__pyx_v_dst.shape[__pyx_v_i]), (__pyx_v_src.shape[__pyx_v_i])); if (unlikely(__pyx_t_6 == ((int)-1))) __PYX_ERR(1, 1299, __pyx_L1_error)
       }
       __pyx_L7:;
 
-      /* "View.MemoryView":1292
+      /* "View.MemoryView":1294
  * 
  *     for i in range(ndim):
  *         if src.shape[i] != dst.shape[i]:             # <<<<<<<<<<<<<<
@@ -20436,7 +20499,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
     }
 
-    /* "View.MemoryView":1299
+    /* "View.MemoryView":1301
  *                 _err_extents(i, dst.shape[i], src.shape[i])
  * 
  *         if src.suboffsets[i] >= 0:             # <<<<<<<<<<<<<<
@@ -20446,16 +20509,16 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
     __pyx_t_2 = (((__pyx_v_src.suboffsets[__pyx_v_i]) >= 0) != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":1300
+      /* "View.MemoryView":1302
  * 
  *         if src.suboffsets[i] >= 0:
  *             _err_dim(ValueError, "Dimension %d is not direct", i)             # <<<<<<<<<<<<<<
  * 
  *     if slices_overlap(&src, &dst, ndim, itemsize):
  */
-      __pyx_t_6 = __pyx_memoryview_err_dim(__pyx_builtin_ValueError, ((char *)"Dimension %d is not direct"), __pyx_v_i); if (unlikely(__pyx_t_6 == ((int)-1))) __PYX_ERR(1, 1300, __pyx_L1_error)
+      __pyx_t_6 = __pyx_memoryview_err_dim(__pyx_builtin_ValueError, ((char *)"Dimension %d is not direct"), __pyx_v_i); if (unlikely(__pyx_t_6 == ((int)-1))) __PYX_ERR(1, 1302, __pyx_L1_error)
 
-      /* "View.MemoryView":1299
+      /* "View.MemoryView":1301
  *                 _err_extents(i, dst.shape[i], src.shape[i])
  * 
  *         if src.suboffsets[i] >= 0:             # <<<<<<<<<<<<<<
@@ -20465,7 +20528,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
     }
   }
 
-  /* "View.MemoryView":1302
+  /* "View.MemoryView":1304
  *             _err_dim(ValueError, "Dimension %d is not direct", i)
  * 
  *     if slices_overlap(&src, &dst, ndim, itemsize):             # <<<<<<<<<<<<<<
@@ -20475,7 +20538,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
   __pyx_t_2 = (__pyx_slices_overlap((&__pyx_v_src), (&__pyx_v_dst), __pyx_v_ndim, __pyx_v_itemsize) != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":1304
+    /* "View.MemoryView":1306
  *     if slices_overlap(&src, &dst, ndim, itemsize):
  * 
  *         if not slice_is_contig(src, order, ndim):             # <<<<<<<<<<<<<<
@@ -20485,7 +20548,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
     __pyx_t_2 = ((!(__pyx_memviewslice_is_contig(__pyx_v_src, __pyx_v_order, __pyx_v_ndim) != 0)) != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":1305
+      /* "View.MemoryView":1307
  * 
  *         if not slice_is_contig(src, order, ndim):
  *             order = get_best_order(&dst, ndim)             # <<<<<<<<<<<<<<
@@ -20494,7 +20557,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
       __pyx_v_order = __pyx_get_best_slice_order((&__pyx_v_dst), __pyx_v_ndim);
 
-      /* "View.MemoryView":1304
+      /* "View.MemoryView":1306
  *     if slices_overlap(&src, &dst, ndim, itemsize):
  * 
  *         if not slice_is_contig(src, order, ndim):             # <<<<<<<<<<<<<<
@@ -20503,17 +20566,17 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
     }
 
-    /* "View.MemoryView":1307
+    /* "View.MemoryView":1309
  *             order = get_best_order(&dst, ndim)
  * 
  *         tmpdata = copy_data_to_temp(&src, &tmp, order, ndim)             # <<<<<<<<<<<<<<
  *         src = tmp
  * 
  */
-    __pyx_t_7 = __pyx_memoryview_copy_data_to_temp((&__pyx_v_src), (&__pyx_v_tmp), __pyx_v_order, __pyx_v_ndim); if (unlikely(__pyx_t_7 == ((void *)NULL))) __PYX_ERR(1, 1307, __pyx_L1_error)
+    __pyx_t_7 = __pyx_memoryview_copy_data_to_temp((&__pyx_v_src), (&__pyx_v_tmp), __pyx_v_order, __pyx_v_ndim); if (unlikely(__pyx_t_7 == ((void *)NULL))) __PYX_ERR(1, 1309, __pyx_L1_error)
     __pyx_v_tmpdata = __pyx_t_7;
 
-    /* "View.MemoryView":1308
+    /* "View.MemoryView":1310
  * 
  *         tmpdata = copy_data_to_temp(&src, &tmp, order, ndim)
  *         src = tmp             # <<<<<<<<<<<<<<
@@ -20522,7 +20585,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
     __pyx_v_src = __pyx_v_tmp;
 
-    /* "View.MemoryView":1302
+    /* "View.MemoryView":1304
  *             _err_dim(ValueError, "Dimension %d is not direct", i)
  * 
  *     if slices_overlap(&src, &dst, ndim, itemsize):             # <<<<<<<<<<<<<<
@@ -20531,7 +20594,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
   }
 
-  /* "View.MemoryView":1310
+  /* "View.MemoryView":1312
  *         src = tmp
  * 
  *     if not broadcasting:             # <<<<<<<<<<<<<<
@@ -20541,7 +20604,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
   __pyx_t_2 = ((!(__pyx_v_broadcasting != 0)) != 0);
   if (__pyx_t_2) {
 
-    /* "View.MemoryView":1313
+    /* "View.MemoryView":1315
  * 
  * 
  *         if slice_is_contig(src, 'C', ndim):             # <<<<<<<<<<<<<<
@@ -20551,7 +20614,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
     __pyx_t_2 = (__pyx_memviewslice_is_contig(__pyx_v_src, 'C', __pyx_v_ndim) != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":1314
+      /* "View.MemoryView":1316
  * 
  *         if slice_is_contig(src, 'C', ndim):
  *             direct_copy = slice_is_contig(dst, 'C', ndim)             # <<<<<<<<<<<<<<
@@ -20560,7 +20623,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
       __pyx_v_direct_copy = __pyx_memviewslice_is_contig(__pyx_v_dst, 'C', __pyx_v_ndim);
 
-      /* "View.MemoryView":1313
+      /* "View.MemoryView":1315
  * 
  * 
  *         if slice_is_contig(src, 'C', ndim):             # <<<<<<<<<<<<<<
@@ -20570,7 +20633,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
       goto __pyx_L12;
     }
 
-    /* "View.MemoryView":1315
+    /* "View.MemoryView":1317
  *         if slice_is_contig(src, 'C', ndim):
  *             direct_copy = slice_is_contig(dst, 'C', ndim)
  *         elif slice_is_contig(src, 'F', ndim):             # <<<<<<<<<<<<<<
@@ -20580,7 +20643,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
     __pyx_t_2 = (__pyx_memviewslice_is_contig(__pyx_v_src, 'F', __pyx_v_ndim) != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":1316
+      /* "View.MemoryView":1318
  *             direct_copy = slice_is_contig(dst, 'C', ndim)
  *         elif slice_is_contig(src, 'F', ndim):
  *             direct_copy = slice_is_contig(dst, 'F', ndim)             # <<<<<<<<<<<<<<
@@ -20589,7 +20652,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
       __pyx_v_direct_copy = __pyx_memviewslice_is_contig(__pyx_v_dst, 'F', __pyx_v_ndim);
 
-      /* "View.MemoryView":1315
+      /* "View.MemoryView":1317
  *         if slice_is_contig(src, 'C', ndim):
  *             direct_copy = slice_is_contig(dst, 'C', ndim)
  *         elif slice_is_contig(src, 'F', ndim):             # <<<<<<<<<<<<<<
@@ -20599,7 +20662,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
     }
     __pyx_L12:;
 
-    /* "View.MemoryView":1318
+    /* "View.MemoryView":1320
  *             direct_copy = slice_is_contig(dst, 'F', ndim)
  * 
  *         if direct_copy:             # <<<<<<<<<<<<<<
@@ -20609,7 +20672,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
     __pyx_t_2 = (__pyx_v_direct_copy != 0);
     if (__pyx_t_2) {
 
-      /* "View.MemoryView":1320
+      /* "View.MemoryView":1322
  *         if direct_copy:
  * 
  *             refcount_copying(&dst, dtype_is_object, ndim, False)             # <<<<<<<<<<<<<<
@@ -20618,7 +20681,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
       __pyx_memoryview_refcount_copying((&__pyx_v_dst), __pyx_v_dtype_is_object, __pyx_v_ndim, 0);
 
-      /* "View.MemoryView":1321
+      /* "View.MemoryView":1323
  * 
  *             refcount_copying(&dst, dtype_is_object, ndim, False)
  *             memcpy(dst.data, src.data, slice_get_size(&src, ndim))             # <<<<<<<<<<<<<<
@@ -20627,7 +20690,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
       (void)(memcpy(__pyx_v_dst.data, __pyx_v_src.data, __pyx_memoryview_slice_get_size((&__pyx_v_src), __pyx_v_ndim)));
 
-      /* "View.MemoryView":1322
+      /* "View.MemoryView":1324
  *             refcount_copying(&dst, dtype_is_object, ndim, False)
  *             memcpy(dst.data, src.data, slice_get_size(&src, ndim))
  *             refcount_copying(&dst, dtype_is_object, ndim, True)             # <<<<<<<<<<<<<<
@@ -20636,7 +20699,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
       __pyx_memoryview_refcount_copying((&__pyx_v_dst), __pyx_v_dtype_is_object, __pyx_v_ndim, 1);
 
-      /* "View.MemoryView":1323
+      /* "View.MemoryView":1325
  *             memcpy(dst.data, src.data, slice_get_size(&src, ndim))
  *             refcount_copying(&dst, dtype_is_object, ndim, True)
  *             free(tmpdata)             # <<<<<<<<<<<<<<
@@ -20645,7 +20708,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
       free(__pyx_v_tmpdata);
 
-      /* "View.MemoryView":1324
+      /* "View.MemoryView":1326
  *             refcount_copying(&dst, dtype_is_object, ndim, True)
  *             free(tmpdata)
  *             return 0             # <<<<<<<<<<<<<<
@@ -20655,7 +20718,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
       __pyx_r = 0;
       goto __pyx_L0;
 
-      /* "View.MemoryView":1318
+      /* "View.MemoryView":1320
  *             direct_copy = slice_is_contig(dst, 'F', ndim)
  * 
  *         if direct_copy:             # <<<<<<<<<<<<<<
@@ -20664,7 +20727,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
     }
 
-    /* "View.MemoryView":1310
+    /* "View.MemoryView":1312
  *         src = tmp
  * 
  *     if not broadcasting:             # <<<<<<<<<<<<<<
@@ -20673,7 +20736,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
   }
 
-  /* "View.MemoryView":1326
+  /* "View.MemoryView":1328
  *             return 0
  * 
  *     if order == 'F' == get_best_order(&dst, ndim):             # <<<<<<<<<<<<<<
@@ -20687,25 +20750,25 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
   __pyx_t_8 = (__pyx_t_2 != 0);
   if (__pyx_t_8) {
 
-    /* "View.MemoryView":1329
+    /* "View.MemoryView":1331
  * 
  * 
  *         transpose_memslice(&src)             # <<<<<<<<<<<<<<
  *         transpose_memslice(&dst)
  * 
  */
-    __pyx_t_5 = __pyx_memslice_transpose((&__pyx_v_src)); if (unlikely(__pyx_t_5 == ((int)0))) __PYX_ERR(1, 1329, __pyx_L1_error)
+    __pyx_t_5 = __pyx_memslice_transpose((&__pyx_v_src)); if (unlikely(__pyx_t_5 == ((int)0))) __PYX_ERR(1, 1331, __pyx_L1_error)
 
-    /* "View.MemoryView":1330
+    /* "View.MemoryView":1332
  * 
  *         transpose_memslice(&src)
  *         transpose_memslice(&dst)             # <<<<<<<<<<<<<<
  * 
  *     refcount_copying(&dst, dtype_is_object, ndim, False)
  */
-    __pyx_t_5 = __pyx_memslice_transpose((&__pyx_v_dst)); if (unlikely(__pyx_t_5 == ((int)0))) __PYX_ERR(1, 1330, __pyx_L1_error)
+    __pyx_t_5 = __pyx_memslice_transpose((&__pyx_v_dst)); if (unlikely(__pyx_t_5 == ((int)0))) __PYX_ERR(1, 1332, __pyx_L1_error)
 
-    /* "View.MemoryView":1326
+    /* "View.MemoryView":1328
  *             return 0
  * 
  *     if order == 'F' == get_best_order(&dst, ndim):             # <<<<<<<<<<<<<<
@@ -20714,7 +20777,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
   }
 
-  /* "View.MemoryView":1332
+  /* "View.MemoryView":1334
  *         transpose_memslice(&dst)
  * 
  *     refcount_copying(&dst, dtype_is_object, ndim, False)             # <<<<<<<<<<<<<<
@@ -20723,7 +20786,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
   __pyx_memoryview_refcount_copying((&__pyx_v_dst), __pyx_v_dtype_is_object, __pyx_v_ndim, 0);
 
-  /* "View.MemoryView":1333
+  /* "View.MemoryView":1335
  * 
  *     refcount_copying(&dst, dtype_is_object, ndim, False)
  *     copy_strided_to_strided(&src, &dst, ndim, itemsize)             # <<<<<<<<<<<<<<
@@ -20732,7 +20795,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
   copy_strided_to_strided((&__pyx_v_src), (&__pyx_v_dst), __pyx_v_ndim, __pyx_v_itemsize);
 
-  /* "View.MemoryView":1334
+  /* "View.MemoryView":1336
  *     refcount_copying(&dst, dtype_is_object, ndim, False)
  *     copy_strided_to_strided(&src, &dst, ndim, itemsize)
  *     refcount_copying(&dst, dtype_is_object, ndim, True)             # <<<<<<<<<<<<<<
@@ -20741,7 +20804,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
   __pyx_memoryview_refcount_copying((&__pyx_v_dst), __pyx_v_dtype_is_object, __pyx_v_ndim, 1);
 
-  /* "View.MemoryView":1336
+  /* "View.MemoryView":1338
  *     refcount_copying(&dst, dtype_is_object, ndim, True)
  * 
  *     free(tmpdata)             # <<<<<<<<<<<<<<
@@ -20750,7 +20813,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
  */
   free(__pyx_v_tmpdata);
 
-  /* "View.MemoryView":1337
+  /* "View.MemoryView":1339
  * 
  *     free(tmpdata)
  *     return 0             # <<<<<<<<<<<<<<
@@ -20760,7 +20823,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
   __pyx_r = 0;
   goto __pyx_L0;
 
-  /* "View.MemoryView":1268
+  /* "View.MemoryView":1270
  * 
  * @cname('__pyx_memoryview_copy_contents')
  * cdef int memoryview_copy_contents(__Pyx_memviewslice src,             # <<<<<<<<<<<<<<
@@ -20784,7 +20847,7 @@ static int __pyx_memoryview_copy_contents(__Pyx_memviewslice __pyx_v_src, __Pyx_
   return __pyx_r;
 }
 
-/* "View.MemoryView":1340
+/* "View.MemoryView":1342
  * 
  * @cname('__pyx_memoryview_broadcast_leading')
  * cdef void broadcast_leading(__Pyx_memviewslice *mslice,             # <<<<<<<<<<<<<<
@@ -20799,7 +20862,7 @@ static void __pyx_memoryview_broadcast_leading(__Pyx_memviewslice *__pyx_v_mslic
   int __pyx_t_2;
   int __pyx_t_3;
 
-  /* "View.MemoryView":1344
+  /* "View.MemoryView":1346
  *                             int ndim_other) nogil:
  *     cdef int i
  *     cdef int offset = ndim_other - ndim             # <<<<<<<<<<<<<<
@@ -20808,7 +20871,7 @@ static void __pyx_memoryview_broadcast_leading(__Pyx_memviewslice *__pyx_v_mslic
  */
   __pyx_v_offset = (__pyx_v_ndim_other - __pyx_v_ndim);
 
-  /* "View.MemoryView":1346
+  /* "View.MemoryView":1348
  *     cdef int offset = ndim_other - ndim
  * 
  *     for i in range(ndim - 1, -1, -1):             # <<<<<<<<<<<<<<
@@ -20818,7 +20881,7 @@ static void __pyx_memoryview_broadcast_leading(__Pyx_memviewslice *__pyx_v_mslic
   for (__pyx_t_1 = (__pyx_v_ndim - 1); __pyx_t_1 > -1; __pyx_t_1-=1) {
     __pyx_v_i = __pyx_t_1;
 
-    /* "View.MemoryView":1347
+    /* "View.MemoryView":1349
  * 
  *     for i in range(ndim - 1, -1, -1):
  *         mslice.shape[i + offset] = mslice.shape[i]             # <<<<<<<<<<<<<<
@@ -20827,7 +20890,7 @@ static void __pyx_memoryview_broadcast_leading(__Pyx_memviewslice *__pyx_v_mslic
  */
     (__pyx_v_mslice->shape[(__pyx_v_i + __pyx_v_offset)]) = (__pyx_v_mslice->shape[__pyx_v_i]);
 
-    /* "View.MemoryView":1348
+    /* "View.MemoryView":1350
  *     for i in range(ndim - 1, -1, -1):
  *         mslice.shape[i + offset] = mslice.shape[i]
  *         mslice.strides[i + offset] = mslice.strides[i]             # <<<<<<<<<<<<<<
@@ -20836,7 +20899,7 @@ static void __pyx_memoryview_broadcast_leading(__Pyx_memviewslice *__pyx_v_mslic
  */
     (__pyx_v_mslice->strides[(__pyx_v_i + __pyx_v_offset)]) = (__pyx_v_mslice->strides[__pyx_v_i]);
 
-    /* "View.MemoryView":1349
+    /* "View.MemoryView":1351
  *         mslice.shape[i + offset] = mslice.shape[i]
  *         mslice.strides[i + offset] = mslice.strides[i]
  *         mslice.suboffsets[i + offset] = mslice.suboffsets[i]             # <<<<<<<<<<<<<<
@@ -20846,7 +20909,7 @@ static void __pyx_memoryview_broadcast_leading(__Pyx_memviewslice *__pyx_v_mslic
     (__pyx_v_mslice->suboffsets[(__pyx_v_i + __pyx_v_offset)]) = (__pyx_v_mslice->suboffsets[__pyx_v_i]);
   }
 
-  /* "View.MemoryView":1351
+  /* "View.MemoryView":1353
  *         mslice.suboffsets[i + offset] = mslice.suboffsets[i]
  * 
  *     for i in range(offset):             # <<<<<<<<<<<<<<
@@ -20858,7 +20921,7 @@ static void __pyx_memoryview_broadcast_leading(__Pyx_memviewslice *__pyx_v_mslic
   for (__pyx_t_3 = 0; __pyx_t_3 < __pyx_t_2; __pyx_t_3+=1) {
     __pyx_v_i = __pyx_t_3;
 
-    /* "View.MemoryView":1352
+    /* "View.MemoryView":1354
  * 
  *     for i in range(offset):
  *         mslice.shape[i] = 1             # <<<<<<<<<<<<<<
@@ -20867,7 +20930,7 @@ static void __pyx_memoryview_broadcast_leading(__Pyx_memviewslice *__pyx_v_mslic
  */
     (__pyx_v_mslice->shape[__pyx_v_i]) = 1;
 
-    /* "View.MemoryView":1353
+    /* "View.MemoryView":1355
  *     for i in range(offset):
  *         mslice.shape[i] = 1
  *         mslice.strides[i] = mslice.strides[0]             # <<<<<<<<<<<<<<
@@ -20876,7 +20939,7 @@ static void __pyx_memoryview_broadcast_leading(__Pyx_memviewslice *__pyx_v_mslic
  */
     (__pyx_v_mslice->strides[__pyx_v_i]) = (__pyx_v_mslice->strides[0]);
 
-    /* "View.MemoryView":1354
+    /* "View.MemoryView":1356
  *         mslice.shape[i] = 1
  *         mslice.strides[i] = mslice.strides[0]
  *         mslice.suboffsets[i] = -1             # <<<<<<<<<<<<<<
@@ -20886,7 +20949,7 @@ static void __pyx_memoryview_broadcast_leading(__Pyx_memviewslice *__pyx_v_mslic
     (__pyx_v_mslice->suboffsets[__pyx_v_i]) = -1L;
   }
 
-  /* "View.MemoryView":1340
+  /* "View.MemoryView":1342
  * 
  * @cname('__pyx_memoryview_broadcast_leading')
  * cdef void broadcast_leading(__Pyx_memviewslice *mslice,             # <<<<<<<<<<<<<<
@@ -20897,7 +20960,7 @@ static void __pyx_memoryview_broadcast_leading(__Pyx_memviewslice *__pyx_v_mslic
   /* function exit code */
 }
 
-/* "View.MemoryView":1362
+/* "View.MemoryView":1364
  * 
  * @cname('__pyx_memoryview_refcount_copying')
  * cdef void refcount_copying(__Pyx_memviewslice *dst, bint dtype_is_object,             # <<<<<<<<<<<<<<
@@ -20908,7 +20971,7 @@ static void __pyx_memoryview_broadcast_leading(__Pyx_memviewslice *__pyx_v_mslic
 static void __pyx_memoryview_refcount_copying(__Pyx_memviewslice *__pyx_v_dst, int __pyx_v_dtype_is_object, int __pyx_v_ndim, int __pyx_v_inc) {
   int __pyx_t_1;
 
-  /* "View.MemoryView":1366
+  /* "View.MemoryView":1368
  * 
  * 
  *     if dtype_is_object:             # <<<<<<<<<<<<<<
@@ -20918,7 +20981,7 @@ static void __pyx_memoryview_refcount_copying(__Pyx_memviewslice *__pyx_v_dst, i
   __pyx_t_1 = (__pyx_v_dtype_is_object != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":1367
+    /* "View.MemoryView":1369
  * 
  *     if dtype_is_object:
  *         refcount_objects_in_slice_with_gil(dst.data, dst.shape,             # <<<<<<<<<<<<<<
@@ -20927,7 +20990,7 @@ static void __pyx_memoryview_refcount_copying(__Pyx_memviewslice *__pyx_v_dst, i
  */
     __pyx_memoryview_refcount_objects_in_slice_with_gil(__pyx_v_dst->data, __pyx_v_dst->shape, __pyx_v_dst->strides, __pyx_v_ndim, __pyx_v_inc);
 
-    /* "View.MemoryView":1366
+    /* "View.MemoryView":1368
  * 
  * 
  *     if dtype_is_object:             # <<<<<<<<<<<<<<
@@ -20936,7 +20999,7 @@ static void __pyx_memoryview_refcount_copying(__Pyx_memviewslice *__pyx_v_dst, i
  */
   }
 
-  /* "View.MemoryView":1362
+  /* "View.MemoryView":1364
  * 
  * @cname('__pyx_memoryview_refcount_copying')
  * cdef void refcount_copying(__Pyx_memviewslice *dst, bint dtype_is_object,             # <<<<<<<<<<<<<<
@@ -20947,7 +21010,7 @@ static void __pyx_memoryview_refcount_copying(__Pyx_memviewslice *__pyx_v_dst, i
   /* function exit code */
 }
 
-/* "View.MemoryView":1371
+/* "View.MemoryView":1373
  * 
  * @cname('__pyx_memoryview_refcount_objects_in_slice_with_gil')
  * cdef void refcount_objects_in_slice_with_gil(char *data, Py_ssize_t *shape,             # <<<<<<<<<<<<<<
@@ -20962,7 +21025,7 @@ static void __pyx_memoryview_refcount_objects_in_slice_with_gil(char *__pyx_v_da
   #endif
   __Pyx_RefNannySetupContext("refcount_objects_in_slice_with_gil", 0);
 
-  /* "View.MemoryView":1374
+  /* "View.MemoryView":1376
  *                                              Py_ssize_t *strides, int ndim,
  *                                              bint inc) with gil:
  *     refcount_objects_in_slice(data, shape, strides, ndim, inc)             # <<<<<<<<<<<<<<
@@ -20971,7 +21034,7 @@ static void __pyx_memoryview_refcount_objects_in_slice_with_gil(char *__pyx_v_da
  */
   __pyx_memoryview_refcount_objects_in_slice(__pyx_v_data, __pyx_v_shape, __pyx_v_strides, __pyx_v_ndim, __pyx_v_inc);
 
-  /* "View.MemoryView":1371
+  /* "View.MemoryView":1373
  * 
  * @cname('__pyx_memoryview_refcount_objects_in_slice_with_gil')
  * cdef void refcount_objects_in_slice_with_gil(char *data, Py_ssize_t *shape,             # <<<<<<<<<<<<<<
@@ -20986,7 +21049,7 @@ static void __pyx_memoryview_refcount_objects_in_slice_with_gil(char *__pyx_v_da
   #endif
 }
 
-/* "View.MemoryView":1377
+/* "View.MemoryView":1379
  * 
  * @cname('__pyx_memoryview_refcount_objects_in_slice')
  * cdef void refcount_objects_in_slice(char *data, Py_ssize_t *shape,             # <<<<<<<<<<<<<<
@@ -21003,7 +21066,7 @@ static void __pyx_memoryview_refcount_objects_in_slice(char *__pyx_v_data, Py_ss
   int __pyx_t_4;
   __Pyx_RefNannySetupContext("refcount_objects_in_slice", 0);
 
-  /* "View.MemoryView":1381
+  /* "View.MemoryView":1383
  *     cdef Py_ssize_t i
  * 
  *     for i in range(shape[0]):             # <<<<<<<<<<<<<<
@@ -21015,7 +21078,7 @@ static void __pyx_memoryview_refcount_objects_in_slice(char *__pyx_v_data, Py_ss
   for (__pyx_t_3 = 0; __pyx_t_3 < __pyx_t_2; __pyx_t_3+=1) {
     __pyx_v_i = __pyx_t_3;
 
-    /* "View.MemoryView":1382
+    /* "View.MemoryView":1384
  * 
  *     for i in range(shape[0]):
  *         if ndim == 1:             # <<<<<<<<<<<<<<
@@ -21025,7 +21088,7 @@ static void __pyx_memoryview_refcount_objects_in_slice(char *__pyx_v_data, Py_ss
     __pyx_t_4 = ((__pyx_v_ndim == 1) != 0);
     if (__pyx_t_4) {
 
-      /* "View.MemoryView":1383
+      /* "View.MemoryView":1385
  *     for i in range(shape[0]):
  *         if ndim == 1:
  *             if inc:             # <<<<<<<<<<<<<<
@@ -21035,7 +21098,7 @@ static void __pyx_memoryview_refcount_objects_in_slice(char *__pyx_v_data, Py_ss
       __pyx_t_4 = (__pyx_v_inc != 0);
       if (__pyx_t_4) {
 
-        /* "View.MemoryView":1384
+        /* "View.MemoryView":1386
  *         if ndim == 1:
  *             if inc:
  *                 Py_INCREF((<PyObject **> data)[0])             # <<<<<<<<<<<<<<
@@ -21044,7 +21107,7 @@ static void __pyx_memoryview_refcount_objects_in_slice(char *__pyx_v_data, Py_ss
  */
         Py_INCREF((((PyObject **)__pyx_v_data)[0]));
 
-        /* "View.MemoryView":1383
+        /* "View.MemoryView":1385
  *     for i in range(shape[0]):
  *         if ndim == 1:
  *             if inc:             # <<<<<<<<<<<<<<
@@ -21054,7 +21117,7 @@ static void __pyx_memoryview_refcount_objects_in_slice(char *__pyx_v_data, Py_ss
         goto __pyx_L6;
       }
 
-      /* "View.MemoryView":1386
+      /* "View.MemoryView":1388
  *                 Py_INCREF((<PyObject **> data)[0])
  *             else:
  *                 Py_DECREF((<PyObject **> data)[0])             # <<<<<<<<<<<<<<
@@ -21066,7 +21129,7 @@ static void __pyx_memoryview_refcount_objects_in_slice(char *__pyx_v_data, Py_ss
       }
       __pyx_L6:;
 
-      /* "View.MemoryView":1382
+      /* "View.MemoryView":1384
  * 
  *     for i in range(shape[0]):
  *         if ndim == 1:             # <<<<<<<<<<<<<<
@@ -21076,7 +21139,7 @@ static void __pyx_memoryview_refcount_objects_in_slice(char *__pyx_v_data, Py_ss
       goto __pyx_L5;
     }
 
-    /* "View.MemoryView":1388
+    /* "View.MemoryView":1390
  *                 Py_DECREF((<PyObject **> data)[0])
  *         else:
  *             refcount_objects_in_slice(data, shape + 1, strides + 1,             # <<<<<<<<<<<<<<
@@ -21085,7 +21148,7 @@ static void __pyx_memoryview_refcount_objects_in_slice(char *__pyx_v_data, Py_ss
  */
     /*else*/ {
 
-      /* "View.MemoryView":1389
+      /* "View.MemoryView":1391
  *         else:
  *             refcount_objects_in_slice(data, shape + 1, strides + 1,
  *                                       ndim - 1, inc)             # <<<<<<<<<<<<<<
@@ -21096,7 +21159,7 @@ static void __pyx_memoryview_refcount_objects_in_slice(char *__pyx_v_data, Py_ss
     }
     __pyx_L5:;
 
-    /* "View.MemoryView":1391
+    /* "View.MemoryView":1393
  *                                       ndim - 1, inc)
  * 
  *         data += strides[0]             # <<<<<<<<<<<<<<
@@ -21106,7 +21169,7 @@ static void __pyx_memoryview_refcount_objects_in_slice(char *__pyx_v_data, Py_ss
     __pyx_v_data = (__pyx_v_data + (__pyx_v_strides[0]));
   }
 
-  /* "View.MemoryView":1377
+  /* "View.MemoryView":1379
  * 
  * @cname('__pyx_memoryview_refcount_objects_in_slice')
  * cdef void refcount_objects_in_slice(char *data, Py_ssize_t *shape,             # <<<<<<<<<<<<<<
@@ -21118,7 +21181,7 @@ static void __pyx_memoryview_refcount_objects_in_slice(char *__pyx_v_data, Py_ss
   __Pyx_RefNannyFinishContext();
 }
 
-/* "View.MemoryView":1397
+/* "View.MemoryView":1399
  * 
  * @cname('__pyx_memoryview_slice_assign_scalar')
  * cdef void slice_assign_scalar(__Pyx_memviewslice *dst, int ndim,             # <<<<<<<<<<<<<<
@@ -21128,7 +21191,7 @@ static void __pyx_memoryview_refcount_objects_in_slice(char *__pyx_v_data, Py_ss
 
 static void __pyx_memoryview_slice_assign_scalar(__Pyx_memviewslice *__pyx_v_dst, int __pyx_v_ndim, size_t __pyx_v_itemsize, void *__pyx_v_item, int __pyx_v_dtype_is_object) {
 
-  /* "View.MemoryView":1400
+  /* "View.MemoryView":1402
  *                               size_t itemsize, void *item,
  *                               bint dtype_is_object) nogil:
  *     refcount_copying(dst, dtype_is_object, ndim, False)             # <<<<<<<<<<<<<<
@@ -21137,7 +21200,7 @@ static void __pyx_memoryview_slice_assign_scalar(__Pyx_memviewslice *__pyx_v_dst
  */
   __pyx_memoryview_refcount_copying(__pyx_v_dst, __pyx_v_dtype_is_object, __pyx_v_ndim, 0);
 
-  /* "View.MemoryView":1401
+  /* "View.MemoryView":1403
  *                               bint dtype_is_object) nogil:
  *     refcount_copying(dst, dtype_is_object, ndim, False)
  *     _slice_assign_scalar(dst.data, dst.shape, dst.strides, ndim,             # <<<<<<<<<<<<<<
@@ -21146,7 +21209,7 @@ static void __pyx_memoryview_slice_assign_scalar(__Pyx_memviewslice *__pyx_v_dst
  */
   __pyx_memoryview__slice_assign_scalar(__pyx_v_dst->data, __pyx_v_dst->shape, __pyx_v_dst->strides, __pyx_v_ndim, __pyx_v_itemsize, __pyx_v_item);
 
-  /* "View.MemoryView":1403
+  /* "View.MemoryView":1405
  *     _slice_assign_scalar(dst.data, dst.shape, dst.strides, ndim,
  *                          itemsize, item)
  *     refcount_copying(dst, dtype_is_object, ndim, True)             # <<<<<<<<<<<<<<
@@ -21155,7 +21218,7 @@ static void __pyx_memoryview_slice_assign_scalar(__Pyx_memviewslice *__pyx_v_dst
  */
   __pyx_memoryview_refcount_copying(__pyx_v_dst, __pyx_v_dtype_is_object, __pyx_v_ndim, 1);
 
-  /* "View.MemoryView":1397
+  /* "View.MemoryView":1399
  * 
  * @cname('__pyx_memoryview_slice_assign_scalar')
  * cdef void slice_assign_scalar(__Pyx_memviewslice *dst, int ndim,             # <<<<<<<<<<<<<<
@@ -21166,7 +21229,7 @@ static void __pyx_memoryview_slice_assign_scalar(__Pyx_memviewslice *__pyx_v_dst
   /* function exit code */
 }
 
-/* "View.MemoryView":1407
+/* "View.MemoryView":1409
  * 
  * @cname('__pyx_memoryview__slice_assign_scalar')
  * cdef void _slice_assign_scalar(char *data, Py_ssize_t *shape,             # <<<<<<<<<<<<<<
@@ -21183,7 +21246,7 @@ static void __pyx_memoryview__slice_assign_scalar(char *__pyx_v_data, Py_ssize_t
   Py_ssize_t __pyx_t_3;
   Py_ssize_t __pyx_t_4;
 
-  /* "View.MemoryView":1411
+  /* "View.MemoryView":1413
  *                               size_t itemsize, void *item) nogil:
  *     cdef Py_ssize_t i
  *     cdef Py_ssize_t stride = strides[0]             # <<<<<<<<<<<<<<
@@ -21192,7 +21255,7 @@ static void __pyx_memoryview__slice_assign_scalar(char *__pyx_v_data, Py_ssize_t
  */
   __pyx_v_stride = (__pyx_v_strides[0]);
 
-  /* "View.MemoryView":1412
+  /* "View.MemoryView":1414
  *     cdef Py_ssize_t i
  *     cdef Py_ssize_t stride = strides[0]
  *     cdef Py_ssize_t extent = shape[0]             # <<<<<<<<<<<<<<
@@ -21201,7 +21264,7 @@ static void __pyx_memoryview__slice_assign_scalar(char *__pyx_v_data, Py_ssize_t
  */
   __pyx_v_extent = (__pyx_v_shape[0]);
 
-  /* "View.MemoryView":1414
+  /* "View.MemoryView":1416
  *     cdef Py_ssize_t extent = shape[0]
  * 
  *     if ndim == 1:             # <<<<<<<<<<<<<<
@@ -21211,7 +21274,7 @@ static void __pyx_memoryview__slice_assign_scalar(char *__pyx_v_data, Py_ssize_t
   __pyx_t_1 = ((__pyx_v_ndim == 1) != 0);
   if (__pyx_t_1) {
 
-    /* "View.MemoryView":1415
+    /* "View.MemoryView":1417
  * 
  *     if ndim == 1:
  *         for i in range(extent):             # <<<<<<<<<<<<<<
@@ -21223,7 +21286,7 @@ static void __pyx_memoryview__slice_assign_scalar(char *__pyx_v_data, Py_ssize_t
     for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) {
       __pyx_v_i = __pyx_t_4;
 
-      /* "View.MemoryView":1416
+      /* "View.MemoryView":1418
  *     if ndim == 1:
  *         for i in range(extent):
  *             memcpy(data, item, itemsize)             # <<<<<<<<<<<<<<
@@ -21232,7 +21295,7 @@ static void __pyx_memoryview__slice_assign_scalar(char *__pyx_v_data, Py_ssize_t
  */
       (void)(memcpy(__pyx_v_data, __pyx_v_item, __pyx_v_itemsize));
 
-      /* "View.MemoryView":1417
+      /* "View.MemoryView":1419
  *         for i in range(extent):
  *             memcpy(data, item, itemsize)
  *             data += stride             # <<<<<<<<<<<<<<
@@ -21242,7 +21305,7 @@ static void __pyx_memoryview__slice_assign_scalar(char *__pyx_v_data, Py_ssize_t
       __pyx_v_data = (__pyx_v_data + __pyx_v_stride);
     }
 
-    /* "View.MemoryView":1414
+    /* "View.MemoryView":1416
  *     cdef Py_ssize_t extent = shape[0]
  * 
  *     if ndim == 1:             # <<<<<<<<<<<<<<
@@ -21252,7 +21315,7 @@ static void __pyx_memoryview__slice_assign_scalar(char *__pyx_v_data, Py_ssize_t
     goto __pyx_L3;
   }
 
-  /* "View.MemoryView":1419
+  /* "View.MemoryView":1421
  *             data += stride
  *     else:
  *         for i in range(extent):             # <<<<<<<<<<<<<<
@@ -21265,7 +21328,7 @@ static void __pyx_memoryview__slice_assign_scalar(char *__pyx_v_data, Py_ssize_t
     for (__pyx_t_4 = 0; __pyx_t_4 < __pyx_t_3; __pyx_t_4+=1) {
       __pyx_v_i = __pyx_t_4;
 
-      /* "View.MemoryView":1420
+      /* "View.MemoryView":1422
  *     else:
  *         for i in range(extent):
  *             _slice_assign_scalar(data, shape + 1, strides + 1,             # <<<<<<<<<<<<<<
@@ -21274,7 +21337,7 @@ static void __pyx_memoryview__slice_assign_scalar(char *__pyx_v_data, Py_ssize_t
  */
       __pyx_memoryview__slice_assign_scalar(__pyx_v_data, (__pyx_v_shape + 1), (__pyx_v_strides + 1), (__pyx_v_ndim - 1), __pyx_v_itemsize, __pyx_v_item);
 
-      /* "View.MemoryView":1422
+      /* "View.MemoryView":1424
  *             _slice_assign_scalar(data, shape + 1, strides + 1,
  *                                 ndim - 1, itemsize, item)
  *             data += stride             # <<<<<<<<<<<<<<
@@ -21286,7 +21349,7 @@ static void __pyx_memoryview__slice_assign_scalar(char *__pyx_v_data, Py_ssize_t
   }
   __pyx_L3:;
 
-  /* "View.MemoryView":1407
+  /* "View.MemoryView":1409
  * 
  * @cname('__pyx_memoryview__slice_assign_scalar')
  * cdef void _slice_assign_scalar(char *data, Py_ssize_t *shape,             # <<<<<<<<<<<<<<
@@ -23798,9 +23861,9 @@ static CYTHON_SMALL_CODE int __Pyx_InitCachedBuiltins(void) {
   __pyx_builtin_range = __Pyx_GetBuiltinName(__pyx_n_s_range); if (!__pyx_builtin_range) __PYX_ERR(0, 194, __pyx_L1_error)
   __pyx_builtin_KeyError = __Pyx_GetBuiltinName(__pyx_n_s_KeyError); if (!__pyx_builtin_KeyError) __PYX_ERR(0, 308, __pyx_L1_error)
   __pyx_builtin_RuntimeError = __Pyx_GetBuiltinName(__pyx_n_s_RuntimeError); if (!__pyx_builtin_RuntimeError) __PYX_ERR(0, 435, __pyx_L1_error)
-  __pyx_builtin_Ellipsis = __Pyx_GetBuiltinName(__pyx_n_s_Ellipsis); if (!__pyx_builtin_Ellipsis) __PYX_ERR(1, 404, __pyx_L1_error)
-  __pyx_builtin_id = __Pyx_GetBuiltinName(__pyx_n_s_id); if (!__pyx_builtin_id) __PYX_ERR(1, 613, __pyx_L1_error)
-  __pyx_builtin_IndexError = __Pyx_GetBuiltinName(__pyx_n_s_IndexError); if (!__pyx_builtin_IndexError) __PYX_ERR(1, 832, __pyx_L1_error)
+  __pyx_builtin_Ellipsis = __Pyx_GetBuiltinName(__pyx_n_s_Ellipsis); if (!__pyx_builtin_Ellipsis) __PYX_ERR(1, 406, __pyx_L1_error)
+  __pyx_builtin_id = __Pyx_GetBuiltinName(__pyx_n_s_id); if (!__pyx_builtin_id) __PYX_ERR(1, 615, __pyx_L1_error)
+  __pyx_builtin_IndexError = __Pyx_GetBuiltinName(__pyx_n_s_IndexError); if (!__pyx_builtin_IndexError) __PYX_ERR(1, 834, __pyx_L1_error)
   return 0;
   __pyx_L1_error:;
   return -1;
@@ -23950,58 +24013,58 @@ static CYTHON_SMALL_CODE int __Pyx_InitCachedConstants(void) {
   __Pyx_GOTREF(__pyx_tuple__15);
   __Pyx_GIVEREF(__pyx_tuple__15);
 
-  /* "View.MemoryView":133
+  /* "View.MemoryView":134
  * 
  *         if not self.ndim:
  *             raise ValueError("Empty shape tuple for cython.array")             # <<<<<<<<<<<<<<
  * 
  *         if itemsize <= 0:
  */
-  __pyx_tuple__16 = PyTuple_Pack(1, __pyx_kp_s_Empty_shape_tuple_for_cython_arr); if (unlikely(!__pyx_tuple__16)) __PYX_ERR(1, 133, __pyx_L1_error)
+  __pyx_tuple__16 = PyTuple_Pack(1, __pyx_kp_s_Empty_shape_tuple_for_cython_arr); if (unlikely(!__pyx_tuple__16)) __PYX_ERR(1, 134, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_tuple__16);
   __Pyx_GIVEREF(__pyx_tuple__16);
 
-  /* "View.MemoryView":136
+  /* "View.MemoryView":137
  * 
  *         if itemsize <= 0:
  *             raise ValueError("itemsize <= 0 for cython.array")             # <<<<<<<<<<<<<<
  * 
  *         if not isinstance(format, bytes):
  */
-  __pyx_tuple__17 = PyTuple_Pack(1, __pyx_kp_s_itemsize_0_for_cython_array); if (unlikely(!__pyx_tuple__17)) __PYX_ERR(1, 136, __pyx_L1_error)
+  __pyx_tuple__17 = PyTuple_Pack(1, __pyx_kp_s_itemsize_0_for_cython_array); if (unlikely(!__pyx_tuple__17)) __PYX_ERR(1, 137, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_tuple__17);
   __Pyx_GIVEREF(__pyx_tuple__17);
 
-  /* "View.MemoryView":148
+  /* "View.MemoryView":149
  * 
  *         if not self._shape:
  *             raise MemoryError("unable to allocate shape and strides.")             # <<<<<<<<<<<<<<
  * 
  * 
  */
-  __pyx_tuple__18 = PyTuple_Pack(1, __pyx_kp_s_unable_to_allocate_shape_and_str); if (unlikely(!__pyx_tuple__18)) __PYX_ERR(1, 148, __pyx_L1_error)
+  __pyx_tuple__18 = PyTuple_Pack(1, __pyx_kp_s_unable_to_allocate_shape_and_str); if (unlikely(!__pyx_tuple__18)) __PYX_ERR(1, 149, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_tuple__18);
   __Pyx_GIVEREF(__pyx_tuple__18);
 
-  /* "View.MemoryView":176
+  /* "View.MemoryView":177
  *             self.data = <char *>malloc(self.len)
  *             if not self.data:
  *                 raise MemoryError("unable to allocate array data.")             # <<<<<<<<<<<<<<
  * 
  *             if self.dtype_is_object:
  */
-  __pyx_tuple__19 = PyTuple_Pack(1, __pyx_kp_s_unable_to_allocate_array_data); if (unlikely(!__pyx_tuple__19)) __PYX_ERR(1, 176, __pyx_L1_error)
+  __pyx_tuple__19 = PyTuple_Pack(1, __pyx_kp_s_unable_to_allocate_array_data); if (unlikely(!__pyx_tuple__19)) __PYX_ERR(1, 177, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_tuple__19);
   __Pyx_GIVEREF(__pyx_tuple__19);
 
-  /* "View.MemoryView":192
+  /* "View.MemoryView":193
  *             bufmode = PyBUF_F_CONTIGUOUS | PyBUF_ANY_CONTIGUOUS
  *         if not (flags & bufmode):
  *             raise ValueError("Can only create a buffer that is contiguous in memory.")             # <<<<<<<<<<<<<<
  *         info.buf = self.data
  *         info.len = self.len
  */
-  __pyx_tuple__20 = PyTuple_Pack(1, __pyx_kp_s_Can_only_create_a_buffer_that_is); if (unlikely(!__pyx_tuple__20)) __PYX_ERR(1, 192, __pyx_L1_error)
+  __pyx_tuple__20 = PyTuple_Pack(1, __pyx_kp_s_Can_only_create_a_buffer_that_is); if (unlikely(!__pyx_tuple__20)) __PYX_ERR(1, 193, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_tuple__20);
   __Pyx_GIVEREF(__pyx_tuple__20);
 
@@ -24024,58 +24087,58 @@ static CYTHON_SMALL_CODE int __Pyx_InitCachedConstants(void) {
   __Pyx_GOTREF(__pyx_tuple__22);
   __Pyx_GIVEREF(__pyx_tuple__22);
 
-  /* "View.MemoryView":418
+  /* "View.MemoryView":420
  *     def __setitem__(memoryview self, object index, object value):
  *         if self.view.readonly:
  *             raise TypeError("Cannot assign to read-only memoryview")             # <<<<<<<<<<<<<<
  * 
  *         have_slices, index = _unellipsify(index, self.view.ndim)
  */
-  __pyx_tuple__23 = PyTuple_Pack(1, __pyx_kp_s_Cannot_assign_to_read_only_memor); if (unlikely(!__pyx_tuple__23)) __PYX_ERR(1, 418, __pyx_L1_error)
+  __pyx_tuple__23 = PyTuple_Pack(1, __pyx_kp_s_Cannot_assign_to_read_only_memor); if (unlikely(!__pyx_tuple__23)) __PYX_ERR(1, 420, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_tuple__23);
   __Pyx_GIVEREF(__pyx_tuple__23);
 
-  /* "View.MemoryView":495
+  /* "View.MemoryView":497
  *             result = struct.unpack(self.view.format, bytesitem)
  *         except struct.error:
  *             raise ValueError("Unable to convert item to object")             # <<<<<<<<<<<<<<
  *         else:
  *             if len(self.view.format) == 1:
  */
-  __pyx_tuple__24 = PyTuple_Pack(1, __pyx_kp_s_Unable_to_convert_item_to_object); if (unlikely(!__pyx_tuple__24)) __PYX_ERR(1, 495, __pyx_L1_error)
+  __pyx_tuple__24 = PyTuple_Pack(1, __pyx_kp_s_Unable_to_convert_item_to_object); if (unlikely(!__pyx_tuple__24)) __PYX_ERR(1, 497, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_tuple__24);
   __Pyx_GIVEREF(__pyx_tuple__24);
 
-  /* "View.MemoryView":520
+  /* "View.MemoryView":522
  *     def __getbuffer__(self, Py_buffer *info, int flags):
  *         if flags & PyBUF_WRITABLE and self.view.readonly:
  *             raise ValueError("Cannot create writable memory view from read-only memoryview")             # <<<<<<<<<<<<<<
  * 
  *         if flags & PyBUF_ND:
  */
-  __pyx_tuple__25 = PyTuple_Pack(1, __pyx_kp_s_Cannot_create_writable_memory_vi); if (unlikely(!__pyx_tuple__25)) __PYX_ERR(1, 520, __pyx_L1_error)
+  __pyx_tuple__25 = PyTuple_Pack(1, __pyx_kp_s_Cannot_create_writable_memory_vi); if (unlikely(!__pyx_tuple__25)) __PYX_ERR(1, 522, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_tuple__25);
   __Pyx_GIVEREF(__pyx_tuple__25);
 
-  /* "View.MemoryView":570
+  /* "View.MemoryView":572
  *         if self.view.strides == NULL:
  * 
  *             raise ValueError("Buffer view does not expose strides")             # <<<<<<<<<<<<<<
  * 
  *         return tuple([stride for stride in self.view.strides[:self.view.ndim]])
  */
-  __pyx_tuple__26 = PyTuple_Pack(1, __pyx_kp_s_Buffer_view_does_not_expose_stri); if (unlikely(!__pyx_tuple__26)) __PYX_ERR(1, 570, __pyx_L1_error)
+  __pyx_tuple__26 = PyTuple_Pack(1, __pyx_kp_s_Buffer_view_does_not_expose_stri); if (unlikely(!__pyx_tuple__26)) __PYX_ERR(1, 572, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_tuple__26);
   __Pyx_GIVEREF(__pyx_tuple__26);
 
-  /* "View.MemoryView":577
+  /* "View.MemoryView":579
  *     def suboffsets(self):
  *         if self.view.suboffsets == NULL:
  *             return (-1,) * self.view.ndim             # <<<<<<<<<<<<<<
  * 
  *         return tuple([suboffset for suboffset in self.view.suboffsets[:self.view.ndim]])
  */
-  __pyx_tuple__27 = PyTuple_New(1); if (unlikely(!__pyx_tuple__27)) __PYX_ERR(1, 577, __pyx_L1_error)
+  __pyx_tuple__27 = PyTuple_New(1); if (unlikely(!__pyx_tuple__27)) __PYX_ERR(1, 579, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_tuple__27);
   __Pyx_INCREF(__pyx_int_neg_1);
   __Pyx_GIVEREF(__pyx_int_neg_1);
@@ -24101,25 +24164,25 @@ static CYTHON_SMALL_CODE int __Pyx_InitCachedConstants(void) {
   __Pyx_GOTREF(__pyx_tuple__29);
   __Pyx_GIVEREF(__pyx_tuple__29);
 
-  /* "View.MemoryView":682
+  /* "View.MemoryView":684
  *         if item is Ellipsis:
  *             if not seen_ellipsis:
  *                 result.extend([slice(None)] * (ndim - len(tup) + 1))             # <<<<<<<<<<<<<<
  *                 seen_ellipsis = True
  *             else:
  */
-  __pyx_slice__30 = PySlice_New(Py_None, Py_None, Py_None); if (unlikely(!__pyx_slice__30)) __PYX_ERR(1, 682, __pyx_L1_error)
+  __pyx_slice__30 = PySlice_New(Py_None, Py_None, Py_None); if (unlikely(!__pyx_slice__30)) __PYX_ERR(1, 684, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_slice__30);
   __Pyx_GIVEREF(__pyx_slice__30);
 
-  /* "View.MemoryView":703
+  /* "View.MemoryView":705
  *     for suboffset in suboffsets[:ndim]:
  *         if suboffset >= 0:
  *             raise ValueError("Indirect dimensions not supported")             # <<<<<<<<<<<<<<
  * 
  * 
  */
-  __pyx_tuple__31 = PyTuple_Pack(1, __pyx_kp_s_Indirect_dimensions_not_supporte); if (unlikely(!__pyx_tuple__31)) __PYX_ERR(1, 703, __pyx_L1_error)
+  __pyx_tuple__31 = PyTuple_Pack(1, __pyx_kp_s_Indirect_dimensions_not_supporte); if (unlikely(!__pyx_tuple__31)) __PYX_ERR(1, 705, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_tuple__31);
   __Pyx_GIVEREF(__pyx_tuple__31);
 
@@ -24145,58 +24208,58 @@ static CYTHON_SMALL_CODE int __Pyx_InitCachedConstants(void) {
   __Pyx_GOTREF(__pyx_tuple__34);
   __Pyx_GIVEREF(__pyx_tuple__34);
 
-  /* "View.MemoryView":286
+  /* "View.MemoryView":287
  *         return self.name
  * 
  * cdef generic = Enum("<strided and direct or indirect>")             # <<<<<<<<<<<<<<
  * cdef strided = Enum("<strided and direct>") # default
  * cdef indirect = Enum("<strided and indirect>")
  */
-  __pyx_tuple__36 = PyTuple_Pack(1, __pyx_kp_s_strided_and_direct_or_indirect); if (unlikely(!__pyx_tuple__36)) __PYX_ERR(1, 286, __pyx_L1_error)
+  __pyx_tuple__36 = PyTuple_Pack(1, __pyx_kp_s_strided_and_direct_or_indirect); if (unlikely(!__pyx_tuple__36)) __PYX_ERR(1, 287, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_tuple__36);
   __Pyx_GIVEREF(__pyx_tuple__36);
 
-  /* "View.MemoryView":287
+  /* "View.MemoryView":288
  * 
  * cdef generic = Enum("<strided and direct or indirect>")
  * cdef strided = Enum("<strided and direct>") # default             # <<<<<<<<<<<<<<
  * cdef indirect = Enum("<strided and indirect>")
  * 
  */
-  __pyx_tuple__37 = PyTuple_Pack(1, __pyx_kp_s_strided_and_direct); if (unlikely(!__pyx_tuple__37)) __PYX_ERR(1, 287, __pyx_L1_error)
+  __pyx_tuple__37 = PyTuple_Pack(1, __pyx_kp_s_strided_and_direct); if (unlikely(!__pyx_tuple__37)) __PYX_ERR(1, 288, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_tuple__37);
   __Pyx_GIVEREF(__pyx_tuple__37);
 
-  /* "View.MemoryView":288
+  /* "View.MemoryView":289
  * cdef generic = Enum("<strided and direct or indirect>")
  * cdef strided = Enum("<strided and direct>") # default
  * cdef indirect = Enum("<strided and indirect>")             # <<<<<<<<<<<<<<
  * 
  * 
  */
-  __pyx_tuple__38 = PyTuple_Pack(1, __pyx_kp_s_strided_and_indirect); if (unlikely(!__pyx_tuple__38)) __PYX_ERR(1, 288, __pyx_L1_error)
+  __pyx_tuple__38 = PyTuple_Pack(1, __pyx_kp_s_strided_and_indirect); if (unlikely(!__pyx_tuple__38)) __PYX_ERR(1, 289, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_tuple__38);
   __Pyx_GIVEREF(__pyx_tuple__38);
 
-  /* "View.MemoryView":291
+  /* "View.MemoryView":292
  * 
  * 
  * cdef contiguous = Enum("<contiguous and direct>")             # <<<<<<<<<<<<<<
  * cdef indirect_contiguous = Enum("<contiguous and indirect>")
  * 
  */
-  __pyx_tuple__39 = PyTuple_Pack(1, __pyx_kp_s_contiguous_and_direct); if (unlikely(!__pyx_tuple__39)) __PYX_ERR(1, 291, __pyx_L1_error)
+  __pyx_tuple__39 = PyTuple_Pack(1, __pyx_kp_s_contiguous_and_direct); if (unlikely(!__pyx_tuple__39)) __PYX_ERR(1, 292, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_tuple__39);
   __Pyx_GIVEREF(__pyx_tuple__39);
 
-  /* "View.MemoryView":292
+  /* "View.MemoryView":293
  * 
  * cdef contiguous = Enum("<contiguous and direct>")
  * cdef indirect_contiguous = Enum("<contiguous and indirect>")             # <<<<<<<<<<<<<<
  * 
  * 
  */
-  __pyx_tuple__40 = PyTuple_Pack(1, __pyx_kp_s_contiguous_and_indirect); if (unlikely(!__pyx_tuple__40)) __PYX_ERR(1, 292, __pyx_L1_error)
+  __pyx_tuple__40 = PyTuple_Pack(1, __pyx_kp_s_contiguous_and_indirect); if (unlikely(!__pyx_tuple__40)) __PYX_ERR(1, 293, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_tuple__40);
   __Pyx_GIVEREF(__pyx_tuple__40);
 
@@ -24374,21 +24437,21 @@ static int __Pyx_modinit_type_init_code(void) {
   __pyx_ptype_9csimdjson___pyx_scope_struct_4_get_implementations = &__pyx_type_9csimdjson___pyx_scope_struct_4_get_implementations;
   __pyx_vtabptr_array = &__pyx_vtable_array;
   __pyx_vtable_array.get_memview = (PyObject *(*)(struct __pyx_array_obj *))__pyx_array_get_memview;
-  if (PyType_Ready(&__pyx_type___pyx_array) < 0) __PYX_ERR(1, 105, __pyx_L1_error)
+  if (PyType_Ready(&__pyx_type___pyx_array) < 0) __PYX_ERR(1, 106, __pyx_L1_error)
   #if PY_VERSION_HEX < 0x030800B1
   __pyx_type___pyx_array.tp_print = 0;
   #endif
-  if (__Pyx_SetVtable(__pyx_type___pyx_array.tp_dict, __pyx_vtabptr_array) < 0) __PYX_ERR(1, 105, __pyx_L1_error)
-  if (__Pyx_setup_reduce((PyObject*)&__pyx_type___pyx_array) < 0) __PYX_ERR(1, 105, __pyx_L1_error)
+  if (__Pyx_SetVtable(__pyx_type___pyx_array.tp_dict, __pyx_vtabptr_array) < 0) __PYX_ERR(1, 106, __pyx_L1_error)
+  if (__Pyx_setup_reduce((PyObject*)&__pyx_type___pyx_array) < 0) __PYX_ERR(1, 106, __pyx_L1_error)
   __pyx_array_type = &__pyx_type___pyx_array;
-  if (PyType_Ready(&__pyx_type___pyx_MemviewEnum) < 0) __PYX_ERR(1, 279, __pyx_L1_error)
+  if (PyType_Ready(&__pyx_type___pyx_MemviewEnum) < 0) __PYX_ERR(1, 280, __pyx_L1_error)
   #if PY_VERSION_HEX < 0x030800B1
   __pyx_type___pyx_MemviewEnum.tp_print = 0;
   #endif
   if ((CYTHON_USE_TYPE_SLOTS && CYTHON_USE_PYTYPE_LOOKUP) && likely(!__pyx_type___pyx_MemviewEnum.tp_dictoffset && __pyx_type___pyx_MemviewEnum.tp_getattro == PyObject_GenericGetAttr)) {
     __pyx_type___pyx_MemviewEnum.tp_getattro = __Pyx_PyObject_GenericGetAttr;
   }
-  if (__Pyx_setup_reduce((PyObject*)&__pyx_type___pyx_MemviewEnum) < 0) __PYX_ERR(1, 279, __pyx_L1_error)
+  if (__Pyx_setup_reduce((PyObject*)&__pyx_type___pyx_MemviewEnum) < 0) __PYX_ERR(1, 280, __pyx_L1_error)
   __pyx_MemviewEnum_type = &__pyx_type___pyx_MemviewEnum;
   __pyx_vtabptr_memoryview = &__pyx_vtable_memoryview;
   __pyx_vtable_memoryview.get_item_pointer = (char *(*)(struct __pyx_memoryview_obj *, PyObject *))__pyx_memoryview_get_item_pointer;
@@ -24398,30 +24461,30 @@ static int __Pyx_modinit_type_init_code(void) {
   __pyx_vtable_memoryview.setitem_indexed = (PyObject *(*)(struct __pyx_memoryview_obj *, PyObject *, PyObject *))__pyx_memoryview_setitem_indexed;
   __pyx_vtable_memoryview.convert_item_to_object = (PyObject *(*)(struct __pyx_memoryview_obj *, char *))__pyx_memoryview_convert_item_to_object;
   __pyx_vtable_memoryview.assign_item_from_object = (PyObject *(*)(struct __pyx_memoryview_obj *, char *, PyObject *))__pyx_memoryview_assign_item_from_object;
-  if (PyType_Ready(&__pyx_type___pyx_memoryview) < 0) __PYX_ERR(1, 330, __pyx_L1_error)
+  if (PyType_Ready(&__pyx_type___pyx_memoryview) < 0) __PYX_ERR(1, 331, __pyx_L1_error)
   #if PY_VERSION_HEX < 0x030800B1
   __pyx_type___pyx_memoryview.tp_print = 0;
   #endif
   if ((CYTHON_USE_TYPE_SLOTS && CYTHON_USE_PYTYPE_LOOKUP) && likely(!__pyx_type___pyx_memoryview.tp_dictoffset && __pyx_type___pyx_memoryview.tp_getattro == PyObject_GenericGetAttr)) {
     __pyx_type___pyx_memoryview.tp_getattro = __Pyx_PyObject_GenericGetAttr;
   }
-  if (__Pyx_SetVtable(__pyx_type___pyx_memoryview.tp_dict, __pyx_vtabptr_memoryview) < 0) __PYX_ERR(1, 330, __pyx_L1_error)
-  if (__Pyx_setup_reduce((PyObject*)&__pyx_type___pyx_memoryview) < 0) __PYX_ERR(1, 330, __pyx_L1_error)
+  if (__Pyx_SetVtable(__pyx_type___pyx_memoryview.tp_dict, __pyx_vtabptr_memoryview) < 0) __PYX_ERR(1, 331, __pyx_L1_error)
+  if (__Pyx_setup_reduce((PyObject*)&__pyx_type___pyx_memoryview) < 0) __PYX_ERR(1, 331, __pyx_L1_error)
   __pyx_memoryview_type = &__pyx_type___pyx_memoryview;
   __pyx_vtabptr__memoryviewslice = &__pyx_vtable__memoryviewslice;
   __pyx_vtable__memoryviewslice.__pyx_base = *__pyx_vtabptr_memoryview;
   __pyx_vtable__memoryviewslice.__pyx_base.convert_item_to_object = (PyObject *(*)(struct __pyx_memoryview_obj *, char *))__pyx_memoryviewslice_convert_item_to_object;
   __pyx_vtable__memoryviewslice.__pyx_base.assign_item_from_object = (PyObject *(*)(struct __pyx_memoryview_obj *, char *, PyObject *))__pyx_memoryviewslice_assign_item_from_object;
   __pyx_type___pyx_memoryviewslice.tp_base = __pyx_memoryview_type;
-  if (PyType_Ready(&__pyx_type___pyx_memoryviewslice) < 0) __PYX_ERR(1, 965, __pyx_L1_error)
+  if (PyType_Ready(&__pyx_type___pyx_memoryviewslice) < 0) __PYX_ERR(1, 967, __pyx_L1_error)
   #if PY_VERSION_HEX < 0x030800B1
   __pyx_type___pyx_memoryviewslice.tp_print = 0;
   #endif
   if ((CYTHON_USE_TYPE_SLOTS && CYTHON_USE_PYTYPE_LOOKUP) && likely(!__pyx_type___pyx_memoryviewslice.tp_dictoffset && __pyx_type___pyx_memoryviewslice.tp_getattro == PyObject_GenericGetAttr)) {
     __pyx_type___pyx_memoryviewslice.tp_getattro = __Pyx_PyObject_GenericGetAttr;
   }
-  if (__Pyx_SetVtable(__pyx_type___pyx_memoryviewslice.tp_dict, __pyx_vtabptr__memoryviewslice) < 0) __PYX_ERR(1, 965, __pyx_L1_error)
-  if (__Pyx_setup_reduce((PyObject*)&__pyx_type___pyx_memoryviewslice) < 0) __PYX_ERR(1, 965, __pyx_L1_error)
+  if (__Pyx_SetVtable(__pyx_type___pyx_memoryviewslice.tp_dict, __pyx_vtabptr__memoryviewslice) < 0) __PYX_ERR(1, 967, __pyx_L1_error)
+  if (__Pyx_setup_reduce((PyObject*)&__pyx_type___pyx_memoryviewslice) < 0) __PYX_ERR(1, 967, __pyx_L1_error)
   __pyx_memoryviewslice_type = &__pyx_type___pyx_memoryviewslice;
   __Pyx_RefNannyFinishContext();
   return 0;
@@ -24812,90 +24875,90 @@ if (!__Pyx_RefNanny) {
   if (PyDict_SetItem(__pyx_d, __pyx_n_s_test, __pyx_t_4) < 0) __PYX_ERR(0, 1, __pyx_L1_error)
   __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
 
-  /* "View.MemoryView":209
+  /* "View.MemoryView":210
  *         info.obj = self
  * 
  *     __pyx_getbuffer = capsule(<void *> &__pyx_array_getbuffer, "getbuffer(obj, view, flags)")             # <<<<<<<<<<<<<<
  * 
  *     def __dealloc__(array self):
  */
-  __pyx_t_4 = __pyx_capsule_create(((void *)(&__pyx_array_getbuffer)), ((char *)"getbuffer(obj, view, flags)")); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 209, __pyx_L1_error)
+  __pyx_t_4 = __pyx_capsule_create(((void *)(&__pyx_array_getbuffer)), ((char *)"getbuffer(obj, view, flags)")); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 210, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_4);
-  if (PyDict_SetItem((PyObject *)__pyx_array_type->tp_dict, __pyx_n_s_pyx_getbuffer, __pyx_t_4) < 0) __PYX_ERR(1, 209, __pyx_L1_error)
+  if (PyDict_SetItem((PyObject *)__pyx_array_type->tp_dict, __pyx_n_s_pyx_getbuffer, __pyx_t_4) < 0) __PYX_ERR(1, 210, __pyx_L1_error)
   __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
   PyType_Modified(__pyx_array_type);
 
-  /* "View.MemoryView":286
+  /* "View.MemoryView":287
  *         return self.name
  * 
  * cdef generic = Enum("<strided and direct or indirect>")             # <<<<<<<<<<<<<<
  * cdef strided = Enum("<strided and direct>") # default
  * cdef indirect = Enum("<strided and indirect>")
  */
-  __pyx_t_4 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__36, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 286, __pyx_L1_error)
+  __pyx_t_4 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__36, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 287, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_4);
   __Pyx_XGOTREF(generic);
   __Pyx_DECREF_SET(generic, __pyx_t_4);
   __Pyx_GIVEREF(__pyx_t_4);
   __pyx_t_4 = 0;
 
-  /* "View.MemoryView":287
+  /* "View.MemoryView":288
  * 
  * cdef generic = Enum("<strided and direct or indirect>")
  * cdef strided = Enum("<strided and direct>") # default             # <<<<<<<<<<<<<<
  * cdef indirect = Enum("<strided and indirect>")
  * 
  */
-  __pyx_t_4 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__37, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 287, __pyx_L1_error)
+  __pyx_t_4 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__37, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 288, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_4);
   __Pyx_XGOTREF(strided);
   __Pyx_DECREF_SET(strided, __pyx_t_4);
   __Pyx_GIVEREF(__pyx_t_4);
   __pyx_t_4 = 0;
 
-  /* "View.MemoryView":288
+  /* "View.MemoryView":289
  * cdef generic = Enum("<strided and direct or indirect>")
  * cdef strided = Enum("<strided and direct>") # default
  * cdef indirect = Enum("<strided and indirect>")             # <<<<<<<<<<<<<<
  * 
  * 
  */
-  __pyx_t_4 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__38, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 288, __pyx_L1_error)
+  __pyx_t_4 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__38, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 289, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_4);
   __Pyx_XGOTREF(indirect);
   __Pyx_DECREF_SET(indirect, __pyx_t_4);
   __Pyx_GIVEREF(__pyx_t_4);
   __pyx_t_4 = 0;
 
-  /* "View.MemoryView":291
+  /* "View.MemoryView":292
  * 
  * 
  * cdef contiguous = Enum("<contiguous and direct>")             # <<<<<<<<<<<<<<
  * cdef indirect_contiguous = Enum("<contiguous and indirect>")
  * 
  */
-  __pyx_t_4 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__39, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 291, __pyx_L1_error)
+  __pyx_t_4 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__39, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 292, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_4);
   __Pyx_XGOTREF(contiguous);
   __Pyx_DECREF_SET(contiguous, __pyx_t_4);
   __Pyx_GIVEREF(__pyx_t_4);
   __pyx_t_4 = 0;
 
-  /* "View.MemoryView":292
+  /* "View.MemoryView":293
  * 
  * cdef contiguous = Enum("<contiguous and direct>")
  * cdef indirect_contiguous = Enum("<contiguous and indirect>")             # <<<<<<<<<<<<<<
  * 
  * 
  */
-  __pyx_t_4 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__40, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 292, __pyx_L1_error)
+  __pyx_t_4 = __Pyx_PyObject_Call(((PyObject *)__pyx_MemviewEnum_type), __pyx_tuple__40, NULL); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 293, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_4);
   __Pyx_XGOTREF(indirect_contiguous);
   __Pyx_DECREF_SET(indirect_contiguous, __pyx_t_4);
   __Pyx_GIVEREF(__pyx_t_4);
   __pyx_t_4 = 0;
 
-  /* "View.MemoryView":316
+  /* "View.MemoryView":317
  * 
  * DEF THREAD_LOCKS_PREALLOCATED = 8
  * cdef int __pyx_memoryview_thread_locks_used = 0             # <<<<<<<<<<<<<<
@@ -24904,7 +24967,7 @@ if (!__Pyx_RefNanny) {
  */
   __pyx_memoryview_thread_locks_used = 0;
 
-  /* "View.MemoryView":317
+  /* "View.MemoryView":318
  * DEF THREAD_LOCKS_PREALLOCATED = 8
  * cdef int __pyx_memoryview_thread_locks_used = 0
  * cdef PyThread_type_lock[THREAD_LOCKS_PREALLOCATED] __pyx_memoryview_thread_locks = [             # <<<<<<<<<<<<<<
@@ -24921,29 +24984,29 @@ if (!__Pyx_RefNanny) {
   __pyx_t_5[7] = PyThread_allocate_lock();
   memcpy(&(__pyx_memoryview_thread_locks[0]), __pyx_t_5, sizeof(__pyx_memoryview_thread_locks[0]) * (8));
 
-  /* "View.MemoryView":549
+  /* "View.MemoryView":551
  *         info.obj = self
  * 
  *     __pyx_getbuffer = capsule(<void *> &__pyx_memoryview_getbuffer, "getbuffer(obj, view, flags)")             # <<<<<<<<<<<<<<
  * 
  * 
  */
-  __pyx_t_4 = __pyx_capsule_create(((void *)(&__pyx_memoryview_getbuffer)), ((char *)"getbuffer(obj, view, flags)")); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 549, __pyx_L1_error)
+  __pyx_t_4 = __pyx_capsule_create(((void *)(&__pyx_memoryview_getbuffer)), ((char *)"getbuffer(obj, view, flags)")); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 551, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_4);
-  if (PyDict_SetItem((PyObject *)__pyx_memoryview_type->tp_dict, __pyx_n_s_pyx_getbuffer, __pyx_t_4) < 0) __PYX_ERR(1, 549, __pyx_L1_error)
+  if (PyDict_SetItem((PyObject *)__pyx_memoryview_type->tp_dict, __pyx_n_s_pyx_getbuffer, __pyx_t_4) < 0) __PYX_ERR(1, 551, __pyx_L1_error)
   __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
   PyType_Modified(__pyx_memoryview_type);
 
-  /* "View.MemoryView":995
+  /* "View.MemoryView":997
  *         return self.from_object
  * 
  *     __pyx_getbuffer = capsule(<void *> &__pyx_memoryview_getbuffer, "getbuffer(obj, view, flags)")             # <<<<<<<<<<<<<<
  * 
  * 
  */
-  __pyx_t_4 = __pyx_capsule_create(((void *)(&__pyx_memoryview_getbuffer)), ((char *)"getbuffer(obj, view, flags)")); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 995, __pyx_L1_error)
+  __pyx_t_4 = __pyx_capsule_create(((void *)(&__pyx_memoryview_getbuffer)), ((char *)"getbuffer(obj, view, flags)")); if (unlikely(!__pyx_t_4)) __PYX_ERR(1, 997, __pyx_L1_error)
   __Pyx_GOTREF(__pyx_t_4);
-  if (PyDict_SetItem((PyObject *)__pyx_memoryviewslice_type->tp_dict, __pyx_n_s_pyx_getbuffer, __pyx_t_4) < 0) __PYX_ERR(1, 995, __pyx_L1_error)
+  if (PyDict_SetItem((PyObject *)__pyx_memoryviewslice_type->tp_dict, __pyx_n_s_pyx_getbuffer, __pyx_t_4) < 0) __PYX_ERR(1, 997, __pyx_L1_error)
   __Pyx_DECREF(__pyx_t_4); __pyx_t_4 = 0;
   PyType_Modified(__pyx_memoryviewslice_type);
 
@@ -25981,7 +26044,7 @@ __Pyx_init_memviewslice(struct __pyx_memoryview_obj *memview,
 static void __pyx_fatalerror(const char *fmt, ...) Py_NO_RETURN {
     va_list vargs;
     char msg[200];
-#ifdef HAVE_STDARG_PROTOTYPES
+#if PY_VERSION_HEX >= 0x030A0000 || defined(HAVE_STDARG_PROTOTYPES)
     va_start(vargs, fmt);
 #else
     va_start(vargs);
@@ -27163,7 +27226,7 @@ static PyObject* __Pyx_PyUnicode_BuildFromAscii(Py_ssize_t ulength, char* chars,
                 padding = PyUnicode_FromOrdinal(padding_char);
                 if (likely(padding) && uoffset > prepend_sign + 1) {
                     PyObject *tmp;
-                    PyObject *repeat = PyInt_FromSize_t(uoffset - prepend_sign);
+                    PyObject *repeat = PyInt_FromSsize_t(uoffset - prepend_sign);
                     if (unlikely(!repeat)) goto done_or_error;
                     tmp = PyNumber_Multiply(padding, repeat);
                     Py_DECREF(repeat);
diff --git a/simdjson/simdjson.cpp b/simdjson/simdjson.cpp
index c98c4b1..611411a 100644
--- a/simdjson/simdjson.cpp
+++ b/simdjson/simdjson.cpp
@@ -1,4 +1,4 @@
-/* auto-generated on 2022-06-02 13:56:20 -0400. Do not edit! */
+/* auto-generated on 2022-11-23 10:31:42 -0500. Do not edit! */
 /* begin file src/simdjson.cpp */
 #include "simdjson.h"
 
@@ -868,9 +868,9 @@ inline char *format_buffer(char *buf, int len, int decimal_exponent,
 
     std::memset(buf + k, '0', static_cast<size_t>(n) - static_cast<size_t>(k));
     // Make it look like a floating-point number (#362, #378)
-    // buf[n + 0] = '.';
-    // buf[n + 1] = '0';
-    return buf + (static_cast<size_t>(n));
+    buf[n + 0] = '.';
+    buf[n + 1] = '0';
+    return buf + (static_cast<size_t>(n)) + 2;
   }
 
   if (0 < n && n <= max_exp) {
@@ -933,10 +933,8 @@ char *to_chars(char *first, const char *last, double value) {
   {
     *first++ = '0';
     // Make it look like a floating-point number (#362, #378)
-    if(negative) {
-      *first++ = '.';
-      *first++ = '0';
-    }
+    *first++ = '.';
+    *first++ = '0';
     return first;
   }
   // Compute v = buffer * 10^decimal_exponent.
@@ -1589,7 +1587,8 @@ namespace internal {
     { INSUFFICIENT_PADDING, "simdjson requires the input JSON string to have at least SIMDJSON_PADDING extra bytes allocated, beyond the string's length. Consider using the simdjson::padded_string class if needed." },
     { INCOMPLETE_ARRAY_OR_OBJECT, "JSON document ended early in the middle of an object or array." },
     { SCALAR_DOCUMENT_AS_VALUE, "A JSON document made of a scalar (number, Boolean, null or string) is treated as a value. Use get_bool(), get_double(), etc. on the document instead. "},
-    { OUT_OF_BOUNDS, "Attempted to access location outside of document."}
+    { OUT_OF_BOUNDS, "Attempted to access location outside of document."},
+    { TRAILING_CONTENT, "Unexpected trailing content in the JSON input."}
   }; // error_messages[]
 
 } // namespace internal
@@ -2677,7 +2676,7 @@ class detect_best_supported_implementation_on_first_use final : public implement
   simdjson_warn_unused bool validate_utf8(const char * buf, size_t len) const noexcept final override {
     return set_best()->validate_utf8(buf, len);
   }
-  simdjson_really_inline detect_best_supported_implementation_on_first_use() noexcept : implementation("best_supported_detector", "Detects the best supported implementation and sets it", 0) {}
+  simdjson_inline detect_best_supported_implementation_on_first_use() noexcept : implementation("best_supported_detector", "Detects the best supported implementation and sets it", 0) {}
 private:
   const implementation *set_best() const noexcept;
 };
@@ -2793,7 +2792,6 @@ simdjson_warn_unused error_code minify(const char *buf, size_t len, char *dst, s
 simdjson_warn_unused bool validate_utf8(const char *buf, size_t len) noexcept {
   return get_active_implementation()->validate_utf8(buf, len);
 }
-
 const implementation * builtin_implementation() {
   static const implementation * builtin_impl = get_available_implementations()[SIMDJSON_STRINGIFY(SIMDJSON_BUILTIN_IMPLEMENTATION)];
   assert(builtin_impl);
@@ -2850,17 +2848,17 @@ namespace {
 using namespace simd;
 
 struct json_character_block {
-  static simdjson_really_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
+  static simdjson_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
 
-  simdjson_really_inline uint64_t whitespace() const noexcept { return _whitespace; }
-  simdjson_really_inline uint64_t op() const noexcept { return _op; }
-  simdjson_really_inline uint64_t scalar() const noexcept { return ~(op() | whitespace()); }
+  simdjson_inline uint64_t whitespace() const noexcept { return _whitespace; }
+  simdjson_inline uint64_t op() const noexcept { return _op; }
+  simdjson_inline uint64_t scalar() const noexcept { return ~(op() | whitespace()); }
 
   uint64_t _whitespace;
   uint64_t _op;
 };
 
-simdjson_really_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
+simdjson_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
   // Functional programming causes trouble with Visual Studio.
   // Keeping this version in comments since it is much nicer:
   // auto v = in.map<uint8_t>([&](simd8<uint8_t> chunk) {
@@ -2914,15 +2912,15 @@ simdjson_really_inline json_character_block json_character_block::classify(const
   return { whitespace, op };
 }
 
-simdjson_really_inline bool is_ascii(const simd8x64<uint8_t>& input) {
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input) {
     simd8<uint8_t> bits = input.reduce_or();
-    return bits.max_val() < 0b10000000u;
+    return bits.max_val() < 0x80u;
 }
 
-simdjson_unused simdjson_really_inline simd8<bool> must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
-    simd8<bool> is_second_byte = prev1 >= uint8_t(0b11000000u);
-    simd8<bool> is_third_byte  = prev2 >= uint8_t(0b11100000u);
-    simd8<bool> is_fourth_byte = prev3 >= uint8_t(0b11110000u);
+simdjson_unused simdjson_inline simd8<bool> must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+    simd8<bool> is_second_byte = prev1 >= uint8_t(0xc0u);
+    simd8<bool> is_third_byte  = prev2 >= uint8_t(0xe0u);
+    simd8<bool> is_fourth_byte = prev3 >= uint8_t(0xf0u);
     // Use ^ instead of | for is_*_byte, because ^ is commutative, and the caller is using ^ as well.
     // This will work fine because we only have to report errors for cases with 0-1 lead bytes.
     // Multiple lead bytes implies 2 overlapping multibyte characters, and if that happens, there is
@@ -2931,9 +2929,9 @@ simdjson_unused simdjson_really_inline simd8<bool> must_be_continuation(const si
     return is_second_byte ^ is_third_byte ^ is_fourth_byte;
 }
 
-simdjson_really_inline simd8<bool> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
-    simd8<bool> is_third_byte  = prev2 >= uint8_t(0b11100000u);
-    simd8<bool> is_fourth_byte = prev3 >= uint8_t(0b11110000u);
+simdjson_inline simd8<bool> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+    simd8<bool> is_third_byte  = prev2 >= uint8_t(0xe0u);
+    simd8<bool> is_fourth_byte = prev3 >= uint8_t(0xf0u);
     return is_third_byte ^ is_fourth_byte;
 }
 
@@ -2949,7 +2947,7 @@ namespace utf8_validation {
 
 using namespace simd;
 
-  simdjson_really_inline simd8<uint8_t> check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+  simdjson_inline simd8<uint8_t> check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
 // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
 // Bit 1 = Too Long (ASCII followed by continuation)
 // Bit 2 = Overlong 3-byte
@@ -3039,7 +3037,7 @@ using namespace simd;
     );
     return (byte_1_high & byte_1_low & byte_2_high);
   }
-  simdjson_really_inline simd8<uint8_t> check_multibyte_lengths(const simd8<uint8_t> input,
+  simdjson_inline simd8<uint8_t> check_multibyte_lengths(const simd8<uint8_t> input,
       const simd8<uint8_t> prev_input, const simd8<uint8_t> sc) {
     simd8<uint8_t> prev2 = input.prev<2>(prev_input);
     simd8<uint8_t> prev3 = input.prev<3>(prev_input);
@@ -3052,7 +3050,7 @@ using namespace simd;
   // Return nonzero if there are incomplete multibyte characters at the end of the block:
   // e.g. if there is a 4-byte character, but it's 3 bytes from the end.
   //
-  simdjson_really_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
+  simdjson_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
     // If the previous input's last 3 bytes match this, they're too short (they ended at EOF):
     // ... 1111____ 111_____ 11______
 #if SIMDJSON_IMPLEMENTATION_ICELAKE
@@ -3064,14 +3062,14 @@ using namespace simd;
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
-      255, 255, 255, 255, 255, 0b11110000u-1, 0b11100000u-1, 0b11000000u-1
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
     };
 #else
     static const uint8_t max_array[32] = {
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
-      255, 255, 255, 255, 255, 0b11110000u-1, 0b11100000u-1, 0b11000000u-1
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
     };
 #endif
     const simd8<uint8_t> max_value(&max_array[sizeof(max_array)-sizeof(simd8<uint8_t>)]);
@@ -3089,7 +3087,7 @@ using namespace simd;
     //
     // Check whether the current bytes are valid UTF-8.
     //
-    simdjson_really_inline void check_utf8_bytes(const simd8<uint8_t> input, const simd8<uint8_t> prev_input) {
+    simdjson_inline void check_utf8_bytes(const simd8<uint8_t> input, const simd8<uint8_t> prev_input) {
       // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ lead bytes
       // (2, 3, 4-byte leads become large positive numbers instead of small negative numbers)
       simd8<uint8_t> prev1 = input.prev<1>(prev_input);
@@ -3100,13 +3098,21 @@ using namespace simd;
     // The only problem that can happen at EOF is that a multibyte character is too short
     // or a byte value too large in the last bytes: check_special_cases only checks for bytes
     // too large in the first of two bytes.
-    simdjson_really_inline void check_eof() {
+    simdjson_inline void check_eof() {
       // If the previous block had incomplete UTF-8 characters at the end, an ASCII block can't
       // possibly finish them.
       this->error |= this->prev_incomplete;
     }
 
-    simdjson_really_inline void check_next_input(const simd8x64<uint8_t>& input) {
+#ifndef SIMDJSON_IF_CONSTEXPR
+#if SIMDJSON_CPLUSPLUS17
+#define SIMDJSON_IF_CONSTEXPR if constexpr
+#else
+#define SIMDJSON_IF_CONSTEXPR if
+#endif
+#endif
+
+    simdjson_inline void check_next_input(const simd8x64<uint8_t>& input) {
       if(simdjson_likely(is_ascii(input))) {
         this->error |= this->prev_incomplete;
       } else {
@@ -3115,12 +3121,12 @@ using namespace simd;
                 ||(simd8x64<uint8_t>::NUM_CHUNKS == 2)
                 || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
                 "We support one, two or four chunks per 64-byte block.");
-        if(simd8x64<uint8_t>::NUM_CHUNKS == 1) {
+        SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 1) {
           this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
-        } if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
           this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
           this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
           this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
           this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
           this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
@@ -3131,7 +3137,7 @@ using namespace simd;
       }
     }
     // do not forget to call check_eof!
-    simdjson_really_inline error_code errors() {
+    simdjson_inline error_code errors() {
       return this->error.any_bits_set_anywhere() ? error_code::UTF8_ERROR : error_code::SUCCESS;
     }
 
@@ -3159,10 +3165,10 @@ namespace {
 template<size_t STEP_SIZE>
 struct buf_block_reader {
 public:
-  simdjson_really_inline buf_block_reader(const uint8_t *_buf, size_t _len);
-  simdjson_really_inline size_t block_index();
-  simdjson_really_inline bool has_full_block() const;
-  simdjson_really_inline const uint8_t *full_block() const;
+  simdjson_inline buf_block_reader(const uint8_t *_buf, size_t _len);
+  simdjson_inline size_t block_index();
+  simdjson_inline bool has_full_block() const;
+  simdjson_inline const uint8_t *full_block() const;
   /**
    * Get the last block, padded with spaces.
    *
@@ -3172,8 +3178,8 @@ struct buf_block_reader {
    *
    * @return the number of effective characters in the last block.
    */
-  simdjson_really_inline size_t get_remainder(uint8_t *dst) const;
-  simdjson_really_inline void advance();
+  simdjson_inline size_t get_remainder(uint8_t *dst) const;
+  simdjson_inline void advance();
 private:
   const uint8_t *buf;
   const size_t len;
@@ -3212,23 +3218,23 @@ simdjson_unused static char * format_mask(uint64_t mask) {
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len) : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, idx{0} {}
+simdjson_inline buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len) : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, idx{0} {}
 
 template<size_t STEP_SIZE>
-simdjson_really_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
+simdjson_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
   return idx < lenminusstep;
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
+simdjson_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
   return &buf[idx];
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
   if(len == idx) { return 0; } // memcpy(dst, null, 0) will trigger an error with some sanitizers
   std::memset(dst, 0x20, STEP_SIZE); // std::memset STEP_SIZE because it's more efficient to write out 8 or 16 bytes at once.
   std::memcpy(dst, buf + idx, len - idx);
@@ -3236,7 +3242,7 @@ simdjson_really_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline void buf_block_reader<STEP_SIZE>::advance() {
+simdjson_inline void buf_block_reader<STEP_SIZE>::advance() {
   idx += STEP_SIZE;
 }
 
@@ -3252,27 +3258,27 @@ namespace stage1 {
 
 struct json_string_block {
   // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
-  simdjson_really_inline json_string_block(uint64_t backslash, uint64_t escaped, uint64_t quote, uint64_t in_string) :
+  simdjson_inline json_string_block(uint64_t backslash, uint64_t escaped, uint64_t quote, uint64_t in_string) :
   _backslash(backslash), _escaped(escaped), _quote(quote), _in_string(in_string) {}
 
   // Escaped characters (characters following an escape() character)
-  simdjson_really_inline uint64_t escaped() const { return _escaped; }
+  simdjson_inline uint64_t escaped() const { return _escaped; }
   // Escape characters (backslashes that are not escaped--i.e. in \\, includes only the first \)
-  simdjson_really_inline uint64_t escape() const { return _backslash & ~_escaped; }
+  simdjson_inline uint64_t escape() const { return _backslash & ~_escaped; }
   // Real (non-backslashed) quotes
-  simdjson_really_inline uint64_t quote() const { return _quote; }
+  simdjson_inline uint64_t quote() const { return _quote; }
   // Start quotes of strings
-  simdjson_really_inline uint64_t string_start() const { return _quote & _in_string; }
+  simdjson_inline uint64_t string_start() const { return _quote & _in_string; }
   // End quotes of strings
-  simdjson_really_inline uint64_t string_end() const { return _quote & ~_in_string; }
+  simdjson_inline uint64_t string_end() const { return _quote & ~_in_string; }
   // Only characters inside the string (not including the quotes)
-  simdjson_really_inline uint64_t string_content() const { return _in_string & ~_quote; }
+  simdjson_inline uint64_t string_content() const { return _in_string & ~_quote; }
   // Return a mask of whether the given characters are inside a string (only works on non-quotes)
-  simdjson_really_inline uint64_t non_quote_inside_string(uint64_t mask) const { return mask & _in_string; }
+  simdjson_inline uint64_t non_quote_inside_string(uint64_t mask) const { return mask & _in_string; }
   // Return a mask of whether the given characters are inside a string (only works on non-quotes)
-  simdjson_really_inline uint64_t non_quote_outside_string(uint64_t mask) const { return mask & ~_in_string; }
+  simdjson_inline uint64_t non_quote_outside_string(uint64_t mask) const { return mask & ~_in_string; }
   // Tail of string (everything except the start quote)
-  simdjson_really_inline uint64_t string_tail() const { return _in_string ^ _quote; }
+  simdjson_inline uint64_t string_tail() const { return _in_string ^ _quote; }
 
   // backslash characters
   uint64_t _backslash;
@@ -3287,14 +3293,14 @@ struct json_string_block {
 // Scans blocks for string characters, storing the state necessary to do so
 class json_string_scanner {
 public:
-  simdjson_really_inline json_string_block next(const simd::simd8x64<uint8_t>& in);
+  simdjson_inline json_string_block next(const simd::simd8x64<uint8_t>& in);
   // Returns either UNCLOSED_STRING or SUCCESS
-  simdjson_really_inline error_code finish();
+  simdjson_inline error_code finish();
 
 private:
   // Intended to be defined by the implementation
-  simdjson_really_inline uint64_t find_escaped(uint64_t escape);
-  simdjson_really_inline uint64_t find_escaped_branchless(uint64_t escape);
+  simdjson_inline uint64_t find_escaped(uint64_t escape);
+  simdjson_inline uint64_t find_escaped_branchless(uint64_t escape);
 
   // Whether the last iteration was still inside a string (all 1's = true, all 0's = false).
   uint64_t prev_in_string = 0ULL;
@@ -3329,7 +3335,7 @@ class json_string_scanner {
 // desired        |   x  | x x  x x  x x  x  x  |
 // text           |  \\\ | \\\"\\\" \\\" \\"\\" |
 //
-simdjson_really_inline uint64_t json_string_scanner::find_escaped_branchless(uint64_t backslash) {
+simdjson_inline uint64_t json_string_scanner::find_escaped_branchless(uint64_t backslash) {
   // If there was overflow, pretend the first character isn't a backslash
   backslash &= ~prev_escaped;
   uint64_t follows_escape = backslash << 1 | prev_escaped;
@@ -3354,7 +3360,7 @@ simdjson_really_inline uint64_t json_string_scanner::find_escaped_branchless(uin
 //
 // Backslash sequences outside of quotes will be detected in stage 2.
 //
-simdjson_really_inline json_string_block json_string_scanner::next(const simd::simd8x64<uint8_t>& in) {
+simdjson_inline json_string_block json_string_scanner::next(const simd::simd8x64<uint8_t>& in) {
   const uint64_t backslash = in.eq('\\');
   const uint64_t escaped = find_escaped(backslash);
   const uint64_t quote = in.eq('"') & ~escaped;
@@ -3387,7 +3393,7 @@ simdjson_really_inline json_string_block json_string_scanner::next(const simd::s
   );
 }
 
-simdjson_really_inline error_code json_string_scanner::finish() {
+simdjson_inline error_code json_string_scanner::finish() {
   if (prev_in_string) {
     return UNCLOSED_STRING;
   }
@@ -3425,25 +3431,25 @@ namespace stage1 {
 struct json_block {
 public:
   // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
-  simdjson_really_inline json_block(json_string_block&& string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  simdjson_inline json_block(json_string_block&& string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
   _string(std::move(string)), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
-  simdjson_really_inline json_block(json_string_block string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  simdjson_inline json_block(json_string_block string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
   _string(string), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
 
   /**
    * The start of structurals.
    * In simdjson prior to v0.3, these were called the pseudo-structural characters.
    **/
-  simdjson_really_inline uint64_t structural_start() const noexcept { return potential_structural_start() & ~_string.string_tail(); }
+  simdjson_inline uint64_t structural_start() const noexcept { return potential_structural_start() & ~_string.string_tail(); }
   /** All JSON whitespace (i.e. not in a string) */
-  simdjson_really_inline uint64_t whitespace() const noexcept { return non_quote_outside_string(_characters.whitespace()); }
+  simdjson_inline uint64_t whitespace() const noexcept { return non_quote_outside_string(_characters.whitespace()); }
 
   // Helpers
 
   /** Whether the given characters are inside a string (only works on non-quotes) */
-  simdjson_really_inline uint64_t non_quote_inside_string(uint64_t mask) const noexcept { return _string.non_quote_inside_string(mask); }
+  simdjson_inline uint64_t non_quote_inside_string(uint64_t mask) const noexcept { return _string.non_quote_inside_string(mask); }
   /** Whether the given characters are outside a string (only works on non-quotes) */
-  simdjson_really_inline uint64_t non_quote_outside_string(uint64_t mask) const noexcept { return _string.non_quote_outside_string(mask); }
+  simdjson_inline uint64_t non_quote_outside_string(uint64_t mask) const noexcept { return _string.non_quote_outside_string(mask); }
 
   // string and escape characters
   json_string_block _string;
@@ -3458,12 +3464,12 @@ struct json_block {
    * structural elements ([,],{,},:, comma) plus scalar starts like 123, true and "abc".
    * They may reside inside a string.
    **/
-  simdjson_really_inline uint64_t potential_structural_start() const noexcept { return _characters.op() | potential_scalar_start(); }
+  simdjson_inline uint64_t potential_structural_start() const noexcept { return _characters.op() | potential_scalar_start(); }
   /**
    * The start of non-operator runs, like 123, true and "abc".
    * It main reside inside a string.
    **/
-  simdjson_really_inline uint64_t potential_scalar_start() const noexcept {
+  simdjson_inline uint64_t potential_scalar_start() const noexcept {
     // The term "scalar" refers to anything except structural characters and white space
     // (so letters, numbers, quotes).
     // Whenever it is preceded by something that is not a structural element ({,},[,],:, ") nor a white-space
@@ -3474,7 +3480,7 @@ struct json_block {
    * Whether the given character is immediately after a non-operator like 123, true.
    * The characters following a quote are not included.
    */
-  simdjson_really_inline uint64_t follows_potential_scalar() const noexcept {
+  simdjson_inline uint64_t follows_potential_scalar() const noexcept {
     // _follows_potential_nonquote_scalar: is defined as marking any character that follows a character
     // that is not a structural element ({,},[,],:, comma) nor a quote (") and that is not a
     // white space.
@@ -3498,10 +3504,10 @@ struct json_block {
  */
 class json_scanner {
 public:
-  json_scanner() {}
-  simdjson_really_inline json_block next(const simd::simd8x64<uint8_t>& in);
+  json_scanner() = default;
+  simdjson_inline json_block next(const simd::simd8x64<uint8_t>& in);
   // Returns either UNCLOSED_STRING or SUCCESS
-  simdjson_really_inline error_code finish();
+  simdjson_inline error_code finish();
 
 private:
   // Whether the last character of the previous iteration is part of a scalar token
@@ -3518,13 +3524,13 @@ class json_scanner {
 //
 //     const uint64_t backslashed_quote = in.eq('"') & immediately_follows(in.eq('\'), prev_backslash);
 //
-simdjson_really_inline uint64_t follows(const uint64_t match, uint64_t &overflow) {
+simdjson_inline uint64_t follows(const uint64_t match, uint64_t &overflow) {
   const uint64_t result = match << 1 | overflow;
   overflow = match >> 63;
   return result;
 }
 
-simdjson_really_inline json_block json_scanner::next(const simd::simd8x64<uint8_t>& in) {
+simdjson_inline json_block json_scanner::next(const simd::simd8x64<uint8_t>& in) {
   json_string_block strings = string_scanner.next(in);
   // identifies the white-space and the structural characters
   json_character_block characters = json_character_block::classify(in);
@@ -3549,7 +3555,7 @@ simdjson_really_inline json_block json_scanner::next(const simd::simd8x64<uint8_
   );
 }
 
-simdjson_really_inline error_code json_scanner::finish() {
+simdjson_inline error_code json_scanner::finish() {
   return string_scanner.finish();
 }
 
@@ -3575,23 +3581,23 @@ class json_minifier {
   static error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept;
 
 private:
-  simdjson_really_inline json_minifier(uint8_t *_dst)
+  simdjson_inline json_minifier(uint8_t *_dst)
   : dst{_dst}
   {}
   template<size_t STEP_SIZE>
-  simdjson_really_inline void step(const uint8_t *block_buf, buf_block_reader<STEP_SIZE> &reader) noexcept;
-  simdjson_really_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block);
-  simdjson_really_inline error_code finish(uint8_t *dst_start, size_t &dst_len);
+  simdjson_inline void step(const uint8_t *block_buf, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block);
+  simdjson_inline error_code finish(uint8_t *dst_start, size_t &dst_len);
   json_scanner scanner{};
   uint8_t *dst;
 };
 
-simdjson_really_inline void json_minifier::next(const simd::simd8x64<uint8_t>& in, const json_block& block) {
+simdjson_inline void json_minifier::next(const simd::simd8x64<uint8_t>& in, const json_block& block) {
   uint64_t mask = block.whitespace();
   dst += in.compress(mask, dst);
 }
 
-simdjson_really_inline error_code json_minifier::finish(uint8_t *dst_start, size_t &dst_len) {
+simdjson_inline error_code json_minifier::finish(uint8_t *dst_start, size_t &dst_len) {
   error_code error = scanner.finish();
   if (error) { dst_len = 0; return error; }
   dst_len = dst - dst_start;
@@ -3599,7 +3605,7 @@ simdjson_really_inline error_code json_minifier::finish(uint8_t *dst_start, size
 }
 
 template<>
-simdjson_really_inline void json_minifier::step<128>(const uint8_t *block_buf, buf_block_reader<128> &reader) noexcept {
+simdjson_inline void json_minifier::step<128>(const uint8_t *block_buf, buf_block_reader<128> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block_buf);
   simd::simd8x64<uint8_t> in_2(block_buf+64);
   json_block block_1 = scanner.next(in_1);
@@ -3610,7 +3616,7 @@ simdjson_really_inline void json_minifier::step<128>(const uint8_t *block_buf, b
 }
 
 template<>
-simdjson_really_inline void json_minifier::step<64>(const uint8_t *block_buf, buf_block_reader<64> &reader) noexcept {
+simdjson_inline void json_minifier::step<64>(const uint8_t *block_buf, buf_block_reader<64> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block_buf);
   json_block block_1 = scanner.next(in_1);
   this->next(block_buf, block_1);
@@ -3683,7 +3689,7 @@ namespace {
   * complete document, therefore the last json buffer location is the end of the
   * batch.
   */
-simdjson_really_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
+simdjson_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
   // Variant: do not count separately, just figure out depth
   if(parser.n_structural_indexes == 0) { return 0; }
   auto arr_cnt = 0;
@@ -3758,7 +3764,7 @@ class bit_indexer {
 public:
   uint32_t *tail;
 
-  simdjson_really_inline bit_indexer(uint32_t *index_buf) : tail(index_buf) {}
+  simdjson_inline bit_indexer(uint32_t *index_buf) : tail(index_buf) {}
 
   // flatten out values in 'bits' assuming that they are are to have values of idx
   // plus their position in the bitvector, and store these indexes at
@@ -3769,9 +3775,9 @@ class bit_indexer {
   // If the kernel sets SIMDJSON_CUSTOM_BIT_INDEXER, then it will provide its own
   // version of the code.
 #ifdef SIMDJSON_CUSTOM_BIT_INDEXER
-  simdjson_really_inline void write(uint32_t idx, uint64_t bits);
+  simdjson_inline void write(uint32_t idx, uint64_t bits);
 #else
-  simdjson_really_inline void write(uint32_t idx, uint64_t bits) {
+  simdjson_inline void write(uint32_t idx, uint64_t bits) {
     // In some instances, the next branch is expensive because it is mispredicted.
     // Unfortunately, in other cases,
     // it helps tremendously.
@@ -3880,11 +3886,11 @@ class json_structural_indexer {
   static error_code index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept;
 
 private:
-  simdjson_really_inline json_structural_indexer(uint32_t *structural_indexes);
+  simdjson_inline json_structural_indexer(uint32_t *structural_indexes);
   template<size_t STEP_SIZE>
-  simdjson_really_inline void step(const uint8_t *block, buf_block_reader<STEP_SIZE> &reader) noexcept;
-  simdjson_really_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx);
-  simdjson_really_inline error_code finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial);
+  simdjson_inline void step(const uint8_t *block, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx);
+  simdjson_inline error_code finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial);
 
   json_scanner scanner{};
   utf8_checker checker{};
@@ -3893,26 +3899,26 @@ class json_structural_indexer {
   uint64_t unescaped_chars_error = 0;
 };
 
-simdjson_really_inline json_structural_indexer::json_structural_indexer(uint32_t *structural_indexes) : indexer{structural_indexes} {}
+simdjson_inline json_structural_indexer::json_structural_indexer(uint32_t *structural_indexes) : indexer{structural_indexes} {}
 
 // Skip the last character if it is partial
-simdjson_really_inline size_t trim_partial_utf8(const uint8_t *buf, size_t len) {
+simdjson_inline size_t trim_partial_utf8(const uint8_t *buf, size_t len) {
   if (simdjson_unlikely(len < 3)) {
     switch (len) {
       case 2:
-        if (buf[len-1] >= 0b11000000) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
-        if (buf[len-2] >= 0b11100000) { return len-2; } // 3- and 4-byte characters with only 2 bytes left
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 2 bytes left
         return len;
       case 1:
-        if (buf[len-1] >= 0b11000000) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
         return len;
       case 0:
         return len;
     }
   }
-  if (buf[len-1] >= 0b11000000) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
-  if (buf[len-2] >= 0b11100000) { return len-2; } // 3- and 4-byte characters with only 1 byte left
-  if (buf[len-3] >= 0b11110000) { return len-3; } // 4-byte characters with only 3 bytes left
+  if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+  if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 1 byte left
+  if (buf[len-3] >= 0xf0) { return len-3; } // 4-byte characters with only 3 bytes left
   return len;
 }
 
@@ -3961,7 +3967,7 @@ error_code json_structural_indexer::index(const uint8_t *buf, size_t len, dom_pa
 }
 
 template<>
-simdjson_really_inline void json_structural_indexer::step<128>(const uint8_t *block, buf_block_reader<128> &reader) noexcept {
+simdjson_inline void json_structural_indexer::step<128>(const uint8_t *block, buf_block_reader<128> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block);
   simd::simd8x64<uint8_t> in_2(block+64);
   json_block block_1 = scanner.next(in_1);
@@ -3972,14 +3978,14 @@ simdjson_really_inline void json_structural_indexer::step<128>(const uint8_t *bl
 }
 
 template<>
-simdjson_really_inline void json_structural_indexer::step<64>(const uint8_t *block, buf_block_reader<64> &reader) noexcept {
+simdjson_inline void json_structural_indexer::step<64>(const uint8_t *block, buf_block_reader<64> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block);
   json_block block_1 = scanner.next(in_1);
   this->next(in_1, block_1, reader.block_index());
   reader.advance();
 }
 
-simdjson_really_inline void json_structural_indexer::next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx) {
+simdjson_inline void json_structural_indexer::next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx) {
   uint64_t unescaped = in.lteq(0x1F);
   checker.check_next_input(in);
   indexer.write(uint32_t(idx-64), prev_structurals); // Output *last* iteration's structurals to the parser
@@ -3987,7 +3993,7 @@ simdjson_really_inline void json_structural_indexer::next(const simd::simd8x64<u
   unescaped_chars_error |= block.non_quote_inside_string(unescaped);
 }
 
-simdjson_really_inline error_code json_structural_indexer::finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial) {
+simdjson_inline error_code json_structural_indexer::finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial) {
   // Write out the final iteration's structurals
   indexer.write(uint32_t(idx-64), prev_structurals);
   error_code error = scanner.finish();
@@ -4134,6 +4140,151 @@ bool generic_validate_utf8(const char * input, size_t length) {
 // Stage 2
 //
 
+/* begin file src/generic/stage2/stringparsing.h */
+// This file contains the common code every implementation uses
+// It is intended to be included multiple times and compiled multiple times
+
+namespace simdjson {
+namespace arm64 {
+namespace {
+/// @private
+namespace stringparsing {
+
+// begin copypasta
+// These chars yield themselves: " \ /
+// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
+// u not handled in this table as it's complex
+static const uint8_t escape_map[256] = {
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
+    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
+    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+};
+
+// handle a unicode codepoint
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr) {
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion isn't valid; we defer the check for this to inside the
+  // multilingual plane check
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) != ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      return false;
+    }
+    uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+
+    // We have already checked that the high surrogate is valid and
+    // (code_point - 0xd800) < 1024.
+    //
+    // Check that code_point_2 is in the range 0xdc00..0xdfff
+    // and that code_point_2 was parsed from valid hex.
+    uint32_t low_bit = code_point_2 - 0xdc00;
+    if (low_bit >> 10) {
+      return false;
+    }
+
+    code_point =
+        (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+    *src_ptr += 6;
+  } else if (code_point >= 0xdc00 && code_point <= 0xdfff) {
+      // If we encounter a low surrogate (not preceded by a high surrogate)
+      // then we have an error.
+      return false;
+  }
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+/**
+ * Unescape a valid UTF-8 string from src to dst, stopping at a final unescaped quote. There
+ * must be an unescaped quote terminating the string. It returns the final output
+ * position as pointer. In case of error (e.g., the string has bad escaped codes),
+ * then null_nullptrptr is returned. It is assumed that the output buffer is large
+ * enough. E.g., if src points at 'joe"', then dst needs to have four free bytes +
+ * SIMDJSON_PADDING bytes.
+ */
+simdjson_warn_unused simdjson_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst) {
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto bs_quote = backslash_and_quote::copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint(&src, &dst)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+  /* can't be reached */
+  return nullptr;
+}
+
+} // namespace stringparsing
+} // unnamed namespace
+} // namespace arm64
+} // namespace simdjson
+/* end file src/generic/stage2/stringparsing.h */
 /* begin file src/generic/stage2/tape_builder.h */
 /* begin file src/generic/stage2/json_iterator.h */
 /* begin file src/generic/stage2/logger.h */
@@ -4159,7 +4310,7 @@ namespace logger {
   static int log_depth; // Not threadsafe. Log only.
 
   // Helper to turn unprintable or newline characters into spaces
-  static simdjson_really_inline char printable_char(char c) {
+  static simdjson_inline char printable_char(char c) {
     if (c >= 0x20) {
       return c;
     } else {
@@ -4168,7 +4319,7 @@ namespace logger {
   }
 
   // Print the header and set up log_start
-  static simdjson_really_inline void log_start() {
+  static simdjson_inline void log_start() {
     if (LOG_ENABLED) {
       log_depth = 0;
       printf("\n");
@@ -4177,7 +4328,7 @@ namespace logger {
     }
   }
 
-  simdjson_unused static simdjson_really_inline void log_string(const char *message) {
+  simdjson_unused static simdjson_inline void log_string(const char *message) {
     if (LOG_ENABLED) {
       printf("%s\n", message);
     }
@@ -4185,7 +4336,7 @@ namespace logger {
 
   // Logs a single line from the stage 2 DOM parser
   template<typename S>
-  static simdjson_really_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
+  static simdjson_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
     if (LOG_ENABLED) {
       printf("| %*s%s%-*s ", log_depth*2, "", title_prefix, LOG_EVENT_LEN - log_depth*2 - int(strlen(title_prefix)), title);
       auto current_index = structurals.at_beginning() ? nullptr : structurals.next_structural-1;
@@ -4263,14 +4414,14 @@ class json_iterator {
    * - increment_count(iter) - each time a value is found in an array or object.
    */
   template<bool STREAMING, typename V>
-  simdjson_warn_unused simdjson_really_inline error_code walk_document(V &visitor) noexcept;
+  simdjson_warn_unused simdjson_inline error_code walk_document(V &visitor) noexcept;
 
   /**
    * Create an iterator capable of walking a JSON document.
    *
    * The document must have already passed through stage 1.
    */
-  simdjson_really_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
+  simdjson_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
 
   /**
    * Look at the next token.
@@ -4279,7 +4430,7 @@ class json_iterator {
    *
    * They may include invalid JSON as well (such as `1.2.3` or `ture`).
    */
-  simdjson_really_inline const uint8_t *peek() const noexcept;
+  simdjson_inline const uint8_t *peek() const noexcept;
   /**
    * Advance to the next token.
    *
@@ -4287,56 +4438,56 @@ class json_iterator {
    *
    * They may include invalid JSON as well (such as `1.2.3` or `ture`).
    */
-  simdjson_really_inline const uint8_t *advance() noexcept;
+  simdjson_inline const uint8_t *advance() noexcept;
   /**
    * Get the remaining length of the document, from the start of the current token.
    */
-  simdjson_really_inline size_t remaining_len() const noexcept;
+  simdjson_inline size_t remaining_len() const noexcept;
   /**
    * Check if we are at the end of the document.
    *
    * If this is true, there are no more tokens.
    */
-  simdjson_really_inline bool at_eof() const noexcept;
+  simdjson_inline bool at_eof() const noexcept;
   /**
    * Check if we are at the beginning of the document.
    */
-  simdjson_really_inline bool at_beginning() const noexcept;
-  simdjson_really_inline uint8_t last_structural() const noexcept;
+  simdjson_inline bool at_beginning() const noexcept;
+  simdjson_inline uint8_t last_structural() const noexcept;
 
   /**
    * Log that a value has been found.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_value(const char *type) const noexcept;
+  simdjson_inline void log_value(const char *type) const noexcept;
   /**
    * Log the start of a multipart value.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_start_value(const char *type) const noexcept;
+  simdjson_inline void log_start_value(const char *type) const noexcept;
   /**
    * Log the end of a multipart value.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_end_value(const char *type) const noexcept;
+  simdjson_inline void log_end_value(const char *type) const noexcept;
   /**
    * Log an error.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_error(const char *error) const noexcept;
+  simdjson_inline void log_error(const char *error) const noexcept;
 
   template<typename V>
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
   template<typename V>
-  simdjson_warn_unused simdjson_really_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
 };
 
 template<bool STREAMING, typename V>
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::walk_document(V &visitor) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::walk_document(V &visitor) noexcept {
   logger::log_start();
 
   //
@@ -4461,52 +4612,52 @@ simdjson_warn_unused simdjson_really_inline error_code json_iterator::walk_docum
 
 } // walk_document()
 
-simdjson_really_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+simdjson_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
   : buf{_dom_parser.buf},
     next_structural{&_dom_parser.structural_indexes[start_structural_index]},
     dom_parser{_dom_parser} {
 }
 
-simdjson_really_inline const uint8_t *json_iterator::peek() const noexcept {
+simdjson_inline const uint8_t *json_iterator::peek() const noexcept {
   return &buf[*(next_structural)];
 }
-simdjson_really_inline const uint8_t *json_iterator::advance() noexcept {
+simdjson_inline const uint8_t *json_iterator::advance() noexcept {
   return &buf[*(next_structural++)];
 }
-simdjson_really_inline size_t json_iterator::remaining_len() const noexcept {
+simdjson_inline size_t json_iterator::remaining_len() const noexcept {
   return dom_parser.len - *(next_structural-1);
 }
 
-simdjson_really_inline bool json_iterator::at_eof() const noexcept {
+simdjson_inline bool json_iterator::at_eof() const noexcept {
   return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
 }
-simdjson_really_inline bool json_iterator::at_beginning() const noexcept {
+simdjson_inline bool json_iterator::at_beginning() const noexcept {
   return next_structural == dom_parser.structural_indexes.get();
 }
-simdjson_really_inline uint8_t json_iterator::last_structural() const noexcept {
+simdjson_inline uint8_t json_iterator::last_structural() const noexcept {
   return buf[dom_parser.structural_indexes[dom_parser.n_structural_indexes - 1]];
 }
 
-simdjson_really_inline void json_iterator::log_value(const char *type) const noexcept {
+simdjson_inline void json_iterator::log_value(const char *type) const noexcept {
   logger::log_line(*this, "", type, "");
 }
 
-simdjson_really_inline void json_iterator::log_start_value(const char *type) const noexcept {
+simdjson_inline void json_iterator::log_start_value(const char *type) const noexcept {
   logger::log_line(*this, "+", type, "");
   if (logger::LOG_ENABLED) { logger::log_depth++; }
 }
 
-simdjson_really_inline void json_iterator::log_end_value(const char *type) const noexcept {
+simdjson_inline void json_iterator::log_end_value(const char *type) const noexcept {
   if (logger::LOG_ENABLED) { logger::log_depth--; }
   logger::log_line(*this, "-", type, "");
 }
 
-simdjson_really_inline void json_iterator::log_error(const char *error) const noexcept {
+simdjson_inline void json_iterator::log_error(const char *error) const noexcept {
   logger::log_line(*this, "", "ERROR", error);
 }
 
 template<typename V>
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
   switch (*value) {
     case '"': return visitor.visit_root_string(*this, value);
     case 't': return visitor.visit_root_true_atom(*this, value);
@@ -4522,7 +4673,7 @@ simdjson_warn_unused simdjson_really_inline error_code json_iterator::visit_root
   }
 }
 template<typename V>
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
   switch (*value) {
     case '"': return visitor.visit_string(*this, value);
     case 't': return visitor.visit_true_atom(*this, value);
@@ -4554,18 +4705,18 @@ struct tape_writer {
   uint64_t *next_tape_loc;
 
   /** Write a signed 64-bit value to tape. */
-  simdjson_really_inline void append_s64(int64_t value) noexcept;
+  simdjson_inline void append_s64(int64_t value) noexcept;
 
   /** Write an unsigned 64-bit value to tape. */
-  simdjson_really_inline void append_u64(uint64_t value) noexcept;
+  simdjson_inline void append_u64(uint64_t value) noexcept;
 
   /** Write a double value to tape. */
-  simdjson_really_inline void append_double(double value) noexcept;
+  simdjson_inline void append_double(double value) noexcept;
 
   /**
    * Append a tape entry (an 8-bit type,and 56 bits worth of value).
    */
-  simdjson_really_inline void append(uint64_t val, internal::tape_type t) noexcept;
+  simdjson_inline void append(uint64_t val, internal::tape_type t) noexcept;
 
   /**
    * Skip the current tape entry without writing.
@@ -4573,24 +4724,24 @@ struct tape_writer {
    * Used to skip the start of the container, since we'll come back later to fill it in when the
    * container ends.
    */
-  simdjson_really_inline void skip() noexcept;
+  simdjson_inline void skip() noexcept;
 
   /**
    * Skip the number of tape entries necessary to write a large u64 or i64.
    */
-  simdjson_really_inline void skip_large_integer() noexcept;
+  simdjson_inline void skip_large_integer() noexcept;
 
   /**
    * Skip the number of tape entries necessary to write a double.
    */
-  simdjson_really_inline void skip_double() noexcept;
+  simdjson_inline void skip_double() noexcept;
 
   /**
    * Write a value to a known location on tape.
    *
    * Used to go back and write out the start of a container after the container ends.
    */
-  simdjson_really_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
+  simdjson_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
 
 private:
   /**
@@ -4598,50 +4749,50 @@ struct tape_writer {
    * all 64 bits, such as double and uint64_t.
    */
   template<typename T>
-  simdjson_really_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
+  simdjson_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
 }; // struct number_writer
 
-simdjson_really_inline void tape_writer::append_s64(int64_t value) noexcept {
+simdjson_inline void tape_writer::append_s64(int64_t value) noexcept {
   append2(0, value, internal::tape_type::INT64);
 }
 
-simdjson_really_inline void tape_writer::append_u64(uint64_t value) noexcept {
+simdjson_inline void tape_writer::append_u64(uint64_t value) noexcept {
   append(0, internal::tape_type::UINT64);
   *next_tape_loc = value;
   next_tape_loc++;
 }
 
 /** Write a double value to tape. */
-simdjson_really_inline void tape_writer::append_double(double value) noexcept {
+simdjson_inline void tape_writer::append_double(double value) noexcept {
   append2(0, value, internal::tape_type::DOUBLE);
 }
 
-simdjson_really_inline void tape_writer::skip() noexcept {
+simdjson_inline void tape_writer::skip() noexcept {
   next_tape_loc++;
 }
 
-simdjson_really_inline void tape_writer::skip_large_integer() noexcept {
+simdjson_inline void tape_writer::skip_large_integer() noexcept {
   next_tape_loc += 2;
 }
 
-simdjson_really_inline void tape_writer::skip_double() noexcept {
+simdjson_inline void tape_writer::skip_double() noexcept {
   next_tape_loc += 2;
 }
 
-simdjson_really_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
+simdjson_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
   *next_tape_loc = val | ((uint64_t(char(t))) << 56);
   next_tape_loc++;
 }
 
 template<typename T>
-simdjson_really_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
+simdjson_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
   append(val, t);
   static_assert(sizeof(val2) == sizeof(*next_tape_loc), "Type is not 64 bits!");
   memcpy(next_tape_loc, &val2, sizeof(val2));
   next_tape_loc++;
 }
 
-simdjson_really_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
+simdjson_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
   tape_loc = val | ((uint64_t(char(t))) << 56);
 }
 
@@ -4658,40 +4809,40 @@ namespace stage2 {
 
 struct tape_builder {
   template<bool STREAMING>
-  simdjson_warn_unused static simdjson_really_inline error_code parse_document(
+  simdjson_warn_unused static simdjson_inline error_code parse_document(
     dom_parser_implementation &dom_parser,
     dom::document &doc) noexcept;
 
   /** Called when a non-empty document starts. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_document_start(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_document_start(json_iterator &iter) noexcept;
   /** Called when a non-empty document ends without error. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_document_end(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_document_end(json_iterator &iter) noexcept;
 
   /** Called when a non-empty array starts. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_array_start(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_array_start(json_iterator &iter) noexcept;
   /** Called when a non-empty array ends. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_array_end(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_array_end(json_iterator &iter) noexcept;
   /** Called when an empty array is found. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_empty_array(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_empty_array(json_iterator &iter) noexcept;
 
   /** Called when a non-empty object starts. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_object_start(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_object_start(json_iterator &iter) noexcept;
   /**
    * Called when a key in a field is encountered.
    *
    * primitive, visit_object_start, visit_empty_object, visit_array_start, or visit_empty_array
    * will be called after this with the field value.
    */
-  simdjson_warn_unused simdjson_really_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
   /** Called when a non-empty object ends. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_object_end(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_object_end(json_iterator &iter) noexcept;
   /** Called when an empty object is found. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_empty_object(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_empty_object(json_iterator &iter) noexcept;
 
   /**
    * Called when a string, number, boolean or null is found.
    */
-  simdjson_warn_unused simdjson_really_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
   /**
    * Called when a string, number, boolean or null is found at the top level of a document (i.e.
    * when there is no array or object and the entire document is a single string, number, boolean or
@@ -4700,22 +4851,22 @@ struct tape_builder {
    * This is separate from primitive() because simdjson's normal primitive parsing routines assume
    * there is at least one more token after the value, which is only true in an array or object.
    */
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
 
-  simdjson_warn_unused simdjson_really_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
 
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
 
   /** Called each time a new field or element in an array or object is found. */
-  simdjson_warn_unused simdjson_really_inline error_code increment_count(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code increment_count(json_iterator &iter) noexcept;
 
   /** Next location to write to tape */
   tape_writer tape;
@@ -4723,18 +4874,18 @@ struct tape_builder {
   /** Next write location in the string buf for stage 2 parsing */
   uint8_t *current_string_buf_loc;
 
-  simdjson_really_inline tape_builder(dom::document &doc) noexcept;
+  simdjson_inline tape_builder(dom::document &doc) noexcept;
 
-  simdjson_really_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
-  simdjson_really_inline void start_container(json_iterator &iter) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
-  simdjson_really_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
-  simdjson_really_inline void on_end_string(uint8_t *dst) noexcept;
+  simdjson_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
+  simdjson_inline void start_container(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_warn_unused simdjson_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
+  simdjson_inline void on_end_string(uint8_t *dst) noexcept;
 }; // class tape_builder
 
 template<bool STREAMING>
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::parse_document(
+simdjson_warn_unused simdjson_inline error_code tape_builder::parse_document(
     dom_parser_implementation &dom_parser,
     dom::document &doc) noexcept {
   dom_parser.doc = &doc;
@@ -4743,56 +4894,56 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::parse_docum
   return iter.walk_document<STREAMING>(builder);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
   return iter.visit_root_primitive(*this, value);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
   return iter.visit_primitive(*this, value);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
   return empty_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
   return empty_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
   start_container(iter);
   return SUCCESS;
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
   start_container(iter);
   return SUCCESS;
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
   start_container(iter);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
   return end_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
   return end_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
   constexpr uint32_t start_tape_index = 0;
   tape.append(start_tape_index, internal::tape_type::ROOT);
   tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter), internal::tape_type::ROOT);
   return SUCCESS;
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
   return visit_string(iter, key, true);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
   iter.dom_parser.open_containers[iter.depth].count++; // we have a key value pair in the object at parser.dom_parser.depth - 1
   return SUCCESS;
 }
 
-simdjson_really_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
+simdjson_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
   iter.log_value(key ? "key" : "string");
   uint8_t *dst = on_start_string(iter);
   dst = stringparsing::parse_string(value+1, dst);
@@ -4804,16 +4955,16 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_strin
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
   return visit_string(iter, value);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("number");
   return numberparsing::parse_number(value, tape);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
   //
   // We need to make a copy to make sure that the string is space terminated.
   // This is not about padding the input, which should already padded up
@@ -4835,42 +4986,42 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_
   return error;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("true");
   if (!atomparsing::is_valid_true_atom(value)) { return T_ATOM_ERROR; }
   tape.append(0, internal::tape_type::TRUE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("true");
   if (!atomparsing::is_valid_true_atom(value, iter.remaining_len())) { return T_ATOM_ERROR; }
   tape.append(0, internal::tape_type::TRUE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("false");
   if (!atomparsing::is_valid_false_atom(value)) { return F_ATOM_ERROR; }
   tape.append(0, internal::tape_type::FALSE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("false");
   if (!atomparsing::is_valid_false_atom(value, iter.remaining_len())) { return F_ATOM_ERROR; }
   tape.append(0, internal::tape_type::FALSE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("null");
   if (!atomparsing::is_valid_null_atom(value)) { return N_ATOM_ERROR; }
   tape.append(0, internal::tape_type::NULL_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("null");
   if (!atomparsing::is_valid_null_atom(value, iter.remaining_len())) { return N_ATOM_ERROR; }
   tape.append(0, internal::tape_type::NULL_VALUE);
@@ -4879,24 +5030,24 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_
 
 // private:
 
-simdjson_really_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
+simdjson_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
   return uint32_t(tape.next_tape_loc - iter.dom_parser.doc->tape.get());
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
   auto start_index = next_tape_index(iter);
   tape.append(start_index+2, start);
   tape.append(start_index, end);
   return SUCCESS;
 }
 
-simdjson_really_inline void tape_builder::start_container(json_iterator &iter) noexcept {
+simdjson_inline void tape_builder::start_container(json_iterator &iter) noexcept {
   iter.dom_parser.open_containers[iter.depth].tape_index = next_tape_index(iter);
   iter.dom_parser.open_containers[iter.depth].count = 0;
   tape.skip(); // We don't actually *write* the start element until the end.
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
   // Write the ending tape element, pointing at the start location
   const uint32_t start_tape_index = iter.dom_parser.open_containers[iter.depth].tape_index;
   tape.append(start_tape_index, end);
@@ -4909,13 +5060,13 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::end_contain
   return SUCCESS;
 }
 
-simdjson_really_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
+simdjson_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
   // we advance the point, accounting for the fact that we have a NULL termination
   tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::STRING);
   return current_string_buf_loc + sizeof(uint32_t);
 }
 
-simdjson_really_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
+simdjson_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
   uint32_t str_length = uint32_t(dst - (current_string_buf_loc + sizeof(uint32_t)));
   // TODO check for overflow in case someone has a crazy string (>=4GB?)
   // But only add the overflow check when the document itself exceeds 4GB
@@ -4941,7 +5092,7 @@ namespace arm64 {
 namespace {
 namespace stage1 {
 
-simdjson_really_inline uint64_t json_string_scanner::find_escaped(uint64_t backslash) {
+simdjson_inline uint64_t json_string_scanner::find_escaped(uint64_t backslash) {
   // On ARM, we don't short-circuit this if there are no backslashes, because the branch gives us no
   // benefit and therefore makes things worse.
   // if (!backslash) { uint64_t escaped = prev_escaped; prev_escaped = 0; return escaped; }
@@ -4973,6 +5124,10 @@ simdjson_warn_unused error_code dom_parser_implementation::stage2_next(dom::docu
   return stage2::tape_builder::parse_document<true>(*this, _doc);
 }
 
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_string(const uint8_t *src, uint8_t *dst) const noexcept {
+  return arm64::stringparsing::parse_string(src, dst);
+}
+
 simdjson_warn_unused error_code dom_parser_implementation::parse(const uint8_t *_buf, size_t _len, dom::document &_doc) noexcept {
   auto error = stage1(_buf, _len, stage1_mode::regular);
   if (error) { return error; }
@@ -5055,7 +5210,7 @@ namespace {
   * complete document, therefore the last json buffer location is the end of the
   * batch.
   */
-simdjson_really_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
+simdjson_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
   // Variant: do not count separately, just figure out depth
   if(parser.n_structural_indexes == 0) { return 0; }
   auto arr_cnt = 0;
@@ -5129,7 +5284,7 @@ namespace stage1 {
 class structural_scanner {
 public:
 
-simdjson_really_inline structural_scanner(dom_parser_implementation &_parser, stage1_mode _partial)
+simdjson_inline structural_scanner(dom_parser_implementation &_parser, stage1_mode _partial)
   : buf{_parser.buf},
     next_structural_index{_parser.structural_indexes.get()},
     parser{_parser},
@@ -5137,18 +5292,18 @@ simdjson_really_inline structural_scanner(dom_parser_implementation &_parser, st
     partial{_partial} {
 }
 
-simdjson_really_inline void add_structural() {
+simdjson_inline void add_structural() {
   *next_structural_index = idx;
   next_structural_index++;
 }
 
-simdjson_really_inline bool is_continuation(uint8_t c) {
-  return (c & 0b11000000) == 0b10000000;
+simdjson_inline bool is_continuation(uint8_t c) {
+  return (c & 0xc0) == 0x80;
 }
 
-simdjson_really_inline void validate_utf8_character() {
+simdjson_inline void validate_utf8_character() {
   // Continuation
-  if (simdjson_unlikely((buf[idx] & 0b01000000) == 0)) {
+  if (simdjson_unlikely((buf[idx] & 0x40) == 0)) {
     // extra continuation
     error = UTF8_ERROR;
     idx++;
@@ -5156,7 +5311,7 @@ simdjson_really_inline void validate_utf8_character() {
   }
 
   // 2-byte
-  if ((buf[idx] & 0b00100000) == 0) {
+  if ((buf[idx] & 0x20) == 0) {
     // missing continuation
     if (simdjson_unlikely(idx+1 > len || !is_continuation(buf[idx+1]))) {
       if (idx+1 > len && is_streaming(partial)) { idx = len; return; }
@@ -5165,13 +5320,13 @@ simdjson_really_inline void validate_utf8_character() {
       return;
     }
     // overlong: 1100000_ 10______
-    if (buf[idx] <= 0b11000001) { error = UTF8_ERROR; }
+    if (buf[idx] <= 0xc1) { error = UTF8_ERROR; }
     idx += 2;
     return;
   }
 
   // 3-byte
-  if ((buf[idx] & 0b00010000) == 0) {
+  if ((buf[idx] & 0x10) == 0) {
     // missing continuation
     if (simdjson_unlikely(idx+2 > len || !is_continuation(buf[idx+1]) || !is_continuation(buf[idx+2]))) {
       if (idx+2 > len && is_streaming(partial)) { idx = len; return; }
@@ -5180,9 +5335,9 @@ simdjson_really_inline void validate_utf8_character() {
       return;
     }
     // overlong: 11100000 100_____ ________
-    if (buf[idx] == 0b11100000 && buf[idx+1] <= 0b10011111) { error = UTF8_ERROR; }
+    if (buf[idx] == 0xe0 && buf[idx+1] <= 0x9f) { error = UTF8_ERROR; }
     // surrogates: U+D800-U+DFFF 11101101 101_____
-    if (buf[idx] == 0b11101101 && buf[idx+1] >= 0b10100000) { error = UTF8_ERROR; }
+    if (buf[idx] == 0xed && buf[idx+1] >= 0xa0) { error = UTF8_ERROR; }
     idx += 3;
     return;
   }
@@ -5196,24 +5351,24 @@ simdjson_really_inline void validate_utf8_character() {
     return;
   }
   // overlong: 11110000 1000____ ________ ________
-  if (buf[idx] == 0b11110000 && buf[idx+1] <= 0b10001111) { error = UTF8_ERROR; }
+  if (buf[idx] == 0xf0 && buf[idx+1] <= 0x8f) { error = UTF8_ERROR; }
   // too large: > U+10FFFF:
   // 11110100 (1001|101_)____
   // 1111(1___|011_|0101) 10______
   // also includes 5, 6, 7 and 8 byte characters:
   // 11111___
-  if (buf[idx] == 0b11110100 && buf[idx+1] >= 0b10010000) { error = UTF8_ERROR; }
-  if (buf[idx] >= 0b11110101) { error = UTF8_ERROR; }
+  if (buf[idx] == 0xf4 && buf[idx+1] >= 0x90) { error = UTF8_ERROR; }
+  if (buf[idx] >= 0xf5) { error = UTF8_ERROR; }
   idx += 4;
 }
 
 // Returns true if the string is unclosed.
-simdjson_really_inline bool validate_string() {
+simdjson_inline bool validate_string() {
   idx++; // skip first quote
   while (idx < len && buf[idx] != '"') {
     if (buf[idx] == '\\') {
       idx += 2;
-    } else if (simdjson_unlikely(buf[idx] & 0b10000000)) {
+    } else if (simdjson_unlikely(buf[idx] & 0x80)) {
       validate_utf8_character();
     } else {
       if (buf[idx] < 0x20) { error = UNESCAPED_CHARS; }
@@ -5224,7 +5379,7 @@ simdjson_really_inline bool validate_string() {
   return false;
 }
 
-simdjson_really_inline bool is_whitespace_or_operator(uint8_t c) {
+simdjson_inline bool is_whitespace_or_operator(uint8_t c) {
   switch (c) {
     case '{': case '}': case '[': case ']': case ',': case ':':
     case ' ': case '\r': case '\n': case '\t':
@@ -5237,7 +5392,7 @@ simdjson_really_inline bool is_whitespace_or_operator(uint8_t c) {
 //
 // Parse the entire input in STEP_SIZE-byte chunks.
 //
-simdjson_really_inline error_code scan() {
+simdjson_inline error_code scan() {
   bool unclosed_string = false;
   for (;idx<len;idx++) {
     switch (buf[idx]) {
@@ -5413,39 +5568,39 @@ simdjson_warn_unused bool implementation::validate_utf8(const char *buf, size_t
       }
     }
     unsigned char byte = data[pos];
-    if (byte < 0b10000000) {
+    if (byte < 0x80) {
       pos++;
       continue;
-    } else if ((byte & 0b11100000) == 0b11000000) {
+    } else if ((byte & 0xe0) == 0xc0) {
       next_pos = pos + 2;
       if (next_pos > len) { return false; }
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return false; }
+      if ((data[pos + 1] & 0xc0) != 0x80) { return false; }
       // range check
-      code_point = (byte & 0b00011111) << 6 | (data[pos + 1] & 0b00111111);
+      code_point = (byte & 0x1f) << 6 | (data[pos + 1] & 0x3f);
       if (code_point < 0x80 || 0x7ff < code_point) { return false; }
-    } else if ((byte & 0b11110000) == 0b11100000) {
+    } else if ((byte & 0xf0) == 0xe0) {
       next_pos = pos + 3;
       if (next_pos > len) { return false; }
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return false; }
-      if ((data[pos + 2] & 0b11000000) != 0b10000000) { return false; }
+      if ((data[pos + 1] & 0xc0) != 0x80) { return false; }
+      if ((data[pos + 2] & 0xc0) != 0x80) { return false; }
       // range check
-      code_point = (byte & 0b00001111) << 12 |
-                   (data[pos + 1] & 0b00111111) << 6 |
-                   (data[pos + 2] & 0b00111111);
+      code_point = (byte & 0x0f) << 12 |
+                   (data[pos + 1] & 0x3f) << 6 |
+                   (data[pos + 2] & 0x3f);
       if (code_point < 0x800 || 0xffff < code_point ||
           (0xd7ff < code_point && code_point < 0xe000)) {
         return false;
       }
-    } else if ((byte & 0b11111000) == 0b11110000) { // 0b11110000
+    } else if ((byte & 0xf8) == 0xf0) { // 0b11110000
       next_pos = pos + 4;
       if (next_pos > len) { return false; }
-      if ((data[pos + 1] & 0b11000000) != 0b10000000) { return false; }
-      if ((data[pos + 2] & 0b11000000) != 0b10000000) { return false; }
-      if ((data[pos + 3] & 0b11000000) != 0b10000000) { return false; }
+      if ((data[pos + 1] & 0xc0) != 0x80) { return false; }
+      if ((data[pos + 2] & 0xc0) != 0x80) { return false; }
+      if ((data[pos + 3] & 0xc0) != 0x80) { return false; }
       // range check
       code_point =
-          (byte & 0b00000111) << 18 | (data[pos + 1] & 0b00111111) << 12 |
-          (data[pos + 2] & 0b00111111) << 6 | (data[pos + 3] & 0b00111111);
+          (byte & 0x07) << 18 | (data[pos + 1] & 0x3f) << 12 |
+          (data[pos + 2] & 0x3f) << 6 | (data[pos + 3] & 0x3f);
       if (code_point <= 0xffff || 0x10ffff < code_point) { return false; }
     } else {
       // we may have a continuation
@@ -5462,6 +5617,151 @@ simdjson_warn_unused bool implementation::validate_utf8(const char *buf, size_t
 //
 // Stage 2
 //
+/* begin file src/generic/stage2/stringparsing.h */
+// This file contains the common code every implementation uses
+// It is intended to be included multiple times and compiled multiple times
+
+namespace simdjson {
+namespace fallback {
+namespace {
+/// @private
+namespace stringparsing {
+
+// begin copypasta
+// These chars yield themselves: " \ /
+// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
+// u not handled in this table as it's complex
+static const uint8_t escape_map[256] = {
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
+    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
+    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+};
+
+// handle a unicode codepoint
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr) {
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion isn't valid; we defer the check for this to inside the
+  // multilingual plane check
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) != ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      return false;
+    }
+    uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+
+    // We have already checked that the high surrogate is valid and
+    // (code_point - 0xd800) < 1024.
+    //
+    // Check that code_point_2 is in the range 0xdc00..0xdfff
+    // and that code_point_2 was parsed from valid hex.
+    uint32_t low_bit = code_point_2 - 0xdc00;
+    if (low_bit >> 10) {
+      return false;
+    }
+
+    code_point =
+        (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+    *src_ptr += 6;
+  } else if (code_point >= 0xdc00 && code_point <= 0xdfff) {
+      // If we encounter a low surrogate (not preceded by a high surrogate)
+      // then we have an error.
+      return false;
+  }
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+/**
+ * Unescape a valid UTF-8 string from src to dst, stopping at a final unescaped quote. There
+ * must be an unescaped quote terminating the string. It returns the final output
+ * position as pointer. In case of error (e.g., the string has bad escaped codes),
+ * then null_nullptrptr is returned. It is assumed that the output buffer is large
+ * enough. E.g., if src points at 'joe"', then dst needs to have four free bytes +
+ * SIMDJSON_PADDING bytes.
+ */
+simdjson_warn_unused simdjson_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst) {
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto bs_quote = backslash_and_quote::copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint(&src, &dst)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+  /* can't be reached */
+  return nullptr;
+}
+
+} // namespace stringparsing
+} // unnamed namespace
+} // namespace fallback
+} // namespace simdjson
+/* end file src/generic/stage2/stringparsing.h */
 /* begin file src/generic/stage2/tape_builder.h */
 /* begin file src/generic/stage2/json_iterator.h */
 /* begin file src/generic/stage2/logger.h */
@@ -5487,7 +5787,7 @@ namespace logger {
   static int log_depth; // Not threadsafe. Log only.
 
   // Helper to turn unprintable or newline characters into spaces
-  static simdjson_really_inline char printable_char(char c) {
+  static simdjson_inline char printable_char(char c) {
     if (c >= 0x20) {
       return c;
     } else {
@@ -5496,7 +5796,7 @@ namespace logger {
   }
 
   // Print the header and set up log_start
-  static simdjson_really_inline void log_start() {
+  static simdjson_inline void log_start() {
     if (LOG_ENABLED) {
       log_depth = 0;
       printf("\n");
@@ -5505,7 +5805,7 @@ namespace logger {
     }
   }
 
-  simdjson_unused static simdjson_really_inline void log_string(const char *message) {
+  simdjson_unused static simdjson_inline void log_string(const char *message) {
     if (LOG_ENABLED) {
       printf("%s\n", message);
     }
@@ -5513,7 +5813,7 @@ namespace logger {
 
   // Logs a single line from the stage 2 DOM parser
   template<typename S>
-  static simdjson_really_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
+  static simdjson_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
     if (LOG_ENABLED) {
       printf("| %*s%s%-*s ", log_depth*2, "", title_prefix, LOG_EVENT_LEN - log_depth*2 - int(strlen(title_prefix)), title);
       auto current_index = structurals.at_beginning() ? nullptr : structurals.next_structural-1;
@@ -5591,14 +5891,14 @@ class json_iterator {
    * - increment_count(iter) - each time a value is found in an array or object.
    */
   template<bool STREAMING, typename V>
-  simdjson_warn_unused simdjson_really_inline error_code walk_document(V &visitor) noexcept;
+  simdjson_warn_unused simdjson_inline error_code walk_document(V &visitor) noexcept;
 
   /**
    * Create an iterator capable of walking a JSON document.
    *
    * The document must have already passed through stage 1.
    */
-  simdjson_really_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
+  simdjson_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
 
   /**
    * Look at the next token.
@@ -5607,7 +5907,7 @@ class json_iterator {
    *
    * They may include invalid JSON as well (such as `1.2.3` or `ture`).
    */
-  simdjson_really_inline const uint8_t *peek() const noexcept;
+  simdjson_inline const uint8_t *peek() const noexcept;
   /**
    * Advance to the next token.
    *
@@ -5615,56 +5915,56 @@ class json_iterator {
    *
    * They may include invalid JSON as well (such as `1.2.3` or `ture`).
    */
-  simdjson_really_inline const uint8_t *advance() noexcept;
+  simdjson_inline const uint8_t *advance() noexcept;
   /**
    * Get the remaining length of the document, from the start of the current token.
    */
-  simdjson_really_inline size_t remaining_len() const noexcept;
+  simdjson_inline size_t remaining_len() const noexcept;
   /**
    * Check if we are at the end of the document.
    *
    * If this is true, there are no more tokens.
    */
-  simdjson_really_inline bool at_eof() const noexcept;
+  simdjson_inline bool at_eof() const noexcept;
   /**
    * Check if we are at the beginning of the document.
    */
-  simdjson_really_inline bool at_beginning() const noexcept;
-  simdjson_really_inline uint8_t last_structural() const noexcept;
+  simdjson_inline bool at_beginning() const noexcept;
+  simdjson_inline uint8_t last_structural() const noexcept;
 
   /**
    * Log that a value has been found.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_value(const char *type) const noexcept;
+  simdjson_inline void log_value(const char *type) const noexcept;
   /**
    * Log the start of a multipart value.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_start_value(const char *type) const noexcept;
+  simdjson_inline void log_start_value(const char *type) const noexcept;
   /**
    * Log the end of a multipart value.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_end_value(const char *type) const noexcept;
+  simdjson_inline void log_end_value(const char *type) const noexcept;
   /**
    * Log an error.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_error(const char *error) const noexcept;
+  simdjson_inline void log_error(const char *error) const noexcept;
 
   template<typename V>
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
   template<typename V>
-  simdjson_warn_unused simdjson_really_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
 };
 
 template<bool STREAMING, typename V>
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::walk_document(V &visitor) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::walk_document(V &visitor) noexcept {
   logger::log_start();
 
   //
@@ -5789,52 +6089,52 @@ simdjson_warn_unused simdjson_really_inline error_code json_iterator::walk_docum
 
 } // walk_document()
 
-simdjson_really_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+simdjson_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
   : buf{_dom_parser.buf},
     next_structural{&_dom_parser.structural_indexes[start_structural_index]},
     dom_parser{_dom_parser} {
 }
 
-simdjson_really_inline const uint8_t *json_iterator::peek() const noexcept {
+simdjson_inline const uint8_t *json_iterator::peek() const noexcept {
   return &buf[*(next_structural)];
 }
-simdjson_really_inline const uint8_t *json_iterator::advance() noexcept {
+simdjson_inline const uint8_t *json_iterator::advance() noexcept {
   return &buf[*(next_structural++)];
 }
-simdjson_really_inline size_t json_iterator::remaining_len() const noexcept {
+simdjson_inline size_t json_iterator::remaining_len() const noexcept {
   return dom_parser.len - *(next_structural-1);
 }
 
-simdjson_really_inline bool json_iterator::at_eof() const noexcept {
+simdjson_inline bool json_iterator::at_eof() const noexcept {
   return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
 }
-simdjson_really_inline bool json_iterator::at_beginning() const noexcept {
+simdjson_inline bool json_iterator::at_beginning() const noexcept {
   return next_structural == dom_parser.structural_indexes.get();
 }
-simdjson_really_inline uint8_t json_iterator::last_structural() const noexcept {
+simdjson_inline uint8_t json_iterator::last_structural() const noexcept {
   return buf[dom_parser.structural_indexes[dom_parser.n_structural_indexes - 1]];
 }
 
-simdjson_really_inline void json_iterator::log_value(const char *type) const noexcept {
+simdjson_inline void json_iterator::log_value(const char *type) const noexcept {
   logger::log_line(*this, "", type, "");
 }
 
-simdjson_really_inline void json_iterator::log_start_value(const char *type) const noexcept {
+simdjson_inline void json_iterator::log_start_value(const char *type) const noexcept {
   logger::log_line(*this, "+", type, "");
   if (logger::LOG_ENABLED) { logger::log_depth++; }
 }
 
-simdjson_really_inline void json_iterator::log_end_value(const char *type) const noexcept {
+simdjson_inline void json_iterator::log_end_value(const char *type) const noexcept {
   if (logger::LOG_ENABLED) { logger::log_depth--; }
   logger::log_line(*this, "-", type, "");
 }
 
-simdjson_really_inline void json_iterator::log_error(const char *error) const noexcept {
+simdjson_inline void json_iterator::log_error(const char *error) const noexcept {
   logger::log_line(*this, "", "ERROR", error);
 }
 
 template<typename V>
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
   switch (*value) {
     case '"': return visitor.visit_root_string(*this, value);
     case 't': return visitor.visit_root_true_atom(*this, value);
@@ -5850,7 +6150,7 @@ simdjson_warn_unused simdjson_really_inline error_code json_iterator::visit_root
   }
 }
 template<typename V>
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
   switch (*value) {
     case '"': return visitor.visit_string(*this, value);
     case 't': return visitor.visit_true_atom(*this, value);
@@ -5882,18 +6182,18 @@ struct tape_writer {
   uint64_t *next_tape_loc;
 
   /** Write a signed 64-bit value to tape. */
-  simdjson_really_inline void append_s64(int64_t value) noexcept;
+  simdjson_inline void append_s64(int64_t value) noexcept;
 
   /** Write an unsigned 64-bit value to tape. */
-  simdjson_really_inline void append_u64(uint64_t value) noexcept;
+  simdjson_inline void append_u64(uint64_t value) noexcept;
 
   /** Write a double value to tape. */
-  simdjson_really_inline void append_double(double value) noexcept;
+  simdjson_inline void append_double(double value) noexcept;
 
   /**
    * Append a tape entry (an 8-bit type,and 56 bits worth of value).
    */
-  simdjson_really_inline void append(uint64_t val, internal::tape_type t) noexcept;
+  simdjson_inline void append(uint64_t val, internal::tape_type t) noexcept;
 
   /**
    * Skip the current tape entry without writing.
@@ -5901,24 +6201,24 @@ struct tape_writer {
    * Used to skip the start of the container, since we'll come back later to fill it in when the
    * container ends.
    */
-  simdjson_really_inline void skip() noexcept;
+  simdjson_inline void skip() noexcept;
 
   /**
    * Skip the number of tape entries necessary to write a large u64 or i64.
    */
-  simdjson_really_inline void skip_large_integer() noexcept;
+  simdjson_inline void skip_large_integer() noexcept;
 
   /**
    * Skip the number of tape entries necessary to write a double.
    */
-  simdjson_really_inline void skip_double() noexcept;
+  simdjson_inline void skip_double() noexcept;
 
   /**
    * Write a value to a known location on tape.
    *
    * Used to go back and write out the start of a container after the container ends.
    */
-  simdjson_really_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
+  simdjson_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
 
 private:
   /**
@@ -5926,50 +6226,50 @@ struct tape_writer {
    * all 64 bits, such as double and uint64_t.
    */
   template<typename T>
-  simdjson_really_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
+  simdjson_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
 }; // struct number_writer
 
-simdjson_really_inline void tape_writer::append_s64(int64_t value) noexcept {
+simdjson_inline void tape_writer::append_s64(int64_t value) noexcept {
   append2(0, value, internal::tape_type::INT64);
 }
 
-simdjson_really_inline void tape_writer::append_u64(uint64_t value) noexcept {
+simdjson_inline void tape_writer::append_u64(uint64_t value) noexcept {
   append(0, internal::tape_type::UINT64);
   *next_tape_loc = value;
   next_tape_loc++;
 }
 
 /** Write a double value to tape. */
-simdjson_really_inline void tape_writer::append_double(double value) noexcept {
+simdjson_inline void tape_writer::append_double(double value) noexcept {
   append2(0, value, internal::tape_type::DOUBLE);
 }
 
-simdjson_really_inline void tape_writer::skip() noexcept {
+simdjson_inline void tape_writer::skip() noexcept {
   next_tape_loc++;
 }
 
-simdjson_really_inline void tape_writer::skip_large_integer() noexcept {
+simdjson_inline void tape_writer::skip_large_integer() noexcept {
   next_tape_loc += 2;
 }
 
-simdjson_really_inline void tape_writer::skip_double() noexcept {
+simdjson_inline void tape_writer::skip_double() noexcept {
   next_tape_loc += 2;
 }
 
-simdjson_really_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
+simdjson_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
   *next_tape_loc = val | ((uint64_t(char(t))) << 56);
   next_tape_loc++;
 }
 
 template<typename T>
-simdjson_really_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
+simdjson_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
   append(val, t);
   static_assert(sizeof(val2) == sizeof(*next_tape_loc), "Type is not 64 bits!");
   memcpy(next_tape_loc, &val2, sizeof(val2));
   next_tape_loc++;
 }
 
-simdjson_really_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
+simdjson_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
   tape_loc = val | ((uint64_t(char(t))) << 56);
 }
 
@@ -5986,40 +6286,40 @@ namespace stage2 {
 
 struct tape_builder {
   template<bool STREAMING>
-  simdjson_warn_unused static simdjson_really_inline error_code parse_document(
+  simdjson_warn_unused static simdjson_inline error_code parse_document(
     dom_parser_implementation &dom_parser,
     dom::document &doc) noexcept;
 
   /** Called when a non-empty document starts. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_document_start(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_document_start(json_iterator &iter) noexcept;
   /** Called when a non-empty document ends without error. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_document_end(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_document_end(json_iterator &iter) noexcept;
 
   /** Called when a non-empty array starts. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_array_start(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_array_start(json_iterator &iter) noexcept;
   /** Called when a non-empty array ends. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_array_end(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_array_end(json_iterator &iter) noexcept;
   /** Called when an empty array is found. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_empty_array(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_empty_array(json_iterator &iter) noexcept;
 
   /** Called when a non-empty object starts. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_object_start(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_object_start(json_iterator &iter) noexcept;
   /**
    * Called when a key in a field is encountered.
    *
    * primitive, visit_object_start, visit_empty_object, visit_array_start, or visit_empty_array
    * will be called after this with the field value.
    */
-  simdjson_warn_unused simdjson_really_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
   /** Called when a non-empty object ends. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_object_end(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_object_end(json_iterator &iter) noexcept;
   /** Called when an empty object is found. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_empty_object(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_empty_object(json_iterator &iter) noexcept;
 
   /**
    * Called when a string, number, boolean or null is found.
    */
-  simdjson_warn_unused simdjson_really_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
   /**
    * Called when a string, number, boolean or null is found at the top level of a document (i.e.
    * when there is no array or object and the entire document is a single string, number, boolean or
@@ -6028,22 +6328,22 @@ struct tape_builder {
    * This is separate from primitive() because simdjson's normal primitive parsing routines assume
    * there is at least one more token after the value, which is only true in an array or object.
    */
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
 
-  simdjson_warn_unused simdjson_really_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
 
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
 
   /** Called each time a new field or element in an array or object is found. */
-  simdjson_warn_unused simdjson_really_inline error_code increment_count(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code increment_count(json_iterator &iter) noexcept;
 
   /** Next location to write to tape */
   tape_writer tape;
@@ -6051,18 +6351,18 @@ struct tape_builder {
   /** Next write location in the string buf for stage 2 parsing */
   uint8_t *current_string_buf_loc;
 
-  simdjson_really_inline tape_builder(dom::document &doc) noexcept;
+  simdjson_inline tape_builder(dom::document &doc) noexcept;
 
-  simdjson_really_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
-  simdjson_really_inline void start_container(json_iterator &iter) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
-  simdjson_really_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
-  simdjson_really_inline void on_end_string(uint8_t *dst) noexcept;
+  simdjson_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
+  simdjson_inline void start_container(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_warn_unused simdjson_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
+  simdjson_inline void on_end_string(uint8_t *dst) noexcept;
 }; // class tape_builder
 
 template<bool STREAMING>
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::parse_document(
+simdjson_warn_unused simdjson_inline error_code tape_builder::parse_document(
     dom_parser_implementation &dom_parser,
     dom::document &doc) noexcept {
   dom_parser.doc = &doc;
@@ -6071,56 +6371,56 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::parse_docum
   return iter.walk_document<STREAMING>(builder);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
   return iter.visit_root_primitive(*this, value);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
   return iter.visit_primitive(*this, value);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
   return empty_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
   return empty_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
   start_container(iter);
   return SUCCESS;
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
   start_container(iter);
   return SUCCESS;
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
   start_container(iter);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
   return end_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
   return end_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
   constexpr uint32_t start_tape_index = 0;
   tape.append(start_tape_index, internal::tape_type::ROOT);
   tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter), internal::tape_type::ROOT);
   return SUCCESS;
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
   return visit_string(iter, key, true);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
   iter.dom_parser.open_containers[iter.depth].count++; // we have a key value pair in the object at parser.dom_parser.depth - 1
   return SUCCESS;
 }
 
-simdjson_really_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
+simdjson_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
   iter.log_value(key ? "key" : "string");
   uint8_t *dst = on_start_string(iter);
   dst = stringparsing::parse_string(value+1, dst);
@@ -6132,16 +6432,16 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_strin
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
   return visit_string(iter, value);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("number");
   return numberparsing::parse_number(value, tape);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
   //
   // We need to make a copy to make sure that the string is space terminated.
   // This is not about padding the input, which should already padded up
@@ -6163,42 +6463,42 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_
   return error;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("true");
   if (!atomparsing::is_valid_true_atom(value)) { return T_ATOM_ERROR; }
   tape.append(0, internal::tape_type::TRUE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("true");
   if (!atomparsing::is_valid_true_atom(value, iter.remaining_len())) { return T_ATOM_ERROR; }
   tape.append(0, internal::tape_type::TRUE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("false");
   if (!atomparsing::is_valid_false_atom(value)) { return F_ATOM_ERROR; }
   tape.append(0, internal::tape_type::FALSE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("false");
   if (!atomparsing::is_valid_false_atom(value, iter.remaining_len())) { return F_ATOM_ERROR; }
   tape.append(0, internal::tape_type::FALSE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("null");
   if (!atomparsing::is_valid_null_atom(value)) { return N_ATOM_ERROR; }
   tape.append(0, internal::tape_type::NULL_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("null");
   if (!atomparsing::is_valid_null_atom(value, iter.remaining_len())) { return N_ATOM_ERROR; }
   tape.append(0, internal::tape_type::NULL_VALUE);
@@ -6207,24 +6507,24 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_
 
 // private:
 
-simdjson_really_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
+simdjson_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
   return uint32_t(tape.next_tape_loc - iter.dom_parser.doc->tape.get());
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
   auto start_index = next_tape_index(iter);
   tape.append(start_index+2, start);
   tape.append(start_index, end);
   return SUCCESS;
 }
 
-simdjson_really_inline void tape_builder::start_container(json_iterator &iter) noexcept {
+simdjson_inline void tape_builder::start_container(json_iterator &iter) noexcept {
   iter.dom_parser.open_containers[iter.depth].tape_index = next_tape_index(iter);
   iter.dom_parser.open_containers[iter.depth].count = 0;
   tape.skip(); // We don't actually *write* the start element until the end.
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
   // Write the ending tape element, pointing at the start location
   const uint32_t start_tape_index = iter.dom_parser.open_containers[iter.depth].tape_index;
   tape.append(start_tape_index, end);
@@ -6237,13 +6537,13 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::end_contain
   return SUCCESS;
 }
 
-simdjson_really_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
+simdjson_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
   // we advance the point, accounting for the fact that we have a NULL termination
   tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::STRING);
   return current_string_buf_loc + sizeof(uint32_t);
 }
 
-simdjson_really_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
+simdjson_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
   uint32_t str_length = uint32_t(dst - (current_string_buf_loc + sizeof(uint32_t)));
   // TODO check for overflow in case someone has a crazy string (>=4GB?)
   // But only add the overflow check when the document itself exceeds 4GB
@@ -6272,6 +6572,10 @@ simdjson_warn_unused error_code dom_parser_implementation::stage2_next(dom::docu
   return stage2::tape_builder::parse_document<true>(*this, _doc);
 }
 
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_string(const uint8_t *src, uint8_t *dst) const noexcept {
+  return fallback::stringparsing::parse_string(src, dst);
+}
+
 simdjson_warn_unused error_code dom_parser_implementation::parse(const uint8_t *_buf, size_t _len, dom::document &_doc) noexcept {
   auto error = stage1(_buf, _len, stage1_mode::regular);
   if (error) { return error; }
@@ -6336,25 +6640,25 @@ namespace {
 using namespace simd;
 
 struct json_character_block {
-  static simdjson_really_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
+  static simdjson_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
   //  ASCII white-space ('\r','\n','\t',' ')
-  simdjson_really_inline uint64_t whitespace() const noexcept;
+  simdjson_inline uint64_t whitespace() const noexcept;
   // non-quote structural characters (comma, colon, braces, brackets)
-  simdjson_really_inline uint64_t op() const noexcept;
+  simdjson_inline uint64_t op() const noexcept;
   // neither a structural character nor a white-space, so letters, numbers and quotes
-  simdjson_really_inline uint64_t scalar() const noexcept;
+  simdjson_inline uint64_t scalar() const noexcept;
 
   uint64_t _whitespace; // ASCII white-space ('\r','\n','\t',' ')
   uint64_t _op; // structural characters (comma, colon, braces, brackets but not quotes)
 };
 
-simdjson_really_inline uint64_t json_character_block::whitespace() const noexcept { return _whitespace; }
-simdjson_really_inline uint64_t json_character_block::op() const noexcept { return _op; }
-simdjson_really_inline uint64_t json_character_block::scalar() const noexcept { return ~(op() | whitespace()); }
+simdjson_inline uint64_t json_character_block::whitespace() const noexcept { return _whitespace; }
+simdjson_inline uint64_t json_character_block::op() const noexcept { return _op; }
+simdjson_inline uint64_t json_character_block::scalar() const noexcept { return ~(op() | whitespace()); }
 
 // This identifies structural characters (comma, colon, braces, brackets),
 // and ASCII white-space ('\r','\n','\t',' ').
-simdjson_really_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
+simdjson_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
   // These lookups rely on the fact that anything < 127 will match the lower 4 bits, which is why
   // we can't use the generic lookup_16.
   const auto whitespace_table = simd8<uint8_t>::repeat_16(' ', 100, 100, 100, 17, 100, 113, 2, 100, '\t', '\n', 112, 100, '\r', 100, 100);
@@ -6404,21 +6708,21 @@ simdjson_really_inline json_character_block json_character_block::classify(const
   return { whitespace, op };
 }
 
-simdjson_really_inline bool is_ascii(const simd8x64<uint8_t>& input) {
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input) {
   return input.reduce_or().is_ascii();
 }
 
-simdjson_unused simdjson_really_inline simd8<bool> must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
-  simd8<uint8_t> is_second_byte = prev1.saturating_sub(0b11000000u-1); // Only 11______ will be > 0
-  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0b11100000u-1); // Only 111_____ will be > 0
-  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0b11110000u-1); // Only 1111____ will be > 0
+simdjson_unused simdjson_inline simd8<bool> must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_second_byte = prev1.saturating_sub(0xc0u-1); // Only 11______ will be > 0
+  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-1); // Only 111_____ will be > 0
+  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-1); // Only 1111____ will be > 0
   // Caller requires a bool (all 1's). All values resulting from the subtraction will be <= 64, so signed comparison is fine.
   return simd8<int8_t>(is_second_byte | is_third_byte | is_fourth_byte) > int8_t(0);
 }
 
-simdjson_really_inline simd8<bool> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
-  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0b11100000u-1); // Only 111_____ will be > 0
-  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0b11110000u-1); // Only 1111____ will be > 0
+simdjson_inline simd8<bool> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-1); // Only 111_____ will be > 0
+  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-1); // Only 1111____ will be > 0
   // Caller requires a bool (all 1's). All values resulting from the subtraction will be <= 64, so signed comparison is fine.
   return simd8<int8_t>(is_third_byte | is_fourth_byte) > int8_t(0);
 }
@@ -6435,7 +6739,7 @@ namespace utf8_validation {
 
 using namespace simd;
 
-  simdjson_really_inline simd8<uint8_t> check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+  simdjson_inline simd8<uint8_t> check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
 // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
 // Bit 1 = Too Long (ASCII followed by continuation)
 // Bit 2 = Overlong 3-byte
@@ -6525,7 +6829,7 @@ using namespace simd;
     );
     return (byte_1_high & byte_1_low & byte_2_high);
   }
-  simdjson_really_inline simd8<uint8_t> check_multibyte_lengths(const simd8<uint8_t> input,
+  simdjson_inline simd8<uint8_t> check_multibyte_lengths(const simd8<uint8_t> input,
       const simd8<uint8_t> prev_input, const simd8<uint8_t> sc) {
     simd8<uint8_t> prev2 = input.prev<2>(prev_input);
     simd8<uint8_t> prev3 = input.prev<3>(prev_input);
@@ -6538,7 +6842,7 @@ using namespace simd;
   // Return nonzero if there are incomplete multibyte characters at the end of the block:
   // e.g. if there is a 4-byte character, but it's 3 bytes from the end.
   //
-  simdjson_really_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
+  simdjson_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
     // If the previous input's last 3 bytes match this, they're too short (they ended at EOF):
     // ... 1111____ 111_____ 11______
 #if SIMDJSON_IMPLEMENTATION_ICELAKE
@@ -6550,14 +6854,14 @@ using namespace simd;
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
-      255, 255, 255, 255, 255, 0b11110000u-1, 0b11100000u-1, 0b11000000u-1
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
     };
 #else
     static const uint8_t max_array[32] = {
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
-      255, 255, 255, 255, 255, 0b11110000u-1, 0b11100000u-1, 0b11000000u-1
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
     };
 #endif
     const simd8<uint8_t> max_value(&max_array[sizeof(max_array)-sizeof(simd8<uint8_t>)]);
@@ -6575,7 +6879,7 @@ using namespace simd;
     //
     // Check whether the current bytes are valid UTF-8.
     //
-    simdjson_really_inline void check_utf8_bytes(const simd8<uint8_t> input, const simd8<uint8_t> prev_input) {
+    simdjson_inline void check_utf8_bytes(const simd8<uint8_t> input, const simd8<uint8_t> prev_input) {
       // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ lead bytes
       // (2, 3, 4-byte leads become large positive numbers instead of small negative numbers)
       simd8<uint8_t> prev1 = input.prev<1>(prev_input);
@@ -6586,13 +6890,21 @@ using namespace simd;
     // The only problem that can happen at EOF is that a multibyte character is too short
     // or a byte value too large in the last bytes: check_special_cases only checks for bytes
     // too large in the first of two bytes.
-    simdjson_really_inline void check_eof() {
+    simdjson_inline void check_eof() {
       // If the previous block had incomplete UTF-8 characters at the end, an ASCII block can't
       // possibly finish them.
       this->error |= this->prev_incomplete;
     }
 
-    simdjson_really_inline void check_next_input(const simd8x64<uint8_t>& input) {
+#ifndef SIMDJSON_IF_CONSTEXPR
+#if SIMDJSON_CPLUSPLUS17
+#define SIMDJSON_IF_CONSTEXPR if constexpr
+#else
+#define SIMDJSON_IF_CONSTEXPR if
+#endif
+#endif
+
+    simdjson_inline void check_next_input(const simd8x64<uint8_t>& input) {
       if(simdjson_likely(is_ascii(input))) {
         this->error |= this->prev_incomplete;
       } else {
@@ -6601,12 +6913,12 @@ using namespace simd;
                 ||(simd8x64<uint8_t>::NUM_CHUNKS == 2)
                 || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
                 "We support one, two or four chunks per 64-byte block.");
-        if(simd8x64<uint8_t>::NUM_CHUNKS == 1) {
+        SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 1) {
           this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
-        } if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
           this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
           this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
           this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
           this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
           this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
@@ -6617,7 +6929,7 @@ using namespace simd;
       }
     }
     // do not forget to call check_eof!
-    simdjson_really_inline error_code errors() {
+    simdjson_inline error_code errors() {
       return this->error.any_bits_set_anywhere() ? error_code::UTF8_ERROR : error_code::SUCCESS;
     }
 
@@ -6647,10 +6959,10 @@ namespace {
 template<size_t STEP_SIZE>
 struct buf_block_reader {
 public:
-  simdjson_really_inline buf_block_reader(const uint8_t *_buf, size_t _len);
-  simdjson_really_inline size_t block_index();
-  simdjson_really_inline bool has_full_block() const;
-  simdjson_really_inline const uint8_t *full_block() const;
+  simdjson_inline buf_block_reader(const uint8_t *_buf, size_t _len);
+  simdjson_inline size_t block_index();
+  simdjson_inline bool has_full_block() const;
+  simdjson_inline const uint8_t *full_block() const;
   /**
    * Get the last block, padded with spaces.
    *
@@ -6660,8 +6972,8 @@ struct buf_block_reader {
    *
    * @return the number of effective characters in the last block.
    */
-  simdjson_really_inline size_t get_remainder(uint8_t *dst) const;
-  simdjson_really_inline void advance();
+  simdjson_inline size_t get_remainder(uint8_t *dst) const;
+  simdjson_inline void advance();
 private:
   const uint8_t *buf;
   const size_t len;
@@ -6700,23 +7012,23 @@ simdjson_unused static char * format_mask(uint64_t mask) {
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len) : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, idx{0} {}
+simdjson_inline buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len) : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, idx{0} {}
 
 template<size_t STEP_SIZE>
-simdjson_really_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
+simdjson_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
   return idx < lenminusstep;
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
+simdjson_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
   return &buf[idx];
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
   if(len == idx) { return 0; } // memcpy(dst, null, 0) will trigger an error with some sanitizers
   std::memset(dst, 0x20, STEP_SIZE); // std::memset STEP_SIZE because it's more efficient to write out 8 or 16 bytes at once.
   std::memcpy(dst, buf + idx, len - idx);
@@ -6724,7 +7036,7 @@ simdjson_really_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline void buf_block_reader<STEP_SIZE>::advance() {
+simdjson_inline void buf_block_reader<STEP_SIZE>::advance() {
   idx += STEP_SIZE;
 }
 
@@ -6740,27 +7052,27 @@ namespace stage1 {
 
 struct json_string_block {
   // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
-  simdjson_really_inline json_string_block(uint64_t backslash, uint64_t escaped, uint64_t quote, uint64_t in_string) :
+  simdjson_inline json_string_block(uint64_t backslash, uint64_t escaped, uint64_t quote, uint64_t in_string) :
   _backslash(backslash), _escaped(escaped), _quote(quote), _in_string(in_string) {}
 
   // Escaped characters (characters following an escape() character)
-  simdjson_really_inline uint64_t escaped() const { return _escaped; }
+  simdjson_inline uint64_t escaped() const { return _escaped; }
   // Escape characters (backslashes that are not escaped--i.e. in \\, includes only the first \)
-  simdjson_really_inline uint64_t escape() const { return _backslash & ~_escaped; }
+  simdjson_inline uint64_t escape() const { return _backslash & ~_escaped; }
   // Real (non-backslashed) quotes
-  simdjson_really_inline uint64_t quote() const { return _quote; }
+  simdjson_inline uint64_t quote() const { return _quote; }
   // Start quotes of strings
-  simdjson_really_inline uint64_t string_start() const { return _quote & _in_string; }
+  simdjson_inline uint64_t string_start() const { return _quote & _in_string; }
   // End quotes of strings
-  simdjson_really_inline uint64_t string_end() const { return _quote & ~_in_string; }
+  simdjson_inline uint64_t string_end() const { return _quote & ~_in_string; }
   // Only characters inside the string (not including the quotes)
-  simdjson_really_inline uint64_t string_content() const { return _in_string & ~_quote; }
+  simdjson_inline uint64_t string_content() const { return _in_string & ~_quote; }
   // Return a mask of whether the given characters are inside a string (only works on non-quotes)
-  simdjson_really_inline uint64_t non_quote_inside_string(uint64_t mask) const { return mask & _in_string; }
+  simdjson_inline uint64_t non_quote_inside_string(uint64_t mask) const { return mask & _in_string; }
   // Return a mask of whether the given characters are inside a string (only works on non-quotes)
-  simdjson_really_inline uint64_t non_quote_outside_string(uint64_t mask) const { return mask & ~_in_string; }
+  simdjson_inline uint64_t non_quote_outside_string(uint64_t mask) const { return mask & ~_in_string; }
   // Tail of string (everything except the start quote)
-  simdjson_really_inline uint64_t string_tail() const { return _in_string ^ _quote; }
+  simdjson_inline uint64_t string_tail() const { return _in_string ^ _quote; }
 
   // backslash characters
   uint64_t _backslash;
@@ -6775,14 +7087,14 @@ struct json_string_block {
 // Scans blocks for string characters, storing the state necessary to do so
 class json_string_scanner {
 public:
-  simdjson_really_inline json_string_block next(const simd::simd8x64<uint8_t>& in);
+  simdjson_inline json_string_block next(const simd::simd8x64<uint8_t>& in);
   // Returns either UNCLOSED_STRING or SUCCESS
-  simdjson_really_inline error_code finish();
+  simdjson_inline error_code finish();
 
 private:
   // Intended to be defined by the implementation
-  simdjson_really_inline uint64_t find_escaped(uint64_t escape);
-  simdjson_really_inline uint64_t find_escaped_branchless(uint64_t escape);
+  simdjson_inline uint64_t find_escaped(uint64_t escape);
+  simdjson_inline uint64_t find_escaped_branchless(uint64_t escape);
 
   // Whether the last iteration was still inside a string (all 1's = true, all 0's = false).
   uint64_t prev_in_string = 0ULL;
@@ -6817,7 +7129,7 @@ class json_string_scanner {
 // desired        |   x  | x x  x x  x x  x  x  |
 // text           |  \\\ | \\\"\\\" \\\" \\"\\" |
 //
-simdjson_really_inline uint64_t json_string_scanner::find_escaped_branchless(uint64_t backslash) {
+simdjson_inline uint64_t json_string_scanner::find_escaped_branchless(uint64_t backslash) {
   // If there was overflow, pretend the first character isn't a backslash
   backslash &= ~prev_escaped;
   uint64_t follows_escape = backslash << 1 | prev_escaped;
@@ -6842,7 +7154,7 @@ simdjson_really_inline uint64_t json_string_scanner::find_escaped_branchless(uin
 //
 // Backslash sequences outside of quotes will be detected in stage 2.
 //
-simdjson_really_inline json_string_block json_string_scanner::next(const simd::simd8x64<uint8_t>& in) {
+simdjson_inline json_string_block json_string_scanner::next(const simd::simd8x64<uint8_t>& in) {
   const uint64_t backslash = in.eq('\\');
   const uint64_t escaped = find_escaped(backslash);
   const uint64_t quote = in.eq('"') & ~escaped;
@@ -6875,7 +7187,7 @@ simdjson_really_inline json_string_block json_string_scanner::next(const simd::s
   );
 }
 
-simdjson_really_inline error_code json_string_scanner::finish() {
+simdjson_inline error_code json_string_scanner::finish() {
   if (prev_in_string) {
     return UNCLOSED_STRING;
   }
@@ -6913,25 +7225,25 @@ namespace stage1 {
 struct json_block {
 public:
   // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
-  simdjson_really_inline json_block(json_string_block&& string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  simdjson_inline json_block(json_string_block&& string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
   _string(std::move(string)), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
-  simdjson_really_inline json_block(json_string_block string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  simdjson_inline json_block(json_string_block string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
   _string(string), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
 
   /**
    * The start of structurals.
    * In simdjson prior to v0.3, these were called the pseudo-structural characters.
    **/
-  simdjson_really_inline uint64_t structural_start() const noexcept { return potential_structural_start() & ~_string.string_tail(); }
+  simdjson_inline uint64_t structural_start() const noexcept { return potential_structural_start() & ~_string.string_tail(); }
   /** All JSON whitespace (i.e. not in a string) */
-  simdjson_really_inline uint64_t whitespace() const noexcept { return non_quote_outside_string(_characters.whitespace()); }
+  simdjson_inline uint64_t whitespace() const noexcept { return non_quote_outside_string(_characters.whitespace()); }
 
   // Helpers
 
   /** Whether the given characters are inside a string (only works on non-quotes) */
-  simdjson_really_inline uint64_t non_quote_inside_string(uint64_t mask) const noexcept { return _string.non_quote_inside_string(mask); }
+  simdjson_inline uint64_t non_quote_inside_string(uint64_t mask) const noexcept { return _string.non_quote_inside_string(mask); }
   /** Whether the given characters are outside a string (only works on non-quotes) */
-  simdjson_really_inline uint64_t non_quote_outside_string(uint64_t mask) const noexcept { return _string.non_quote_outside_string(mask); }
+  simdjson_inline uint64_t non_quote_outside_string(uint64_t mask) const noexcept { return _string.non_quote_outside_string(mask); }
 
   // string and escape characters
   json_string_block _string;
@@ -6946,12 +7258,12 @@ struct json_block {
    * structural elements ([,],{,},:, comma) plus scalar starts like 123, true and "abc".
    * They may reside inside a string.
    **/
-  simdjson_really_inline uint64_t potential_structural_start() const noexcept { return _characters.op() | potential_scalar_start(); }
+  simdjson_inline uint64_t potential_structural_start() const noexcept { return _characters.op() | potential_scalar_start(); }
   /**
    * The start of non-operator runs, like 123, true and "abc".
    * It main reside inside a string.
    **/
-  simdjson_really_inline uint64_t potential_scalar_start() const noexcept {
+  simdjson_inline uint64_t potential_scalar_start() const noexcept {
     // The term "scalar" refers to anything except structural characters and white space
     // (so letters, numbers, quotes).
     // Whenever it is preceded by something that is not a structural element ({,},[,],:, ") nor a white-space
@@ -6962,7 +7274,7 @@ struct json_block {
    * Whether the given character is immediately after a non-operator like 123, true.
    * The characters following a quote are not included.
    */
-  simdjson_really_inline uint64_t follows_potential_scalar() const noexcept {
+  simdjson_inline uint64_t follows_potential_scalar() const noexcept {
     // _follows_potential_nonquote_scalar: is defined as marking any character that follows a character
     // that is not a structural element ({,},[,],:, comma) nor a quote (") and that is not a
     // white space.
@@ -6986,10 +7298,10 @@ struct json_block {
  */
 class json_scanner {
 public:
-  json_scanner() {}
-  simdjson_really_inline json_block next(const simd::simd8x64<uint8_t>& in);
+  json_scanner() = default;
+  simdjson_inline json_block next(const simd::simd8x64<uint8_t>& in);
   // Returns either UNCLOSED_STRING or SUCCESS
-  simdjson_really_inline error_code finish();
+  simdjson_inline error_code finish();
 
 private:
   // Whether the last character of the previous iteration is part of a scalar token
@@ -7006,13 +7318,13 @@ class json_scanner {
 //
 //     const uint64_t backslashed_quote = in.eq('"') & immediately_follows(in.eq('\'), prev_backslash);
 //
-simdjson_really_inline uint64_t follows(const uint64_t match, uint64_t &overflow) {
+simdjson_inline uint64_t follows(const uint64_t match, uint64_t &overflow) {
   const uint64_t result = match << 1 | overflow;
   overflow = match >> 63;
   return result;
 }
 
-simdjson_really_inline json_block json_scanner::next(const simd::simd8x64<uint8_t>& in) {
+simdjson_inline json_block json_scanner::next(const simd::simd8x64<uint8_t>& in) {
   json_string_block strings = string_scanner.next(in);
   // identifies the white-space and the structural characters
   json_character_block characters = json_character_block::classify(in);
@@ -7037,7 +7349,7 @@ simdjson_really_inline json_block json_scanner::next(const simd::simd8x64<uint8_
   );
 }
 
-simdjson_really_inline error_code json_scanner::finish() {
+simdjson_inline error_code json_scanner::finish() {
   return string_scanner.finish();
 }
 
@@ -7063,23 +7375,23 @@ class json_minifier {
   static error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept;
 
 private:
-  simdjson_really_inline json_minifier(uint8_t *_dst)
+  simdjson_inline json_minifier(uint8_t *_dst)
   : dst{_dst}
   {}
   template<size_t STEP_SIZE>
-  simdjson_really_inline void step(const uint8_t *block_buf, buf_block_reader<STEP_SIZE> &reader) noexcept;
-  simdjson_really_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block);
-  simdjson_really_inline error_code finish(uint8_t *dst_start, size_t &dst_len);
+  simdjson_inline void step(const uint8_t *block_buf, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block);
+  simdjson_inline error_code finish(uint8_t *dst_start, size_t &dst_len);
   json_scanner scanner{};
   uint8_t *dst;
 };
 
-simdjson_really_inline void json_minifier::next(const simd::simd8x64<uint8_t>& in, const json_block& block) {
+simdjson_inline void json_minifier::next(const simd::simd8x64<uint8_t>& in, const json_block& block) {
   uint64_t mask = block.whitespace();
   dst += in.compress(mask, dst);
 }
 
-simdjson_really_inline error_code json_minifier::finish(uint8_t *dst_start, size_t &dst_len) {
+simdjson_inline error_code json_minifier::finish(uint8_t *dst_start, size_t &dst_len) {
   error_code error = scanner.finish();
   if (error) { dst_len = 0; return error; }
   dst_len = dst - dst_start;
@@ -7087,7 +7399,7 @@ simdjson_really_inline error_code json_minifier::finish(uint8_t *dst_start, size
 }
 
 template<>
-simdjson_really_inline void json_minifier::step<128>(const uint8_t *block_buf, buf_block_reader<128> &reader) noexcept {
+simdjson_inline void json_minifier::step<128>(const uint8_t *block_buf, buf_block_reader<128> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block_buf);
   simd::simd8x64<uint8_t> in_2(block_buf+64);
   json_block block_1 = scanner.next(in_1);
@@ -7098,7 +7410,7 @@ simdjson_really_inline void json_minifier::step<128>(const uint8_t *block_buf, b
 }
 
 template<>
-simdjson_really_inline void json_minifier::step<64>(const uint8_t *block_buf, buf_block_reader<64> &reader) noexcept {
+simdjson_inline void json_minifier::step<64>(const uint8_t *block_buf, buf_block_reader<64> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block_buf);
   json_block block_1 = scanner.next(in_1);
   this->next(block_buf, block_1);
@@ -7171,7 +7483,7 @@ namespace {
   * complete document, therefore the last json buffer location is the end of the
   * batch.
   */
-simdjson_really_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
+simdjson_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
   // Variant: do not count separately, just figure out depth
   if(parser.n_structural_indexes == 0) { return 0; }
   auto arr_cnt = 0;
@@ -7246,7 +7558,7 @@ class bit_indexer {
 public:
   uint32_t *tail;
 
-  simdjson_really_inline bit_indexer(uint32_t *index_buf) : tail(index_buf) {}
+  simdjson_inline bit_indexer(uint32_t *index_buf) : tail(index_buf) {}
 
   // flatten out values in 'bits' assuming that they are are to have values of idx
   // plus their position in the bitvector, and store these indexes at
@@ -7257,9 +7569,9 @@ class bit_indexer {
   // If the kernel sets SIMDJSON_CUSTOM_BIT_INDEXER, then it will provide its own
   // version of the code.
 #ifdef SIMDJSON_CUSTOM_BIT_INDEXER
-  simdjson_really_inline void write(uint32_t idx, uint64_t bits);
+  simdjson_inline void write(uint32_t idx, uint64_t bits);
 #else
-  simdjson_really_inline void write(uint32_t idx, uint64_t bits) {
+  simdjson_inline void write(uint32_t idx, uint64_t bits) {
     // In some instances, the next branch is expensive because it is mispredicted.
     // Unfortunately, in other cases,
     // it helps tremendously.
@@ -7368,11 +7680,11 @@ class json_structural_indexer {
   static error_code index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept;
 
 private:
-  simdjson_really_inline json_structural_indexer(uint32_t *structural_indexes);
+  simdjson_inline json_structural_indexer(uint32_t *structural_indexes);
   template<size_t STEP_SIZE>
-  simdjson_really_inline void step(const uint8_t *block, buf_block_reader<STEP_SIZE> &reader) noexcept;
-  simdjson_really_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx);
-  simdjson_really_inline error_code finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial);
+  simdjson_inline void step(const uint8_t *block, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx);
+  simdjson_inline error_code finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial);
 
   json_scanner scanner{};
   utf8_checker checker{};
@@ -7381,26 +7693,26 @@ class json_structural_indexer {
   uint64_t unescaped_chars_error = 0;
 };
 
-simdjson_really_inline json_structural_indexer::json_structural_indexer(uint32_t *structural_indexes) : indexer{structural_indexes} {}
+simdjson_inline json_structural_indexer::json_structural_indexer(uint32_t *structural_indexes) : indexer{structural_indexes} {}
 
 // Skip the last character if it is partial
-simdjson_really_inline size_t trim_partial_utf8(const uint8_t *buf, size_t len) {
+simdjson_inline size_t trim_partial_utf8(const uint8_t *buf, size_t len) {
   if (simdjson_unlikely(len < 3)) {
     switch (len) {
       case 2:
-        if (buf[len-1] >= 0b11000000) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
-        if (buf[len-2] >= 0b11100000) { return len-2; } // 3- and 4-byte characters with only 2 bytes left
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 2 bytes left
         return len;
       case 1:
-        if (buf[len-1] >= 0b11000000) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
         return len;
       case 0:
         return len;
     }
   }
-  if (buf[len-1] >= 0b11000000) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
-  if (buf[len-2] >= 0b11100000) { return len-2; } // 3- and 4-byte characters with only 1 byte left
-  if (buf[len-3] >= 0b11110000) { return len-3; } // 4-byte characters with only 3 bytes left
+  if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+  if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 1 byte left
+  if (buf[len-3] >= 0xf0) { return len-3; } // 4-byte characters with only 3 bytes left
   return len;
 }
 
@@ -7449,7 +7761,7 @@ error_code json_structural_indexer::index(const uint8_t *buf, size_t len, dom_pa
 }
 
 template<>
-simdjson_really_inline void json_structural_indexer::step<128>(const uint8_t *block, buf_block_reader<128> &reader) noexcept {
+simdjson_inline void json_structural_indexer::step<128>(const uint8_t *block, buf_block_reader<128> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block);
   simd::simd8x64<uint8_t> in_2(block+64);
   json_block block_1 = scanner.next(in_1);
@@ -7460,14 +7772,14 @@ simdjson_really_inline void json_structural_indexer::step<128>(const uint8_t *bl
 }
 
 template<>
-simdjson_really_inline void json_structural_indexer::step<64>(const uint8_t *block, buf_block_reader<64> &reader) noexcept {
+simdjson_inline void json_structural_indexer::step<64>(const uint8_t *block, buf_block_reader<64> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block);
   json_block block_1 = scanner.next(in_1);
   this->next(in_1, block_1, reader.block_index());
   reader.advance();
 }
 
-simdjson_really_inline void json_structural_indexer::next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx) {
+simdjson_inline void json_structural_indexer::next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx) {
   uint64_t unescaped = in.lteq(0x1F);
   checker.check_next_input(in);
   indexer.write(uint32_t(idx-64), prev_structurals); // Output *last* iteration's structurals to the parser
@@ -7475,7 +7787,7 @@ simdjson_really_inline void json_structural_indexer::next(const simd::simd8x64<u
   unescaped_chars_error |= block.non_quote_inside_string(unescaped);
 }
 
-simdjson_really_inline error_code json_structural_indexer::finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial) {
+simdjson_inline error_code json_structural_indexer::finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial) {
   // Write out the final iteration's structurals
   indexer.write(uint32_t(idx-64), prev_structurals);
   error_code error = scanner.finish();
@@ -7593,7 +7905,7 @@ simdjson_really_inline error_code json_structural_indexer::finish(dom_parser_imp
 // as a workaround, we disable warnings within the following function.
 SIMDJSON_PUSH_DISABLE_ALL_WARNINGS
 namespace simdjson { namespace icelake { namespace { namespace stage1 {
-simdjson_really_inline void bit_indexer::write(uint32_t idx, uint64_t bits) {
+simdjson_inline void bit_indexer::write(uint32_t idx, uint64_t bits) {
     // In some instances, the next branch is expensive because it is mispredicted.
     // Unfortunately, in other cases,
     // it helps tremendously.
@@ -7668,6 +7980,151 @@ bool generic_validate_utf8(const char * input, size_t length) {
 //
 // Stage 2
 //
+/* begin file src/generic/stage2/stringparsing.h */
+// This file contains the common code every implementation uses
+// It is intended to be included multiple times and compiled multiple times
+
+namespace simdjson {
+namespace icelake {
+namespace {
+/// @private
+namespace stringparsing {
+
+// begin copypasta
+// These chars yield themselves: " \ /
+// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
+// u not handled in this table as it's complex
+static const uint8_t escape_map[256] = {
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
+    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
+    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+};
+
+// handle a unicode codepoint
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr) {
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion isn't valid; we defer the check for this to inside the
+  // multilingual plane check
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) != ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      return false;
+    }
+    uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+
+    // We have already checked that the high surrogate is valid and
+    // (code_point - 0xd800) < 1024.
+    //
+    // Check that code_point_2 is in the range 0xdc00..0xdfff
+    // and that code_point_2 was parsed from valid hex.
+    uint32_t low_bit = code_point_2 - 0xdc00;
+    if (low_bit >> 10) {
+      return false;
+    }
+
+    code_point =
+        (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+    *src_ptr += 6;
+  } else if (code_point >= 0xdc00 && code_point <= 0xdfff) {
+      // If we encounter a low surrogate (not preceded by a high surrogate)
+      // then we have an error.
+      return false;
+  }
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+/**
+ * Unescape a valid UTF-8 string from src to dst, stopping at a final unescaped quote. There
+ * must be an unescaped quote terminating the string. It returns the final output
+ * position as pointer. In case of error (e.g., the string has bad escaped codes),
+ * then null_nullptrptr is returned. It is assumed that the output buffer is large
+ * enough. E.g., if src points at 'joe"', then dst needs to have four free bytes +
+ * SIMDJSON_PADDING bytes.
+ */
+simdjson_warn_unused simdjson_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst) {
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto bs_quote = backslash_and_quote::copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint(&src, &dst)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+  /* can't be reached */
+  return nullptr;
+}
+
+} // namespace stringparsing
+} // unnamed namespace
+} // namespace icelake
+} // namespace simdjson
+/* end file src/generic/stage2/stringparsing.h */
 /* begin file src/generic/stage2/tape_builder.h */
 /* begin file src/generic/stage2/json_iterator.h */
 /* begin file src/generic/stage2/logger.h */
@@ -7693,7 +8150,7 @@ namespace logger {
   static int log_depth; // Not threadsafe. Log only.
 
   // Helper to turn unprintable or newline characters into spaces
-  static simdjson_really_inline char printable_char(char c) {
+  static simdjson_inline char printable_char(char c) {
     if (c >= 0x20) {
       return c;
     } else {
@@ -7702,7 +8159,7 @@ namespace logger {
   }
 
   // Print the header and set up log_start
-  static simdjson_really_inline void log_start() {
+  static simdjson_inline void log_start() {
     if (LOG_ENABLED) {
       log_depth = 0;
       printf("\n");
@@ -7711,7 +8168,7 @@ namespace logger {
     }
   }
 
-  simdjson_unused static simdjson_really_inline void log_string(const char *message) {
+  simdjson_unused static simdjson_inline void log_string(const char *message) {
     if (LOG_ENABLED) {
       printf("%s\n", message);
     }
@@ -7719,7 +8176,7 @@ namespace logger {
 
   // Logs a single line from the stage 2 DOM parser
   template<typename S>
-  static simdjson_really_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
+  static simdjson_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
     if (LOG_ENABLED) {
       printf("| %*s%s%-*s ", log_depth*2, "", title_prefix, LOG_EVENT_LEN - log_depth*2 - int(strlen(title_prefix)), title);
       auto current_index = structurals.at_beginning() ? nullptr : structurals.next_structural-1;
@@ -7797,14 +8254,14 @@ class json_iterator {
    * - increment_count(iter) - each time a value is found in an array or object.
    */
   template<bool STREAMING, typename V>
-  simdjson_warn_unused simdjson_really_inline error_code walk_document(V &visitor) noexcept;
+  simdjson_warn_unused simdjson_inline error_code walk_document(V &visitor) noexcept;
 
   /**
    * Create an iterator capable of walking a JSON document.
    *
    * The document must have already passed through stage 1.
    */
-  simdjson_really_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
+  simdjson_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
 
   /**
    * Look at the next token.
@@ -7813,7 +8270,7 @@ class json_iterator {
    *
    * They may include invalid JSON as well (such as `1.2.3` or `ture`).
    */
-  simdjson_really_inline const uint8_t *peek() const noexcept;
+  simdjson_inline const uint8_t *peek() const noexcept;
   /**
    * Advance to the next token.
    *
@@ -7821,56 +8278,56 @@ class json_iterator {
    *
    * They may include invalid JSON as well (such as `1.2.3` or `ture`).
    */
-  simdjson_really_inline const uint8_t *advance() noexcept;
+  simdjson_inline const uint8_t *advance() noexcept;
   /**
    * Get the remaining length of the document, from the start of the current token.
    */
-  simdjson_really_inline size_t remaining_len() const noexcept;
+  simdjson_inline size_t remaining_len() const noexcept;
   /**
    * Check if we are at the end of the document.
    *
    * If this is true, there are no more tokens.
    */
-  simdjson_really_inline bool at_eof() const noexcept;
+  simdjson_inline bool at_eof() const noexcept;
   /**
    * Check if we are at the beginning of the document.
    */
-  simdjson_really_inline bool at_beginning() const noexcept;
-  simdjson_really_inline uint8_t last_structural() const noexcept;
+  simdjson_inline bool at_beginning() const noexcept;
+  simdjson_inline uint8_t last_structural() const noexcept;
 
   /**
    * Log that a value has been found.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_value(const char *type) const noexcept;
+  simdjson_inline void log_value(const char *type) const noexcept;
   /**
    * Log the start of a multipart value.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_start_value(const char *type) const noexcept;
+  simdjson_inline void log_start_value(const char *type) const noexcept;
   /**
    * Log the end of a multipart value.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_end_value(const char *type) const noexcept;
+  simdjson_inline void log_end_value(const char *type) const noexcept;
   /**
    * Log an error.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_error(const char *error) const noexcept;
+  simdjson_inline void log_error(const char *error) const noexcept;
 
   template<typename V>
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
   template<typename V>
-  simdjson_warn_unused simdjson_really_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
 };
 
 template<bool STREAMING, typename V>
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::walk_document(V &visitor) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::walk_document(V &visitor) noexcept {
   logger::log_start();
 
   //
@@ -7995,52 +8452,52 @@ simdjson_warn_unused simdjson_really_inline error_code json_iterator::walk_docum
 
 } // walk_document()
 
-simdjson_really_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+simdjson_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
   : buf{_dom_parser.buf},
     next_structural{&_dom_parser.structural_indexes[start_structural_index]},
     dom_parser{_dom_parser} {
 }
 
-simdjson_really_inline const uint8_t *json_iterator::peek() const noexcept {
+simdjson_inline const uint8_t *json_iterator::peek() const noexcept {
   return &buf[*(next_structural)];
 }
-simdjson_really_inline const uint8_t *json_iterator::advance() noexcept {
+simdjson_inline const uint8_t *json_iterator::advance() noexcept {
   return &buf[*(next_structural++)];
 }
-simdjson_really_inline size_t json_iterator::remaining_len() const noexcept {
+simdjson_inline size_t json_iterator::remaining_len() const noexcept {
   return dom_parser.len - *(next_structural-1);
 }
 
-simdjson_really_inline bool json_iterator::at_eof() const noexcept {
+simdjson_inline bool json_iterator::at_eof() const noexcept {
   return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
 }
-simdjson_really_inline bool json_iterator::at_beginning() const noexcept {
+simdjson_inline bool json_iterator::at_beginning() const noexcept {
   return next_structural == dom_parser.structural_indexes.get();
 }
-simdjson_really_inline uint8_t json_iterator::last_structural() const noexcept {
+simdjson_inline uint8_t json_iterator::last_structural() const noexcept {
   return buf[dom_parser.structural_indexes[dom_parser.n_structural_indexes - 1]];
 }
 
-simdjson_really_inline void json_iterator::log_value(const char *type) const noexcept {
+simdjson_inline void json_iterator::log_value(const char *type) const noexcept {
   logger::log_line(*this, "", type, "");
 }
 
-simdjson_really_inline void json_iterator::log_start_value(const char *type) const noexcept {
+simdjson_inline void json_iterator::log_start_value(const char *type) const noexcept {
   logger::log_line(*this, "+", type, "");
   if (logger::LOG_ENABLED) { logger::log_depth++; }
 }
 
-simdjson_really_inline void json_iterator::log_end_value(const char *type) const noexcept {
+simdjson_inline void json_iterator::log_end_value(const char *type) const noexcept {
   if (logger::LOG_ENABLED) { logger::log_depth--; }
   logger::log_line(*this, "-", type, "");
 }
 
-simdjson_really_inline void json_iterator::log_error(const char *error) const noexcept {
+simdjson_inline void json_iterator::log_error(const char *error) const noexcept {
   logger::log_line(*this, "", "ERROR", error);
 }
 
 template<typename V>
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
   switch (*value) {
     case '"': return visitor.visit_root_string(*this, value);
     case 't': return visitor.visit_root_true_atom(*this, value);
@@ -8056,7 +8513,7 @@ simdjson_warn_unused simdjson_really_inline error_code json_iterator::visit_root
   }
 }
 template<typename V>
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
   switch (*value) {
     case '"': return visitor.visit_string(*this, value);
     case 't': return visitor.visit_true_atom(*this, value);
@@ -8088,18 +8545,18 @@ struct tape_writer {
   uint64_t *next_tape_loc;
 
   /** Write a signed 64-bit value to tape. */
-  simdjson_really_inline void append_s64(int64_t value) noexcept;
+  simdjson_inline void append_s64(int64_t value) noexcept;
 
   /** Write an unsigned 64-bit value to tape. */
-  simdjson_really_inline void append_u64(uint64_t value) noexcept;
+  simdjson_inline void append_u64(uint64_t value) noexcept;
 
   /** Write a double value to tape. */
-  simdjson_really_inline void append_double(double value) noexcept;
+  simdjson_inline void append_double(double value) noexcept;
 
   /**
    * Append a tape entry (an 8-bit type,and 56 bits worth of value).
    */
-  simdjson_really_inline void append(uint64_t val, internal::tape_type t) noexcept;
+  simdjson_inline void append(uint64_t val, internal::tape_type t) noexcept;
 
   /**
    * Skip the current tape entry without writing.
@@ -8107,24 +8564,24 @@ struct tape_writer {
    * Used to skip the start of the container, since we'll come back later to fill it in when the
    * container ends.
    */
-  simdjson_really_inline void skip() noexcept;
+  simdjson_inline void skip() noexcept;
 
   /**
    * Skip the number of tape entries necessary to write a large u64 or i64.
    */
-  simdjson_really_inline void skip_large_integer() noexcept;
+  simdjson_inline void skip_large_integer() noexcept;
 
   /**
    * Skip the number of tape entries necessary to write a double.
    */
-  simdjson_really_inline void skip_double() noexcept;
+  simdjson_inline void skip_double() noexcept;
 
   /**
    * Write a value to a known location on tape.
    *
    * Used to go back and write out the start of a container after the container ends.
    */
-  simdjson_really_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
+  simdjson_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
 
 private:
   /**
@@ -8132,50 +8589,50 @@ struct tape_writer {
    * all 64 bits, such as double and uint64_t.
    */
   template<typename T>
-  simdjson_really_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
+  simdjson_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
 }; // struct number_writer
 
-simdjson_really_inline void tape_writer::append_s64(int64_t value) noexcept {
+simdjson_inline void tape_writer::append_s64(int64_t value) noexcept {
   append2(0, value, internal::tape_type::INT64);
 }
 
-simdjson_really_inline void tape_writer::append_u64(uint64_t value) noexcept {
+simdjson_inline void tape_writer::append_u64(uint64_t value) noexcept {
   append(0, internal::tape_type::UINT64);
   *next_tape_loc = value;
   next_tape_loc++;
 }
 
 /** Write a double value to tape. */
-simdjson_really_inline void tape_writer::append_double(double value) noexcept {
+simdjson_inline void tape_writer::append_double(double value) noexcept {
   append2(0, value, internal::tape_type::DOUBLE);
 }
 
-simdjson_really_inline void tape_writer::skip() noexcept {
+simdjson_inline void tape_writer::skip() noexcept {
   next_tape_loc++;
 }
 
-simdjson_really_inline void tape_writer::skip_large_integer() noexcept {
+simdjson_inline void tape_writer::skip_large_integer() noexcept {
   next_tape_loc += 2;
 }
 
-simdjson_really_inline void tape_writer::skip_double() noexcept {
+simdjson_inline void tape_writer::skip_double() noexcept {
   next_tape_loc += 2;
 }
 
-simdjson_really_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
+simdjson_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
   *next_tape_loc = val | ((uint64_t(char(t))) << 56);
   next_tape_loc++;
 }
 
 template<typename T>
-simdjson_really_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
+simdjson_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
   append(val, t);
   static_assert(sizeof(val2) == sizeof(*next_tape_loc), "Type is not 64 bits!");
   memcpy(next_tape_loc, &val2, sizeof(val2));
   next_tape_loc++;
 }
 
-simdjson_really_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
+simdjson_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
   tape_loc = val | ((uint64_t(char(t))) << 56);
 }
 
@@ -8192,40 +8649,40 @@ namespace stage2 {
 
 struct tape_builder {
   template<bool STREAMING>
-  simdjson_warn_unused static simdjson_really_inline error_code parse_document(
+  simdjson_warn_unused static simdjson_inline error_code parse_document(
     dom_parser_implementation &dom_parser,
     dom::document &doc) noexcept;
 
   /** Called when a non-empty document starts. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_document_start(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_document_start(json_iterator &iter) noexcept;
   /** Called when a non-empty document ends without error. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_document_end(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_document_end(json_iterator &iter) noexcept;
 
   /** Called when a non-empty array starts. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_array_start(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_array_start(json_iterator &iter) noexcept;
   /** Called when a non-empty array ends. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_array_end(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_array_end(json_iterator &iter) noexcept;
   /** Called when an empty array is found. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_empty_array(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_empty_array(json_iterator &iter) noexcept;
 
   /** Called when a non-empty object starts. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_object_start(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_object_start(json_iterator &iter) noexcept;
   /**
    * Called when a key in a field is encountered.
    *
    * primitive, visit_object_start, visit_empty_object, visit_array_start, or visit_empty_array
    * will be called after this with the field value.
    */
-  simdjson_warn_unused simdjson_really_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
   /** Called when a non-empty object ends. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_object_end(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_object_end(json_iterator &iter) noexcept;
   /** Called when an empty object is found. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_empty_object(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_empty_object(json_iterator &iter) noexcept;
 
   /**
    * Called when a string, number, boolean or null is found.
    */
-  simdjson_warn_unused simdjson_really_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
   /**
    * Called when a string, number, boolean or null is found at the top level of a document (i.e.
    * when there is no array or object and the entire document is a single string, number, boolean or
@@ -8234,22 +8691,22 @@ struct tape_builder {
    * This is separate from primitive() because simdjson's normal primitive parsing routines assume
    * there is at least one more token after the value, which is only true in an array or object.
    */
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
 
-  simdjson_warn_unused simdjson_really_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
 
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
 
   /** Called each time a new field or element in an array or object is found. */
-  simdjson_warn_unused simdjson_really_inline error_code increment_count(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code increment_count(json_iterator &iter) noexcept;
 
   /** Next location to write to tape */
   tape_writer tape;
@@ -8257,18 +8714,18 @@ struct tape_builder {
   /** Next write location in the string buf for stage 2 parsing */
   uint8_t *current_string_buf_loc;
 
-  simdjson_really_inline tape_builder(dom::document &doc) noexcept;
+  simdjson_inline tape_builder(dom::document &doc) noexcept;
 
-  simdjson_really_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
-  simdjson_really_inline void start_container(json_iterator &iter) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
-  simdjson_really_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
-  simdjson_really_inline void on_end_string(uint8_t *dst) noexcept;
+  simdjson_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
+  simdjson_inline void start_container(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_warn_unused simdjson_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
+  simdjson_inline void on_end_string(uint8_t *dst) noexcept;
 }; // class tape_builder
 
 template<bool STREAMING>
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::parse_document(
+simdjson_warn_unused simdjson_inline error_code tape_builder::parse_document(
     dom_parser_implementation &dom_parser,
     dom::document &doc) noexcept {
   dom_parser.doc = &doc;
@@ -8277,56 +8734,56 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::parse_docum
   return iter.walk_document<STREAMING>(builder);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
   return iter.visit_root_primitive(*this, value);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
   return iter.visit_primitive(*this, value);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
   return empty_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
   return empty_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
   start_container(iter);
   return SUCCESS;
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
   start_container(iter);
   return SUCCESS;
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
   start_container(iter);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
   return end_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
   return end_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
   constexpr uint32_t start_tape_index = 0;
   tape.append(start_tape_index, internal::tape_type::ROOT);
   tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter), internal::tape_type::ROOT);
   return SUCCESS;
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
   return visit_string(iter, key, true);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
   iter.dom_parser.open_containers[iter.depth].count++; // we have a key value pair in the object at parser.dom_parser.depth - 1
   return SUCCESS;
 }
 
-simdjson_really_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
+simdjson_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
   iter.log_value(key ? "key" : "string");
   uint8_t *dst = on_start_string(iter);
   dst = stringparsing::parse_string(value+1, dst);
@@ -8338,16 +8795,16 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_strin
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
   return visit_string(iter, value);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("number");
   return numberparsing::parse_number(value, tape);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
   //
   // We need to make a copy to make sure that the string is space terminated.
   // This is not about padding the input, which should already padded up
@@ -8369,42 +8826,42 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_
   return error;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("true");
   if (!atomparsing::is_valid_true_atom(value)) { return T_ATOM_ERROR; }
   tape.append(0, internal::tape_type::TRUE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("true");
   if (!atomparsing::is_valid_true_atom(value, iter.remaining_len())) { return T_ATOM_ERROR; }
   tape.append(0, internal::tape_type::TRUE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("false");
   if (!atomparsing::is_valid_false_atom(value)) { return F_ATOM_ERROR; }
   tape.append(0, internal::tape_type::FALSE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("false");
   if (!atomparsing::is_valid_false_atom(value, iter.remaining_len())) { return F_ATOM_ERROR; }
   tape.append(0, internal::tape_type::FALSE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("null");
   if (!atomparsing::is_valid_null_atom(value)) { return N_ATOM_ERROR; }
   tape.append(0, internal::tape_type::NULL_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("null");
   if (!atomparsing::is_valid_null_atom(value, iter.remaining_len())) { return N_ATOM_ERROR; }
   tape.append(0, internal::tape_type::NULL_VALUE);
@@ -8413,24 +8870,24 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_
 
 // private:
 
-simdjson_really_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
+simdjson_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
   return uint32_t(tape.next_tape_loc - iter.dom_parser.doc->tape.get());
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
   auto start_index = next_tape_index(iter);
   tape.append(start_index+2, start);
   tape.append(start_index, end);
   return SUCCESS;
 }
 
-simdjson_really_inline void tape_builder::start_container(json_iterator &iter) noexcept {
+simdjson_inline void tape_builder::start_container(json_iterator &iter) noexcept {
   iter.dom_parser.open_containers[iter.depth].tape_index = next_tape_index(iter);
   iter.dom_parser.open_containers[iter.depth].count = 0;
   tape.skip(); // We don't actually *write* the start element until the end.
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
   // Write the ending tape element, pointing at the start location
   const uint32_t start_tape_index = iter.dom_parser.open_containers[iter.depth].tape_index;
   tape.append(start_tape_index, end);
@@ -8443,13 +8900,13 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::end_contain
   return SUCCESS;
 }
 
-simdjson_really_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
+simdjson_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
   // we advance the point, accounting for the fact that we have a NULL termination
   tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::STRING);
   return current_string_buf_loc + sizeof(uint32_t);
 }
 
-simdjson_really_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
+simdjson_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
   uint32_t str_length = uint32_t(dst - (current_string_buf_loc + sizeof(uint32_t)));
   // TODO check for overflow in case someone has a crazy string (>=4GB?)
   // But only add the overflow check when the document itself exceeds 4GB
@@ -8475,7 +8932,7 @@ namespace icelake {
 namespace {
 namespace stage1 {
 
-simdjson_really_inline uint64_t json_string_scanner::find_escaped(uint64_t backslash) {
+simdjson_inline uint64_t json_string_scanner::find_escaped(uint64_t backslash) {
   if (!backslash) { uint64_t escaped = prev_escaped; prev_escaped = 0; return escaped; }
   return find_escaped_branchless(backslash);
 }
@@ -8505,6 +8962,10 @@ simdjson_warn_unused error_code dom_parser_implementation::stage2_next(dom::docu
   return stage2::tape_builder::parse_document<true>(*this, _doc);
 }
 
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_string(const uint8_t *src, uint8_t *dst) const noexcept {
+  return icelake::stringparsing::parse_string(src, dst);
+}
+
 simdjson_warn_unused error_code dom_parser_implementation::parse(const uint8_t *_buf, size_t _len, dom::document &_doc) noexcept {
   auto error = stage1(_buf, _len, stage1_mode::regular);
   if (error) { return error; }
@@ -8570,25 +9031,25 @@ namespace {
 using namespace simd;
 
 struct json_character_block {
-  static simdjson_really_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
+  static simdjson_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
   //  ASCII white-space ('\r','\n','\t',' ')
-  simdjson_really_inline uint64_t whitespace() const noexcept;
+  simdjson_inline uint64_t whitespace() const noexcept;
   // non-quote structural characters (comma, colon, braces, brackets)
-  simdjson_really_inline uint64_t op() const noexcept;
+  simdjson_inline uint64_t op() const noexcept;
   // neither a structural character nor a white-space, so letters, numbers and quotes
-  simdjson_really_inline uint64_t scalar() const noexcept;
+  simdjson_inline uint64_t scalar() const noexcept;
 
   uint64_t _whitespace; // ASCII white-space ('\r','\n','\t',' ')
   uint64_t _op; // structural characters (comma, colon, braces, brackets but not quotes)
 };
 
-simdjson_really_inline uint64_t json_character_block::whitespace() const noexcept { return _whitespace; }
-simdjson_really_inline uint64_t json_character_block::op() const noexcept { return _op; }
-simdjson_really_inline uint64_t json_character_block::scalar() const noexcept { return ~(op() | whitespace()); }
+simdjson_inline uint64_t json_character_block::whitespace() const noexcept { return _whitespace; }
+simdjson_inline uint64_t json_character_block::op() const noexcept { return _op; }
+simdjson_inline uint64_t json_character_block::scalar() const noexcept { return ~(op() | whitespace()); }
 
 // This identifies structural characters (comma, colon, braces, brackets),
 // and ASCII white-space ('\r','\n','\t',' ').
-simdjson_really_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
+simdjson_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
   // These lookups rely on the fact that anything < 127 will match the lower 4 bits, which is why
   // we can't use the generic lookup_16.
   const auto whitespace_table = simd8<uint8_t>::repeat_16(' ', 100, 100, 100, 17, 100, 113, 2, 100, '\t', '\n', 112, 100, '\r', 100, 100);
@@ -8641,21 +9102,21 @@ simdjson_really_inline json_character_block json_character_block::classify(const
   return { whitespace, op };
 }
 
-simdjson_really_inline bool is_ascii(const simd8x64<uint8_t>& input) {
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input) {
   return input.reduce_or().is_ascii();
 }
 
-simdjson_unused simdjson_really_inline simd8<bool> must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
-  simd8<uint8_t> is_second_byte = prev1.saturating_sub(0b11000000u-1); // Only 11______ will be > 0
-  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0b11100000u-1); // Only 111_____ will be > 0
-  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0b11110000u-1); // Only 1111____ will be > 0
+simdjson_unused simdjson_inline simd8<bool> must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_second_byte = prev1.saturating_sub(0xc0u-1); // Only 11______ will be > 0
+  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-1); // Only 111_____ will be > 0
+  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-1); // Only 1111____ will be > 0
   // Caller requires a bool (all 1's). All values resulting from the subtraction will be <= 64, so signed comparison is fine.
   return simd8<int8_t>(is_second_byte | is_third_byte | is_fourth_byte) > int8_t(0);
 }
 
-simdjson_really_inline simd8<bool> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
-  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0b11100000u-1); // Only 111_____ will be > 0
-  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0b11110000u-1); // Only 1111____ will be > 0
+simdjson_inline simd8<bool> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-1); // Only 111_____ will be > 0
+  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-1); // Only 1111____ will be > 0
   // Caller requires a bool (all 1's). All values resulting from the subtraction will be <= 64, so signed comparison is fine.
   return simd8<int8_t>(is_third_byte | is_fourth_byte) > int8_t(0);
 }
@@ -8672,7 +9133,7 @@ namespace utf8_validation {
 
 using namespace simd;
 
-  simdjson_really_inline simd8<uint8_t> check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+  simdjson_inline simd8<uint8_t> check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
 // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
 // Bit 1 = Too Long (ASCII followed by continuation)
 // Bit 2 = Overlong 3-byte
@@ -8762,7 +9223,7 @@ using namespace simd;
     );
     return (byte_1_high & byte_1_low & byte_2_high);
   }
-  simdjson_really_inline simd8<uint8_t> check_multibyte_lengths(const simd8<uint8_t> input,
+  simdjson_inline simd8<uint8_t> check_multibyte_lengths(const simd8<uint8_t> input,
       const simd8<uint8_t> prev_input, const simd8<uint8_t> sc) {
     simd8<uint8_t> prev2 = input.prev<2>(prev_input);
     simd8<uint8_t> prev3 = input.prev<3>(prev_input);
@@ -8775,7 +9236,7 @@ using namespace simd;
   // Return nonzero if there are incomplete multibyte characters at the end of the block:
   // e.g. if there is a 4-byte character, but it's 3 bytes from the end.
   //
-  simdjson_really_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
+  simdjson_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
     // If the previous input's last 3 bytes match this, they're too short (they ended at EOF):
     // ... 1111____ 111_____ 11______
 #if SIMDJSON_IMPLEMENTATION_ICELAKE
@@ -8787,14 +9248,14 @@ using namespace simd;
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
-      255, 255, 255, 255, 255, 0b11110000u-1, 0b11100000u-1, 0b11000000u-1
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
     };
 #else
     static const uint8_t max_array[32] = {
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
-      255, 255, 255, 255, 255, 0b11110000u-1, 0b11100000u-1, 0b11000000u-1
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
     };
 #endif
     const simd8<uint8_t> max_value(&max_array[sizeof(max_array)-sizeof(simd8<uint8_t>)]);
@@ -8812,7 +9273,7 @@ using namespace simd;
     //
     // Check whether the current bytes are valid UTF-8.
     //
-    simdjson_really_inline void check_utf8_bytes(const simd8<uint8_t> input, const simd8<uint8_t> prev_input) {
+    simdjson_inline void check_utf8_bytes(const simd8<uint8_t> input, const simd8<uint8_t> prev_input) {
       // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ lead bytes
       // (2, 3, 4-byte leads become large positive numbers instead of small negative numbers)
       simd8<uint8_t> prev1 = input.prev<1>(prev_input);
@@ -8823,13 +9284,21 @@ using namespace simd;
     // The only problem that can happen at EOF is that a multibyte character is too short
     // or a byte value too large in the last bytes: check_special_cases only checks for bytes
     // too large in the first of two bytes.
-    simdjson_really_inline void check_eof() {
+    simdjson_inline void check_eof() {
       // If the previous block had incomplete UTF-8 characters at the end, an ASCII block can't
       // possibly finish them.
       this->error |= this->prev_incomplete;
     }
 
-    simdjson_really_inline void check_next_input(const simd8x64<uint8_t>& input) {
+#ifndef SIMDJSON_IF_CONSTEXPR
+#if SIMDJSON_CPLUSPLUS17
+#define SIMDJSON_IF_CONSTEXPR if constexpr
+#else
+#define SIMDJSON_IF_CONSTEXPR if
+#endif
+#endif
+
+    simdjson_inline void check_next_input(const simd8x64<uint8_t>& input) {
       if(simdjson_likely(is_ascii(input))) {
         this->error |= this->prev_incomplete;
       } else {
@@ -8838,12 +9307,12 @@ using namespace simd;
                 ||(simd8x64<uint8_t>::NUM_CHUNKS == 2)
                 || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
                 "We support one, two or four chunks per 64-byte block.");
-        if(simd8x64<uint8_t>::NUM_CHUNKS == 1) {
+        SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 1) {
           this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
-        } if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
           this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
           this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
           this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
           this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
           this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
@@ -8854,7 +9323,7 @@ using namespace simd;
       }
     }
     // do not forget to call check_eof!
-    simdjson_really_inline error_code errors() {
+    simdjson_inline error_code errors() {
       return this->error.any_bits_set_anywhere() ? error_code::UTF8_ERROR : error_code::SUCCESS;
     }
 
@@ -8882,10 +9351,10 @@ namespace {
 template<size_t STEP_SIZE>
 struct buf_block_reader {
 public:
-  simdjson_really_inline buf_block_reader(const uint8_t *_buf, size_t _len);
-  simdjson_really_inline size_t block_index();
-  simdjson_really_inline bool has_full_block() const;
-  simdjson_really_inline const uint8_t *full_block() const;
+  simdjson_inline buf_block_reader(const uint8_t *_buf, size_t _len);
+  simdjson_inline size_t block_index();
+  simdjson_inline bool has_full_block() const;
+  simdjson_inline const uint8_t *full_block() const;
   /**
    * Get the last block, padded with spaces.
    *
@@ -8895,8 +9364,8 @@ struct buf_block_reader {
    *
    * @return the number of effective characters in the last block.
    */
-  simdjson_really_inline size_t get_remainder(uint8_t *dst) const;
-  simdjson_really_inline void advance();
+  simdjson_inline size_t get_remainder(uint8_t *dst) const;
+  simdjson_inline void advance();
 private:
   const uint8_t *buf;
   const size_t len;
@@ -8935,23 +9404,23 @@ simdjson_unused static char * format_mask(uint64_t mask) {
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len) : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, idx{0} {}
+simdjson_inline buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len) : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, idx{0} {}
 
 template<size_t STEP_SIZE>
-simdjson_really_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
+simdjson_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
   return idx < lenminusstep;
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
+simdjson_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
   return &buf[idx];
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
   if(len == idx) { return 0; } // memcpy(dst, null, 0) will trigger an error with some sanitizers
   std::memset(dst, 0x20, STEP_SIZE); // std::memset STEP_SIZE because it's more efficient to write out 8 or 16 bytes at once.
   std::memcpy(dst, buf + idx, len - idx);
@@ -8959,7 +9428,7 @@ simdjson_really_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline void buf_block_reader<STEP_SIZE>::advance() {
+simdjson_inline void buf_block_reader<STEP_SIZE>::advance() {
   idx += STEP_SIZE;
 }
 
@@ -8975,27 +9444,27 @@ namespace stage1 {
 
 struct json_string_block {
   // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
-  simdjson_really_inline json_string_block(uint64_t backslash, uint64_t escaped, uint64_t quote, uint64_t in_string) :
+  simdjson_inline json_string_block(uint64_t backslash, uint64_t escaped, uint64_t quote, uint64_t in_string) :
   _backslash(backslash), _escaped(escaped), _quote(quote), _in_string(in_string) {}
 
   // Escaped characters (characters following an escape() character)
-  simdjson_really_inline uint64_t escaped() const { return _escaped; }
+  simdjson_inline uint64_t escaped() const { return _escaped; }
   // Escape characters (backslashes that are not escaped--i.e. in \\, includes only the first \)
-  simdjson_really_inline uint64_t escape() const { return _backslash & ~_escaped; }
+  simdjson_inline uint64_t escape() const { return _backslash & ~_escaped; }
   // Real (non-backslashed) quotes
-  simdjson_really_inline uint64_t quote() const { return _quote; }
+  simdjson_inline uint64_t quote() const { return _quote; }
   // Start quotes of strings
-  simdjson_really_inline uint64_t string_start() const { return _quote & _in_string; }
+  simdjson_inline uint64_t string_start() const { return _quote & _in_string; }
   // End quotes of strings
-  simdjson_really_inline uint64_t string_end() const { return _quote & ~_in_string; }
+  simdjson_inline uint64_t string_end() const { return _quote & ~_in_string; }
   // Only characters inside the string (not including the quotes)
-  simdjson_really_inline uint64_t string_content() const { return _in_string & ~_quote; }
+  simdjson_inline uint64_t string_content() const { return _in_string & ~_quote; }
   // Return a mask of whether the given characters are inside a string (only works on non-quotes)
-  simdjson_really_inline uint64_t non_quote_inside_string(uint64_t mask) const { return mask & _in_string; }
+  simdjson_inline uint64_t non_quote_inside_string(uint64_t mask) const { return mask & _in_string; }
   // Return a mask of whether the given characters are inside a string (only works on non-quotes)
-  simdjson_really_inline uint64_t non_quote_outside_string(uint64_t mask) const { return mask & ~_in_string; }
+  simdjson_inline uint64_t non_quote_outside_string(uint64_t mask) const { return mask & ~_in_string; }
   // Tail of string (everything except the start quote)
-  simdjson_really_inline uint64_t string_tail() const { return _in_string ^ _quote; }
+  simdjson_inline uint64_t string_tail() const { return _in_string ^ _quote; }
 
   // backslash characters
   uint64_t _backslash;
@@ -9010,14 +9479,14 @@ struct json_string_block {
 // Scans blocks for string characters, storing the state necessary to do so
 class json_string_scanner {
 public:
-  simdjson_really_inline json_string_block next(const simd::simd8x64<uint8_t>& in);
+  simdjson_inline json_string_block next(const simd::simd8x64<uint8_t>& in);
   // Returns either UNCLOSED_STRING or SUCCESS
-  simdjson_really_inline error_code finish();
+  simdjson_inline error_code finish();
 
 private:
   // Intended to be defined by the implementation
-  simdjson_really_inline uint64_t find_escaped(uint64_t escape);
-  simdjson_really_inline uint64_t find_escaped_branchless(uint64_t escape);
+  simdjson_inline uint64_t find_escaped(uint64_t escape);
+  simdjson_inline uint64_t find_escaped_branchless(uint64_t escape);
 
   // Whether the last iteration was still inside a string (all 1's = true, all 0's = false).
   uint64_t prev_in_string = 0ULL;
@@ -9052,7 +9521,7 @@ class json_string_scanner {
 // desired        |   x  | x x  x x  x x  x  x  |
 // text           |  \\\ | \\\"\\\" \\\" \\"\\" |
 //
-simdjson_really_inline uint64_t json_string_scanner::find_escaped_branchless(uint64_t backslash) {
+simdjson_inline uint64_t json_string_scanner::find_escaped_branchless(uint64_t backslash) {
   // If there was overflow, pretend the first character isn't a backslash
   backslash &= ~prev_escaped;
   uint64_t follows_escape = backslash << 1 | prev_escaped;
@@ -9077,7 +9546,7 @@ simdjson_really_inline uint64_t json_string_scanner::find_escaped_branchless(uin
 //
 // Backslash sequences outside of quotes will be detected in stage 2.
 //
-simdjson_really_inline json_string_block json_string_scanner::next(const simd::simd8x64<uint8_t>& in) {
+simdjson_inline json_string_block json_string_scanner::next(const simd::simd8x64<uint8_t>& in) {
   const uint64_t backslash = in.eq('\\');
   const uint64_t escaped = find_escaped(backslash);
   const uint64_t quote = in.eq('"') & ~escaped;
@@ -9110,7 +9579,7 @@ simdjson_really_inline json_string_block json_string_scanner::next(const simd::s
   );
 }
 
-simdjson_really_inline error_code json_string_scanner::finish() {
+simdjson_inline error_code json_string_scanner::finish() {
   if (prev_in_string) {
     return UNCLOSED_STRING;
   }
@@ -9148,25 +9617,25 @@ namespace stage1 {
 struct json_block {
 public:
   // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
-  simdjson_really_inline json_block(json_string_block&& string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  simdjson_inline json_block(json_string_block&& string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
   _string(std::move(string)), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
-  simdjson_really_inline json_block(json_string_block string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  simdjson_inline json_block(json_string_block string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
   _string(string), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
 
   /**
    * The start of structurals.
    * In simdjson prior to v0.3, these were called the pseudo-structural characters.
    **/
-  simdjson_really_inline uint64_t structural_start() const noexcept { return potential_structural_start() & ~_string.string_tail(); }
+  simdjson_inline uint64_t structural_start() const noexcept { return potential_structural_start() & ~_string.string_tail(); }
   /** All JSON whitespace (i.e. not in a string) */
-  simdjson_really_inline uint64_t whitespace() const noexcept { return non_quote_outside_string(_characters.whitespace()); }
+  simdjson_inline uint64_t whitespace() const noexcept { return non_quote_outside_string(_characters.whitespace()); }
 
   // Helpers
 
   /** Whether the given characters are inside a string (only works on non-quotes) */
-  simdjson_really_inline uint64_t non_quote_inside_string(uint64_t mask) const noexcept { return _string.non_quote_inside_string(mask); }
+  simdjson_inline uint64_t non_quote_inside_string(uint64_t mask) const noexcept { return _string.non_quote_inside_string(mask); }
   /** Whether the given characters are outside a string (only works on non-quotes) */
-  simdjson_really_inline uint64_t non_quote_outside_string(uint64_t mask) const noexcept { return _string.non_quote_outside_string(mask); }
+  simdjson_inline uint64_t non_quote_outside_string(uint64_t mask) const noexcept { return _string.non_quote_outside_string(mask); }
 
   // string and escape characters
   json_string_block _string;
@@ -9181,12 +9650,12 @@ struct json_block {
    * structural elements ([,],{,},:, comma) plus scalar starts like 123, true and "abc".
    * They may reside inside a string.
    **/
-  simdjson_really_inline uint64_t potential_structural_start() const noexcept { return _characters.op() | potential_scalar_start(); }
+  simdjson_inline uint64_t potential_structural_start() const noexcept { return _characters.op() | potential_scalar_start(); }
   /**
    * The start of non-operator runs, like 123, true and "abc".
    * It main reside inside a string.
    **/
-  simdjson_really_inline uint64_t potential_scalar_start() const noexcept {
+  simdjson_inline uint64_t potential_scalar_start() const noexcept {
     // The term "scalar" refers to anything except structural characters and white space
     // (so letters, numbers, quotes).
     // Whenever it is preceded by something that is not a structural element ({,},[,],:, ") nor a white-space
@@ -9197,7 +9666,7 @@ struct json_block {
    * Whether the given character is immediately after a non-operator like 123, true.
    * The characters following a quote are not included.
    */
-  simdjson_really_inline uint64_t follows_potential_scalar() const noexcept {
+  simdjson_inline uint64_t follows_potential_scalar() const noexcept {
     // _follows_potential_nonquote_scalar: is defined as marking any character that follows a character
     // that is not a structural element ({,},[,],:, comma) nor a quote (") and that is not a
     // white space.
@@ -9221,10 +9690,10 @@ struct json_block {
  */
 class json_scanner {
 public:
-  json_scanner() {}
-  simdjson_really_inline json_block next(const simd::simd8x64<uint8_t>& in);
+  json_scanner() = default;
+  simdjson_inline json_block next(const simd::simd8x64<uint8_t>& in);
   // Returns either UNCLOSED_STRING or SUCCESS
-  simdjson_really_inline error_code finish();
+  simdjson_inline error_code finish();
 
 private:
   // Whether the last character of the previous iteration is part of a scalar token
@@ -9241,13 +9710,13 @@ class json_scanner {
 //
 //     const uint64_t backslashed_quote = in.eq('"') & immediately_follows(in.eq('\'), prev_backslash);
 //
-simdjson_really_inline uint64_t follows(const uint64_t match, uint64_t &overflow) {
+simdjson_inline uint64_t follows(const uint64_t match, uint64_t &overflow) {
   const uint64_t result = match << 1 | overflow;
   overflow = match >> 63;
   return result;
 }
 
-simdjson_really_inline json_block json_scanner::next(const simd::simd8x64<uint8_t>& in) {
+simdjson_inline json_block json_scanner::next(const simd::simd8x64<uint8_t>& in) {
   json_string_block strings = string_scanner.next(in);
   // identifies the white-space and the structural characters
   json_character_block characters = json_character_block::classify(in);
@@ -9272,7 +9741,7 @@ simdjson_really_inline json_block json_scanner::next(const simd::simd8x64<uint8_
   );
 }
 
-simdjson_really_inline error_code json_scanner::finish() {
+simdjson_inline error_code json_scanner::finish() {
   return string_scanner.finish();
 }
 
@@ -9298,23 +9767,23 @@ class json_minifier {
   static error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept;
 
 private:
-  simdjson_really_inline json_minifier(uint8_t *_dst)
+  simdjson_inline json_minifier(uint8_t *_dst)
   : dst{_dst}
   {}
   template<size_t STEP_SIZE>
-  simdjson_really_inline void step(const uint8_t *block_buf, buf_block_reader<STEP_SIZE> &reader) noexcept;
-  simdjson_really_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block);
-  simdjson_really_inline error_code finish(uint8_t *dst_start, size_t &dst_len);
+  simdjson_inline void step(const uint8_t *block_buf, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block);
+  simdjson_inline error_code finish(uint8_t *dst_start, size_t &dst_len);
   json_scanner scanner{};
   uint8_t *dst;
 };
 
-simdjson_really_inline void json_minifier::next(const simd::simd8x64<uint8_t>& in, const json_block& block) {
+simdjson_inline void json_minifier::next(const simd::simd8x64<uint8_t>& in, const json_block& block) {
   uint64_t mask = block.whitespace();
   dst += in.compress(mask, dst);
 }
 
-simdjson_really_inline error_code json_minifier::finish(uint8_t *dst_start, size_t &dst_len) {
+simdjson_inline error_code json_minifier::finish(uint8_t *dst_start, size_t &dst_len) {
   error_code error = scanner.finish();
   if (error) { dst_len = 0; return error; }
   dst_len = dst - dst_start;
@@ -9322,7 +9791,7 @@ simdjson_really_inline error_code json_minifier::finish(uint8_t *dst_start, size
 }
 
 template<>
-simdjson_really_inline void json_minifier::step<128>(const uint8_t *block_buf, buf_block_reader<128> &reader) noexcept {
+simdjson_inline void json_minifier::step<128>(const uint8_t *block_buf, buf_block_reader<128> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block_buf);
   simd::simd8x64<uint8_t> in_2(block_buf+64);
   json_block block_1 = scanner.next(in_1);
@@ -9333,7 +9802,7 @@ simdjson_really_inline void json_minifier::step<128>(const uint8_t *block_buf, b
 }
 
 template<>
-simdjson_really_inline void json_minifier::step<64>(const uint8_t *block_buf, buf_block_reader<64> &reader) noexcept {
+simdjson_inline void json_minifier::step<64>(const uint8_t *block_buf, buf_block_reader<64> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block_buf);
   json_block block_1 = scanner.next(in_1);
   this->next(block_buf, block_1);
@@ -9406,7 +9875,7 @@ namespace {
   * complete document, therefore the last json buffer location is the end of the
   * batch.
   */
-simdjson_really_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
+simdjson_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
   // Variant: do not count separately, just figure out depth
   if(parser.n_structural_indexes == 0) { return 0; }
   auto arr_cnt = 0;
@@ -9481,7 +9950,7 @@ class bit_indexer {
 public:
   uint32_t *tail;
 
-  simdjson_really_inline bit_indexer(uint32_t *index_buf) : tail(index_buf) {}
+  simdjson_inline bit_indexer(uint32_t *index_buf) : tail(index_buf) {}
 
   // flatten out values in 'bits' assuming that they are are to have values of idx
   // plus their position in the bitvector, and store these indexes at
@@ -9492,9 +9961,9 @@ class bit_indexer {
   // If the kernel sets SIMDJSON_CUSTOM_BIT_INDEXER, then it will provide its own
   // version of the code.
 #ifdef SIMDJSON_CUSTOM_BIT_INDEXER
-  simdjson_really_inline void write(uint32_t idx, uint64_t bits);
+  simdjson_inline void write(uint32_t idx, uint64_t bits);
 #else
-  simdjson_really_inline void write(uint32_t idx, uint64_t bits) {
+  simdjson_inline void write(uint32_t idx, uint64_t bits) {
     // In some instances, the next branch is expensive because it is mispredicted.
     // Unfortunately, in other cases,
     // it helps tremendously.
@@ -9603,11 +10072,11 @@ class json_structural_indexer {
   static error_code index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept;
 
 private:
-  simdjson_really_inline json_structural_indexer(uint32_t *structural_indexes);
+  simdjson_inline json_structural_indexer(uint32_t *structural_indexes);
   template<size_t STEP_SIZE>
-  simdjson_really_inline void step(const uint8_t *block, buf_block_reader<STEP_SIZE> &reader) noexcept;
-  simdjson_really_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx);
-  simdjson_really_inline error_code finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial);
+  simdjson_inline void step(const uint8_t *block, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx);
+  simdjson_inline error_code finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial);
 
   json_scanner scanner{};
   utf8_checker checker{};
@@ -9616,26 +10085,26 @@ class json_structural_indexer {
   uint64_t unescaped_chars_error = 0;
 };
 
-simdjson_really_inline json_structural_indexer::json_structural_indexer(uint32_t *structural_indexes) : indexer{structural_indexes} {}
+simdjson_inline json_structural_indexer::json_structural_indexer(uint32_t *structural_indexes) : indexer{structural_indexes} {}
 
 // Skip the last character if it is partial
-simdjson_really_inline size_t trim_partial_utf8(const uint8_t *buf, size_t len) {
+simdjson_inline size_t trim_partial_utf8(const uint8_t *buf, size_t len) {
   if (simdjson_unlikely(len < 3)) {
     switch (len) {
       case 2:
-        if (buf[len-1] >= 0b11000000) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
-        if (buf[len-2] >= 0b11100000) { return len-2; } // 3- and 4-byte characters with only 2 bytes left
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 2 bytes left
         return len;
       case 1:
-        if (buf[len-1] >= 0b11000000) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
         return len;
       case 0:
         return len;
     }
   }
-  if (buf[len-1] >= 0b11000000) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
-  if (buf[len-2] >= 0b11100000) { return len-2; } // 3- and 4-byte characters with only 1 byte left
-  if (buf[len-3] >= 0b11110000) { return len-3; } // 4-byte characters with only 3 bytes left
+  if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+  if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 1 byte left
+  if (buf[len-3] >= 0xf0) { return len-3; } // 4-byte characters with only 3 bytes left
   return len;
 }
 
@@ -9684,7 +10153,7 @@ error_code json_structural_indexer::index(const uint8_t *buf, size_t len, dom_pa
 }
 
 template<>
-simdjson_really_inline void json_structural_indexer::step<128>(const uint8_t *block, buf_block_reader<128> &reader) noexcept {
+simdjson_inline void json_structural_indexer::step<128>(const uint8_t *block, buf_block_reader<128> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block);
   simd::simd8x64<uint8_t> in_2(block+64);
   json_block block_1 = scanner.next(in_1);
@@ -9695,14 +10164,14 @@ simdjson_really_inline void json_structural_indexer::step<128>(const uint8_t *bl
 }
 
 template<>
-simdjson_really_inline void json_structural_indexer::step<64>(const uint8_t *block, buf_block_reader<64> &reader) noexcept {
+simdjson_inline void json_structural_indexer::step<64>(const uint8_t *block, buf_block_reader<64> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block);
   json_block block_1 = scanner.next(in_1);
   this->next(in_1, block_1, reader.block_index());
   reader.advance();
 }
 
-simdjson_really_inline void json_structural_indexer::next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx) {
+simdjson_inline void json_structural_indexer::next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx) {
   uint64_t unescaped = in.lteq(0x1F);
   checker.check_next_input(in);
   indexer.write(uint32_t(idx-64), prev_structurals); // Output *last* iteration's structurals to the parser
@@ -9710,7 +10179,7 @@ simdjson_really_inline void json_structural_indexer::next(const simd::simd8x64<u
   unescaped_chars_error |= block.non_quote_inside_string(unescaped);
 }
 
-simdjson_really_inline error_code json_structural_indexer::finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial) {
+simdjson_inline error_code json_structural_indexer::finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial) {
   // Write out the final iteration's structurals
   indexer.write(uint32_t(idx-64), prev_structurals);
   error_code error = scanner.finish();
@@ -9856,6 +10325,151 @@ bool generic_validate_utf8(const char * input, size_t length) {
 //
 // Stage 2
 //
+/* begin file src/generic/stage2/stringparsing.h */
+// This file contains the common code every implementation uses
+// It is intended to be included multiple times and compiled multiple times
+
+namespace simdjson {
+namespace haswell {
+namespace {
+/// @private
+namespace stringparsing {
+
+// begin copypasta
+// These chars yield themselves: " \ /
+// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
+// u not handled in this table as it's complex
+static const uint8_t escape_map[256] = {
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
+    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
+    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+};
+
+// handle a unicode codepoint
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr) {
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion isn't valid; we defer the check for this to inside the
+  // multilingual plane check
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) != ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      return false;
+    }
+    uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+
+    // We have already checked that the high surrogate is valid and
+    // (code_point - 0xd800) < 1024.
+    //
+    // Check that code_point_2 is in the range 0xdc00..0xdfff
+    // and that code_point_2 was parsed from valid hex.
+    uint32_t low_bit = code_point_2 - 0xdc00;
+    if (low_bit >> 10) {
+      return false;
+    }
+
+    code_point =
+        (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+    *src_ptr += 6;
+  } else if (code_point >= 0xdc00 && code_point <= 0xdfff) {
+      // If we encounter a low surrogate (not preceded by a high surrogate)
+      // then we have an error.
+      return false;
+  }
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+/**
+ * Unescape a valid UTF-8 string from src to dst, stopping at a final unescaped quote. There
+ * must be an unescaped quote terminating the string. It returns the final output
+ * position as pointer. In case of error (e.g., the string has bad escaped codes),
+ * then null_nullptrptr is returned. It is assumed that the output buffer is large
+ * enough. E.g., if src points at 'joe"', then dst needs to have four free bytes +
+ * SIMDJSON_PADDING bytes.
+ */
+simdjson_warn_unused simdjson_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst) {
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto bs_quote = backslash_and_quote::copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint(&src, &dst)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+  /* can't be reached */
+  return nullptr;
+}
+
+} // namespace stringparsing
+} // unnamed namespace
+} // namespace haswell
+} // namespace simdjson
+/* end file src/generic/stage2/stringparsing.h */
 /* begin file src/generic/stage2/tape_builder.h */
 /* begin file src/generic/stage2/json_iterator.h */
 /* begin file src/generic/stage2/logger.h */
@@ -9881,7 +10495,7 @@ namespace logger {
   static int log_depth; // Not threadsafe. Log only.
 
   // Helper to turn unprintable or newline characters into spaces
-  static simdjson_really_inline char printable_char(char c) {
+  static simdjson_inline char printable_char(char c) {
     if (c >= 0x20) {
       return c;
     } else {
@@ -9890,7 +10504,7 @@ namespace logger {
   }
 
   // Print the header and set up log_start
-  static simdjson_really_inline void log_start() {
+  static simdjson_inline void log_start() {
     if (LOG_ENABLED) {
       log_depth = 0;
       printf("\n");
@@ -9899,7 +10513,7 @@ namespace logger {
     }
   }
 
-  simdjson_unused static simdjson_really_inline void log_string(const char *message) {
+  simdjson_unused static simdjson_inline void log_string(const char *message) {
     if (LOG_ENABLED) {
       printf("%s\n", message);
     }
@@ -9907,7 +10521,7 @@ namespace logger {
 
   // Logs a single line from the stage 2 DOM parser
   template<typename S>
-  static simdjson_really_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
+  static simdjson_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
     if (LOG_ENABLED) {
       printf("| %*s%s%-*s ", log_depth*2, "", title_prefix, LOG_EVENT_LEN - log_depth*2 - int(strlen(title_prefix)), title);
       auto current_index = structurals.at_beginning() ? nullptr : structurals.next_structural-1;
@@ -9985,14 +10599,14 @@ class json_iterator {
    * - increment_count(iter) - each time a value is found in an array or object.
    */
   template<bool STREAMING, typename V>
-  simdjson_warn_unused simdjson_really_inline error_code walk_document(V &visitor) noexcept;
+  simdjson_warn_unused simdjson_inline error_code walk_document(V &visitor) noexcept;
 
   /**
    * Create an iterator capable of walking a JSON document.
    *
    * The document must have already passed through stage 1.
    */
-  simdjson_really_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
+  simdjson_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
 
   /**
    * Look at the next token.
@@ -10001,7 +10615,7 @@ class json_iterator {
    *
    * They may include invalid JSON as well (such as `1.2.3` or `ture`).
    */
-  simdjson_really_inline const uint8_t *peek() const noexcept;
+  simdjson_inline const uint8_t *peek() const noexcept;
   /**
    * Advance to the next token.
    *
@@ -10009,56 +10623,56 @@ class json_iterator {
    *
    * They may include invalid JSON as well (such as `1.2.3` or `ture`).
    */
-  simdjson_really_inline const uint8_t *advance() noexcept;
+  simdjson_inline const uint8_t *advance() noexcept;
   /**
    * Get the remaining length of the document, from the start of the current token.
    */
-  simdjson_really_inline size_t remaining_len() const noexcept;
+  simdjson_inline size_t remaining_len() const noexcept;
   /**
    * Check if we are at the end of the document.
    *
    * If this is true, there are no more tokens.
    */
-  simdjson_really_inline bool at_eof() const noexcept;
+  simdjson_inline bool at_eof() const noexcept;
   /**
    * Check if we are at the beginning of the document.
    */
-  simdjson_really_inline bool at_beginning() const noexcept;
-  simdjson_really_inline uint8_t last_structural() const noexcept;
+  simdjson_inline bool at_beginning() const noexcept;
+  simdjson_inline uint8_t last_structural() const noexcept;
 
   /**
    * Log that a value has been found.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_value(const char *type) const noexcept;
+  simdjson_inline void log_value(const char *type) const noexcept;
   /**
    * Log the start of a multipart value.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_start_value(const char *type) const noexcept;
+  simdjson_inline void log_start_value(const char *type) const noexcept;
   /**
    * Log the end of a multipart value.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_end_value(const char *type) const noexcept;
+  simdjson_inline void log_end_value(const char *type) const noexcept;
   /**
    * Log an error.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_error(const char *error) const noexcept;
+  simdjson_inline void log_error(const char *error) const noexcept;
 
   template<typename V>
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
   template<typename V>
-  simdjson_warn_unused simdjson_really_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
 };
 
 template<bool STREAMING, typename V>
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::walk_document(V &visitor) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::walk_document(V &visitor) noexcept {
   logger::log_start();
 
   //
@@ -10183,52 +10797,52 @@ simdjson_warn_unused simdjson_really_inline error_code json_iterator::walk_docum
 
 } // walk_document()
 
-simdjson_really_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+simdjson_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
   : buf{_dom_parser.buf},
     next_structural{&_dom_parser.structural_indexes[start_structural_index]},
     dom_parser{_dom_parser} {
 }
 
-simdjson_really_inline const uint8_t *json_iterator::peek() const noexcept {
+simdjson_inline const uint8_t *json_iterator::peek() const noexcept {
   return &buf[*(next_structural)];
 }
-simdjson_really_inline const uint8_t *json_iterator::advance() noexcept {
+simdjson_inline const uint8_t *json_iterator::advance() noexcept {
   return &buf[*(next_structural++)];
 }
-simdjson_really_inline size_t json_iterator::remaining_len() const noexcept {
+simdjson_inline size_t json_iterator::remaining_len() const noexcept {
   return dom_parser.len - *(next_structural-1);
 }
 
-simdjson_really_inline bool json_iterator::at_eof() const noexcept {
+simdjson_inline bool json_iterator::at_eof() const noexcept {
   return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
 }
-simdjson_really_inline bool json_iterator::at_beginning() const noexcept {
+simdjson_inline bool json_iterator::at_beginning() const noexcept {
   return next_structural == dom_parser.structural_indexes.get();
 }
-simdjson_really_inline uint8_t json_iterator::last_structural() const noexcept {
+simdjson_inline uint8_t json_iterator::last_structural() const noexcept {
   return buf[dom_parser.structural_indexes[dom_parser.n_structural_indexes - 1]];
 }
 
-simdjson_really_inline void json_iterator::log_value(const char *type) const noexcept {
+simdjson_inline void json_iterator::log_value(const char *type) const noexcept {
   logger::log_line(*this, "", type, "");
 }
 
-simdjson_really_inline void json_iterator::log_start_value(const char *type) const noexcept {
+simdjson_inline void json_iterator::log_start_value(const char *type) const noexcept {
   logger::log_line(*this, "+", type, "");
   if (logger::LOG_ENABLED) { logger::log_depth++; }
 }
 
-simdjson_really_inline void json_iterator::log_end_value(const char *type) const noexcept {
+simdjson_inline void json_iterator::log_end_value(const char *type) const noexcept {
   if (logger::LOG_ENABLED) { logger::log_depth--; }
   logger::log_line(*this, "-", type, "");
 }
 
-simdjson_really_inline void json_iterator::log_error(const char *error) const noexcept {
+simdjson_inline void json_iterator::log_error(const char *error) const noexcept {
   logger::log_line(*this, "", "ERROR", error);
 }
 
 template<typename V>
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
   switch (*value) {
     case '"': return visitor.visit_root_string(*this, value);
     case 't': return visitor.visit_root_true_atom(*this, value);
@@ -10244,7 +10858,7 @@ simdjson_warn_unused simdjson_really_inline error_code json_iterator::visit_root
   }
 }
 template<typename V>
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
   switch (*value) {
     case '"': return visitor.visit_string(*this, value);
     case 't': return visitor.visit_true_atom(*this, value);
@@ -10276,18 +10890,18 @@ struct tape_writer {
   uint64_t *next_tape_loc;
 
   /** Write a signed 64-bit value to tape. */
-  simdjson_really_inline void append_s64(int64_t value) noexcept;
+  simdjson_inline void append_s64(int64_t value) noexcept;
 
   /** Write an unsigned 64-bit value to tape. */
-  simdjson_really_inline void append_u64(uint64_t value) noexcept;
+  simdjson_inline void append_u64(uint64_t value) noexcept;
 
   /** Write a double value to tape. */
-  simdjson_really_inline void append_double(double value) noexcept;
+  simdjson_inline void append_double(double value) noexcept;
 
   /**
    * Append a tape entry (an 8-bit type,and 56 bits worth of value).
    */
-  simdjson_really_inline void append(uint64_t val, internal::tape_type t) noexcept;
+  simdjson_inline void append(uint64_t val, internal::tape_type t) noexcept;
 
   /**
    * Skip the current tape entry without writing.
@@ -10295,24 +10909,24 @@ struct tape_writer {
    * Used to skip the start of the container, since we'll come back later to fill it in when the
    * container ends.
    */
-  simdjson_really_inline void skip() noexcept;
+  simdjson_inline void skip() noexcept;
 
   /**
    * Skip the number of tape entries necessary to write a large u64 or i64.
    */
-  simdjson_really_inline void skip_large_integer() noexcept;
+  simdjson_inline void skip_large_integer() noexcept;
 
   /**
    * Skip the number of tape entries necessary to write a double.
    */
-  simdjson_really_inline void skip_double() noexcept;
+  simdjson_inline void skip_double() noexcept;
 
   /**
    * Write a value to a known location on tape.
    *
    * Used to go back and write out the start of a container after the container ends.
    */
-  simdjson_really_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
+  simdjson_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
 
 private:
   /**
@@ -10320,50 +10934,50 @@ struct tape_writer {
    * all 64 bits, such as double and uint64_t.
    */
   template<typename T>
-  simdjson_really_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
+  simdjson_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
 }; // struct number_writer
 
-simdjson_really_inline void tape_writer::append_s64(int64_t value) noexcept {
+simdjson_inline void tape_writer::append_s64(int64_t value) noexcept {
   append2(0, value, internal::tape_type::INT64);
 }
 
-simdjson_really_inline void tape_writer::append_u64(uint64_t value) noexcept {
+simdjson_inline void tape_writer::append_u64(uint64_t value) noexcept {
   append(0, internal::tape_type::UINT64);
   *next_tape_loc = value;
   next_tape_loc++;
 }
 
 /** Write a double value to tape. */
-simdjson_really_inline void tape_writer::append_double(double value) noexcept {
+simdjson_inline void tape_writer::append_double(double value) noexcept {
   append2(0, value, internal::tape_type::DOUBLE);
 }
 
-simdjson_really_inline void tape_writer::skip() noexcept {
+simdjson_inline void tape_writer::skip() noexcept {
   next_tape_loc++;
 }
 
-simdjson_really_inline void tape_writer::skip_large_integer() noexcept {
+simdjson_inline void tape_writer::skip_large_integer() noexcept {
   next_tape_loc += 2;
 }
 
-simdjson_really_inline void tape_writer::skip_double() noexcept {
+simdjson_inline void tape_writer::skip_double() noexcept {
   next_tape_loc += 2;
 }
 
-simdjson_really_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
+simdjson_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
   *next_tape_loc = val | ((uint64_t(char(t))) << 56);
   next_tape_loc++;
 }
 
 template<typename T>
-simdjson_really_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
+simdjson_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
   append(val, t);
   static_assert(sizeof(val2) == sizeof(*next_tape_loc), "Type is not 64 bits!");
   memcpy(next_tape_loc, &val2, sizeof(val2));
   next_tape_loc++;
 }
 
-simdjson_really_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
+simdjson_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
   tape_loc = val | ((uint64_t(char(t))) << 56);
 }
 
@@ -10380,40 +10994,40 @@ namespace stage2 {
 
 struct tape_builder {
   template<bool STREAMING>
-  simdjson_warn_unused static simdjson_really_inline error_code parse_document(
+  simdjson_warn_unused static simdjson_inline error_code parse_document(
     dom_parser_implementation &dom_parser,
     dom::document &doc) noexcept;
 
   /** Called when a non-empty document starts. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_document_start(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_document_start(json_iterator &iter) noexcept;
   /** Called when a non-empty document ends without error. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_document_end(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_document_end(json_iterator &iter) noexcept;
 
   /** Called when a non-empty array starts. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_array_start(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_array_start(json_iterator &iter) noexcept;
   /** Called when a non-empty array ends. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_array_end(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_array_end(json_iterator &iter) noexcept;
   /** Called when an empty array is found. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_empty_array(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_empty_array(json_iterator &iter) noexcept;
 
   /** Called when a non-empty object starts. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_object_start(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_object_start(json_iterator &iter) noexcept;
   /**
    * Called when a key in a field is encountered.
    *
    * primitive, visit_object_start, visit_empty_object, visit_array_start, or visit_empty_array
    * will be called after this with the field value.
    */
-  simdjson_warn_unused simdjson_really_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
   /** Called when a non-empty object ends. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_object_end(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_object_end(json_iterator &iter) noexcept;
   /** Called when an empty object is found. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_empty_object(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_empty_object(json_iterator &iter) noexcept;
 
   /**
    * Called when a string, number, boolean or null is found.
    */
-  simdjson_warn_unused simdjson_really_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
   /**
    * Called when a string, number, boolean or null is found at the top level of a document (i.e.
    * when there is no array or object and the entire document is a single string, number, boolean or
@@ -10422,22 +11036,22 @@ struct tape_builder {
    * This is separate from primitive() because simdjson's normal primitive parsing routines assume
    * there is at least one more token after the value, which is only true in an array or object.
    */
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
 
-  simdjson_warn_unused simdjson_really_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
 
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
 
   /** Called each time a new field or element in an array or object is found. */
-  simdjson_warn_unused simdjson_really_inline error_code increment_count(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code increment_count(json_iterator &iter) noexcept;
 
   /** Next location to write to tape */
   tape_writer tape;
@@ -10445,18 +11059,18 @@ struct tape_builder {
   /** Next write location in the string buf for stage 2 parsing */
   uint8_t *current_string_buf_loc;
 
-  simdjson_really_inline tape_builder(dom::document &doc) noexcept;
+  simdjson_inline tape_builder(dom::document &doc) noexcept;
 
-  simdjson_really_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
-  simdjson_really_inline void start_container(json_iterator &iter) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
-  simdjson_really_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
-  simdjson_really_inline void on_end_string(uint8_t *dst) noexcept;
+  simdjson_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
+  simdjson_inline void start_container(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_warn_unused simdjson_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
+  simdjson_inline void on_end_string(uint8_t *dst) noexcept;
 }; // class tape_builder
 
 template<bool STREAMING>
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::parse_document(
+simdjson_warn_unused simdjson_inline error_code tape_builder::parse_document(
     dom_parser_implementation &dom_parser,
     dom::document &doc) noexcept {
   dom_parser.doc = &doc;
@@ -10465,56 +11079,56 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::parse_docum
   return iter.walk_document<STREAMING>(builder);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
   return iter.visit_root_primitive(*this, value);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
   return iter.visit_primitive(*this, value);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
   return empty_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
   return empty_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
   start_container(iter);
   return SUCCESS;
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
   start_container(iter);
   return SUCCESS;
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
   start_container(iter);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
   return end_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
   return end_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
   constexpr uint32_t start_tape_index = 0;
   tape.append(start_tape_index, internal::tape_type::ROOT);
   tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter), internal::tape_type::ROOT);
   return SUCCESS;
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
   return visit_string(iter, key, true);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
   iter.dom_parser.open_containers[iter.depth].count++; // we have a key value pair in the object at parser.dom_parser.depth - 1
   return SUCCESS;
 }
 
-simdjson_really_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
+simdjson_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
   iter.log_value(key ? "key" : "string");
   uint8_t *dst = on_start_string(iter);
   dst = stringparsing::parse_string(value+1, dst);
@@ -10526,16 +11140,16 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_strin
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
   return visit_string(iter, value);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("number");
   return numberparsing::parse_number(value, tape);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
   //
   // We need to make a copy to make sure that the string is space terminated.
   // This is not about padding the input, which should already padded up
@@ -10557,42 +11171,42 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_
   return error;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("true");
   if (!atomparsing::is_valid_true_atom(value)) { return T_ATOM_ERROR; }
   tape.append(0, internal::tape_type::TRUE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("true");
   if (!atomparsing::is_valid_true_atom(value, iter.remaining_len())) { return T_ATOM_ERROR; }
   tape.append(0, internal::tape_type::TRUE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("false");
   if (!atomparsing::is_valid_false_atom(value)) { return F_ATOM_ERROR; }
   tape.append(0, internal::tape_type::FALSE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("false");
   if (!atomparsing::is_valid_false_atom(value, iter.remaining_len())) { return F_ATOM_ERROR; }
   tape.append(0, internal::tape_type::FALSE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("null");
   if (!atomparsing::is_valid_null_atom(value)) { return N_ATOM_ERROR; }
   tape.append(0, internal::tape_type::NULL_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("null");
   if (!atomparsing::is_valid_null_atom(value, iter.remaining_len())) { return N_ATOM_ERROR; }
   tape.append(0, internal::tape_type::NULL_VALUE);
@@ -10601,24 +11215,24 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_
 
 // private:
 
-simdjson_really_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
+simdjson_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
   return uint32_t(tape.next_tape_loc - iter.dom_parser.doc->tape.get());
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
   auto start_index = next_tape_index(iter);
   tape.append(start_index+2, start);
   tape.append(start_index, end);
   return SUCCESS;
 }
 
-simdjson_really_inline void tape_builder::start_container(json_iterator &iter) noexcept {
+simdjson_inline void tape_builder::start_container(json_iterator &iter) noexcept {
   iter.dom_parser.open_containers[iter.depth].tape_index = next_tape_index(iter);
   iter.dom_parser.open_containers[iter.depth].count = 0;
   tape.skip(); // We don't actually *write* the start element until the end.
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
   // Write the ending tape element, pointing at the start location
   const uint32_t start_tape_index = iter.dom_parser.open_containers[iter.depth].tape_index;
   tape.append(start_tape_index, end);
@@ -10631,13 +11245,13 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::end_contain
   return SUCCESS;
 }
 
-simdjson_really_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
+simdjson_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
   // we advance the point, accounting for the fact that we have a NULL termination
   tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::STRING);
   return current_string_buf_loc + sizeof(uint32_t);
 }
 
-simdjson_really_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
+simdjson_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
   uint32_t str_length = uint32_t(dst - (current_string_buf_loc + sizeof(uint32_t)));
   // TODO check for overflow in case someone has a crazy string (>=4GB?)
   // But only add the overflow check when the document itself exceeds 4GB
@@ -10663,7 +11277,7 @@ namespace haswell {
 namespace {
 namespace stage1 {
 
-simdjson_really_inline uint64_t json_string_scanner::find_escaped(uint64_t backslash) {
+simdjson_inline uint64_t json_string_scanner::find_escaped(uint64_t backslash) {
   if (!backslash) { uint64_t escaped = prev_escaped; prev_escaped = 0; return escaped; }
   return find_escaped_branchless(backslash);
 }
@@ -10693,6 +11307,10 @@ simdjson_warn_unused error_code dom_parser_implementation::stage2_next(dom::docu
   return stage2::tape_builder::parse_document<true>(*this, _doc);
 }
 
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_string(const uint8_t *src, uint8_t *dst) const noexcept {
+  return haswell::stringparsing::parse_string(src, dst);
+}
+
 simdjson_warn_unused error_code dom_parser_implementation::parse(const uint8_t *_buf, size_t _len, dom::document &_doc) noexcept {
   auto error = stage1(_buf, _len, stage1_mode::regular);
   if (error) { return error; }
@@ -10753,17 +11371,17 @@ namespace {
 using namespace simd;
 
 struct json_character_block {
-  static simdjson_really_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
+  static simdjson_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
 
-  simdjson_really_inline uint64_t whitespace() const noexcept { return _whitespace; }
-  simdjson_really_inline uint64_t op() const noexcept { return _op; }
-  simdjson_really_inline uint64_t scalar() const noexcept { return ~(op() | whitespace()); }
+  simdjson_inline uint64_t whitespace() const noexcept { return _whitespace; }
+  simdjson_inline uint64_t op() const noexcept { return _op; }
+  simdjson_inline uint64_t scalar() const noexcept { return ~(op() | whitespace()); }
 
   uint64_t _whitespace;
   uint64_t _op;
 };
 
-simdjson_really_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
+simdjson_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
   const simd8<uint8_t> table1(16, 0, 0, 0, 0, 0, 0, 0, 0, 8, 12, 1, 2, 9, 0, 0);
   const simd8<uint8_t> table2(8, 0, 18, 4, 0, 1, 0, 1, 0, 0, 0, 3, 2, 1, 0, 0);
 
@@ -10791,22 +11409,22 @@ simdjson_really_inline json_character_block json_character_block::classify(const
   return { whitespace, op };
 }
 
-simdjson_really_inline bool is_ascii(const simd8x64<uint8_t>& input) {
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input) {
   // careful: 0x80 is not ascii.
-  return input.reduce_or().saturating_sub(0b01111111u).bits_not_set_anywhere();
+  return input.reduce_or().saturating_sub(0x7fu).bits_not_set_anywhere();
 }
 
-simdjson_unused simdjson_really_inline simd8<bool> must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
-  simd8<uint8_t> is_second_byte = prev1.saturating_sub(0b11000000u-1); // Only 11______ will be > 0
-  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0b11100000u-1); // Only 111_____ will be > 0
-  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0b11110000u-1); // Only 1111____ will be > 0
+simdjson_unused simdjson_inline simd8<bool> must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_second_byte = prev1.saturating_sub(0xc0u-1); // Only 11______ will be > 0
+  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-1); // Only 111_____ will be > 0
+  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-1); // Only 1111____ will be > 0
   // Caller requires a bool (all 1's). All values resulting from the subtraction will be <= 64, so signed comparison is fine.
   return simd8<int8_t>(is_second_byte | is_third_byte | is_fourth_byte) > int8_t(0);
 }
 
-simdjson_really_inline simd8<bool> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
-  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0b11100000u-1); // Only 111_____ will be > 0
-  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0b11110000u-1); // Only 1111____ will be > 0
+simdjson_inline simd8<bool> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-1); // Only 111_____ will be > 0
+  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-1); // Only 1111____ will be > 0
   // Caller requires a bool (all 1's). All values resulting from the subtraction will be <= 64, so signed comparison is fine.
   return simd8<int8_t>(is_third_byte | is_fourth_byte) > int8_t(0);
 }
@@ -10823,7 +11441,7 @@ namespace utf8_validation {
 
 using namespace simd;
 
-  simdjson_really_inline simd8<uint8_t> check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+  simdjson_inline simd8<uint8_t> check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
 // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
 // Bit 1 = Too Long (ASCII followed by continuation)
 // Bit 2 = Overlong 3-byte
@@ -10913,7 +11531,7 @@ using namespace simd;
     );
     return (byte_1_high & byte_1_low & byte_2_high);
   }
-  simdjson_really_inline simd8<uint8_t> check_multibyte_lengths(const simd8<uint8_t> input,
+  simdjson_inline simd8<uint8_t> check_multibyte_lengths(const simd8<uint8_t> input,
       const simd8<uint8_t> prev_input, const simd8<uint8_t> sc) {
     simd8<uint8_t> prev2 = input.prev<2>(prev_input);
     simd8<uint8_t> prev3 = input.prev<3>(prev_input);
@@ -10926,7 +11544,7 @@ using namespace simd;
   // Return nonzero if there are incomplete multibyte characters at the end of the block:
   // e.g. if there is a 4-byte character, but it's 3 bytes from the end.
   //
-  simdjson_really_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
+  simdjson_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
     // If the previous input's last 3 bytes match this, they're too short (they ended at EOF):
     // ... 1111____ 111_____ 11______
 #if SIMDJSON_IMPLEMENTATION_ICELAKE
@@ -10938,14 +11556,14 @@ using namespace simd;
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
-      255, 255, 255, 255, 255, 0b11110000u-1, 0b11100000u-1, 0b11000000u-1
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
     };
 #else
     static const uint8_t max_array[32] = {
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
-      255, 255, 255, 255, 255, 0b11110000u-1, 0b11100000u-1, 0b11000000u-1
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
     };
 #endif
     const simd8<uint8_t> max_value(&max_array[sizeof(max_array)-sizeof(simd8<uint8_t>)]);
@@ -10963,7 +11581,7 @@ using namespace simd;
     //
     // Check whether the current bytes are valid UTF-8.
     //
-    simdjson_really_inline void check_utf8_bytes(const simd8<uint8_t> input, const simd8<uint8_t> prev_input) {
+    simdjson_inline void check_utf8_bytes(const simd8<uint8_t> input, const simd8<uint8_t> prev_input) {
       // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ lead bytes
       // (2, 3, 4-byte leads become large positive numbers instead of small negative numbers)
       simd8<uint8_t> prev1 = input.prev<1>(prev_input);
@@ -10974,13 +11592,21 @@ using namespace simd;
     // The only problem that can happen at EOF is that a multibyte character is too short
     // or a byte value too large in the last bytes: check_special_cases only checks for bytes
     // too large in the first of two bytes.
-    simdjson_really_inline void check_eof() {
+    simdjson_inline void check_eof() {
       // If the previous block had incomplete UTF-8 characters at the end, an ASCII block can't
       // possibly finish them.
       this->error |= this->prev_incomplete;
     }
 
-    simdjson_really_inline void check_next_input(const simd8x64<uint8_t>& input) {
+#ifndef SIMDJSON_IF_CONSTEXPR
+#if SIMDJSON_CPLUSPLUS17
+#define SIMDJSON_IF_CONSTEXPR if constexpr
+#else
+#define SIMDJSON_IF_CONSTEXPR if
+#endif
+#endif
+
+    simdjson_inline void check_next_input(const simd8x64<uint8_t>& input) {
       if(simdjson_likely(is_ascii(input))) {
         this->error |= this->prev_incomplete;
       } else {
@@ -10989,12 +11615,12 @@ using namespace simd;
                 ||(simd8x64<uint8_t>::NUM_CHUNKS == 2)
                 || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
                 "We support one, two or four chunks per 64-byte block.");
-        if(simd8x64<uint8_t>::NUM_CHUNKS == 1) {
+        SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 1) {
           this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
-        } if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
           this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
           this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
           this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
           this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
           this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
@@ -11005,7 +11631,7 @@ using namespace simd;
       }
     }
     // do not forget to call check_eof!
-    simdjson_really_inline error_code errors() {
+    simdjson_inline error_code errors() {
       return this->error.any_bits_set_anywhere() ? error_code::UTF8_ERROR : error_code::SUCCESS;
     }
 
@@ -11033,10 +11659,10 @@ namespace {
 template<size_t STEP_SIZE>
 struct buf_block_reader {
 public:
-  simdjson_really_inline buf_block_reader(const uint8_t *_buf, size_t _len);
-  simdjson_really_inline size_t block_index();
-  simdjson_really_inline bool has_full_block() const;
-  simdjson_really_inline const uint8_t *full_block() const;
+  simdjson_inline buf_block_reader(const uint8_t *_buf, size_t _len);
+  simdjson_inline size_t block_index();
+  simdjson_inline bool has_full_block() const;
+  simdjson_inline const uint8_t *full_block() const;
   /**
    * Get the last block, padded with spaces.
    *
@@ -11046,8 +11672,8 @@ struct buf_block_reader {
    *
    * @return the number of effective characters in the last block.
    */
-  simdjson_really_inline size_t get_remainder(uint8_t *dst) const;
-  simdjson_really_inline void advance();
+  simdjson_inline size_t get_remainder(uint8_t *dst) const;
+  simdjson_inline void advance();
 private:
   const uint8_t *buf;
   const size_t len;
@@ -11086,23 +11712,23 @@ simdjson_unused static char * format_mask(uint64_t mask) {
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len) : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, idx{0} {}
+simdjson_inline buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len) : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, idx{0} {}
 
 template<size_t STEP_SIZE>
-simdjson_really_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
+simdjson_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
   return idx < lenminusstep;
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
+simdjson_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
   return &buf[idx];
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
   if(len == idx) { return 0; } // memcpy(dst, null, 0) will trigger an error with some sanitizers
   std::memset(dst, 0x20, STEP_SIZE); // std::memset STEP_SIZE because it's more efficient to write out 8 or 16 bytes at once.
   std::memcpy(dst, buf + idx, len - idx);
@@ -11110,7 +11736,7 @@ simdjson_really_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline void buf_block_reader<STEP_SIZE>::advance() {
+simdjson_inline void buf_block_reader<STEP_SIZE>::advance() {
   idx += STEP_SIZE;
 }
 
@@ -11126,27 +11752,27 @@ namespace stage1 {
 
 struct json_string_block {
   // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
-  simdjson_really_inline json_string_block(uint64_t backslash, uint64_t escaped, uint64_t quote, uint64_t in_string) :
+  simdjson_inline json_string_block(uint64_t backslash, uint64_t escaped, uint64_t quote, uint64_t in_string) :
   _backslash(backslash), _escaped(escaped), _quote(quote), _in_string(in_string) {}
 
   // Escaped characters (characters following an escape() character)
-  simdjson_really_inline uint64_t escaped() const { return _escaped; }
+  simdjson_inline uint64_t escaped() const { return _escaped; }
   // Escape characters (backslashes that are not escaped--i.e. in \\, includes only the first \)
-  simdjson_really_inline uint64_t escape() const { return _backslash & ~_escaped; }
+  simdjson_inline uint64_t escape() const { return _backslash & ~_escaped; }
   // Real (non-backslashed) quotes
-  simdjson_really_inline uint64_t quote() const { return _quote; }
+  simdjson_inline uint64_t quote() const { return _quote; }
   // Start quotes of strings
-  simdjson_really_inline uint64_t string_start() const { return _quote & _in_string; }
+  simdjson_inline uint64_t string_start() const { return _quote & _in_string; }
   // End quotes of strings
-  simdjson_really_inline uint64_t string_end() const { return _quote & ~_in_string; }
+  simdjson_inline uint64_t string_end() const { return _quote & ~_in_string; }
   // Only characters inside the string (not including the quotes)
-  simdjson_really_inline uint64_t string_content() const { return _in_string & ~_quote; }
+  simdjson_inline uint64_t string_content() const { return _in_string & ~_quote; }
   // Return a mask of whether the given characters are inside a string (only works on non-quotes)
-  simdjson_really_inline uint64_t non_quote_inside_string(uint64_t mask) const { return mask & _in_string; }
+  simdjson_inline uint64_t non_quote_inside_string(uint64_t mask) const { return mask & _in_string; }
   // Return a mask of whether the given characters are inside a string (only works on non-quotes)
-  simdjson_really_inline uint64_t non_quote_outside_string(uint64_t mask) const { return mask & ~_in_string; }
+  simdjson_inline uint64_t non_quote_outside_string(uint64_t mask) const { return mask & ~_in_string; }
   // Tail of string (everything except the start quote)
-  simdjson_really_inline uint64_t string_tail() const { return _in_string ^ _quote; }
+  simdjson_inline uint64_t string_tail() const { return _in_string ^ _quote; }
 
   // backslash characters
   uint64_t _backslash;
@@ -11161,14 +11787,14 @@ struct json_string_block {
 // Scans blocks for string characters, storing the state necessary to do so
 class json_string_scanner {
 public:
-  simdjson_really_inline json_string_block next(const simd::simd8x64<uint8_t>& in);
+  simdjson_inline json_string_block next(const simd::simd8x64<uint8_t>& in);
   // Returns either UNCLOSED_STRING or SUCCESS
-  simdjson_really_inline error_code finish();
+  simdjson_inline error_code finish();
 
 private:
   // Intended to be defined by the implementation
-  simdjson_really_inline uint64_t find_escaped(uint64_t escape);
-  simdjson_really_inline uint64_t find_escaped_branchless(uint64_t escape);
+  simdjson_inline uint64_t find_escaped(uint64_t escape);
+  simdjson_inline uint64_t find_escaped_branchless(uint64_t escape);
 
   // Whether the last iteration was still inside a string (all 1's = true, all 0's = false).
   uint64_t prev_in_string = 0ULL;
@@ -11203,7 +11829,7 @@ class json_string_scanner {
 // desired        |   x  | x x  x x  x x  x  x  |
 // text           |  \\\ | \\\"\\\" \\\" \\"\\" |
 //
-simdjson_really_inline uint64_t json_string_scanner::find_escaped_branchless(uint64_t backslash) {
+simdjson_inline uint64_t json_string_scanner::find_escaped_branchless(uint64_t backslash) {
   // If there was overflow, pretend the first character isn't a backslash
   backslash &= ~prev_escaped;
   uint64_t follows_escape = backslash << 1 | prev_escaped;
@@ -11228,7 +11854,7 @@ simdjson_really_inline uint64_t json_string_scanner::find_escaped_branchless(uin
 //
 // Backslash sequences outside of quotes will be detected in stage 2.
 //
-simdjson_really_inline json_string_block json_string_scanner::next(const simd::simd8x64<uint8_t>& in) {
+simdjson_inline json_string_block json_string_scanner::next(const simd::simd8x64<uint8_t>& in) {
   const uint64_t backslash = in.eq('\\');
   const uint64_t escaped = find_escaped(backslash);
   const uint64_t quote = in.eq('"') & ~escaped;
@@ -11261,7 +11887,7 @@ simdjson_really_inline json_string_block json_string_scanner::next(const simd::s
   );
 }
 
-simdjson_really_inline error_code json_string_scanner::finish() {
+simdjson_inline error_code json_string_scanner::finish() {
   if (prev_in_string) {
     return UNCLOSED_STRING;
   }
@@ -11299,25 +11925,25 @@ namespace stage1 {
 struct json_block {
 public:
   // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
-  simdjson_really_inline json_block(json_string_block&& string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  simdjson_inline json_block(json_string_block&& string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
   _string(std::move(string)), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
-  simdjson_really_inline json_block(json_string_block string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  simdjson_inline json_block(json_string_block string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
   _string(string), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
 
   /**
    * The start of structurals.
    * In simdjson prior to v0.3, these were called the pseudo-structural characters.
    **/
-  simdjson_really_inline uint64_t structural_start() const noexcept { return potential_structural_start() & ~_string.string_tail(); }
+  simdjson_inline uint64_t structural_start() const noexcept { return potential_structural_start() & ~_string.string_tail(); }
   /** All JSON whitespace (i.e. not in a string) */
-  simdjson_really_inline uint64_t whitespace() const noexcept { return non_quote_outside_string(_characters.whitespace()); }
+  simdjson_inline uint64_t whitespace() const noexcept { return non_quote_outside_string(_characters.whitespace()); }
 
   // Helpers
 
   /** Whether the given characters are inside a string (only works on non-quotes) */
-  simdjson_really_inline uint64_t non_quote_inside_string(uint64_t mask) const noexcept { return _string.non_quote_inside_string(mask); }
+  simdjson_inline uint64_t non_quote_inside_string(uint64_t mask) const noexcept { return _string.non_quote_inside_string(mask); }
   /** Whether the given characters are outside a string (only works on non-quotes) */
-  simdjson_really_inline uint64_t non_quote_outside_string(uint64_t mask) const noexcept { return _string.non_quote_outside_string(mask); }
+  simdjson_inline uint64_t non_quote_outside_string(uint64_t mask) const noexcept { return _string.non_quote_outside_string(mask); }
 
   // string and escape characters
   json_string_block _string;
@@ -11332,12 +11958,12 @@ struct json_block {
    * structural elements ([,],{,},:, comma) plus scalar starts like 123, true and "abc".
    * They may reside inside a string.
    **/
-  simdjson_really_inline uint64_t potential_structural_start() const noexcept { return _characters.op() | potential_scalar_start(); }
+  simdjson_inline uint64_t potential_structural_start() const noexcept { return _characters.op() | potential_scalar_start(); }
   /**
    * The start of non-operator runs, like 123, true and "abc".
    * It main reside inside a string.
    **/
-  simdjson_really_inline uint64_t potential_scalar_start() const noexcept {
+  simdjson_inline uint64_t potential_scalar_start() const noexcept {
     // The term "scalar" refers to anything except structural characters and white space
     // (so letters, numbers, quotes).
     // Whenever it is preceded by something that is not a structural element ({,},[,],:, ") nor a white-space
@@ -11348,7 +11974,7 @@ struct json_block {
    * Whether the given character is immediately after a non-operator like 123, true.
    * The characters following a quote are not included.
    */
-  simdjson_really_inline uint64_t follows_potential_scalar() const noexcept {
+  simdjson_inline uint64_t follows_potential_scalar() const noexcept {
     // _follows_potential_nonquote_scalar: is defined as marking any character that follows a character
     // that is not a structural element ({,},[,],:, comma) nor a quote (") and that is not a
     // white space.
@@ -11372,10 +11998,10 @@ struct json_block {
  */
 class json_scanner {
 public:
-  json_scanner() {}
-  simdjson_really_inline json_block next(const simd::simd8x64<uint8_t>& in);
+  json_scanner() = default;
+  simdjson_inline json_block next(const simd::simd8x64<uint8_t>& in);
   // Returns either UNCLOSED_STRING or SUCCESS
-  simdjson_really_inline error_code finish();
+  simdjson_inline error_code finish();
 
 private:
   // Whether the last character of the previous iteration is part of a scalar token
@@ -11392,13 +12018,13 @@ class json_scanner {
 //
 //     const uint64_t backslashed_quote = in.eq('"') & immediately_follows(in.eq('\'), prev_backslash);
 //
-simdjson_really_inline uint64_t follows(const uint64_t match, uint64_t &overflow) {
+simdjson_inline uint64_t follows(const uint64_t match, uint64_t &overflow) {
   const uint64_t result = match << 1 | overflow;
   overflow = match >> 63;
   return result;
 }
 
-simdjson_really_inline json_block json_scanner::next(const simd::simd8x64<uint8_t>& in) {
+simdjson_inline json_block json_scanner::next(const simd::simd8x64<uint8_t>& in) {
   json_string_block strings = string_scanner.next(in);
   // identifies the white-space and the structural characters
   json_character_block characters = json_character_block::classify(in);
@@ -11423,7 +12049,7 @@ simdjson_really_inline json_block json_scanner::next(const simd::simd8x64<uint8_
   );
 }
 
-simdjson_really_inline error_code json_scanner::finish() {
+simdjson_inline error_code json_scanner::finish() {
   return string_scanner.finish();
 }
 
@@ -11449,23 +12075,23 @@ class json_minifier {
   static error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept;
 
 private:
-  simdjson_really_inline json_minifier(uint8_t *_dst)
+  simdjson_inline json_minifier(uint8_t *_dst)
   : dst{_dst}
   {}
   template<size_t STEP_SIZE>
-  simdjson_really_inline void step(const uint8_t *block_buf, buf_block_reader<STEP_SIZE> &reader) noexcept;
-  simdjson_really_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block);
-  simdjson_really_inline error_code finish(uint8_t *dst_start, size_t &dst_len);
+  simdjson_inline void step(const uint8_t *block_buf, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block);
+  simdjson_inline error_code finish(uint8_t *dst_start, size_t &dst_len);
   json_scanner scanner{};
   uint8_t *dst;
 };
 
-simdjson_really_inline void json_minifier::next(const simd::simd8x64<uint8_t>& in, const json_block& block) {
+simdjson_inline void json_minifier::next(const simd::simd8x64<uint8_t>& in, const json_block& block) {
   uint64_t mask = block.whitespace();
   dst += in.compress(mask, dst);
 }
 
-simdjson_really_inline error_code json_minifier::finish(uint8_t *dst_start, size_t &dst_len) {
+simdjson_inline error_code json_minifier::finish(uint8_t *dst_start, size_t &dst_len) {
   error_code error = scanner.finish();
   if (error) { dst_len = 0; return error; }
   dst_len = dst - dst_start;
@@ -11473,7 +12099,7 @@ simdjson_really_inline error_code json_minifier::finish(uint8_t *dst_start, size
 }
 
 template<>
-simdjson_really_inline void json_minifier::step<128>(const uint8_t *block_buf, buf_block_reader<128> &reader) noexcept {
+simdjson_inline void json_minifier::step<128>(const uint8_t *block_buf, buf_block_reader<128> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block_buf);
   simd::simd8x64<uint8_t> in_2(block_buf+64);
   json_block block_1 = scanner.next(in_1);
@@ -11484,7 +12110,7 @@ simdjson_really_inline void json_minifier::step<128>(const uint8_t *block_buf, b
 }
 
 template<>
-simdjson_really_inline void json_minifier::step<64>(const uint8_t *block_buf, buf_block_reader<64> &reader) noexcept {
+simdjson_inline void json_minifier::step<64>(const uint8_t *block_buf, buf_block_reader<64> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block_buf);
   json_block block_1 = scanner.next(in_1);
   this->next(block_buf, block_1);
@@ -11557,7 +12183,7 @@ namespace {
   * complete document, therefore the last json buffer location is the end of the
   * batch.
   */
-simdjson_really_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
+simdjson_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
   // Variant: do not count separately, just figure out depth
   if(parser.n_structural_indexes == 0) { return 0; }
   auto arr_cnt = 0;
@@ -11632,7 +12258,7 @@ class bit_indexer {
 public:
   uint32_t *tail;
 
-  simdjson_really_inline bit_indexer(uint32_t *index_buf) : tail(index_buf) {}
+  simdjson_inline bit_indexer(uint32_t *index_buf) : tail(index_buf) {}
 
   // flatten out values in 'bits' assuming that they are are to have values of idx
   // plus their position in the bitvector, and store these indexes at
@@ -11643,9 +12269,9 @@ class bit_indexer {
   // If the kernel sets SIMDJSON_CUSTOM_BIT_INDEXER, then it will provide its own
   // version of the code.
 #ifdef SIMDJSON_CUSTOM_BIT_INDEXER
-  simdjson_really_inline void write(uint32_t idx, uint64_t bits);
+  simdjson_inline void write(uint32_t idx, uint64_t bits);
 #else
-  simdjson_really_inline void write(uint32_t idx, uint64_t bits) {
+  simdjson_inline void write(uint32_t idx, uint64_t bits) {
     // In some instances, the next branch is expensive because it is mispredicted.
     // Unfortunately, in other cases,
     // it helps tremendously.
@@ -11754,11 +12380,11 @@ class json_structural_indexer {
   static error_code index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept;
 
 private:
-  simdjson_really_inline json_structural_indexer(uint32_t *structural_indexes);
+  simdjson_inline json_structural_indexer(uint32_t *structural_indexes);
   template<size_t STEP_SIZE>
-  simdjson_really_inline void step(const uint8_t *block, buf_block_reader<STEP_SIZE> &reader) noexcept;
-  simdjson_really_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx);
-  simdjson_really_inline error_code finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial);
+  simdjson_inline void step(const uint8_t *block, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx);
+  simdjson_inline error_code finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial);
 
   json_scanner scanner{};
   utf8_checker checker{};
@@ -11767,26 +12393,26 @@ class json_structural_indexer {
   uint64_t unescaped_chars_error = 0;
 };
 
-simdjson_really_inline json_structural_indexer::json_structural_indexer(uint32_t *structural_indexes) : indexer{structural_indexes} {}
+simdjson_inline json_structural_indexer::json_structural_indexer(uint32_t *structural_indexes) : indexer{structural_indexes} {}
 
 // Skip the last character if it is partial
-simdjson_really_inline size_t trim_partial_utf8(const uint8_t *buf, size_t len) {
+simdjson_inline size_t trim_partial_utf8(const uint8_t *buf, size_t len) {
   if (simdjson_unlikely(len < 3)) {
     switch (len) {
       case 2:
-        if (buf[len-1] >= 0b11000000) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
-        if (buf[len-2] >= 0b11100000) { return len-2; } // 3- and 4-byte characters with only 2 bytes left
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 2 bytes left
         return len;
       case 1:
-        if (buf[len-1] >= 0b11000000) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
         return len;
       case 0:
         return len;
     }
   }
-  if (buf[len-1] >= 0b11000000) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
-  if (buf[len-2] >= 0b11100000) { return len-2; } // 3- and 4-byte characters with only 1 byte left
-  if (buf[len-3] >= 0b11110000) { return len-3; } // 4-byte characters with only 3 bytes left
+  if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+  if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 1 byte left
+  if (buf[len-3] >= 0xf0) { return len-3; } // 4-byte characters with only 3 bytes left
   return len;
 }
 
@@ -11835,7 +12461,7 @@ error_code json_structural_indexer::index(const uint8_t *buf, size_t len, dom_pa
 }
 
 template<>
-simdjson_really_inline void json_structural_indexer::step<128>(const uint8_t *block, buf_block_reader<128> &reader) noexcept {
+simdjson_inline void json_structural_indexer::step<128>(const uint8_t *block, buf_block_reader<128> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block);
   simd::simd8x64<uint8_t> in_2(block+64);
   json_block block_1 = scanner.next(in_1);
@@ -11846,14 +12472,14 @@ simdjson_really_inline void json_structural_indexer::step<128>(const uint8_t *bl
 }
 
 template<>
-simdjson_really_inline void json_structural_indexer::step<64>(const uint8_t *block, buf_block_reader<64> &reader) noexcept {
+simdjson_inline void json_structural_indexer::step<64>(const uint8_t *block, buf_block_reader<64> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block);
   json_block block_1 = scanner.next(in_1);
   this->next(in_1, block_1, reader.block_index());
   reader.advance();
 }
 
-simdjson_really_inline void json_structural_indexer::next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx) {
+simdjson_inline void json_structural_indexer::next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx) {
   uint64_t unescaped = in.lteq(0x1F);
   checker.check_next_input(in);
   indexer.write(uint32_t(idx-64), prev_structurals); // Output *last* iteration's structurals to the parser
@@ -11861,7 +12487,7 @@ simdjson_really_inline void json_structural_indexer::next(const simd::simd8x64<u
   unescaped_chars_error |= block.non_quote_inside_string(unescaped);
 }
 
-simdjson_really_inline error_code json_structural_indexer::finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial) {
+simdjson_inline error_code json_structural_indexer::finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial) {
   // Write out the final iteration's structurals
   indexer.write(uint32_t(idx-64), prev_structurals);
   error_code error = scanner.finish();
@@ -12007,7 +12633,151 @@ bool generic_validate_utf8(const char * input, size_t length) {
 //
 // Stage 2
 //
+/* begin file src/generic/stage2/stringparsing.h */
+// This file contains the common code every implementation uses
+// It is intended to be included multiple times and compiled multiple times
 
+namespace simdjson {
+namespace ppc64 {
+namespace {
+/// @private
+namespace stringparsing {
+
+// begin copypasta
+// These chars yield themselves: " \ /
+// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
+// u not handled in this table as it's complex
+static const uint8_t escape_map[256] = {
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
+    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
+    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+};
+
+// handle a unicode codepoint
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr) {
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion isn't valid; we defer the check for this to inside the
+  // multilingual plane check
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) != ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      return false;
+    }
+    uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+
+    // We have already checked that the high surrogate is valid and
+    // (code_point - 0xd800) < 1024.
+    //
+    // Check that code_point_2 is in the range 0xdc00..0xdfff
+    // and that code_point_2 was parsed from valid hex.
+    uint32_t low_bit = code_point_2 - 0xdc00;
+    if (low_bit >> 10) {
+      return false;
+    }
+
+    code_point =
+        (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+    *src_ptr += 6;
+  } else if (code_point >= 0xdc00 && code_point <= 0xdfff) {
+      // If we encounter a low surrogate (not preceded by a high surrogate)
+      // then we have an error.
+      return false;
+  }
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+/**
+ * Unescape a valid UTF-8 string from src to dst, stopping at a final unescaped quote. There
+ * must be an unescaped quote terminating the string. It returns the final output
+ * position as pointer. In case of error (e.g., the string has bad escaped codes),
+ * then null_nullptrptr is returned. It is assumed that the output buffer is large
+ * enough. E.g., if src points at 'joe"', then dst needs to have four free bytes +
+ * SIMDJSON_PADDING bytes.
+ */
+simdjson_warn_unused simdjson_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst) {
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto bs_quote = backslash_and_quote::copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint(&src, &dst)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+  /* can't be reached */
+  return nullptr;
+}
+
+} // namespace stringparsing
+} // unnamed namespace
+} // namespace ppc64
+} // namespace simdjson
+/* end file src/generic/stage2/stringparsing.h */
 /* begin file src/generic/stage2/tape_builder.h */
 /* begin file src/generic/stage2/json_iterator.h */
 /* begin file src/generic/stage2/logger.h */
@@ -12033,7 +12803,7 @@ namespace logger {
   static int log_depth; // Not threadsafe. Log only.
 
   // Helper to turn unprintable or newline characters into spaces
-  static simdjson_really_inline char printable_char(char c) {
+  static simdjson_inline char printable_char(char c) {
     if (c >= 0x20) {
       return c;
     } else {
@@ -12042,7 +12812,7 @@ namespace logger {
   }
 
   // Print the header and set up log_start
-  static simdjson_really_inline void log_start() {
+  static simdjson_inline void log_start() {
     if (LOG_ENABLED) {
       log_depth = 0;
       printf("\n");
@@ -12051,7 +12821,7 @@ namespace logger {
     }
   }
 
-  simdjson_unused static simdjson_really_inline void log_string(const char *message) {
+  simdjson_unused static simdjson_inline void log_string(const char *message) {
     if (LOG_ENABLED) {
       printf("%s\n", message);
     }
@@ -12059,7 +12829,7 @@ namespace logger {
 
   // Logs a single line from the stage 2 DOM parser
   template<typename S>
-  static simdjson_really_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
+  static simdjson_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
     if (LOG_ENABLED) {
       printf("| %*s%s%-*s ", log_depth*2, "", title_prefix, LOG_EVENT_LEN - log_depth*2 - int(strlen(title_prefix)), title);
       auto current_index = structurals.at_beginning() ? nullptr : structurals.next_structural-1;
@@ -12137,14 +12907,14 @@ class json_iterator {
    * - increment_count(iter) - each time a value is found in an array or object.
    */
   template<bool STREAMING, typename V>
-  simdjson_warn_unused simdjson_really_inline error_code walk_document(V &visitor) noexcept;
+  simdjson_warn_unused simdjson_inline error_code walk_document(V &visitor) noexcept;
 
   /**
    * Create an iterator capable of walking a JSON document.
    *
    * The document must have already passed through stage 1.
    */
-  simdjson_really_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
+  simdjson_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
 
   /**
    * Look at the next token.
@@ -12153,7 +12923,7 @@ class json_iterator {
    *
    * They may include invalid JSON as well (such as `1.2.3` or `ture`).
    */
-  simdjson_really_inline const uint8_t *peek() const noexcept;
+  simdjson_inline const uint8_t *peek() const noexcept;
   /**
    * Advance to the next token.
    *
@@ -12161,56 +12931,56 @@ class json_iterator {
    *
    * They may include invalid JSON as well (such as `1.2.3` or `ture`).
    */
-  simdjson_really_inline const uint8_t *advance() noexcept;
+  simdjson_inline const uint8_t *advance() noexcept;
   /**
    * Get the remaining length of the document, from the start of the current token.
    */
-  simdjson_really_inline size_t remaining_len() const noexcept;
+  simdjson_inline size_t remaining_len() const noexcept;
   /**
    * Check if we are at the end of the document.
    *
    * If this is true, there are no more tokens.
    */
-  simdjson_really_inline bool at_eof() const noexcept;
+  simdjson_inline bool at_eof() const noexcept;
   /**
    * Check if we are at the beginning of the document.
    */
-  simdjson_really_inline bool at_beginning() const noexcept;
-  simdjson_really_inline uint8_t last_structural() const noexcept;
+  simdjson_inline bool at_beginning() const noexcept;
+  simdjson_inline uint8_t last_structural() const noexcept;
 
   /**
    * Log that a value has been found.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_value(const char *type) const noexcept;
+  simdjson_inline void log_value(const char *type) const noexcept;
   /**
    * Log the start of a multipart value.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_start_value(const char *type) const noexcept;
+  simdjson_inline void log_start_value(const char *type) const noexcept;
   /**
    * Log the end of a multipart value.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_end_value(const char *type) const noexcept;
+  simdjson_inline void log_end_value(const char *type) const noexcept;
   /**
    * Log an error.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_error(const char *error) const noexcept;
+  simdjson_inline void log_error(const char *error) const noexcept;
 
   template<typename V>
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
   template<typename V>
-  simdjson_warn_unused simdjson_really_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
 };
 
 template<bool STREAMING, typename V>
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::walk_document(V &visitor) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::walk_document(V &visitor) noexcept {
   logger::log_start();
 
   //
@@ -12335,52 +13105,52 @@ simdjson_warn_unused simdjson_really_inline error_code json_iterator::walk_docum
 
 } // walk_document()
 
-simdjson_really_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+simdjson_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
   : buf{_dom_parser.buf},
     next_structural{&_dom_parser.structural_indexes[start_structural_index]},
     dom_parser{_dom_parser} {
 }
 
-simdjson_really_inline const uint8_t *json_iterator::peek() const noexcept {
+simdjson_inline const uint8_t *json_iterator::peek() const noexcept {
   return &buf[*(next_structural)];
 }
-simdjson_really_inline const uint8_t *json_iterator::advance() noexcept {
+simdjson_inline const uint8_t *json_iterator::advance() noexcept {
   return &buf[*(next_structural++)];
 }
-simdjson_really_inline size_t json_iterator::remaining_len() const noexcept {
+simdjson_inline size_t json_iterator::remaining_len() const noexcept {
   return dom_parser.len - *(next_structural-1);
 }
 
-simdjson_really_inline bool json_iterator::at_eof() const noexcept {
+simdjson_inline bool json_iterator::at_eof() const noexcept {
   return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
 }
-simdjson_really_inline bool json_iterator::at_beginning() const noexcept {
+simdjson_inline bool json_iterator::at_beginning() const noexcept {
   return next_structural == dom_parser.structural_indexes.get();
 }
-simdjson_really_inline uint8_t json_iterator::last_structural() const noexcept {
+simdjson_inline uint8_t json_iterator::last_structural() const noexcept {
   return buf[dom_parser.structural_indexes[dom_parser.n_structural_indexes - 1]];
 }
 
-simdjson_really_inline void json_iterator::log_value(const char *type) const noexcept {
+simdjson_inline void json_iterator::log_value(const char *type) const noexcept {
   logger::log_line(*this, "", type, "");
 }
 
-simdjson_really_inline void json_iterator::log_start_value(const char *type) const noexcept {
+simdjson_inline void json_iterator::log_start_value(const char *type) const noexcept {
   logger::log_line(*this, "+", type, "");
   if (logger::LOG_ENABLED) { logger::log_depth++; }
 }
 
-simdjson_really_inline void json_iterator::log_end_value(const char *type) const noexcept {
+simdjson_inline void json_iterator::log_end_value(const char *type) const noexcept {
   if (logger::LOG_ENABLED) { logger::log_depth--; }
   logger::log_line(*this, "-", type, "");
 }
 
-simdjson_really_inline void json_iterator::log_error(const char *error) const noexcept {
+simdjson_inline void json_iterator::log_error(const char *error) const noexcept {
   logger::log_line(*this, "", "ERROR", error);
 }
 
 template<typename V>
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
   switch (*value) {
     case '"': return visitor.visit_root_string(*this, value);
     case 't': return visitor.visit_root_true_atom(*this, value);
@@ -12396,7 +13166,7 @@ simdjson_warn_unused simdjson_really_inline error_code json_iterator::visit_root
   }
 }
 template<typename V>
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
   switch (*value) {
     case '"': return visitor.visit_string(*this, value);
     case 't': return visitor.visit_true_atom(*this, value);
@@ -12428,18 +13198,18 @@ struct tape_writer {
   uint64_t *next_tape_loc;
 
   /** Write a signed 64-bit value to tape. */
-  simdjson_really_inline void append_s64(int64_t value) noexcept;
+  simdjson_inline void append_s64(int64_t value) noexcept;
 
   /** Write an unsigned 64-bit value to tape. */
-  simdjson_really_inline void append_u64(uint64_t value) noexcept;
+  simdjson_inline void append_u64(uint64_t value) noexcept;
 
   /** Write a double value to tape. */
-  simdjson_really_inline void append_double(double value) noexcept;
+  simdjson_inline void append_double(double value) noexcept;
 
   /**
    * Append a tape entry (an 8-bit type,and 56 bits worth of value).
    */
-  simdjson_really_inline void append(uint64_t val, internal::tape_type t) noexcept;
+  simdjson_inline void append(uint64_t val, internal::tape_type t) noexcept;
 
   /**
    * Skip the current tape entry without writing.
@@ -12447,24 +13217,24 @@ struct tape_writer {
    * Used to skip the start of the container, since we'll come back later to fill it in when the
    * container ends.
    */
-  simdjson_really_inline void skip() noexcept;
+  simdjson_inline void skip() noexcept;
 
   /**
    * Skip the number of tape entries necessary to write a large u64 or i64.
    */
-  simdjson_really_inline void skip_large_integer() noexcept;
+  simdjson_inline void skip_large_integer() noexcept;
 
   /**
    * Skip the number of tape entries necessary to write a double.
    */
-  simdjson_really_inline void skip_double() noexcept;
+  simdjson_inline void skip_double() noexcept;
 
   /**
    * Write a value to a known location on tape.
    *
    * Used to go back and write out the start of a container after the container ends.
    */
-  simdjson_really_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
+  simdjson_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
 
 private:
   /**
@@ -12472,50 +13242,50 @@ struct tape_writer {
    * all 64 bits, such as double and uint64_t.
    */
   template<typename T>
-  simdjson_really_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
+  simdjson_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
 }; // struct number_writer
 
-simdjson_really_inline void tape_writer::append_s64(int64_t value) noexcept {
+simdjson_inline void tape_writer::append_s64(int64_t value) noexcept {
   append2(0, value, internal::tape_type::INT64);
 }
 
-simdjson_really_inline void tape_writer::append_u64(uint64_t value) noexcept {
+simdjson_inline void tape_writer::append_u64(uint64_t value) noexcept {
   append(0, internal::tape_type::UINT64);
   *next_tape_loc = value;
   next_tape_loc++;
 }
 
 /** Write a double value to tape. */
-simdjson_really_inline void tape_writer::append_double(double value) noexcept {
+simdjson_inline void tape_writer::append_double(double value) noexcept {
   append2(0, value, internal::tape_type::DOUBLE);
 }
 
-simdjson_really_inline void tape_writer::skip() noexcept {
+simdjson_inline void tape_writer::skip() noexcept {
   next_tape_loc++;
 }
 
-simdjson_really_inline void tape_writer::skip_large_integer() noexcept {
+simdjson_inline void tape_writer::skip_large_integer() noexcept {
   next_tape_loc += 2;
 }
 
-simdjson_really_inline void tape_writer::skip_double() noexcept {
+simdjson_inline void tape_writer::skip_double() noexcept {
   next_tape_loc += 2;
 }
 
-simdjson_really_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
+simdjson_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
   *next_tape_loc = val | ((uint64_t(char(t))) << 56);
   next_tape_loc++;
 }
 
 template<typename T>
-simdjson_really_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
+simdjson_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
   append(val, t);
   static_assert(sizeof(val2) == sizeof(*next_tape_loc), "Type is not 64 bits!");
   memcpy(next_tape_loc, &val2, sizeof(val2));
   next_tape_loc++;
 }
 
-simdjson_really_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
+simdjson_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
   tape_loc = val | ((uint64_t(char(t))) << 56);
 }
 
@@ -12532,40 +13302,40 @@ namespace stage2 {
 
 struct tape_builder {
   template<bool STREAMING>
-  simdjson_warn_unused static simdjson_really_inline error_code parse_document(
+  simdjson_warn_unused static simdjson_inline error_code parse_document(
     dom_parser_implementation &dom_parser,
     dom::document &doc) noexcept;
 
   /** Called when a non-empty document starts. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_document_start(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_document_start(json_iterator &iter) noexcept;
   /** Called when a non-empty document ends without error. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_document_end(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_document_end(json_iterator &iter) noexcept;
 
   /** Called when a non-empty array starts. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_array_start(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_array_start(json_iterator &iter) noexcept;
   /** Called when a non-empty array ends. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_array_end(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_array_end(json_iterator &iter) noexcept;
   /** Called when an empty array is found. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_empty_array(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_empty_array(json_iterator &iter) noexcept;
 
   /** Called when a non-empty object starts. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_object_start(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_object_start(json_iterator &iter) noexcept;
   /**
    * Called when a key in a field is encountered.
    *
    * primitive, visit_object_start, visit_empty_object, visit_array_start, or visit_empty_array
    * will be called after this with the field value.
    */
-  simdjson_warn_unused simdjson_really_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
   /** Called when a non-empty object ends. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_object_end(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_object_end(json_iterator &iter) noexcept;
   /** Called when an empty object is found. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_empty_object(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_empty_object(json_iterator &iter) noexcept;
 
   /**
    * Called when a string, number, boolean or null is found.
    */
-  simdjson_warn_unused simdjson_really_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
   /**
    * Called when a string, number, boolean or null is found at the top level of a document (i.e.
    * when there is no array or object and the entire document is a single string, number, boolean or
@@ -12574,22 +13344,22 @@ struct tape_builder {
    * This is separate from primitive() because simdjson's normal primitive parsing routines assume
    * there is at least one more token after the value, which is only true in an array or object.
    */
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
 
-  simdjson_warn_unused simdjson_really_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
 
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
 
   /** Called each time a new field or element in an array or object is found. */
-  simdjson_warn_unused simdjson_really_inline error_code increment_count(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code increment_count(json_iterator &iter) noexcept;
 
   /** Next location to write to tape */
   tape_writer tape;
@@ -12597,18 +13367,18 @@ struct tape_builder {
   /** Next write location in the string buf for stage 2 parsing */
   uint8_t *current_string_buf_loc;
 
-  simdjson_really_inline tape_builder(dom::document &doc) noexcept;
+  simdjson_inline tape_builder(dom::document &doc) noexcept;
 
-  simdjson_really_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
-  simdjson_really_inline void start_container(json_iterator &iter) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
-  simdjson_really_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
-  simdjson_really_inline void on_end_string(uint8_t *dst) noexcept;
+  simdjson_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
+  simdjson_inline void start_container(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_warn_unused simdjson_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
+  simdjson_inline void on_end_string(uint8_t *dst) noexcept;
 }; // class tape_builder
 
 template<bool STREAMING>
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::parse_document(
+simdjson_warn_unused simdjson_inline error_code tape_builder::parse_document(
     dom_parser_implementation &dom_parser,
     dom::document &doc) noexcept {
   dom_parser.doc = &doc;
@@ -12617,56 +13387,56 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::parse_docum
   return iter.walk_document<STREAMING>(builder);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
   return iter.visit_root_primitive(*this, value);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
   return iter.visit_primitive(*this, value);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
   return empty_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
   return empty_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
   start_container(iter);
   return SUCCESS;
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
   start_container(iter);
   return SUCCESS;
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
   start_container(iter);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
   return end_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
   return end_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
   constexpr uint32_t start_tape_index = 0;
   tape.append(start_tape_index, internal::tape_type::ROOT);
   tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter), internal::tape_type::ROOT);
   return SUCCESS;
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
   return visit_string(iter, key, true);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
   iter.dom_parser.open_containers[iter.depth].count++; // we have a key value pair in the object at parser.dom_parser.depth - 1
   return SUCCESS;
 }
 
-simdjson_really_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
+simdjson_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
   iter.log_value(key ? "key" : "string");
   uint8_t *dst = on_start_string(iter);
   dst = stringparsing::parse_string(value+1, dst);
@@ -12678,16 +13448,16 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_strin
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
   return visit_string(iter, value);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("number");
   return numberparsing::parse_number(value, tape);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
   //
   // We need to make a copy to make sure that the string is space terminated.
   // This is not about padding the input, which should already padded up
@@ -12709,42 +13479,42 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_
   return error;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("true");
   if (!atomparsing::is_valid_true_atom(value)) { return T_ATOM_ERROR; }
   tape.append(0, internal::tape_type::TRUE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("true");
   if (!atomparsing::is_valid_true_atom(value, iter.remaining_len())) { return T_ATOM_ERROR; }
   tape.append(0, internal::tape_type::TRUE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("false");
   if (!atomparsing::is_valid_false_atom(value)) { return F_ATOM_ERROR; }
   tape.append(0, internal::tape_type::FALSE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("false");
   if (!atomparsing::is_valid_false_atom(value, iter.remaining_len())) { return F_ATOM_ERROR; }
   tape.append(0, internal::tape_type::FALSE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("null");
   if (!atomparsing::is_valid_null_atom(value)) { return N_ATOM_ERROR; }
   tape.append(0, internal::tape_type::NULL_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("null");
   if (!atomparsing::is_valid_null_atom(value, iter.remaining_len())) { return N_ATOM_ERROR; }
   tape.append(0, internal::tape_type::NULL_VALUE);
@@ -12753,24 +13523,24 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_
 
 // private:
 
-simdjson_really_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
+simdjson_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
   return uint32_t(tape.next_tape_loc - iter.dom_parser.doc->tape.get());
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
   auto start_index = next_tape_index(iter);
   tape.append(start_index+2, start);
   tape.append(start_index, end);
   return SUCCESS;
 }
 
-simdjson_really_inline void tape_builder::start_container(json_iterator &iter) noexcept {
+simdjson_inline void tape_builder::start_container(json_iterator &iter) noexcept {
   iter.dom_parser.open_containers[iter.depth].tape_index = next_tape_index(iter);
   iter.dom_parser.open_containers[iter.depth].count = 0;
   tape.skip(); // We don't actually *write* the start element until the end.
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
   // Write the ending tape element, pointing at the start location
   const uint32_t start_tape_index = iter.dom_parser.open_containers[iter.depth].tape_index;
   tape.append(start_tape_index, end);
@@ -12783,13 +13553,13 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::end_contain
   return SUCCESS;
 }
 
-simdjson_really_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
+simdjson_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
   // we advance the point, accounting for the fact that we have a NULL termination
   tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::STRING);
   return current_string_buf_loc + sizeof(uint32_t);
 }
 
-simdjson_really_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
+simdjson_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
   uint32_t str_length = uint32_t(dst - (current_string_buf_loc + sizeof(uint32_t)));
   // TODO check for overflow in case someone has a crazy string (>=4GB?)
   // But only add the overflow check when the document itself exceeds 4GB
@@ -12815,7 +13585,7 @@ namespace ppc64 {
 namespace {
 namespace stage1 {
 
-simdjson_really_inline uint64_t json_string_scanner::find_escaped(uint64_t backslash) {
+simdjson_inline uint64_t json_string_scanner::find_escaped(uint64_t backslash) {
   // On PPC, we don't short-circuit this if there are no backslashes, because the branch gives us no
   // benefit and therefore makes things worse.
   // if (!backslash) { uint64_t escaped = prev_escaped; prev_escaped = 0; return escaped; }
@@ -12847,6 +13617,10 @@ simdjson_warn_unused error_code dom_parser_implementation::stage2_next(dom::docu
   return stage2::tape_builder::parse_document<true>(*this, _doc);
 }
 
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_string(const uint8_t *src, uint8_t *dst) const noexcept {
+  return ppc64::stringparsing::parse_string(src, dst);
+}
+
 simdjson_warn_unused error_code dom_parser_implementation::parse(const uint8_t *_buf, size_t _len, dom::document &_doc) noexcept {
   auto error = stage1(_buf, _len, stage1_mode::regular);
   if (error) { return error; }
@@ -12910,17 +13684,17 @@ namespace {
 using namespace simd;
 
 struct json_character_block {
-  static simdjson_really_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
+  static simdjson_inline json_character_block classify(const simd::simd8x64<uint8_t>& in);
 
-  simdjson_really_inline uint64_t whitespace() const noexcept { return _whitespace; }
-  simdjson_really_inline uint64_t op() const noexcept { return _op; }
-  simdjson_really_inline uint64_t scalar() const noexcept { return ~(op() | whitespace()); }
+  simdjson_inline uint64_t whitespace() const noexcept { return _whitespace; }
+  simdjson_inline uint64_t op() const noexcept { return _op; }
+  simdjson_inline uint64_t scalar() const noexcept { return ~(op() | whitespace()); }
 
   uint64_t _whitespace;
   uint64_t _op;
 };
 
-simdjson_really_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
+simdjson_inline json_character_block json_character_block::classify(const simd::simd8x64<uint8_t>& in) {
   // These lookups rely on the fact that anything < 127 will match the lower 4 bits, which is why
   // we can't use the generic lookup_16.
   auto whitespace_table = simd8<uint8_t>::repeat_16(' ', 100, 100, 100, 17, 100, 113, 2, 100, '\t', '\n', 112, 100, '\r', 100, 100);
@@ -12979,21 +13753,21 @@ simdjson_really_inline json_character_block json_character_block::classify(const
     return { whitespace, op };
 }
 
-simdjson_really_inline bool is_ascii(const simd8x64<uint8_t>& input) {
+simdjson_inline bool is_ascii(const simd8x64<uint8_t>& input) {
   return input.reduce_or().is_ascii();
 }
 
-simdjson_unused simdjson_really_inline simd8<bool> must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
-  simd8<uint8_t> is_second_byte = prev1.saturating_sub(0b11000000u-1); // Only 11______ will be > 0
-  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0b11100000u-1); // Only 111_____ will be > 0
-  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0b11110000u-1); // Only 1111____ will be > 0
+simdjson_unused simdjson_inline simd8<bool> must_be_continuation(const simd8<uint8_t> prev1, const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_second_byte = prev1.saturating_sub(0xc0u-1); // Only 11______ will be > 0
+  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-1); // Only 111_____ will be > 0
+  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-1); // Only 1111____ will be > 0
   // Caller requires a bool (all 1's). All values resulting from the subtraction will be <= 64, so signed comparison is fine.
   return simd8<int8_t>(is_second_byte | is_third_byte | is_fourth_byte) > int8_t(0);
 }
 
-simdjson_really_inline simd8<bool> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
-  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0b11100000u-1); // Only 111_____ will be > 0
-  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0b11110000u-1); // Only 1111____ will be > 0
+simdjson_inline simd8<bool> must_be_2_3_continuation(const simd8<uint8_t> prev2, const simd8<uint8_t> prev3) {
+  simd8<uint8_t> is_third_byte  = prev2.saturating_sub(0xe0u-1); // Only 111_____ will be > 0
+  simd8<uint8_t> is_fourth_byte = prev3.saturating_sub(0xf0u-1); // Only 1111____ will be > 0
   // Caller requires a bool (all 1's). All values resulting from the subtraction will be <= 64, so signed comparison is fine.
   return simd8<int8_t>(is_third_byte | is_fourth_byte) > int8_t(0);
 }
@@ -13010,7 +13784,7 @@ namespace utf8_validation {
 
 using namespace simd;
 
-  simdjson_really_inline simd8<uint8_t> check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
+  simdjson_inline simd8<uint8_t> check_special_cases(const simd8<uint8_t> input, const simd8<uint8_t> prev1) {
 // Bit 0 = Too Short (lead byte/ASCII followed by lead byte/ASCII)
 // Bit 1 = Too Long (ASCII followed by continuation)
 // Bit 2 = Overlong 3-byte
@@ -13100,7 +13874,7 @@ using namespace simd;
     );
     return (byte_1_high & byte_1_low & byte_2_high);
   }
-  simdjson_really_inline simd8<uint8_t> check_multibyte_lengths(const simd8<uint8_t> input,
+  simdjson_inline simd8<uint8_t> check_multibyte_lengths(const simd8<uint8_t> input,
       const simd8<uint8_t> prev_input, const simd8<uint8_t> sc) {
     simd8<uint8_t> prev2 = input.prev<2>(prev_input);
     simd8<uint8_t> prev3 = input.prev<3>(prev_input);
@@ -13113,7 +13887,7 @@ using namespace simd;
   // Return nonzero if there are incomplete multibyte characters at the end of the block:
   // e.g. if there is a 4-byte character, but it's 3 bytes from the end.
   //
-  simdjson_really_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
+  simdjson_inline simd8<uint8_t> is_incomplete(const simd8<uint8_t> input) {
     // If the previous input's last 3 bytes match this, they're too short (they ended at EOF):
     // ... 1111____ 111_____ 11______
 #if SIMDJSON_IMPLEMENTATION_ICELAKE
@@ -13125,14 +13899,14 @@ using namespace simd;
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
-      255, 255, 255, 255, 255, 0b11110000u-1, 0b11100000u-1, 0b11000000u-1
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
     };
 #else
     static const uint8_t max_array[32] = {
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
       255, 255, 255, 255, 255, 255, 255, 255,
-      255, 255, 255, 255, 255, 0b11110000u-1, 0b11100000u-1, 0b11000000u-1
+      255, 255, 255, 255, 255, 0xf0u-1, 0xe0u-1, 0xc0u-1
     };
 #endif
     const simd8<uint8_t> max_value(&max_array[sizeof(max_array)-sizeof(simd8<uint8_t>)]);
@@ -13150,7 +13924,7 @@ using namespace simd;
     //
     // Check whether the current bytes are valid UTF-8.
     //
-    simdjson_really_inline void check_utf8_bytes(const simd8<uint8_t> input, const simd8<uint8_t> prev_input) {
+    simdjson_inline void check_utf8_bytes(const simd8<uint8_t> input, const simd8<uint8_t> prev_input) {
       // Flip prev1...prev3 so we can easily determine if they are 2+, 3+ or 4+ lead bytes
       // (2, 3, 4-byte leads become large positive numbers instead of small negative numbers)
       simd8<uint8_t> prev1 = input.prev<1>(prev_input);
@@ -13161,13 +13935,21 @@ using namespace simd;
     // The only problem that can happen at EOF is that a multibyte character is too short
     // or a byte value too large in the last bytes: check_special_cases only checks for bytes
     // too large in the first of two bytes.
-    simdjson_really_inline void check_eof() {
+    simdjson_inline void check_eof() {
       // If the previous block had incomplete UTF-8 characters at the end, an ASCII block can't
       // possibly finish them.
       this->error |= this->prev_incomplete;
     }
 
-    simdjson_really_inline void check_next_input(const simd8x64<uint8_t>& input) {
+#ifndef SIMDJSON_IF_CONSTEXPR
+#if SIMDJSON_CPLUSPLUS17
+#define SIMDJSON_IF_CONSTEXPR if constexpr
+#else
+#define SIMDJSON_IF_CONSTEXPR if
+#endif
+#endif
+
+    simdjson_inline void check_next_input(const simd8x64<uint8_t>& input) {
       if(simdjson_likely(is_ascii(input))) {
         this->error |= this->prev_incomplete;
       } else {
@@ -13176,12 +13958,12 @@ using namespace simd;
                 ||(simd8x64<uint8_t>::NUM_CHUNKS == 2)
                 || (simd8x64<uint8_t>::NUM_CHUNKS == 4),
                 "We support one, two or four chunks per 64-byte block.");
-        if(simd8x64<uint8_t>::NUM_CHUNKS == 1) {
+        SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 1) {
           this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
-        } if(simd8x64<uint8_t>::NUM_CHUNKS == 2) {
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 2) {
           this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
           this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
-        } else if(simd8x64<uint8_t>::NUM_CHUNKS == 4) {
+        } else SIMDJSON_IF_CONSTEXPR (simd8x64<uint8_t>::NUM_CHUNKS == 4) {
           this->check_utf8_bytes(input.chunks[0], this->prev_input_block);
           this->check_utf8_bytes(input.chunks[1], input.chunks[0]);
           this->check_utf8_bytes(input.chunks[2], input.chunks[1]);
@@ -13192,7 +13974,7 @@ using namespace simd;
       }
     }
     // do not forget to call check_eof!
-    simdjson_really_inline error_code errors() {
+    simdjson_inline error_code errors() {
       return this->error.any_bits_set_anywhere() ? error_code::UTF8_ERROR : error_code::SUCCESS;
     }
 
@@ -13220,10 +14002,10 @@ namespace {
 template<size_t STEP_SIZE>
 struct buf_block_reader {
 public:
-  simdjson_really_inline buf_block_reader(const uint8_t *_buf, size_t _len);
-  simdjson_really_inline size_t block_index();
-  simdjson_really_inline bool has_full_block() const;
-  simdjson_really_inline const uint8_t *full_block() const;
+  simdjson_inline buf_block_reader(const uint8_t *_buf, size_t _len);
+  simdjson_inline size_t block_index();
+  simdjson_inline bool has_full_block() const;
+  simdjson_inline const uint8_t *full_block() const;
   /**
    * Get the last block, padded with spaces.
    *
@@ -13233,8 +14015,8 @@ struct buf_block_reader {
    *
    * @return the number of effective characters in the last block.
    */
-  simdjson_really_inline size_t get_remainder(uint8_t *dst) const;
-  simdjson_really_inline void advance();
+  simdjson_inline size_t get_remainder(uint8_t *dst) const;
+  simdjson_inline void advance();
 private:
   const uint8_t *buf;
   const size_t len;
@@ -13273,23 +14055,23 @@ simdjson_unused static char * format_mask(uint64_t mask) {
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len) : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, idx{0} {}
+simdjson_inline buf_block_reader<STEP_SIZE>::buf_block_reader(const uint8_t *_buf, size_t _len) : buf{_buf}, len{_len}, lenminusstep{len < STEP_SIZE ? 0 : len - STEP_SIZE}, idx{0} {}
 
 template<size_t STEP_SIZE>
-simdjson_really_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::block_index() { return idx; }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
+simdjson_inline bool buf_block_reader<STEP_SIZE>::has_full_block() const {
   return idx < lenminusstep;
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
+simdjson_inline const uint8_t *buf_block_reader<STEP_SIZE>::full_block() const {
   return &buf[idx];
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
+simdjson_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t *dst) const {
   if(len == idx) { return 0; } // memcpy(dst, null, 0) will trigger an error with some sanitizers
   std::memset(dst, 0x20, STEP_SIZE); // std::memset STEP_SIZE because it's more efficient to write out 8 or 16 bytes at once.
   std::memcpy(dst, buf + idx, len - idx);
@@ -13297,7 +14079,7 @@ simdjson_really_inline size_t buf_block_reader<STEP_SIZE>::get_remainder(uint8_t
 }
 
 template<size_t STEP_SIZE>
-simdjson_really_inline void buf_block_reader<STEP_SIZE>::advance() {
+simdjson_inline void buf_block_reader<STEP_SIZE>::advance() {
   idx += STEP_SIZE;
 }
 
@@ -13313,27 +14095,27 @@ namespace stage1 {
 
 struct json_string_block {
   // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
-  simdjson_really_inline json_string_block(uint64_t backslash, uint64_t escaped, uint64_t quote, uint64_t in_string) :
+  simdjson_inline json_string_block(uint64_t backslash, uint64_t escaped, uint64_t quote, uint64_t in_string) :
   _backslash(backslash), _escaped(escaped), _quote(quote), _in_string(in_string) {}
 
   // Escaped characters (characters following an escape() character)
-  simdjson_really_inline uint64_t escaped() const { return _escaped; }
+  simdjson_inline uint64_t escaped() const { return _escaped; }
   // Escape characters (backslashes that are not escaped--i.e. in \\, includes only the first \)
-  simdjson_really_inline uint64_t escape() const { return _backslash & ~_escaped; }
+  simdjson_inline uint64_t escape() const { return _backslash & ~_escaped; }
   // Real (non-backslashed) quotes
-  simdjson_really_inline uint64_t quote() const { return _quote; }
+  simdjson_inline uint64_t quote() const { return _quote; }
   // Start quotes of strings
-  simdjson_really_inline uint64_t string_start() const { return _quote & _in_string; }
+  simdjson_inline uint64_t string_start() const { return _quote & _in_string; }
   // End quotes of strings
-  simdjson_really_inline uint64_t string_end() const { return _quote & ~_in_string; }
+  simdjson_inline uint64_t string_end() const { return _quote & ~_in_string; }
   // Only characters inside the string (not including the quotes)
-  simdjson_really_inline uint64_t string_content() const { return _in_string & ~_quote; }
+  simdjson_inline uint64_t string_content() const { return _in_string & ~_quote; }
   // Return a mask of whether the given characters are inside a string (only works on non-quotes)
-  simdjson_really_inline uint64_t non_quote_inside_string(uint64_t mask) const { return mask & _in_string; }
+  simdjson_inline uint64_t non_quote_inside_string(uint64_t mask) const { return mask & _in_string; }
   // Return a mask of whether the given characters are inside a string (only works on non-quotes)
-  simdjson_really_inline uint64_t non_quote_outside_string(uint64_t mask) const { return mask & ~_in_string; }
+  simdjson_inline uint64_t non_quote_outside_string(uint64_t mask) const { return mask & ~_in_string; }
   // Tail of string (everything except the start quote)
-  simdjson_really_inline uint64_t string_tail() const { return _in_string ^ _quote; }
+  simdjson_inline uint64_t string_tail() const { return _in_string ^ _quote; }
 
   // backslash characters
   uint64_t _backslash;
@@ -13348,14 +14130,14 @@ struct json_string_block {
 // Scans blocks for string characters, storing the state necessary to do so
 class json_string_scanner {
 public:
-  simdjson_really_inline json_string_block next(const simd::simd8x64<uint8_t>& in);
+  simdjson_inline json_string_block next(const simd::simd8x64<uint8_t>& in);
   // Returns either UNCLOSED_STRING or SUCCESS
-  simdjson_really_inline error_code finish();
+  simdjson_inline error_code finish();
 
 private:
   // Intended to be defined by the implementation
-  simdjson_really_inline uint64_t find_escaped(uint64_t escape);
-  simdjson_really_inline uint64_t find_escaped_branchless(uint64_t escape);
+  simdjson_inline uint64_t find_escaped(uint64_t escape);
+  simdjson_inline uint64_t find_escaped_branchless(uint64_t escape);
 
   // Whether the last iteration was still inside a string (all 1's = true, all 0's = false).
   uint64_t prev_in_string = 0ULL;
@@ -13390,7 +14172,7 @@ class json_string_scanner {
 // desired        |   x  | x x  x x  x x  x  x  |
 // text           |  \\\ | \\\"\\\" \\\" \\"\\" |
 //
-simdjson_really_inline uint64_t json_string_scanner::find_escaped_branchless(uint64_t backslash) {
+simdjson_inline uint64_t json_string_scanner::find_escaped_branchless(uint64_t backslash) {
   // If there was overflow, pretend the first character isn't a backslash
   backslash &= ~prev_escaped;
   uint64_t follows_escape = backslash << 1 | prev_escaped;
@@ -13415,7 +14197,7 @@ simdjson_really_inline uint64_t json_string_scanner::find_escaped_branchless(uin
 //
 // Backslash sequences outside of quotes will be detected in stage 2.
 //
-simdjson_really_inline json_string_block json_string_scanner::next(const simd::simd8x64<uint8_t>& in) {
+simdjson_inline json_string_block json_string_scanner::next(const simd::simd8x64<uint8_t>& in) {
   const uint64_t backslash = in.eq('\\');
   const uint64_t escaped = find_escaped(backslash);
   const uint64_t quote = in.eq('"') & ~escaped;
@@ -13448,7 +14230,7 @@ simdjson_really_inline json_string_block json_string_scanner::next(const simd::s
   );
 }
 
-simdjson_really_inline error_code json_string_scanner::finish() {
+simdjson_inline error_code json_string_scanner::finish() {
   if (prev_in_string) {
     return UNCLOSED_STRING;
   }
@@ -13486,25 +14268,25 @@ namespace stage1 {
 struct json_block {
 public:
   // We spell out the constructors in the hope of resolving inlining issues with Visual Studio 2017
-  simdjson_really_inline json_block(json_string_block&& string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  simdjson_inline json_block(json_string_block&& string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
   _string(std::move(string)), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
-  simdjson_really_inline json_block(json_string_block string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
+  simdjson_inline json_block(json_string_block string, json_character_block characters, uint64_t follows_potential_nonquote_scalar) :
   _string(string), _characters(characters), _follows_potential_nonquote_scalar(follows_potential_nonquote_scalar) {}
 
   /**
    * The start of structurals.
    * In simdjson prior to v0.3, these were called the pseudo-structural characters.
    **/
-  simdjson_really_inline uint64_t structural_start() const noexcept { return potential_structural_start() & ~_string.string_tail(); }
+  simdjson_inline uint64_t structural_start() const noexcept { return potential_structural_start() & ~_string.string_tail(); }
   /** All JSON whitespace (i.e. not in a string) */
-  simdjson_really_inline uint64_t whitespace() const noexcept { return non_quote_outside_string(_characters.whitespace()); }
+  simdjson_inline uint64_t whitespace() const noexcept { return non_quote_outside_string(_characters.whitespace()); }
 
   // Helpers
 
   /** Whether the given characters are inside a string (only works on non-quotes) */
-  simdjson_really_inline uint64_t non_quote_inside_string(uint64_t mask) const noexcept { return _string.non_quote_inside_string(mask); }
+  simdjson_inline uint64_t non_quote_inside_string(uint64_t mask) const noexcept { return _string.non_quote_inside_string(mask); }
   /** Whether the given characters are outside a string (only works on non-quotes) */
-  simdjson_really_inline uint64_t non_quote_outside_string(uint64_t mask) const noexcept { return _string.non_quote_outside_string(mask); }
+  simdjson_inline uint64_t non_quote_outside_string(uint64_t mask) const noexcept { return _string.non_quote_outside_string(mask); }
 
   // string and escape characters
   json_string_block _string;
@@ -13519,12 +14301,12 @@ struct json_block {
    * structural elements ([,],{,},:, comma) plus scalar starts like 123, true and "abc".
    * They may reside inside a string.
    **/
-  simdjson_really_inline uint64_t potential_structural_start() const noexcept { return _characters.op() | potential_scalar_start(); }
+  simdjson_inline uint64_t potential_structural_start() const noexcept { return _characters.op() | potential_scalar_start(); }
   /**
    * The start of non-operator runs, like 123, true and "abc".
    * It main reside inside a string.
    **/
-  simdjson_really_inline uint64_t potential_scalar_start() const noexcept {
+  simdjson_inline uint64_t potential_scalar_start() const noexcept {
     // The term "scalar" refers to anything except structural characters and white space
     // (so letters, numbers, quotes).
     // Whenever it is preceded by something that is not a structural element ({,},[,],:, ") nor a white-space
@@ -13535,7 +14317,7 @@ struct json_block {
    * Whether the given character is immediately after a non-operator like 123, true.
    * The characters following a quote are not included.
    */
-  simdjson_really_inline uint64_t follows_potential_scalar() const noexcept {
+  simdjson_inline uint64_t follows_potential_scalar() const noexcept {
     // _follows_potential_nonquote_scalar: is defined as marking any character that follows a character
     // that is not a structural element ({,},[,],:, comma) nor a quote (") and that is not a
     // white space.
@@ -13559,10 +14341,10 @@ struct json_block {
  */
 class json_scanner {
 public:
-  json_scanner() {}
-  simdjson_really_inline json_block next(const simd::simd8x64<uint8_t>& in);
+  json_scanner() = default;
+  simdjson_inline json_block next(const simd::simd8x64<uint8_t>& in);
   // Returns either UNCLOSED_STRING or SUCCESS
-  simdjson_really_inline error_code finish();
+  simdjson_inline error_code finish();
 
 private:
   // Whether the last character of the previous iteration is part of a scalar token
@@ -13579,13 +14361,13 @@ class json_scanner {
 //
 //     const uint64_t backslashed_quote = in.eq('"') & immediately_follows(in.eq('\'), prev_backslash);
 //
-simdjson_really_inline uint64_t follows(const uint64_t match, uint64_t &overflow) {
+simdjson_inline uint64_t follows(const uint64_t match, uint64_t &overflow) {
   const uint64_t result = match << 1 | overflow;
   overflow = match >> 63;
   return result;
 }
 
-simdjson_really_inline json_block json_scanner::next(const simd::simd8x64<uint8_t>& in) {
+simdjson_inline json_block json_scanner::next(const simd::simd8x64<uint8_t>& in) {
   json_string_block strings = string_scanner.next(in);
   // identifies the white-space and the structural characters
   json_character_block characters = json_character_block::classify(in);
@@ -13610,7 +14392,7 @@ simdjson_really_inline json_block json_scanner::next(const simd::simd8x64<uint8_
   );
 }
 
-simdjson_really_inline error_code json_scanner::finish() {
+simdjson_inline error_code json_scanner::finish() {
   return string_scanner.finish();
 }
 
@@ -13636,23 +14418,23 @@ class json_minifier {
   static error_code minify(const uint8_t *buf, size_t len, uint8_t *dst, size_t &dst_len) noexcept;
 
 private:
-  simdjson_really_inline json_minifier(uint8_t *_dst)
+  simdjson_inline json_minifier(uint8_t *_dst)
   : dst{_dst}
   {}
   template<size_t STEP_SIZE>
-  simdjson_really_inline void step(const uint8_t *block_buf, buf_block_reader<STEP_SIZE> &reader) noexcept;
-  simdjson_really_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block);
-  simdjson_really_inline error_code finish(uint8_t *dst_start, size_t &dst_len);
+  simdjson_inline void step(const uint8_t *block_buf, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block);
+  simdjson_inline error_code finish(uint8_t *dst_start, size_t &dst_len);
   json_scanner scanner{};
   uint8_t *dst;
 };
 
-simdjson_really_inline void json_minifier::next(const simd::simd8x64<uint8_t>& in, const json_block& block) {
+simdjson_inline void json_minifier::next(const simd::simd8x64<uint8_t>& in, const json_block& block) {
   uint64_t mask = block.whitespace();
   dst += in.compress(mask, dst);
 }
 
-simdjson_really_inline error_code json_minifier::finish(uint8_t *dst_start, size_t &dst_len) {
+simdjson_inline error_code json_minifier::finish(uint8_t *dst_start, size_t &dst_len) {
   error_code error = scanner.finish();
   if (error) { dst_len = 0; return error; }
   dst_len = dst - dst_start;
@@ -13660,7 +14442,7 @@ simdjson_really_inline error_code json_minifier::finish(uint8_t *dst_start, size
 }
 
 template<>
-simdjson_really_inline void json_minifier::step<128>(const uint8_t *block_buf, buf_block_reader<128> &reader) noexcept {
+simdjson_inline void json_minifier::step<128>(const uint8_t *block_buf, buf_block_reader<128> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block_buf);
   simd::simd8x64<uint8_t> in_2(block_buf+64);
   json_block block_1 = scanner.next(in_1);
@@ -13671,7 +14453,7 @@ simdjson_really_inline void json_minifier::step<128>(const uint8_t *block_buf, b
 }
 
 template<>
-simdjson_really_inline void json_minifier::step<64>(const uint8_t *block_buf, buf_block_reader<64> &reader) noexcept {
+simdjson_inline void json_minifier::step<64>(const uint8_t *block_buf, buf_block_reader<64> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block_buf);
   json_block block_1 = scanner.next(in_1);
   this->next(block_buf, block_1);
@@ -13744,7 +14526,7 @@ namespace {
   * complete document, therefore the last json buffer location is the end of the
   * batch.
   */
-simdjson_really_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
+simdjson_inline uint32_t find_next_document_index(dom_parser_implementation &parser) {
   // Variant: do not count separately, just figure out depth
   if(parser.n_structural_indexes == 0) { return 0; }
   auto arr_cnt = 0;
@@ -13819,7 +14601,7 @@ class bit_indexer {
 public:
   uint32_t *tail;
 
-  simdjson_really_inline bit_indexer(uint32_t *index_buf) : tail(index_buf) {}
+  simdjson_inline bit_indexer(uint32_t *index_buf) : tail(index_buf) {}
 
   // flatten out values in 'bits' assuming that they are are to have values of idx
   // plus their position in the bitvector, and store these indexes at
@@ -13830,9 +14612,9 @@ class bit_indexer {
   // If the kernel sets SIMDJSON_CUSTOM_BIT_INDEXER, then it will provide its own
   // version of the code.
 #ifdef SIMDJSON_CUSTOM_BIT_INDEXER
-  simdjson_really_inline void write(uint32_t idx, uint64_t bits);
+  simdjson_inline void write(uint32_t idx, uint64_t bits);
 #else
-  simdjson_really_inline void write(uint32_t idx, uint64_t bits) {
+  simdjson_inline void write(uint32_t idx, uint64_t bits) {
     // In some instances, the next branch is expensive because it is mispredicted.
     // Unfortunately, in other cases,
     // it helps tremendously.
@@ -13941,11 +14723,11 @@ class json_structural_indexer {
   static error_code index(const uint8_t *buf, size_t len, dom_parser_implementation &parser, stage1_mode partial) noexcept;
 
 private:
-  simdjson_really_inline json_structural_indexer(uint32_t *structural_indexes);
+  simdjson_inline json_structural_indexer(uint32_t *structural_indexes);
   template<size_t STEP_SIZE>
-  simdjson_really_inline void step(const uint8_t *block, buf_block_reader<STEP_SIZE> &reader) noexcept;
-  simdjson_really_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx);
-  simdjson_really_inline error_code finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial);
+  simdjson_inline void step(const uint8_t *block, buf_block_reader<STEP_SIZE> &reader) noexcept;
+  simdjson_inline void next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx);
+  simdjson_inline error_code finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial);
 
   json_scanner scanner{};
   utf8_checker checker{};
@@ -13954,26 +14736,26 @@ class json_structural_indexer {
   uint64_t unescaped_chars_error = 0;
 };
 
-simdjson_really_inline json_structural_indexer::json_structural_indexer(uint32_t *structural_indexes) : indexer{structural_indexes} {}
+simdjson_inline json_structural_indexer::json_structural_indexer(uint32_t *structural_indexes) : indexer{structural_indexes} {}
 
 // Skip the last character if it is partial
-simdjson_really_inline size_t trim_partial_utf8(const uint8_t *buf, size_t len) {
+simdjson_inline size_t trim_partial_utf8(const uint8_t *buf, size_t len) {
   if (simdjson_unlikely(len < 3)) {
     switch (len) {
       case 2:
-        if (buf[len-1] >= 0b11000000) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
-        if (buf[len-2] >= 0b11100000) { return len-2; } // 3- and 4-byte characters with only 2 bytes left
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 2 bytes left
         return len;
       case 1:
-        if (buf[len-1] >= 0b11000000) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+        if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
         return len;
       case 0:
         return len;
     }
   }
-  if (buf[len-1] >= 0b11000000) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
-  if (buf[len-2] >= 0b11100000) { return len-2; } // 3- and 4-byte characters with only 1 byte left
-  if (buf[len-3] >= 0b11110000) { return len-3; } // 4-byte characters with only 3 bytes left
+  if (buf[len-1] >= 0xc0) { return len-1; } // 2-, 3- and 4-byte characters with only 1 byte left
+  if (buf[len-2] >= 0xe0) { return len-2; } // 3- and 4-byte characters with only 1 byte left
+  if (buf[len-3] >= 0xf0) { return len-3; } // 4-byte characters with only 3 bytes left
   return len;
 }
 
@@ -14022,7 +14804,7 @@ error_code json_structural_indexer::index(const uint8_t *buf, size_t len, dom_pa
 }
 
 template<>
-simdjson_really_inline void json_structural_indexer::step<128>(const uint8_t *block, buf_block_reader<128> &reader) noexcept {
+simdjson_inline void json_structural_indexer::step<128>(const uint8_t *block, buf_block_reader<128> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block);
   simd::simd8x64<uint8_t> in_2(block+64);
   json_block block_1 = scanner.next(in_1);
@@ -14033,14 +14815,14 @@ simdjson_really_inline void json_structural_indexer::step<128>(const uint8_t *bl
 }
 
 template<>
-simdjson_really_inline void json_structural_indexer::step<64>(const uint8_t *block, buf_block_reader<64> &reader) noexcept {
+simdjson_inline void json_structural_indexer::step<64>(const uint8_t *block, buf_block_reader<64> &reader) noexcept {
   simd::simd8x64<uint8_t> in_1(block);
   json_block block_1 = scanner.next(in_1);
   this->next(in_1, block_1, reader.block_index());
   reader.advance();
 }
 
-simdjson_really_inline void json_structural_indexer::next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx) {
+simdjson_inline void json_structural_indexer::next(const simd::simd8x64<uint8_t>& in, const json_block& block, size_t idx) {
   uint64_t unescaped = in.lteq(0x1F);
   checker.check_next_input(in);
   indexer.write(uint32_t(idx-64), prev_structurals); // Output *last* iteration's structurals to the parser
@@ -14048,7 +14830,7 @@ simdjson_really_inline void json_structural_indexer::next(const simd::simd8x64<u
   unescaped_chars_error |= block.non_quote_inside_string(unescaped);
 }
 
-simdjson_really_inline error_code json_structural_indexer::finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial) {
+simdjson_inline error_code json_structural_indexer::finish(dom_parser_implementation &parser, size_t idx, size_t len, stage1_mode partial) {
   // Write out the final iteration's structurals
   indexer.write(uint32_t(idx-64), prev_structurals);
   error_code error = scanner.finish();
@@ -14194,6 +14976,151 @@ bool generic_validate_utf8(const char * input, size_t length) {
 //
 // Stage 2
 //
+/* begin file src/generic/stage2/stringparsing.h */
+// This file contains the common code every implementation uses
+// It is intended to be included multiple times and compiled multiple times
+
+namespace simdjson {
+namespace westmere {
+namespace {
+/// @private
+namespace stringparsing {
+
+// begin copypasta
+// These chars yield themselves: " \ /
+// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
+// u not handled in this table as it's complex
+static const uint8_t escape_map[256] = {
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
+    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
+    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
+};
+
+// handle a unicode codepoint
+// write appropriate values into dest
+// src will advance 6 bytes or 12 bytes
+// dest will advance a variable amount (return via pointer)
+// return true if the unicode codepoint was valid
+// We work in little-endian then swap at write time
+simdjson_warn_unused
+simdjson_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
+                                            uint8_t **dst_ptr) {
+  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
+  // conversion isn't valid; we defer the check for this to inside the
+  // multilingual plane check
+  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
+  *src_ptr += 6;
+
+  // If we found a high surrogate, we must
+  // check for low surrogate for characters
+  // outside the Basic
+  // Multilingual Plane.
+  if (code_point >= 0xd800 && code_point < 0xdc00) {
+    const uint8_t *src_data = *src_ptr;
+    /* Compiler optimizations convert this to a single 16-bit load and compare on most platforms */
+    if (((src_data[0] << 8) | src_data[1]) != ((static_cast<uint8_t> ('\\') << 8) | static_cast<uint8_t> ('u'))) {
+      return false;
+    }
+    uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(src_data + 2);
+
+    // We have already checked that the high surrogate is valid and
+    // (code_point - 0xd800) < 1024.
+    //
+    // Check that code_point_2 is in the range 0xdc00..0xdfff
+    // and that code_point_2 was parsed from valid hex.
+    uint32_t low_bit = code_point_2 - 0xdc00;
+    if (low_bit >> 10) {
+      return false;
+    }
+
+    code_point =
+        (((code_point - 0xd800) << 10) | low_bit) + 0x10000;
+    *src_ptr += 6;
+  } else if (code_point >= 0xdc00 && code_point <= 0xdfff) {
+      // If we encounter a low surrogate (not preceded by a high surrogate)
+      // then we have an error.
+      return false;
+  }
+  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
+  *dst_ptr += offset;
+  return offset > 0;
+}
+
+/**
+ * Unescape a valid UTF-8 string from src to dst, stopping at a final unescaped quote. There
+ * must be an unescaped quote terminating the string. It returns the final output
+ * position as pointer. In case of error (e.g., the string has bad escaped codes),
+ * then null_nullptrptr is returned. It is assumed that the output buffer is large
+ * enough. E.g., if src points at 'joe"', then dst needs to have four free bytes +
+ * SIMDJSON_PADDING bytes.
+ */
+simdjson_warn_unused simdjson_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst) {
+  while (1) {
+    // Copy the next n bytes, and find the backslash and quote in them.
+    auto bs_quote = backslash_and_quote::copy_and_find(src, dst);
+    // If the next thing is the end quote, copy and return
+    if (bs_quote.has_quote_first()) {
+      // we encountered quotes first. Move dst to point to quotes and exit
+      return dst + bs_quote.quote_index();
+    }
+    if (bs_quote.has_backslash()) {
+      /* find out where the backspace is */
+      auto bs_dist = bs_quote.backslash_index();
+      uint8_t escape_char = src[bs_dist + 1];
+      /* we encountered backslash first. Handle backslash */
+      if (escape_char == 'u') {
+        /* move src/dst up to the start; they will be further adjusted
+           within the unicode codepoint handling code. */
+        src += bs_dist;
+        dst += bs_dist;
+        if (!handle_unicode_codepoint(&src, &dst)) {
+          return nullptr;
+        }
+      } else {
+        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
+         * write bs_dist+1 characters to output
+         * note this may reach beyond the part of the buffer we've actually
+         * seen. I think this is ok */
+        uint8_t escape_result = escape_map[escape_char];
+        if (escape_result == 0u) {
+          return nullptr; /* bogus escape value is an error */
+        }
+        dst[bs_dist] = escape_result;
+        src += bs_dist + 2;
+        dst += bs_dist + 1;
+      }
+    } else {
+      /* they are the same. Since they can't co-occur, it means we
+       * encountered neither. */
+      src += backslash_and_quote::BYTES_PROCESSED;
+      dst += backslash_and_quote::BYTES_PROCESSED;
+    }
+  }
+  /* can't be reached */
+  return nullptr;
+}
+
+} // namespace stringparsing
+} // unnamed namespace
+} // namespace westmere
+} // namespace simdjson
+/* end file src/generic/stage2/stringparsing.h */
 /* begin file src/generic/stage2/tape_builder.h */
 /* begin file src/generic/stage2/json_iterator.h */
 /* begin file src/generic/stage2/logger.h */
@@ -14219,7 +15146,7 @@ namespace logger {
   static int log_depth; // Not threadsafe. Log only.
 
   // Helper to turn unprintable or newline characters into spaces
-  static simdjson_really_inline char printable_char(char c) {
+  static simdjson_inline char printable_char(char c) {
     if (c >= 0x20) {
       return c;
     } else {
@@ -14228,7 +15155,7 @@ namespace logger {
   }
 
   // Print the header and set up log_start
-  static simdjson_really_inline void log_start() {
+  static simdjson_inline void log_start() {
     if (LOG_ENABLED) {
       log_depth = 0;
       printf("\n");
@@ -14237,7 +15164,7 @@ namespace logger {
     }
   }
 
-  simdjson_unused static simdjson_really_inline void log_string(const char *message) {
+  simdjson_unused static simdjson_inline void log_string(const char *message) {
     if (LOG_ENABLED) {
       printf("%s\n", message);
     }
@@ -14245,7 +15172,7 @@ namespace logger {
 
   // Logs a single line from the stage 2 DOM parser
   template<typename S>
-  static simdjson_really_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
+  static simdjson_inline void log_line(S &structurals, const char *title_prefix, const char *title, const char *detail) {
     if (LOG_ENABLED) {
       printf("| %*s%s%-*s ", log_depth*2, "", title_prefix, LOG_EVENT_LEN - log_depth*2 - int(strlen(title_prefix)), title);
       auto current_index = structurals.at_beginning() ? nullptr : structurals.next_structural-1;
@@ -14323,14 +15250,14 @@ class json_iterator {
    * - increment_count(iter) - each time a value is found in an array or object.
    */
   template<bool STREAMING, typename V>
-  simdjson_warn_unused simdjson_really_inline error_code walk_document(V &visitor) noexcept;
+  simdjson_warn_unused simdjson_inline error_code walk_document(V &visitor) noexcept;
 
   /**
    * Create an iterator capable of walking a JSON document.
    *
    * The document must have already passed through stage 1.
    */
-  simdjson_really_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
+  simdjson_inline json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index);
 
   /**
    * Look at the next token.
@@ -14339,7 +15266,7 @@ class json_iterator {
    *
    * They may include invalid JSON as well (such as `1.2.3` or `ture`).
    */
-  simdjson_really_inline const uint8_t *peek() const noexcept;
+  simdjson_inline const uint8_t *peek() const noexcept;
   /**
    * Advance to the next token.
    *
@@ -14347,56 +15274,56 @@ class json_iterator {
    *
    * They may include invalid JSON as well (such as `1.2.3` or `ture`).
    */
-  simdjson_really_inline const uint8_t *advance() noexcept;
+  simdjson_inline const uint8_t *advance() noexcept;
   /**
    * Get the remaining length of the document, from the start of the current token.
    */
-  simdjson_really_inline size_t remaining_len() const noexcept;
+  simdjson_inline size_t remaining_len() const noexcept;
   /**
    * Check if we are at the end of the document.
    *
    * If this is true, there are no more tokens.
    */
-  simdjson_really_inline bool at_eof() const noexcept;
+  simdjson_inline bool at_eof() const noexcept;
   /**
    * Check if we are at the beginning of the document.
    */
-  simdjson_really_inline bool at_beginning() const noexcept;
-  simdjson_really_inline uint8_t last_structural() const noexcept;
+  simdjson_inline bool at_beginning() const noexcept;
+  simdjson_inline uint8_t last_structural() const noexcept;
 
   /**
    * Log that a value has been found.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_value(const char *type) const noexcept;
+  simdjson_inline void log_value(const char *type) const noexcept;
   /**
    * Log the start of a multipart value.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_start_value(const char *type) const noexcept;
+  simdjson_inline void log_start_value(const char *type) const noexcept;
   /**
    * Log the end of a multipart value.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_end_value(const char *type) const noexcept;
+  simdjson_inline void log_end_value(const char *type) const noexcept;
   /**
    * Log an error.
    *
    * Set LOG_ENABLED=true in logger.h to see logging.
    */
-  simdjson_really_inline void log_error(const char *error) const noexcept;
+  simdjson_inline void log_error(const char *error) const noexcept;
 
   template<typename V>
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(V &visitor, const uint8_t *value) noexcept;
   template<typename V>
-  simdjson_warn_unused simdjson_really_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(V &visitor, const uint8_t *value) noexcept;
 };
 
 template<bool STREAMING, typename V>
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::walk_document(V &visitor) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::walk_document(V &visitor) noexcept {
   logger::log_start();
 
   //
@@ -14521,52 +15448,52 @@ simdjson_warn_unused simdjson_really_inline error_code json_iterator::walk_docum
 
 } // walk_document()
 
-simdjson_really_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
+simdjson_inline json_iterator::json_iterator(dom_parser_implementation &_dom_parser, size_t start_structural_index)
   : buf{_dom_parser.buf},
     next_structural{&_dom_parser.structural_indexes[start_structural_index]},
     dom_parser{_dom_parser} {
 }
 
-simdjson_really_inline const uint8_t *json_iterator::peek() const noexcept {
+simdjson_inline const uint8_t *json_iterator::peek() const noexcept {
   return &buf[*(next_structural)];
 }
-simdjson_really_inline const uint8_t *json_iterator::advance() noexcept {
+simdjson_inline const uint8_t *json_iterator::advance() noexcept {
   return &buf[*(next_structural++)];
 }
-simdjson_really_inline size_t json_iterator::remaining_len() const noexcept {
+simdjson_inline size_t json_iterator::remaining_len() const noexcept {
   return dom_parser.len - *(next_structural-1);
 }
 
-simdjson_really_inline bool json_iterator::at_eof() const noexcept {
+simdjson_inline bool json_iterator::at_eof() const noexcept {
   return next_structural == &dom_parser.structural_indexes[dom_parser.n_structural_indexes];
 }
-simdjson_really_inline bool json_iterator::at_beginning() const noexcept {
+simdjson_inline bool json_iterator::at_beginning() const noexcept {
   return next_structural == dom_parser.structural_indexes.get();
 }
-simdjson_really_inline uint8_t json_iterator::last_structural() const noexcept {
+simdjson_inline uint8_t json_iterator::last_structural() const noexcept {
   return buf[dom_parser.structural_indexes[dom_parser.n_structural_indexes - 1]];
 }
 
-simdjson_really_inline void json_iterator::log_value(const char *type) const noexcept {
+simdjson_inline void json_iterator::log_value(const char *type) const noexcept {
   logger::log_line(*this, "", type, "");
 }
 
-simdjson_really_inline void json_iterator::log_start_value(const char *type) const noexcept {
+simdjson_inline void json_iterator::log_start_value(const char *type) const noexcept {
   logger::log_line(*this, "+", type, "");
   if (logger::LOG_ENABLED) { logger::log_depth++; }
 }
 
-simdjson_really_inline void json_iterator::log_end_value(const char *type) const noexcept {
+simdjson_inline void json_iterator::log_end_value(const char *type) const noexcept {
   if (logger::LOG_ENABLED) { logger::log_depth--; }
   logger::log_line(*this, "-", type, "");
 }
 
-simdjson_really_inline void json_iterator::log_error(const char *error) const noexcept {
+simdjson_inline void json_iterator::log_error(const char *error) const noexcept {
   logger::log_line(*this, "", "ERROR", error);
 }
 
 template<typename V>
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_root_primitive(V &visitor, const uint8_t *value) noexcept {
   switch (*value) {
     case '"': return visitor.visit_root_string(*this, value);
     case 't': return visitor.visit_root_true_atom(*this, value);
@@ -14582,7 +15509,7 @@ simdjson_warn_unused simdjson_really_inline error_code json_iterator::visit_root
   }
 }
 template<typename V>
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::visit_primitive(V &visitor, const uint8_t *value) noexcept {
   switch (*value) {
     case '"': return visitor.visit_string(*this, value);
     case 't': return visitor.visit_true_atom(*this, value);
@@ -14614,18 +15541,18 @@ struct tape_writer {
   uint64_t *next_tape_loc;
 
   /** Write a signed 64-bit value to tape. */
-  simdjson_really_inline void append_s64(int64_t value) noexcept;
+  simdjson_inline void append_s64(int64_t value) noexcept;
 
   /** Write an unsigned 64-bit value to tape. */
-  simdjson_really_inline void append_u64(uint64_t value) noexcept;
+  simdjson_inline void append_u64(uint64_t value) noexcept;
 
   /** Write a double value to tape. */
-  simdjson_really_inline void append_double(double value) noexcept;
+  simdjson_inline void append_double(double value) noexcept;
 
   /**
    * Append a tape entry (an 8-bit type,and 56 bits worth of value).
    */
-  simdjson_really_inline void append(uint64_t val, internal::tape_type t) noexcept;
+  simdjson_inline void append(uint64_t val, internal::tape_type t) noexcept;
 
   /**
    * Skip the current tape entry without writing.
@@ -14633,24 +15560,24 @@ struct tape_writer {
    * Used to skip the start of the container, since we'll come back later to fill it in when the
    * container ends.
    */
-  simdjson_really_inline void skip() noexcept;
+  simdjson_inline void skip() noexcept;
 
   /**
    * Skip the number of tape entries necessary to write a large u64 or i64.
    */
-  simdjson_really_inline void skip_large_integer() noexcept;
+  simdjson_inline void skip_large_integer() noexcept;
 
   /**
    * Skip the number of tape entries necessary to write a double.
    */
-  simdjson_really_inline void skip_double() noexcept;
+  simdjson_inline void skip_double() noexcept;
 
   /**
    * Write a value to a known location on tape.
    *
    * Used to go back and write out the start of a container after the container ends.
    */
-  simdjson_really_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
+  simdjson_inline static void write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept;
 
 private:
   /**
@@ -14658,50 +15585,50 @@ struct tape_writer {
    * all 64 bits, such as double and uint64_t.
    */
   template<typename T>
-  simdjson_really_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
+  simdjson_inline void append2(uint64_t val, T val2, internal::tape_type t) noexcept;
 }; // struct number_writer
 
-simdjson_really_inline void tape_writer::append_s64(int64_t value) noexcept {
+simdjson_inline void tape_writer::append_s64(int64_t value) noexcept {
   append2(0, value, internal::tape_type::INT64);
 }
 
-simdjson_really_inline void tape_writer::append_u64(uint64_t value) noexcept {
+simdjson_inline void tape_writer::append_u64(uint64_t value) noexcept {
   append(0, internal::tape_type::UINT64);
   *next_tape_loc = value;
   next_tape_loc++;
 }
 
 /** Write a double value to tape. */
-simdjson_really_inline void tape_writer::append_double(double value) noexcept {
+simdjson_inline void tape_writer::append_double(double value) noexcept {
   append2(0, value, internal::tape_type::DOUBLE);
 }
 
-simdjson_really_inline void tape_writer::skip() noexcept {
+simdjson_inline void tape_writer::skip() noexcept {
   next_tape_loc++;
 }
 
-simdjson_really_inline void tape_writer::skip_large_integer() noexcept {
+simdjson_inline void tape_writer::skip_large_integer() noexcept {
   next_tape_loc += 2;
 }
 
-simdjson_really_inline void tape_writer::skip_double() noexcept {
+simdjson_inline void tape_writer::skip_double() noexcept {
   next_tape_loc += 2;
 }
 
-simdjson_really_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
+simdjson_inline void tape_writer::append(uint64_t val, internal::tape_type t) noexcept {
   *next_tape_loc = val | ((uint64_t(char(t))) << 56);
   next_tape_loc++;
 }
 
 template<typename T>
-simdjson_really_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
+simdjson_inline void tape_writer::append2(uint64_t val, T val2, internal::tape_type t) noexcept {
   append(val, t);
   static_assert(sizeof(val2) == sizeof(*next_tape_loc), "Type is not 64 bits!");
   memcpy(next_tape_loc, &val2, sizeof(val2));
   next_tape_loc++;
 }
 
-simdjson_really_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
+simdjson_inline void tape_writer::write(uint64_t &tape_loc, uint64_t val, internal::tape_type t) noexcept {
   tape_loc = val | ((uint64_t(char(t))) << 56);
 }
 
@@ -14718,40 +15645,40 @@ namespace stage2 {
 
 struct tape_builder {
   template<bool STREAMING>
-  simdjson_warn_unused static simdjson_really_inline error_code parse_document(
+  simdjson_warn_unused static simdjson_inline error_code parse_document(
     dom_parser_implementation &dom_parser,
     dom::document &doc) noexcept;
 
   /** Called when a non-empty document starts. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_document_start(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_document_start(json_iterator &iter) noexcept;
   /** Called when a non-empty document ends without error. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_document_end(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_document_end(json_iterator &iter) noexcept;
 
   /** Called when a non-empty array starts. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_array_start(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_array_start(json_iterator &iter) noexcept;
   /** Called when a non-empty array ends. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_array_end(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_array_end(json_iterator &iter) noexcept;
   /** Called when an empty array is found. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_empty_array(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_empty_array(json_iterator &iter) noexcept;
 
   /** Called when a non-empty object starts. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_object_start(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_object_start(json_iterator &iter) noexcept;
   /**
    * Called when a key in a field is encountered.
    *
    * primitive, visit_object_start, visit_empty_object, visit_array_start, or visit_empty_array
    * will be called after this with the field value.
    */
-  simdjson_warn_unused simdjson_really_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_key(json_iterator &iter, const uint8_t *key) noexcept;
   /** Called when a non-empty object ends. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_object_end(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_object_end(json_iterator &iter) noexcept;
   /** Called when an empty object is found. */
-  simdjson_warn_unused simdjson_really_inline error_code visit_empty_object(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_empty_object(json_iterator &iter) noexcept;
 
   /**
    * Called when a string, number, boolean or null is found.
    */
-  simdjson_warn_unused simdjson_really_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_primitive(json_iterator &iter, const uint8_t *value) noexcept;
   /**
    * Called when a string, number, boolean or null is found at the top level of a document (i.e.
    * when there is no array or object and the entire document is a single string, number, boolean or
@@ -14760,22 +15687,22 @@ struct tape_builder {
    * This is separate from primitive() because simdjson's normal primitive parsing routines assume
    * there is at least one more token after the value, which is only true in an array or object.
    */
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept;
 
-  simdjson_warn_unused simdjson_really_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_string(json_iterator &iter, const uint8_t *value, bool key = false) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
 
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_string(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_number(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept;
+  simdjson_warn_unused simdjson_inline error_code visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept;
 
   /** Called each time a new field or element in an array or object is found. */
-  simdjson_warn_unused simdjson_really_inline error_code increment_count(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code increment_count(json_iterator &iter) noexcept;
 
   /** Next location to write to tape */
   tape_writer tape;
@@ -14783,18 +15710,18 @@ struct tape_builder {
   /** Next write location in the string buf for stage 2 parsing */
   uint8_t *current_string_buf_loc;
 
-  simdjson_really_inline tape_builder(dom::document &doc) noexcept;
+  simdjson_inline tape_builder(dom::document &doc) noexcept;
 
-  simdjson_really_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
-  simdjson_really_inline void start_container(json_iterator &iter) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
-  simdjson_warn_unused simdjson_really_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
-  simdjson_really_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
-  simdjson_really_inline void on_end_string(uint8_t *dst) noexcept;
+  simdjson_inline uint32_t next_tape_index(json_iterator &iter) const noexcept;
+  simdjson_inline void start_container(json_iterator &iter) noexcept;
+  simdjson_warn_unused simdjson_inline error_code end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_warn_unused simdjson_inline error_code empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept;
+  simdjson_inline uint8_t *on_start_string(json_iterator &iter) noexcept;
+  simdjson_inline void on_end_string(uint8_t *dst) noexcept;
 }; // class tape_builder
 
 template<bool STREAMING>
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::parse_document(
+simdjson_warn_unused simdjson_inline error_code tape_builder::parse_document(
     dom_parser_implementation &dom_parser,
     dom::document &doc) noexcept {
   dom_parser.doc = &doc;
@@ -14803,56 +15730,56 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::parse_docum
   return iter.walk_document<STREAMING>(builder);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_primitive(json_iterator &iter, const uint8_t *value) noexcept {
   return iter.visit_root_primitive(*this, value);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_primitive(json_iterator &iter, const uint8_t *value) noexcept {
   return iter.visit_primitive(*this, value);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_object(json_iterator &iter) noexcept {
   return empty_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_empty_array(json_iterator &iter) noexcept {
   return empty_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_start(json_iterator &iter) noexcept {
   start_container(iter);
   return SUCCESS;
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_start(json_iterator &iter) noexcept {
   start_container(iter);
   return SUCCESS;
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_start(json_iterator &iter) noexcept {
   start_container(iter);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_object_end(json_iterator &iter) noexcept {
   return end_container(iter, internal::tape_type::START_OBJECT, internal::tape_type::END_OBJECT);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_array_end(json_iterator &iter) noexcept {
   return end_container(iter, internal::tape_type::START_ARRAY, internal::tape_type::END_ARRAY);
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_document_end(json_iterator &iter) noexcept {
   constexpr uint32_t start_tape_index = 0;
   tape.append(start_tape_index, internal::tape_type::ROOT);
   tape_writer::write(iter.dom_parser.doc->tape[start_tape_index], next_tape_index(iter), internal::tape_type::ROOT);
   return SUCCESS;
 }
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_key(json_iterator &iter, const uint8_t *key) noexcept {
   return visit_string(iter, key, true);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::increment_count(json_iterator &iter) noexcept {
   iter.dom_parser.open_containers[iter.depth].count++; // we have a key value pair in the object at parser.dom_parser.depth - 1
   return SUCCESS;
 }
 
-simdjson_really_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
+simdjson_inline tape_builder::tape_builder(dom::document &doc) noexcept : tape{doc.tape.get()}, current_string_buf_loc{doc.string_buf.get()} {}
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_string(json_iterator &iter, const uint8_t *value, bool key) noexcept {
   iter.log_value(key ? "key" : "string");
   uint8_t *dst = on_start_string(iter);
   dst = stringparsing::parse_string(value+1, dst);
@@ -14864,16 +15791,16 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_strin
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_string(json_iterator &iter, const uint8_t *value) noexcept {
   return visit_string(iter, value);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_number(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("number");
   return numberparsing::parse_number(value, tape);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_number(json_iterator &iter, const uint8_t *value) noexcept {
   //
   // We need to make a copy to make sure that the string is space terminated.
   // This is not about padding the input, which should already padded up
@@ -14895,42 +15822,42 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_
   return error;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("true");
   if (!atomparsing::is_valid_true_atom(value)) { return T_ATOM_ERROR; }
   tape.append(0, internal::tape_type::TRUE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_true_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("true");
   if (!atomparsing::is_valid_true_atom(value, iter.remaining_len())) { return T_ATOM_ERROR; }
   tape.append(0, internal::tape_type::TRUE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("false");
   if (!atomparsing::is_valid_false_atom(value)) { return F_ATOM_ERROR; }
   tape.append(0, internal::tape_type::FALSE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_false_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("false");
   if (!atomparsing::is_valid_false_atom(value, iter.remaining_len())) { return F_ATOM_ERROR; }
   tape.append(0, internal::tape_type::FALSE_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("null");
   if (!atomparsing::is_valid_null_atom(value)) { return N_ATOM_ERROR; }
   tape.append(0, internal::tape_type::NULL_VALUE);
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::visit_root_null_atom(json_iterator &iter, const uint8_t *value) noexcept {
   iter.log_value("null");
   if (!atomparsing::is_valid_null_atom(value, iter.remaining_len())) { return N_ATOM_ERROR; }
   tape.append(0, internal::tape_type::NULL_VALUE);
@@ -14939,24 +15866,24 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::visit_root_
 
 // private:
 
-simdjson_really_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
+simdjson_inline uint32_t tape_builder::next_tape_index(json_iterator &iter) const noexcept {
   return uint32_t(tape.next_tape_loc - iter.dom_parser.doc->tape.get());
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::empty_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
   auto start_index = next_tape_index(iter);
   tape.append(start_index+2, start);
   tape.append(start_index, end);
   return SUCCESS;
 }
 
-simdjson_really_inline void tape_builder::start_container(json_iterator &iter) noexcept {
+simdjson_inline void tape_builder::start_container(json_iterator &iter) noexcept {
   iter.dom_parser.open_containers[iter.depth].tape_index = next_tape_index(iter);
   iter.dom_parser.open_containers[iter.depth].count = 0;
   tape.skip(); // We don't actually *write* the start element until the end.
 }
 
-simdjson_warn_unused simdjson_really_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
+simdjson_warn_unused simdjson_inline error_code tape_builder::end_container(json_iterator &iter, internal::tape_type start, internal::tape_type end) noexcept {
   // Write the ending tape element, pointing at the start location
   const uint32_t start_tape_index = iter.dom_parser.open_containers[iter.depth].tape_index;
   tape.append(start_tape_index, end);
@@ -14969,13 +15896,13 @@ simdjson_warn_unused simdjson_really_inline error_code tape_builder::end_contain
   return SUCCESS;
 }
 
-simdjson_really_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
+simdjson_inline uint8_t *tape_builder::on_start_string(json_iterator &iter) noexcept {
   // we advance the point, accounting for the fact that we have a NULL termination
   tape.append(current_string_buf_loc - iter.dom_parser.doc->string_buf.get(), internal::tape_type::STRING);
   return current_string_buf_loc + sizeof(uint32_t);
 }
 
-simdjson_really_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
+simdjson_inline void tape_builder::on_end_string(uint8_t *dst) noexcept {
   uint32_t str_length = uint32_t(dst - (current_string_buf_loc + sizeof(uint32_t)));
   // TODO check for overflow in case someone has a crazy string (>=4GB?)
   // But only add the overflow check when the document itself exceeds 4GB
@@ -15002,7 +15929,7 @@ namespace westmere {
 namespace {
 namespace stage1 {
 
-simdjson_really_inline uint64_t json_string_scanner::find_escaped(uint64_t backslash) {
+simdjson_inline uint64_t json_string_scanner::find_escaped(uint64_t backslash) {
   if (!backslash) { uint64_t escaped = prev_escaped; prev_escaped = 0; return escaped; }
   return find_escaped_branchless(backslash);
 }
@@ -15032,6 +15959,10 @@ simdjson_warn_unused error_code dom_parser_implementation::stage2_next(dom::docu
   return stage2::tape_builder::parse_document<true>(*this, _doc);
 }
 
+simdjson_warn_unused uint8_t *dom_parser_implementation::parse_string(const uint8_t *src, uint8_t *dst) const noexcept {
+  return westmere::stringparsing::parse_string(src, dst);
+}
+
 simdjson_warn_unused error_code dom_parser_implementation::parse(const uint8_t *_buf, size_t _len, dom::document &_doc) noexcept {
   auto error = stage1(_buf, _len, stage1_mode::regular);
   if (error) { return error; }
diff --git a/simdjson/simdjson.h b/simdjson/simdjson.h
index 67cb53b..f9d8081 100644
--- a/simdjson/simdjson.h
+++ b/simdjson/simdjson.h
@@ -1,4 +1,4 @@
-/* auto-generated on 2022-06-02 13:56:20 -0400. Do not edit! */
+/* auto-generated on 2022-11-23 10:31:42 -0500. Do not edit! */
 /* begin file include/simdjson.h */
 #ifndef SIMDJSON_H
 #define SIMDJSON_H
@@ -43,14 +43,14 @@
 #define SIMDJSON_SIMDJSON_VERSION_H
 
 /** The version of simdjson being used (major.minor.revision) */
-#define SIMDJSON_VERSION 2.0.3
+#define SIMDJSON_VERSION 3.0.1
 
 namespace simdjson {
 enum {
   /**
    * The major version (MAJOR.minor.revision) of simdjson being used.
    */
-  SIMDJSON_VERSION_MAJOR = 2,
+  SIMDJSON_VERSION_MAJOR = 3,
   /**
    * The minor version (major.MINOR.revision) of simdjson being used.
    */
@@ -58,7 +58,7 @@ enum {
   /**
    * The revision (major.minor.REVISION) of simdjson being used.
    */
-  SIMDJSON_VERSION_REVISION = 3
+  SIMDJSON_VERSION_REVISION = 1
 };
 } // namespace simdjson
 
@@ -329,12 +329,12 @@ constexpr size_t SIMDJSON_MAXSIZE_BYTES = 0xFFFFFFFF;
 /**
  * The amount of padding needed in a buffer to parse JSON.
  *
- * the input buf should be readable up to buf + SIMDJSON_PADDING
+ * The input buf should be readable up to buf + SIMDJSON_PADDING
  * this is a stopgap; there should be a better description of the
  * main loop and its behavior that abstracts over this
  * See https://github.com/simdjson/simdjson/issues/174
  */
-constexpr size_t SIMDJSON_PADDING = 32;
+constexpr size_t SIMDJSON_PADDING = 64;
 
 /**
  * By default, simdjson supports this many nested objects and arrays.
@@ -460,6 +460,18 @@ constexpr size_t DEFAULT_MAX_DEPTH = 1024;
 
 #endif // MSC_VER
 
+#if defined(simdjson_inline)
+  // Prefer the user's definition of simdjson_inline; don't define it ourselves.
+#elif defined(__GNUC__) && !defined(__OPTIMIZE__)
+  // If optimizations are disabled, forcing inlining can lead to significant
+  // code bloat and high compile times. Don't use simdjson_really_inline for
+  // unoptimized builds.
+  #define simdjson_inline inline
+#else
+  // Force inlining for most simdjson functions.
+  #define simdjson_inline simdjson_really_inline
+#endif
+
 #if defined(SIMDJSON_VISUAL_STUDIO)
     /**
      * Windows users need to do some extra work when building
@@ -2197,11 +2209,30 @@ namespace std {
 /// If EXPR is an error, returns it.
 #define SIMDJSON_TRY(EXPR) { auto _err = (EXPR); if (_err) { return _err; } }
 
+// Unless the programmer has already set SIMDJSON_DEVELOPMENT_CHECKS,
+// we want to set it under debug builds. We detect a debug build
+// under Visual Studio when the _DEBUG macro is set. Under the other
+// compilers, we use the fact that they define __OPTIMIZE__ whenever
+// they allow optimizations.
+// It is possible that this could miss some cases where SIMDJSON_DEVELOPMENT_CHECKS
+// is helpful, but the programmer can set the macro SIMDJSON_DEVELOPMENT_CHECKS.
+// It could also wrongly set SIMDJSON_DEVELOPMENT_CHECKS (e.g., if the programmer
+// sets _DEBUG in a release build under Visual Studio, or if some compiler fails to
+// set the __OPTIMIZE__ macro).
 #ifndef SIMDJSON_DEVELOPMENT_CHECKS
-#ifndef NDEBUG
-#define SIMDJSON_DEVELOPMENT_CHECKS
-#endif
-#endif
+#ifdef _MSC_VER
+// Visual Studio seems to set _DEBUG for debug builds.
+#ifdef _DEBUG
+#define SIMDJSON_DEVELOPMENT_CHECKS 1
+#endif // _DEBUG
+#else // _MSC_VER
+// All other compilers appear to set __OPTIMIZE__ to a positive integer
+// when the compiler is optimizing.
+#ifndef __OPTIMIZE__
+#define SIMDJSON_DEVELOPMENT_CHECKS 1
+#endif // __OPTIMIZE__
+#endif // _MSC_VER
+#endif // SIMDJSON_DEVELOPMENT_CHECKS
 
 // The SIMDJSON_CHECK_EOF macro is a feature flag for the "don't require padding"
 // feature.
@@ -2222,6 +2253,13 @@ namespace std {
 # define simdjson_fallthrough do {} while (0)  /* fallthrough */
 #endif // simdjson_fallthrough
 
+
+#if SIMDJSON_DEVELOPMENT_CHECKS
+#define SIMDJSON_DEVELOPMENT_ASSERT(expr) do { assert ((expr)); } while (0)
+#else
+#define SIMDJSON_DEVELOPMENT_ASSERT(expr) do { } while (0)
+#endif
+
 #endif // SIMDJSON_COMMON_DEFS_H
 /* end file include/simdjson/common_defs.h */
 
@@ -2249,7 +2287,7 @@ enum error_code {
   SUCCESS = 0,                ///< No error
   CAPACITY,                   ///< This parser can't support a document that big
   MEMALLOC,                   ///< Error allocating memory, most likely out of memory
-  TAPE_ERROR,                 ///< Something went wrong while writing to the tape (stage 2), this is a generic error
+  TAPE_ERROR,                 ///< Something went wrong, this is a generic error
   DEPTH_ERROR,                ///< Your document exceeds the user-specified depth limitation
   STRING_ERROR,               ///< Problem while parsing a string
   T_ATOM_ERROR,               ///< Problem while parsing an atom starting with the letter 't'
@@ -2276,6 +2314,7 @@ enum error_code {
   INCOMPLETE_ARRAY_OR_OBJECT, ///< The document ends early.
   SCALAR_DOCUMENT_AS_VALUE,   ///< A scalar document is treated as a value.
   OUT_OF_BOUNDS,              ///< Attempted to access location outside of document.
+  TRAILING_CONTENT,           ///< Unexpected trailing content in the JSON input
   NUM_ERROR_CODES
 };
 
@@ -2342,22 +2381,22 @@ struct simdjson_result_base : protected std::pair<T, error_code> {
   /**
    * Create a new empty result with error = UNINITIALIZED.
    */
-  simdjson_really_inline simdjson_result_base() noexcept;
+  simdjson_inline simdjson_result_base() noexcept;
 
   /**
    * Create a new error result.
    */
-  simdjson_really_inline simdjson_result_base(error_code error) noexcept;
+  simdjson_inline simdjson_result_base(error_code error) noexcept;
 
   /**
    * Create a new successful result.
    */
-  simdjson_really_inline simdjson_result_base(T &&value) noexcept;
+  simdjson_inline simdjson_result_base(T &&value) noexcept;
 
   /**
    * Create a new result with both things (use if you don't want to branch when creating the result).
    */
-  simdjson_really_inline simdjson_result_base(T &&value, error_code error) noexcept;
+  simdjson_inline simdjson_result_base(T &&value, error_code error) noexcept;
 
   /**
    * Move the value and the error to the provided variables.
@@ -2365,26 +2404,19 @@ struct simdjson_result_base : protected std::pair<T, error_code> {
    * @param value The variable to assign the value to. May not be set if there is an error.
    * @param error The variable to assign the error to. Set to SUCCESS if there is no error.
    */
-  simdjson_really_inline void tie(T &value, error_code &error) && noexcept;
+  simdjson_inline void tie(T &value, error_code &error) && noexcept;
 
   /**
    * Move the value to the provided variable.
    *
    * @param value The variable to assign the value to. May not be set if there is an error.
    */
-  simdjson_really_inline error_code get(T &value) && noexcept;
-
-  /**
-   * Move the value to the provided variable.
-   *
-   * @param value The variable to assign the value to. May not be set if there is an error.
-   */
-  simdjson_really_inline const T &value(error_code &error) const & noexcept;
+  simdjson_inline error_code get(T &value) && noexcept;
 
   /**
    * The error.
    */
-  simdjson_really_inline error_code error() const noexcept;
+  simdjson_inline error_code error() const noexcept;
 
 #if SIMDJSON_EXCEPTIONS
 
@@ -2393,41 +2425,41 @@ struct simdjson_result_base : protected std::pair<T, error_code> {
    *
    * @throw simdjson_error if there was an error.
    */
-  simdjson_really_inline T& value() & noexcept(false);
+  simdjson_inline T& value() & noexcept(false);
 
   /**
    * Take the result value (move it).
    *
    * @throw simdjson_error if there was an error.
    */
-  simdjson_really_inline T&& value() && noexcept(false);
+  simdjson_inline T&& value() && noexcept(false);
 
   /**
    * Take the result value (move it).
    *
    * @throw simdjson_error if there was an error.
    */
-  simdjson_really_inline T&& take_value() && noexcept(false);
+  simdjson_inline T&& take_value() && noexcept(false);
 
   /**
    * Cast to the value (will throw on error).
    *
    * @throw simdjson_error if there was an error.
    */
-  simdjson_really_inline operator T&&() && noexcept(false);
+  simdjson_inline operator T&&() && noexcept(false);
 #endif // SIMDJSON_EXCEPTIONS
 
   /**
    * Get the result value. This function is safe if and only
    * the error() method returns a value that evaluates to false.
    */
-  simdjson_really_inline const T& value_unsafe() const& noexcept;
+  simdjson_inline const T& value_unsafe() const& noexcept;
 
   /**
    * Take the result value (move it). This function is safe if and only
    * the error() method returns a value that evaluates to false.
    */
-  simdjson_really_inline T&& value_unsafe() && noexcept;
+  simdjson_inline T&& value_unsafe() && noexcept;
 
 }; // struct simdjson_result_base
 
@@ -2443,19 +2475,19 @@ struct simdjson_result : public internal::simdjson_result_base<T> {
   /**
    * @private Create a new empty result with error = UNINITIALIZED.
    */
-  simdjson_really_inline simdjson_result() noexcept;
+  simdjson_inline simdjson_result() noexcept;
   /**
    * @private Create a new error result.
    */
-  simdjson_really_inline simdjson_result(T &&value) noexcept;
+  simdjson_inline simdjson_result(T &&value) noexcept;
   /**
    * @private Create a new successful result.
    */
-  simdjson_really_inline simdjson_result(error_code error_code) noexcept;
+  simdjson_inline simdjson_result(error_code error_code) noexcept;
   /**
    * @private Create a new result with both things (use if you don't want to branch when creating the result).
    */
-  simdjson_really_inline simdjson_result(T &&value, error_code error) noexcept;
+  simdjson_inline simdjson_result(T &&value, error_code error) noexcept;
 
   /**
    * Move the value and the error to the provided variables.
@@ -2463,19 +2495,19 @@ struct simdjson_result : public internal::simdjson_result_base<T> {
    * @param value The variable to assign the value to. May not be set if there is an error.
    * @param error The variable to assign the error to. Set to SUCCESS if there is no error.
    */
-  simdjson_really_inline void tie(T &value, error_code &error) && noexcept;
+  simdjson_inline void tie(T &value, error_code &error) && noexcept;
 
   /**
    * Move the value to the provided variable.
    *
    * @param value The variable to assign the value to. May not be set if there is an error.
    */
-  simdjson_warn_unused simdjson_really_inline error_code get(T &value) && noexcept;
+  simdjson_warn_unused simdjson_inline error_code get(T &value) && noexcept;
 
   /**
    * The error.
    */
-  simdjson_really_inline error_code error() const noexcept;
+  simdjson_inline error_code error() const noexcept;
 
 #if SIMDJSON_EXCEPTIONS
 
@@ -2484,41 +2516,41 @@ struct simdjson_result : public internal::simdjson_result_base<T> {
    *
    * @throw simdjson_error if there was an error.
    */
-  simdjson_really_inline T& value() & noexcept(false);
+  simdjson_inline T& value() & noexcept(false);
 
   /**
    * Take the result value (move it).
    *
    * @throw simdjson_error if there was an error.
    */
-  simdjson_really_inline T&& value() && noexcept(false);
+  simdjson_inline T&& value() && noexcept(false);
 
   /**
    * Take the result value (move it).
    *
    * @throw simdjson_error if there was an error.
    */
-  simdjson_really_inline T&& take_value() && noexcept(false);
+  simdjson_inline T&& take_value() && noexcept(false);
 
   /**
    * Cast to the value (will throw on error).
    *
    * @throw simdjson_error if there was an error.
    */
-  simdjson_really_inline operator T&&() && noexcept(false);
+  simdjson_inline operator T&&() && noexcept(false);
 #endif // SIMDJSON_EXCEPTIONS
 
   /**
    * Get the result value. This function is safe if and only
    * the error() method returns a value that evaluates to false.
    */
-  simdjson_really_inline const T& value_unsafe() const& noexcept;
+  simdjson_inline const T& value_unsafe() const& noexcept;
 
   /**
    * Take the result value (move it). This function is safe if and only
    * the error() method returns a value that evaluates to false.
    */
-  simdjson_really_inline T&& value_unsafe() && noexcept;
+  simdjson_inline T&& value_unsafe() && noexcept;
 
 }; // struct simdjson_result
 
@@ -2926,6 +2958,22 @@ class dom_parser_implementation {
    */
   simdjson_warn_unused virtual error_code stage2_next(dom::document &doc) noexcept = 0;
 
+  /**
+   * Unescape a valid UTF-8 string from src to dst, stopping at a final unescaped quote. There
+   * must be an unescaped quote terminating the string. It returns the final output
+   * position as pointer. In case of error (e.g., the string has bad escaped codes),
+   * then null_nullptrptr is returned. It is assumed that the output buffer is large
+   * enough. E.g., if src points at 'joe"', then dst needs to have four free bytes +
+   * SIMDJSON_PADDING bytes.
+   *
+   * Overridden by each implementation.
+   *
+   * @param str pointer to the beginning of a valid UTF-8 JSON string, must end with an unescaped quote.
+   * @param dst pointer to a destination buffer, it must point a region in memory of sufficient size.
+   * @return end of the of the written region (exclusive) or nullptr in case of error.
+   */
+  simdjson_warn_unused virtual uint8_t *parse_string(const uint8_t *src, uint8_t *dst) const noexcept = 0;
+
   /**
    * Change the capacity of this parser.
    *
@@ -2968,14 +3016,14 @@ class dom_parser_implementation {
    *
    * @return Current capacity, in bytes.
    */
-  simdjson_really_inline size_t capacity() const noexcept;
+  simdjson_inline size_t capacity() const noexcept;
 
   /**
    * The maximum level of nested object and arrays supported by this parser.
    *
    * @return Maximum depth, in bytes.
    */
-  simdjson_really_inline size_t max_depth() const noexcept;
+  simdjson_inline size_t max_depth() const noexcept;
 
   /**
    * Ensure this parser has enough memory to process JSON documents up to `capacity` bytes in length
@@ -2987,6 +3035,7 @@ class dom_parser_implementation {
    */
   simdjson_warn_unused inline error_code allocate(size_t capacity, size_t max_depth) noexcept;
 
+
 protected:
   /**
    * The maximum document length this parser supports.
@@ -3003,23 +3052,23 @@ class dom_parser_implementation {
   size_t _max_depth{0};
 
   // Declaring these so that subclasses can use them to implement their constructors.
-  simdjson_really_inline dom_parser_implementation() noexcept;
-  simdjson_really_inline dom_parser_implementation(dom_parser_implementation &&other) noexcept;
-  simdjson_really_inline dom_parser_implementation &operator=(dom_parser_implementation &&other) noexcept;
+  simdjson_inline dom_parser_implementation() noexcept;
+  simdjson_inline dom_parser_implementation(dom_parser_implementation &&other) noexcept;
+  simdjson_inline dom_parser_implementation &operator=(dom_parser_implementation &&other) noexcept;
 
-  simdjson_really_inline dom_parser_implementation(const dom_parser_implementation &) noexcept = delete;
-  simdjson_really_inline dom_parser_implementation &operator=(const dom_parser_implementation &other) noexcept = delete;
+  simdjson_inline dom_parser_implementation(const dom_parser_implementation &) noexcept = delete;
+  simdjson_inline dom_parser_implementation &operator=(const dom_parser_implementation &other) noexcept = delete;
 }; // class dom_parser_implementation
 
-simdjson_really_inline dom_parser_implementation::dom_parser_implementation() noexcept = default;
-simdjson_really_inline dom_parser_implementation::dom_parser_implementation(dom_parser_implementation &&other) noexcept = default;
-simdjson_really_inline dom_parser_implementation &dom_parser_implementation::operator=(dom_parser_implementation &&other) noexcept = default;
+simdjson_inline dom_parser_implementation::dom_parser_implementation() noexcept = default;
+simdjson_inline dom_parser_implementation::dom_parser_implementation(dom_parser_implementation &&other) noexcept = default;
+simdjson_inline dom_parser_implementation &dom_parser_implementation::operator=(dom_parser_implementation &&other) noexcept = default;
 
-simdjson_really_inline size_t dom_parser_implementation::capacity() const noexcept {
+simdjson_inline size_t dom_parser_implementation::capacity() const noexcept {
   return _capacity;
 }
 
-simdjson_really_inline size_t dom_parser_implementation::max_depth() const noexcept {
+simdjson_inline size_t dom_parser_implementation::max_depth() const noexcept {
   return _max_depth;
 }
 
@@ -3159,7 +3208,7 @@ constexpr uint32_t cpuid_avx512pf_bit = 1 << 26;    ///< @private bit 26 of EBX
 constexpr uint32_t cpuid_avx512er_bit = 1 << 27;    ///< @private bit 27 of EBX for EAX=0x7
 constexpr uint32_t cpuid_avx512cd_bit = 1 << 28;    ///< @private bit 28 of EBX for EAX=0x7
 constexpr uint32_t cpuid_avx512bw_bit = 1 << 30;    ///< @private bit 30 of EBX for EAX=0x7
-constexpr uint32_t cpuid_avx512vl_bit = 1 << 31;    ///< @private bit 31 of EBX for EAX=0x7
+constexpr uint32_t cpuid_avx512vl_bit = 1U << 31;    ///< @private bit 31 of EBX for EAX=0x7
 constexpr uint32_t cpuid_avx512vbmi2_bit = 1 << 6;  ///< @private bit 6 of ECX for EAX=0x7
 constexpr uint32_t cpuid_sse42_bit = 1 << 20;       ///< @private bit 20 of ECX for EAX=0x1
 constexpr uint32_t cpuid_pclmulqdq_bit = 1 << 1;    ///< @private bit  1 of ECX for EAX=0x1
@@ -3287,15 +3336,13 @@ namespace simdjson {
  * @return true if the string is valid UTF-8.
  */
 simdjson_warn_unused bool validate_utf8(const char * buf, size_t len) noexcept;
-
-
 /**
  * Validate the UTF-8 string.
  *
  * @param sv the string_view to validate.
  * @return true if the string is valid UTF-8.
  */
-simdjson_really_inline simdjson_warn_unused bool validate_utf8(const std::string_view sv) noexcept {
+simdjson_inline simdjson_warn_unused bool validate_utf8(const std::string_view sv) noexcept {
   return validate_utf8(sv.data(), sv.size());
 }
 
@@ -3305,7 +3352,7 @@ simdjson_really_inline simdjson_warn_unused bool validate_utf8(const std::string
  * @param p the string to validate.
  * @return true if the string is valid UTF-8.
  */
-simdjson_really_inline simdjson_warn_unused bool validate_utf8(const std::string& s) noexcept {
+simdjson_inline simdjson_warn_unused bool validate_utf8(const std::string& s) noexcept {
   return validate_utf8(s.data(), s.size());
 }
 
@@ -3328,7 +3375,7 @@ class implementation {
    *     const implementation *impl = simdjson::get_active_implementation();
    *     cout << "simdjson is optimized for " << impl->name() << "(" << impl->description() << ")" << endl;
    *
-   * @return the name of the implementation, e.g. "haswell", "westmere", "arm64"
+   * @return the name of the implementation, e.g. "haswell", "westmere", "arm64".
    */
   virtual const std::string &name() const { return _name; }
 
@@ -3338,7 +3385,7 @@ class implementation {
    *     const implementation *impl = simdjson::get_active_implementation();
    *     cout << "simdjson is optimized for " << impl->name() << "(" << impl->description() << ")" << endl;
    *
-   * @return the name of the implementation, e.g. "haswell", "westmere", "arm64"
+   * @return the description of the implementation, e.g. "Intel/AMD AVX2", "Intel/AMD SSE4.2", "ARM NEON".
    */
   virtual const std::string &description() const { return _description; }
 
@@ -3347,8 +3394,7 @@ class implementation {
    * and the current CPU match. This function may poll the current CPU/system
    * and should therefore not be called too often if performance is a concern.
    *
-   *
-   * @return true if the implementation can be safely used on the current system (determined at runtime)
+   * @return true if the implementation can be safely used on the current system (determined at runtime).
    */
   bool supported_by_runtime_system() const;
 
@@ -3357,7 +3403,7 @@ class implementation {
    *
    * The instruction sets this implementation is compiled against.
    *
-   * @return a mask of all required `internal::instruction_set::` values
+   * @return a mask of all required `internal::instruction_set::` values.
    */
   virtual uint32_t required_instruction_sets() const { return _required_instruction_sets; };
 
@@ -3370,7 +3416,7 @@ class implementation {
    * @param capacity The largest document that will be passed to the parser.
    * @param max_depth The maximum JSON object/array nesting this parser is expected to handle.
    * @param dst The place to put the resulting parser implementation.
-   * @return the name of the implementation, e.g. "haswell", "westmere", "arm64"
+   * @return the error code, or SUCCESS if there was no error.
    */
   virtual error_code create_dom_parser_implementation(
     size_t capacity,
@@ -3407,7 +3453,7 @@ class implementation {
 
 protected:
   /** @private Construct an implementation with the given name and description. For subclasses. */
-  simdjson_really_inline implementation(
+  simdjson_inline implementation(
     std::string_view name,
     std::string_view description,
     uint32_t required_instruction_sets
@@ -3445,7 +3491,7 @@ namespace internal {
 class available_implementation_list {
 public:
   /** Get the list of available implementations compiled into simdjson */
-  simdjson_really_inline available_implementation_list() {}
+  simdjson_inline available_implementation_list() {}
   /** Number of implementations */
   size_t size() const noexcept;
   /** STL const begin() iterator */
@@ -3572,7 +3618,7 @@ namespace internal {
 //
 
 template<typename T>
-simdjson_really_inline void simdjson_result_base<T>::tie(T &value, error_code &error) && noexcept {
+simdjson_inline void simdjson_result_base<T>::tie(T &value, error_code &error) && noexcept {
   error = this->second;
   if (!error) {
     value = std::forward<simdjson_result_base<T>>(*this).first;
@@ -3580,64 +3626,64 @@ simdjson_really_inline void simdjson_result_base<T>::tie(T &value, error_code &e
 }
 
 template<typename T>
-simdjson_warn_unused simdjson_really_inline error_code simdjson_result_base<T>::get(T &value) && noexcept {
+simdjson_warn_unused simdjson_inline error_code simdjson_result_base<T>::get(T &value) && noexcept {
   error_code error;
   std::forward<simdjson_result_base<T>>(*this).tie(value, error);
   return error;
 }
 
 template<typename T>
-simdjson_really_inline error_code simdjson_result_base<T>::error() const noexcept {
+simdjson_inline error_code simdjson_result_base<T>::error() const noexcept {
   return this->second;
 }
 
 #if SIMDJSON_EXCEPTIONS
 
 template<typename T>
-simdjson_really_inline T& simdjson_result_base<T>::value() & noexcept(false) {
+simdjson_inline T& simdjson_result_base<T>::value() & noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return this->first;
 }
 
 template<typename T>
-simdjson_really_inline T&& simdjson_result_base<T>::value() && noexcept(false) {
+simdjson_inline T&& simdjson_result_base<T>::value() && noexcept(false) {
   return std::forward<simdjson_result_base<T>>(*this).take_value();
 }
 
 template<typename T>
-simdjson_really_inline T&& simdjson_result_base<T>::take_value() && noexcept(false) {
+simdjson_inline T&& simdjson_result_base<T>::take_value() && noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return std::forward<T>(this->first);
 }
 
 template<typename T>
-simdjson_really_inline simdjson_result_base<T>::operator T&&() && noexcept(false) {
+simdjson_inline simdjson_result_base<T>::operator T&&() && noexcept(false) {
   return std::forward<simdjson_result_base<T>>(*this).take_value();
 }
 
 #endif // SIMDJSON_EXCEPTIONS
 
 template<typename T>
-simdjson_really_inline const T& simdjson_result_base<T>::value_unsafe() const& noexcept {
+simdjson_inline const T& simdjson_result_base<T>::value_unsafe() const& noexcept {
   return this->first;
 }
 
 template<typename T>
-simdjson_really_inline T&& simdjson_result_base<T>::value_unsafe() && noexcept {
+simdjson_inline T&& simdjson_result_base<T>::value_unsafe() && noexcept {
   return std::forward<T>(this->first);
 }
 
 template<typename T>
-simdjson_really_inline simdjson_result_base<T>::simdjson_result_base(T &&value, error_code error) noexcept
+simdjson_inline simdjson_result_base<T>::simdjson_result_base(T &&value, error_code error) noexcept
     : std::pair<T, error_code>(std::forward<T>(value), error) {}
 template<typename T>
-simdjson_really_inline simdjson_result_base<T>::simdjson_result_base(error_code error) noexcept
+simdjson_inline simdjson_result_base<T>::simdjson_result_base(error_code error) noexcept
     : simdjson_result_base(T{}, error) {}
 template<typename T>
-simdjson_really_inline simdjson_result_base<T>::simdjson_result_base(T &&value) noexcept
+simdjson_inline simdjson_result_base<T>::simdjson_result_base(T &&value) noexcept
     : simdjson_result_base(std::forward<T>(value), SUCCESS) {}
 template<typename T>
-simdjson_really_inline simdjson_result_base<T>::simdjson_result_base() noexcept
+simdjson_inline simdjson_result_base<T>::simdjson_result_base() noexcept
     : simdjson_result_base(T{}, UNINITIALIZED) {}
 
 } // namespace internal
@@ -3647,65 +3693,65 @@ simdjson_really_inline simdjson_result_base<T>::simdjson_result_base() noexcept
 ///
 
 template<typename T>
-simdjson_really_inline void simdjson_result<T>::tie(T &value, error_code &error) && noexcept {
+simdjson_inline void simdjson_result<T>::tie(T &value, error_code &error) && noexcept {
   std::forward<internal::simdjson_result_base<T>>(*this).tie(value, error);
 }
 
 template<typename T>
-simdjson_warn_unused simdjson_really_inline error_code simdjson_result<T>::get(T &value) && noexcept {
+simdjson_warn_unused simdjson_inline error_code simdjson_result<T>::get(T &value) && noexcept {
   return std::forward<internal::simdjson_result_base<T>>(*this).get(value);
 }
 
 template<typename T>
-simdjson_really_inline error_code simdjson_result<T>::error() const noexcept {
+simdjson_inline error_code simdjson_result<T>::error() const noexcept {
   return internal::simdjson_result_base<T>::error();
 }
 
 #if SIMDJSON_EXCEPTIONS
 
 template<typename T>
-simdjson_really_inline T& simdjson_result<T>::value() & noexcept(false) {
+simdjson_inline T& simdjson_result<T>::value() & noexcept(false) {
   return internal::simdjson_result_base<T>::value();
 }
 
 template<typename T>
-simdjson_really_inline T&& simdjson_result<T>::value() && noexcept(false) {
+simdjson_inline T&& simdjson_result<T>::value() && noexcept(false) {
   return std::forward<internal::simdjson_result_base<T>>(*this).value();
 }
 
 template<typename T>
-simdjson_really_inline T&& simdjson_result<T>::take_value() && noexcept(false) {
+simdjson_inline T&& simdjson_result<T>::take_value() && noexcept(false) {
   return std::forward<internal::simdjson_result_base<T>>(*this).take_value();
 }
 
 template<typename T>
-simdjson_really_inline simdjson_result<T>::operator T&&() && noexcept(false) {
+simdjson_inline simdjson_result<T>::operator T&&() && noexcept(false) {
   return std::forward<internal::simdjson_result_base<T>>(*this).take_value();
 }
 
 #endif // SIMDJSON_EXCEPTIONS
 
 template<typename T>
-simdjson_really_inline const T& simdjson_result<T>::value_unsafe() const& noexcept {
+simdjson_inline const T& simdjson_result<T>::value_unsafe() const& noexcept {
   return internal::simdjson_result_base<T>::value_unsafe();
 }
 
 template<typename T>
-simdjson_really_inline T&& simdjson_result<T>::value_unsafe() && noexcept {
+simdjson_inline T&& simdjson_result<T>::value_unsafe() && noexcept {
   return std::forward<internal::simdjson_result_base<T>>(*this).value_unsafe();
 }
 
 template<typename T>
-simdjson_really_inline simdjson_result<T>::simdjson_result(T &&value, error_code error) noexcept
+simdjson_inline simdjson_result<T>::simdjson_result(T &&value, error_code error) noexcept
     : internal::simdjson_result_base<T>(std::forward<T>(value), error) {}
 template<typename T>
-simdjson_really_inline simdjson_result<T>::simdjson_result(error_code error) noexcept
+simdjson_inline simdjson_result<T>::simdjson_result(error_code error) noexcept
     : internal::simdjson_result_base<T>(error) {}
 template<typename T>
-simdjson_really_inline simdjson_result<T>::simdjson_result(T &&value) noexcept
+simdjson_inline simdjson_result<T>::simdjson_result(T &&value) noexcept
     : internal::simdjson_result_base<T>(std::forward<T>(value)) {}
 template<typename T>
-simdjson_really_inline simdjson_result<T>::simdjson_result() noexcept
+simdjson_inline simdjson_result<T>::simdjson_result() noexcept
     : internal::simdjson_result_base<T>() {}
 
 } // namespace simdjson
@@ -3758,7 +3804,7 @@ inline char *allocate_padded_buffer(size_t length) noexcept {
 } // namespace internal
 
 
-inline padded_string::padded_string() noexcept {}
+inline padded_string::padded_string() noexcept = default;
 inline padded_string::padded_string(size_t length) noexcept
     : viable_size(length), data_ptr(internal::allocate_padded_buffer(length)) {
 }
@@ -3840,7 +3886,13 @@ inline simdjson_result<padded_string> padded_string::load(std::string_view filen
   }
 
   // Get the file size
-  if(std::fseek(fp, 0, SEEK_END) < 0) {
+  int ret;
+#if defined(SIMDJSON_VISUAL_STUDIO) && !SIMDJSON_IS_32BITS
+  ret = _fseeki64(fp, 0, SEEK_END);
+#else
+  ret = std::fseek(fp, 0, SEEK_END);
+#endif // _WIN64
+  if(ret < 0) {
     std::fclose(fp);
     return IO_ERROR;
   }
@@ -3984,25 +4036,26 @@ constexpr const uint32_t JSON_COUNT_MASK = 0xFFFFFF;
  */
 class tape_ref {
 public:
-  simdjson_really_inline tape_ref() noexcept;
-  simdjson_really_inline tape_ref(const dom::document *doc, size_t json_index) noexcept;
+  simdjson_inline tape_ref() noexcept;
+  simdjson_inline tape_ref(const dom::document *doc, size_t json_index) noexcept;
   inline size_t after_element() const noexcept;
-  simdjson_really_inline tape_type tape_ref_type() const noexcept;
-  simdjson_really_inline uint64_t tape_value() const noexcept;
-  simdjson_really_inline bool is_double() const noexcept;
-  simdjson_really_inline bool is_int64() const noexcept;
-  simdjson_really_inline bool is_uint64() const noexcept;
-  simdjson_really_inline bool is_false() const noexcept;
-  simdjson_really_inline bool is_true() const noexcept;
-  simdjson_really_inline bool is_null_on_tape() const noexcept;// different name to avoid clash with is_null.
-  simdjson_really_inline uint32_t matching_brace_index() const noexcept;
-  simdjson_really_inline uint32_t scope_count() const noexcept;
+  simdjson_inline tape_type tape_ref_type() const noexcept;
+  simdjson_inline uint64_t tape_value() const noexcept;
+  simdjson_inline bool is_double() const noexcept;
+  simdjson_inline bool is_int64() const noexcept;
+  simdjson_inline bool is_uint64() const noexcept;
+  simdjson_inline bool is_false() const noexcept;
+  simdjson_inline bool is_true() const noexcept;
+  simdjson_inline bool is_null_on_tape() const noexcept;// different name to avoid clash with is_null.
+  simdjson_inline uint32_t matching_brace_index() const noexcept;
+  simdjson_inline uint32_t scope_count() const noexcept;
   template<typename T>
-  simdjson_really_inline T next_tape_value() const noexcept;
-  simdjson_really_inline uint32_t get_string_length() const noexcept;
-  simdjson_really_inline const char * get_c_str() const noexcept;
+  simdjson_inline T next_tape_value() const noexcept;
+  simdjson_inline uint32_t get_string_length() const noexcept;
+  simdjson_inline const char * get_c_str() const noexcept;
   inline std::string_view get_string_view() const noexcept;
-  simdjson_really_inline bool is_document_root() const noexcept;
+  simdjson_inline bool is_document_root() const noexcept;
+  simdjson_inline bool usable() const noexcept;
 
   /** The document this element references. */
   const dom::document *doc;
@@ -4034,7 +4087,7 @@ class element;
 class array {
 public:
   /** Create a new, invalid array */
-  simdjson_really_inline array() noexcept;
+  simdjson_inline array() noexcept;
 
   class iterator {
   public:
@@ -4074,7 +4127,7 @@ class array {
     iterator(const iterator&) noexcept = default;
     iterator& operator=(const iterator&) noexcept = default;
   private:
-    simdjson_really_inline iterator(const internal::tape_ref &tape) noexcept;
+    simdjson_inline iterator(const internal::tape_ref &tape) noexcept;
     internal::tape_ref tape;
     friend class array;
   };
@@ -4145,7 +4198,7 @@ class array {
   inline simdjson_result<element> at(size_t index) const noexcept;
 
 private:
-  simdjson_really_inline array(const internal::tape_ref &tape) noexcept;
+  simdjson_inline array(const internal::tape_ref &tape) noexcept;
   internal::tape_ref tape;
   friend class element;
   friend struct simdjson_result<element>;
@@ -4160,9 +4213,9 @@ class array {
 template<>
 struct simdjson_result<dom::array> : public internal::simdjson_result_base<dom::array> {
 public:
-  simdjson_really_inline simdjson_result() noexcept; ///< @private
-  simdjson_really_inline simdjson_result(dom::array value) noexcept; ///< @private
-  simdjson_really_inline simdjson_result(error_code error) noexcept; ///< @private
+  simdjson_inline simdjson_result() noexcept; ///< @private
+  simdjson_inline simdjson_result(dom::array value) noexcept; ///< @private
+  simdjson_inline simdjson_result(error_code error) noexcept; ///< @private
 
   inline simdjson_result<dom::element> at_pointer(std::string_view json_pointer) const noexcept;
   inline simdjson_result<dom::element> at(size_t index) const noexcept;
@@ -4357,20 +4410,20 @@ class parser {
    *    to allocate an initial capacity, call allocate() after constructing the parser.
    *    Defaults to SIMDJSON_MAXSIZE_BYTES (the largest single document simdjson can process).
    */
-  simdjson_really_inline explicit parser(size_t max_capacity = SIMDJSON_MAXSIZE_BYTES) noexcept;
+  simdjson_inline explicit parser(size_t max_capacity = SIMDJSON_MAXSIZE_BYTES) noexcept;
   /**
    * Take another parser's buffers and state.
    *
    * @param other The parser to take. Its capacity is zeroed.
    */
-  simdjson_really_inline parser(parser &&other) noexcept;
+  simdjson_inline parser(parser &&other) noexcept;
   parser(const parser &) = delete; ///< @private Disallow copying
   /**
    * Take another parser's buffers and state.
    *
    * @param other The parser to take. Its capacity is zeroed.
    */
-  simdjson_really_inline parser &operator=(parser &&other) noexcept;
+  simdjson_inline parser &operator=(parser &&other) noexcept;
   parser &operator=(const parser &) = delete; ///< @private Disallow copying
 
   /** Deallocate the JSON parser. */
@@ -4480,17 +4533,20 @@ class parser {
   inline simdjson_result<element> parse(const uint8_t *buf, size_t len, bool realloc_if_needed = true) & noexcept;
   inline simdjson_result<element> parse(const uint8_t *buf, size_t len, bool realloc_if_needed = true) && =delete;
   /** @overload parse(const uint8_t *buf, size_t len, bool realloc_if_needed) */
-  simdjson_really_inline simdjson_result<element> parse(const char *buf, size_t len, bool realloc_if_needed = true) & noexcept;
-  simdjson_really_inline simdjson_result<element> parse(const char *buf, size_t len, bool realloc_if_needed = true) && =delete;
+  simdjson_inline simdjson_result<element> parse(const char *buf, size_t len, bool realloc_if_needed = true) & noexcept;
+  simdjson_inline simdjson_result<element> parse(const char *buf, size_t len, bool realloc_if_needed = true) && =delete;
+  /** @overload parse(const uint8_t *buf, size_t len, bool realloc_if_needed) */
+  simdjson_inline simdjson_result<element> parse(const std::string &s) & noexcept;
+  simdjson_inline simdjson_result<element> parse(const std::string &s) && =delete;
   /** @overload parse(const uint8_t *buf, size_t len, bool realloc_if_needed) */
-  simdjson_really_inline simdjson_result<element> parse(const std::string &s) & noexcept;
-  simdjson_really_inline simdjson_result<element> parse(const std::string &s) && =delete;
+  simdjson_inline simdjson_result<element> parse(const padded_string &s) & noexcept;
+  simdjson_inline simdjson_result<element> parse(const padded_string &s) && =delete;
   /** @overload parse(const uint8_t *buf, size_t len, bool realloc_if_needed) */
-  simdjson_really_inline simdjson_result<element> parse(const padded_string &s) & noexcept;
-  simdjson_really_inline simdjson_result<element> parse(const padded_string &s) && =delete;
+  simdjson_inline simdjson_result<element> parse(const padded_string_view &v) & noexcept;
+  simdjson_inline simdjson_result<element> parse(const padded_string_view &v) && =delete;
 
   /** @private We do not want to allow implicit conversion from C string to std::string. */
-  simdjson_really_inline simdjson_result<element> parse(const char *buf) noexcept = delete;
+  simdjson_inline simdjson_result<element> parse(const char *buf) noexcept = delete;
 
   /**
    * Parse a JSON document into a provide document instance and return a temporary reference to it.
@@ -4539,17 +4595,17 @@ class parser {
   inline simdjson_result<element> parse_into_document(document& doc, const uint8_t *buf, size_t len, bool realloc_if_needed = true) & noexcept;
   inline simdjson_result<element> parse_into_document(document& doc, const uint8_t *buf, size_t len, bool realloc_if_needed = true) && =delete;
   /** @overload parse_into_document(const uint8_t *buf, size_t len, bool realloc_if_needed) */
-  simdjson_really_inline simdjson_result<element> parse_into_document(document& doc, const char *buf, size_t len, bool realloc_if_needed = true) & noexcept;
-  simdjson_really_inline simdjson_result<element> parse_into_document(document& doc, const char *buf, size_t len, bool realloc_if_needed = true) && =delete;
+  simdjson_inline simdjson_result<element> parse_into_document(document& doc, const char *buf, size_t len, bool realloc_if_needed = true) & noexcept;
+  simdjson_inline simdjson_result<element> parse_into_document(document& doc, const char *buf, size_t len, bool realloc_if_needed = true) && =delete;
   /** @overload parse_into_document(const uint8_t *buf, size_t len, bool realloc_if_needed) */
-  simdjson_really_inline simdjson_result<element> parse_into_document(document& doc, const std::string &s) & noexcept;
-  simdjson_really_inline simdjson_result<element> parse_into_document(document& doc, const std::string &s) && =delete;
+  simdjson_inline simdjson_result<element> parse_into_document(document& doc, const std::string &s) & noexcept;
+  simdjson_inline simdjson_result<element> parse_into_document(document& doc, const std::string &s) && =delete;
   /** @overload parse_into_document(const uint8_t *buf, size_t len, bool realloc_if_needed) */
-  simdjson_really_inline simdjson_result<element> parse_into_document(document& doc, const padded_string &s) & noexcept;
-  simdjson_really_inline simdjson_result<element> parse_into_document(document& doc, const padded_string &s) && =delete;
+  simdjson_inline simdjson_result<element> parse_into_document(document& doc, const padded_string &s) & noexcept;
+  simdjson_inline simdjson_result<element> parse_into_document(document& doc, const padded_string &s) && =delete;
 
   /** @private We do not want to allow implicit conversion from C string to std::string. */
-  simdjson_really_inline simdjson_result<element> parse_into_document(document& doc, const char *buf) noexcept = delete;
+  simdjson_inline simdjson_result<element> parse_into_document(document& doc, const char *buf) noexcept = delete;
 
   /**
    * Load a file containing many JSON documents.
@@ -4761,7 +4817,7 @@ class parser {
    *
    * @return Current capacity, in bytes.
    */
-  simdjson_really_inline size_t capacity() const noexcept;
+  simdjson_inline size_t capacity() const noexcept;
 
   /**
    * The largest document this parser can automatically support.
@@ -4770,14 +4826,14 @@ class parser {
    *
    * @return Maximum capacity, in bytes.
    */
-  simdjson_really_inline size_t max_capacity() const noexcept;
+  simdjson_inline size_t max_capacity() const noexcept;
 
   /**
    * The maximum level of nested object and arrays supported by this parser.
    *
    * @return Maximum depth, in bytes.
    */
-  simdjson_really_inline size_t max_depth() const noexcept;
+  simdjson_inline size_t max_depth() const noexcept;
 
   /**
    * Set max_capacity. This is the largest document this parser can automatically support.
@@ -4793,7 +4849,7 @@ class parser {
    *
    * @param max_capacity The new maximum capacity, in bytes.
    */
-  simdjson_really_inline void set_max_capacity(size_t max_capacity) noexcept;
+  simdjson_inline void set_max_capacity(size_t max_capacity) noexcept;
 
 #ifdef SIMDJSON_THREADS_ENABLED
   /**
@@ -4975,13 +5031,13 @@ class document_stream {
    *  error = parser.parse_many(json).get(docs);
    *  ```
    */
-  simdjson_really_inline document_stream() noexcept;
+  simdjson_inline document_stream() noexcept;
   /** Move one document_stream to another. */
-  simdjson_really_inline document_stream(document_stream &&other) noexcept = default;
+  simdjson_inline document_stream(document_stream &&other) noexcept = default;
   /** Move one document_stream to another. */
-  simdjson_really_inline document_stream &operator=(document_stream &&other) noexcept = default;
+  simdjson_inline document_stream &operator=(document_stream &&other) noexcept = default;
 
-  simdjson_really_inline ~document_stream() noexcept;
+  simdjson_inline ~document_stream() noexcept;
   /**
    * Returns the input size in bytes.
    */
@@ -5020,11 +5076,11 @@ class document_stream {
     /**
      * Default constructor.
      */
-    simdjson_really_inline iterator() noexcept;
+    simdjson_inline iterator() noexcept;
     /**
      * Get the current document (or error).
      */
-    simdjson_really_inline reference operator*() noexcept;
+    simdjson_inline reference operator*() noexcept;
     /**
      * Advance to the next document (prefix).
      */
@@ -5033,7 +5089,7 @@ class document_stream {
      * Check if we're at the end yet.
      * @param other the end iterator to compare to.
      */
-    simdjson_really_inline bool operator!=(const iterator &other) const noexcept;
+    simdjson_inline bool operator!=(const iterator &other) const noexcept;
     /**
      * @private
      *
@@ -5049,7 +5105,7 @@ class document_stream {
      * may change in future versions of simdjson: we find the API somewhat
      * awkward and we would like to offer something friendlier.
      */
-     simdjson_really_inline size_t current_index() const noexcept;
+     simdjson_inline size_t current_index() const noexcept;
     /**
      * @private
      *
@@ -5069,10 +5125,10 @@ class document_stream {
      * may change in future versions of simdjson: we find the API somewhat
      * awkward and we would like to offer something friendlier.
      */
-     simdjson_really_inline std::string_view source() const noexcept;
+     simdjson_inline std::string_view source() const noexcept;
 
   private:
-    simdjson_really_inline iterator(document_stream *s, bool finished) noexcept;
+    simdjson_inline iterator(document_stream *s, bool finished) noexcept;
     /** The document_stream we're iterating through. */
     document_stream* stream;
     /** Whether we're finished or not. */
@@ -5083,11 +5139,11 @@ class document_stream {
   /**
    * Start iterating the documents in the stream.
    */
-  simdjson_really_inline iterator begin() noexcept;
+  simdjson_inline iterator begin() noexcept;
   /**
    * The end of the stream, for iterator comparison purposes.
    */
-  simdjson_really_inline iterator end() noexcept;
+  simdjson_inline iterator end() noexcept;
 
 private:
 
@@ -5103,7 +5159,7 @@ class document_stream {
    * @param len is the length of the raw byte buffer in bytes
    * @param batch_size is the size of the windows (must be strictly greater or equal to the largest JSON document)
    */
-  simdjson_really_inline document_stream(
+  simdjson_inline document_stream(
     dom::parser &parser,
     const uint8_t *buf,
     size_t len,
@@ -5197,19 +5253,19 @@ class document_stream {
 template<>
 struct simdjson_result<dom::document_stream> : public internal::simdjson_result_base<dom::document_stream> {
 public:
-  simdjson_really_inline simdjson_result() noexcept; ///< @private
-  simdjson_really_inline simdjson_result(error_code error) noexcept; ///< @private
-  simdjson_really_inline simdjson_result(dom::document_stream &&value) noexcept; ///< @private
+  simdjson_inline simdjson_result() noexcept; ///< @private
+  simdjson_inline simdjson_result(error_code error) noexcept; ///< @private
+  simdjson_inline simdjson_result(dom::document_stream &&value) noexcept; ///< @private
 
 #if SIMDJSON_EXCEPTIONS
-  simdjson_really_inline dom::document_stream::iterator begin() noexcept(false);
-  simdjson_really_inline dom::document_stream::iterator end() noexcept(false);
+  simdjson_inline dom::document_stream::iterator begin() noexcept(false);
+  simdjson_inline dom::document_stream::iterator end() noexcept(false);
 #else // SIMDJSON_EXCEPTIONS
 #ifndef SIMDJSON_DISABLE_DEPRECATED_API
   [[deprecated("parse_many() and load_many() may return errors. Use document_stream stream; error = parser.parse_many().get(doc); instead.")]]
-  simdjson_really_inline dom::document_stream::iterator begin() noexcept;
+  simdjson_inline dom::document_stream::iterator begin() noexcept;
   [[deprecated("parse_many() and load_many() may return errors. Use document_stream stream; error = parser.parse_many().get(doc); instead.")]]
-  simdjson_really_inline dom::document_stream::iterator end() noexcept;
+  simdjson_inline dom::document_stream::iterator end() noexcept;
 #endif // SIMDJSON_DISABLE_DEPRECATED_API
 #endif // SIMDJSON_EXCEPTIONS
 }; // struct simdjson_result<dom::document_stream>
@@ -5258,10 +5314,10 @@ enum class element_type {
 class element {
 public:
   /** Create a new, invalid element. */
-  simdjson_really_inline element() noexcept;
+  simdjson_inline element() noexcept;
 
   /** The type of this element. */
-  simdjson_really_inline element_type type() const noexcept;
+  simdjson_inline element_type type() const noexcept;
 
   /**
    * Cast this element to an array.
@@ -5413,7 +5469,7 @@ class element {
    * @tparam T bool, double, uint64_t, int64_t, std::string_view, const char *, dom::array, dom::object
    */
   template<typename T>
-  simdjson_really_inline bool is() const noexcept;
+  simdjson_inline bool is() const noexcept;
 
   /**
    * Get the value as the provided type (T).
@@ -5458,7 +5514,7 @@ class element {
    * @returns The error that occurred, or SUCCESS if there was no error.
    */
   template<typename T>
-  simdjson_warn_unused simdjson_really_inline error_code get(T &value) const noexcept;
+  simdjson_warn_unused simdjson_inline error_code get(T &value) const noexcept;
 
   /**
    * Get the value as the provided type (T), setting error if it's not the given type.
@@ -5682,7 +5738,7 @@ class element {
   inline bool dump_raw_tape(std::ostream &out) const noexcept;
 
 private:
-  simdjson_really_inline element(const internal::tape_ref &tape) noexcept;
+  simdjson_inline element(const internal::tape_ref &tape) noexcept;
   internal::tape_ref tape;
   friend class document;
   friend class object;
@@ -5699,59 +5755,59 @@ class element {
 template<>
 struct simdjson_result<dom::element> : public internal::simdjson_result_base<dom::element> {
 public:
-  simdjson_really_inline simdjson_result() noexcept; ///< @private
-  simdjson_really_inline simdjson_result(dom::element &&value) noexcept; ///< @private
-  simdjson_really_inline simdjson_result(error_code error) noexcept; ///< @private
+  simdjson_inline simdjson_result() noexcept; ///< @private
+  simdjson_inline simdjson_result(dom::element &&value) noexcept; ///< @private
+  simdjson_inline simdjson_result(error_code error) noexcept; ///< @private
 
-  simdjson_really_inline simdjson_result<dom::element_type> type() const noexcept;
+  simdjson_inline simdjson_result<dom::element_type> type() const noexcept;
   template<typename T>
-  simdjson_really_inline bool is() const noexcept;
+  simdjson_inline bool is() const noexcept;
   template<typename T>
-  simdjson_really_inline simdjson_result<T> get() const noexcept;
+  simdjson_inline simdjson_result<T> get() const noexcept;
   template<typename T>
-  simdjson_warn_unused simdjson_really_inline error_code get(T &value) const noexcept;
-
-  simdjson_really_inline simdjson_result<dom::array> get_array() const noexcept;
-  simdjson_really_inline simdjson_result<dom::object> get_object() const noexcept;
-  simdjson_really_inline simdjson_result<const char *> get_c_str() const noexcept;
-  simdjson_really_inline simdjson_result<size_t> get_string_length() const noexcept;
-  simdjson_really_inline simdjson_result<std::string_view> get_string() const noexcept;
-  simdjson_really_inline simdjson_result<int64_t> get_int64() const noexcept;
-  simdjson_really_inline simdjson_result<uint64_t> get_uint64() const noexcept;
-  simdjson_really_inline simdjson_result<double> get_double() const noexcept;
-  simdjson_really_inline simdjson_result<bool> get_bool() const noexcept;
-
-  simdjson_really_inline bool is_array() const noexcept;
-  simdjson_really_inline bool is_object() const noexcept;
-  simdjson_really_inline bool is_string() const noexcept;
-  simdjson_really_inline bool is_int64() const noexcept;
-  simdjson_really_inline bool is_uint64() const noexcept;
-  simdjson_really_inline bool is_double() const noexcept;
-  simdjson_really_inline bool is_number() const noexcept;
-  simdjson_really_inline bool is_bool() const noexcept;
-  simdjson_really_inline bool is_null() const noexcept;
-
-  simdjson_really_inline simdjson_result<dom::element> operator[](std::string_view key) const noexcept;
-  simdjson_really_inline simdjson_result<dom::element> operator[](const char *key) const noexcept;
-  simdjson_really_inline simdjson_result<dom::element> at_pointer(const std::string_view json_pointer) const noexcept;
+  simdjson_warn_unused simdjson_inline error_code get(T &value) const noexcept;
+
+  simdjson_inline simdjson_result<dom::array> get_array() const noexcept;
+  simdjson_inline simdjson_result<dom::object> get_object() const noexcept;
+  simdjson_inline simdjson_result<const char *> get_c_str() const noexcept;
+  simdjson_inline simdjson_result<size_t> get_string_length() const noexcept;
+  simdjson_inline simdjson_result<std::string_view> get_string() const noexcept;
+  simdjson_inline simdjson_result<int64_t> get_int64() const noexcept;
+  simdjson_inline simdjson_result<uint64_t> get_uint64() const noexcept;
+  simdjson_inline simdjson_result<double> get_double() const noexcept;
+  simdjson_inline simdjson_result<bool> get_bool() const noexcept;
+
+  simdjson_inline bool is_array() const noexcept;
+  simdjson_inline bool is_object() const noexcept;
+  simdjson_inline bool is_string() const noexcept;
+  simdjson_inline bool is_int64() const noexcept;
+  simdjson_inline bool is_uint64() const noexcept;
+  simdjson_inline bool is_double() const noexcept;
+  simdjson_inline bool is_number() const noexcept;
+  simdjson_inline bool is_bool() const noexcept;
+  simdjson_inline bool is_null() const noexcept;
+
+  simdjson_inline simdjson_result<dom::element> operator[](std::string_view key) const noexcept;
+  simdjson_inline simdjson_result<dom::element> operator[](const char *key) const noexcept;
+  simdjson_inline simdjson_result<dom::element> at_pointer(const std::string_view json_pointer) const noexcept;
   [[deprecated("For standard compliance, use at_pointer instead, and prefix your pointers with a slash '/', see RFC6901 ")]]
-  simdjson_really_inline simdjson_result<dom::element> at(const std::string_view json_pointer) const noexcept;
-  simdjson_really_inline simdjson_result<dom::element> at(size_t index) const noexcept;
-  simdjson_really_inline simdjson_result<dom::element> at_key(std::string_view key) const noexcept;
-  simdjson_really_inline simdjson_result<dom::element> at_key_case_insensitive(std::string_view key) const noexcept;
+  simdjson_inline simdjson_result<dom::element> at(const std::string_view json_pointer) const noexcept;
+  simdjson_inline simdjson_result<dom::element> at(size_t index) const noexcept;
+  simdjson_inline simdjson_result<dom::element> at_key(std::string_view key) const noexcept;
+  simdjson_inline simdjson_result<dom::element> at_key_case_insensitive(std::string_view key) const noexcept;
 
 #if SIMDJSON_EXCEPTIONS
-  simdjson_really_inline operator bool() const noexcept(false);
-  simdjson_really_inline explicit operator const char*() const noexcept(false);
-  simdjson_really_inline operator std::string_view() const noexcept(false);
-  simdjson_really_inline operator uint64_t() const noexcept(false);
-  simdjson_really_inline operator int64_t() const noexcept(false);
-  simdjson_really_inline operator double() const noexcept(false);
-  simdjson_really_inline operator dom::array() const noexcept(false);
-  simdjson_really_inline operator dom::object() const noexcept(false);
-
-  simdjson_really_inline dom::array::iterator begin() const noexcept(false);
-  simdjson_really_inline dom::array::iterator end() const noexcept(false);
+  simdjson_inline operator bool() const noexcept(false);
+  simdjson_inline explicit operator const char*() const noexcept(false);
+  simdjson_inline operator std::string_view() const noexcept(false);
+  simdjson_inline operator uint64_t() const noexcept(false);
+  simdjson_inline operator int64_t() const noexcept(false);
+  simdjson_inline operator double() const noexcept(false);
+  simdjson_inline operator dom::array() const noexcept(false);
+  simdjson_inline operator dom::object() const noexcept(false);
+
+  simdjson_inline dom::array::iterator begin() const noexcept(false);
+  simdjson_inline dom::array::iterator end() const noexcept(false);
 #endif // SIMDJSON_EXCEPTIONS
 };
 
@@ -5782,7 +5838,7 @@ class key_value_pair;
 class object {
 public:
   /** Create a new, invalid object */
-  simdjson_really_inline object() noexcept;
+  simdjson_inline object() noexcept;
 
   class iterator {
   public:
@@ -5852,7 +5908,7 @@ class object {
     iterator(const iterator&) noexcept = default;
     iterator& operator=(const iterator&) noexcept = default;
   private:
-    simdjson_really_inline iterator(const internal::tape_ref &tape) noexcept;
+    simdjson_inline iterator(const internal::tape_ref &tape) noexcept;
 
     internal::tape_ref tape;
 
@@ -5966,7 +6022,7 @@ class object {
   inline simdjson_result<element> at_key_case_insensitive(std::string_view key) const noexcept;
 
 private:
-  simdjson_really_inline object(const internal::tape_ref &tape) noexcept;
+  simdjson_inline object(const internal::tape_ref &tape) noexcept;
 
   internal::tape_ref tape;
 
@@ -5987,7 +6043,7 @@ class key_value_pair {
   element value;
 
 private:
-  simdjson_really_inline key_value_pair(std::string_view _key, element _value) noexcept;
+  simdjson_inline key_value_pair(std::string_view _key, element _value) noexcept;
   friend class object;
 };
 
@@ -5997,9 +6053,9 @@ class key_value_pair {
 template<>
 struct simdjson_result<dom::object> : public internal::simdjson_result_base<dom::object> {
 public:
-  simdjson_really_inline simdjson_result() noexcept; ///< @private
-  simdjson_really_inline simdjson_result(dom::object value) noexcept; ///< @private
-  simdjson_really_inline simdjson_result(error_code error) noexcept; ///< @private
+  simdjson_inline simdjson_result() noexcept; ///< @private
+  simdjson_inline simdjson_result(dom::object value) noexcept; ///< @private
+  simdjson_inline simdjson_result(error_code error) noexcept; ///< @private
 
   inline simdjson_result<dom::element> operator[](std::string_view key) const noexcept;
   inline simdjson_result<dom::element> operator[](const char *key) const noexcept;
@@ -6069,7 +6125,7 @@ class string_builder {
   /** Append an object to the builder (to be printed) **/
   inline void append(simdjson::dom::object value);
   /** Reset the builder (so that it would print the empty string) **/
-  simdjson_really_inline void clear();
+  simdjson_inline void clear();
   /**
    * Get access to the string. The string_view is owned by the builder
    * and it is invalid to use it after the string_builder has been
@@ -6077,9 +6133,9 @@ class string_builder {
    * However you can make a copy of the string_view on memory that you
    * own.
    */
-  simdjson_really_inline std::string_view str() const;
+  simdjson_inline std::string_view str() const;
   /** Append a key_value_pair to the builder (to be printed) **/
-  simdjson_really_inline void append(simdjson::dom::key_value_pair value);
+  simdjson_inline void append(simdjson::dom::key_value_pair value);
 private:
   formatter format{};
 };
@@ -6093,43 +6149,43 @@ class mini_formatter {
 public:
   mini_formatter() = default;
   /** Add a comma **/
-  simdjson_really_inline void comma();
+  simdjson_inline void comma();
   /** Start an array, prints [ **/
-  simdjson_really_inline void start_array();
+  simdjson_inline void start_array();
   /** End an array, prints ] **/
-  simdjson_really_inline void end_array();
+  simdjson_inline void end_array();
   /** Start an array, prints { **/
-  simdjson_really_inline void start_object();
+  simdjson_inline void start_object();
   /** Start an array, prints } **/
-  simdjson_really_inline void end_object();
+  simdjson_inline void end_object();
   /** Prints a true **/
-  simdjson_really_inline void true_atom();
+  simdjson_inline void true_atom();
   /** Prints a false **/
-  simdjson_really_inline void false_atom();
+  simdjson_inline void false_atom();
   /** Prints a null **/
-  simdjson_really_inline void null_atom();
+  simdjson_inline void null_atom();
   /** Prints a number **/
-  simdjson_really_inline void number(int64_t x);
+  simdjson_inline void number(int64_t x);
   /** Prints a number **/
-  simdjson_really_inline void number(uint64_t x);
+  simdjson_inline void number(uint64_t x);
   /** Prints a number **/
-  simdjson_really_inline void number(double x);
+  simdjson_inline void number(double x);
   /** Prints a key (string + colon) **/
-  simdjson_really_inline void key(std::string_view unescaped);
+  simdjson_inline void key(std::string_view unescaped);
   /** Prints a string. The string is escaped as needed. **/
-  simdjson_really_inline void string(std::string_view unescaped);
+  simdjson_inline void string(std::string_view unescaped);
   /** Clears out the content. **/
-  simdjson_really_inline void clear();
+  simdjson_inline void clear();
   /**
    * Get access to the buffer, it is owned by the instance, but
    * the user can make a copy.
    **/
-  simdjson_really_inline std::string_view str() const;
+  simdjson_inline std::string_view str() const;
 
 private:
   // implementation details (subject to change)
   /** Prints one character **/
-  simdjson_really_inline void one_char(char c);
+  simdjson_inline void one_char(char c);
   /** Backing buffer **/
   std::vector<char> buffer{}; // not ideal!
 };
@@ -6758,11 +6814,11 @@ namespace simdjson {
 //
 // simdjson_result<dom::array> inline implementation
 //
-simdjson_really_inline simdjson_result<dom::array>::simdjson_result() noexcept
+simdjson_inline simdjson_result<dom::array>::simdjson_result() noexcept
     : internal::simdjson_result_base<dom::array>() {}
-simdjson_really_inline simdjson_result<dom::array>::simdjson_result(dom::array value) noexcept
+simdjson_inline simdjson_result<dom::array>::simdjson_result(dom::array value) noexcept
     : internal::simdjson_result_base<dom::array>(std::forward<dom::array>(value)) {}
-simdjson_really_inline simdjson_result<dom::array>::simdjson_result(error_code error) noexcept
+simdjson_inline simdjson_result<dom::array>::simdjson_result(error_code error) noexcept
     : internal::simdjson_result_base<dom::array>(error) {}
 
 #if SIMDJSON_EXCEPTIONS
@@ -6796,21 +6852,26 @@ namespace dom {
 //
 // array inline implementation
 //
-simdjson_really_inline array::array() noexcept : tape{} {}
-simdjson_really_inline array::array(const internal::tape_ref &_tape) noexcept : tape{_tape} {}
+simdjson_inline array::array() noexcept : tape{} {}
+simdjson_inline array::array(const internal::tape_ref &_tape) noexcept : tape{_tape} {}
 inline array::iterator array::begin() const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   return internal::tape_ref(tape.doc, tape.json_index + 1);
 }
 inline array::iterator array::end() const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   return internal::tape_ref(tape.doc, tape.after_element() - 1);
 }
 inline size_t array::size() const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   return tape.scope_count();
 }
 inline size_t array::number_of_slots() const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   return tape.matching_brace_index() - tape.json_index;
 }
 inline simdjson_result<element> array::at_pointer(std::string_view json_pointer) const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   if(json_pointer.empty()) { // an empty string means that we return the current node
       return element(this->tape); // copy the current node
   } else if(json_pointer[0] != '/') { // otherwise there is an error
@@ -6851,6 +6912,7 @@ inline simdjson_result<element> array::at_pointer(std::string_view json_pointer)
 }
 
 inline simdjson_result<element> array::at(size_t index) const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   size_t i=0;
   for (auto element : *this) {
     if (i == index) { return element; }
@@ -6862,7 +6924,7 @@ inline simdjson_result<element> array::at(size_t index) const noexcept {
 //
 // array::iterator inline implementation
 //
-simdjson_really_inline array::iterator::iterator(const internal::tape_ref &_tape) noexcept : tape{_tape} { }
+simdjson_inline array::iterator::iterator(const internal::tape_ref &_tape) noexcept : tape{_tape} { }
 inline element array::iterator::operator*() const noexcept {
   return element(tape);
 }
@@ -6911,11 +6973,11 @@ namespace simdjson {
 //
 // simdjson_result<dom::element> inline implementation
 //
-simdjson_really_inline simdjson_result<dom::element>::simdjson_result() noexcept
+simdjson_inline simdjson_result<dom::element>::simdjson_result() noexcept
     : internal::simdjson_result_base<dom::element>() {}
-simdjson_really_inline simdjson_result<dom::element>::simdjson_result(dom::element &&value) noexcept
+simdjson_inline simdjson_result<dom::element>::simdjson_result(dom::element &&value) noexcept
     : internal::simdjson_result_base<dom::element>(std::forward<dom::element>(value)) {}
-simdjson_really_inline simdjson_result<dom::element>::simdjson_result(error_code error) noexcept
+simdjson_inline simdjson_result<dom::element>::simdjson_result(error_code error) noexcept
     : internal::simdjson_result_base<dom::element>(error) {}
 inline simdjson_result<dom::element_type> simdjson_result<dom::element>::type() const noexcept {
   if (error()) { return error(); }
@@ -6923,101 +6985,101 @@ inline simdjson_result<dom::element_type> simdjson_result<dom::element>::type()
 }
 
 template<typename T>
-simdjson_really_inline bool simdjson_result<dom::element>::is() const noexcept {
+simdjson_inline bool simdjson_result<dom::element>::is() const noexcept {
   return !error() && first.is<T>();
 }
 template<typename T>
-simdjson_really_inline simdjson_result<T> simdjson_result<dom::element>::get() const noexcept {
+simdjson_inline simdjson_result<T> simdjson_result<dom::element>::get() const noexcept {
   if (error()) { return error(); }
   return first.get<T>();
 }
 template<typename T>
-simdjson_warn_unused simdjson_really_inline error_code simdjson_result<dom::element>::get(T &value) const noexcept {
+simdjson_warn_unused simdjson_inline error_code simdjson_result<dom::element>::get(T &value) const noexcept {
   if (error()) { return error(); }
   return first.get<T>(value);
 }
 
-simdjson_really_inline simdjson_result<dom::array> simdjson_result<dom::element>::get_array() const noexcept {
+simdjson_inline simdjson_result<dom::array> simdjson_result<dom::element>::get_array() const noexcept {
   if (error()) { return error(); }
   return first.get_array();
 }
-simdjson_really_inline simdjson_result<dom::object> simdjson_result<dom::element>::get_object() const noexcept {
+simdjson_inline simdjson_result<dom::object> simdjson_result<dom::element>::get_object() const noexcept {
   if (error()) { return error(); }
   return first.get_object();
 }
-simdjson_really_inline simdjson_result<const char *> simdjson_result<dom::element>::get_c_str() const noexcept {
+simdjson_inline simdjson_result<const char *> simdjson_result<dom::element>::get_c_str() const noexcept {
   if (error()) { return error(); }
   return first.get_c_str();
 }
-simdjson_really_inline simdjson_result<size_t> simdjson_result<dom::element>::get_string_length() const noexcept {
+simdjson_inline simdjson_result<size_t> simdjson_result<dom::element>::get_string_length() const noexcept {
   if (error()) { return error(); }
   return first.get_string_length();
 }
-simdjson_really_inline simdjson_result<std::string_view> simdjson_result<dom::element>::get_string() const noexcept {
+simdjson_inline simdjson_result<std::string_view> simdjson_result<dom::element>::get_string() const noexcept {
   if (error()) { return error(); }
   return first.get_string();
 }
-simdjson_really_inline simdjson_result<int64_t> simdjson_result<dom::element>::get_int64() const noexcept {
+simdjson_inline simdjson_result<int64_t> simdjson_result<dom::element>::get_int64() const noexcept {
   if (error()) { return error(); }
   return first.get_int64();
 }
-simdjson_really_inline simdjson_result<uint64_t> simdjson_result<dom::element>::get_uint64() const noexcept {
+simdjson_inline simdjson_result<uint64_t> simdjson_result<dom::element>::get_uint64() const noexcept {
   if (error()) { return error(); }
   return first.get_uint64();
 }
-simdjson_really_inline simdjson_result<double> simdjson_result<dom::element>::get_double() const noexcept {
+simdjson_inline simdjson_result<double> simdjson_result<dom::element>::get_double() const noexcept {
   if (error()) { return error(); }
   return first.get_double();
 }
-simdjson_really_inline simdjson_result<bool> simdjson_result<dom::element>::get_bool() const noexcept {
+simdjson_inline simdjson_result<bool> simdjson_result<dom::element>::get_bool() const noexcept {
   if (error()) { return error(); }
   return first.get_bool();
 }
 
-simdjson_really_inline bool simdjson_result<dom::element>::is_array() const noexcept {
+simdjson_inline bool simdjson_result<dom::element>::is_array() const noexcept {
   return !error() && first.is_array();
 }
-simdjson_really_inline bool simdjson_result<dom::element>::is_object() const noexcept {
+simdjson_inline bool simdjson_result<dom::element>::is_object() const noexcept {
   return !error() && first.is_object();
 }
-simdjson_really_inline bool simdjson_result<dom::element>::is_string() const noexcept {
+simdjson_inline bool simdjson_result<dom::element>::is_string() const noexcept {
   return !error() && first.is_string();
 }
-simdjson_really_inline bool simdjson_result<dom::element>::is_int64() const noexcept {
+simdjson_inline bool simdjson_result<dom::element>::is_int64() const noexcept {
   return !error() && first.is_int64();
 }
-simdjson_really_inline bool simdjson_result<dom::element>::is_uint64() const noexcept {
+simdjson_inline bool simdjson_result<dom::element>::is_uint64() const noexcept {
   return !error() && first.is_uint64();
 }
-simdjson_really_inline bool simdjson_result<dom::element>::is_double() const noexcept {
+simdjson_inline bool simdjson_result<dom::element>::is_double() const noexcept {
   return !error() && first.is_double();
 }
-simdjson_really_inline bool simdjson_result<dom::element>::is_number() const noexcept {
+simdjson_inline bool simdjson_result<dom::element>::is_number() const noexcept {
   return !error() && first.is_number();
 }
-simdjson_really_inline bool simdjson_result<dom::element>::is_bool() const noexcept {
+simdjson_inline bool simdjson_result<dom::element>::is_bool() const noexcept {
   return !error() && first.is_bool();
 }
 
-simdjson_really_inline bool simdjson_result<dom::element>::is_null() const noexcept {
+simdjson_inline bool simdjson_result<dom::element>::is_null() const noexcept {
   return !error() && first.is_null();
 }
 
-simdjson_really_inline simdjson_result<dom::element> simdjson_result<dom::element>::operator[](std::string_view key) const noexcept {
+simdjson_inline simdjson_result<dom::element> simdjson_result<dom::element>::operator[](std::string_view key) const noexcept {
   if (error()) { return error(); }
   return first[key];
 }
-simdjson_really_inline simdjson_result<dom::element> simdjson_result<dom::element>::operator[](const char *key) const noexcept {
+simdjson_inline simdjson_result<dom::element> simdjson_result<dom::element>::operator[](const char *key) const noexcept {
   if (error()) { return error(); }
   return first[key];
 }
-simdjson_really_inline simdjson_result<dom::element> simdjson_result<dom::element>::at_pointer(const std::string_view json_pointer) const noexcept {
+simdjson_inline simdjson_result<dom::element> simdjson_result<dom::element>::at_pointer(const std::string_view json_pointer) const noexcept {
   if (error()) { return error(); }
   return first.at_pointer(json_pointer);
 }
 #ifndef SIMDJSON_DISABLE_DEPRECATED_API
 [[deprecated("For standard compliance, use at_pointer instead, and prefix your pointers with a slash '/', see RFC6901 ")]]
-simdjson_really_inline simdjson_result<dom::element> simdjson_result<dom::element>::at(const std::string_view json_pointer) const noexcept {
+simdjson_inline simdjson_result<dom::element> simdjson_result<dom::element>::at(const std::string_view json_pointer) const noexcept {
 SIMDJSON_PUSH_DISABLE_WARNINGS
 SIMDJSON_DISABLE_DEPRECATED_WARNING
   if (error()) { return error(); }
@@ -7025,51 +7087,51 @@ SIMDJSON_DISABLE_DEPRECATED_WARNING
 SIMDJSON_POP_DISABLE_WARNINGS
 }
 #endif // SIMDJSON_DISABLE_DEPRECATED_API
-simdjson_really_inline simdjson_result<dom::element> simdjson_result<dom::element>::at(size_t index) const noexcept {
+simdjson_inline simdjson_result<dom::element> simdjson_result<dom::element>::at(size_t index) const noexcept {
   if (error()) { return error(); }
   return first.at(index);
 }
-simdjson_really_inline simdjson_result<dom::element> simdjson_result<dom::element>::at_key(std::string_view key) const noexcept {
+simdjson_inline simdjson_result<dom::element> simdjson_result<dom::element>::at_key(std::string_view key) const noexcept {
   if (error()) { return error(); }
   return first.at_key(key);
 }
-simdjson_really_inline simdjson_result<dom::element> simdjson_result<dom::element>::at_key_case_insensitive(std::string_view key) const noexcept {
+simdjson_inline simdjson_result<dom::element> simdjson_result<dom::element>::at_key_case_insensitive(std::string_view key) const noexcept {
   if (error()) { return error(); }
   return first.at_key_case_insensitive(key);
 }
 
 #if SIMDJSON_EXCEPTIONS
 
-simdjson_really_inline simdjson_result<dom::element>::operator bool() const noexcept(false) {
+simdjson_inline simdjson_result<dom::element>::operator bool() const noexcept(false) {
   return get<bool>();
 }
-simdjson_really_inline simdjson_result<dom::element>::operator const char *() const noexcept(false) {
+simdjson_inline simdjson_result<dom::element>::operator const char *() const noexcept(false) {
   return get<const char *>();
 }
-simdjson_really_inline simdjson_result<dom::element>::operator std::string_view() const noexcept(false) {
+simdjson_inline simdjson_result<dom::element>::operator std::string_view() const noexcept(false) {
   return get<std::string_view>();
 }
-simdjson_really_inline simdjson_result<dom::element>::operator uint64_t() const noexcept(false) {
+simdjson_inline simdjson_result<dom::element>::operator uint64_t() const noexcept(false) {
   return get<uint64_t>();
 }
-simdjson_really_inline simdjson_result<dom::element>::operator int64_t() const noexcept(false) {
+simdjson_inline simdjson_result<dom::element>::operator int64_t() const noexcept(false) {
   return get<int64_t>();
 }
-simdjson_really_inline simdjson_result<dom::element>::operator double() const noexcept(false) {
+simdjson_inline simdjson_result<dom::element>::operator double() const noexcept(false) {
   return get<double>();
 }
-simdjson_really_inline simdjson_result<dom::element>::operator dom::array() const noexcept(false) {
+simdjson_inline simdjson_result<dom::element>::operator dom::array() const noexcept(false) {
   return get<dom::array>();
 }
-simdjson_really_inline simdjson_result<dom::element>::operator dom::object() const noexcept(false) {
+simdjson_inline simdjson_result<dom::element>::operator dom::object() const noexcept(false) {
   return get<dom::object>();
 }
 
-simdjson_really_inline dom::array::iterator simdjson_result<dom::element>::begin() const noexcept(false) {
+simdjson_inline dom::array::iterator simdjson_result<dom::element>::begin() const noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first.begin();
 }
-simdjson_really_inline dom::array::iterator simdjson_result<dom::element>::end() const noexcept(false) {
+simdjson_inline dom::array::iterator simdjson_result<dom::element>::end() const noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first.end();
 }
@@ -7081,15 +7143,17 @@ namespace dom {
 //
 // element inline implementation
 //
-simdjson_really_inline element::element() noexcept : tape{} {}
-simdjson_really_inline element::element(const internal::tape_ref &_tape) noexcept : tape{_tape} { }
+simdjson_inline element::element() noexcept : tape{} {}
+simdjson_inline element::element(const internal::tape_ref &_tape) noexcept : tape{_tape} { }
 
 inline element_type element::type() const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   auto tape_type = tape.tape_ref_type();
   return tape_type == internal::tape_type::FALSE_VALUE ? element_type::BOOL : static_cast<element_type>(tape_type);
 }
 
 inline simdjson_result<bool> element::get_bool() const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   if(tape.is_true()) {
     return true;
   } else if(tape.is_false()) {
@@ -7098,6 +7162,7 @@ inline simdjson_result<bool> element::get_bool() const noexcept {
   return INCORRECT_TYPE;
 }
 inline simdjson_result<const char *> element::get_c_str() const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   switch (tape.tape_ref_type()) {
     case internal::tape_type::STRING: {
       return tape.get_c_str();
@@ -7107,6 +7172,7 @@ inline simdjson_result<const char *> element::get_c_str() const noexcept {
   }
 }
 inline simdjson_result<size_t> element::get_string_length() const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   switch (tape.tape_ref_type()) {
     case internal::tape_type::STRING: {
       return tape.get_string_length();
@@ -7116,6 +7182,7 @@ inline simdjson_result<size_t> element::get_string_length() const noexcept {
   }
 }
 inline simdjson_result<std::string_view> element::get_string() const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   switch (tape.tape_ref_type()) {
     case internal::tape_type::STRING:
       return tape.get_string_view();
@@ -7124,6 +7191,7 @@ inline simdjson_result<std::string_view> element::get_string() const noexcept {
   }
 }
 inline simdjson_result<uint64_t> element::get_uint64() const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   if(simdjson_unlikely(!tape.is_uint64())) { // branch rarely taken
     if(tape.is_int64()) {
       int64_t result = tape.next_tape_value<int64_t>();
@@ -7137,6 +7205,7 @@ inline simdjson_result<uint64_t> element::get_uint64() const noexcept {
   return tape.next_tape_value<int64_t>();
 }
 inline simdjson_result<int64_t> element::get_int64() const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   if(simdjson_unlikely(!tape.is_int64())) { // branch rarely taken
     if(tape.is_uint64()) {
       uint64_t result = tape.next_tape_value<uint64_t>();
@@ -7151,6 +7220,7 @@ inline simdjson_result<int64_t> element::get_int64() const noexcept {
   return tape.next_tape_value<int64_t>();
 }
 inline simdjson_result<double> element::get_double() const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   // Performance considerations:
   // 1. Querying tape_ref_type() implies doing a shift, it is fast to just do a straight
   //   comparison.
@@ -7172,6 +7242,7 @@ inline simdjson_result<double> element::get_double() const noexcept {
   return tape.next_tape_value<double>();
 }
 inline simdjson_result<array> element::get_array() const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   switch (tape.tape_ref_type()) {
     case internal::tape_type::START_ARRAY:
       return array(tape);
@@ -7180,6 +7251,7 @@ inline simdjson_result<array> element::get_array() const noexcept {
   }
 }
 inline simdjson_result<object> element::get_object() const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   switch (tape.tape_ref_type()) {
     case internal::tape_type::START_OBJECT:
       return object(tape);
@@ -7189,12 +7261,12 @@ inline simdjson_result<object> element::get_object() const noexcept {
 }
 
 template<typename T>
-simdjson_warn_unused simdjson_really_inline error_code element::get(T &value) const noexcept {
+simdjson_warn_unused simdjson_inline error_code element::get(T &value) const noexcept {
   return get<T>().get(value);
 }
 // An element-specific version prevents recursion with simdjson_result::get<element>(value)
 template<>
-simdjson_warn_unused simdjson_really_inline error_code element::get<element>(element &value) const noexcept {
+simdjson_warn_unused simdjson_inline error_code element::get<element>(element &value) const noexcept {
   value = element(tape);
   return SUCCESS;
 }
@@ -7204,7 +7276,7 @@ inline void element::tie(T &value, error_code &error) && noexcept {
 }
 
 template<typename T>
-simdjson_really_inline bool element::is() const noexcept {
+simdjson_inline bool element::is() const noexcept {
   auto result = get<T>();
   return !result.error();
 }
@@ -7259,6 +7331,7 @@ inline simdjson_result<element> element::operator[](const char *key) const noexc
 }
 
 inline simdjson_result<element> element::at_pointer(std::string_view json_pointer) const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   switch (tape.tape_ref_type()) {
     case internal::tape_type::START_OBJECT:
       return object(tape).at_pointer(json_pointer);
@@ -7294,6 +7367,7 @@ inline simdjson_result<element> element::at_key_case_insensitive(std::string_vie
 }
 
 inline bool element::dump_raw_tape(std::ostream &out) const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   return tape.doc->dump_raw_tape(out);
 }
 
@@ -7422,7 +7496,7 @@ inline void stage1_worker::run(document_stream * ds, dom::parser * stage1, size_
 }
 #endif
 
-simdjson_really_inline document_stream::document_stream(
+simdjson_inline document_stream::document_stream(
   dom::parser &_parser,
   const uint8_t *_buf,
   size_t _len,
@@ -7444,7 +7518,7 @@ simdjson_really_inline document_stream::document_stream(
 #endif
 }
 
-simdjson_really_inline document_stream::document_stream() noexcept
+simdjson_inline document_stream::document_stream() noexcept
   : parser{nullptr},
     buf{nullptr},
     len{0},
@@ -7456,31 +7530,31 @@ simdjson_really_inline document_stream::document_stream() noexcept
 {
 }
 
-simdjson_really_inline document_stream::~document_stream() noexcept {
+simdjson_inline document_stream::~document_stream() noexcept {
 #ifdef SIMDJSON_THREADS_ENABLED
   worker.reset();
 #endif
 }
 
-simdjson_really_inline document_stream::iterator::iterator() noexcept
+simdjson_inline document_stream::iterator::iterator() noexcept
   : stream{nullptr}, finished{true} {
 }
 
-simdjson_really_inline document_stream::iterator document_stream::begin() noexcept {
+simdjson_inline document_stream::iterator document_stream::begin() noexcept {
   start();
   // If there are no documents, we're finished.
   return iterator(this, error == EMPTY);
 }
 
-simdjson_really_inline document_stream::iterator document_stream::end() noexcept {
+simdjson_inline document_stream::iterator document_stream::end() noexcept {
   return iterator(this, true);
 }
 
-simdjson_really_inline document_stream::iterator::iterator(document_stream* _stream, bool is_end) noexcept
+simdjson_inline document_stream::iterator::iterator(document_stream* _stream, bool is_end) noexcept
   : stream{_stream}, finished{is_end} {
 }
 
-simdjson_really_inline document_stream::iterator::reference document_stream::iterator::operator*() noexcept {
+simdjson_inline document_stream::iterator::reference document_stream::iterator::operator*() noexcept {
   // Note that in case of error, we do not yet mark
   // the iterator as "finished": this detection is done
   // in the operator++ function since it is possible
@@ -7490,7 +7564,7 @@ simdjson_really_inline document_stream::iterator::reference document_stream::ite
   return stream->parser->doc.root();
 }
 
-simdjson_really_inline document_stream::iterator& document_stream::iterator::operator++() noexcept {
+simdjson_inline document_stream::iterator& document_stream::iterator::operator++() noexcept {
   // If there is an error, then we want the iterator
   // to be finished, no matter what. (E.g., we do not
   // keep generating documents with errors, or go beyond
@@ -7517,7 +7591,7 @@ simdjson_really_inline document_stream::iterator& document_stream::iterator::ope
   return *this;
 }
 
-simdjson_really_inline bool document_stream::iterator::operator!=(const document_stream::iterator &other) const noexcept {
+simdjson_inline bool document_stream::iterator::operator!=(const document_stream::iterator &other) const noexcept {
   return finished != other.finished;
 }
 
@@ -7548,11 +7622,11 @@ inline void document_stream::start() noexcept {
   next();
 }
 
-simdjson_really_inline size_t document_stream::iterator::current_index() const noexcept {
+simdjson_inline size_t document_stream::iterator::current_index() const noexcept {
   return stream->doc_index;
 }
 
-simdjson_really_inline std::string_view document_stream::iterator::source() const noexcept {
+simdjson_inline std::string_view document_stream::iterator::source() const noexcept {
   const char* start = reinterpret_cast<const char*>(stream->buf) + current_index();
   bool object_or_array = ((*start == '[') || (*start == '{'));
   if(object_or_array) {
@@ -7645,32 +7719,32 @@ inline void document_stream::start_stage1_thread() noexcept {
 
 } // namespace dom
 
-simdjson_really_inline simdjson_result<dom::document_stream>::simdjson_result() noexcept
+simdjson_inline simdjson_result<dom::document_stream>::simdjson_result() noexcept
   : simdjson_result_base() {
 }
-simdjson_really_inline simdjson_result<dom::document_stream>::simdjson_result(error_code error) noexcept
+simdjson_inline simdjson_result<dom::document_stream>::simdjson_result(error_code error) noexcept
   : simdjson_result_base(error) {
 }
-simdjson_really_inline simdjson_result<dom::document_stream>::simdjson_result(dom::document_stream &&value) noexcept
+simdjson_inline simdjson_result<dom::document_stream>::simdjson_result(dom::document_stream &&value) noexcept
   : simdjson_result_base(std::forward<dom::document_stream>(value)) {
 }
 
 #if SIMDJSON_EXCEPTIONS
-simdjson_really_inline dom::document_stream::iterator simdjson_result<dom::document_stream>::begin() noexcept(false) {
+simdjson_inline dom::document_stream::iterator simdjson_result<dom::document_stream>::begin() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first.begin();
 }
-simdjson_really_inline dom::document_stream::iterator simdjson_result<dom::document_stream>::end() noexcept(false) {
+simdjson_inline dom::document_stream::iterator simdjson_result<dom::document_stream>::end() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first.end();
 }
 #else // SIMDJSON_EXCEPTIONS
 #ifndef SIMDJSON_DISABLE_DEPRECATED_API
-simdjson_really_inline dom::document_stream::iterator simdjson_result<dom::document_stream>::begin() noexcept {
+simdjson_inline dom::document_stream::iterator simdjson_result<dom::document_stream>::begin() noexcept {
   first.error = error();
   return first.begin();
 }
-simdjson_really_inline dom::document_stream::iterator simdjson_result<dom::document_stream>::end() noexcept {
+simdjson_inline dom::document_stream::iterator simdjson_result<dom::document_stream>::end() noexcept {
   first.error = error();
   return first.end();
 }
@@ -7848,11 +7922,11 @@ namespace simdjson {
 //
 // simdjson_result<dom::object> inline implementation
 //
-simdjson_really_inline simdjson_result<dom::object>::simdjson_result() noexcept
+simdjson_inline simdjson_result<dom::object>::simdjson_result() noexcept
     : internal::simdjson_result_base<dom::object>() {}
-simdjson_really_inline simdjson_result<dom::object>::simdjson_result(dom::object value) noexcept
+simdjson_inline simdjson_result<dom::object>::simdjson_result(dom::object value) noexcept
     : internal::simdjson_result_base<dom::object>(std::forward<dom::object>(value)) {}
-simdjson_really_inline simdjson_result<dom::object>::simdjson_result(error_code error) noexcept
+simdjson_inline simdjson_result<dom::object>::simdjson_result(error_code error) noexcept
     : internal::simdjson_result_base<dom::object>(error) {}
 
 inline simdjson_result<dom::element> simdjson_result<dom::object>::operator[](std::string_view key) const noexcept {
@@ -7898,15 +7972,18 @@ namespace dom {
 //
 // object inline implementation
 //
-simdjson_really_inline object::object() noexcept : tape{} {}
-simdjson_really_inline object::object(const internal::tape_ref &_tape) noexcept : tape{_tape} { }
+simdjson_inline object::object() noexcept : tape{} {}
+simdjson_inline object::object(const internal::tape_ref &_tape) noexcept : tape{_tape} { }
 inline object::iterator object::begin() const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   return internal::tape_ref(tape.doc, tape.json_index + 1);
 }
 inline object::iterator object::end() const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   return internal::tape_ref(tape.doc, tape.after_element() - 1);
 }
 inline size_t object::size() const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   return tape.scope_count();
 }
 
@@ -7917,6 +7994,7 @@ inline simdjson_result<element> object::operator[](const char *key) const noexce
   return at_key(key);
 }
 inline simdjson_result<element> object::at_pointer(std::string_view json_pointer) const noexcept {
+  SIMDJSON_DEVELOPMENT_ASSERT(tape.usable()); // https://github.com/simdjson/simdjson/issues/1914
   if(json_pointer.empty()) { // an empty string means that we return the current node
       return element(this->tape); // copy the current node
   } else if(json_pointer[0] != '/') { // otherwise there is an error
@@ -7985,7 +8063,7 @@ inline simdjson_result<element> object::at_key_case_insensitive(std::string_view
 //
 // object::iterator inline implementation
 //
-simdjson_really_inline object::iterator::iterator(const internal::tape_ref &_tape) noexcept : tape{_tape} { }
+simdjson_inline object::iterator::iterator(const internal::tape_ref &_tape) noexcept : tape{_tape} { }
 inline const key_value_pair object::iterator::operator*() const noexcept {
   return key_value_pair(key(), value());
 }
@@ -8589,12 +8667,12 @@ namespace dom {
 //
 // parser inline implementation
 //
-simdjson_really_inline parser::parser(size_t max_capacity) noexcept
+simdjson_inline parser::parser(size_t max_capacity) noexcept
   : _max_capacity{max_capacity},
     loaded_bytes(nullptr) {
 }
-simdjson_really_inline parser::parser(parser &&other) noexcept = default;
-simdjson_really_inline parser &parser::operator=(parser &&other) noexcept = default;
+simdjson_inline parser::parser(parser &&other) noexcept = default;
+simdjson_inline parser &parser::operator=(parser &&other) noexcept = default;
 
 inline bool parser::is_valid() const noexcept { return valid; }
 inline int parser::get_error_code() const noexcept { return error; }
@@ -8616,7 +8694,13 @@ inline simdjson_result<size_t> parser::read_file(const std::string &path) noexce
   }
 
   // Get the file size
-  if(std::fseek(fp, 0, SEEK_END) < 0) {
+  int ret;
+#if defined(SIMDJSON_VISUAL_STUDIO) && !SIMDJSON_IS_32BITS
+  ret = _fseeki64(fp, 0, SEEK_END);
+#else
+  ret = std::fseek(fp, 0, SEEK_END);
+#endif // _WIN64
+  if(ret < 0) {
     std::fclose(fp);
     return IO_ERROR;
   }
@@ -8692,13 +8776,13 @@ inline simdjson_result<element> parser::parse_into_document(document& provided_d
   return provided_doc.root();
 }
 
-simdjson_really_inline simdjson_result<element> parser::parse_into_document(document& provided_doc, const char *buf, size_t len, bool realloc_if_needed) & noexcept {
+simdjson_inline simdjson_result<element> parser::parse_into_document(document& provided_doc, const char *buf, size_t len, bool realloc_if_needed) & noexcept {
   return parse_into_document(provided_doc, reinterpret_cast<const uint8_t *>(buf), len, realloc_if_needed);
 }
-simdjson_really_inline simdjson_result<element> parser::parse_into_document(document& provided_doc, const std::string &s) & noexcept {
+simdjson_inline simdjson_result<element> parser::parse_into_document(document& provided_doc, const std::string &s) & noexcept {
   return parse_into_document(provided_doc, s.data(), s.length(), s.capacity() - s.length() < SIMDJSON_PADDING);
 }
-simdjson_really_inline simdjson_result<element> parser::parse_into_document(document& provided_doc, const padded_string &s) & noexcept {
+simdjson_inline simdjson_result<element> parser::parse_into_document(document& provided_doc, const padded_string &s) & noexcept {
   return parse_into_document(provided_doc, s.data(), s.length(), false);
 }
 
@@ -8707,15 +8791,18 @@ inline simdjson_result<element> parser::parse(const uint8_t *buf, size_t len, bo
   return parse_into_document(doc, buf, len, realloc_if_needed);
 }
 
-simdjson_really_inline simdjson_result<element> parser::parse(const char *buf, size_t len, bool realloc_if_needed) & noexcept {
+simdjson_inline simdjson_result<element> parser::parse(const char *buf, size_t len, bool realloc_if_needed) & noexcept {
   return parse(reinterpret_cast<const uint8_t *>(buf), len, realloc_if_needed);
 }
-simdjson_really_inline simdjson_result<element> parser::parse(const std::string &s) & noexcept {
+simdjson_inline simdjson_result<element> parser::parse(const std::string &s) & noexcept {
   return parse(s.data(), s.length(), s.capacity() - s.length() < SIMDJSON_PADDING);
 }
-simdjson_really_inline simdjson_result<element> parser::parse(const padded_string &s) & noexcept {
+simdjson_inline simdjson_result<element> parser::parse(const padded_string &s) & noexcept {
   return parse(s.data(), s.length(), false);
 }
+simdjson_inline simdjson_result<element> parser::parse(const padded_string_view &v) & noexcept {
+  return parse(v.data(), v.length(), false);
+}
 
 inline simdjson_result<document_stream> parser::parse_many(const uint8_t *buf, size_t len, size_t batch_size) noexcept {
   if(batch_size < MINIMAL_BATCH_SIZE) { batch_size = MINIMAL_BATCH_SIZE; }
@@ -8731,13 +8818,13 @@ inline simdjson_result<document_stream> parser::parse_many(const padded_string &
   return parse_many(s.data(), s.length(), batch_size);
 }
 
-simdjson_really_inline size_t parser::capacity() const noexcept {
+simdjson_inline size_t parser::capacity() const noexcept {
   return implementation ? implementation->capacity() : 0;
 }
-simdjson_really_inline size_t parser::max_capacity() const noexcept {
+simdjson_inline size_t parser::max_capacity() const noexcept {
   return _max_capacity;
 }
-simdjson_really_inline size_t parser::max_depth() const noexcept {
+simdjson_inline size_t parser::max_depth() const noexcept {
   return implementation ? implementation->max_depth() : DEFAULT_MAX_DEPTH;
 }
 
@@ -8790,7 +8877,7 @@ inline error_code parser::ensure_capacity(document& target_document, size_t desi
   return SUCCESS;
 }
 
-simdjson_really_inline void parser::set_max_capacity(size_t max_capacity) noexcept {
+simdjson_inline void parser::set_max_capacity(size_t max_capacity) noexcept {
   if(max_capacity < MINIMAL_DOCUMENT_CAPACITY) {
     _max_capacity = max_capacity;
   } else {
@@ -8815,39 +8902,41 @@ namespace internal {
 //
 // tape_ref inline implementation
 //
-simdjson_really_inline tape_ref::tape_ref() noexcept : doc{nullptr}, json_index{0} {}
-simdjson_really_inline tape_ref::tape_ref(const dom::document *_doc, size_t _json_index) noexcept : doc{_doc}, json_index{_json_index} {}
+simdjson_inline tape_ref::tape_ref() noexcept : doc{nullptr}, json_index{0} {}
+simdjson_inline tape_ref::tape_ref(const dom::document *_doc, size_t _json_index) noexcept : doc{_doc}, json_index{_json_index} {}
 
 
-simdjson_really_inline bool tape_ref::is_document_root() const noexcept {
+simdjson_inline bool tape_ref::is_document_root() const noexcept {
   return json_index == 1; // should we ever change the structure of the tape, this should get updated.
 }
-
+simdjson_inline bool tape_ref::usable() const noexcept {
+  return doc != nullptr; // when the document pointer is null, this tape_ref is uninitialized (should not be accessed).
+}
 // Some value types have a specific on-tape word value. It can be faster
 // to check the type by doing a word-to-word comparison instead of extracting the
 // most significant 8 bits.
 
-simdjson_really_inline bool tape_ref::is_double() const noexcept {
+simdjson_inline bool tape_ref::is_double() const noexcept {
   constexpr uint64_t tape_double = uint64_t(tape_type::DOUBLE)<<56;
   return doc->tape[json_index] == tape_double;
 }
-simdjson_really_inline bool tape_ref::is_int64() const noexcept {
+simdjson_inline bool tape_ref::is_int64() const noexcept {
   constexpr uint64_t tape_int64 = uint64_t(tape_type::INT64)<<56;
   return doc->tape[json_index] == tape_int64;
 }
-simdjson_really_inline bool tape_ref::is_uint64() const noexcept {
+simdjson_inline bool tape_ref::is_uint64() const noexcept {
   constexpr uint64_t tape_uint64 = uint64_t(tape_type::UINT64)<<56;
   return doc->tape[json_index] == tape_uint64;
 }
-simdjson_really_inline bool tape_ref::is_false() const noexcept {
+simdjson_inline bool tape_ref::is_false() const noexcept {
   constexpr uint64_t tape_false = uint64_t(tape_type::FALSE_VALUE)<<56;
   return doc->tape[json_index] == tape_false;
 }
-simdjson_really_inline bool tape_ref::is_true() const noexcept {
+simdjson_inline bool tape_ref::is_true() const noexcept {
   constexpr uint64_t tape_true = uint64_t(tape_type::TRUE_VALUE)<<56;
   return doc->tape[json_index] == tape_true;
 }
-simdjson_really_inline bool tape_ref::is_null_on_tape() const noexcept {
+simdjson_inline bool tape_ref::is_null_on_tape() const noexcept {
   constexpr uint64_t tape_null = uint64_t(tape_type::NULL_VALUE)<<56;
   return doc->tape[json_index] == tape_null;
 }
@@ -8865,21 +8954,21 @@ inline size_t tape_ref::after_element() const noexcept {
       return json_index + 1;
   }
 }
-simdjson_really_inline tape_type tape_ref::tape_ref_type() const noexcept {
+simdjson_inline tape_type tape_ref::tape_ref_type() const noexcept {
   return static_cast<tape_type>(doc->tape[json_index] >> 56);
 }
-simdjson_really_inline uint64_t internal::tape_ref::tape_value() const noexcept {
+simdjson_inline uint64_t internal::tape_ref::tape_value() const noexcept {
   return doc->tape[json_index] & internal::JSON_VALUE_MASK;
 }
-simdjson_really_inline uint32_t internal::tape_ref::matching_brace_index() const noexcept {
+simdjson_inline uint32_t internal::tape_ref::matching_brace_index() const noexcept {
   return uint32_t(doc->tape[json_index]);
 }
-simdjson_really_inline uint32_t internal::tape_ref::scope_count() const noexcept {
+simdjson_inline uint32_t internal::tape_ref::scope_count() const noexcept {
   return uint32_t((doc->tape[json_index] >> 32) & internal::JSON_COUNT_MASK);
 }
 
 template<typename T>
-simdjson_really_inline T tape_ref::next_tape_value() const noexcept {
+simdjson_inline T tape_ref::next_tape_value() const noexcept {
   static_assert(sizeof(T) == sizeof(uint64_t), "next_tape_value() template parameter must be 64-bit");
   // Though the following is tempting...
   //  return *reinterpret_cast<const T*>(&doc->tape[json_index + 1]);
@@ -8890,14 +8979,14 @@ simdjson_really_inline T tape_ref::next_tape_value() const noexcept {
   return x;
 }
 
-simdjson_really_inline uint32_t internal::tape_ref::get_string_length() const noexcept {
+simdjson_inline uint32_t internal::tape_ref::get_string_length() const noexcept {
   size_t string_buf_index = size_t(tape_value());
   uint32_t len;
   std::memcpy(&len, &doc->string_buf[string_buf_index], sizeof(len));
   return len;
 }
 
-simdjson_really_inline const char * internal::tape_ref::get_c_str() const noexcept {
+simdjson_inline const char * internal::tape_ref::get_c_str() const noexcept {
   size_t string_buf_index = size_t(tape_value());
   return reinterpret_cast<const char *>(&doc->string_buf[string_buf_index + sizeof(uint32_t)]);
 }
@@ -9023,19 +9112,19 @@ namespace internal {
  * Minifier/formatter code.
  **/
 
-simdjson_really_inline void mini_formatter::number(uint64_t x) {
+simdjson_inline void mini_formatter::number(uint64_t x) {
   char number_buffer[24];
   char *newp = fast_itoa(number_buffer, x);
   buffer.insert(buffer.end(), number_buffer, newp);
 }
 
-simdjson_really_inline void mini_formatter::number(int64_t x) {
+simdjson_inline void mini_formatter::number(int64_t x) {
   char number_buffer[24];
   char *newp = fast_itoa(number_buffer, x);
   buffer.insert(buffer.end(), number_buffer, newp);
 }
 
-simdjson_really_inline void mini_formatter::number(double x) {
+simdjson_inline void mini_formatter::number(double x) {
   char number_buffer[24];
   // Currently, passing the nullptr to the second argument is
   // safe because our implementation does not check the second
@@ -9044,31 +9133,31 @@ simdjson_really_inline void mini_formatter::number(double x) {
   buffer.insert(buffer.end(), number_buffer, newp);
 }
 
-simdjson_really_inline void mini_formatter::start_array() { one_char('['); }
-simdjson_really_inline void mini_formatter::end_array() { one_char(']'); }
-simdjson_really_inline void mini_formatter::start_object() { one_char('{'); }
-simdjson_really_inline void mini_formatter::end_object() { one_char('}'); }
-simdjson_really_inline void mini_formatter::comma() { one_char(','); }
+simdjson_inline void mini_formatter::start_array() { one_char('['); }
+simdjson_inline void mini_formatter::end_array() { one_char(']'); }
+simdjson_inline void mini_formatter::start_object() { one_char('{'); }
+simdjson_inline void mini_formatter::end_object() { one_char('}'); }
+simdjson_inline void mini_formatter::comma() { one_char(','); }
 
 
-simdjson_really_inline void mini_formatter::true_atom() {
+simdjson_inline void mini_formatter::true_atom() {
   const char * s = "true";
   buffer.insert(buffer.end(), s, s + 4);
 }
-simdjson_really_inline void mini_formatter::false_atom() {
+simdjson_inline void mini_formatter::false_atom() {
   const char * s = "false";
   buffer.insert(buffer.end(), s, s + 5);
 }
-simdjson_really_inline void mini_formatter::null_atom() {
+simdjson_inline void mini_formatter::null_atom() {
   const char * s = "null";
   buffer.insert(buffer.end(), s, s + 4);
 }
-simdjson_really_inline void mini_formatter::one_char(char c) { buffer.push_back(c); }
-simdjson_really_inline void mini_formatter::key(std::string_view unescaped) {
+simdjson_inline void mini_formatter::one_char(char c) { buffer.push_back(c); }
+simdjson_inline void mini_formatter::key(std::string_view unescaped) {
   string(unescaped);
   one_char(':');
 }
-simdjson_really_inline void mini_formatter::string(std::string_view unescaped) {
+simdjson_inline void mini_formatter::string(std::string_view unescaped) {
   one_char('\"');
   size_t i = 0;
   // Fast path for the case where we have no control character, no ", and no backslash.
@@ -9151,7 +9240,7 @@ inline void mini_formatter::clear() {
   buffer.clear();
 }
 
-simdjson_really_inline std::string_view mini_formatter::str() const {
+simdjson_inline std::string_view mini_formatter::str() const {
   return std::string_view(buffer.data(), buffer.size());
 }
 
@@ -9322,18 +9411,18 @@ inline void string_builder<serializer>::append(simdjson::dom::array value) {
 }
 
 template <class serializer>
-simdjson_really_inline void string_builder<serializer>::append(simdjson::dom::key_value_pair kv) {
+simdjson_inline void string_builder<serializer>::append(simdjson::dom::key_value_pair kv) {
   format.key(kv.key);
   append(kv.value);
 }
 
 template <class serializer>
-simdjson_really_inline void string_builder<serializer>::clear() {
+simdjson_inline void string_builder<serializer>::clear() {
   format.clear();
 }
 
 template <class serializer>
-simdjson_really_inline std::string_view string_builder<serializer>::str() const {
+simdjson_inline std::string_view string_builder<serializer>::str() const {
   return format.str();
 }
 
@@ -9592,7 +9681,7 @@ using namespace simdjson::dom;
 
 class implementation final : public simdjson::implementation {
 public:
-  simdjson_really_inline implementation() : simdjson::implementation("arm64", "ARM NEON", internal::instruction_set::NEON) {}
+  simdjson_inline implementation() : simdjson::implementation("arm64", "ARM NEON", internal::instruction_set::NEON) {}
   simdjson_warn_unused error_code create_dom_parser_implementation(
     size_t capacity,
     size_t max_length,
@@ -9650,10 +9739,11 @@ class dom_parser_implementation final : public internal::dom_parser_implementati
   simdjson_warn_unused error_code stage1(const uint8_t *buf, size_t len, stage1_mode partial) noexcept final;
   simdjson_warn_unused error_code stage2(dom::document &doc) noexcept final;
   simdjson_warn_unused error_code stage2_next(dom::document &doc) noexcept final;
+  simdjson_warn_unused uint8_t *parse_string(const uint8_t *src, uint8_t *dst) const noexcept final;
   inline simdjson_warn_unused error_code set_capacity(size_t capacity) noexcept final;
   inline simdjson_warn_unused error_code set_max_depth(size_t max_depth) noexcept final;
 private:
-  simdjson_really_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
+  simdjson_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
 
 };
 
@@ -9702,6 +9792,8 @@ inline simdjson_warn_unused error_code dom_parser_implementation::set_max_depth(
 // you use visual studio or other compilers.
 #include <arm_neon.h>
 
+static_assert(sizeof(uint8x16_t) <= simdjson::SIMDJSON_PADDING, "insufficient padding for arm64");
+
 #endif //  SIMDJSON_ARM64_INTRINSICS_H
 /* end file include/simdjson/arm64/intrinsics.h */
 /* begin file include/simdjson/arm64/bitmanipulation.h */
@@ -9716,7 +9808,7 @@ namespace {
 // but the algorithms do not end up using the returned value.
 // Sadly, sanitizers are not smart enough to figure it out.
 SIMDJSON_NO_SANITIZE_UNDEFINED
-simdjson_really_inline int trailing_zeroes(uint64_t input_num) {
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
   unsigned long ret;
   // Search the mask data from least significant bit (LSB)
@@ -9729,12 +9821,12 @@ simdjson_really_inline int trailing_zeroes(uint64_t input_num) {
 }
 
 /* result might be undefined when input_num is zero */
-simdjson_really_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
   return input_num & (input_num-1);
 }
 
 /* result might be undefined when input_num is zero */
-simdjson_really_inline int leading_zeroes(uint64_t input_num) {
+simdjson_inline int leading_zeroes(uint64_t input_num) {
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
   unsigned long leading_zero = 0;
   // Search the mask data from most significant bit (MSB)
@@ -9749,7 +9841,7 @@ simdjson_really_inline int leading_zeroes(uint64_t input_num) {
 }
 
 /* result might be undefined when input_num is zero */
-simdjson_really_inline int count_ones(uint64_t input_num) {
+simdjson_inline int count_ones(uint64_t input_num) {
    return vaddv_u8(vcnt_u8(vcreate_u8(input_num)));
 }
 
@@ -9768,7 +9860,7 @@ simdjson_really_inline int count_ones(uint64_t input_num) {
 #define SIMDJSON_PREFER_REVERSE_BITS 1
 
 /* reverse the bits */
-simdjson_really_inline uint64_t reverse_bits(uint64_t input_num) {
+simdjson_inline uint64_t reverse_bits(uint64_t input_num) {
   uint64_t rev_bits;
   __asm("rbit %0, %1" : "=r"(rev_bits) : "r"(input_num));
   return rev_bits;
@@ -9781,13 +9873,13 @@ simdjson_really_inline uint64_t reverse_bits(uint64_t input_num) {
  * of such undefined behavior is never used.
  **/
 SIMDJSON_NO_SANITIZE_UNDEFINED
-simdjson_really_inline uint64_t zero_leading_bit(uint64_t rev_bits, int leading_zeroes) {
+simdjson_inline uint64_t zero_leading_bit(uint64_t rev_bits, int leading_zeroes) {
   return rev_bits ^ (uint64_t(0x8000000000000000) >> leading_zeroes);
 }
 
 #endif
 
-simdjson_really_inline bool add_overflow(uint64_t value1, uint64_t value2, uint64_t *result) {
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2, uint64_t *result) {
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
   *result = value1 + value2;
   return *result < value1;
@@ -9816,7 +9908,7 @@ namespace {
 //
 // For example, prefix_xor(00100100) == 00011100
 //
-simdjson_really_inline uint64_t prefix_xor(uint64_t bitmask) {
+simdjson_inline uint64_t prefix_xor(uint64_t bitmask) {
   /////////////
   // We could do this with PMULL, but it is apparently slow.
   //
@@ -9873,7 +9965,7 @@ namespace {
  * You should not use this function except for compile-time constants:
  * it is not efficient.
  */
-simdjson_really_inline uint8x16_t make_uint8x16_t(uint8_t x1,  uint8_t x2,  uint8_t x3,  uint8_t x4,
+simdjson_inline uint8x16_t make_uint8x16_t(uint8_t x1,  uint8_t x2,  uint8_t x3,  uint8_t x4,
                                          uint8_t x5,  uint8_t x6,  uint8_t x7,  uint8_t x8,
                                          uint8_t x9,  uint8_t x10, uint8_t x11, uint8_t x12,
                                          uint8_t x13, uint8_t x14, uint8_t x15, uint8_t x16) {
@@ -9902,7 +9994,7 @@ simdjson_really_inline uint8x16_t make_uint8x16_t(uint8_t x1,  uint8_t x2,  uint
   return x;
 }
 
-simdjson_really_inline uint8x8_t make_uint8x8_t(uint8_t x1,  uint8_t x2,  uint8_t x3,  uint8_t x4,
+simdjson_inline uint8x8_t make_uint8x8_t(uint8_t x1,  uint8_t x2,  uint8_t x3,  uint8_t x4,
                                          uint8_t x5,  uint8_t x6,  uint8_t x7,  uint8_t x8) {
   uint8x8_t x{};
   x = vset_lane_u8(x1, x, 0);
@@ -9917,7 +10009,7 @@ simdjson_really_inline uint8x8_t make_uint8x8_t(uint8_t x1,  uint8_t x2,  uint8_
 }
 
 // We have to do the same work for make_int8x16_t
-simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x3,  int8_t x4,
+simdjson_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x3,  int8_t x4,
                                        int8_t x5,  int8_t x6,  int8_t x7,  int8_t x8,
                                        int8_t x9,  int8_t x10, int8_t x11, int8_t x12,
                                        int8_t x13, int8_t x14, int8_t x15, int8_t x16) {
@@ -9963,24 +10055,24 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
     static const int SIZE = sizeof(value);
 
     // Conversion from/to SIMD register
-    simdjson_really_inline base_u8(const uint8x16_t _value) : value(_value) {}
-    simdjson_really_inline operator const uint8x16_t&() const { return this->value; }
-    simdjson_really_inline operator uint8x16_t&() { return this->value; }
+    simdjson_inline base_u8(const uint8x16_t _value) : value(_value) {}
+    simdjson_inline operator const uint8x16_t&() const { return this->value; }
+    simdjson_inline operator uint8x16_t&() { return this->value; }
 
     // Bit operations
-    simdjson_really_inline simd8<T> operator|(const simd8<T> other) const { return vorrq_u8(*this, other); }
-    simdjson_really_inline simd8<T> operator&(const simd8<T> other) const { return vandq_u8(*this, other); }
-    simdjson_really_inline simd8<T> operator^(const simd8<T> other) const { return veorq_u8(*this, other); }
-    simdjson_really_inline simd8<T> bit_andnot(const simd8<T> other) const { return vbicq_u8(*this, other); }
-    simdjson_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
-    simdjson_really_inline simd8<T>& operator|=(const simd8<T> other) { auto this_cast = static_cast<simd8<T>*>(this); *this_cast = *this_cast | other; return *this_cast; }
-    simdjson_really_inline simd8<T>& operator&=(const simd8<T> other) { auto this_cast = static_cast<simd8<T>*>(this); *this_cast = *this_cast & other; return *this_cast; }
-    simdjson_really_inline simd8<T>& operator^=(const simd8<T> other) { auto this_cast = static_cast<simd8<T>*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+    simdjson_inline simd8<T> operator|(const simd8<T> other) const { return vorrq_u8(*this, other); }
+    simdjson_inline simd8<T> operator&(const simd8<T> other) const { return vandq_u8(*this, other); }
+    simdjson_inline simd8<T> operator^(const simd8<T> other) const { return veorq_u8(*this, other); }
+    simdjson_inline simd8<T> bit_andnot(const simd8<T> other) const { return vbicq_u8(*this, other); }
+    simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+    simdjson_inline simd8<T>& operator|=(const simd8<T> other) { auto this_cast = static_cast<simd8<T>*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline simd8<T>& operator&=(const simd8<T> other) { auto this_cast = static_cast<simd8<T>*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline simd8<T>& operator^=(const simd8<T> other) { auto this_cast = static_cast<simd8<T>*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
 
-    friend simdjson_really_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) { return vceqq_u8(lhs, rhs); }
+    friend simdjson_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) { return vceqq_u8(lhs, rhs); }
 
     template<int N=1>
-    simdjson_really_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+    simdjson_inline simd8<T> prev(const simd8<T> prev_chunk) const {
       return vextq_u8(prev_chunk, *this, 16 - N);
     }
   };
@@ -9991,17 +10083,17 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
     typedef uint16_t bitmask_t;
     typedef uint32_t bitmask2_t;
 
-    static simdjson_really_inline simd8<bool> splat(bool _value) { return vmovq_n_u8(uint8_t(-(!!_value))); }
+    static simdjson_inline simd8<bool> splat(bool _value) { return vmovq_n_u8(uint8_t(-(!!_value))); }
 
-    simdjson_really_inline simd8(const uint8x16_t _value) : base_u8<bool>(_value) {}
+    simdjson_inline simd8(const uint8x16_t _value) : base_u8<bool>(_value) {}
     // False constructor
-    simdjson_really_inline simd8() : simd8(vdupq_n_u8(0)) {}
+    simdjson_inline simd8() : simd8(vdupq_n_u8(0)) {}
     // Splat constructor
-    simdjson_really_inline simd8(bool _value) : simd8(splat(_value)) {}
+    simdjson_inline simd8(bool _value) : simd8(splat(_value)) {}
 
     // We return uint32_t instead of uint16_t because that seems to be more efficient for most
     // purposes (cutting it down to uint16_t costs performance in some compilers).
-    simdjson_really_inline uint32_t to_bitmask() const {
+    simdjson_inline uint32_t to_bitmask() const {
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
       const uint8x16_t bit_mask =  make_uint8x16_t(0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
                                                    0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80);
@@ -10015,26 +10107,26 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
       tmp = vpaddq_u8(tmp, tmp);
       return vgetq_lane_u16(vreinterpretq_u16_u8(tmp), 0);
     }
-    simdjson_really_inline bool any() const { return vmaxvq_u8(*this) != 0; }
+    simdjson_inline bool any() const { return vmaxvq_u8(*this) != 0; }
   };
 
   // Unsigned bytes
   template<>
   struct simd8<uint8_t>: base_u8<uint8_t> {
-    static simdjson_really_inline uint8x16_t splat(uint8_t _value) { return vmovq_n_u8(_value); }
-    static simdjson_really_inline uint8x16_t zero() { return vdupq_n_u8(0); }
-    static simdjson_really_inline uint8x16_t load(const uint8_t* values) { return vld1q_u8(values); }
+    static simdjson_inline uint8x16_t splat(uint8_t _value) { return vmovq_n_u8(_value); }
+    static simdjson_inline uint8x16_t zero() { return vdupq_n_u8(0); }
+    static simdjson_inline uint8x16_t load(const uint8_t* values) { return vld1q_u8(values); }
 
-    simdjson_really_inline simd8(const uint8x16_t _value) : base_u8<uint8_t>(_value) {}
+    simdjson_inline simd8(const uint8x16_t _value) : base_u8<uint8_t>(_value) {}
     // Zero constructor
-    simdjson_really_inline simd8() : simd8(zero()) {}
+    simdjson_inline simd8() : simd8(zero()) {}
     // Array constructor
-    simdjson_really_inline simd8(const uint8_t values[16]) : simd8(load(values)) {}
+    simdjson_inline simd8(const uint8_t values[16]) : simd8(load(values)) {}
     // Splat constructor
-    simdjson_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
     // Member-by-member initialization
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
-    simdjson_really_inline simd8(
+    simdjson_inline simd8(
       uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
       uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
     ) : simd8(make_uint8x16_t(
@@ -10042,7 +10134,7 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
       v8, v9, v10,v11,v12,v13,v14,v15
     )) {}
 #else
-    simdjson_really_inline simd8(
+    simdjson_inline simd8(
       uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
       uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
     ) : simd8(uint8x16_t{
@@ -10052,7 +10144,7 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
 #endif
 
     // Repeat 16 values as many times as necessary (usually for lookup tables)
-    simdjson_really_inline static simd8<uint8_t> repeat_16(
+    simdjson_inline static simd8<uint8_t> repeat_16(
       uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
       uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
     ) {
@@ -10063,44 +10155,44 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
     }
 
     // Store to array
-    simdjson_really_inline void store(uint8_t dst[16]) const { return vst1q_u8(dst, *this); }
+    simdjson_inline void store(uint8_t dst[16]) const { return vst1q_u8(dst, *this); }
 
     // Saturated math
-    simdjson_really_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return vqaddq_u8(*this, other); }
-    simdjson_really_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return vqsubq_u8(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return vqaddq_u8(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return vqsubq_u8(*this, other); }
 
     // Addition/subtraction are the same for signed and unsigned
-    simdjson_really_inline simd8<uint8_t> operator+(const simd8<uint8_t> other) const { return vaddq_u8(*this, other); }
-    simdjson_really_inline simd8<uint8_t> operator-(const simd8<uint8_t> other) const { return vsubq_u8(*this, other); }
-    simdjson_really_inline simd8<uint8_t>& operator+=(const simd8<uint8_t> other) { *this = *this + other; return *this; }
-    simdjson_really_inline simd8<uint8_t>& operator-=(const simd8<uint8_t> other) { *this = *this - other; return *this; }
+    simdjson_inline simd8<uint8_t> operator+(const simd8<uint8_t> other) const { return vaddq_u8(*this, other); }
+    simdjson_inline simd8<uint8_t> operator-(const simd8<uint8_t> other) const { return vsubq_u8(*this, other); }
+    simdjson_inline simd8<uint8_t>& operator+=(const simd8<uint8_t> other) { *this = *this + other; return *this; }
+    simdjson_inline simd8<uint8_t>& operator-=(const simd8<uint8_t> other) { *this = *this - other; return *this; }
 
     // Order-specific operations
-    simdjson_really_inline uint8_t max_val() const { return vmaxvq_u8(*this); }
-    simdjson_really_inline uint8_t min_val() const { return vminvq_u8(*this); }
-    simdjson_really_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return vmaxq_u8(*this, other); }
-    simdjson_really_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return vminq_u8(*this, other); }
-    simdjson_really_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return vcleq_u8(*this, other); }
-    simdjson_really_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return vcgeq_u8(*this, other); }
-    simdjson_really_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return vcltq_u8(*this, other); }
-    simdjson_really_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return vcgtq_u8(*this, other); }
+    simdjson_inline uint8_t max_val() const { return vmaxvq_u8(*this); }
+    simdjson_inline uint8_t min_val() const { return vminvq_u8(*this); }
+    simdjson_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return vmaxq_u8(*this, other); }
+    simdjson_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return vminq_u8(*this, other); }
+    simdjson_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return vcleq_u8(*this, other); }
+    simdjson_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return vcgeq_u8(*this, other); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return vcltq_u8(*this, other); }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return vcgtq_u8(*this, other); }
     // Same as >, but instead of guaranteeing all 1's == true, false = 0 and true = nonzero. For ARM, returns all 1's.
-    simdjson_really_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return simd8<uint8_t>(*this > other); }
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return simd8<uint8_t>(*this > other); }
     // Same as <, but instead of guaranteeing all 1's == true, false = 0 and true = nonzero. For ARM, returns all 1's.
-    simdjson_really_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return simd8<uint8_t>(*this < other); }
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return simd8<uint8_t>(*this < other); }
 
     // Bit-specific operations
-    simdjson_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return vtstq_u8(*this, bits); }
-    simdjson_really_inline bool any_bits_set_anywhere() const { return this->max_val() != 0; }
-    simdjson_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return (*this & bits).any_bits_set_anywhere(); }
+    simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return vtstq_u8(*this, bits); }
+    simdjson_inline bool any_bits_set_anywhere() const { return this->max_val() != 0; }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return (*this & bits).any_bits_set_anywhere(); }
     template<int N>
-    simdjson_really_inline simd8<uint8_t> shr() const { return vshrq_n_u8(*this, N); }
+    simdjson_inline simd8<uint8_t> shr() const { return vshrq_n_u8(*this, N); }
     template<int N>
-    simdjson_really_inline simd8<uint8_t> shl() const { return vshlq_n_u8(*this, N); }
+    simdjson_inline simd8<uint8_t> shl() const { return vshlq_n_u8(*this, N); }
 
     // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
     template<typename L>
-    simdjson_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
       return lookup_table.apply_lookup_16_to(*this);
     }
 
@@ -10113,7 +10205,7 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
     // signature simd8<L> compress(uint16_t mask) would be
     // sensible, but the AVX ISA makes this kind of approach difficult.
     template<typename L>
-    simdjson_really_inline void compress(uint16_t mask, L * output) const {
+    simdjson_inline void compress(uint16_t mask, L * output) const {
       using internal::thintable_epi8;
       using internal::BitsSetTable256mul2;
       using internal::pshufb_combine_table;
@@ -10150,7 +10242,7 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
     // Copies all bytes corresponding to a 0 in the low half of the mask (interpreted as a
     // bitset) to output1, then those corresponding to a 0 in the high half to output2.
     template<typename L>
-    simdjson_really_inline void compress_halves(uint16_t mask, L *output1, L *output2) const {
+    simdjson_inline void compress_halves(uint16_t mask, L *output1, L *output2) const {
       using internal::thintable_epi8;
       uint8_t mask1 = uint8_t(mask); // least significant 8 bits
       uint8_t mask2 = uint8_t(mask >> 8); // most significant 8 bits
@@ -10169,7 +10261,7 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
     }
 
     template<typename L>
-    simdjson_really_inline simd8<L> lookup_16(
+    simdjson_inline simd8<L> lookup_16(
         L replace0,  L replace1,  L replace2,  L replace3,
         L replace4,  L replace5,  L replace6,  L replace7,
         L replace8,  L replace9,  L replace10, L replace11,
@@ -10183,7 +10275,7 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
     }
 
     template<typename T>
-    simdjson_really_inline simd8<uint8_t> apply_lookup_16_to(const simd8<T> original) {
+    simdjson_inline simd8<uint8_t> apply_lookup_16_to(const simd8<T> original) {
       return vqtbl1q_u8(*this, simd8<uint8_t>(original));
     }
   };
@@ -10193,24 +10285,24 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
   struct simd8<int8_t> {
     int8x16_t value;
 
-    static simdjson_really_inline simd8<int8_t> splat(int8_t _value) { return vmovq_n_s8(_value); }
-    static simdjson_really_inline simd8<int8_t> zero() { return vdupq_n_s8(0); }
-    static simdjson_really_inline simd8<int8_t> load(const int8_t values[16]) { return vld1q_s8(values); }
+    static simdjson_inline simd8<int8_t> splat(int8_t _value) { return vmovq_n_s8(_value); }
+    static simdjson_inline simd8<int8_t> zero() { return vdupq_n_s8(0); }
+    static simdjson_inline simd8<int8_t> load(const int8_t values[16]) { return vld1q_s8(values); }
 
     // Conversion from/to SIMD register
-    simdjson_really_inline simd8(const int8x16_t _value) : value{_value} {}
-    simdjson_really_inline operator const int8x16_t&() const { return this->value; }
-    simdjson_really_inline operator int8x16_t&() { return this->value; }
+    simdjson_inline simd8(const int8x16_t _value) : value{_value} {}
+    simdjson_inline operator const int8x16_t&() const { return this->value; }
+    simdjson_inline operator int8x16_t&() { return this->value; }
 
     // Zero constructor
-    simdjson_really_inline simd8() : simd8(zero()) {}
+    simdjson_inline simd8() : simd8(zero()) {}
     // Splat constructor
-    simdjson_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
     // Array constructor
-    simdjson_really_inline simd8(const int8_t* values) : simd8(load(values)) {}
+    simdjson_inline simd8(const int8_t* values) : simd8(load(values)) {}
     // Member-by-member initialization
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
-    simdjson_really_inline simd8(
+    simdjson_inline simd8(
       int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3, int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
       int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
     ) : simd8(make_int8x16_t(
@@ -10218,7 +10310,7 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
       v8, v9, v10,v11,v12,v13,v14,v15
     )) {}
 #else
-    simdjson_really_inline simd8(
+    simdjson_inline simd8(
       int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3, int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
       int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
     ) : simd8(int8x16_t{
@@ -10227,7 +10319,7 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
     }) {}
 #endif
     // Repeat 16 values as many times as necessary (usually for lookup tables)
-    simdjson_really_inline static simd8<int8_t> repeat_16(
+    simdjson_inline static simd8<int8_t> repeat_16(
       int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
       int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
     ) {
@@ -10238,7 +10330,7 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
     }
 
     // Store to array
-    simdjson_really_inline void store(int8_t dst[16]) const { return vst1q_s8(dst, *this); }
+    simdjson_inline void store(int8_t dst[16]) const { return vst1q_s8(dst, *this); }
 
     // Explicit conversion to/from unsigned
     //
@@ -10246,35 +10338,35 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
     // In theory, we could check this occurrence with std::same_as and std::enabled_if but it is C++14
     // and relatively ugly and hard to read.
 #ifndef SIMDJSON_REGULAR_VISUAL_STUDIO
-    simdjson_really_inline explicit simd8(const uint8x16_t other): simd8(vreinterpretq_s8_u8(other)) {}
+    simdjson_inline explicit simd8(const uint8x16_t other): simd8(vreinterpretq_s8_u8(other)) {}
 #endif
-    simdjson_really_inline explicit operator simd8<uint8_t>() const { return vreinterpretq_u8_s8(this->value); }
+    simdjson_inline explicit operator simd8<uint8_t>() const { return vreinterpretq_u8_s8(this->value); }
 
     // Math
-    simdjson_really_inline simd8<int8_t> operator+(const simd8<int8_t> other) const { return vaddq_s8(*this, other); }
-    simdjson_really_inline simd8<int8_t> operator-(const simd8<int8_t> other) const { return vsubq_s8(*this, other); }
-    simdjson_really_inline simd8<int8_t>& operator+=(const simd8<int8_t> other) { *this = *this + other; return *this; }
-    simdjson_really_inline simd8<int8_t>& operator-=(const simd8<int8_t> other) { *this = *this - other; return *this; }
+    simdjson_inline simd8<int8_t> operator+(const simd8<int8_t> other) const { return vaddq_s8(*this, other); }
+    simdjson_inline simd8<int8_t> operator-(const simd8<int8_t> other) const { return vsubq_s8(*this, other); }
+    simdjson_inline simd8<int8_t>& operator+=(const simd8<int8_t> other) { *this = *this + other; return *this; }
+    simdjson_inline simd8<int8_t>& operator-=(const simd8<int8_t> other) { *this = *this - other; return *this; }
 
     // Order-sensitive comparisons
-    simdjson_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return vmaxq_s8(*this, other); }
-    simdjson_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return vminq_s8(*this, other); }
-    simdjson_really_inline simd8<bool> operator>(const simd8<int8_t> other) const { return vcgtq_s8(*this, other); }
-    simdjson_really_inline simd8<bool> operator<(const simd8<int8_t> other) const { return vcltq_s8(*this, other); }
-    simdjson_really_inline simd8<bool> operator==(const simd8<int8_t> other) const { return vceqq_s8(*this, other); }
+    simdjson_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return vmaxq_s8(*this, other); }
+    simdjson_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return vminq_s8(*this, other); }
+    simdjson_inline simd8<bool> operator>(const simd8<int8_t> other) const { return vcgtq_s8(*this, other); }
+    simdjson_inline simd8<bool> operator<(const simd8<int8_t> other) const { return vcltq_s8(*this, other); }
+    simdjson_inline simd8<bool> operator==(const simd8<int8_t> other) const { return vceqq_s8(*this, other); }
 
     template<int N=1>
-    simdjson_really_inline simd8<int8_t> prev(const simd8<int8_t> prev_chunk) const {
+    simdjson_inline simd8<int8_t> prev(const simd8<int8_t> prev_chunk) const {
       return vextq_s8(prev_chunk, *this, 16 - N);
     }
 
     // Perform a lookup assuming no value is larger than 16
     template<typename L>
-    simdjson_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
       return lookup_table.apply_lookup_16_to(*this);
     }
     template<typename L>
-    simdjson_really_inline simd8<L> lookup_16(
+    simdjson_inline simd8<L> lookup_16(
         L replace0,  L replace1,  L replace2,  L replace3,
         L replace4,  L replace5,  L replace6,  L replace7,
         L replace8,  L replace9,  L replace10, L replace11,
@@ -10288,7 +10380,7 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
     }
 
     template<typename T>
-    simdjson_really_inline simd8<int8_t> apply_lookup_16_to(const simd8<T> original) {
+    simdjson_inline simd8<int8_t> apply_lookup_16_to(const simd8<T> original) {
       return vqtbl1q_s8(*this, simd8<uint8_t>(original));
     }
   };
@@ -10303,22 +10395,22 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
     simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
     simd8x64() = delete; // no default constructor allowed
 
-    simdjson_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1, const simd8<T> chunk2, const simd8<T> chunk3) : chunks{chunk0, chunk1, chunk2, chunk3} {}
-    simdjson_really_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+16), simd8<T>::load(ptr+32), simd8<T>::load(ptr+48)} {}
+    simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1, const simd8<T> chunk2, const simd8<T> chunk3) : chunks{chunk0, chunk1, chunk2, chunk3} {}
+    simdjson_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+16), simd8<T>::load(ptr+32), simd8<T>::load(ptr+48)} {}
 
-    simdjson_really_inline void store(T ptr[64]) const {
+    simdjson_inline void store(T ptr[64]) const {
       this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
       this->chunks[1].store(ptr+sizeof(simd8<T>)*1);
       this->chunks[2].store(ptr+sizeof(simd8<T>)*2);
       this->chunks[3].store(ptr+sizeof(simd8<T>)*3);
     }
 
-    simdjson_really_inline simd8<T> reduce_or() const {
+    simdjson_inline simd8<T> reduce_or() const {
       return (this->chunks[0] | this->chunks[1]) | (this->chunks[2] | this->chunks[3]);
     }
 
 
-    simdjson_really_inline uint64_t compress(uint64_t mask, T * output) const {
+    simdjson_inline uint64_t compress(uint64_t mask, T * output) const {
       uint64_t popcounts = vget_lane_u64(vreinterpret_u64_u8(vcnt_u8(vcreate_u8(~mask))), 0);
       // compute the prefix sum of the popcounts of each byte
       uint64_t offsets = popcounts * 0x0101010101010101;
@@ -10329,7 +10421,7 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
       return offsets >> 56;
     }
 
-    simdjson_really_inline uint64_t to_bitmask() const {
+    simdjson_inline uint64_t to_bitmask() const {
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
       const uint8x16_t bit_mask = make_uint8x16_t(
         0x01, 0x02, 0x4, 0x8, 0x10, 0x20, 0x40, 0x80,
@@ -10349,7 +10441,7 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
       return vgetq_lane_u64(vreinterpretq_u64_u8(sum0), 0);
     }
 
-    simdjson_really_inline uint64_t eq(const T m) const {
+    simdjson_inline uint64_t eq(const T m) const {
       const simd8<T> mask = simd8<T>::splat(m);
       return  simd8x64<bool>(
         this->chunks[0] == mask,
@@ -10359,7 +10451,7 @@ simdjson_really_inline int8x16_t make_int8x16_t(int8_t x1,  int8_t x2,  int8_t x
       ).to_bitmask();
     }
 
-    simdjson_really_inline uint64_t lteq(const T m) const {
+    simdjson_inline uint64_t lteq(const T m) const {
       const simd8<T> mask = simd8<T>::splat(m);
       return  simd8x64<bool>(
         this->chunks[0] <= mask,
@@ -10386,11 +10478,11 @@ namespace jsoncharutils {
 
 // return non-zero if not a structural or whitespace char
 // zero otherwise
-simdjson_really_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
+simdjson_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
   return internal::structural_or_whitespace_negated[c];
 }
 
-simdjson_really_inline uint32_t is_structural_or_whitespace(uint8_t c) {
+simdjson_inline uint32_t is_structural_or_whitespace(uint8_t c) {
   return internal::structural_or_whitespace[c];
 }
 
@@ -10421,7 +10513,7 @@ static inline uint32_t hex_to_u32_nocheck(
 //
 // Note: we assume that surrogates are treated separately
 //
-simdjson_really_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
+simdjson_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
   if (cp <= 0x7F) {
     c[0] = uint8_t(cp);
     return 1; // ascii
@@ -10453,10 +10545,10 @@ simdjson_really_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
 #ifdef SIMDJSON_IS_32BITS // _umul128 for x86, arm
 // this is a slow emulation routine for 32-bit
 //
-static simdjson_really_inline uint64_t __emulu(uint32_t x, uint32_t y) {
+static simdjson_inline uint64_t __emulu(uint32_t x, uint32_t y) {
   return x * (uint64_t)y;
 }
-static simdjson_really_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+static simdjson_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
   uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
   uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
   uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
@@ -10470,7 +10562,7 @@ static simdjson_really_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64
 
 using internal::value128;
 
-simdjson_really_inline value128 full_multiplication(uint64_t value1, uint64_t value2) {
+simdjson_inline value128 full_multiplication(uint64_t value1, uint64_t value2) {
   value128 answer;
 #if defined(SIMDJSON_REGULAR_VISUAL_STUDIO) || defined(SIMDJSON_IS_32BITS)
 #ifdef _M_ARM64
@@ -10506,13 +10598,13 @@ namespace atomparsing {
 // You might think that using memcpy makes this function expensive, but you'd be wrong.
 // All decent optimizing compilers (GCC, clang, Visual Studio) will compile string_to_uint32("false");
 // to the compile-time constant 1936482662.
-simdjson_really_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
+simdjson_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
 
 
 // Again in str4ncmp we use a memcpy to avoid undefined behavior. The memcpy may appear expensive.
 // Yet all decent optimizing compilers will compile memcpy to a single instruction, just about.
 simdjson_warn_unused
-simdjson_really_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
+simdjson_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
   uint32_t srcval; // we want to avoid unaligned 32-bit loads (undefined in C/C++)
   static_assert(sizeof(uint32_t) <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be larger than 4 bytes");
   std::memcpy(&srcval, src, sizeof(uint32_t));
@@ -10520,36 +10612,36 @@ simdjson_really_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_true_atom(const uint8_t *src) {
+simdjson_inline bool is_valid_true_atom(const uint8_t *src) {
   return (str4ncmp(src, "true") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
+simdjson_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
   if (len > 4) { return is_valid_true_atom(src); }
   else if (len == 4) { return !str4ncmp(src, "true"); }
   else { return false; }
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_false_atom(const uint8_t *src) {
+simdjson_inline bool is_valid_false_atom(const uint8_t *src) {
   return (str4ncmp(src+1, "alse") | jsoncharutils::is_not_structural_or_whitespace(src[5])) == 0;
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
+simdjson_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
   if (len > 5) { return is_valid_false_atom(src); }
   else if (len == 5) { return !str4ncmp(src+1, "alse"); }
   else { return false; }
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_null_atom(const uint8_t *src) {
+simdjson_inline bool is_valid_null_atom(const uint8_t *src) {
   return (str4ncmp(src, "null") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
+simdjson_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
   if (len > 4) { return is_valid_null_atom(src); }
   else if (len == 4) { return !str4ncmp(src, "null"); }
   else { return false; }
@@ -10575,18 +10667,18 @@ using namespace simd;
 struct backslash_and_quote {
 public:
   static constexpr uint32_t BYTES_PROCESSED = 32;
-  simdjson_really_inline static backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+  simdjson_inline static backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
 
-  simdjson_really_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
-  simdjson_really_inline bool has_backslash() { return bs_bits != 0; }
-  simdjson_really_inline int quote_index() { return trailing_zeroes(quote_bits); }
-  simdjson_really_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+  simdjson_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_inline bool has_backslash() { return bs_bits != 0; }
+  simdjson_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_inline int backslash_index() { return trailing_zeroes(bs_bits); }
 
   uint32_t bs_bits;
   uint32_t quote_bits;
 }; // struct backslash_and_quote
 
-simdjson_really_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
   // this can read up to 31 bytes beyond the buffer size, but we require
   // SIMDJSON_PADDING of padding
   static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
@@ -10608,147 +10700,6 @@ simdjson_really_inline backslash_and_quote backslash_and_quote::copy_and_find(co
 } // namespace arm64
 } // namespace simdjson
 
-/* begin file include/simdjson/generic/stringparsing.h */
-// This file contains the common code every implementation uses
-// It is intended to be included multiple times and compiled multiple times
-
-namespace simdjson {
-namespace arm64 {
-namespace {
-/// @private
-namespace stringparsing {
-
-// begin copypasta
-// These chars yield themselves: " \ /
-// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
-// u not handled in this table as it's complex
-static const uint8_t escape_map[256] = {
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
-    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
-    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
-
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-};
-
-// handle a unicode codepoint
-// write appropriate values into dest
-// src will advance 6 bytes or 12 bytes
-// dest will advance a variable amount (return via pointer)
-// return true if the unicode codepoint was valid
-// We work in little-endian then swap at write time
-simdjson_warn_unused
-simdjson_really_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
-                                            uint8_t **dst_ptr) {
-  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
-  // conversion isn't valid; we defer the check for this to inside the
-  // multilingual plane check
-  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
-  *src_ptr += 6;
-  // check for low surrogate for characters outside the Basic
-  // Multilingual Plane.
-  if (code_point >= 0xd800 && code_point < 0xdc00) {
-    if (((*src_ptr)[0] != '\\') || (*src_ptr)[1] != 'u') {
-      return false;
-    }
-    uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
-
-    // if the first code point is invalid we will get here, as we will go past
-    // the check for being outside the Basic Multilingual plane. If we don't
-    // find a \u immediately afterwards we fail out anyhow, but if we do,
-    // this check catches both the case of the first code point being invalid
-    // or the second code point being invalid.
-    if ((code_point | code_point_2) >> 16) {
-      return false;
-    }
-
-    code_point =
-        (((code_point - 0xd800) << 10) | (code_point_2 - 0xdc00)) + 0x10000;
-    *src_ptr += 6;
-  }
-  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
-  *dst_ptr += offset;
-  return offset > 0;
-}
-
-/**
- * Unescape a string from src to dst, stopping at a final unescaped quote. E.g., if src points at 'joe"', then
- * dst needs to have four free bytes.
- */
-simdjson_warn_unused simdjson_really_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst) {
-  while (1) {
-    // Copy the next n bytes, and find the backslash and quote in them.
-    auto bs_quote = backslash_and_quote::copy_and_find(src, dst);
-    // If the next thing is the end quote, copy and return
-    if (bs_quote.has_quote_first()) {
-      // we encountered quotes first. Move dst to point to quotes and exit
-      return dst + bs_quote.quote_index();
-    }
-    if (bs_quote.has_backslash()) {
-      /* find out where the backspace is */
-      auto bs_dist = bs_quote.backslash_index();
-      uint8_t escape_char = src[bs_dist + 1];
-      /* we encountered backslash first. Handle backslash */
-      if (escape_char == 'u') {
-        /* move src/dst up to the start; they will be further adjusted
-           within the unicode codepoint handling code. */
-        src += bs_dist;
-        dst += bs_dist;
-        if (!handle_unicode_codepoint(&src, &dst)) {
-          return nullptr;
-        }
-      } else {
-        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
-         * write bs_dist+1 characters to output
-         * note this may reach beyond the part of the buffer we've actually
-         * seen. I think this is ok */
-        uint8_t escape_result = escape_map[escape_char];
-        if (escape_result == 0u) {
-          return nullptr; /* bogus escape value is an error */
-        }
-        dst[bs_dist] = escape_result;
-        src += bs_dist + 2;
-        dst += bs_dist + 1;
-      }
-    } else {
-      /* they are the same. Since they can't co-occur, it means we
-       * encountered neither. */
-      src += backslash_and_quote::BYTES_PROCESSED;
-      dst += backslash_and_quote::BYTES_PROCESSED;
-    }
-  }
-  /* can't be reached */
-  return nullptr;
-}
-
-simdjson_unused simdjson_warn_unused simdjson_really_inline error_code parse_string_to_buffer(const uint8_t *src, uint8_t *&current_string_buf_loc, std::string_view &s) {
-  if (*(src++) != '"') { return STRING_ERROR; }
-  auto end = stringparsing::parse_string(src, current_string_buf_loc);
-  if (!end) { return STRING_ERROR; }
-  s = std::string_view(reinterpret_cast<const char *>(current_string_buf_loc), end-current_string_buf_loc);
-  current_string_buf_loc = end;
-  return SUCCESS;
-}
-
-} // namespace stringparsing
-} // unnamed namespace
-} // namespace arm64
-} // namespace simdjson
-/* end file include/simdjson/generic/stringparsing.h */
-
 #endif // SIMDJSON_ARM64_STRINGPARSING_H
 /* end file include/simdjson/arm64/stringparsing.h */
 /* begin file include/simdjson/arm64/numberparsing.h */
@@ -10761,7 +10712,7 @@ namespace {
 
 // we don't have SSE, so let us use a scalar function
 // credit: https://johnnylee-sde.github.io/Fast-numeric-string-to-int/
-static simdjson_really_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
   uint64_t val;
   std::memcpy(&val, chars, sizeof(uint64_t));
   val = (val & 0x0F0F0F0F0F0F0F0F) * 2561 >> 8;
@@ -10814,7 +10765,7 @@ namespace {
 // Convert a mantissa, an exponent and a sign bit into an ieee64 double.
 // The real_exponent needs to be in [0, 2046] (technically real_exponent = 2047 would be acceptable).
 // The mantissa should be in [0,1<<53). The bit at index (1ULL << 52) while be zeroed.
-simdjson_really_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
+simdjson_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
     double d;
     mantissa &= ~(1ULL << 52);
     mantissa |= real_exponent << 52;
@@ -10829,7 +10780,7 @@ simdjson_really_inline double to_double(uint64_t mantissa, uint64_t real_exponen
 // set to false. This should work *most of the time* (like 99% of the time).
 // We assume that power is in the [smallest_power,
 // largest_power] interval: the caller is responsible for this check.
-simdjson_really_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
+simdjson_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
   // we start with a fast path
   // It was described in
   // Clinger WD. How to read floating point numbers accurately.
@@ -10889,7 +10840,7 @@ simdjson_really_inline bool compute_float_64(int64_t power, uint64_t i, bool neg
   // In the slow path, we need to adjust i so that it is > 1<<63 which is always
   // possible, except if i == 0, so we handle i == 0 separately.
   if(i == 0) {
-    d = 0.0;
+    d = negative ? -0.0 : 0.0;
     return true;
   }
 
@@ -11004,7 +10955,7 @@ simdjson_really_inline bool compute_float_64(int64_t power, uint64_t i, bool neg
   if (simdjson_unlikely(real_exponent <= 0)) { // we have a subnormal?
     // Here have that real_exponent <= 0 so -real_exponent >= 0
     if(-real_exponent + 1 >= 64) { // if we have more than 64 bits below the minimum exponent, you have a zero for sure.
-      d = 0.0;
+      d = negative ? -0.0 : 0.0;
       return true;
     }
     // next line is safe because -real_exponent + 1 < 0
@@ -11112,7 +11063,7 @@ static bool parse_float_fallback(const uint8_t *ptr, const uint8_t *end_ptr, dou
 // check quickly whether the next 8 chars are made of digits
 // at a glance, it looks better than Mula's
 // http://0x80.pl/articles/swar-digits-validate.html
-simdjson_really_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
+simdjson_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
   uint64_t val;
   // this can read up to 7 bytes beyond the buffer size, but we require
   // SIMDJSON_PADDING of padding
@@ -11139,7 +11090,7 @@ error_code slow_float_parsing(simdjson_unused const uint8_t * src, W writer) {
 
 template<typename I>
 SIMDJSON_NO_SANITIZE_UNDEFINED // We deliberately allow overflow here and check later
-simdjson_really_inline bool parse_digit(const uint8_t c, I &i) {
+simdjson_inline bool parse_digit(const uint8_t c, I &i) {
   const uint8_t digit = static_cast<uint8_t>(c - '0');
   if (digit > 9) {
     return false;
@@ -11149,7 +11100,7 @@ simdjson_really_inline bool parse_digit(const uint8_t c, I &i) {
   return true;
 }
 
-simdjson_really_inline error_code parse_decimal(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
+simdjson_inline error_code parse_decimal(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
   // we continue with the fiction that we have an integer. If the
   // floating point number is representable as x * 10^z for some integer
   // z that fits in 53 bits, then we will be able to convert back the
@@ -11177,7 +11128,7 @@ simdjson_really_inline error_code parse_decimal(simdjson_unused const uint8_t *c
   return SUCCESS;
 }
 
-simdjson_really_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
+simdjson_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
   // Exp Sign: -123.456e[-]78
   bool neg_exp = ('-' == *p);
   if (neg_exp || '+' == *p) { p++; } // Skip + as well
@@ -11228,7 +11179,7 @@ simdjson_really_inline error_code parse_exponent(simdjson_unused const uint8_t *
   return SUCCESS;
 }
 
-simdjson_really_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
+simdjson_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
   // It is possible that the integer had an overflow.
   // We have to handle the case where we have 0.0000somenumber.
   const uint8_t *start = start_digits;
@@ -11238,7 +11189,7 @@ simdjson_really_inline size_t significant_digits(const uint8_t * start_digits, s
 }
 
 template<typename W>
-simdjson_really_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
+simdjson_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
   // If we frequently had to deal with long strings of digits,
   // we could extend our code by using a 128-bit integer instead
   // of a 64-bit integer. However, this is uncommon in practice.
@@ -11274,7 +11225,8 @@ simdjson_really_inline error_code write_float(const uint8_t *const src, bool neg
     static_assert(simdjson::internal::smallest_power <= -342, "smallest_power is not small enough");
     //
     if((exponent < simdjson::internal::smallest_power) || (i == 0)) {
-      WRITE_DOUBLE(0, src, writer);
+      // E.g. Parse "-0.0e-999" into the same value as "-0.0". See https://en.wikipedia.org/wiki/Signed_zero
+      WRITE_DOUBLE(negative ? -0.0 : 0.0, src, writer);
       return SUCCESS;
     } else { // (exponent > largest_power) and (i != 0)
       // We have, for sure, an infinite value and simdjson refuses to parse infinite values.
@@ -11294,20 +11246,20 @@ simdjson_really_inline error_code write_float(const uint8_t *const src, bool neg
 #ifdef SIMDJSON_SKIPNUMBERPARSING
 
 template<typename W>
-simdjson_really_inline error_code parse_number(const uint8_t *const, W &writer) {
+simdjson_inline error_code parse_number(const uint8_t *const, W &writer) {
   writer.append_s64(0);        // always write zero
   return SUCCESS;              // always succeeds
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
-simdjson_unused simdjson_really_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
-simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept { return ondemand::number_type::signed_integer; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept { return ondemand::number_type::signed_integer; }
 #else
 
 // parse the number at src
@@ -11320,13 +11272,13 @@ simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> ge
 //
 // Our objective is accurate parsing (ULP of 0) at high speed.
 template<typename W>
-simdjson_really_inline error_code parse_number(const uint8_t *const src, W &writer) {
+simdjson_inline error_code parse_number(const uint8_t *const src, W &writer) {
 
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  const uint8_t *p = src + negative;
+  const uint8_t *p = src + uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -11470,7 +11422,7 @@ const uint8_t integer_string_finisher[256] = {
     NUMBER_ERROR};
 
 // Parse any number from 0 to 18,446,744,073,709,551,615
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
   const uint8_t *p = src;
   //
   // Parse the integer part.
@@ -11520,7 +11472,7 @@ simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(
 
 // Parse any number from 0 to 18,446,744,073,709,551,615
 // Never read at src_end or beyond
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
   const uint8_t *p = src;
   //
   // Parse the integer part.
@@ -11568,7 +11520,7 @@ simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(
 }
 
 // Parse any number from 0 to 18,446,744,073,709,551,615
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
   const uint8_t *p = src + 1;
   //
   // Parse the integer part.
@@ -11618,12 +11570,12 @@ simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned_
 }
 
 // Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  const uint8_t *p = src + negative;
+  const uint8_t *p = src + uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -11661,13 +11613,13 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(co
 
 // Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
 // Never read at src_end or beyond
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
   //
   // Check for minus sign
   //
   if(src == src_end) { return NUMBER_ERROR; }
   bool negative = (*src == '-');
-  const uint8_t *p = src + negative;
+  const uint8_t *p = src + uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -11704,24 +11656,24 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(co
 }
 
 // Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*(src + 1) == '-');
-  const uint8_t *p = src + negative + 1;
+  src += uint8_t(negative) + 1;
 
   //
   // Parse the integer part.
   //
   // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
-  const uint8_t *const start_digits = p;
+  const uint8_t *const start_digits = src;
   uint64_t i = 0;
-  while (parse_digit(*p, i)) { p++; }
+  while (parse_digit(*src, i)) { src++; }
 
   // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
   // Optimization note: size_t is expected to be unsigned.
-  size_t digit_count = size_t(p - start_digits);
+  size_t digit_count = size_t(src - start_digits);
   // We go from
   // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
   // so we can never represent numbers that have more than 19 digits.
@@ -11733,11 +11685,11 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in
   // Here digit_count > 0.
   if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
   // We can do the following...
-  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
-  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // if (!jsoncharutils::is_structural_or_whitespace(*src)) {
+  //  return (*src == '.' || *src == 'e' || *src == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
   // }
   // as a single table lookup:
-  if(*p != '"') { return NUMBER_ERROR; }
+  if(*src != '"') { return NUMBER_ERROR; }
   // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
   // Performance note: This check is only needed when digit_count == longest_digit_count but it is
   // so cheap that we might as well always make it.
@@ -11745,12 +11697,12 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in
   return negative ? (~i+1) : i;
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -11817,19 +11769,19 @@ simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(cons
   if (simdjson_likely(!overflow)) {
     if (compute_float_64(exponent, i, negative, d)) { return d; }
   }
-  if (!parse_float_fallback(src-negative, &d)) {
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
     return NUMBER_ERROR;
   }
   return d;
 }
 
-simdjson_unused simdjson_really_inline bool is_negative(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept {
   return (*src == '-');
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
   const uint8_t *p = src;
   while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
   if ( p == src ) { return NUMBER_ERROR; }
@@ -11837,9 +11789,9 @@ simdjson_unused simdjson_really_inline simdjson_result<bool> is_integer(const ui
   return false;
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept {
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
   const uint8_t *p = src;
   while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
   if ( p == src ) { return NUMBER_ERROR; }
@@ -11863,13 +11815,13 @@ simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> ge
 }
 
 // Never read at src_end or beyond
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
   if(src == src_end) { return NUMBER_ERROR; }
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -11938,18 +11890,18 @@ simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(cons
   if (simdjson_likely(!overflow)) {
     if (compute_float_64(exponent, i, negative, d)) { return d; }
   }
-  if (!parse_float_fallback(src-negative, src_end, &d)) {
+  if (!parse_float_fallback(src - uint8_t(negative), src_end, &d)) {
     return NUMBER_ERROR;
   }
   return d;
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*(src + 1) == '-');
-  src += negative + 1;
+  src += uint8_t(negative) + 1;
 
   //
   // Parse the integer part.
@@ -12016,7 +11968,7 @@ simdjson_unused simdjson_really_inline simdjson_result<double> parse_double_in_s
   if (simdjson_likely(!overflow)) {
     if (compute_float_64(exponent, i, negative, d)) { return d; }
   }
-  if (!parse_float_fallback(src-negative, &d)) {
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
     return NUMBER_ERROR;
   }
   return d;
@@ -12069,7 +12021,7 @@ using namespace simdjson::dom;
 
 class implementation final : public simdjson::implementation {
 public:
-  simdjson_really_inline implementation() : simdjson::implementation(
+  simdjson_inline implementation() : simdjson::implementation(
       "fallback",
       "Generic fallback implementation",
       0
@@ -12131,10 +12083,11 @@ class dom_parser_implementation final : public internal::dom_parser_implementati
   simdjson_warn_unused error_code stage1(const uint8_t *buf, size_t len, stage1_mode partial) noexcept final;
   simdjson_warn_unused error_code stage2(dom::document &doc) noexcept final;
   simdjson_warn_unused error_code stage2_next(dom::document &doc) noexcept final;
+  simdjson_warn_unused uint8_t *parse_string(const uint8_t *src, uint8_t *dst) const noexcept final;
   inline simdjson_warn_unused error_code set_capacity(size_t capacity) noexcept final;
   inline simdjson_warn_unused error_code set_max_depth(size_t max_depth) noexcept final;
 private:
-  simdjson_really_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
+  simdjson_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
 
 };
 
@@ -12203,7 +12156,7 @@ static unsigned char _BitScanReverse64(unsigned long* ret, uint64_t x) {
 #endif
 
 /* result might be undefined when input_num is zero */
-simdjson_really_inline int leading_zeroes(uint64_t input_num) {
+simdjson_inline int leading_zeroes(uint64_t input_num) {
 #ifdef _MSC_VER
   unsigned long leading_zero = 0;
   // Search the mask data from most significant bit (MSB)
@@ -12232,11 +12185,11 @@ namespace jsoncharutils {
 
 // return non-zero if not a structural or whitespace char
 // zero otherwise
-simdjson_really_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
+simdjson_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
   return internal::structural_or_whitespace_negated[c];
 }
 
-simdjson_really_inline uint32_t is_structural_or_whitespace(uint8_t c) {
+simdjson_inline uint32_t is_structural_or_whitespace(uint8_t c) {
   return internal::structural_or_whitespace[c];
 }
 
@@ -12267,7 +12220,7 @@ static inline uint32_t hex_to_u32_nocheck(
 //
 // Note: we assume that surrogates are treated separately
 //
-simdjson_really_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
+simdjson_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
   if (cp <= 0x7F) {
     c[0] = uint8_t(cp);
     return 1; // ascii
@@ -12299,10 +12252,10 @@ simdjson_really_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
 #ifdef SIMDJSON_IS_32BITS // _umul128 for x86, arm
 // this is a slow emulation routine for 32-bit
 //
-static simdjson_really_inline uint64_t __emulu(uint32_t x, uint32_t y) {
+static simdjson_inline uint64_t __emulu(uint32_t x, uint32_t y) {
   return x * (uint64_t)y;
 }
-static simdjson_really_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+static simdjson_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
   uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
   uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
   uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
@@ -12316,7 +12269,7 @@ static simdjson_really_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64
 
 using internal::value128;
 
-simdjson_really_inline value128 full_multiplication(uint64_t value1, uint64_t value2) {
+simdjson_inline value128 full_multiplication(uint64_t value1, uint64_t value2) {
   value128 answer;
 #if defined(SIMDJSON_REGULAR_VISUAL_STUDIO) || defined(SIMDJSON_IS_32BITS)
 #ifdef _M_ARM64
@@ -12352,13 +12305,13 @@ namespace atomparsing {
 // You might think that using memcpy makes this function expensive, but you'd be wrong.
 // All decent optimizing compilers (GCC, clang, Visual Studio) will compile string_to_uint32("false");
 // to the compile-time constant 1936482662.
-simdjson_really_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
+simdjson_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
 
 
 // Again in str4ncmp we use a memcpy to avoid undefined behavior. The memcpy may appear expensive.
 // Yet all decent optimizing compilers will compile memcpy to a single instruction, just about.
 simdjson_warn_unused
-simdjson_really_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
+simdjson_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
   uint32_t srcval; // we want to avoid unaligned 32-bit loads (undefined in C/C++)
   static_assert(sizeof(uint32_t) <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be larger than 4 bytes");
   std::memcpy(&srcval, src, sizeof(uint32_t));
@@ -12366,36 +12319,36 @@ simdjson_really_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_true_atom(const uint8_t *src) {
+simdjson_inline bool is_valid_true_atom(const uint8_t *src) {
   return (str4ncmp(src, "true") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
+simdjson_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
   if (len > 4) { return is_valid_true_atom(src); }
   else if (len == 4) { return !str4ncmp(src, "true"); }
   else { return false; }
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_false_atom(const uint8_t *src) {
+simdjson_inline bool is_valid_false_atom(const uint8_t *src) {
   return (str4ncmp(src+1, "alse") | jsoncharutils::is_not_structural_or_whitespace(src[5])) == 0;
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
+simdjson_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
   if (len > 5) { return is_valid_false_atom(src); }
   else if (len == 5) { return !str4ncmp(src+1, "alse"); }
   else { return false; }
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_null_atom(const uint8_t *src) {
+simdjson_inline bool is_valid_null_atom(const uint8_t *src) {
   return (str4ncmp(src, "null") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
+simdjson_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
   if (len > 4) { return is_valid_null_atom(src); }
   else if (len == 4) { return !str4ncmp(src, "null"); }
   else { return false; }
@@ -12419,17 +12372,17 @@ namespace {
 struct backslash_and_quote {
 public:
   static constexpr uint32_t BYTES_PROCESSED = 1;
-  simdjson_really_inline static backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+  simdjson_inline static backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
 
-  simdjson_really_inline bool has_quote_first() { return c == '"'; }
-  simdjson_really_inline bool has_backslash() { return c == '\\'; }
-  simdjson_really_inline int quote_index() { return c == '"' ? 0 : 1; }
-  simdjson_really_inline int backslash_index() { return c == '\\' ? 0 : 1; }
+  simdjson_inline bool has_quote_first() { return c == '"'; }
+  simdjson_inline bool has_backslash() { return c == '\\'; }
+  simdjson_inline int quote_index() { return c == '"' ? 0 : 1; }
+  simdjson_inline int backslash_index() { return c == '\\' ? 0 : 1; }
 
   uint8_t c;
 }; // struct backslash_and_quote
 
-simdjson_really_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
   // store to dest unconditionally - we can overwrite the bits we don't like later
   dst[0] = src[0];
   return { src[0] };
@@ -12439,147 +12392,6 @@ simdjson_really_inline backslash_and_quote backslash_and_quote::copy_and_find(co
 } // namespace fallback
 } // namespace simdjson
 
-/* begin file include/simdjson/generic/stringparsing.h */
-// This file contains the common code every implementation uses
-// It is intended to be included multiple times and compiled multiple times
-
-namespace simdjson {
-namespace fallback {
-namespace {
-/// @private
-namespace stringparsing {
-
-// begin copypasta
-// These chars yield themselves: " \ /
-// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
-// u not handled in this table as it's complex
-static const uint8_t escape_map[256] = {
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
-    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
-    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
-
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-};
-
-// handle a unicode codepoint
-// write appropriate values into dest
-// src will advance 6 bytes or 12 bytes
-// dest will advance a variable amount (return via pointer)
-// return true if the unicode codepoint was valid
-// We work in little-endian then swap at write time
-simdjson_warn_unused
-simdjson_really_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
-                                            uint8_t **dst_ptr) {
-  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
-  // conversion isn't valid; we defer the check for this to inside the
-  // multilingual plane check
-  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
-  *src_ptr += 6;
-  // check for low surrogate for characters outside the Basic
-  // Multilingual Plane.
-  if (code_point >= 0xd800 && code_point < 0xdc00) {
-    if (((*src_ptr)[0] != '\\') || (*src_ptr)[1] != 'u') {
-      return false;
-    }
-    uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
-
-    // if the first code point is invalid we will get here, as we will go past
-    // the check for being outside the Basic Multilingual plane. If we don't
-    // find a \u immediately afterwards we fail out anyhow, but if we do,
-    // this check catches both the case of the first code point being invalid
-    // or the second code point being invalid.
-    if ((code_point | code_point_2) >> 16) {
-      return false;
-    }
-
-    code_point =
-        (((code_point - 0xd800) << 10) | (code_point_2 - 0xdc00)) + 0x10000;
-    *src_ptr += 6;
-  }
-  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
-  *dst_ptr += offset;
-  return offset > 0;
-}
-
-/**
- * Unescape a string from src to dst, stopping at a final unescaped quote. E.g., if src points at 'joe"', then
- * dst needs to have four free bytes.
- */
-simdjson_warn_unused simdjson_really_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst) {
-  while (1) {
-    // Copy the next n bytes, and find the backslash and quote in them.
-    auto bs_quote = backslash_and_quote::copy_and_find(src, dst);
-    // If the next thing is the end quote, copy and return
-    if (bs_quote.has_quote_first()) {
-      // we encountered quotes first. Move dst to point to quotes and exit
-      return dst + bs_quote.quote_index();
-    }
-    if (bs_quote.has_backslash()) {
-      /* find out where the backspace is */
-      auto bs_dist = bs_quote.backslash_index();
-      uint8_t escape_char = src[bs_dist + 1];
-      /* we encountered backslash first. Handle backslash */
-      if (escape_char == 'u') {
-        /* move src/dst up to the start; they will be further adjusted
-           within the unicode codepoint handling code. */
-        src += bs_dist;
-        dst += bs_dist;
-        if (!handle_unicode_codepoint(&src, &dst)) {
-          return nullptr;
-        }
-      } else {
-        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
-         * write bs_dist+1 characters to output
-         * note this may reach beyond the part of the buffer we've actually
-         * seen. I think this is ok */
-        uint8_t escape_result = escape_map[escape_char];
-        if (escape_result == 0u) {
-          return nullptr; /* bogus escape value is an error */
-        }
-        dst[bs_dist] = escape_result;
-        src += bs_dist + 2;
-        dst += bs_dist + 1;
-      }
-    } else {
-      /* they are the same. Since they can't co-occur, it means we
-       * encountered neither. */
-      src += backslash_and_quote::BYTES_PROCESSED;
-      dst += backslash_and_quote::BYTES_PROCESSED;
-    }
-  }
-  /* can't be reached */
-  return nullptr;
-}
-
-simdjson_unused simdjson_warn_unused simdjson_really_inline error_code parse_string_to_buffer(const uint8_t *src, uint8_t *&current_string_buf_loc, std::string_view &s) {
-  if (*(src++) != '"') { return STRING_ERROR; }
-  auto end = stringparsing::parse_string(src, current_string_buf_loc);
-  if (!end) { return STRING_ERROR; }
-  s = std::string_view(reinterpret_cast<const char *>(current_string_buf_loc), end-current_string_buf_loc);
-  current_string_buf_loc = end;
-  return SUCCESS;
-}
-
-} // namespace stringparsing
-} // unnamed namespace
-} // namespace fallback
-} // namespace simdjson
-/* end file include/simdjson/generic/stringparsing.h */
-
 #endif // SIMDJSON_FALLBACK_STRINGPARSING_H
 /* end file include/simdjson/fallback/stringparsing.h */
 /* begin file include/simdjson/fallback/numberparsing.h */
@@ -12597,14 +12409,14 @@ namespace simdjson {
 namespace fallback {
 namespace {
 // credit: https://johnnylee-sde.github.io/Fast-numeric-string-to-int/
-static simdjson_really_inline uint32_t parse_eight_digits_unrolled(const char *chars) {
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const char *chars) {
   uint64_t val;
   memcpy(&val, chars, sizeof(uint64_t));
   val = (val & 0x0F0F0F0F0F0F0F0F) * 2561 >> 8;
   val = (val & 0x00FF00FF00FF00FF) * 6553601 >> 16;
   return uint32_t((val & 0x0000FFFF0000FFFF) * 42949672960001 >> 32);
 }
-static simdjson_really_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
   return parse_eight_digits_unrolled(reinterpret_cast<const char *>(chars));
 }
 
@@ -12653,7 +12465,7 @@ namespace {
 // Convert a mantissa, an exponent and a sign bit into an ieee64 double.
 // The real_exponent needs to be in [0, 2046] (technically real_exponent = 2047 would be acceptable).
 // The mantissa should be in [0,1<<53). The bit at index (1ULL << 52) while be zeroed.
-simdjson_really_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
+simdjson_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
     double d;
     mantissa &= ~(1ULL << 52);
     mantissa |= real_exponent << 52;
@@ -12668,7 +12480,7 @@ simdjson_really_inline double to_double(uint64_t mantissa, uint64_t real_exponen
 // set to false. This should work *most of the time* (like 99% of the time).
 // We assume that power is in the [smallest_power,
 // largest_power] interval: the caller is responsible for this check.
-simdjson_really_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
+simdjson_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
   // we start with a fast path
   // It was described in
   // Clinger WD. How to read floating point numbers accurately.
@@ -12728,7 +12540,7 @@ simdjson_really_inline bool compute_float_64(int64_t power, uint64_t i, bool neg
   // In the slow path, we need to adjust i so that it is > 1<<63 which is always
   // possible, except if i == 0, so we handle i == 0 separately.
   if(i == 0) {
-    d = 0.0;
+    d = negative ? -0.0 : 0.0;
     return true;
   }
 
@@ -12843,7 +12655,7 @@ simdjson_really_inline bool compute_float_64(int64_t power, uint64_t i, bool neg
   if (simdjson_unlikely(real_exponent <= 0)) { // we have a subnormal?
     // Here have that real_exponent <= 0 so -real_exponent >= 0
     if(-real_exponent + 1 >= 64) { // if we have more than 64 bits below the minimum exponent, you have a zero for sure.
-      d = 0.0;
+      d = negative ? -0.0 : 0.0;
       return true;
     }
     // next line is safe because -real_exponent + 1 < 0
@@ -12951,7 +12763,7 @@ static bool parse_float_fallback(const uint8_t *ptr, const uint8_t *end_ptr, dou
 // check quickly whether the next 8 chars are made of digits
 // at a glance, it looks better than Mula's
 // http://0x80.pl/articles/swar-digits-validate.html
-simdjson_really_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
+simdjson_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
   uint64_t val;
   // this can read up to 7 bytes beyond the buffer size, but we require
   // SIMDJSON_PADDING of padding
@@ -12978,7 +12790,7 @@ error_code slow_float_parsing(simdjson_unused const uint8_t * src, W writer) {
 
 template<typename I>
 SIMDJSON_NO_SANITIZE_UNDEFINED // We deliberately allow overflow here and check later
-simdjson_really_inline bool parse_digit(const uint8_t c, I &i) {
+simdjson_inline bool parse_digit(const uint8_t c, I &i) {
   const uint8_t digit = static_cast<uint8_t>(c - '0');
   if (digit > 9) {
     return false;
@@ -12988,7 +12800,7 @@ simdjson_really_inline bool parse_digit(const uint8_t c, I &i) {
   return true;
 }
 
-simdjson_really_inline error_code parse_decimal(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
+simdjson_inline error_code parse_decimal(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
   // we continue with the fiction that we have an integer. If the
   // floating point number is representable as x * 10^z for some integer
   // z that fits in 53 bits, then we will be able to convert back the
@@ -13016,7 +12828,7 @@ simdjson_really_inline error_code parse_decimal(simdjson_unused const uint8_t *c
   return SUCCESS;
 }
 
-simdjson_really_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
+simdjson_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
   // Exp Sign: -123.456e[-]78
   bool neg_exp = ('-' == *p);
   if (neg_exp || '+' == *p) { p++; } // Skip + as well
@@ -13067,7 +12879,7 @@ simdjson_really_inline error_code parse_exponent(simdjson_unused const uint8_t *
   return SUCCESS;
 }
 
-simdjson_really_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
+simdjson_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
   // It is possible that the integer had an overflow.
   // We have to handle the case where we have 0.0000somenumber.
   const uint8_t *start = start_digits;
@@ -13077,7 +12889,7 @@ simdjson_really_inline size_t significant_digits(const uint8_t * start_digits, s
 }
 
 template<typename W>
-simdjson_really_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
+simdjson_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
   // If we frequently had to deal with long strings of digits,
   // we could extend our code by using a 128-bit integer instead
   // of a 64-bit integer. However, this is uncommon in practice.
@@ -13113,7 +12925,8 @@ simdjson_really_inline error_code write_float(const uint8_t *const src, bool neg
     static_assert(simdjson::internal::smallest_power <= -342, "smallest_power is not small enough");
     //
     if((exponent < simdjson::internal::smallest_power) || (i == 0)) {
-      WRITE_DOUBLE(0, src, writer);
+      // E.g. Parse "-0.0e-999" into the same value as "-0.0". See https://en.wikipedia.org/wiki/Signed_zero
+      WRITE_DOUBLE(negative ? -0.0 : 0.0, src, writer);
       return SUCCESS;
     } else { // (exponent > largest_power) and (i != 0)
       // We have, for sure, an infinite value and simdjson refuses to parse infinite values.
@@ -13133,20 +12946,20 @@ simdjson_really_inline error_code write_float(const uint8_t *const src, bool neg
 #ifdef SIMDJSON_SKIPNUMBERPARSING
 
 template<typename W>
-simdjson_really_inline error_code parse_number(const uint8_t *const, W &writer) {
+simdjson_inline error_code parse_number(const uint8_t *const, W &writer) {
   writer.append_s64(0);        // always write zero
   return SUCCESS;              // always succeeds
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
-simdjson_unused simdjson_really_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
-simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept { return ondemand::number_type::signed_integer; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept { return ondemand::number_type::signed_integer; }
 #else
 
 // parse the number at src
@@ -13159,13 +12972,13 @@ simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> ge
 //
 // Our objective is accurate parsing (ULP of 0) at high speed.
 template<typename W>
-simdjson_really_inline error_code parse_number(const uint8_t *const src, W &writer) {
+simdjson_inline error_code parse_number(const uint8_t *const src, W &writer) {
 
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  const uint8_t *p = src + negative;
+  const uint8_t *p = src + uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -13309,7 +13122,7 @@ const uint8_t integer_string_finisher[256] = {
     NUMBER_ERROR};
 
 // Parse any number from 0 to 18,446,744,073,709,551,615
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
   const uint8_t *p = src;
   //
   // Parse the integer part.
@@ -13359,7 +13172,7 @@ simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(
 
 // Parse any number from 0 to 18,446,744,073,709,551,615
 // Never read at src_end or beyond
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
   const uint8_t *p = src;
   //
   // Parse the integer part.
@@ -13407,7 +13220,7 @@ simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(
 }
 
 // Parse any number from 0 to 18,446,744,073,709,551,615
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
   const uint8_t *p = src + 1;
   //
   // Parse the integer part.
@@ -13457,12 +13270,12 @@ simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned_
 }
 
 // Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  const uint8_t *p = src + negative;
+  const uint8_t *p = src + uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -13500,13 +13313,13 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(co
 
 // Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
 // Never read at src_end or beyond
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
   //
   // Check for minus sign
   //
   if(src == src_end) { return NUMBER_ERROR; }
   bool negative = (*src == '-');
-  const uint8_t *p = src + negative;
+  const uint8_t *p = src + uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -13543,24 +13356,24 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(co
 }
 
 // Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*(src + 1) == '-');
-  const uint8_t *p = src + negative + 1;
+  src += uint8_t(negative) + 1;
 
   //
   // Parse the integer part.
   //
   // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
-  const uint8_t *const start_digits = p;
+  const uint8_t *const start_digits = src;
   uint64_t i = 0;
-  while (parse_digit(*p, i)) { p++; }
+  while (parse_digit(*src, i)) { src++; }
 
   // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
   // Optimization note: size_t is expected to be unsigned.
-  size_t digit_count = size_t(p - start_digits);
+  size_t digit_count = size_t(src - start_digits);
   // We go from
   // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
   // so we can never represent numbers that have more than 19 digits.
@@ -13572,11 +13385,11 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in
   // Here digit_count > 0.
   if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
   // We can do the following...
-  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
-  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // if (!jsoncharutils::is_structural_or_whitespace(*src)) {
+  //  return (*src == '.' || *src == 'e' || *src == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
   // }
   // as a single table lookup:
-  if(*p != '"') { return NUMBER_ERROR; }
+  if(*src != '"') { return NUMBER_ERROR; }
   // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
   // Performance note: This check is only needed when digit_count == longest_digit_count but it is
   // so cheap that we might as well always make it.
@@ -13584,12 +13397,12 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in
   return negative ? (~i+1) : i;
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -13656,19 +13469,19 @@ simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(cons
   if (simdjson_likely(!overflow)) {
     if (compute_float_64(exponent, i, negative, d)) { return d; }
   }
-  if (!parse_float_fallback(src-negative, &d)) {
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
     return NUMBER_ERROR;
   }
   return d;
 }
 
-simdjson_unused simdjson_really_inline bool is_negative(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept {
   return (*src == '-');
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
   const uint8_t *p = src;
   while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
   if ( p == src ) { return NUMBER_ERROR; }
@@ -13676,9 +13489,9 @@ simdjson_unused simdjson_really_inline simdjson_result<bool> is_integer(const ui
   return false;
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept {
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
   const uint8_t *p = src;
   while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
   if ( p == src ) { return NUMBER_ERROR; }
@@ -13702,13 +13515,13 @@ simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> ge
 }
 
 // Never read at src_end or beyond
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
   if(src == src_end) { return NUMBER_ERROR; }
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -13777,18 +13590,18 @@ simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(cons
   if (simdjson_likely(!overflow)) {
     if (compute_float_64(exponent, i, negative, d)) { return d; }
   }
-  if (!parse_float_fallback(src-negative, src_end, &d)) {
+  if (!parse_float_fallback(src - uint8_t(negative), src_end, &d)) {
     return NUMBER_ERROR;
   }
   return d;
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*(src + 1) == '-');
-  src += negative + 1;
+  src += uint8_t(negative) + 1;
 
   //
   // Parse the integer part.
@@ -13855,7 +13668,7 @@ simdjson_unused simdjson_really_inline simdjson_result<double> parse_double_in_s
   if (simdjson_likely(!overflow)) {
     if (compute_float_64(exponent, i, negative, d)) { return d; }
   }
-  if (!parse_float_fallback(src-negative, &d)) {
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
     return NUMBER_ERROR;
   }
   return d;
@@ -13916,7 +13729,7 @@ using namespace simdjson;
 
 class implementation final : public simdjson::implementation {
 public:
-  simdjson_really_inline implementation() : simdjson::implementation(
+  simdjson_inline implementation() : simdjson::implementation(
       "icelake",
       "Intel/AMD AVX512",
       internal::instruction_set::AVX2 | internal::instruction_set::PCLMULQDQ | internal::instruction_set::BMI1 | internal::instruction_set::BMI2 | internal::instruction_set::AVX512F | internal::instruction_set::AVX512DQ | internal::instruction_set::AVX512CD | internal::instruction_set::AVX512BW | internal::instruction_set::AVX512VL | internal::instruction_set::AVX512VBMI2
@@ -13990,6 +13803,8 @@ class implementation final : public simdjson::implementation {
 #endif //  _blsr_u64
 #endif // SIMDJSON_CLANG_VISUAL_STUDIO
 
+static_assert(sizeof(__m512i) <= simdjson::SIMDJSON_PADDING, "insufficient padding for icelake");
+
 #endif // SIMDJSON_ICELAKE_INTRINSICS_H
 /* end file include/simdjson/icelake/intrinsics.h */
 
@@ -14039,10 +13854,11 @@ class dom_parser_implementation final : public internal::dom_parser_implementati
   simdjson_warn_unused error_code stage1(const uint8_t *buf, size_t len, stage1_mode partial) noexcept final;
   simdjson_warn_unused error_code stage2(dom::document &doc) noexcept final;
   simdjson_warn_unused error_code stage2_next(dom::document &doc) noexcept final;
+  simdjson_warn_unused uint8_t *parse_string(const uint8_t *src, uint8_t *dst) const noexcept final;
   inline simdjson_warn_unused error_code set_capacity(size_t capacity) noexcept final;
   inline simdjson_warn_unused error_code set_max_depth(size_t max_depth) noexcept final;
 private:
-  simdjson_really_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
+  simdjson_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
 
 };
 
@@ -14095,7 +13911,7 @@ namespace {
 // but the algorithms do not end up using the returned value.
 // Sadly, sanitizers are not smart enough to figure it out.
 SIMDJSON_NO_SANITIZE_UNDEFINED
-simdjson_really_inline int trailing_zeroes(uint64_t input_num) {
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
   return (int)_tzcnt_u64(input_num);
 #else // SIMDJSON_REGULAR_VISUAL_STUDIO
@@ -14109,27 +13925,27 @@ simdjson_really_inline int trailing_zeroes(uint64_t input_num) {
 }
 
 /* result might be undefined when input_num is zero */
-simdjson_really_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
   return _blsr_u64(input_num);
 }
 
 /* result might be undefined when input_num is zero */
-simdjson_really_inline int leading_zeroes(uint64_t input_num) {
+simdjson_inline int leading_zeroes(uint64_t input_num) {
   return int(_lzcnt_u64(input_num));
 }
 
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
-simdjson_really_inline unsigned __int64 count_ones(uint64_t input_num) {
+simdjson_inline unsigned __int64 count_ones(uint64_t input_num) {
   // note: we do not support legacy 32-bit Windows
   return __popcnt64(input_num);// Visual Studio wants two underscores
 }
 #else
-simdjson_really_inline long long int count_ones(uint64_t input_num) {
+simdjson_inline long long int count_ones(uint64_t input_num) {
   return _popcnt64(input_num);
 }
 #endif
 
-simdjson_really_inline bool add_overflow(uint64_t value1, uint64_t value2,
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2,
                                 uint64_t *result) {
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
   return _addcarry_u64(0, value1, value2,
@@ -14159,7 +13975,7 @@ namespace {
 //
 // For example, prefix_xor(00100100) == 00011100
 //
-simdjson_really_inline uint64_t prefix_xor(const uint64_t bitmask) {
+simdjson_inline uint64_t prefix_xor(const uint64_t bitmask) {
   // There should be no such thing with a processor supporting avx2
   // but not clmul.
   __m128i all_ones = _mm_set1_epi8('\xFF');
@@ -14215,23 +14031,23 @@ namespace simd {
     __m512i value;
 
     // Zero constructor
-    simdjson_really_inline base() : value{__m512i()} {}
+    simdjson_inline base() : value{__m512i()} {}
 
     // Conversion from SIMD register
-    simdjson_really_inline base(const __m512i _value) : value(_value) {}
+    simdjson_inline base(const __m512i _value) : value(_value) {}
 
     // Conversion to SIMD register
-    simdjson_really_inline operator const __m512i&() const { return this->value; }
-    simdjson_really_inline operator __m512i&() { return this->value; }
+    simdjson_inline operator const __m512i&() const { return this->value; }
+    simdjson_inline operator __m512i&() { return this->value; }
 
     // Bit operations
-    simdjson_really_inline Child operator|(const Child other) const { return _mm512_or_si512(*this, other); }
-    simdjson_really_inline Child operator&(const Child other) const { return _mm512_and_si512(*this, other); }
-    simdjson_really_inline Child operator^(const Child other) const { return _mm512_xor_si512(*this, other); }
-    simdjson_really_inline Child bit_andnot(const Child other) const { return _mm512_andnot_si512(other, *this); }
-    simdjson_really_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
-    simdjson_really_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
-    simdjson_really_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+    simdjson_inline Child operator|(const Child other) const { return _mm512_or_si512(*this, other); }
+    simdjson_inline Child operator&(const Child other) const { return _mm512_and_si512(*this, other); }
+    simdjson_inline Child operator^(const Child other) const { return _mm512_xor_si512(*this, other); }
+    simdjson_inline Child bit_andnot(const Child other) const { return _mm512_andnot_si512(other, *this); }
+    simdjson_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
   };
 
   // Forward-declared so they can be used by splat and friends.
@@ -14243,15 +14059,17 @@ namespace simd {
     typedef uint32_t bitmask_t;
     typedef uint64_t bitmask2_t;
 
-    simdjson_really_inline base8() : base<simd8<T>>() {}
-    simdjson_really_inline base8(const __m512i _value) : base<simd8<T>>(_value) {}
+    simdjson_inline base8() : base<simd8<T>>() {}
+    simdjson_inline base8(const __m512i _value) : base<simd8<T>>(_value) {}
 
-    simdjson_really_inline uint64_t operator==(const simd8<T> other) const { return _mm512_cmpeq_epi8_mask(*this, other); }
+    friend simdjson_really_inline uint64_t operator==(const simd8<T> lhs, const simd8<T> rhs) {
+      return _mm512_cmpeq_epi8_mask(lhs, rhs);
+    }
 
     static const int SIZE = sizeof(base<T>::value);
 
     template<int N=1>
-    simdjson_really_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+    simdjson_inline simd8<T> prev(const simd8<T> prev_chunk) const {
 #if SIMDJSON_GCC8
      // workaround for compilers unable to figure out that 16 - N is a constant (GCC 8)
       constexpr int shift = 16 - N;
@@ -14265,25 +14083,25 @@ namespace simd {
   // SIMD byte mask type (returned by things like eq and gt)
   template<>
   struct simd8<bool>: base8<bool> {
-    static simdjson_really_inline simd8<bool> splat(bool _value) { return _mm512_set1_epi8(uint8_t(-(!!_value))); }
+    static simdjson_inline simd8<bool> splat(bool _value) { return _mm512_set1_epi8(uint8_t(-(!!_value))); }
 
-    simdjson_really_inline simd8<bool>() : base8() {}
-    simdjson_really_inline simd8<bool>(const __m512i _value) : base8<bool>(_value) {}
+    simdjson_inline simd8<bool>() : base8() {}
+    simdjson_inline simd8<bool>(const __m512i _value) : base8<bool>(_value) {}
     // Splat constructor
-    simdjson_really_inline simd8<bool>(bool _value) : base8<bool>(splat(_value)) {}
-    simdjson_really_inline bool any() const { return !!_mm512_test_epi8_mask (*this, *this); }
-    simdjson_really_inline simd8<bool> operator~() const { return *this ^ true; }
+    simdjson_inline simd8<bool>(bool _value) : base8<bool>(splat(_value)) {}
+    simdjson_inline bool any() const { return !!_mm512_test_epi8_mask (*this, *this); }
+    simdjson_inline simd8<bool> operator~() const { return *this ^ true; }
   };
 
   template<typename T>
   struct base8_numeric: base8<T> {
-    static simdjson_really_inline simd8<T> splat(T _value) { return _mm512_set1_epi8(_value); }
-    static simdjson_really_inline simd8<T> zero() { return _mm512_setzero_si512(); }
-    static simdjson_really_inline simd8<T> load(const T values[64]) {
+    static simdjson_inline simd8<T> splat(T _value) { return _mm512_set1_epi8(_value); }
+    static simdjson_inline simd8<T> zero() { return _mm512_setzero_si512(); }
+    static simdjson_inline simd8<T> load(const T values[64]) {
       return _mm512_loadu_si512(reinterpret_cast<const __m512i *>(values));
     }
     // Repeat 16 values as many times as necessary (usually for lookup tables)
-    static simdjson_really_inline simd8<T> repeat_16(
+    static simdjson_inline simd8<T> repeat_16(
       T v0,  T v1,  T v2,  T v3,  T v4,  T v5,  T v6,  T v7,
       T v8,  T v9,  T v10, T v11, T v12, T v13, T v14, T v15
     ) {
@@ -14299,24 +14117,24 @@ namespace simd {
       );
     }
 
-    simdjson_really_inline base8_numeric() : base8<T>() {}
-    simdjson_really_inline base8_numeric(const __m512i _value) : base8<T>(_value) {}
+    simdjson_inline base8_numeric() : base8<T>() {}
+    simdjson_inline base8_numeric(const __m512i _value) : base8<T>(_value) {}
 
     // Store to array
-    simdjson_really_inline void store(T dst[64]) const { return _mm512_storeu_si512(reinterpret_cast<__m512i *>(dst), *this); }
+    simdjson_inline void store(T dst[64]) const { return _mm512_storeu_si512(reinterpret_cast<__m512i *>(dst), *this); }
 
     // Addition/subtraction are the same for signed and unsigned
-    simdjson_really_inline simd8<T> operator+(const simd8<T> other) const { return _mm512_add_epi8(*this, other); }
-    simdjson_really_inline simd8<T> operator-(const simd8<T> other) const { return _mm512_sub_epi8(*this, other); }
-    simdjson_really_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
-    simdjson_really_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
+    simdjson_inline simd8<T> operator+(const simd8<T> other) const { return _mm512_add_epi8(*this, other); }
+    simdjson_inline simd8<T> operator-(const simd8<T> other) const { return _mm512_sub_epi8(*this, other); }
+    simdjson_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
+    simdjson_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
 
     // Override to distinguish from bool version
-    simdjson_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+    simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
 
     // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
     template<typename L>
-    simdjson_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
       return _mm512_shuffle_epi8(lookup_table, *this);
     }
 
@@ -14328,12 +14146,12 @@ namespace simd {
     // signature simd8<L> compress(uint32_t mask) would be
     // sensible, but the AVX ISA makes this kind of approach difficult.
     template<typename L>
-    simdjson_really_inline void compress(uint64_t mask, L * output) const {
+    simdjson_inline void compress(uint64_t mask, L * output) const {
       _mm512_mask_compressstoreu_epi8 (output,~mask,*this);
     }
 
     template<typename L>
-    simdjson_really_inline simd8<L> lookup_16(
+    simdjson_inline simd8<L> lookup_16(
         L replace0,  L replace1,  L replace2,  L replace3,
         L replace4,  L replace5,  L replace6,  L replace7,
         L replace8,  L replace9,  L replace10, L replace11,
@@ -14350,14 +14168,14 @@ namespace simd {
   // Signed bytes
   template<>
   struct simd8<int8_t> : base8_numeric<int8_t> {
-    simdjson_really_inline simd8() : base8_numeric<int8_t>() {}
-    simdjson_really_inline simd8(const __m512i _value) : base8_numeric<int8_t>(_value) {}
+    simdjson_inline simd8() : base8_numeric<int8_t>() {}
+    simdjson_inline simd8(const __m512i _value) : base8_numeric<int8_t>(_value) {}
     // Splat constructor
-    simdjson_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
     // Array constructor
-    simdjson_really_inline simd8(const int8_t values[64]) : simd8(load(values)) {}
+    simdjson_inline simd8(const int8_t values[64]) : simd8(load(values)) {}
     // Member-by-member initialization
-    simdjson_really_inline simd8(
+    simdjson_inline simd8(
       int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
       int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15,
       int8_t v16, int8_t v17, int8_t v18, int8_t v19, int8_t v20, int8_t v21, int8_t v22, int8_t v23,
@@ -14378,7 +14196,7 @@ namespace simd {
     )) {}
 
     // Repeat 16 values as many times as necessary (usually for lookup tables)
-    simdjson_really_inline static simd8<int8_t> repeat_16(
+    simdjson_inline static simd8<int8_t> repeat_16(
       int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
       int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
     ) {
@@ -14395,24 +14213,24 @@ namespace simd {
     }
 
     // Order-sensitive comparisons
-    simdjson_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return _mm512_max_epi8(*this, other); }
-    simdjson_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return _mm512_min_epi8(*this, other); }
+    simdjson_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return _mm512_max_epi8(*this, other); }
+    simdjson_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return _mm512_min_epi8(*this, other); }
 
-    simdjson_really_inline simd8<bool> operator>(const simd8<int8_t> other) const { return _mm512_maskz_abs_epi8(_mm512_cmpgt_epi8_mask(*this, other),_mm512_set1_epi8(uint8_t(0x80))); }
-    simdjson_really_inline simd8<bool> operator<(const simd8<int8_t> other) const { return _mm512_maskz_abs_epi8(_mm512_cmpgt_epi8_mask(other, *this),_mm512_set1_epi8(uint8_t(0x80))); }
+    simdjson_inline simd8<bool> operator>(const simd8<int8_t> other) const { return _mm512_maskz_abs_epi8(_mm512_cmpgt_epi8_mask(*this, other),_mm512_set1_epi8(uint8_t(0x80))); }
+    simdjson_inline simd8<bool> operator<(const simd8<int8_t> other) const { return _mm512_maskz_abs_epi8(_mm512_cmpgt_epi8_mask(other, *this),_mm512_set1_epi8(uint8_t(0x80))); }
   };
 
   // Unsigned bytes
   template<>
   struct simd8<uint8_t>: base8_numeric<uint8_t> {
-    simdjson_really_inline simd8() : base8_numeric<uint8_t>() {}
-    simdjson_really_inline simd8(const __m512i _value) : base8_numeric<uint8_t>(_value) {}
+    simdjson_inline simd8() : base8_numeric<uint8_t>() {}
+    simdjson_inline simd8(const __m512i _value) : base8_numeric<uint8_t>(_value) {}
     // Splat constructor
-    simdjson_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
     // Array constructor
-    simdjson_really_inline simd8(const uint8_t values[64]) : simd8(load(values)) {}
+    simdjson_inline simd8(const uint8_t values[64]) : simd8(load(values)) {}
     // Member-by-member initialization
-    simdjson_really_inline simd8(
+    simdjson_inline simd8(
       uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
       uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15,
       uint8_t v16, uint8_t v17, uint8_t v18, uint8_t v19, uint8_t v20, uint8_t v21, uint8_t v22, uint8_t v23,
@@ -14433,7 +14251,7 @@ namespace simd {
     )) {}
 
     // Repeat 16 values as many times as necessary (usually for lookup tables)
-    simdjson_really_inline static simd8<uint8_t> repeat_16(
+    simdjson_inline static simd8<uint8_t> repeat_16(
       uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
       uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
     ) {
@@ -14450,42 +14268,42 @@ namespace simd {
     }
 
     // Saturated math
-    simdjson_really_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return _mm512_adds_epu8(*this, other); }
-    simdjson_really_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return _mm512_subs_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return _mm512_adds_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return _mm512_subs_epu8(*this, other); }
 
     // Order-specific operations
-    simdjson_really_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return _mm512_max_epu8(*this, other); }
-    simdjson_really_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return _mm512_min_epu8(other, *this); }
+    simdjson_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return _mm512_max_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return _mm512_min_epu8(other, *this); }
     // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
-    simdjson_really_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
     // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
-    simdjson_really_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
-    simdjson_really_inline uint64_t operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
-    simdjson_really_inline uint64_t operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
-    simdjson_really_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
-    simdjson_really_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->lt_bits(other).any_bits_set(); }
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
+    simdjson_inline uint64_t operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
+    simdjson_inline uint64_t operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->lt_bits(other).any_bits_set(); }
 
     // Bit-specific operations
-    simdjson_really_inline simd8<bool> bits_not_set() const { return _mm512_mask_blend_epi8(*this == uint8_t(0), _mm512_set1_epi8(0), _mm512_set1_epi8(-1)); }
-    simdjson_really_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
-    simdjson_really_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
-    simdjson_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
+    simdjson_inline simd8<bool> bits_not_set() const { return _mm512_mask_blend_epi8(*this == uint8_t(0), _mm512_set1_epi8(0), _mm512_set1_epi8(-1)); }
+    simdjson_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
 
-    simdjson_really_inline bool is_ascii() const { return _mm512_movepi8_mask(*this) == 0; }
-    simdjson_really_inline bool bits_not_set_anywhere() const {
+    simdjson_inline bool is_ascii() const { return _mm512_movepi8_mask(*this) == 0; }
+    simdjson_inline bool bits_not_set_anywhere() const {
       return !_mm512_test_epi8_mask(*this, *this);
     }
-    simdjson_really_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
-    simdjson_really_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const { return !_mm512_test_epi8_mask(*this, bits); }
-    simdjson_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
+    simdjson_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
+    simdjson_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const { return !_mm512_test_epi8_mask(*this, bits); }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
     template<int N>
-    simdjson_really_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(_mm512_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); }
+    simdjson_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(_mm512_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); }
     template<int N>
-    simdjson_really_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(_mm512_slli_epi16(*this, N)) & uint8_t(0xFFu << N); }
+    simdjson_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(_mm512_slli_epi16(*this, N)) & uint8_t(0xFFu << N); }
     // Get one of the bits and make a bitmask out of it.
     // e.g. value.get_bit<7>() gets the high bit
     template<int N>
-    simdjson_really_inline uint64_t get_bit() const { return _mm512_movepi8_mask(_mm512_slli_epi16(*this, 7-N)); }
+    simdjson_inline uint64_t get_bit() const { return _mm512_movepi8_mask(_mm512_slli_epi16(*this, 7-N)); }
   };
 
   template<typename T>
@@ -14498,40 +14316,40 @@ namespace simd {
     simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
     simd8x64() = delete; // no default constructor allowed
 
-    simdjson_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1) : chunks{chunk0, chunk1} {}
-    simdjson_really_inline simd8x64(const simd8<T> chunk0) : chunks{chunk0} {}
-    simdjson_really_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr)} {}
+    simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1) : chunks{chunk0, chunk1} {}
+    simdjson_inline simd8x64(const simd8<T> chunk0) : chunks{chunk0} {}
+    simdjson_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr)} {}
 
-    simdjson_really_inline uint64_t compress(uint64_t mask, T * output) const {
+    simdjson_inline uint64_t compress(uint64_t mask, T * output) const {
       this->chunks[0].compress(mask, output);
       return 64 - count_ones(mask);
     }
 
-    simdjson_really_inline void store(T ptr[64]) const {
+    simdjson_inline void store(T ptr[64]) const {
       this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
     }
 
-    simdjson_really_inline simd8<T> reduce_or() const {
+    simdjson_inline simd8<T> reduce_or() const {
       return this->chunks[0];
     }
 
-    simdjson_really_inline simd8x64<T> bit_or(const T m) const {
+    simdjson_inline simd8x64<T> bit_or(const T m) const {
       const simd8<T> mask = simd8<T>::splat(m);
       return simd8x64<T>(
         this->chunks[0] | mask
       );
     }
 
-    simdjson_really_inline uint64_t eq(const T m) const {
+    simdjson_inline uint64_t eq(const T m) const {
       const simd8<T> mask = simd8<T>::splat(m);
       return this->chunks[0] == mask;
     }
 
-    simdjson_really_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+    simdjson_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
       return this->chunks[0] == other.chunks[0];
     }
 
-    simdjson_really_inline uint64_t lteq(const T m) const {
+    simdjson_inline uint64_t lteq(const T m) const {
       const simd8<T> mask = simd8<T>::splat(m);
       return this->chunks[0] <= mask;
     }
@@ -14554,11 +14372,11 @@ namespace jsoncharutils {
 
 // return non-zero if not a structural or whitespace char
 // zero otherwise
-simdjson_really_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
+simdjson_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
   return internal::structural_or_whitespace_negated[c];
 }
 
-simdjson_really_inline uint32_t is_structural_or_whitespace(uint8_t c) {
+simdjson_inline uint32_t is_structural_or_whitespace(uint8_t c) {
   return internal::structural_or_whitespace[c];
 }
 
@@ -14589,7 +14407,7 @@ static inline uint32_t hex_to_u32_nocheck(
 //
 // Note: we assume that surrogates are treated separately
 //
-simdjson_really_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
+simdjson_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
   if (cp <= 0x7F) {
     c[0] = uint8_t(cp);
     return 1; // ascii
@@ -14621,10 +14439,10 @@ simdjson_really_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
 #ifdef SIMDJSON_IS_32BITS // _umul128 for x86, arm
 // this is a slow emulation routine for 32-bit
 //
-static simdjson_really_inline uint64_t __emulu(uint32_t x, uint32_t y) {
+static simdjson_inline uint64_t __emulu(uint32_t x, uint32_t y) {
   return x * (uint64_t)y;
 }
-static simdjson_really_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+static simdjson_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
   uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
   uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
   uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
@@ -14638,7 +14456,7 @@ static simdjson_really_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64
 
 using internal::value128;
 
-simdjson_really_inline value128 full_multiplication(uint64_t value1, uint64_t value2) {
+simdjson_inline value128 full_multiplication(uint64_t value1, uint64_t value2) {
   value128 answer;
 #if defined(SIMDJSON_REGULAR_VISUAL_STUDIO) || defined(SIMDJSON_IS_32BITS)
 #ifdef _M_ARM64
@@ -14674,13 +14492,13 @@ namespace atomparsing {
 // You might think that using memcpy makes this function expensive, but you'd be wrong.
 // All decent optimizing compilers (GCC, clang, Visual Studio) will compile string_to_uint32("false");
 // to the compile-time constant 1936482662.
-simdjson_really_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
+simdjson_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
 
 
 // Again in str4ncmp we use a memcpy to avoid undefined behavior. The memcpy may appear expensive.
 // Yet all decent optimizing compilers will compile memcpy to a single instruction, just about.
 simdjson_warn_unused
-simdjson_really_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
+simdjson_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
   uint32_t srcval; // we want to avoid unaligned 32-bit loads (undefined in C/C++)
   static_assert(sizeof(uint32_t) <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be larger than 4 bytes");
   std::memcpy(&srcval, src, sizeof(uint32_t));
@@ -14688,36 +14506,36 @@ simdjson_really_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_true_atom(const uint8_t *src) {
+simdjson_inline bool is_valid_true_atom(const uint8_t *src) {
   return (str4ncmp(src, "true") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
+simdjson_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
   if (len > 4) { return is_valid_true_atom(src); }
   else if (len == 4) { return !str4ncmp(src, "true"); }
   else { return false; }
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_false_atom(const uint8_t *src) {
+simdjson_inline bool is_valid_false_atom(const uint8_t *src) {
   return (str4ncmp(src+1, "alse") | jsoncharutils::is_not_structural_or_whitespace(src[5])) == 0;
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
+simdjson_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
   if (len > 5) { return is_valid_false_atom(src); }
   else if (len == 5) { return !str4ncmp(src+1, "alse"); }
   else { return false; }
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_null_atom(const uint8_t *src) {
+simdjson_inline bool is_valid_null_atom(const uint8_t *src) {
   return (str4ncmp(src, "null") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
+simdjson_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
   if (len > 4) { return is_valid_null_atom(src); }
   else if (len == 4) { return !str4ncmp(src, "null"); }
   else { return false; }
@@ -14743,18 +14561,18 @@ using namespace simd;
 struct backslash_and_quote {
 public:
   static constexpr uint32_t BYTES_PROCESSED = 32;
-  simdjson_really_inline static backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+  simdjson_inline static backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
 
-  simdjson_really_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
-  simdjson_really_inline bool has_backslash() { return ((quote_bits - 1) & bs_bits) != 0; }
-  simdjson_really_inline int quote_index() { return trailing_zeroes(quote_bits); }
-  simdjson_really_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+  simdjson_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_inline bool has_backslash() { return ((quote_bits - 1) & bs_bits) != 0; }
+  simdjson_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_inline int backslash_index() { return trailing_zeroes(bs_bits); }
 
   uint64_t bs_bits;
   uint64_t quote_bits;
 }; // struct backslash_and_quote
 
-simdjson_really_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
   // this can read up to 15 bytes beyond the buffer size, but we require
   // SIMDJSON_PADDING of padding
   static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
@@ -14771,147 +14589,6 @@ simdjson_really_inline backslash_and_quote backslash_and_quote::copy_and_find(co
 } // namespace icelake
 } // namespace simdjson
 
-/* begin file include/simdjson/generic/stringparsing.h */
-// This file contains the common code every implementation uses
-// It is intended to be included multiple times and compiled multiple times
-
-namespace simdjson {
-namespace icelake {
-namespace {
-/// @private
-namespace stringparsing {
-
-// begin copypasta
-// These chars yield themselves: " \ /
-// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
-// u not handled in this table as it's complex
-static const uint8_t escape_map[256] = {
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
-    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
-    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
-
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-};
-
-// handle a unicode codepoint
-// write appropriate values into dest
-// src will advance 6 bytes or 12 bytes
-// dest will advance a variable amount (return via pointer)
-// return true if the unicode codepoint was valid
-// We work in little-endian then swap at write time
-simdjson_warn_unused
-simdjson_really_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
-                                            uint8_t **dst_ptr) {
-  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
-  // conversion isn't valid; we defer the check for this to inside the
-  // multilingual plane check
-  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
-  *src_ptr += 6;
-  // check for low surrogate for characters outside the Basic
-  // Multilingual Plane.
-  if (code_point >= 0xd800 && code_point < 0xdc00) {
-    if (((*src_ptr)[0] != '\\') || (*src_ptr)[1] != 'u') {
-      return false;
-    }
-    uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
-
-    // if the first code point is invalid we will get here, as we will go past
-    // the check for being outside the Basic Multilingual plane. If we don't
-    // find a \u immediately afterwards we fail out anyhow, but if we do,
-    // this check catches both the case of the first code point being invalid
-    // or the second code point being invalid.
-    if ((code_point | code_point_2) >> 16) {
-      return false;
-    }
-
-    code_point =
-        (((code_point - 0xd800) << 10) | (code_point_2 - 0xdc00)) + 0x10000;
-    *src_ptr += 6;
-  }
-  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
-  *dst_ptr += offset;
-  return offset > 0;
-}
-
-/**
- * Unescape a string from src to dst, stopping at a final unescaped quote. E.g., if src points at 'joe"', then
- * dst needs to have four free bytes.
- */
-simdjson_warn_unused simdjson_really_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst) {
-  while (1) {
-    // Copy the next n bytes, and find the backslash and quote in them.
-    auto bs_quote = backslash_and_quote::copy_and_find(src, dst);
-    // If the next thing is the end quote, copy and return
-    if (bs_quote.has_quote_first()) {
-      // we encountered quotes first. Move dst to point to quotes and exit
-      return dst + bs_quote.quote_index();
-    }
-    if (bs_quote.has_backslash()) {
-      /* find out where the backspace is */
-      auto bs_dist = bs_quote.backslash_index();
-      uint8_t escape_char = src[bs_dist + 1];
-      /* we encountered backslash first. Handle backslash */
-      if (escape_char == 'u') {
-        /* move src/dst up to the start; they will be further adjusted
-           within the unicode codepoint handling code. */
-        src += bs_dist;
-        dst += bs_dist;
-        if (!handle_unicode_codepoint(&src, &dst)) {
-          return nullptr;
-        }
-      } else {
-        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
-         * write bs_dist+1 characters to output
-         * note this may reach beyond the part of the buffer we've actually
-         * seen. I think this is ok */
-        uint8_t escape_result = escape_map[escape_char];
-        if (escape_result == 0u) {
-          return nullptr; /* bogus escape value is an error */
-        }
-        dst[bs_dist] = escape_result;
-        src += bs_dist + 2;
-        dst += bs_dist + 1;
-      }
-    } else {
-      /* they are the same. Since they can't co-occur, it means we
-       * encountered neither. */
-      src += backslash_and_quote::BYTES_PROCESSED;
-      dst += backslash_and_quote::BYTES_PROCESSED;
-    }
-  }
-  /* can't be reached */
-  return nullptr;
-}
-
-simdjson_unused simdjson_warn_unused simdjson_really_inline error_code parse_string_to_buffer(const uint8_t *src, uint8_t *&current_string_buf_loc, std::string_view &s) {
-  if (*(src++) != '"') { return STRING_ERROR; }
-  auto end = stringparsing::parse_string(src, current_string_buf_loc);
-  if (!end) { return STRING_ERROR; }
-  s = std::string_view(reinterpret_cast<const char *>(current_string_buf_loc), end-current_string_buf_loc);
-  current_string_buf_loc = end;
-  return SUCCESS;
-}
-
-} // namespace stringparsing
-} // unnamed namespace
-} // namespace icelake
-} // namespace simdjson
-/* end file include/simdjson/generic/stringparsing.h */
-
 #endif // SIMDJSON_ICELAKE_STRINGPARSING_H
 /* end file include/simdjson/icelake/stringparsing.h */
 /* begin file include/simdjson/icelake/numberparsing.h */
@@ -14922,7 +14599,7 @@ namespace simdjson {
 namespace icelake {
 namespace {
 
-static simdjson_really_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
   // this actually computes *16* values so we are being wasteful.
   const __m128i ascii0 = _mm_set1_epi8('0');
   const __m128i mul_1_10 =
@@ -14985,7 +14662,7 @@ namespace {
 // Convert a mantissa, an exponent and a sign bit into an ieee64 double.
 // The real_exponent needs to be in [0, 2046] (technically real_exponent = 2047 would be acceptable).
 // The mantissa should be in [0,1<<53). The bit at index (1ULL << 52) while be zeroed.
-simdjson_really_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
+simdjson_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
     double d;
     mantissa &= ~(1ULL << 52);
     mantissa |= real_exponent << 52;
@@ -15000,7 +14677,7 @@ simdjson_really_inline double to_double(uint64_t mantissa, uint64_t real_exponen
 // set to false. This should work *most of the time* (like 99% of the time).
 // We assume that power is in the [smallest_power,
 // largest_power] interval: the caller is responsible for this check.
-simdjson_really_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
+simdjson_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
   // we start with a fast path
   // It was described in
   // Clinger WD. How to read floating point numbers accurately.
@@ -15060,7 +14737,7 @@ simdjson_really_inline bool compute_float_64(int64_t power, uint64_t i, bool neg
   // In the slow path, we need to adjust i so that it is > 1<<63 which is always
   // possible, except if i == 0, so we handle i == 0 separately.
   if(i == 0) {
-    d = 0.0;
+    d = negative ? -0.0 : 0.0;
     return true;
   }
 
@@ -15175,7 +14852,7 @@ simdjson_really_inline bool compute_float_64(int64_t power, uint64_t i, bool neg
   if (simdjson_unlikely(real_exponent <= 0)) { // we have a subnormal?
     // Here have that real_exponent <= 0 so -real_exponent >= 0
     if(-real_exponent + 1 >= 64) { // if we have more than 64 bits below the minimum exponent, you have a zero for sure.
-      d = 0.0;
+      d = negative ? -0.0 : 0.0;
       return true;
     }
     // next line is safe because -real_exponent + 1 < 0
@@ -15283,7 +14960,7 @@ static bool parse_float_fallback(const uint8_t *ptr, const uint8_t *end_ptr, dou
 // check quickly whether the next 8 chars are made of digits
 // at a glance, it looks better than Mula's
 // http://0x80.pl/articles/swar-digits-validate.html
-simdjson_really_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
+simdjson_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
   uint64_t val;
   // this can read up to 7 bytes beyond the buffer size, but we require
   // SIMDJSON_PADDING of padding
@@ -15310,7 +14987,7 @@ error_code slow_float_parsing(simdjson_unused const uint8_t * src, W writer) {
 
 template<typename I>
 SIMDJSON_NO_SANITIZE_UNDEFINED // We deliberately allow overflow here and check later
-simdjson_really_inline bool parse_digit(const uint8_t c, I &i) {
+simdjson_inline bool parse_digit(const uint8_t c, I &i) {
   const uint8_t digit = static_cast<uint8_t>(c - '0');
   if (digit > 9) {
     return false;
@@ -15320,7 +14997,7 @@ simdjson_really_inline bool parse_digit(const uint8_t c, I &i) {
   return true;
 }
 
-simdjson_really_inline error_code parse_decimal(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
+simdjson_inline error_code parse_decimal(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
   // we continue with the fiction that we have an integer. If the
   // floating point number is representable as x * 10^z for some integer
   // z that fits in 53 bits, then we will be able to convert back the
@@ -15348,7 +15025,7 @@ simdjson_really_inline error_code parse_decimal(simdjson_unused const uint8_t *c
   return SUCCESS;
 }
 
-simdjson_really_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
+simdjson_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
   // Exp Sign: -123.456e[-]78
   bool neg_exp = ('-' == *p);
   if (neg_exp || '+' == *p) { p++; } // Skip + as well
@@ -15399,7 +15076,7 @@ simdjson_really_inline error_code parse_exponent(simdjson_unused const uint8_t *
   return SUCCESS;
 }
 
-simdjson_really_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
+simdjson_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
   // It is possible that the integer had an overflow.
   // We have to handle the case where we have 0.0000somenumber.
   const uint8_t *start = start_digits;
@@ -15409,7 +15086,7 @@ simdjson_really_inline size_t significant_digits(const uint8_t * start_digits, s
 }
 
 template<typename W>
-simdjson_really_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
+simdjson_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
   // If we frequently had to deal with long strings of digits,
   // we could extend our code by using a 128-bit integer instead
   // of a 64-bit integer. However, this is uncommon in practice.
@@ -15445,7 +15122,8 @@ simdjson_really_inline error_code write_float(const uint8_t *const src, bool neg
     static_assert(simdjson::internal::smallest_power <= -342, "smallest_power is not small enough");
     //
     if((exponent < simdjson::internal::smallest_power) || (i == 0)) {
-      WRITE_DOUBLE(0, src, writer);
+      // E.g. Parse "-0.0e-999" into the same value as "-0.0". See https://en.wikipedia.org/wiki/Signed_zero
+      WRITE_DOUBLE(negative ? -0.0 : 0.0, src, writer);
       return SUCCESS;
     } else { // (exponent > largest_power) and (i != 0)
       // We have, for sure, an infinite value and simdjson refuses to parse infinite values.
@@ -15465,20 +15143,20 @@ simdjson_really_inline error_code write_float(const uint8_t *const src, bool neg
 #ifdef SIMDJSON_SKIPNUMBERPARSING
 
 template<typename W>
-simdjson_really_inline error_code parse_number(const uint8_t *const, W &writer) {
+simdjson_inline error_code parse_number(const uint8_t *const, W &writer) {
   writer.append_s64(0);        // always write zero
   return SUCCESS;              // always succeeds
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
-simdjson_unused simdjson_really_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
-simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept { return ondemand::number_type::signed_integer; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept { return ondemand::number_type::signed_integer; }
 #else
 
 // parse the number at src
@@ -15491,13 +15169,13 @@ simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> ge
 //
 // Our objective is accurate parsing (ULP of 0) at high speed.
 template<typename W>
-simdjson_really_inline error_code parse_number(const uint8_t *const src, W &writer) {
+simdjson_inline error_code parse_number(const uint8_t *const src, W &writer) {
 
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  const uint8_t *p = src + negative;
+  const uint8_t *p = src + uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -15641,7 +15319,7 @@ const uint8_t integer_string_finisher[256] = {
     NUMBER_ERROR};
 
 // Parse any number from 0 to 18,446,744,073,709,551,615
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
   const uint8_t *p = src;
   //
   // Parse the integer part.
@@ -15691,7 +15369,7 @@ simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(
 
 // Parse any number from 0 to 18,446,744,073,709,551,615
 // Never read at src_end or beyond
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
   const uint8_t *p = src;
   //
   // Parse the integer part.
@@ -15739,7 +15417,7 @@ simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(
 }
 
 // Parse any number from 0 to 18,446,744,073,709,551,615
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
   const uint8_t *p = src + 1;
   //
   // Parse the integer part.
@@ -15789,12 +15467,12 @@ simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned_
 }
 
 // Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  const uint8_t *p = src + negative;
+  const uint8_t *p = src + uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -15832,13 +15510,13 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(co
 
 // Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
 // Never read at src_end or beyond
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
   //
   // Check for minus sign
   //
   if(src == src_end) { return NUMBER_ERROR; }
   bool negative = (*src == '-');
-  const uint8_t *p = src + negative;
+  const uint8_t *p = src + uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -15875,24 +15553,24 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(co
 }
 
 // Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*(src + 1) == '-');
-  const uint8_t *p = src + negative + 1;
+  src += uint8_t(negative) + 1;
 
   //
   // Parse the integer part.
   //
   // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
-  const uint8_t *const start_digits = p;
+  const uint8_t *const start_digits = src;
   uint64_t i = 0;
-  while (parse_digit(*p, i)) { p++; }
+  while (parse_digit(*src, i)) { src++; }
 
   // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
   // Optimization note: size_t is expected to be unsigned.
-  size_t digit_count = size_t(p - start_digits);
+  size_t digit_count = size_t(src - start_digits);
   // We go from
   // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
   // so we can never represent numbers that have more than 19 digits.
@@ -15904,11 +15582,11 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in
   // Here digit_count > 0.
   if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
   // We can do the following...
-  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
-  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // if (!jsoncharutils::is_structural_or_whitespace(*src)) {
+  //  return (*src == '.' || *src == 'e' || *src == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
   // }
   // as a single table lookup:
-  if(*p != '"') { return NUMBER_ERROR; }
+  if(*src != '"') { return NUMBER_ERROR; }
   // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
   // Performance note: This check is only needed when digit_count == longest_digit_count but it is
   // so cheap that we might as well always make it.
@@ -15916,12 +15594,12 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in
   return negative ? (~i+1) : i;
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -15988,19 +15666,19 @@ simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(cons
   if (simdjson_likely(!overflow)) {
     if (compute_float_64(exponent, i, negative, d)) { return d; }
   }
-  if (!parse_float_fallback(src-negative, &d)) {
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
     return NUMBER_ERROR;
   }
   return d;
 }
 
-simdjson_unused simdjson_really_inline bool is_negative(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept {
   return (*src == '-');
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
   const uint8_t *p = src;
   while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
   if ( p == src ) { return NUMBER_ERROR; }
@@ -16008,9 +15686,9 @@ simdjson_unused simdjson_really_inline simdjson_result<bool> is_integer(const ui
   return false;
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept {
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
   const uint8_t *p = src;
   while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
   if ( p == src ) { return NUMBER_ERROR; }
@@ -16034,13 +15712,13 @@ simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> ge
 }
 
 // Never read at src_end or beyond
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
   if(src == src_end) { return NUMBER_ERROR; }
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -16109,18 +15787,18 @@ simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(cons
   if (simdjson_likely(!overflow)) {
     if (compute_float_64(exponent, i, negative, d)) { return d; }
   }
-  if (!parse_float_fallback(src-negative, src_end, &d)) {
+  if (!parse_float_fallback(src - uint8_t(negative), src_end, &d)) {
     return NUMBER_ERROR;
   }
   return d;
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*(src + 1) == '-');
-  src += negative + 1;
+  src += uint8_t(negative) + 1;
 
   //
   // Parse the integer part.
@@ -16187,7 +15865,7 @@ simdjson_unused simdjson_really_inline simdjson_result<double> parse_double_in_s
   if (simdjson_likely(!overflow)) {
     if (compute_float_64(exponent, i, negative, d)) { return d; }
   }
-  if (!parse_float_fallback(src-negative, &d)) {
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
     return NUMBER_ERROR;
   }
   return d;
@@ -16249,7 +15927,7 @@ using namespace simdjson;
 
 class implementation final : public simdjson::implementation {
 public:
-  simdjson_really_inline implementation() : simdjson::implementation(
+  simdjson_inline implementation() : simdjson::implementation(
       "haswell",
       "Intel/AMD AVX2",
       internal::instruction_set::AVX2 | internal::instruction_set::PCLMULQDQ | internal::instruction_set::BMI1 | internal::instruction_set::BMI2
@@ -16315,6 +15993,8 @@ class implementation final : public simdjson::implementation {
 #endif //  _blsr_u64
 #endif // SIMDJSON_CLANG_VISUAL_STUDIO
 
+static_assert(sizeof(__m256i) <= simdjson::SIMDJSON_PADDING, "insufficient padding for haswell kernel.");
+
 #endif // SIMDJSON_HASWELL_INTRINSICS_H
 /* end file include/simdjson/haswell/intrinsics.h */
 
@@ -16364,10 +16044,11 @@ class dom_parser_implementation final : public internal::dom_parser_implementati
   simdjson_warn_unused error_code stage1(const uint8_t *buf, size_t len, stage1_mode partial) noexcept final;
   simdjson_warn_unused error_code stage2(dom::document &doc) noexcept final;
   simdjson_warn_unused error_code stage2_next(dom::document &doc) noexcept final;
+  simdjson_warn_unused uint8_t *parse_string(const uint8_t *src, uint8_t *dst) const noexcept final;
   inline simdjson_warn_unused error_code set_capacity(size_t capacity) noexcept final;
   inline simdjson_warn_unused error_code set_max_depth(size_t max_depth) noexcept final;
 private:
-  simdjson_really_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
+  simdjson_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
 
 };
 
@@ -16420,7 +16101,7 @@ namespace {
 // but the algorithms do not end up using the returned value.
 // Sadly, sanitizers are not smart enough to figure it out.
 SIMDJSON_NO_SANITIZE_UNDEFINED
-simdjson_really_inline int trailing_zeroes(uint64_t input_num) {
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
   return (int)_tzcnt_u64(input_num);
 #else // SIMDJSON_REGULAR_VISUAL_STUDIO
@@ -16434,27 +16115,27 @@ simdjson_really_inline int trailing_zeroes(uint64_t input_num) {
 }
 
 /* result might be undefined when input_num is zero */
-simdjson_really_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
   return _blsr_u64(input_num);
 }
 
 /* result might be undefined when input_num is zero */
-simdjson_really_inline int leading_zeroes(uint64_t input_num) {
+simdjson_inline int leading_zeroes(uint64_t input_num) {
   return int(_lzcnt_u64(input_num));
 }
 
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
-simdjson_really_inline unsigned __int64 count_ones(uint64_t input_num) {
+simdjson_inline unsigned __int64 count_ones(uint64_t input_num) {
   // note: we do not support legacy 32-bit Windows
   return __popcnt64(input_num);// Visual Studio wants two underscores
 }
 #else
-simdjson_really_inline long long int count_ones(uint64_t input_num) {
+simdjson_inline long long int count_ones(uint64_t input_num) {
   return _popcnt64(input_num);
 }
 #endif
 
-simdjson_really_inline bool add_overflow(uint64_t value1, uint64_t value2,
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2,
                                 uint64_t *result) {
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
   return _addcarry_u64(0, value1, value2,
@@ -16484,7 +16165,7 @@ namespace {
 //
 // For example, prefix_xor(00100100) == 00011100
 //
-simdjson_really_inline uint64_t prefix_xor(const uint64_t bitmask) {
+simdjson_inline uint64_t prefix_xor(const uint64_t bitmask) {
   // There should be no such thing with a processor supporting avx2
   // but not clmul.
   __m128i all_ones = _mm_set1_epi8('\xFF');
@@ -16514,23 +16195,23 @@ namespace simd {
     __m256i value;
 
     // Zero constructor
-    simdjson_really_inline base() : value{__m256i()} {}
+    simdjson_inline base() : value{__m256i()} {}
 
     // Conversion from SIMD register
-    simdjson_really_inline base(const __m256i _value) : value(_value) {}
+    simdjson_inline base(const __m256i _value) : value(_value) {}
 
     // Conversion to SIMD register
-    simdjson_really_inline operator const __m256i&() const { return this->value; }
-    simdjson_really_inline operator __m256i&() { return this->value; }
+    simdjson_inline operator const __m256i&() const { return this->value; }
+    simdjson_inline operator __m256i&() { return this->value; }
 
     // Bit operations
-    simdjson_really_inline Child operator|(const Child other) const { return _mm256_or_si256(*this, other); }
-    simdjson_really_inline Child operator&(const Child other) const { return _mm256_and_si256(*this, other); }
-    simdjson_really_inline Child operator^(const Child other) const { return _mm256_xor_si256(*this, other); }
-    simdjson_really_inline Child bit_andnot(const Child other) const { return _mm256_andnot_si256(other, *this); }
-    simdjson_really_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
-    simdjson_really_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
-    simdjson_really_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+    simdjson_inline Child operator|(const Child other) const { return _mm256_or_si256(*this, other); }
+    simdjson_inline Child operator&(const Child other) const { return _mm256_and_si256(*this, other); }
+    simdjson_inline Child operator^(const Child other) const { return _mm256_xor_si256(*this, other); }
+    simdjson_inline Child bit_andnot(const Child other) const { return _mm256_andnot_si256(other, *this); }
+    simdjson_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
   };
 
   // Forward-declared so they can be used by splat and friends.
@@ -16542,15 +16223,15 @@ namespace simd {
     typedef uint32_t bitmask_t;
     typedef uint64_t bitmask2_t;
 
-    simdjson_really_inline base8() : base<simd8<T>>() {}
-    simdjson_really_inline base8(const __m256i _value) : base<simd8<T>>(_value) {}
+    simdjson_inline base8() : base<simd8<T>>() {}
+    simdjson_inline base8(const __m256i _value) : base<simd8<T>>(_value) {}
 
     friend simdjson_really_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) { return _mm256_cmpeq_epi8(lhs, rhs); }
 
     static const int SIZE = sizeof(base<T>::value);
 
     template<int N=1>
-    simdjson_really_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+    simdjson_inline simd8<T> prev(const simd8<T> prev_chunk) const {
       return _mm256_alignr_epi8(*this, _mm256_permute2x128_si256(prev_chunk, *this, 0x21), 16 - N);
     }
   };
@@ -16558,27 +16239,27 @@ namespace simd {
   // SIMD byte mask type (returned by things like eq and gt)
   template<>
   struct simd8<bool>: base8<bool> {
-    static simdjson_really_inline simd8<bool> splat(bool _value) { return _mm256_set1_epi8(uint8_t(-(!!_value))); }
+    static simdjson_inline simd8<bool> splat(bool _value) { return _mm256_set1_epi8(uint8_t(-(!!_value))); }
 
-    simdjson_really_inline simd8<bool>() : base8() {}
-    simdjson_really_inline simd8<bool>(const __m256i _value) : base8<bool>(_value) {}
+    simdjson_inline simd8<bool>() : base8() {}
+    simdjson_inline simd8<bool>(const __m256i _value) : base8<bool>(_value) {}
     // Splat constructor
-    simdjson_really_inline simd8<bool>(bool _value) : base8<bool>(splat(_value)) {}
+    simdjson_inline simd8<bool>(bool _value) : base8<bool>(splat(_value)) {}
 
-    simdjson_really_inline int to_bitmask() const { return _mm256_movemask_epi8(*this); }
-    simdjson_really_inline bool any() const { return !_mm256_testz_si256(*this, *this); }
-    simdjson_really_inline simd8<bool> operator~() const { return *this ^ true; }
+    simdjson_inline int to_bitmask() const { return _mm256_movemask_epi8(*this); }
+    simdjson_inline bool any() const { return !_mm256_testz_si256(*this, *this); }
+    simdjson_inline simd8<bool> operator~() const { return *this ^ true; }
   };
 
   template<typename T>
   struct base8_numeric: base8<T> {
-    static simdjson_really_inline simd8<T> splat(T _value) { return _mm256_set1_epi8(_value); }
-    static simdjson_really_inline simd8<T> zero() { return _mm256_setzero_si256(); }
-    static simdjson_really_inline simd8<T> load(const T values[32]) {
+    static simdjson_inline simd8<T> splat(T _value) { return _mm256_set1_epi8(_value); }
+    static simdjson_inline simd8<T> zero() { return _mm256_setzero_si256(); }
+    static simdjson_inline simd8<T> load(const T values[32]) {
       return _mm256_loadu_si256(reinterpret_cast<const __m256i *>(values));
     }
     // Repeat 16 values as many times as necessary (usually for lookup tables)
-    static simdjson_really_inline simd8<T> repeat_16(
+    static simdjson_inline simd8<T> repeat_16(
       T v0,  T v1,  T v2,  T v3,  T v4,  T v5,  T v6,  T v7,
       T v8,  T v9,  T v10, T v11, T v12, T v13, T v14, T v15
     ) {
@@ -16590,24 +16271,24 @@ namespace simd {
       );
     }
 
-    simdjson_really_inline base8_numeric() : base8<T>() {}
-    simdjson_really_inline base8_numeric(const __m256i _value) : base8<T>(_value) {}
+    simdjson_inline base8_numeric() : base8<T>() {}
+    simdjson_inline base8_numeric(const __m256i _value) : base8<T>(_value) {}
 
     // Store to array
-    simdjson_really_inline void store(T dst[32]) const { return _mm256_storeu_si256(reinterpret_cast<__m256i *>(dst), *this); }
+    simdjson_inline void store(T dst[32]) const { return _mm256_storeu_si256(reinterpret_cast<__m256i *>(dst), *this); }
 
     // Addition/subtraction are the same for signed and unsigned
-    simdjson_really_inline simd8<T> operator+(const simd8<T> other) const { return _mm256_add_epi8(*this, other); }
-    simdjson_really_inline simd8<T> operator-(const simd8<T> other) const { return _mm256_sub_epi8(*this, other); }
-    simdjson_really_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
-    simdjson_really_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
+    simdjson_inline simd8<T> operator+(const simd8<T> other) const { return _mm256_add_epi8(*this, other); }
+    simdjson_inline simd8<T> operator-(const simd8<T> other) const { return _mm256_sub_epi8(*this, other); }
+    simdjson_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
+    simdjson_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
 
     // Override to distinguish from bool version
-    simdjson_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+    simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
 
     // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
     template<typename L>
-    simdjson_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
       return _mm256_shuffle_epi8(lookup_table, *this);
     }
 
@@ -16619,7 +16300,7 @@ namespace simd {
     // signature simd8<L> compress(uint32_t mask) would be
     // sensible, but the AVX ISA makes this kind of approach difficult.
     template<typename L>
-    simdjson_really_inline void compress(uint32_t mask, L * output) const {
+    simdjson_inline void compress(uint32_t mask, L * output) const {
       using internal::thintable_epi8;
       using internal::BitsSetTable256mul2;
       using internal::pshufb_combine_table;
@@ -16665,7 +16346,7 @@ namespace simd {
     }
 
     template<typename L>
-    simdjson_really_inline simd8<L> lookup_16(
+    simdjson_inline simd8<L> lookup_16(
         L replace0,  L replace1,  L replace2,  L replace3,
         L replace4,  L replace5,  L replace6,  L replace7,
         L replace8,  L replace9,  L replace10, L replace11,
@@ -16682,14 +16363,14 @@ namespace simd {
   // Signed bytes
   template<>
   struct simd8<int8_t> : base8_numeric<int8_t> {
-    simdjson_really_inline simd8() : base8_numeric<int8_t>() {}
-    simdjson_really_inline simd8(const __m256i _value) : base8_numeric<int8_t>(_value) {}
+    simdjson_inline simd8() : base8_numeric<int8_t>() {}
+    simdjson_inline simd8(const __m256i _value) : base8_numeric<int8_t>(_value) {}
     // Splat constructor
-    simdjson_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
     // Array constructor
-    simdjson_really_inline simd8(const int8_t values[32]) : simd8(load(values)) {}
+    simdjson_inline simd8(const int8_t values[32]) : simd8(load(values)) {}
     // Member-by-member initialization
-    simdjson_really_inline simd8(
+    simdjson_inline simd8(
       int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
       int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15,
       int8_t v16, int8_t v17, int8_t v18, int8_t v19, int8_t v20, int8_t v21, int8_t v22, int8_t v23,
@@ -16701,7 +16382,7 @@ namespace simd {
       v24,v25,v26,v27,v28,v29,v30,v31
     )) {}
     // Repeat 16 values as many times as necessary (usually for lookup tables)
-    simdjson_really_inline static simd8<int8_t> repeat_16(
+    simdjson_inline static simd8<int8_t> repeat_16(
       int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
       int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
     ) {
@@ -16714,23 +16395,23 @@ namespace simd {
     }
 
     // Order-sensitive comparisons
-    simdjson_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return _mm256_max_epi8(*this, other); }
-    simdjson_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return _mm256_min_epi8(*this, other); }
-    simdjson_really_inline simd8<bool> operator>(const simd8<int8_t> other) const { return _mm256_cmpgt_epi8(*this, other); }
-    simdjson_really_inline simd8<bool> operator<(const simd8<int8_t> other) const { return _mm256_cmpgt_epi8(other, *this); }
+    simdjson_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return _mm256_max_epi8(*this, other); }
+    simdjson_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return _mm256_min_epi8(*this, other); }
+    simdjson_inline simd8<bool> operator>(const simd8<int8_t> other) const { return _mm256_cmpgt_epi8(*this, other); }
+    simdjson_inline simd8<bool> operator<(const simd8<int8_t> other) const { return _mm256_cmpgt_epi8(other, *this); }
   };
 
   // Unsigned bytes
   template<>
   struct simd8<uint8_t>: base8_numeric<uint8_t> {
-    simdjson_really_inline simd8() : base8_numeric<uint8_t>() {}
-    simdjson_really_inline simd8(const __m256i _value) : base8_numeric<uint8_t>(_value) {}
+    simdjson_inline simd8() : base8_numeric<uint8_t>() {}
+    simdjson_inline simd8(const __m256i _value) : base8_numeric<uint8_t>(_value) {}
     // Splat constructor
-    simdjson_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
     // Array constructor
-    simdjson_really_inline simd8(const uint8_t values[32]) : simd8(load(values)) {}
+    simdjson_inline simd8(const uint8_t values[32]) : simd8(load(values)) {}
     // Member-by-member initialization
-    simdjson_really_inline simd8(
+    simdjson_inline simd8(
       uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
       uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15,
       uint8_t v16, uint8_t v17, uint8_t v18, uint8_t v19, uint8_t v20, uint8_t v21, uint8_t v22, uint8_t v23,
@@ -16742,7 +16423,7 @@ namespace simd {
       v24,v25,v26,v27,v28,v29,v30,v31
     )) {}
     // Repeat 16 values as many times as necessary (usually for lookup tables)
-    simdjson_really_inline static simd8<uint8_t> repeat_16(
+    simdjson_inline static simd8<uint8_t> repeat_16(
       uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
       uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
     ) {
@@ -16755,39 +16436,39 @@ namespace simd {
     }
 
     // Saturated math
-    simdjson_really_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return _mm256_adds_epu8(*this, other); }
-    simdjson_really_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return _mm256_subs_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return _mm256_adds_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return _mm256_subs_epu8(*this, other); }
 
     // Order-specific operations
-    simdjson_really_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return _mm256_max_epu8(*this, other); }
-    simdjson_really_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return _mm256_min_epu8(other, *this); }
+    simdjson_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return _mm256_max_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return _mm256_min_epu8(other, *this); }
     // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
-    simdjson_really_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
     // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
-    simdjson_really_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
-    simdjson_really_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
-    simdjson_really_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
-    simdjson_really_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
-    simdjson_really_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->lt_bits(other).any_bits_set(); }
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
+    simdjson_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->lt_bits(other).any_bits_set(); }
 
     // Bit-specific operations
-    simdjson_really_inline simd8<bool> bits_not_set() const { return *this == uint8_t(0); }
-    simdjson_really_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
-    simdjson_really_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
-    simdjson_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
-    simdjson_really_inline bool is_ascii() const { return _mm256_movemask_epi8(*this) == 0; }
-    simdjson_really_inline bool bits_not_set_anywhere() const { return _mm256_testz_si256(*this, *this); }
-    simdjson_really_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
-    simdjson_really_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const { return _mm256_testz_si256(*this, bits); }
-    simdjson_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
+    simdjson_inline simd8<bool> bits_not_set() const { return *this == uint8_t(0); }
+    simdjson_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
+    simdjson_inline bool is_ascii() const { return _mm256_movemask_epi8(*this) == 0; }
+    simdjson_inline bool bits_not_set_anywhere() const { return _mm256_testz_si256(*this, *this); }
+    simdjson_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
+    simdjson_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const { return _mm256_testz_si256(*this, bits); }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
     template<int N>
-    simdjson_really_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(_mm256_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); }
+    simdjson_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(_mm256_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); }
     template<int N>
-    simdjson_really_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(_mm256_slli_epi16(*this, N)) & uint8_t(0xFFu << N); }
+    simdjson_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(_mm256_slli_epi16(*this, N)) & uint8_t(0xFFu << N); }
     // Get one of the bits and make a bitmask out of it.
     // e.g. value.get_bit<7>() gets the high bit
     template<int N>
-    simdjson_really_inline int get_bit() const { return _mm256_movemask_epi8(_mm256_slli_epi16(*this, 7-N)); }
+    simdjson_inline int get_bit() const { return _mm256_movemask_epi8(_mm256_slli_epi16(*this, 7-N)); }
   };
 
   template<typename T>
@@ -16800,10 +16481,10 @@ namespace simd {
     simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
     simd8x64() = delete; // no default constructor allowed
 
-    simdjson_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1) : chunks{chunk0, chunk1} {}
-    simdjson_really_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+32)} {}
+    simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1) : chunks{chunk0, chunk1} {}
+    simdjson_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+32)} {}
 
-    simdjson_really_inline uint64_t compress(uint64_t mask, T * output) const {
+    simdjson_inline uint64_t compress(uint64_t mask, T * output) const {
       uint32_t mask1 = uint32_t(mask);
       uint32_t mask2 = uint32_t(mask >> 32);
       this->chunks[0].compress(mask1, output);
@@ -16811,22 +16492,22 @@ namespace simd {
       return 64 - count_ones(mask);
     }
 
-    simdjson_really_inline void store(T ptr[64]) const {
+    simdjson_inline void store(T ptr[64]) const {
       this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
       this->chunks[1].store(ptr+sizeof(simd8<T>)*1);
     }
 
-    simdjson_really_inline uint64_t to_bitmask() const {
+    simdjson_inline uint64_t to_bitmask() const {
       uint64_t r_lo = uint32_t(this->chunks[0].to_bitmask());
       uint64_t r_hi =                       this->chunks[1].to_bitmask();
       return r_lo | (r_hi << 32);
     }
 
-    simdjson_really_inline simd8<T> reduce_or() const {
+    simdjson_inline simd8<T> reduce_or() const {
       return this->chunks[0] | this->chunks[1];
     }
 
-    simdjson_really_inline simd8x64<T> bit_or(const T m) const {
+    simdjson_inline simd8x64<T> bit_or(const T m) const {
       const simd8<T> mask = simd8<T>::splat(m);
       return simd8x64<T>(
         this->chunks[0] | mask,
@@ -16834,7 +16515,7 @@ namespace simd {
       );
     }
 
-    simdjson_really_inline uint64_t eq(const T m) const {
+    simdjson_inline uint64_t eq(const T m) const {
       const simd8<T> mask = simd8<T>::splat(m);
       return  simd8x64<bool>(
         this->chunks[0] == mask,
@@ -16842,14 +16523,14 @@ namespace simd {
       ).to_bitmask();
     }
 
-    simdjson_really_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+    simdjson_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
       return  simd8x64<bool>(
         this->chunks[0] == other.chunks[0],
         this->chunks[1] == other.chunks[1]
       ).to_bitmask();
     }
 
-    simdjson_really_inline uint64_t lteq(const T m) const {
+    simdjson_inline uint64_t lteq(const T m) const {
       const simd8<T> mask = simd8<T>::splat(m);
       return  simd8x64<bool>(
         this->chunks[0] <= mask,
@@ -16875,11 +16556,11 @@ namespace jsoncharutils {
 
 // return non-zero if not a structural or whitespace char
 // zero otherwise
-simdjson_really_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
+simdjson_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
   return internal::structural_or_whitespace_negated[c];
 }
 
-simdjson_really_inline uint32_t is_structural_or_whitespace(uint8_t c) {
+simdjson_inline uint32_t is_structural_or_whitespace(uint8_t c) {
   return internal::structural_or_whitespace[c];
 }
 
@@ -16910,7 +16591,7 @@ static inline uint32_t hex_to_u32_nocheck(
 //
 // Note: we assume that surrogates are treated separately
 //
-simdjson_really_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
+simdjson_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
   if (cp <= 0x7F) {
     c[0] = uint8_t(cp);
     return 1; // ascii
@@ -16942,10 +16623,10 @@ simdjson_really_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
 #ifdef SIMDJSON_IS_32BITS // _umul128 for x86, arm
 // this is a slow emulation routine for 32-bit
 //
-static simdjson_really_inline uint64_t __emulu(uint32_t x, uint32_t y) {
+static simdjson_inline uint64_t __emulu(uint32_t x, uint32_t y) {
   return x * (uint64_t)y;
 }
-static simdjson_really_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+static simdjson_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
   uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
   uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
   uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
@@ -16959,7 +16640,7 @@ static simdjson_really_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64
 
 using internal::value128;
 
-simdjson_really_inline value128 full_multiplication(uint64_t value1, uint64_t value2) {
+simdjson_inline value128 full_multiplication(uint64_t value1, uint64_t value2) {
   value128 answer;
 #if defined(SIMDJSON_REGULAR_VISUAL_STUDIO) || defined(SIMDJSON_IS_32BITS)
 #ifdef _M_ARM64
@@ -16995,13 +16676,13 @@ namespace atomparsing {
 // You might think that using memcpy makes this function expensive, but you'd be wrong.
 // All decent optimizing compilers (GCC, clang, Visual Studio) will compile string_to_uint32("false");
 // to the compile-time constant 1936482662.
-simdjson_really_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
+simdjson_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
 
 
 // Again in str4ncmp we use a memcpy to avoid undefined behavior. The memcpy may appear expensive.
 // Yet all decent optimizing compilers will compile memcpy to a single instruction, just about.
 simdjson_warn_unused
-simdjson_really_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
+simdjson_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
   uint32_t srcval; // we want to avoid unaligned 32-bit loads (undefined in C/C++)
   static_assert(sizeof(uint32_t) <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be larger than 4 bytes");
   std::memcpy(&srcval, src, sizeof(uint32_t));
@@ -17009,36 +16690,36 @@ simdjson_really_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_true_atom(const uint8_t *src) {
+simdjson_inline bool is_valid_true_atom(const uint8_t *src) {
   return (str4ncmp(src, "true") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
+simdjson_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
   if (len > 4) { return is_valid_true_atom(src); }
   else if (len == 4) { return !str4ncmp(src, "true"); }
   else { return false; }
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_false_atom(const uint8_t *src) {
+simdjson_inline bool is_valid_false_atom(const uint8_t *src) {
   return (str4ncmp(src+1, "alse") | jsoncharutils::is_not_structural_or_whitespace(src[5])) == 0;
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
+simdjson_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
   if (len > 5) { return is_valid_false_atom(src); }
   else if (len == 5) { return !str4ncmp(src+1, "alse"); }
   else { return false; }
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_null_atom(const uint8_t *src) {
+simdjson_inline bool is_valid_null_atom(const uint8_t *src) {
   return (str4ncmp(src, "null") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
+simdjson_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
   if (len > 4) { return is_valid_null_atom(src); }
   else if (len == 4) { return !str4ncmp(src, "null"); }
   else { return false; }
@@ -17064,18 +16745,18 @@ using namespace simd;
 struct backslash_and_quote {
 public:
   static constexpr uint32_t BYTES_PROCESSED = 32;
-  simdjson_really_inline static backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+  simdjson_inline static backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
 
-  simdjson_really_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
-  simdjson_really_inline bool has_backslash() { return ((quote_bits - 1) & bs_bits) != 0; }
-  simdjson_really_inline int quote_index() { return trailing_zeroes(quote_bits); }
-  simdjson_really_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+  simdjson_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_inline bool has_backslash() { return ((quote_bits - 1) & bs_bits) != 0; }
+  simdjson_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_inline int backslash_index() { return trailing_zeroes(bs_bits); }
 
   uint32_t bs_bits;
   uint32_t quote_bits;
 }; // struct backslash_and_quote
 
-simdjson_really_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
   // this can read up to 15 bytes beyond the buffer size, but we require
   // SIMDJSON_PADDING of padding
   static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
@@ -17092,147 +16773,6 @@ simdjson_really_inline backslash_and_quote backslash_and_quote::copy_and_find(co
 } // namespace haswell
 } // namespace simdjson
 
-/* begin file include/simdjson/generic/stringparsing.h */
-// This file contains the common code every implementation uses
-// It is intended to be included multiple times and compiled multiple times
-
-namespace simdjson {
-namespace haswell {
-namespace {
-/// @private
-namespace stringparsing {
-
-// begin copypasta
-// These chars yield themselves: " \ /
-// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
-// u not handled in this table as it's complex
-static const uint8_t escape_map[256] = {
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
-    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
-    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
-
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-};
-
-// handle a unicode codepoint
-// write appropriate values into dest
-// src will advance 6 bytes or 12 bytes
-// dest will advance a variable amount (return via pointer)
-// return true if the unicode codepoint was valid
-// We work in little-endian then swap at write time
-simdjson_warn_unused
-simdjson_really_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
-                                            uint8_t **dst_ptr) {
-  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
-  // conversion isn't valid; we defer the check for this to inside the
-  // multilingual plane check
-  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
-  *src_ptr += 6;
-  // check for low surrogate for characters outside the Basic
-  // Multilingual Plane.
-  if (code_point >= 0xd800 && code_point < 0xdc00) {
-    if (((*src_ptr)[0] != '\\') || (*src_ptr)[1] != 'u') {
-      return false;
-    }
-    uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
-
-    // if the first code point is invalid we will get here, as we will go past
-    // the check for being outside the Basic Multilingual plane. If we don't
-    // find a \u immediately afterwards we fail out anyhow, but if we do,
-    // this check catches both the case of the first code point being invalid
-    // or the second code point being invalid.
-    if ((code_point | code_point_2) >> 16) {
-      return false;
-    }
-
-    code_point =
-        (((code_point - 0xd800) << 10) | (code_point_2 - 0xdc00)) + 0x10000;
-    *src_ptr += 6;
-  }
-  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
-  *dst_ptr += offset;
-  return offset > 0;
-}
-
-/**
- * Unescape a string from src to dst, stopping at a final unescaped quote. E.g., if src points at 'joe"', then
- * dst needs to have four free bytes.
- */
-simdjson_warn_unused simdjson_really_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst) {
-  while (1) {
-    // Copy the next n bytes, and find the backslash and quote in them.
-    auto bs_quote = backslash_and_quote::copy_and_find(src, dst);
-    // If the next thing is the end quote, copy and return
-    if (bs_quote.has_quote_first()) {
-      // we encountered quotes first. Move dst to point to quotes and exit
-      return dst + bs_quote.quote_index();
-    }
-    if (bs_quote.has_backslash()) {
-      /* find out where the backspace is */
-      auto bs_dist = bs_quote.backslash_index();
-      uint8_t escape_char = src[bs_dist + 1];
-      /* we encountered backslash first. Handle backslash */
-      if (escape_char == 'u') {
-        /* move src/dst up to the start; they will be further adjusted
-           within the unicode codepoint handling code. */
-        src += bs_dist;
-        dst += bs_dist;
-        if (!handle_unicode_codepoint(&src, &dst)) {
-          return nullptr;
-        }
-      } else {
-        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
-         * write bs_dist+1 characters to output
-         * note this may reach beyond the part of the buffer we've actually
-         * seen. I think this is ok */
-        uint8_t escape_result = escape_map[escape_char];
-        if (escape_result == 0u) {
-          return nullptr; /* bogus escape value is an error */
-        }
-        dst[bs_dist] = escape_result;
-        src += bs_dist + 2;
-        dst += bs_dist + 1;
-      }
-    } else {
-      /* they are the same. Since they can't co-occur, it means we
-       * encountered neither. */
-      src += backslash_and_quote::BYTES_PROCESSED;
-      dst += backslash_and_quote::BYTES_PROCESSED;
-    }
-  }
-  /* can't be reached */
-  return nullptr;
-}
-
-simdjson_unused simdjson_warn_unused simdjson_really_inline error_code parse_string_to_buffer(const uint8_t *src, uint8_t *&current_string_buf_loc, std::string_view &s) {
-  if (*(src++) != '"') { return STRING_ERROR; }
-  auto end = stringparsing::parse_string(src, current_string_buf_loc);
-  if (!end) { return STRING_ERROR; }
-  s = std::string_view(reinterpret_cast<const char *>(current_string_buf_loc), end-current_string_buf_loc);
-  current_string_buf_loc = end;
-  return SUCCESS;
-}
-
-} // namespace stringparsing
-} // unnamed namespace
-} // namespace haswell
-} // namespace simdjson
-/* end file include/simdjson/generic/stringparsing.h */
-
 #endif // SIMDJSON_HASWELL_STRINGPARSING_H
 /* end file include/simdjson/haswell/stringparsing.h */
 /* begin file include/simdjson/haswell/numberparsing.h */
@@ -17243,7 +16783,7 @@ namespace simdjson {
 namespace haswell {
 namespace {
 
-static simdjson_really_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
   // this actually computes *16* values so we are being wasteful.
   const __m128i ascii0 = _mm_set1_epi8('0');
   const __m128i mul_1_10 =
@@ -17306,7 +16846,7 @@ namespace {
 // Convert a mantissa, an exponent and a sign bit into an ieee64 double.
 // The real_exponent needs to be in [0, 2046] (technically real_exponent = 2047 would be acceptable).
 // The mantissa should be in [0,1<<53). The bit at index (1ULL << 52) while be zeroed.
-simdjson_really_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
+simdjson_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
     double d;
     mantissa &= ~(1ULL << 52);
     mantissa |= real_exponent << 52;
@@ -17321,7 +16861,7 @@ simdjson_really_inline double to_double(uint64_t mantissa, uint64_t real_exponen
 // set to false. This should work *most of the time* (like 99% of the time).
 // We assume that power is in the [smallest_power,
 // largest_power] interval: the caller is responsible for this check.
-simdjson_really_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
+simdjson_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
   // we start with a fast path
   // It was described in
   // Clinger WD. How to read floating point numbers accurately.
@@ -17381,7 +16921,7 @@ simdjson_really_inline bool compute_float_64(int64_t power, uint64_t i, bool neg
   // In the slow path, we need to adjust i so that it is > 1<<63 which is always
   // possible, except if i == 0, so we handle i == 0 separately.
   if(i == 0) {
-    d = 0.0;
+    d = negative ? -0.0 : 0.0;
     return true;
   }
 
@@ -17496,7 +17036,7 @@ simdjson_really_inline bool compute_float_64(int64_t power, uint64_t i, bool neg
   if (simdjson_unlikely(real_exponent <= 0)) { // we have a subnormal?
     // Here have that real_exponent <= 0 so -real_exponent >= 0
     if(-real_exponent + 1 >= 64) { // if we have more than 64 bits below the minimum exponent, you have a zero for sure.
-      d = 0.0;
+      d = negative ? -0.0 : 0.0;
       return true;
     }
     // next line is safe because -real_exponent + 1 < 0
@@ -17604,7 +17144,7 @@ static bool parse_float_fallback(const uint8_t *ptr, const uint8_t *end_ptr, dou
 // check quickly whether the next 8 chars are made of digits
 // at a glance, it looks better than Mula's
 // http://0x80.pl/articles/swar-digits-validate.html
-simdjson_really_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
+simdjson_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
   uint64_t val;
   // this can read up to 7 bytes beyond the buffer size, but we require
   // SIMDJSON_PADDING of padding
@@ -17631,7 +17171,7 @@ error_code slow_float_parsing(simdjson_unused const uint8_t * src, W writer) {
 
 template<typename I>
 SIMDJSON_NO_SANITIZE_UNDEFINED // We deliberately allow overflow here and check later
-simdjson_really_inline bool parse_digit(const uint8_t c, I &i) {
+simdjson_inline bool parse_digit(const uint8_t c, I &i) {
   const uint8_t digit = static_cast<uint8_t>(c - '0');
   if (digit > 9) {
     return false;
@@ -17641,7 +17181,7 @@ simdjson_really_inline bool parse_digit(const uint8_t c, I &i) {
   return true;
 }
 
-simdjson_really_inline error_code parse_decimal(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
+simdjson_inline error_code parse_decimal(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
   // we continue with the fiction that we have an integer. If the
   // floating point number is representable as x * 10^z for some integer
   // z that fits in 53 bits, then we will be able to convert back the
@@ -17669,7 +17209,7 @@ simdjson_really_inline error_code parse_decimal(simdjson_unused const uint8_t *c
   return SUCCESS;
 }
 
-simdjson_really_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
+simdjson_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
   // Exp Sign: -123.456e[-]78
   bool neg_exp = ('-' == *p);
   if (neg_exp || '+' == *p) { p++; } // Skip + as well
@@ -17720,7 +17260,7 @@ simdjson_really_inline error_code parse_exponent(simdjson_unused const uint8_t *
   return SUCCESS;
 }
 
-simdjson_really_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
+simdjson_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
   // It is possible that the integer had an overflow.
   // We have to handle the case where we have 0.0000somenumber.
   const uint8_t *start = start_digits;
@@ -17730,7 +17270,7 @@ simdjson_really_inline size_t significant_digits(const uint8_t * start_digits, s
 }
 
 template<typename W>
-simdjson_really_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
+simdjson_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
   // If we frequently had to deal with long strings of digits,
   // we could extend our code by using a 128-bit integer instead
   // of a 64-bit integer. However, this is uncommon in practice.
@@ -17766,7 +17306,8 @@ simdjson_really_inline error_code write_float(const uint8_t *const src, bool neg
     static_assert(simdjson::internal::smallest_power <= -342, "smallest_power is not small enough");
     //
     if((exponent < simdjson::internal::smallest_power) || (i == 0)) {
-      WRITE_DOUBLE(0, src, writer);
+      // E.g. Parse "-0.0e-999" into the same value as "-0.0". See https://en.wikipedia.org/wiki/Signed_zero
+      WRITE_DOUBLE(negative ? -0.0 : 0.0, src, writer);
       return SUCCESS;
     } else { // (exponent > largest_power) and (i != 0)
       // We have, for sure, an infinite value and simdjson refuses to parse infinite values.
@@ -17786,20 +17327,20 @@ simdjson_really_inline error_code write_float(const uint8_t *const src, bool neg
 #ifdef SIMDJSON_SKIPNUMBERPARSING
 
 template<typename W>
-simdjson_really_inline error_code parse_number(const uint8_t *const, W &writer) {
+simdjson_inline error_code parse_number(const uint8_t *const, W &writer) {
   writer.append_s64(0);        // always write zero
   return SUCCESS;              // always succeeds
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
-simdjson_unused simdjson_really_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
-simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept { return ondemand::number_type::signed_integer; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept { return ondemand::number_type::signed_integer; }
 #else
 
 // parse the number at src
@@ -17812,13 +17353,13 @@ simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> ge
 //
 // Our objective is accurate parsing (ULP of 0) at high speed.
 template<typename W>
-simdjson_really_inline error_code parse_number(const uint8_t *const src, W &writer) {
+simdjson_inline error_code parse_number(const uint8_t *const src, W &writer) {
 
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  const uint8_t *p = src + negative;
+  const uint8_t *p = src + uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -17962,7 +17503,7 @@ const uint8_t integer_string_finisher[256] = {
     NUMBER_ERROR};
 
 // Parse any number from 0 to 18,446,744,073,709,551,615
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
   const uint8_t *p = src;
   //
   // Parse the integer part.
@@ -18012,7 +17553,7 @@ simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(
 
 // Parse any number from 0 to 18,446,744,073,709,551,615
 // Never read at src_end or beyond
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
   const uint8_t *p = src;
   //
   // Parse the integer part.
@@ -18060,7 +17601,7 @@ simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(
 }
 
 // Parse any number from 0 to 18,446,744,073,709,551,615
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
   const uint8_t *p = src + 1;
   //
   // Parse the integer part.
@@ -18110,12 +17651,12 @@ simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned_
 }
 
 // Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  const uint8_t *p = src + negative;
+  const uint8_t *p = src + uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -18153,13 +17694,13 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(co
 
 // Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
 // Never read at src_end or beyond
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
   //
   // Check for minus sign
   //
   if(src == src_end) { return NUMBER_ERROR; }
   bool negative = (*src == '-');
-  const uint8_t *p = src + negative;
+  const uint8_t *p = src + uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -18196,24 +17737,24 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(co
 }
 
 // Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*(src + 1) == '-');
-  const uint8_t *p = src + negative + 1;
+  src += uint8_t(negative) + 1;
 
   //
   // Parse the integer part.
   //
   // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
-  const uint8_t *const start_digits = p;
+  const uint8_t *const start_digits = src;
   uint64_t i = 0;
-  while (parse_digit(*p, i)) { p++; }
+  while (parse_digit(*src, i)) { src++; }
 
   // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
   // Optimization note: size_t is expected to be unsigned.
-  size_t digit_count = size_t(p - start_digits);
+  size_t digit_count = size_t(src - start_digits);
   // We go from
   // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
   // so we can never represent numbers that have more than 19 digits.
@@ -18225,11 +17766,11 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in
   // Here digit_count > 0.
   if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
   // We can do the following...
-  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
-  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // if (!jsoncharutils::is_structural_or_whitespace(*src)) {
+  //  return (*src == '.' || *src == 'e' || *src == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
   // }
   // as a single table lookup:
-  if(*p != '"') { return NUMBER_ERROR; }
+  if(*src != '"') { return NUMBER_ERROR; }
   // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
   // Performance note: This check is only needed when digit_count == longest_digit_count but it is
   // so cheap that we might as well always make it.
@@ -18237,12 +17778,12 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in
   return negative ? (~i+1) : i;
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -18309,19 +17850,19 @@ simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(cons
   if (simdjson_likely(!overflow)) {
     if (compute_float_64(exponent, i, negative, d)) { return d; }
   }
-  if (!parse_float_fallback(src-negative, &d)) {
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
     return NUMBER_ERROR;
   }
   return d;
 }
 
-simdjson_unused simdjson_really_inline bool is_negative(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept {
   return (*src == '-');
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
   const uint8_t *p = src;
   while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
   if ( p == src ) { return NUMBER_ERROR; }
@@ -18329,9 +17870,9 @@ simdjson_unused simdjson_really_inline simdjson_result<bool> is_integer(const ui
   return false;
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept {
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
   const uint8_t *p = src;
   while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
   if ( p == src ) { return NUMBER_ERROR; }
@@ -18355,13 +17896,13 @@ simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> ge
 }
 
 // Never read at src_end or beyond
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
   if(src == src_end) { return NUMBER_ERROR; }
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -18430,18 +17971,18 @@ simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(cons
   if (simdjson_likely(!overflow)) {
     if (compute_float_64(exponent, i, negative, d)) { return d; }
   }
-  if (!parse_float_fallback(src-negative, src_end, &d)) {
+  if (!parse_float_fallback(src - uint8_t(negative), src_end, &d)) {
     return NUMBER_ERROR;
   }
   return d;
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*(src + 1) == '-');
-  src += negative + 1;
+  src += uint8_t(negative) + 1;
 
   //
   // Parse the integer part.
@@ -18508,7 +18049,7 @@ simdjson_unused simdjson_really_inline simdjson_result<double> parse_double_in_s
   if (simdjson_likely(!overflow)) {
     if (compute_float_64(exponent, i, negative, d)) { return d; }
   }
-  if (!parse_float_fallback(src-negative, &d)) {
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
     return NUMBER_ERROR;
   }
   return d;
@@ -18561,7 +18102,7 @@ using namespace simdjson::dom;
 
 class implementation final : public simdjson::implementation {
 public:
-  simdjson_really_inline implementation()
+  simdjson_inline implementation()
       : simdjson::implementation("ppc64", "PPC64 ALTIVEC",
                                  internal::instruction_set::ALTIVEC) {}
   simdjson_warn_unused error_code create_dom_parser_implementation(
@@ -18623,10 +18164,11 @@ class dom_parser_implementation final : public internal::dom_parser_implementati
   simdjson_warn_unused error_code stage1(const uint8_t *buf, size_t len, stage1_mode partial) noexcept final;
   simdjson_warn_unused error_code stage2(dom::document &doc) noexcept final;
   simdjson_warn_unused error_code stage2_next(dom::document &doc) noexcept final;
+  simdjson_warn_unused uint8_t *parse_string(const uint8_t *src, uint8_t *dst) const noexcept final;
   inline simdjson_warn_unused error_code set_capacity(size_t capacity) noexcept final;
   inline simdjson_warn_unused error_code set_max_depth(size_t max_depth) noexcept final;
 private:
-  simdjson_really_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
+  simdjson_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
 
 };
 
@@ -18685,6 +18227,8 @@ inline simdjson_warn_unused error_code dom_parser_implementation::set_max_depth(
 #undef vector
 #endif
 
+static_assert(sizeof(__vector unsigned char) <= simdjson::SIMDJSON_PADDING, "insufficient padding for ppc64");
+
 #endif //  SIMDJSON_PPC64_INTRINSICS_H
 /* end file include/simdjson/ppc64/intrinsics.h */
 /* begin file include/simdjson/ppc64/bitmanipulation.h */
@@ -18699,7 +18243,7 @@ namespace {
 // but the algorithms do not end up using the returned value.
 // Sadly, sanitizers are not smart enough to figure it out.
 SIMDJSON_NO_SANITIZE_UNDEFINED
-simdjson_really_inline int trailing_zeroes(uint64_t input_num) {
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
   unsigned long ret;
   // Search the mask data from least significant bit (LSB)
@@ -18712,12 +18256,12 @@ simdjson_really_inline int trailing_zeroes(uint64_t input_num) {
 }
 
 /* result might be undefined when input_num is zero */
-simdjson_really_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
   return input_num & (input_num - 1);
 }
 
 /* result might be undefined when input_num is zero */
-simdjson_really_inline int leading_zeroes(uint64_t input_num) {
+simdjson_inline int leading_zeroes(uint64_t input_num) {
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
   unsigned long leading_zero = 0;
   // Search the mask data from most significant bit (MSB)
@@ -18732,17 +18276,17 @@ simdjson_really_inline int leading_zeroes(uint64_t input_num) {
 }
 
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
-simdjson_really_inline int count_ones(uint64_t input_num) {
+simdjson_inline int count_ones(uint64_t input_num) {
   // note: we do not support legacy 32-bit Windows
   return __popcnt64(input_num); // Visual Studio wants two underscores
 }
 #else
-simdjson_really_inline int count_ones(uint64_t input_num) {
+simdjson_inline int count_ones(uint64_t input_num) {
   return __builtin_popcountll(input_num);
 }
 #endif
 
-simdjson_really_inline bool add_overflow(uint64_t value1, uint64_t value2,
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2,
                                          uint64_t *result) {
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
   *result = value1 + value2;
@@ -18773,7 +18317,7 @@ namespace {
 //
 // For example, prefix_xor(00100100) == 00011100
 //
-simdjson_really_inline uint64_t prefix_xor(uint64_t bitmask) {
+simdjson_inline uint64_t prefix_xor(uint64_t bitmask) {
   // You can use the version below, however gcc sometimes miscompiles
   // vec_pmsum_be, it happens somewhere around between 8 and 9th version.
   // The performance boost was not noticeable, falling back to a usual
@@ -18820,41 +18364,41 @@ template <typename Child> struct base {
   __m128i value;
 
   // Zero constructor
-  simdjson_really_inline base() : value{__m128i()} {}
+  simdjson_inline base() : value{__m128i()} {}
 
   // Conversion from SIMD register
-  simdjson_really_inline base(const __m128i _value) : value(_value) {}
+  simdjson_inline base(const __m128i _value) : value(_value) {}
 
   // Conversion to SIMD register
-  simdjson_really_inline operator const __m128i &() const {
+  simdjson_inline operator const __m128i &() const {
     return this->value;
   }
-  simdjson_really_inline operator __m128i &() { return this->value; }
+  simdjson_inline operator __m128i &() { return this->value; }
 
   // Bit operations
-  simdjson_really_inline Child operator|(const Child other) const {
+  simdjson_inline Child operator|(const Child other) const {
     return vec_or(this->value, (__m128i)other);
   }
-  simdjson_really_inline Child operator&(const Child other) const {
+  simdjson_inline Child operator&(const Child other) const {
     return vec_and(this->value, (__m128i)other);
   }
-  simdjson_really_inline Child operator^(const Child other) const {
+  simdjson_inline Child operator^(const Child other) const {
     return vec_xor(this->value, (__m128i)other);
   }
-  simdjson_really_inline Child bit_andnot(const Child other) const {
+  simdjson_inline Child bit_andnot(const Child other) const {
     return vec_andc(this->value, (__m128i)other);
   }
-  simdjson_really_inline Child &operator|=(const Child other) {
+  simdjson_inline Child &operator|=(const Child other) {
     auto this_cast = static_cast<Child*>(this);
     *this_cast = *this_cast | other;
     return *this_cast;
   }
-  simdjson_really_inline Child &operator&=(const Child other) {
+  simdjson_inline Child &operator&=(const Child other) {
     auto this_cast = static_cast<Child*>(this);
     *this_cast = *this_cast & other;
     return *this_cast;
   }
-  simdjson_really_inline Child &operator^=(const Child other) {
+  simdjson_inline Child &operator^=(const Child other) {
     auto this_cast = static_cast<Child*>(this);
     *this_cast = *this_cast ^ other;
     return *this_cast;
@@ -18869,17 +18413,17 @@ struct base8 : base<simd8<T>> {
   typedef uint16_t bitmask_t;
   typedef uint32_t bitmask2_t;
 
-  simdjson_really_inline base8() : base<simd8<T>>() {}
-  simdjson_really_inline base8(const __m128i _value) : base<simd8<T>>(_value) {}
+  simdjson_inline base8() : base<simd8<T>>() {}
+  simdjson_inline base8(const __m128i _value) : base<simd8<T>>(_value) {}
 
-  friend simdjson_really_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) {
+  friend simdjson_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) {
     return (__m128i)vec_cmpeq(lhs.value, (__m128i)rhs);
   }
 
   static const int SIZE = sizeof(base<simd8<T>>::value);
 
   template <int N = 1>
-  simdjson_really_inline simd8<T> prev(simd8<T> prev_chunk) const {
+  simdjson_inline simd8<T> prev(simd8<T> prev_chunk) const {
     __m128i chunk = this->value;
 #ifdef __LITTLE_ENDIAN__
     chunk = (__m128i)vec_reve(this->value);
@@ -18895,18 +18439,18 @@ struct base8 : base<simd8<T>> {
 
 // SIMD byte mask type (returned by things like eq and gt)
 template <> struct simd8<bool> : base8<bool> {
-  static simdjson_really_inline simd8<bool> splat(bool _value) {
+  static simdjson_inline simd8<bool> splat(bool _value) {
     return (__m128i)vec_splats((unsigned char)(-(!!_value)));
   }
 
-  simdjson_really_inline simd8<bool>() : base8() {}
-  simdjson_really_inline simd8<bool>(const __m128i _value)
+  simdjson_inline simd8<bool>() : base8() {}
+  simdjson_inline simd8<bool>(const __m128i _value)
       : base8<bool>(_value) {}
   // Splat constructor
-  simdjson_really_inline simd8<bool>(bool _value)
+  simdjson_inline simd8<bool>(bool _value)
       : base8<bool>(splat(_value)) {}
 
-  simdjson_really_inline int to_bitmask() const {
+  simdjson_inline int to_bitmask() const {
     __vector unsigned long long result;
     const __m128i perm_mask = {0x78, 0x70, 0x68, 0x60, 0x58, 0x50, 0x48, 0x40,
                                0x38, 0x30, 0x28, 0x20, 0x18, 0x10, 0x08, 0x00};
@@ -18919,25 +18463,25 @@ template <> struct simd8<bool> : base8<bool> {
     return static_cast<int>(result[0]);
 #endif
   }
-  simdjson_really_inline bool any() const {
+  simdjson_inline bool any() const {
     return !vec_all_eq(this->value, (__m128i)vec_splats(0));
   }
-  simdjson_really_inline simd8<bool> operator~() const {
+  simdjson_inline simd8<bool> operator~() const {
     return this->value ^ (__m128i)splat(true);
   }
 };
 
 template <typename T> struct base8_numeric : base8<T> {
-  static simdjson_really_inline simd8<T> splat(T value) {
+  static simdjson_inline simd8<T> splat(T value) {
     (void)value;
     return (__m128i)vec_splats(value);
   }
-  static simdjson_really_inline simd8<T> zero() { return splat(0); }
-  static simdjson_really_inline simd8<T> load(const T values[16]) {
+  static simdjson_inline simd8<T> zero() { return splat(0); }
+  static simdjson_inline simd8<T> load(const T values[16]) {
     return (__m128i)(vec_vsx_ld(0, reinterpret_cast<const uint8_t *>(values)));
   }
   // Repeat 16 values as many times as necessary (usually for lookup tables)
-  static simdjson_really_inline simd8<T> repeat_16(T v0, T v1, T v2, T v3, T v4,
+  static simdjson_inline simd8<T> repeat_16(T v0, T v1, T v2, T v3, T v4,
                                                    T v5, T v6, T v7, T v8, T v9,
                                                    T v10, T v11, T v12, T v13,
                                                    T v14, T v15) {
@@ -18945,30 +18489,30 @@ template <typename T> struct base8_numeric : base8<T> {
                     v14, v15);
   }
 
-  simdjson_really_inline base8_numeric() : base8<T>() {}
-  simdjson_really_inline base8_numeric(const __m128i _value)
+  simdjson_inline base8_numeric() : base8<T>() {}
+  simdjson_inline base8_numeric(const __m128i _value)
       : base8<T>(_value) {}
 
   // Store to array
-  simdjson_really_inline void store(T dst[16]) const {
+  simdjson_inline void store(T dst[16]) const {
     vec_vsx_st(this->value, 0, reinterpret_cast<__m128i *>(dst));
   }
 
   // Override to distinguish from bool version
-  simdjson_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+  simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
 
   // Addition/subtraction are the same for signed and unsigned
-  simdjson_really_inline simd8<T> operator+(const simd8<T> other) const {
+  simdjson_inline simd8<T> operator+(const simd8<T> other) const {
     return (__m128i)((__m128i)this->value + (__m128i)other);
   }
-  simdjson_really_inline simd8<T> operator-(const simd8<T> other) const {
+  simdjson_inline simd8<T> operator-(const simd8<T> other) const {
     return (__m128i)((__m128i)this->value - (__m128i)other);
   }
-  simdjson_really_inline simd8<T> &operator+=(const simd8<T> other) {
+  simdjson_inline simd8<T> &operator+=(const simd8<T> other) {
     *this = *this + other;
     return *static_cast<simd8<T> *>(this);
   }
-  simdjson_really_inline simd8<T> &operator-=(const simd8<T> other) {
+  simdjson_inline simd8<T> &operator-=(const simd8<T> other) {
     *this = *this - other;
     return *static_cast<simd8<T> *>(this);
   }
@@ -18976,7 +18520,7 @@ template <typename T> struct base8_numeric : base8<T> {
   // Perform a lookup assuming the value is between 0 and 16 (undefined behavior
   // for out of range values)
   template <typename L>
-  simdjson_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+  simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
     return (__m128i)vec_perm((__m128i)lookup_table, (__m128i)lookup_table, this->value);
   }
 
@@ -18987,7 +18531,7 @@ template <typename T> struct base8_numeric : base8<T> {
   // seems like a function with the signature simd8<L> compress(uint32_t mask)
   // would be sensible, but the AVX ISA makes this kind of approach difficult.
   template <typename L>
-  simdjson_really_inline void compress(uint16_t mask, L *output) const {
+  simdjson_inline void compress(uint16_t mask, L *output) const {
     using internal::BitsSetTable256mul2;
     using internal::pshufb_combine_table;
     using internal::thintable_epi8;
@@ -19026,7 +18570,7 @@ template <typename T> struct base8_numeric : base8<T> {
   }
 
   template <typename L>
-  simdjson_really_inline simd8<L>
+  simdjson_inline simd8<L>
   lookup_16(L replace0, L replace1, L replace2, L replace3, L replace4,
             L replace5, L replace6, L replace7, L replace8, L replace9,
             L replace10, L replace11, L replace12, L replace13, L replace14,
@@ -19040,15 +18584,15 @@ template <typename T> struct base8_numeric : base8<T> {
 
 // Signed bytes
 template <> struct simd8<int8_t> : base8_numeric<int8_t> {
-  simdjson_really_inline simd8() : base8_numeric<int8_t>() {}
-  simdjson_really_inline simd8(const __m128i _value)
+  simdjson_inline simd8() : base8_numeric<int8_t>() {}
+  simdjson_inline simd8(const __m128i _value)
       : base8_numeric<int8_t>(_value) {}
   // Splat constructor
-  simdjson_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+  simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
   // Array constructor
-  simdjson_really_inline simd8(const int8_t *values) : simd8(load(values)) {}
+  simdjson_inline simd8(const int8_t *values) : simd8(load(values)) {}
   // Member-by-member initialization
-  simdjson_really_inline simd8(int8_t v0, int8_t v1, int8_t v2, int8_t v3,
+  simdjson_inline simd8(int8_t v0, int8_t v1, int8_t v2, int8_t v3,
                                int8_t v4, int8_t v5, int8_t v6, int8_t v7,
                                int8_t v8, int8_t v9, int8_t v10, int8_t v11,
                                int8_t v12, int8_t v13, int8_t v14, int8_t v15)
@@ -19056,7 +18600,7 @@ template <> struct simd8<int8_t> : base8_numeric<int8_t> {
                                               v8, v9, v10, v11, v12, v13, v14,
                                               v15}) {}
   // Repeat 16 values as many times as necessary (usually for lookup tables)
-  simdjson_really_inline static simd8<int8_t>
+  simdjson_inline static simd8<int8_t>
   repeat_16(int8_t v0, int8_t v1, int8_t v2, int8_t v3, int8_t v4, int8_t v5,
             int8_t v6, int8_t v7, int8_t v8, int8_t v9, int8_t v10, int8_t v11,
             int8_t v12, int8_t v13, int8_t v14, int8_t v15) {
@@ -19065,22 +18609,22 @@ template <> struct simd8<int8_t> : base8_numeric<int8_t> {
   }
 
   // Order-sensitive comparisons
-  simdjson_really_inline simd8<int8_t>
+  simdjson_inline simd8<int8_t>
   max_val(const simd8<int8_t> other) const {
     return (__m128i)vec_max((__vector signed char)this->value,
                             (__vector signed char)(__m128i)other);
   }
-  simdjson_really_inline simd8<int8_t>
+  simdjson_inline simd8<int8_t>
   min_val(const simd8<int8_t> other) const {
     return (__m128i)vec_min((__vector signed char)this->value,
                             (__vector signed char)(__m128i)other);
   }
-  simdjson_really_inline simd8<bool>
+  simdjson_inline simd8<bool>
   operator>(const simd8<int8_t> other) const {
     return (__m128i)vec_cmpgt((__vector signed char)this->value,
                               (__vector signed char)(__m128i)other);
   }
-  simdjson_really_inline simd8<bool>
+  simdjson_inline simd8<bool>
   operator<(const simd8<int8_t> other) const {
     return (__m128i)vec_cmplt((__vector signed char)this->value,
                               (__vector signed char)(__m128i)other);
@@ -19089,22 +18633,22 @@ template <> struct simd8<int8_t> : base8_numeric<int8_t> {
 
 // Unsigned bytes
 template <> struct simd8<uint8_t> : base8_numeric<uint8_t> {
-  simdjson_really_inline simd8() : base8_numeric<uint8_t>() {}
-  simdjson_really_inline simd8(const __m128i _value)
+  simdjson_inline simd8() : base8_numeric<uint8_t>() {}
+  simdjson_inline simd8(const __m128i _value)
       : base8_numeric<uint8_t>(_value) {}
   // Splat constructor
-  simdjson_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+  simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
   // Array constructor
-  simdjson_really_inline simd8(const uint8_t *values) : simd8(load(values)) {}
+  simdjson_inline simd8(const uint8_t *values) : simd8(load(values)) {}
   // Member-by-member initialization
-  simdjson_really_inline
+  simdjson_inline
   simd8(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4, uint8_t v5,
         uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9, uint8_t v10,
         uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15)
       : simd8((__m128i){v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12,
                         v13, v14, v15}) {}
   // Repeat 16 values as many times as necessary (usually for lookup tables)
-  simdjson_really_inline static simd8<uint8_t>
+  simdjson_inline static simd8<uint8_t>
   repeat_16(uint8_t v0, uint8_t v1, uint8_t v2, uint8_t v3, uint8_t v4,
             uint8_t v5, uint8_t v6, uint8_t v7, uint8_t v8, uint8_t v9,
             uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14,
@@ -19114,82 +18658,82 @@ template <> struct simd8<uint8_t> : base8_numeric<uint8_t> {
   }
 
   // Saturated math
-  simdjson_really_inline simd8<uint8_t>
+  simdjson_inline simd8<uint8_t>
   saturating_add(const simd8<uint8_t> other) const {
     return (__m128i)vec_adds(this->value, (__m128i)other);
   }
-  simdjson_really_inline simd8<uint8_t>
+  simdjson_inline simd8<uint8_t>
   saturating_sub(const simd8<uint8_t> other) const {
     return (__m128i)vec_subs(this->value, (__m128i)other);
   }
 
   // Order-specific operations
-  simdjson_really_inline simd8<uint8_t>
+  simdjson_inline simd8<uint8_t>
   max_val(const simd8<uint8_t> other) const {
     return (__m128i)vec_max(this->value, (__m128i)other);
   }
-  simdjson_really_inline simd8<uint8_t>
+  simdjson_inline simd8<uint8_t>
   min_val(const simd8<uint8_t> other) const {
     return (__m128i)vec_min(this->value, (__m128i)other);
   }
   // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
-  simdjson_really_inline simd8<uint8_t>
+  simdjson_inline simd8<uint8_t>
   gt_bits(const simd8<uint8_t> other) const {
     return this->saturating_sub(other);
   }
   // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
-  simdjson_really_inline simd8<uint8_t>
+  simdjson_inline simd8<uint8_t>
   lt_bits(const simd8<uint8_t> other) const {
     return other.saturating_sub(*this);
   }
-  simdjson_really_inline simd8<bool>
+  simdjson_inline simd8<bool>
   operator<=(const simd8<uint8_t> other) const {
     return other.max_val(*this) == other;
   }
-  simdjson_really_inline simd8<bool>
+  simdjson_inline simd8<bool>
   operator>=(const simd8<uint8_t> other) const {
     return other.min_val(*this) == other;
   }
-  simdjson_really_inline simd8<bool>
+  simdjson_inline simd8<bool>
   operator>(const simd8<uint8_t> other) const {
     return this->gt_bits(other).any_bits_set();
   }
-  simdjson_really_inline simd8<bool>
+  simdjson_inline simd8<bool>
   operator<(const simd8<uint8_t> other) const {
     return this->gt_bits(other).any_bits_set();
   }
 
   // Bit-specific operations
-  simdjson_really_inline simd8<bool> bits_not_set() const {
+  simdjson_inline simd8<bool> bits_not_set() const {
     return (__m128i)vec_cmpeq(this->value, (__m128i)vec_splats(uint8_t(0)));
   }
-  simdjson_really_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const {
+  simdjson_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const {
     return (*this & bits).bits_not_set();
   }
-  simdjson_really_inline simd8<bool> any_bits_set() const {
+  simdjson_inline simd8<bool> any_bits_set() const {
     return ~this->bits_not_set();
   }
-  simdjson_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const {
+  simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const {
     return ~this->bits_not_set(bits);
   }
-  simdjson_really_inline bool bits_not_set_anywhere() const {
+  simdjson_inline bool bits_not_set_anywhere() const {
     return vec_all_eq(this->value, (__m128i)vec_splats(0));
   }
-  simdjson_really_inline bool any_bits_set_anywhere() const {
+  simdjson_inline bool any_bits_set_anywhere() const {
     return !bits_not_set_anywhere();
   }
-  simdjson_really_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const {
+  simdjson_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const {
     return vec_all_eq(vec_and(this->value, (__m128i)bits),
                       (__m128i)vec_splats(0));
   }
-  simdjson_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const {
+  simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const {
     return !bits_not_set_anywhere(bits);
   }
-  template <int N> simdjson_really_inline simd8<uint8_t> shr() const {
+  template <int N> simdjson_inline simd8<uint8_t> shr() const {
     return simd8<uint8_t>(
         (__m128i)vec_sr(this->value, (__m128i)vec_splat_u8(N)));
   }
-  template <int N> simdjson_really_inline simd8<uint8_t> shl() const {
+  template <int N> simdjson_inline simd8<uint8_t> shl() const {
     return simd8<uint8_t>(
         (__m128i)vec_sl(this->value, (__m128i)vec_splat_u8(N)));
   }
@@ -19206,26 +18750,26 @@ template <typename T> struct simd8x64 {
   operator=(const simd8<T>& other) = delete; // no assignment allowed
   simd8x64() = delete;                      // no default constructor allowed
 
-  simdjson_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1,
+  simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1,
                                   const simd8<T> chunk2, const simd8<T> chunk3)
       : chunks{chunk0, chunk1, chunk2, chunk3} {}
-  simdjson_really_inline simd8x64(const T ptr[64])
+  simdjson_inline simd8x64(const T ptr[64])
       : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr + 16),
                simd8<T>::load(ptr + 32), simd8<T>::load(ptr + 48)} {}
 
-  simdjson_really_inline void store(T ptr[64]) const {
+  simdjson_inline void store(T ptr[64]) const {
     this->chunks[0].store(ptr + sizeof(simd8<T>) * 0);
     this->chunks[1].store(ptr + sizeof(simd8<T>) * 1);
     this->chunks[2].store(ptr + sizeof(simd8<T>) * 2);
     this->chunks[3].store(ptr + sizeof(simd8<T>) * 3);
   }
 
-  simdjson_really_inline simd8<T> reduce_or() const {
+  simdjson_inline simd8<T> reduce_or() const {
     return (this->chunks[0] | this->chunks[1]) |
            (this->chunks[2] | this->chunks[3]);
   }
 
-  simdjson_really_inline uint64_t compress(uint64_t mask, T *output) const {
+  simdjson_inline uint64_t compress(uint64_t mask, T *output) const {
     this->chunks[0].compress(uint16_t(mask), output);
     this->chunks[1].compress(uint16_t(mask >> 16),
                              output + 16 - count_ones(mask & 0xFFFF));
@@ -19236,7 +18780,7 @@ template <typename T> struct simd8x64 {
     return 64 - count_ones(mask);
   }
 
-  simdjson_really_inline uint64_t to_bitmask() const {
+  simdjson_inline uint64_t to_bitmask() const {
     uint64_t r0 = uint32_t(this->chunks[0].to_bitmask());
     uint64_t r1 = this->chunks[1].to_bitmask();
     uint64_t r2 = this->chunks[2].to_bitmask();
@@ -19244,14 +18788,14 @@ template <typename T> struct simd8x64 {
     return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48);
   }
 
-  simdjson_really_inline uint64_t eq(const T m) const {
+  simdjson_inline uint64_t eq(const T m) const {
     const simd8<T> mask = simd8<T>::splat(m);
     return simd8x64<bool>(this->chunks[0] == mask, this->chunks[1] == mask,
                           this->chunks[2] == mask, this->chunks[3] == mask)
         .to_bitmask();
   }
 
-  simdjson_really_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+  simdjson_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
     return simd8x64<bool>(this->chunks[0] == other.chunks[0],
                           this->chunks[1] == other.chunks[1],
                           this->chunks[2] == other.chunks[2],
@@ -19259,7 +18803,7 @@ template <typename T> struct simd8x64 {
         .to_bitmask();
   }
 
-  simdjson_really_inline uint64_t lteq(const T m) const {
+  simdjson_inline uint64_t lteq(const T m) const {
     const simd8<T> mask = simd8<T>::splat(m);
     return simd8x64<bool>(this->chunks[0] <= mask, this->chunks[1] <= mask,
                           this->chunks[2] <= mask, this->chunks[3] <= mask)
@@ -19283,11 +18827,11 @@ namespace jsoncharutils {
 
 // return non-zero if not a structural or whitespace char
 // zero otherwise
-simdjson_really_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
+simdjson_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
   return internal::structural_or_whitespace_negated[c];
 }
 
-simdjson_really_inline uint32_t is_structural_or_whitespace(uint8_t c) {
+simdjson_inline uint32_t is_structural_or_whitespace(uint8_t c) {
   return internal::structural_or_whitespace[c];
 }
 
@@ -19318,7 +18862,7 @@ static inline uint32_t hex_to_u32_nocheck(
 //
 // Note: we assume that surrogates are treated separately
 //
-simdjson_really_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
+simdjson_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
   if (cp <= 0x7F) {
     c[0] = uint8_t(cp);
     return 1; // ascii
@@ -19350,10 +18894,10 @@ simdjson_really_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
 #ifdef SIMDJSON_IS_32BITS // _umul128 for x86, arm
 // this is a slow emulation routine for 32-bit
 //
-static simdjson_really_inline uint64_t __emulu(uint32_t x, uint32_t y) {
+static simdjson_inline uint64_t __emulu(uint32_t x, uint32_t y) {
   return x * (uint64_t)y;
 }
-static simdjson_really_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+static simdjson_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
   uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
   uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
   uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
@@ -19367,7 +18911,7 @@ static simdjson_really_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64
 
 using internal::value128;
 
-simdjson_really_inline value128 full_multiplication(uint64_t value1, uint64_t value2) {
+simdjson_inline value128 full_multiplication(uint64_t value1, uint64_t value2) {
   value128 answer;
 #if defined(SIMDJSON_REGULAR_VISUAL_STUDIO) || defined(SIMDJSON_IS_32BITS)
 #ifdef _M_ARM64
@@ -19403,13 +18947,13 @@ namespace atomparsing {
 // You might think that using memcpy makes this function expensive, but you'd be wrong.
 // All decent optimizing compilers (GCC, clang, Visual Studio) will compile string_to_uint32("false");
 // to the compile-time constant 1936482662.
-simdjson_really_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
+simdjson_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
 
 
 // Again in str4ncmp we use a memcpy to avoid undefined behavior. The memcpy may appear expensive.
 // Yet all decent optimizing compilers will compile memcpy to a single instruction, just about.
 simdjson_warn_unused
-simdjson_really_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
+simdjson_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
   uint32_t srcval; // we want to avoid unaligned 32-bit loads (undefined in C/C++)
   static_assert(sizeof(uint32_t) <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be larger than 4 bytes");
   std::memcpy(&srcval, src, sizeof(uint32_t));
@@ -19417,36 +18961,36 @@ simdjson_really_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_true_atom(const uint8_t *src) {
+simdjson_inline bool is_valid_true_atom(const uint8_t *src) {
   return (str4ncmp(src, "true") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
+simdjson_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
   if (len > 4) { return is_valid_true_atom(src); }
   else if (len == 4) { return !str4ncmp(src, "true"); }
   else { return false; }
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_false_atom(const uint8_t *src) {
+simdjson_inline bool is_valid_false_atom(const uint8_t *src) {
   return (str4ncmp(src+1, "alse") | jsoncharutils::is_not_structural_or_whitespace(src[5])) == 0;
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
+simdjson_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
   if (len > 5) { return is_valid_false_atom(src); }
   else if (len == 5) { return !str4ncmp(src+1, "alse"); }
   else { return false; }
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_null_atom(const uint8_t *src) {
+simdjson_inline bool is_valid_null_atom(const uint8_t *src) {
   return (str4ncmp(src, "null") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
+simdjson_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
   if (len > 4) { return is_valid_null_atom(src); }
   else if (len == 4) { return !str4ncmp(src, "null"); }
   else { return false; }
@@ -19472,17 +19016,17 @@ using namespace simd;
 struct backslash_and_quote {
 public:
   static constexpr uint32_t BYTES_PROCESSED = 32;
-  simdjson_really_inline static backslash_and_quote
+  simdjson_inline static backslash_and_quote
   copy_and_find(const uint8_t *src, uint8_t *dst);
 
-  simdjson_really_inline bool has_quote_first() {
+  simdjson_inline bool has_quote_first() {
     return ((bs_bits - 1) & quote_bits) != 0;
   }
-  simdjson_really_inline bool has_backslash() { return bs_bits != 0; }
-  simdjson_really_inline int quote_index() {
+  simdjson_inline bool has_backslash() { return bs_bits != 0; }
+  simdjson_inline int quote_index() {
     return trailing_zeroes(quote_bits);
   }
-  simdjson_really_inline int backslash_index() {
+  simdjson_inline int backslash_index() {
     return trailing_zeroes(bs_bits);
   }
 
@@ -19490,7 +19034,7 @@ struct backslash_and_quote {
   uint32_t quote_bits;
 }; // struct backslash_and_quote
 
-simdjson_really_inline backslash_and_quote
+simdjson_inline backslash_and_quote
 backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
   // this can read up to 31 bytes beyond the buffer size, but we require
   // SIMDJSON_PADDING of padding
@@ -19517,147 +19061,6 @@ backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
 } // namespace ppc64
 } // namespace simdjson
 
-/* begin file include/simdjson/generic/stringparsing.h */
-// This file contains the common code every implementation uses
-// It is intended to be included multiple times and compiled multiple times
-
-namespace simdjson {
-namespace ppc64 {
-namespace {
-/// @private
-namespace stringparsing {
-
-// begin copypasta
-// These chars yield themselves: " \ /
-// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
-// u not handled in this table as it's complex
-static const uint8_t escape_map[256] = {
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
-    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
-    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
-
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-};
-
-// handle a unicode codepoint
-// write appropriate values into dest
-// src will advance 6 bytes or 12 bytes
-// dest will advance a variable amount (return via pointer)
-// return true if the unicode codepoint was valid
-// We work in little-endian then swap at write time
-simdjson_warn_unused
-simdjson_really_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
-                                            uint8_t **dst_ptr) {
-  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
-  // conversion isn't valid; we defer the check for this to inside the
-  // multilingual plane check
-  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
-  *src_ptr += 6;
-  // check for low surrogate for characters outside the Basic
-  // Multilingual Plane.
-  if (code_point >= 0xd800 && code_point < 0xdc00) {
-    if (((*src_ptr)[0] != '\\') || (*src_ptr)[1] != 'u') {
-      return false;
-    }
-    uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
-
-    // if the first code point is invalid we will get here, as we will go past
-    // the check for being outside the Basic Multilingual plane. If we don't
-    // find a \u immediately afterwards we fail out anyhow, but if we do,
-    // this check catches both the case of the first code point being invalid
-    // or the second code point being invalid.
-    if ((code_point | code_point_2) >> 16) {
-      return false;
-    }
-
-    code_point =
-        (((code_point - 0xd800) << 10) | (code_point_2 - 0xdc00)) + 0x10000;
-    *src_ptr += 6;
-  }
-  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
-  *dst_ptr += offset;
-  return offset > 0;
-}
-
-/**
- * Unescape a string from src to dst, stopping at a final unescaped quote. E.g., if src points at 'joe"', then
- * dst needs to have four free bytes.
- */
-simdjson_warn_unused simdjson_really_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst) {
-  while (1) {
-    // Copy the next n bytes, and find the backslash and quote in them.
-    auto bs_quote = backslash_and_quote::copy_and_find(src, dst);
-    // If the next thing is the end quote, copy and return
-    if (bs_quote.has_quote_first()) {
-      // we encountered quotes first. Move dst to point to quotes and exit
-      return dst + bs_quote.quote_index();
-    }
-    if (bs_quote.has_backslash()) {
-      /* find out where the backspace is */
-      auto bs_dist = bs_quote.backslash_index();
-      uint8_t escape_char = src[bs_dist + 1];
-      /* we encountered backslash first. Handle backslash */
-      if (escape_char == 'u') {
-        /* move src/dst up to the start; they will be further adjusted
-           within the unicode codepoint handling code. */
-        src += bs_dist;
-        dst += bs_dist;
-        if (!handle_unicode_codepoint(&src, &dst)) {
-          return nullptr;
-        }
-      } else {
-        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
-         * write bs_dist+1 characters to output
-         * note this may reach beyond the part of the buffer we've actually
-         * seen. I think this is ok */
-        uint8_t escape_result = escape_map[escape_char];
-        if (escape_result == 0u) {
-          return nullptr; /* bogus escape value is an error */
-        }
-        dst[bs_dist] = escape_result;
-        src += bs_dist + 2;
-        dst += bs_dist + 1;
-      }
-    } else {
-      /* they are the same. Since they can't co-occur, it means we
-       * encountered neither. */
-      src += backslash_and_quote::BYTES_PROCESSED;
-      dst += backslash_and_quote::BYTES_PROCESSED;
-    }
-  }
-  /* can't be reached */
-  return nullptr;
-}
-
-simdjson_unused simdjson_warn_unused simdjson_really_inline error_code parse_string_to_buffer(const uint8_t *src, uint8_t *&current_string_buf_loc, std::string_view &s) {
-  if (*(src++) != '"') { return STRING_ERROR; }
-  auto end = stringparsing::parse_string(src, current_string_buf_loc);
-  if (!end) { return STRING_ERROR; }
-  s = std::string_view(reinterpret_cast<const char *>(current_string_buf_loc), end-current_string_buf_loc);
-  current_string_buf_loc = end;
-  return SUCCESS;
-}
-
-} // namespace stringparsing
-} // unnamed namespace
-} // namespace ppc64
-} // namespace simdjson
-/* end file include/simdjson/generic/stringparsing.h */
-
 #endif // SIMDJSON_PPC64_STRINGPARSING_H
 /* end file include/simdjson/ppc64/stringparsing.h */
 /* begin file include/simdjson/ppc64/numberparsing.h */
@@ -19676,7 +19079,7 @@ namespace {
 
 // we don't have appropriate instructions, so let us use a scalar function
 // credit: https://johnnylee-sde.github.io/Fast-numeric-string-to-int/
-static simdjson_really_inline uint32_t
+static simdjson_inline uint32_t
 parse_eight_digits_unrolled(const uint8_t *chars) {
   uint64_t val;
   std::memcpy(&val, chars, sizeof(uint64_t));
@@ -19737,7 +19140,7 @@ namespace {
 // Convert a mantissa, an exponent and a sign bit into an ieee64 double.
 // The real_exponent needs to be in [0, 2046] (technically real_exponent = 2047 would be acceptable).
 // The mantissa should be in [0,1<<53). The bit at index (1ULL << 52) while be zeroed.
-simdjson_really_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
+simdjson_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
     double d;
     mantissa &= ~(1ULL << 52);
     mantissa |= real_exponent << 52;
@@ -19752,7 +19155,7 @@ simdjson_really_inline double to_double(uint64_t mantissa, uint64_t real_exponen
 // set to false. This should work *most of the time* (like 99% of the time).
 // We assume that power is in the [smallest_power,
 // largest_power] interval: the caller is responsible for this check.
-simdjson_really_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
+simdjson_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
   // we start with a fast path
   // It was described in
   // Clinger WD. How to read floating point numbers accurately.
@@ -19812,7 +19215,7 @@ simdjson_really_inline bool compute_float_64(int64_t power, uint64_t i, bool neg
   // In the slow path, we need to adjust i so that it is > 1<<63 which is always
   // possible, except if i == 0, so we handle i == 0 separately.
   if(i == 0) {
-    d = 0.0;
+    d = negative ? -0.0 : 0.0;
     return true;
   }
 
@@ -19927,7 +19330,7 @@ simdjson_really_inline bool compute_float_64(int64_t power, uint64_t i, bool neg
   if (simdjson_unlikely(real_exponent <= 0)) { // we have a subnormal?
     // Here have that real_exponent <= 0 so -real_exponent >= 0
     if(-real_exponent + 1 >= 64) { // if we have more than 64 bits below the minimum exponent, you have a zero for sure.
-      d = 0.0;
+      d = negative ? -0.0 : 0.0;
       return true;
     }
     // next line is safe because -real_exponent + 1 < 0
@@ -20035,7 +19438,7 @@ static bool parse_float_fallback(const uint8_t *ptr, const uint8_t *end_ptr, dou
 // check quickly whether the next 8 chars are made of digits
 // at a glance, it looks better than Mula's
 // http://0x80.pl/articles/swar-digits-validate.html
-simdjson_really_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
+simdjson_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
   uint64_t val;
   // this can read up to 7 bytes beyond the buffer size, but we require
   // SIMDJSON_PADDING of padding
@@ -20062,7 +19465,7 @@ error_code slow_float_parsing(simdjson_unused const uint8_t * src, W writer) {
 
 template<typename I>
 SIMDJSON_NO_SANITIZE_UNDEFINED // We deliberately allow overflow here and check later
-simdjson_really_inline bool parse_digit(const uint8_t c, I &i) {
+simdjson_inline bool parse_digit(const uint8_t c, I &i) {
   const uint8_t digit = static_cast<uint8_t>(c - '0');
   if (digit > 9) {
     return false;
@@ -20072,7 +19475,7 @@ simdjson_really_inline bool parse_digit(const uint8_t c, I &i) {
   return true;
 }
 
-simdjson_really_inline error_code parse_decimal(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
+simdjson_inline error_code parse_decimal(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
   // we continue with the fiction that we have an integer. If the
   // floating point number is representable as x * 10^z for some integer
   // z that fits in 53 bits, then we will be able to convert back the
@@ -20100,7 +19503,7 @@ simdjson_really_inline error_code parse_decimal(simdjson_unused const uint8_t *c
   return SUCCESS;
 }
 
-simdjson_really_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
+simdjson_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
   // Exp Sign: -123.456e[-]78
   bool neg_exp = ('-' == *p);
   if (neg_exp || '+' == *p) { p++; } // Skip + as well
@@ -20151,7 +19554,7 @@ simdjson_really_inline error_code parse_exponent(simdjson_unused const uint8_t *
   return SUCCESS;
 }
 
-simdjson_really_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
+simdjson_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
   // It is possible that the integer had an overflow.
   // We have to handle the case where we have 0.0000somenumber.
   const uint8_t *start = start_digits;
@@ -20161,7 +19564,7 @@ simdjson_really_inline size_t significant_digits(const uint8_t * start_digits, s
 }
 
 template<typename W>
-simdjson_really_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
+simdjson_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
   // If we frequently had to deal with long strings of digits,
   // we could extend our code by using a 128-bit integer instead
   // of a 64-bit integer. However, this is uncommon in practice.
@@ -20197,7 +19600,8 @@ simdjson_really_inline error_code write_float(const uint8_t *const src, bool neg
     static_assert(simdjson::internal::smallest_power <= -342, "smallest_power is not small enough");
     //
     if((exponent < simdjson::internal::smallest_power) || (i == 0)) {
-      WRITE_DOUBLE(0, src, writer);
+      // E.g. Parse "-0.0e-999" into the same value as "-0.0". See https://en.wikipedia.org/wiki/Signed_zero
+      WRITE_DOUBLE(negative ? -0.0 : 0.0, src, writer);
       return SUCCESS;
     } else { // (exponent > largest_power) and (i != 0)
       // We have, for sure, an infinite value and simdjson refuses to parse infinite values.
@@ -20217,20 +19621,20 @@ simdjson_really_inline error_code write_float(const uint8_t *const src, bool neg
 #ifdef SIMDJSON_SKIPNUMBERPARSING
 
 template<typename W>
-simdjson_really_inline error_code parse_number(const uint8_t *const, W &writer) {
+simdjson_inline error_code parse_number(const uint8_t *const, W &writer) {
   writer.append_s64(0);        // always write zero
   return SUCCESS;              // always succeeds
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
-simdjson_unused simdjson_really_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
-simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept { return ondemand::number_type::signed_integer; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept { return ondemand::number_type::signed_integer; }
 #else
 
 // parse the number at src
@@ -20243,13 +19647,13 @@ simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> ge
 //
 // Our objective is accurate parsing (ULP of 0) at high speed.
 template<typename W>
-simdjson_really_inline error_code parse_number(const uint8_t *const src, W &writer) {
+simdjson_inline error_code parse_number(const uint8_t *const src, W &writer) {
 
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  const uint8_t *p = src + negative;
+  const uint8_t *p = src + uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -20393,7 +19797,7 @@ const uint8_t integer_string_finisher[256] = {
     NUMBER_ERROR};
 
 // Parse any number from 0 to 18,446,744,073,709,551,615
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
   const uint8_t *p = src;
   //
   // Parse the integer part.
@@ -20443,7 +19847,7 @@ simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(
 
 // Parse any number from 0 to 18,446,744,073,709,551,615
 // Never read at src_end or beyond
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
   const uint8_t *p = src;
   //
   // Parse the integer part.
@@ -20491,7 +19895,7 @@ simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(
 }
 
 // Parse any number from 0 to 18,446,744,073,709,551,615
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
   const uint8_t *p = src + 1;
   //
   // Parse the integer part.
@@ -20541,12 +19945,12 @@ simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned_
 }
 
 // Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  const uint8_t *p = src + negative;
+  const uint8_t *p = src + uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -20584,13 +19988,13 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(co
 
 // Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
 // Never read at src_end or beyond
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
   //
   // Check for minus sign
   //
   if(src == src_end) { return NUMBER_ERROR; }
   bool negative = (*src == '-');
-  const uint8_t *p = src + negative;
+  const uint8_t *p = src + uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -20627,24 +20031,24 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(co
 }
 
 // Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*(src + 1) == '-');
-  const uint8_t *p = src + negative + 1;
+  src += uint8_t(negative) + 1;
 
   //
   // Parse the integer part.
   //
   // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
-  const uint8_t *const start_digits = p;
+  const uint8_t *const start_digits = src;
   uint64_t i = 0;
-  while (parse_digit(*p, i)) { p++; }
+  while (parse_digit(*src, i)) { src++; }
 
   // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
   // Optimization note: size_t is expected to be unsigned.
-  size_t digit_count = size_t(p - start_digits);
+  size_t digit_count = size_t(src - start_digits);
   // We go from
   // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
   // so we can never represent numbers that have more than 19 digits.
@@ -20656,11 +20060,11 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in
   // Here digit_count > 0.
   if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
   // We can do the following...
-  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
-  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // if (!jsoncharutils::is_structural_or_whitespace(*src)) {
+  //  return (*src == '.' || *src == 'e' || *src == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
   // }
   // as a single table lookup:
-  if(*p != '"') { return NUMBER_ERROR; }
+  if(*src != '"') { return NUMBER_ERROR; }
   // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
   // Performance note: This check is only needed when digit_count == longest_digit_count but it is
   // so cheap that we might as well always make it.
@@ -20668,12 +20072,12 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in
   return negative ? (~i+1) : i;
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -20740,19 +20144,19 @@ simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(cons
   if (simdjson_likely(!overflow)) {
     if (compute_float_64(exponent, i, negative, d)) { return d; }
   }
-  if (!parse_float_fallback(src-negative, &d)) {
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
     return NUMBER_ERROR;
   }
   return d;
 }
 
-simdjson_unused simdjson_really_inline bool is_negative(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept {
   return (*src == '-');
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
   const uint8_t *p = src;
   while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
   if ( p == src ) { return NUMBER_ERROR; }
@@ -20760,9 +20164,9 @@ simdjson_unused simdjson_really_inline simdjson_result<bool> is_integer(const ui
   return false;
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept {
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
   const uint8_t *p = src;
   while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
   if ( p == src ) { return NUMBER_ERROR; }
@@ -20786,13 +20190,13 @@ simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> ge
 }
 
 // Never read at src_end or beyond
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
   if(src == src_end) { return NUMBER_ERROR; }
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -20861,18 +20265,18 @@ simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(cons
   if (simdjson_likely(!overflow)) {
     if (compute_float_64(exponent, i, negative, d)) { return d; }
   }
-  if (!parse_float_fallback(src-negative, src_end, &d)) {
+  if (!parse_float_fallback(src - uint8_t(negative), src_end, &d)) {
     return NUMBER_ERROR;
   }
   return d;
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*(src + 1) == '-');
-  src += negative + 1;
+  src += uint8_t(negative) + 1;
 
   //
   // Parse the integer part.
@@ -20939,7 +20343,7 @@ simdjson_unused simdjson_really_inline simdjson_result<double> parse_double_in_s
   if (simdjson_likely(!overflow)) {
     if (compute_float_64(exponent, i, negative, d)) { return d; }
   }
-  if (!parse_float_fallback(src-negative, &d)) {
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
     return NUMBER_ERROR;
   }
   return d;
@@ -21004,7 +20408,7 @@ using namespace simdjson::dom;
 
 class implementation final : public simdjson::implementation {
 public:
-  simdjson_really_inline implementation() : simdjson::implementation("westmere", "Intel/AMD SSE4.2", internal::instruction_set::SSE42 | internal::instruction_set::PCLMULQDQ) {}
+  simdjson_inline implementation() : simdjson::implementation("westmere", "Intel/AMD SSE4.2", internal::instruction_set::SSE42 | internal::instruction_set::PCLMULQDQ) {}
   simdjson_warn_unused error_code create_dom_parser_implementation(
     size_t capacity,
     size_t max_length,
@@ -21044,7 +20448,7 @@ class implementation final : public simdjson::implementation {
 #include <wmmintrin.h>  // for  _mm_clmulepi64_si128
 #endif
 
-
+static_assert(sizeof(__m128i) <= simdjson::SIMDJSON_PADDING, "insufficient padding for westmere");
 
 #endif // SIMDJSON_WESTMERE_INTRINSICS_H
 /* end file include/simdjson/westmere/intrinsics.h */
@@ -21095,10 +20499,11 @@ class dom_parser_implementation final : public internal::dom_parser_implementati
   simdjson_warn_unused error_code stage1(const uint8_t *buf, size_t len, stage1_mode partial) noexcept final;
   simdjson_warn_unused error_code stage2(dom::document &doc) noexcept final;
   simdjson_warn_unused error_code stage2_next(dom::document &doc) noexcept final;
+  simdjson_warn_unused uint8_t *parse_string(const uint8_t *src, uint8_t *dst) const noexcept final;
   inline simdjson_warn_unused error_code set_capacity(size_t capacity) noexcept final;
   inline simdjson_warn_unused error_code set_max_depth(size_t max_depth) noexcept final;
 private:
-  simdjson_really_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
+  simdjson_inline simdjson_warn_unused error_code set_capacity_stage1(size_t capacity);
 
 };
 
@@ -21151,7 +20556,7 @@ namespace {
 // but the algorithms do not end up using the returned value.
 // Sadly, sanitizers are not smart enough to figure it out.
 SIMDJSON_NO_SANITIZE_UNDEFINED
-simdjson_really_inline int trailing_zeroes(uint64_t input_num) {
+simdjson_inline int trailing_zeroes(uint64_t input_num) {
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
   unsigned long ret;
   // Search the mask data from least significant bit (LSB)
@@ -21164,12 +20569,12 @@ simdjson_really_inline int trailing_zeroes(uint64_t input_num) {
 }
 
 /* result might be undefined when input_num is zero */
-simdjson_really_inline uint64_t clear_lowest_bit(uint64_t input_num) {
+simdjson_inline uint64_t clear_lowest_bit(uint64_t input_num) {
   return input_num & (input_num-1);
 }
 
 /* result might be undefined when input_num is zero */
-simdjson_really_inline int leading_zeroes(uint64_t input_num) {
+simdjson_inline int leading_zeroes(uint64_t input_num) {
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
   unsigned long leading_zero = 0;
   // Search the mask data from most significant bit (MSB)
@@ -21184,17 +20589,17 @@ simdjson_really_inline int leading_zeroes(uint64_t input_num) {
 }
 
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
-simdjson_really_inline unsigned __int64 count_ones(uint64_t input_num) {
+simdjson_inline unsigned __int64 count_ones(uint64_t input_num) {
   // note: we do not support legacy 32-bit Windows
   return __popcnt64(input_num);// Visual Studio wants two underscores
 }
 #else
-simdjson_really_inline long long int count_ones(uint64_t input_num) {
+simdjson_inline long long int count_ones(uint64_t input_num) {
   return _popcnt64(input_num);
 }
 #endif
 
-simdjson_really_inline bool add_overflow(uint64_t value1, uint64_t value2,
+simdjson_inline bool add_overflow(uint64_t value1, uint64_t value2,
                                 uint64_t *result) {
 #ifdef SIMDJSON_REGULAR_VISUAL_STUDIO
   return _addcarry_u64(0, value1, value2,
@@ -21224,7 +20629,7 @@ namespace {
 //
 // For example, prefix_xor(00100100) == 00011100
 //
-simdjson_really_inline uint64_t prefix_xor(const uint64_t bitmask) {
+simdjson_inline uint64_t prefix_xor(const uint64_t bitmask) {
   // There should be no such thing with a processing supporting avx2
   // but not clmul.
   __m128i all_ones = _mm_set1_epi8('\xFF');
@@ -21253,23 +20658,23 @@ namespace simd {
     __m128i value;
 
     // Zero constructor
-    simdjson_really_inline base() : value{__m128i()} {}
+    simdjson_inline base() : value{__m128i()} {}
 
     // Conversion from SIMD register
-    simdjson_really_inline base(const __m128i _value) : value(_value) {}
+    simdjson_inline base(const __m128i _value) : value(_value) {}
 
     // Conversion to SIMD register
-    simdjson_really_inline operator const __m128i&() const { return this->value; }
-    simdjson_really_inline operator __m128i&() { return this->value; }
+    simdjson_inline operator const __m128i&() const { return this->value; }
+    simdjson_inline operator __m128i&() { return this->value; }
 
     // Bit operations
-    simdjson_really_inline Child operator|(const Child other) const { return _mm_or_si128(*this, other); }
-    simdjson_really_inline Child operator&(const Child other) const { return _mm_and_si128(*this, other); }
-    simdjson_really_inline Child operator^(const Child other) const { return _mm_xor_si128(*this, other); }
-    simdjson_really_inline Child bit_andnot(const Child other) const { return _mm_andnot_si128(other, *this); }
-    simdjson_really_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
-    simdjson_really_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
-    simdjson_really_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
+    simdjson_inline Child operator|(const Child other) const { return _mm_or_si128(*this, other); }
+    simdjson_inline Child operator&(const Child other) const { return _mm_and_si128(*this, other); }
+    simdjson_inline Child operator^(const Child other) const { return _mm_xor_si128(*this, other); }
+    simdjson_inline Child bit_andnot(const Child other) const { return _mm_andnot_si128(other, *this); }
+    simdjson_inline Child& operator|=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast | other; return *this_cast; }
+    simdjson_inline Child& operator&=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast & other; return *this_cast; }
+    simdjson_inline Child& operator^=(const Child other) { auto this_cast = static_cast<Child*>(this); *this_cast = *this_cast ^ other; return *this_cast; }
   };
 
   // Forward-declared so they can be used by splat and friends.
@@ -21281,15 +20686,15 @@ namespace simd {
     typedef uint16_t bitmask_t;
     typedef uint32_t bitmask2_t;
 
-    simdjson_really_inline base8() : base<simd8<T>>() {}
-    simdjson_really_inline base8(const __m128i _value) : base<simd8<T>>(_value) {}
+    simdjson_inline base8() : base<simd8<T>>() {}
+    simdjson_inline base8(const __m128i _value) : base<simd8<T>>(_value) {}
 
-    friend simdjson_really_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) { return _mm_cmpeq_epi8(lhs, rhs); }
+    friend simdjson_inline Mask operator==(const simd8<T> lhs, const simd8<T> rhs) { return _mm_cmpeq_epi8(lhs, rhs); }
 
     static const int SIZE = sizeof(base<simd8<T>>::value);
 
     template<int N=1>
-    simdjson_really_inline simd8<T> prev(const simd8<T> prev_chunk) const {
+    simdjson_inline simd8<T> prev(const simd8<T> prev_chunk) const {
       return _mm_alignr_epi8(*this, prev_chunk, 16 - N);
     }
   };
@@ -21297,27 +20702,27 @@ namespace simd {
   // SIMD byte mask type (returned by things like eq and gt)
   template<>
   struct simd8<bool>: base8<bool> {
-    static simdjson_really_inline simd8<bool> splat(bool _value) { return _mm_set1_epi8(uint8_t(-(!!_value))); }
+    static simdjson_inline simd8<bool> splat(bool _value) { return _mm_set1_epi8(uint8_t(-(!!_value))); }
 
-    simdjson_really_inline simd8<bool>() : base8() {}
-    simdjson_really_inline simd8<bool>(const __m128i _value) : base8<bool>(_value) {}
+    simdjson_inline simd8<bool>() : base8() {}
+    simdjson_inline simd8<bool>(const __m128i _value) : base8<bool>(_value) {}
     // Splat constructor
-    simdjson_really_inline simd8<bool>(bool _value) : base8<bool>(splat(_value)) {}
+    simdjson_inline simd8<bool>(bool _value) : base8<bool>(splat(_value)) {}
 
-    simdjson_really_inline int to_bitmask() const { return _mm_movemask_epi8(*this); }
-    simdjson_really_inline bool any() const { return !_mm_testz_si128(*this, *this); }
-    simdjson_really_inline simd8<bool> operator~() const { return *this ^ true; }
+    simdjson_inline int to_bitmask() const { return _mm_movemask_epi8(*this); }
+    simdjson_inline bool any() const { return !_mm_testz_si128(*this, *this); }
+    simdjson_inline simd8<bool> operator~() const { return *this ^ true; }
   };
 
   template<typename T>
   struct base8_numeric: base8<T> {
-    static simdjson_really_inline simd8<T> splat(T _value) { return _mm_set1_epi8(_value); }
-    static simdjson_really_inline simd8<T> zero() { return _mm_setzero_si128(); }
-    static simdjson_really_inline simd8<T> load(const T values[16]) {
+    static simdjson_inline simd8<T> splat(T _value) { return _mm_set1_epi8(_value); }
+    static simdjson_inline simd8<T> zero() { return _mm_setzero_si128(); }
+    static simdjson_inline simd8<T> load(const T values[16]) {
       return _mm_loadu_si128(reinterpret_cast<const __m128i *>(values));
     }
     // Repeat 16 values as many times as necessary (usually for lookup tables)
-    static simdjson_really_inline simd8<T> repeat_16(
+    static simdjson_inline simd8<T> repeat_16(
       T v0,  T v1,  T v2,  T v3,  T v4,  T v5,  T v6,  T v7,
       T v8,  T v9,  T v10, T v11, T v12, T v13, T v14, T v15
     ) {
@@ -21327,24 +20732,24 @@ namespace simd {
       );
     }
 
-    simdjson_really_inline base8_numeric() : base8<T>() {}
-    simdjson_really_inline base8_numeric(const __m128i _value) : base8<T>(_value) {}
+    simdjson_inline base8_numeric() : base8<T>() {}
+    simdjson_inline base8_numeric(const __m128i _value) : base8<T>(_value) {}
 
     // Store to array
-    simdjson_really_inline void store(T dst[16]) const { return _mm_storeu_si128(reinterpret_cast<__m128i *>(dst), *this); }
+    simdjson_inline void store(T dst[16]) const { return _mm_storeu_si128(reinterpret_cast<__m128i *>(dst), *this); }
 
     // Override to distinguish from bool version
-    simdjson_really_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
+    simdjson_inline simd8<T> operator~() const { return *this ^ 0xFFu; }
 
     // Addition/subtraction are the same for signed and unsigned
-    simdjson_really_inline simd8<T> operator+(const simd8<T> other) const { return _mm_add_epi8(*this, other); }
-    simdjson_really_inline simd8<T> operator-(const simd8<T> other) const { return _mm_sub_epi8(*this, other); }
-    simdjson_really_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
-    simdjson_really_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
+    simdjson_inline simd8<T> operator+(const simd8<T> other) const { return _mm_add_epi8(*this, other); }
+    simdjson_inline simd8<T> operator-(const simd8<T> other) const { return _mm_sub_epi8(*this, other); }
+    simdjson_inline simd8<T>& operator+=(const simd8<T> other) { *this = *this + other; return *static_cast<simd8<T>*>(this); }
+    simdjson_inline simd8<T>& operator-=(const simd8<T> other) { *this = *this - other; return *static_cast<simd8<T>*>(this); }
 
     // Perform a lookup assuming the value is between 0 and 16 (undefined behavior for out of range values)
     template<typename L>
-    simdjson_really_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
+    simdjson_inline simd8<L> lookup_16(simd8<L> lookup_table) const {
       return _mm_shuffle_epi8(lookup_table, *this);
     }
 
@@ -21356,7 +20761,7 @@ namespace simd {
     // signature simd8<L> compress(uint32_t mask) would be
     // sensible, but the AVX ISA makes this kind of approach difficult.
     template<typename L>
-    simdjson_really_inline void compress(uint16_t mask, L * output) const {
+    simdjson_inline void compress(uint16_t mask, L * output) const {
       using internal::thintable_epi8;
       using internal::BitsSetTable256mul2;
       using internal::pshufb_combine_table;
@@ -21387,7 +20792,7 @@ namespace simd {
     }
 
     template<typename L>
-    simdjson_really_inline simd8<L> lookup_16(
+    simdjson_inline simd8<L> lookup_16(
         L replace0,  L replace1,  L replace2,  L replace3,
         L replace4,  L replace5,  L replace6,  L replace7,
         L replace8,  L replace9,  L replace10, L replace11,
@@ -21404,14 +20809,14 @@ namespace simd {
   // Signed bytes
   template<>
   struct simd8<int8_t> : base8_numeric<int8_t> {
-    simdjson_really_inline simd8() : base8_numeric<int8_t>() {}
-    simdjson_really_inline simd8(const __m128i _value) : base8_numeric<int8_t>(_value) {}
+    simdjson_inline simd8() : base8_numeric<int8_t>() {}
+    simdjson_inline simd8(const __m128i _value) : base8_numeric<int8_t>(_value) {}
     // Splat constructor
-    simdjson_really_inline simd8(int8_t _value) : simd8(splat(_value)) {}
+    simdjson_inline simd8(int8_t _value) : simd8(splat(_value)) {}
     // Array constructor
-    simdjson_really_inline simd8(const int8_t* values) : simd8(load(values)) {}
+    simdjson_inline simd8(const int8_t* values) : simd8(load(values)) {}
     // Member-by-member initialization
-    simdjson_really_inline simd8(
+    simdjson_inline simd8(
       int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
       int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
     ) : simd8(_mm_setr_epi8(
@@ -21419,7 +20824,7 @@ namespace simd {
       v8, v9, v10,v11,v12,v13,v14,v15
     )) {}
     // Repeat 16 values as many times as necessary (usually for lookup tables)
-    simdjson_really_inline static simd8<int8_t> repeat_16(
+    simdjson_inline static simd8<int8_t> repeat_16(
       int8_t v0,  int8_t v1,  int8_t v2,  int8_t v3,  int8_t v4,  int8_t v5,  int8_t v6,  int8_t v7,
       int8_t v8,  int8_t v9,  int8_t v10, int8_t v11, int8_t v12, int8_t v13, int8_t v14, int8_t v15
     ) {
@@ -21430,23 +20835,23 @@ namespace simd {
     }
 
     // Order-sensitive comparisons
-    simdjson_really_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return _mm_max_epi8(*this, other); }
-    simdjson_really_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return _mm_min_epi8(*this, other); }
-    simdjson_really_inline simd8<bool> operator>(const simd8<int8_t> other) const { return _mm_cmpgt_epi8(*this, other); }
-    simdjson_really_inline simd8<bool> operator<(const simd8<int8_t> other) const { return _mm_cmpgt_epi8(other, *this); }
+    simdjson_inline simd8<int8_t> max_val(const simd8<int8_t> other) const { return _mm_max_epi8(*this, other); }
+    simdjson_inline simd8<int8_t> min_val(const simd8<int8_t> other) const { return _mm_min_epi8(*this, other); }
+    simdjson_inline simd8<bool> operator>(const simd8<int8_t> other) const { return _mm_cmpgt_epi8(*this, other); }
+    simdjson_inline simd8<bool> operator<(const simd8<int8_t> other) const { return _mm_cmpgt_epi8(other, *this); }
   };
 
   // Unsigned bytes
   template<>
   struct simd8<uint8_t>: base8_numeric<uint8_t> {
-    simdjson_really_inline simd8() : base8_numeric<uint8_t>() {}
-    simdjson_really_inline simd8(const __m128i _value) : base8_numeric<uint8_t>(_value) {}
+    simdjson_inline simd8() : base8_numeric<uint8_t>() {}
+    simdjson_inline simd8(const __m128i _value) : base8_numeric<uint8_t>(_value) {}
     // Splat constructor
-    simdjson_really_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
+    simdjson_inline simd8(uint8_t _value) : simd8(splat(_value)) {}
     // Array constructor
-    simdjson_really_inline simd8(const uint8_t* values) : simd8(load(values)) {}
+    simdjson_inline simd8(const uint8_t* values) : simd8(load(values)) {}
     // Member-by-member initialization
-    simdjson_really_inline simd8(
+    simdjson_inline simd8(
       uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
       uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
     ) : simd8(_mm_setr_epi8(
@@ -21454,7 +20859,7 @@ namespace simd {
       v8, v9, v10,v11,v12,v13,v14,v15
     )) {}
     // Repeat 16 values as many times as necessary (usually for lookup tables)
-    simdjson_really_inline static simd8<uint8_t> repeat_16(
+    simdjson_inline static simd8<uint8_t> repeat_16(
       uint8_t v0,  uint8_t v1,  uint8_t v2,  uint8_t v3,  uint8_t v4,  uint8_t v5,  uint8_t v6,  uint8_t v7,
       uint8_t v8,  uint8_t v9,  uint8_t v10, uint8_t v11, uint8_t v12, uint8_t v13, uint8_t v14, uint8_t v15
     ) {
@@ -21465,39 +20870,39 @@ namespace simd {
     }
 
     // Saturated math
-    simdjson_really_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return _mm_adds_epu8(*this, other); }
-    simdjson_really_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return _mm_subs_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_add(const simd8<uint8_t> other) const { return _mm_adds_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> saturating_sub(const simd8<uint8_t> other) const { return _mm_subs_epu8(*this, other); }
 
     // Order-specific operations
-    simdjson_really_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return _mm_max_epu8(*this, other); }
-    simdjson_really_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return _mm_min_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> max_val(const simd8<uint8_t> other) const { return _mm_max_epu8(*this, other); }
+    simdjson_inline simd8<uint8_t> min_val(const simd8<uint8_t> other) const { return _mm_min_epu8(*this, other); }
     // Same as >, but only guarantees true is nonzero (< guarantees true = -1)
-    simdjson_really_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
+    simdjson_inline simd8<uint8_t> gt_bits(const simd8<uint8_t> other) const { return this->saturating_sub(other); }
     // Same as <, but only guarantees true is nonzero (< guarantees true = -1)
-    simdjson_really_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
-    simdjson_really_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
-    simdjson_really_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
-    simdjson_really_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
-    simdjson_really_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+    simdjson_inline simd8<uint8_t> lt_bits(const simd8<uint8_t> other) const { return other.saturating_sub(*this); }
+    simdjson_inline simd8<bool> operator<=(const simd8<uint8_t> other) const { return other.max_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>=(const simd8<uint8_t> other) const { return other.min_val(*this) == other; }
+    simdjson_inline simd8<bool> operator>(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
+    simdjson_inline simd8<bool> operator<(const simd8<uint8_t> other) const { return this->gt_bits(other).any_bits_set(); }
 
     // Bit-specific operations
-    simdjson_really_inline simd8<bool> bits_not_set() const { return *this == uint8_t(0); }
-    simdjson_really_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
-    simdjson_really_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
-    simdjson_really_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
-    simdjson_really_inline bool is_ascii() const { return _mm_movemask_epi8(*this) == 0; }
-    simdjson_really_inline bool bits_not_set_anywhere() const { return _mm_testz_si128(*this, *this); }
-    simdjson_really_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
-    simdjson_really_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const { return _mm_testz_si128(*this, bits); }
-    simdjson_really_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
+    simdjson_inline simd8<bool> bits_not_set() const { return *this == uint8_t(0); }
+    simdjson_inline simd8<bool> bits_not_set(simd8<uint8_t> bits) const { return (*this & bits).bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set() const { return ~this->bits_not_set(); }
+    simdjson_inline simd8<bool> any_bits_set(simd8<uint8_t> bits) const { return ~this->bits_not_set(bits); }
+    simdjson_inline bool is_ascii() const { return _mm_movemask_epi8(*this) == 0; }
+    simdjson_inline bool bits_not_set_anywhere() const { return _mm_testz_si128(*this, *this); }
+    simdjson_inline bool any_bits_set_anywhere() const { return !bits_not_set_anywhere(); }
+    simdjson_inline bool bits_not_set_anywhere(simd8<uint8_t> bits) const { return _mm_testz_si128(*this, bits); }
+    simdjson_inline bool any_bits_set_anywhere(simd8<uint8_t> bits) const { return !bits_not_set_anywhere(bits); }
     template<int N>
-    simdjson_really_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(_mm_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); }
+    simdjson_inline simd8<uint8_t> shr() const { return simd8<uint8_t>(_mm_srli_epi16(*this, N)) & uint8_t(0xFFu >> N); }
     template<int N>
-    simdjson_really_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(_mm_slli_epi16(*this, N)) & uint8_t(0xFFu << N); }
+    simdjson_inline simd8<uint8_t> shl() const { return simd8<uint8_t>(_mm_slli_epi16(*this, N)) & uint8_t(0xFFu << N); }
     // Get one of the bits and make a bitmask out of it.
     // e.g. value.get_bit<7>() gets the high bit
     template<int N>
-    simdjson_really_inline int get_bit() const { return _mm_movemask_epi8(_mm_slli_epi16(*this, 7-N)); }
+    simdjson_inline int get_bit() const { return _mm_movemask_epi8(_mm_slli_epi16(*this, 7-N)); }
   };
 
   template<typename T>
@@ -21510,21 +20915,21 @@ namespace simd {
     simd8x64<T>& operator=(const simd8<T>& other) = delete; // no assignment allowed
     simd8x64() = delete; // no default constructor allowed
 
-    simdjson_really_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1, const simd8<T> chunk2, const simd8<T> chunk3) : chunks{chunk0, chunk1, chunk2, chunk3} {}
-    simdjson_really_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+16), simd8<T>::load(ptr+32), simd8<T>::load(ptr+48)} {}
+    simdjson_inline simd8x64(const simd8<T> chunk0, const simd8<T> chunk1, const simd8<T> chunk2, const simd8<T> chunk3) : chunks{chunk0, chunk1, chunk2, chunk3} {}
+    simdjson_inline simd8x64(const T ptr[64]) : chunks{simd8<T>::load(ptr), simd8<T>::load(ptr+16), simd8<T>::load(ptr+32), simd8<T>::load(ptr+48)} {}
 
-    simdjson_really_inline void store(T ptr[64]) const {
+    simdjson_inline void store(T ptr[64]) const {
       this->chunks[0].store(ptr+sizeof(simd8<T>)*0);
       this->chunks[1].store(ptr+sizeof(simd8<T>)*1);
       this->chunks[2].store(ptr+sizeof(simd8<T>)*2);
       this->chunks[3].store(ptr+sizeof(simd8<T>)*3);
     }
 
-    simdjson_really_inline simd8<T> reduce_or() const {
+    simdjson_inline simd8<T> reduce_or() const {
       return (this->chunks[0] | this->chunks[1]) | (this->chunks[2] | this->chunks[3]);
     }
 
-    simdjson_really_inline uint64_t compress(uint64_t mask, T * output) const {
+    simdjson_inline uint64_t compress(uint64_t mask, T * output) const {
       this->chunks[0].compress(uint16_t(mask), output);
       this->chunks[1].compress(uint16_t(mask >> 16), output + 16 - count_ones(mask & 0xFFFF));
       this->chunks[2].compress(uint16_t(mask >> 32), output + 32 - count_ones(mask & 0xFFFFFFFF));
@@ -21532,7 +20937,7 @@ namespace simd {
       return 64 - count_ones(mask);
     }
 
-    simdjson_really_inline uint64_t to_bitmask() const {
+    simdjson_inline uint64_t to_bitmask() const {
       uint64_t r0 = uint32_t(this->chunks[0].to_bitmask() );
       uint64_t r1 =          this->chunks[1].to_bitmask() ;
       uint64_t r2 =          this->chunks[2].to_bitmask() ;
@@ -21540,7 +20945,7 @@ namespace simd {
       return r0 | (r1 << 16) | (r2 << 32) | (r3 << 48);
     }
 
-    simdjson_really_inline uint64_t eq(const T m) const {
+    simdjson_inline uint64_t eq(const T m) const {
       const simd8<T> mask = simd8<T>::splat(m);
       return  simd8x64<bool>(
         this->chunks[0] == mask,
@@ -21550,7 +20955,7 @@ namespace simd {
       ).to_bitmask();
     }
 
-    simdjson_really_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
+    simdjson_inline uint64_t eq(const simd8x64<uint8_t> &other) const {
       return  simd8x64<bool>(
         this->chunks[0] == other.chunks[0],
         this->chunks[1] == other.chunks[1],
@@ -21559,7 +20964,7 @@ namespace simd {
       ).to_bitmask();
     }
 
-    simdjson_really_inline uint64_t lteq(const T m) const {
+    simdjson_inline uint64_t lteq(const T m) const {
       const simd8<T> mask = simd8<T>::splat(m);
       return  simd8x64<bool>(
         this->chunks[0] <= mask,
@@ -21586,11 +20991,11 @@ namespace jsoncharutils {
 
 // return non-zero if not a structural or whitespace char
 // zero otherwise
-simdjson_really_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
+simdjson_inline uint32_t is_not_structural_or_whitespace(uint8_t c) {
   return internal::structural_or_whitespace_negated[c];
 }
 
-simdjson_really_inline uint32_t is_structural_or_whitespace(uint8_t c) {
+simdjson_inline uint32_t is_structural_or_whitespace(uint8_t c) {
   return internal::structural_or_whitespace[c];
 }
 
@@ -21621,7 +21026,7 @@ static inline uint32_t hex_to_u32_nocheck(
 //
 // Note: we assume that surrogates are treated separately
 //
-simdjson_really_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
+simdjson_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
   if (cp <= 0x7F) {
     c[0] = uint8_t(cp);
     return 1; // ascii
@@ -21653,10 +21058,10 @@ simdjson_really_inline size_t codepoint_to_utf8(uint32_t cp, uint8_t *c) {
 #ifdef SIMDJSON_IS_32BITS // _umul128 for x86, arm
 // this is a slow emulation routine for 32-bit
 //
-static simdjson_really_inline uint64_t __emulu(uint32_t x, uint32_t y) {
+static simdjson_inline uint64_t __emulu(uint32_t x, uint32_t y) {
   return x * (uint64_t)y;
 }
-static simdjson_really_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
+static simdjson_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64_t *hi) {
   uint64_t ad = __emulu((uint32_t)(ab >> 32), (uint32_t)cd);
   uint64_t bd = __emulu((uint32_t)ab, (uint32_t)cd);
   uint64_t adbc = ad + __emulu((uint32_t)ab, (uint32_t)(cd >> 32));
@@ -21670,7 +21075,7 @@ static simdjson_really_inline uint64_t _umul128(uint64_t ab, uint64_t cd, uint64
 
 using internal::value128;
 
-simdjson_really_inline value128 full_multiplication(uint64_t value1, uint64_t value2) {
+simdjson_inline value128 full_multiplication(uint64_t value1, uint64_t value2) {
   value128 answer;
 #if defined(SIMDJSON_REGULAR_VISUAL_STUDIO) || defined(SIMDJSON_IS_32BITS)
 #ifdef _M_ARM64
@@ -21706,13 +21111,13 @@ namespace atomparsing {
 // You might think that using memcpy makes this function expensive, but you'd be wrong.
 // All decent optimizing compilers (GCC, clang, Visual Studio) will compile string_to_uint32("false");
 // to the compile-time constant 1936482662.
-simdjson_really_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
+simdjson_inline uint32_t string_to_uint32(const char* str) { uint32_t val; std::memcpy(&val, str, sizeof(uint32_t)); return val; }
 
 
 // Again in str4ncmp we use a memcpy to avoid undefined behavior. The memcpy may appear expensive.
 // Yet all decent optimizing compilers will compile memcpy to a single instruction, just about.
 simdjson_warn_unused
-simdjson_really_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
+simdjson_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
   uint32_t srcval; // we want to avoid unaligned 32-bit loads (undefined in C/C++)
   static_assert(sizeof(uint32_t) <= SIMDJSON_PADDING, "SIMDJSON_PADDING must be larger than 4 bytes");
   std::memcpy(&srcval, src, sizeof(uint32_t));
@@ -21720,36 +21125,36 @@ simdjson_really_inline uint32_t str4ncmp(const uint8_t *src, const char* atom) {
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_true_atom(const uint8_t *src) {
+simdjson_inline bool is_valid_true_atom(const uint8_t *src) {
   return (str4ncmp(src, "true") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
+simdjson_inline bool is_valid_true_atom(const uint8_t *src, size_t len) {
   if (len > 4) { return is_valid_true_atom(src); }
   else if (len == 4) { return !str4ncmp(src, "true"); }
   else { return false; }
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_false_atom(const uint8_t *src) {
+simdjson_inline bool is_valid_false_atom(const uint8_t *src) {
   return (str4ncmp(src+1, "alse") | jsoncharutils::is_not_structural_or_whitespace(src[5])) == 0;
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
+simdjson_inline bool is_valid_false_atom(const uint8_t *src, size_t len) {
   if (len > 5) { return is_valid_false_atom(src); }
   else if (len == 5) { return !str4ncmp(src+1, "alse"); }
   else { return false; }
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_null_atom(const uint8_t *src) {
+simdjson_inline bool is_valid_null_atom(const uint8_t *src) {
   return (str4ncmp(src, "null") | jsoncharutils::is_not_structural_or_whitespace(src[4])) == 0;
 }
 
 simdjson_warn_unused
-simdjson_really_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
+simdjson_inline bool is_valid_null_atom(const uint8_t *src, size_t len) {
   if (len > 4) { return is_valid_null_atom(src); }
   else if (len == 4) { return !str4ncmp(src, "null"); }
   else { return false; }
@@ -21774,18 +21179,18 @@ using namespace simd;
 struct backslash_and_quote {
 public:
   static constexpr uint32_t BYTES_PROCESSED = 32;
-  simdjson_really_inline static backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
+  simdjson_inline static backslash_and_quote copy_and_find(const uint8_t *src, uint8_t *dst);
 
-  simdjson_really_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
-  simdjson_really_inline bool has_backslash() { return bs_bits != 0; }
-  simdjson_really_inline int quote_index() { return trailing_zeroes(quote_bits); }
-  simdjson_really_inline int backslash_index() { return trailing_zeroes(bs_bits); }
+  simdjson_inline bool has_quote_first() { return ((bs_bits - 1) & quote_bits) != 0; }
+  simdjson_inline bool has_backslash() { return bs_bits != 0; }
+  simdjson_inline int quote_index() { return trailing_zeroes(quote_bits); }
+  simdjson_inline int backslash_index() { return trailing_zeroes(bs_bits); }
 
   uint32_t bs_bits;
   uint32_t quote_bits;
 }; // struct backslash_and_quote
 
-simdjson_really_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
+simdjson_inline backslash_and_quote backslash_and_quote::copy_and_find(const uint8_t *src, uint8_t *dst) {
   // this can read up to 31 bytes beyond the buffer size, but we require
   // SIMDJSON_PADDING of padding
   static_assert(SIMDJSON_PADDING >= (BYTES_PROCESSED - 1), "backslash and quote finder must process fewer than SIMDJSON_PADDING bytes");
@@ -21804,147 +21209,6 @@ simdjson_really_inline backslash_and_quote backslash_and_quote::copy_and_find(co
 } // namespace westmere
 } // namespace simdjson
 
-/* begin file include/simdjson/generic/stringparsing.h */
-// This file contains the common code every implementation uses
-// It is intended to be included multiple times and compiled multiple times
-
-namespace simdjson {
-namespace westmere {
-namespace {
-/// @private
-namespace stringparsing {
-
-// begin copypasta
-// These chars yield themselves: " \ /
-// b -> backspace, f -> formfeed, n -> newline, r -> cr, t -> horizontal tab
-// u not handled in this table as it's complex
-static const uint8_t escape_map[256] = {
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x0.
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0x22, 0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0x2f,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x4.
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0x5c, 0, 0,    0, // 0x5.
-    0, 0, 0x08, 0, 0,    0, 0x0c, 0, 0, 0, 0, 0, 0,    0, 0x0a, 0, // 0x6.
-    0, 0, 0x0d, 0, 0x09, 0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0, // 0x7.
-
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-    0, 0, 0,    0, 0,    0, 0,    0, 0, 0, 0, 0, 0,    0, 0,    0,
-};
-
-// handle a unicode codepoint
-// write appropriate values into dest
-// src will advance 6 bytes or 12 bytes
-// dest will advance a variable amount (return via pointer)
-// return true if the unicode codepoint was valid
-// We work in little-endian then swap at write time
-simdjson_warn_unused
-simdjson_really_inline bool handle_unicode_codepoint(const uint8_t **src_ptr,
-                                            uint8_t **dst_ptr) {
-  // jsoncharutils::hex_to_u32_nocheck fills high 16 bits of the return value with 1s if the
-  // conversion isn't valid; we defer the check for this to inside the
-  // multilingual plane check
-  uint32_t code_point = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
-  *src_ptr += 6;
-  // check for low surrogate for characters outside the Basic
-  // Multilingual Plane.
-  if (code_point >= 0xd800 && code_point < 0xdc00) {
-    if (((*src_ptr)[0] != '\\') || (*src_ptr)[1] != 'u') {
-      return false;
-    }
-    uint32_t code_point_2 = jsoncharutils::hex_to_u32_nocheck(*src_ptr + 2);
-
-    // if the first code point is invalid we will get here, as we will go past
-    // the check for being outside the Basic Multilingual plane. If we don't
-    // find a \u immediately afterwards we fail out anyhow, but if we do,
-    // this check catches both the case of the first code point being invalid
-    // or the second code point being invalid.
-    if ((code_point | code_point_2) >> 16) {
-      return false;
-    }
-
-    code_point =
-        (((code_point - 0xd800) << 10) | (code_point_2 - 0xdc00)) + 0x10000;
-    *src_ptr += 6;
-  }
-  size_t offset = jsoncharutils::codepoint_to_utf8(code_point, *dst_ptr);
-  *dst_ptr += offset;
-  return offset > 0;
-}
-
-/**
- * Unescape a string from src to dst, stopping at a final unescaped quote. E.g., if src points at 'joe"', then
- * dst needs to have four free bytes.
- */
-simdjson_warn_unused simdjson_really_inline uint8_t *parse_string(const uint8_t *src, uint8_t *dst) {
-  while (1) {
-    // Copy the next n bytes, and find the backslash and quote in them.
-    auto bs_quote = backslash_and_quote::copy_and_find(src, dst);
-    // If the next thing is the end quote, copy and return
-    if (bs_quote.has_quote_first()) {
-      // we encountered quotes first. Move dst to point to quotes and exit
-      return dst + bs_quote.quote_index();
-    }
-    if (bs_quote.has_backslash()) {
-      /* find out where the backspace is */
-      auto bs_dist = bs_quote.backslash_index();
-      uint8_t escape_char = src[bs_dist + 1];
-      /* we encountered backslash first. Handle backslash */
-      if (escape_char == 'u') {
-        /* move src/dst up to the start; they will be further adjusted
-           within the unicode codepoint handling code. */
-        src += bs_dist;
-        dst += bs_dist;
-        if (!handle_unicode_codepoint(&src, &dst)) {
-          return nullptr;
-        }
-      } else {
-        /* simple 1:1 conversion. Will eat bs_dist+2 characters in input and
-         * write bs_dist+1 characters to output
-         * note this may reach beyond the part of the buffer we've actually
-         * seen. I think this is ok */
-        uint8_t escape_result = escape_map[escape_char];
-        if (escape_result == 0u) {
-          return nullptr; /* bogus escape value is an error */
-        }
-        dst[bs_dist] = escape_result;
-        src += bs_dist + 2;
-        dst += bs_dist + 1;
-      }
-    } else {
-      /* they are the same. Since they can't co-occur, it means we
-       * encountered neither. */
-      src += backslash_and_quote::BYTES_PROCESSED;
-      dst += backslash_and_quote::BYTES_PROCESSED;
-    }
-  }
-  /* can't be reached */
-  return nullptr;
-}
-
-simdjson_unused simdjson_warn_unused simdjson_really_inline error_code parse_string_to_buffer(const uint8_t *src, uint8_t *&current_string_buf_loc, std::string_view &s) {
-  if (*(src++) != '"') { return STRING_ERROR; }
-  auto end = stringparsing::parse_string(src, current_string_buf_loc);
-  if (!end) { return STRING_ERROR; }
-  s = std::string_view(reinterpret_cast<const char *>(current_string_buf_loc), end-current_string_buf_loc);
-  current_string_buf_loc = end;
-  return SUCCESS;
-}
-
-} // namespace stringparsing
-} // unnamed namespace
-} // namespace westmere
-} // namespace simdjson
-/* end file include/simdjson/generic/stringparsing.h */
-
 #endif // SIMDJSON_WESTMERE_STRINGPARSING_H
 /* end file include/simdjson/westmere/stringparsing.h */
 /* begin file include/simdjson/westmere/numberparsing.h */
@@ -21955,7 +21219,7 @@ namespace simdjson {
 namespace westmere {
 namespace {
 
-static simdjson_really_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
+static simdjson_inline uint32_t parse_eight_digits_unrolled(const uint8_t *chars) {
   // this actually computes *16* values so we are being wasteful.
   const __m128i ascii0 = _mm_set1_epi8('0');
   const __m128i mul_1_10 =
@@ -22018,7 +21282,7 @@ namespace {
 // Convert a mantissa, an exponent and a sign bit into an ieee64 double.
 // The real_exponent needs to be in [0, 2046] (technically real_exponent = 2047 would be acceptable).
 // The mantissa should be in [0,1<<53). The bit at index (1ULL << 52) while be zeroed.
-simdjson_really_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
+simdjson_inline double to_double(uint64_t mantissa, uint64_t real_exponent, bool negative) {
     double d;
     mantissa &= ~(1ULL << 52);
     mantissa |= real_exponent << 52;
@@ -22033,7 +21297,7 @@ simdjson_really_inline double to_double(uint64_t mantissa, uint64_t real_exponen
 // set to false. This should work *most of the time* (like 99% of the time).
 // We assume that power is in the [smallest_power,
 // largest_power] interval: the caller is responsible for this check.
-simdjson_really_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
+simdjson_inline bool compute_float_64(int64_t power, uint64_t i, bool negative, double &d) {
   // we start with a fast path
   // It was described in
   // Clinger WD. How to read floating point numbers accurately.
@@ -22093,7 +21357,7 @@ simdjson_really_inline bool compute_float_64(int64_t power, uint64_t i, bool neg
   // In the slow path, we need to adjust i so that it is > 1<<63 which is always
   // possible, except if i == 0, so we handle i == 0 separately.
   if(i == 0) {
-    d = 0.0;
+    d = negative ? -0.0 : 0.0;
     return true;
   }
 
@@ -22208,7 +21472,7 @@ simdjson_really_inline bool compute_float_64(int64_t power, uint64_t i, bool neg
   if (simdjson_unlikely(real_exponent <= 0)) { // we have a subnormal?
     // Here have that real_exponent <= 0 so -real_exponent >= 0
     if(-real_exponent + 1 >= 64) { // if we have more than 64 bits below the minimum exponent, you have a zero for sure.
-      d = 0.0;
+      d = negative ? -0.0 : 0.0;
       return true;
     }
     // next line is safe because -real_exponent + 1 < 0
@@ -22316,7 +21580,7 @@ static bool parse_float_fallback(const uint8_t *ptr, const uint8_t *end_ptr, dou
 // check quickly whether the next 8 chars are made of digits
 // at a glance, it looks better than Mula's
 // http://0x80.pl/articles/swar-digits-validate.html
-simdjson_really_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
+simdjson_inline bool is_made_of_eight_digits_fast(const uint8_t *chars) {
   uint64_t val;
   // this can read up to 7 bytes beyond the buffer size, but we require
   // SIMDJSON_PADDING of padding
@@ -22343,7 +21607,7 @@ error_code slow_float_parsing(simdjson_unused const uint8_t * src, W writer) {
 
 template<typename I>
 SIMDJSON_NO_SANITIZE_UNDEFINED // We deliberately allow overflow here and check later
-simdjson_really_inline bool parse_digit(const uint8_t c, I &i) {
+simdjson_inline bool parse_digit(const uint8_t c, I &i) {
   const uint8_t digit = static_cast<uint8_t>(c - '0');
   if (digit > 9) {
     return false;
@@ -22353,7 +21617,7 @@ simdjson_really_inline bool parse_digit(const uint8_t c, I &i) {
   return true;
 }
 
-simdjson_really_inline error_code parse_decimal(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
+simdjson_inline error_code parse_decimal(simdjson_unused const uint8_t *const src, const uint8_t *&p, uint64_t &i, int64_t &exponent) {
   // we continue with the fiction that we have an integer. If the
   // floating point number is representable as x * 10^z for some integer
   // z that fits in 53 bits, then we will be able to convert back the
@@ -22381,7 +21645,7 @@ simdjson_really_inline error_code parse_decimal(simdjson_unused const uint8_t *c
   return SUCCESS;
 }
 
-simdjson_really_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
+simdjson_inline error_code parse_exponent(simdjson_unused const uint8_t *const src, const uint8_t *&p, int64_t &exponent) {
   // Exp Sign: -123.456e[-]78
   bool neg_exp = ('-' == *p);
   if (neg_exp || '+' == *p) { p++; } // Skip + as well
@@ -22432,7 +21696,7 @@ simdjson_really_inline error_code parse_exponent(simdjson_unused const uint8_t *
   return SUCCESS;
 }
 
-simdjson_really_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
+simdjson_inline size_t significant_digits(const uint8_t * start_digits, size_t digit_count) {
   // It is possible that the integer had an overflow.
   // We have to handle the case where we have 0.0000somenumber.
   const uint8_t *start = start_digits;
@@ -22442,7 +21706,7 @@ simdjson_really_inline size_t significant_digits(const uint8_t * start_digits, s
 }
 
 template<typename W>
-simdjson_really_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
+simdjson_inline error_code write_float(const uint8_t *const src, bool negative, uint64_t i, const uint8_t * start_digits, size_t digit_count, int64_t exponent, W &writer) {
   // If we frequently had to deal with long strings of digits,
   // we could extend our code by using a 128-bit integer instead
   // of a 64-bit integer. However, this is uncommon in practice.
@@ -22478,7 +21742,8 @@ simdjson_really_inline error_code write_float(const uint8_t *const src, bool neg
     static_assert(simdjson::internal::smallest_power <= -342, "smallest_power is not small enough");
     //
     if((exponent < simdjson::internal::smallest_power) || (i == 0)) {
-      WRITE_DOUBLE(0, src, writer);
+      // E.g. Parse "-0.0e-999" into the same value as "-0.0". See https://en.wikipedia.org/wiki/Signed_zero
+      WRITE_DOUBLE(negative ? -0.0 : 0.0, src, writer);
       return SUCCESS;
     } else { // (exponent > largest_power) and (i != 0)
       // We have, for sure, an infinite value and simdjson refuses to parse infinite values.
@@ -22498,20 +21763,20 @@ simdjson_really_inline error_code write_float(const uint8_t *const src, bool neg
 #ifdef SIMDJSON_SKIPNUMBERPARSING
 
 template<typename W>
-simdjson_really_inline error_code parse_number(const uint8_t *const, W &writer) {
+simdjson_inline error_code parse_number(const uint8_t *const, W &writer) {
   writer.append_s64(0);        // always write zero
   return SUCCESS;              // always succeeds
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
-simdjson_unused simdjson_really_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
-simdjson_unused simdjson_really_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
-simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept { return ondemand::number_type::signed_integer; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * const src) noexcept { return 0; }
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept  { return false; }
+simdjson_unused simdjson_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept { return ondemand::number_type::signed_integer; }
 #else
 
 // parse the number at src
@@ -22524,13 +21789,13 @@ simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> ge
 //
 // Our objective is accurate parsing (ULP of 0) at high speed.
 template<typename W>
-simdjson_really_inline error_code parse_number(const uint8_t *const src, W &writer) {
+simdjson_inline error_code parse_number(const uint8_t *const src, W &writer) {
 
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  const uint8_t *p = src + negative;
+  const uint8_t *p = src + uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -22674,7 +21939,7 @@ const uint8_t integer_string_finisher[256] = {
     NUMBER_ERROR};
 
 // Parse any number from 0 to 18,446,744,073,709,551,615
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src) noexcept {
   const uint8_t *p = src;
   //
   // Parse the integer part.
@@ -22724,7 +21989,7 @@ simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(
 
 // Parse any number from 0 to 18,446,744,073,709,551,615
 // Never read at src_end or beyond
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned(const uint8_t * const src, const uint8_t * const src_end) noexcept {
   const uint8_t *p = src;
   //
   // Parse the integer part.
@@ -22772,7 +22037,7 @@ simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned(
 }
 
 // Parse any number from 0 to 18,446,744,073,709,551,615
-simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<uint64_t> parse_unsigned_in_string(const uint8_t * const src) noexcept {
   const uint8_t *p = src + 1;
   //
   // Parse the integer part.
@@ -22822,12 +22087,12 @@ simdjson_unused simdjson_really_inline simdjson_result<uint64_t> parse_unsigned_
 }
 
 // Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t *src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  const uint8_t *p = src + negative;
+  const uint8_t *p = src + uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -22865,13 +22130,13 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(co
 
 // Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
 // Never read at src_end or beyond
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer(const uint8_t * const src, const uint8_t * const src_end) noexcept {
   //
   // Check for minus sign
   //
   if(src == src_end) { return NUMBER_ERROR; }
   bool negative = (*src == '-');
-  const uint8_t *p = src + negative;
+  const uint8_t *p = src + uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -22908,24 +22173,24 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer(co
 }
 
 // Parse any number from  -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
-simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<int64_t> parse_integer_in_string(const uint8_t *src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*(src + 1) == '-');
-  const uint8_t *p = src + negative + 1;
+  src += uint8_t(negative) + 1;
 
   //
   // Parse the integer part.
   //
   // PERF NOTE: we don't use is_made_of_eight_digits_fast because large integers like 123456789 are rare
-  const uint8_t *const start_digits = p;
+  const uint8_t *const start_digits = src;
   uint64_t i = 0;
-  while (parse_digit(*p, i)) { p++; }
+  while (parse_digit(*src, i)) { src++; }
 
   // If there were no digits, or if the integer starts with 0 and has more than one digit, it's an error.
   // Optimization note: size_t is expected to be unsigned.
-  size_t digit_count = size_t(p - start_digits);
+  size_t digit_count = size_t(src - start_digits);
   // We go from
   // -9,223,372,036,854,775,808 to 9,223,372,036,854,775,807
   // so we can never represent numbers that have more than 19 digits.
@@ -22937,11 +22202,11 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in
   // Here digit_count > 0.
   if (('0' == *start_digits) && (digit_count > 1)) { return NUMBER_ERROR; }
   // We can do the following...
-  // if (!jsoncharutils::is_structural_or_whitespace(*p)) {
-  //  return (*p == '.' || *p == 'e' || *p == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
+  // if (!jsoncharutils::is_structural_or_whitespace(*src)) {
+  //  return (*src == '.' || *src == 'e' || *src == 'E') ? INCORRECT_TYPE : NUMBER_ERROR;
   // }
   // as a single table lookup:
-  if(*p != '"') { return NUMBER_ERROR; }
+  if(*src != '"') { return NUMBER_ERROR; }
   // Negative numbers have can go down to - INT64_MAX - 1 whereas positive numbers are limited to INT64_MAX.
   // Performance note: This check is only needed when digit_count == longest_digit_count but it is
   // so cheap that we might as well always make it.
@@ -22949,12 +22214,12 @@ simdjson_unused simdjson_really_inline simdjson_result<int64_t> parse_integer_in
   return negative ? (~i+1) : i;
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -23021,19 +22286,19 @@ simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(cons
   if (simdjson_likely(!overflow)) {
     if (compute_float_64(exponent, i, negative, d)) { return d; }
   }
-  if (!parse_float_fallback(src-negative, &d)) {
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
     return NUMBER_ERROR;
   }
   return d;
 }
 
-simdjson_unused simdjson_really_inline bool is_negative(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline bool is_negative(const uint8_t * src) noexcept {
   return (*src == '-');
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<bool> is_integer(const uint8_t * src) noexcept {
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
   const uint8_t *p = src;
   while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
   if ( p == src ) { return NUMBER_ERROR; }
@@ -23041,9 +22306,9 @@ simdjson_unused simdjson_really_inline simdjson_result<bool> is_integer(const ui
   return false;
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<ondemand::number_type> get_number_type(const uint8_t * src) noexcept {
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
   const uint8_t *p = src;
   while(static_cast<uint8_t>(*p - '0') <= 9) { p++; }
   if ( p == src ) { return NUMBER_ERROR; }
@@ -23067,13 +22332,13 @@ simdjson_unused simdjson_really_inline simdjson_result<ondemand::number_type> ge
 }
 
 // Never read at src_end or beyond
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
+simdjson_unused simdjson_inline simdjson_result<double> parse_double(const uint8_t * src, const uint8_t * const src_end) noexcept {
   if(src == src_end) { return NUMBER_ERROR; }
   //
   // Check for minus sign
   //
   bool negative = (*src == '-');
-  src += negative;
+  src += uint8_t(negative);
 
   //
   // Parse the integer part.
@@ -23142,18 +22407,18 @@ simdjson_unused simdjson_really_inline simdjson_result<double> parse_double(cons
   if (simdjson_likely(!overflow)) {
     if (compute_float_64(exponent, i, negative, d)) { return d; }
   }
-  if (!parse_float_fallback(src-negative, src_end, &d)) {
+  if (!parse_float_fallback(src - uint8_t(negative), src_end, &d)) {
     return NUMBER_ERROR;
   }
   return d;
 }
 
-simdjson_unused simdjson_really_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
+simdjson_unused simdjson_inline simdjson_result<double> parse_double_in_string(const uint8_t * src) noexcept {
   //
   // Check for minus sign
   //
   bool negative = (*(src + 1) == '-');
-  src += negative + 1;
+  src += uint8_t(negative) + 1;
 
   //
   // Parse the integer part.
@@ -23220,7 +22485,7 @@ simdjson_unused simdjson_really_inline simdjson_result<double> parse_double_in_s
   if (simdjson_likely(!overflow)) {
     if (compute_float_64(exponent, i, negative, d)) { return d; }
   }
-  if (!parse_float_fallback(src-negative, &d)) {
+  if (!parse_float_fallback(src - uint8_t(negative), &d)) {
     return NUMBER_ERROR;
   }
   return d;
@@ -23309,22 +22574,22 @@ struct implementation_simdjson_result_base {
   /**
    * Create a new empty result with error = UNINITIALIZED.
    */
-  simdjson_really_inline implementation_simdjson_result_base() noexcept = default;
+  simdjson_inline implementation_simdjson_result_base() noexcept = default;
 
   /**
    * Create a new error result.
    */
-  simdjson_really_inline implementation_simdjson_result_base(error_code error) noexcept;
+  simdjson_inline implementation_simdjson_result_base(error_code error) noexcept;
 
   /**
    * Create a new successful result.
    */
-  simdjson_really_inline implementation_simdjson_result_base(T &&value) noexcept;
+  simdjson_inline implementation_simdjson_result_base(T &&value) noexcept;
 
   /**
    * Create a new result with both things (use if you don't want to branch when creating the result).
    */
-  simdjson_really_inline implementation_simdjson_result_base(T &&value, error_code error) noexcept;
+  simdjson_inline implementation_simdjson_result_base(T &&value, error_code error) noexcept;
 
   /**
    * Move the value and the error to the provided variables.
@@ -23332,19 +22597,19 @@ struct implementation_simdjson_result_base {
    * @param value The variable to assign the value to. May not be set if there is an error.
    * @param error The variable to assign the error to. Set to SUCCESS if there is no error.
    */
-  simdjson_really_inline void tie(T &value, error_code &error) && noexcept;
+  simdjson_inline void tie(T &value, error_code &error) && noexcept;
 
   /**
    * Move the value to the provided variable.
    *
    * @param value The variable to assign the value to. May not be set if there is an error.
    */
-  simdjson_really_inline error_code get(T &value) && noexcept;
+  simdjson_inline error_code get(T &value) && noexcept;
 
   /**
    * The error.
    */
-  simdjson_really_inline error_code error() const noexcept;
+  simdjson_inline error_code error() const noexcept;
 
 #if SIMDJSON_EXCEPTIONS
 
@@ -23353,28 +22618,28 @@ struct implementation_simdjson_result_base {
    *
    * @throw simdjson_error if there was an error.
    */
-  simdjson_really_inline T& value() & noexcept(false);
+  simdjson_inline T& value() & noexcept(false);
 
   /**
    * Take the result value (move it).
    *
    * @throw simdjson_error if there was an error.
    */
-  simdjson_really_inline T&& value() && noexcept(false);
+  simdjson_inline T&& value() && noexcept(false);
 
   /**
    * Take the result value (move it).
    *
    * @throw simdjson_error if there was an error.
    */
-  simdjson_really_inline T&& take_value() && noexcept(false);
+  simdjson_inline T&& take_value() && noexcept(false);
 
   /**
    * Cast to the value (will throw on error).
    *
    * @throw simdjson_error if there was an error.
    */
-  simdjson_really_inline operator T&&() && noexcept(false);
+  simdjson_inline operator T&&() && noexcept(false);
 
 
 #endif // SIMDJSON_EXCEPTIONS
@@ -23383,17 +22648,17 @@ struct implementation_simdjson_result_base {
    * Get the result value. This function is safe if and only
    * the error() method returns a value that evaluates to false.
    */
-  simdjson_really_inline const T& value_unsafe() const& noexcept;
+  simdjson_inline const T& value_unsafe() const& noexcept;
   /**
    * Get the result value. This function is safe if and only
    * the error() method returns a value that evaluates to false.
    */
-  simdjson_really_inline T& value_unsafe() & noexcept;
+  simdjson_inline T& value_unsafe() & noexcept;
   /**
    * Take the result value (move it). This function is safe if and only
    * the error() method returns a value that evaluates to false.
    */
-  simdjson_really_inline T&& value_unsafe() && noexcept;
+  simdjson_inline T&& value_unsafe() && noexcept;
 protected:
   /** users should never directly access first and second. **/
   T first{}; /** Users should never directly access 'first'. **/
@@ -23457,46 +22722,46 @@ struct number {
    *        unsigned_integer         /// a positive integer larger or equal to 1<<63
    *    };
    */
-  simdjson_really_inline number_type get_number_type() const noexcept;
+  simdjson_inline number_type get_number_type() const noexcept;
   /**
    * return true if the automatically determined type of
    * the number is number_type::unsigned_integer.
    */
-  simdjson_really_inline bool is_uint64() const noexcept;
+  simdjson_inline bool is_uint64() const noexcept;
   /**
    * return the value as a uint64_t, only valid if is_uint64() is true.
    */
-  simdjson_really_inline uint64_t get_uint64() const noexcept;
-  simdjson_really_inline operator uint64_t() const noexcept;
+  simdjson_inline uint64_t get_uint64() const noexcept;
+  simdjson_inline operator uint64_t() const noexcept;
 
   /**
    * return true if the automatically determined type of
    * the number is number_type::signed_integer.
    */
-  simdjson_really_inline bool is_int64() const noexcept;
+  simdjson_inline bool is_int64() const noexcept;
   /**
    * return the value as a int64_t, only valid if is_int64() is true.
    */
-  simdjson_really_inline int64_t get_int64() const noexcept;
-  simdjson_really_inline operator int64_t() const noexcept;
+  simdjson_inline int64_t get_int64() const noexcept;
+  simdjson_inline operator int64_t() const noexcept;
 
 
   /**
    * return true if the automatically determined type of
    * the number is number_type::floating_point_number.
    */
-  simdjson_really_inline bool is_double() const noexcept;
+  simdjson_inline bool is_double() const noexcept;
   /**
    * return the value as a double, only valid if is_double() is true.
    */
-  simdjson_really_inline double get_double() const noexcept;
-  simdjson_really_inline operator double() const noexcept;
+  simdjson_inline double get_double() const noexcept;
+  simdjson_inline operator double() const noexcept;
 
   /**
    * Convert the number to a double. Though it always succeed, the conversion
    * may be lossy if the number cannot be represented exactly.
    */
-  simdjson_really_inline double as_double() const noexcept;
+  simdjson_inline double as_double() const noexcept;
 
 
 protected:
@@ -23513,13 +22778,13 @@ struct number {
   template<typename W>
   friend error_code numberparsing::slow_float_parsing(simdjson_unused const uint8_t * src, W writer);
   /** Store a signed 64-bit value to the number. */
-  simdjson_really_inline void append_s64(int64_t value) noexcept;
+  simdjson_inline void append_s64(int64_t value) noexcept;
   /** Store an unsigned 64-bit value to the number. */
-  simdjson_really_inline void append_u64(uint64_t value) noexcept;
+  simdjson_inline void append_u64(uint64_t value) noexcept;
   /** Store a double value to the number. */
-  simdjson_really_inline void append_double(double value) noexcept;
+  simdjson_inline void append_double(double value) noexcept;
   /** Specifies that the value is a double, but leave it undefined. */
-  simdjson_really_inline void skip_double() noexcept;
+  simdjson_inline void skip_double() noexcept;
   /**
    * End of friend declarations.
    */
@@ -23567,10 +22832,10 @@ namespace simdjson {
 template<>
 struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type> : public SIMDJSON_BUILTIN_IMPLEMENTATION::implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type> {
 public:
-  simdjson_really_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type &&value) noexcept; ///< @private
-  simdjson_really_inline simdjson_result(error_code error) noexcept; ///< @private
-  simdjson_really_inline simdjson_result() noexcept = default;
-  simdjson_really_inline ~simdjson_result() noexcept = default; ///< @private
+  simdjson_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type &&value) noexcept; ///< @private
+  simdjson_inline simdjson_result(error_code error) noexcept; ///< @private
+  simdjson_inline simdjson_result() noexcept = default;
+  simdjson_inline ~simdjson_result() noexcept = default; ///< @private
 };
 
 } // namespace simdjson
@@ -23648,12 +22913,19 @@ class json_iterator;
  *
  * This class is deliberately simplistic and has little functionality. You can
  * compare a raw_json_string instance with an unescaped C string, but
- * that is pretty much all you can do.
+ * that is nearly all you can do.
+ *
+ * The raw_json_string is unescaped. If you wish to write an unescaped version of it to your own
+ * buffer, you may do so using the parser.unescape(string, buff) method, using an ondemand::parser
+ * instance. Doing so requires you to have a sufficiently large buffer.
+ *
+ * The raw_json_string instances originate typically from field instance which in turn represent
+ * key-value pairs from object instances. From a field instance, you get the raw_json_string
+ * instance by calling key(). You can, if you want a more usable string_view instance, call
+ * the unescaped_key() method on the field instance. You may also create a raw_json_string from
+ * any other string value, with the value.get_raw_json_string() method. Again, you can get
+ * a more usable string_view instance by calling get_string().
  *
- * They originate typically from field instance which in turn represent key-value pairs from
- * object instances. From a field instance, you get the raw_json_string instance by calling key().
- * You can, if you want a more usable string_view instance, call the unescaped_key() method
- * on the field instance.
  */
 class raw_json_string {
 public:
@@ -23662,7 +22934,7 @@ class raw_json_string {
    *
    * Exists so you can declare a variable and later assign to it before use.
    */
-  simdjson_really_inline raw_json_string() noexcept = default;
+  simdjson_inline raw_json_string() noexcept = default;
 
   /**
    * Create a new invalid raw_json_string pointed at the given location in the JSON.
@@ -23671,14 +22943,14 @@ class raw_json_string {
    *
    * It *must* be terminated by a ", and be a valid JSON string.
    */
-  simdjson_really_inline raw_json_string(const uint8_t * _buf) noexcept;
+  simdjson_inline raw_json_string(const uint8_t * _buf) noexcept;
   /**
    * Get the raw pointer to the beginning of the string in the JSON (just after the ").
    *
    * It is possible for this function to return a null pointer if the instance
    * has outlived its existence.
    */
-  simdjson_really_inline const char * raw() const noexcept;
+  simdjson_inline const char * raw() const noexcept;
 
   /**
    * This compares the current instance to the std::string_view target: returns true if
@@ -23694,7 +22966,7 @@ class raw_json_string {
    * Performance: the comparison may be done using memcmp which may be efficient
    * for long strings.
    */
-  simdjson_really_inline bool unsafe_is_equal(size_t length, std::string_view target) const noexcept;
+  simdjson_inline bool unsafe_is_equal(size_t length, std::string_view target) const noexcept;
 
   /**
    * This compares the current instance to the std::string_view target: returns true if
@@ -23711,7 +22983,7 @@ class raw_json_string {
    *   static_assert(raw_json_string::is_free_from_unescaped_quote(target), "");
    *   s.unsafe_is_equal(target);
    */
-  simdjson_really_inline bool unsafe_is_equal(std::string_view target) const noexcept;
+  simdjson_inline bool unsafe_is_equal(std::string_view target) const noexcept;
 
   /**
    * This compares the current instance to the C string target: returns true if
@@ -23725,27 +22997,27 @@ class raw_json_string {
    *   static_assert(raw_json_string::is_free_from_unescaped_quote(target), "");
    *   s.unsafe_is_equal(target);
    */
-  simdjson_really_inline bool unsafe_is_equal(const char* target) const noexcept;
+  simdjson_inline bool unsafe_is_equal(const char* target) const noexcept;
 
   /**
    * This compares the current instance to the std::string_view target: returns true if
    * they are byte-by-byte equal (no escaping is done).
    */
-  simdjson_really_inline bool is_equal(std::string_view target) const noexcept;
+  simdjson_inline bool is_equal(std::string_view target) const noexcept;
 
   /**
    * This compares the current instance to the C string target: returns true if
    * they are byte-by-byte equal (no escaping is done).
    */
-  simdjson_really_inline bool is_equal(const char* target) const noexcept;
+  simdjson_inline bool is_equal(const char* target) const noexcept;
 
   /**
    * Returns true if target is free from unescaped quote. If target is known at
    * compile-time, we might expect the computation to happen at compile time with
    * many compilers (not all!).
    */
-  static simdjson_really_inline bool is_free_from_unescaped_quote(std::string_view target) noexcept;
-  static simdjson_really_inline bool is_free_from_unescaped_quote(const char* target) noexcept;
+  static simdjson_inline bool is_free_from_unescaped_quote(std::string_view target) noexcept;
+  static simdjson_inline bool is_free_from_unescaped_quote(const char* target) noexcept;
 
 private:
 
@@ -23754,27 +23026,13 @@ class raw_json_string {
    * This will set the inner pointer to zero, effectively making
    * this instance unusable.
    */
-  simdjson_really_inline void consume() noexcept { buf = nullptr; }
+  simdjson_inline void consume() noexcept { buf = nullptr; }
 
   /**
    * Checks whether the inner pointer is non-null and thus usable.
    */
-  simdjson_really_inline simdjson_warn_unused bool alive() const noexcept { return buf != nullptr; }
+  simdjson_inline simdjson_warn_unused bool alive() const noexcept { return buf != nullptr; }
 
-  /**
-   * Unescape this JSON string, replacing \\ with \, \n with newline, etc.
-   *
-   * ## IMPORTANT: string_view lifetime
-   *
-   * The string_view is only valid as long as the bytes in dst.
-   *
-   * @param dst A pointer to a buffer at least large enough to write this string as well as a \0.
-   *            dst will be updated to the next unused location (just after the \0 written out at
-   *            the end of this string).
-   * @return A string_view pointing at the unescaped string in dst
-   * @error STRING_ERROR if escapes are incorrect.
-   */
-  simdjson_really_inline simdjson_warn_unused simdjson_result<std::string_view> unescape(uint8_t *&dst) const noexcept;
   /**
    * Unescape this JSON string, replacing \\ with \, \n with newline, etc.
    *
@@ -23784,24 +23042,25 @@ class raw_json_string {
    *
    * @param iter A json_iterator, which contains a buffer where the string will be written.
    */
-  simdjson_really_inline simdjson_warn_unused simdjson_result<std::string_view> unescape(json_iterator &iter) const noexcept;
+  simdjson_inline simdjson_warn_unused simdjson_result<std::string_view> unescape(json_iterator &iter) const noexcept;
 
   const uint8_t * buf{};
   friend class object;
   friend class field;
+  friend class parser;
   friend struct simdjson_result<raw_json_string>;
 };
 
-simdjson_unused simdjson_really_inline std::ostream &operator<<(std::ostream &, const raw_json_string &) noexcept;
+simdjson_unused simdjson_inline std::ostream &operator<<(std::ostream &, const raw_json_string &) noexcept;
 
 /**
  * Comparisons between raw_json_string and std::string_view instances are potentially unsafe: the user is responsible
  * for providing a string with no unescaped quote. Note that unescaped quotes cannot be present in valid JSON strings.
  */
-simdjson_unused simdjson_really_inline bool operator==(const raw_json_string &a, std::string_view c) noexcept;
-simdjson_unused simdjson_really_inline bool operator==(std::string_view c, const raw_json_string &a) noexcept;
-simdjson_unused simdjson_really_inline bool operator!=(const raw_json_string &a, std::string_view c) noexcept;
-simdjson_unused simdjson_really_inline bool operator!=(std::string_view c, const raw_json_string &a) noexcept;
+simdjson_unused simdjson_inline bool operator==(const raw_json_string &a, std::string_view c) noexcept;
+simdjson_unused simdjson_inline bool operator==(std::string_view c, const raw_json_string &a) noexcept;
+simdjson_unused simdjson_inline bool operator!=(const raw_json_string &a, std::string_view c) noexcept;
+simdjson_unused simdjson_inline bool operator!=(std::string_view c, const raw_json_string &a) noexcept;
 
 
 } // namespace ondemand
@@ -23813,14 +23072,13 @@ namespace simdjson {
 template<>
 struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string> : public SIMDJSON_BUILTIN_IMPLEMENTATION::implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string> {
 public:
-  simdjson_really_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string &&value) noexcept; ///< @private
-  simdjson_really_inline simdjson_result(error_code error) noexcept; ///< @private
-  simdjson_really_inline simdjson_result() noexcept = default;
-  simdjson_really_inline ~simdjson_result() noexcept = default; ///< @private
-
-  simdjson_really_inline simdjson_result<const char *> raw() const noexcept;
-  simdjson_really_inline simdjson_warn_unused simdjson_result<std::string_view> unescape(uint8_t *&dst) const noexcept;
-  simdjson_really_inline simdjson_warn_unused simdjson_result<std::string_view> unescape(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_iterator &iter) const noexcept;
+  simdjson_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string &&value) noexcept; ///< @private
+  simdjson_inline simdjson_result(error_code error) noexcept; ///< @private
+  simdjson_inline simdjson_result() noexcept = default;
+  simdjson_inline ~simdjson_result() noexcept = default; ///< @private
+
+  simdjson_inline simdjson_result<const char *> raw() const noexcept;
+  simdjson_inline simdjson_warn_unused simdjson_result<std::string_view> unescape(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_iterator &iter) const noexcept;
 };
 
 } // namespace simdjson
@@ -23843,20 +23101,20 @@ class token_iterator {
    *
    * Exists so you can declare a variable and later assign to it before use.
    */
-  simdjson_really_inline token_iterator() noexcept = default;
-  simdjson_really_inline token_iterator(token_iterator &&other) noexcept = default;
-  simdjson_really_inline token_iterator &operator=(token_iterator &&other) noexcept = default;
-  simdjson_really_inline token_iterator(const token_iterator &other) noexcept = default;
-  simdjson_really_inline token_iterator &operator=(const token_iterator &other) noexcept = default;
+  simdjson_inline token_iterator() noexcept = default;
+  simdjson_inline token_iterator(token_iterator &&other) noexcept = default;
+  simdjson_inline token_iterator &operator=(token_iterator &&other) noexcept = default;
+  simdjson_inline token_iterator(const token_iterator &other) noexcept = default;
+  simdjson_inline token_iterator &operator=(const token_iterator &other) noexcept = default;
 
   /**
    * Advance to the next token (returning the current one).
    */
-  simdjson_really_inline const uint8_t *return_current_and_advance() noexcept;
+  simdjson_inline const uint8_t *return_current_and_advance() noexcept;
   /**
    * Reports the current offset in bytes from the start of the underlying buffer.
    */
-  simdjson_really_inline uint32_t current_offset() const noexcept;
+  simdjson_inline uint32_t current_offset() const noexcept;
   /**
    * Get the JSON text for a given token (relative).
    *
@@ -23868,7 +23126,7 @@ class token_iterator {
    * TODO consider a string_view, assuming the length will get stripped out by the optimizer when
    * it isn't used ...
    */
-  simdjson_really_inline const uint8_t *peek(int32_t delta=0) const noexcept;
+  simdjson_inline const uint8_t *peek(int32_t delta=0) const noexcept;
   /**
    * Get the maximum length of the JSON text for a given token.
    *
@@ -23877,7 +23135,7 @@ class token_iterator {
    * @param delta The relative position of the token to retrieve. e.g. 0 = current token,
    *              1 = next token, -1 = prev token.
    */
-  simdjson_really_inline uint32_t peek_length(int32_t delta=0) const noexcept;
+  simdjson_inline uint32_t peek_length(int32_t delta=0) const noexcept;
 
   /**
    * Get the JSON text for a given token.
@@ -23887,7 +23145,7 @@ class token_iterator {
    * @param position The position of the token.
    *
    */
-  simdjson_really_inline const uint8_t *peek(token_position position) const noexcept;
+  simdjson_inline const uint8_t *peek(token_position position) const noexcept;
   /**
    * Get the maximum length of the JSON text for a given token.
    *
@@ -23895,29 +23153,29 @@ class token_iterator {
    *
    * @param position The position of the token.
    */
-  simdjson_really_inline uint32_t peek_length(token_position position) const noexcept;
+  simdjson_inline uint32_t peek_length(token_position position) const noexcept;
 
   /**
    * Return the current index.
    */
-  simdjson_really_inline token_position position() const noexcept;
+  simdjson_inline token_position position() const noexcept;
   /**
    * Reset to a previously saved index.
    */
-  simdjson_really_inline void set_position(token_position target_position) noexcept;
+  simdjson_inline void set_position(token_position target_position) noexcept;
 
   // NOTE: we don't support a full C++ iterator interface, because we expect people to make
   // different calls to advance the iterator based on *their own* state.
 
-  simdjson_really_inline bool operator==(const token_iterator &other) const noexcept;
-  simdjson_really_inline bool operator!=(const token_iterator &other) const noexcept;
-  simdjson_really_inline bool operator>(const token_iterator &other) const noexcept;
-  simdjson_really_inline bool operator>=(const token_iterator &other) const noexcept;
-  simdjson_really_inline bool operator<(const token_iterator &other) const noexcept;
-  simdjson_really_inline bool operator<=(const token_iterator &other) const noexcept;
+  simdjson_inline bool operator==(const token_iterator &other) const noexcept;
+  simdjson_inline bool operator!=(const token_iterator &other) const noexcept;
+  simdjson_inline bool operator>(const token_iterator &other) const noexcept;
+  simdjson_inline bool operator>=(const token_iterator &other) const noexcept;
+  simdjson_inline bool operator<(const token_iterator &other) const noexcept;
+  simdjson_inline bool operator<=(const token_iterator &other) const noexcept;
 
 protected:
-  simdjson_really_inline token_iterator(const uint8_t *buf, token_position position) noexcept;
+  simdjson_inline token_iterator(const uint8_t *buf, token_position position) noexcept;
 
   /**
    * Get the index of the JSON text for a given token (relative).
@@ -23927,7 +23185,7 @@ class token_iterator {
    * @param delta The relative position of the token to retrieve. e.g. 0 = current token,
    *              1 = next token, -1 = prev token.
    */
-  simdjson_really_inline uint32_t peek_index(int32_t delta=0) const noexcept;
+  simdjson_inline uint32_t peek_index(int32_t delta=0) const noexcept;
   /**
    * Get the index of the JSON text for a given token.
    *
@@ -23936,7 +23194,7 @@ class token_iterator {
    * @param position The position of the token.
    *
    */
-  simdjson_really_inline uint32_t peek_index(token_position position) const noexcept;
+  simdjson_inline uint32_t peek_index(token_position position) const noexcept;
 
   const uint8_t *buf{};
   token_position _position{};
@@ -23944,8 +23202,8 @@ class token_iterator {
   friend class json_iterator;
   friend class value_iterator;
   friend class object;
-  friend simdjson_really_inline void logger::log_line(const json_iterator &iter, const char *title_prefix, const char *title, std::string_view detail, int delta, int depth_delta) noexcept;
-  friend simdjson_really_inline void logger::log_line(const json_iterator &iter, token_position index, depth_t depth, const char *title_prefix, const char *title, std::string_view detail) noexcept;
+  friend simdjson_inline void logger::log_line(const json_iterator &iter, const char *title_prefix, const char *title, std::string_view detail, int delta, int depth_delta) noexcept;
+  friend simdjson_inline void logger::log_line(const json_iterator &iter, token_position index, depth_t depth, const char *title_prefix, const char *title, std::string_view detail) noexcept;
 };
 
 } // namespace ondemand
@@ -23957,10 +23215,10 @@ namespace simdjson {
 template<>
 struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::token_iterator> : public SIMDJSON_BUILTIN_IMPLEMENTATION::implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::token_iterator> {
 public:
-  simdjson_really_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::token_iterator &&value) noexcept; ///< @private
-  simdjson_really_inline simdjson_result(error_code error) noexcept; ///< @private
-  simdjson_really_inline simdjson_result() noexcept = default;
-  simdjson_really_inline ~simdjson_result() noexcept = default; ///< @private
+  simdjson_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::token_iterator &&value) noexcept; ///< @private
+  simdjson_inline simdjson_result(error_code error) noexcept; ///< @private
+  simdjson_inline simdjson_result() noexcept = default;
+  simdjson_inline ~simdjson_result() noexcept = default; ///< @private
 };
 
 } // namespace simdjson
@@ -24029,20 +23287,20 @@ class json_iterator {
   bool _streaming{false};
 
 public:
-  simdjson_really_inline json_iterator() noexcept = default;
-  simdjson_really_inline json_iterator(json_iterator &&other) noexcept;
-  simdjson_really_inline json_iterator &operator=(json_iterator &&other) noexcept;
-  simdjson_really_inline explicit json_iterator(const json_iterator &other) noexcept = default;
-  simdjson_really_inline json_iterator &operator=(const json_iterator &other) noexcept = default;
+  simdjson_inline json_iterator() noexcept = default;
+  simdjson_inline json_iterator(json_iterator &&other) noexcept;
+  simdjson_inline json_iterator &operator=(json_iterator &&other) noexcept;
+  simdjson_inline explicit json_iterator(const json_iterator &other) noexcept = default;
+  simdjson_inline json_iterator &operator=(const json_iterator &other) noexcept = default;
   /**
    * Skips a JSON value, whether it is a scalar, array or object.
    */
-  simdjson_warn_unused simdjson_really_inline error_code skip_child(depth_t parent_depth) noexcept;
+  simdjson_warn_unused simdjson_inline error_code skip_child(depth_t parent_depth) noexcept;
 
   /**
    * Tell whether the iterator is still at the start
    */
-  simdjson_really_inline bool at_root() const noexcept;
+  simdjson_inline bool at_root() const noexcept;
 
   /**
    * Tell whether we should be expected to run in streaming
@@ -24050,53 +23308,61 @@ class json_iterator {
    * that does not affect how the iterator works. It is used by
    * start_root_array() and start_root_object().
    */
-  simdjson_really_inline bool streaming() const noexcept;
+  simdjson_inline bool streaming() const noexcept;
 
   /**
    * Get the root value iterator
    */
-  simdjson_really_inline token_position root_position() const noexcept;
+  simdjson_inline token_position root_position() const noexcept;
   /**
    * Assert that we are at the document depth (== 1)
    */
-  simdjson_really_inline void assert_at_document_depth() const noexcept;
+  simdjson_inline void assert_at_document_depth() const noexcept;
   /**
    * Assert that we are at the root of the document
    */
-  simdjson_really_inline void assert_at_root() const noexcept;
+  simdjson_inline void assert_at_root() const noexcept;
 
   /**
    * Tell whether the iterator is at the EOF mark
    */
-  simdjson_really_inline bool at_end() const noexcept;
+  simdjson_inline bool at_end() const noexcept;
 
   /**
    * Tell whether the iterator is live (has not been moved).
    */
-  simdjson_really_inline bool is_alive() const noexcept;
+  simdjson_inline bool is_alive() const noexcept;
 
   /**
    * Abandon this iterator, setting depth to 0 (as if the document is finished).
    */
-  simdjson_really_inline void abandon() noexcept;
+  simdjson_inline void abandon() noexcept;
 
   /**
    * Advance the current token without modifying depth.
    */
-  simdjson_really_inline const uint8_t *return_current_and_advance() noexcept;
+  simdjson_inline const uint8_t *return_current_and_advance() noexcept;
+
+  /**
+   * Returns true if there is a single token in the index (i.e., it is
+   * a JSON with a scalar value such as a single number).
+   *
+   * @return whether there is a single token
+   */
+  simdjson_inline bool is_single_token() const noexcept;
 
   /**
    * Assert that there are at least the given number of tokens left.
    *
    * Has no effect in release builds.
    */
-  simdjson_really_inline void assert_more_tokens(uint32_t required_tokens=1) const noexcept;
+  simdjson_inline void assert_more_tokens(uint32_t required_tokens=1) const noexcept;
   /**
    * Assert that the given position addresses an actual token (is within bounds).
    *
    * Has no effect in release builds.
    */
-  simdjson_really_inline void assert_valid_position(token_position position) const noexcept;
+  simdjson_inline void assert_valid_position(token_position position) const noexcept;
   /**
    * Get the JSON text for a given token (relative).
    *
@@ -24107,7 +23373,7 @@ class json_iterator {
    * TODO consider a string_view, assuming the length will get stripped out by the optimizer when
    * it isn't used ...
    */
-  simdjson_really_inline const uint8_t *peek(int32_t delta=0) const noexcept;
+  simdjson_inline const uint8_t *peek(int32_t delta=0) const noexcept;
   /**
    * Get the maximum length of the JSON text for the current token (or relative).
    *
@@ -24115,7 +23381,7 @@ class json_iterator {
    *
    * @param delta The relative position of the token to retrieve. e.g. 0 = next token, -1 = prev token.
    */
-  simdjson_really_inline uint32_t peek_length(int32_t delta=0) const noexcept;
+  simdjson_inline uint32_t peek_length(int32_t delta=0) const noexcept;
   /**
    * Get a pointer to the current location in the input buffer.
    *
@@ -24124,7 +23390,7 @@ class json_iterator {
    * You may be pointing outside of the input buffer: it is not generally
    * safe to dereference this pointer.
    */
-  simdjson_really_inline const uint8_t *unsafe_pointer() const noexcept;
+  simdjson_inline const uint8_t *unsafe_pointer() const noexcept;
   /**
    * Get the JSON text for a given token.
    *
@@ -24135,7 +23401,7 @@ class json_iterator {
    * TODO consider a string_view, assuming the length will get stripped out by the optimizer when
    * it isn't used ...
    */
-  simdjson_really_inline const uint8_t *peek(token_position position) const noexcept;
+  simdjson_inline const uint8_t *peek(token_position position) const noexcept;
   /**
    * Get the maximum length of the JSON text for the current token (or relative).
    *
@@ -24143,7 +23409,7 @@ class json_iterator {
    *
    * @param position The position of the token to retrieve.
    */
-  simdjson_really_inline uint32_t peek_length(token_position position) const noexcept;
+  simdjson_inline uint32_t peek_length(token_position position) const noexcept;
   /**
    * Get the JSON text for the last token in the document.
    *
@@ -24152,7 +23418,7 @@ class json_iterator {
    * TODO consider a string_view, assuming the length will get stripped out by the optimizer when
    * it isn't used ...
    */
-  simdjson_really_inline const uint8_t *peek_last() const noexcept;
+  simdjson_inline const uint8_t *peek_last() const noexcept;
 
   /**
    * Ascend one level.
@@ -24161,7 +23427,7 @@ class json_iterator {
    *
    * @param parent_depth the expected parent depth.
    */
-  simdjson_really_inline void ascend_to(depth_t parent_depth) noexcept;
+  simdjson_inline void ascend_to(depth_t parent_depth) noexcept;
 
   /**
    * Descend one level.
@@ -24170,18 +23436,18 @@ class json_iterator {
    *
    * @param child_depth the expected child depth.
    */
-  simdjson_really_inline void descend_to(depth_t child_depth) noexcept;
-  simdjson_really_inline void descend_to(depth_t child_depth, int32_t delta) noexcept;
+  simdjson_inline void descend_to(depth_t child_depth) noexcept;
+  simdjson_inline void descend_to(depth_t child_depth, int32_t delta) noexcept;
 
   /**
    * Get current depth.
    */
-  simdjson_really_inline depth_t depth() const noexcept;
+  simdjson_inline depth_t depth() const noexcept;
 
   /**
    * Get current (writeable) location in the string buffer.
    */
-  simdjson_really_inline uint8_t *&string_buf_loc() noexcept;
+  simdjson_inline uint8_t *&string_buf_loc() noexcept;
 
   /**
    * Report an unrecoverable error, preventing further iteration.
@@ -24189,23 +23455,31 @@ class json_iterator {
    * @param error The error to report. Must not be SUCCESS, UNINITIALIZED, INCORRECT_TYPE, or NO_SUCH_FIELD.
    * @param message An error message to report with the error.
    */
-  simdjson_really_inline error_code report_error(error_code error, const char *message) noexcept;
+  simdjson_inline error_code report_error(error_code error, const char *message) noexcept;
 
   /**
    * Log error, but don't stop iteration.
    * @param error The error to report. Must be INCORRECT_TYPE, or NO_SUCH_FIELD.
    * @param message An error message to report with the error.
    */
-  simdjson_really_inline error_code optional_error(error_code error, const char *message) noexcept;
+  simdjson_inline error_code optional_error(error_code error, const char *message) noexcept;
 
-  template<int N> simdjson_warn_unused simdjson_really_inline bool copy_to_buffer(const uint8_t *json, uint32_t max_len, uint8_t (&tmpbuf)[N]) noexcept;
+  template<int N> simdjson_warn_unused simdjson_inline bool copy_to_buffer(const uint8_t *json, uint32_t max_len, uint8_t (&tmpbuf)[N]) noexcept;
 
-  simdjson_really_inline token_position position() const noexcept;
-  simdjson_really_inline void reenter_child(token_position position, depth_t child_depth) noexcept;
-#ifdef SIMDJSON_DEVELOPMENT_CHECKS
-  simdjson_really_inline token_position start_position(depth_t depth) const noexcept;
-  simdjson_really_inline void set_start_position(depth_t depth, token_position position) noexcept;
+  simdjson_inline token_position position() const noexcept;
+  /**
+   * Write the raw_json_string to the string buffer and return a string_view.
+   * Each raw_json_string should be unescaped once, or else the string buffer might
+   * overflow.
+   */
+  simdjson_inline simdjson_result<std::string_view> unescape(raw_json_string in) noexcept;
+  simdjson_inline void reenter_child(token_position position, depth_t child_depth) noexcept;
+
+#if SIMDJSON_DEVELOPMENT_CHECKS
+  simdjson_inline token_position start_position(depth_t depth) const noexcept;
+  simdjson_inline void set_start_position(depth_t depth, token_position position) noexcept;
 #endif
+
   /* Useful for debugging and logging purposes. */
   inline std::string to_string() const noexcept;
 
@@ -24219,14 +23493,21 @@ class json_iterator {
    * as if it had just been created.
    */
   inline void rewind() noexcept;
+  /**
+   * This checks whether the {,},[,] are balanced so that the document
+   * ends with proper zero depth. This requires scanning the whole document
+   * and it may be expensive. It is expected that it will be rarely called.
+   * It does not attempt to match { with } and [ with ].
+   */
+  inline bool balanced() const noexcept;
 protected:
-  simdjson_really_inline json_iterator(const uint8_t *buf, ondemand::parser *parser) noexcept;
+  simdjson_inline json_iterator(const uint8_t *buf, ondemand::parser *parser) noexcept;
   /// The last token before the end
-  simdjson_really_inline token_position last_position() const noexcept;
+  simdjson_inline token_position last_position() const noexcept;
   /// The token *at* the end. This points at gibberish and should only be used for comparison.
-  simdjson_really_inline token_position end_position() const noexcept;
+  simdjson_inline token_position end_position() const noexcept;
   /// The end of the buffer.
-  simdjson_really_inline token_position end() const noexcept;
+  simdjson_inline token_position end() const noexcept;
 
   friend class document;
   friend class document_stream;
@@ -24236,8 +23517,8 @@ class json_iterator {
   friend class raw_json_string;
   friend class parser;
   friend class value_iterator;
-  friend simdjson_really_inline void logger::log_line(const json_iterator &iter, const char *title_prefix, const char *title, std::string_view detail, int delta, int depth_delta) noexcept;
-  friend simdjson_really_inline void logger::log_line(const json_iterator &iter, token_position index, depth_t depth, const char *title_prefix, const char *title, std::string_view detail) noexcept;
+  friend simdjson_inline void logger::log_line(const json_iterator &iter, const char *title_prefix, const char *title, std::string_view detail, int delta, int depth_delta) noexcept;
+  friend simdjson_inline void logger::log_line(const json_iterator &iter, token_position index, depth_t depth, const char *title_prefix, const char *title, std::string_view detail) noexcept;
 }; // json_iterator
 
 } // namespace ondemand
@@ -24249,10 +23530,10 @@ namespace simdjson {
 template<>
 struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_iterator> : public SIMDJSON_BUILTIN_IMPLEMENTATION::implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_iterator> {
 public:
-  simdjson_really_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_iterator &&value) noexcept; ///< @private
-  simdjson_really_inline simdjson_result(error_code error) noexcept; ///< @private
+  simdjson_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_iterator &&value) noexcept; ///< @private
+  simdjson_inline simdjson_result(error_code error) noexcept; ///< @private
 
-  simdjson_really_inline simdjson_result() noexcept = default;
+  simdjson_inline simdjson_result() noexcept = default;
 };
 
 } // namespace simdjson
@@ -24289,61 +23570,61 @@ class value_iterator {
   token_position _start_position{};
 
 public:
-  simdjson_really_inline value_iterator() noexcept = default;
+  simdjson_inline value_iterator() noexcept = default;
 
   /**
    * Denote that we're starting a document.
    */
-  simdjson_really_inline void start_document() noexcept;
+  simdjson_inline void start_document() noexcept;
 
   /**
    * Skips a non-iterated or partially-iterated JSON value, whether it is a scalar, array or object.
    *
    * Optimized for scalars.
    */
-  simdjson_warn_unused simdjson_really_inline error_code skip_child() noexcept;
+  simdjson_warn_unused simdjson_inline error_code skip_child() noexcept;
 
   /**
    * Tell whether the iterator is at the EOF mark
    */
-  simdjson_really_inline bool at_end() const noexcept;
+  simdjson_inline bool at_end() const noexcept;
 
   /**
    * Tell whether the iterator is at the start of the value
    */
-  simdjson_really_inline bool at_start() const noexcept;
+  simdjson_inline bool at_start() const noexcept;
 
   /**
    * Tell whether the value is open--if the value has not been used, or the array/object is still open.
    */
-  simdjson_really_inline bool is_open() const noexcept;
+  simdjson_inline bool is_open() const noexcept;
 
   /**
    * Tell whether the value is at an object's first field (just after the {).
    */
-  simdjson_really_inline bool at_first_field() const noexcept;
+  simdjson_inline bool at_first_field() const noexcept;
 
   /**
    * Abandon all iteration.
    */
-  simdjson_really_inline void abandon() noexcept;
+  simdjson_inline void abandon() noexcept;
 
   /**
    * Get the child value as a value_iterator.
    */
-  simdjson_really_inline value_iterator child_value() const noexcept;
+  simdjson_inline value_iterator child_value() const noexcept;
 
   /**
    * Get the depth of this value.
    */
-  simdjson_really_inline int32_t depth() const noexcept;
+  simdjson_inline int32_t depth() const noexcept;
 
   /**
    * Get the JSON type of this value.
    *
    * @error TAPE_ERROR when the JSON value is a bad token like "}" "," or "alse".
    */
-  simdjson_really_inline simdjson_result<json_type> type() const noexcept;
+  simdjson_inline simdjson_result<json_type> type() const noexcept;
 
   /**
    * @addtogroup object Object iteration
@@ -24360,7 +23641,7 @@ class value_iterator {
    * @returns Whether the object had any fields (returns false for empty).
    * @error INCORRECT_TYPE if there is no opening {
    */
-  simdjson_warn_unused simdjson_really_inline simdjson_result<bool> start_object() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<bool> start_object() noexcept;
   /**
    * Start an object iteration from the root.
    *
@@ -24368,7 +23649,7 @@ class value_iterator {
    * @error INCORRECT_TYPE if there is no opening {
    * @error TAPE_ERROR if there is no matching } at end of document
    */
-  simdjson_warn_unused simdjson_really_inline simdjson_result<bool> start_root_object() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<bool> start_root_object() noexcept;
 
   /**
    * Start an object iteration after the user has already checked and moved past the {.
@@ -24379,7 +23660,7 @@ class value_iterator {
    * @error INCOMPLETE_ARRAY_OR_OBJECT If there are no more tokens (implying the *parent*
    *        array or object is incomplete).
    */
-  simdjson_warn_unused simdjson_really_inline simdjson_result<bool> started_object() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<bool> started_object() noexcept;
   /**
    * Start an object iteration from the root, after the user has already checked and moved past the {.
    *
@@ -24389,7 +23670,7 @@ class value_iterator {
    * @error INCOMPLETE_ARRAY_OR_OBJECT If there are no more tokens (implying the *parent*
    *        array or object is incomplete).
    */
-  simdjson_warn_unused simdjson_really_inline simdjson_result<bool> started_root_object() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<bool> started_root_object() noexcept;
 
   /**
    * Moves to the next field in an object.
@@ -24401,17 +23682,17 @@ class value_iterator {
    * @error TAPE_ERROR If there is a comma missing between fields.
    * @error TAPE_ERROR If there is a comma, but not enough tokens remaining to have a key, :, and value.
    */
-  simdjson_warn_unused simdjson_really_inline simdjson_result<bool> has_next_field() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<bool> has_next_field() noexcept;
 
   /**
    * Get the current field's key.
    */
-  simdjson_warn_unused simdjson_really_inline simdjson_result<raw_json_string> field_key() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<raw_json_string> field_key() noexcept;
 
   /**
    * Pass the : in the field and move to its value.
    */
-  simdjson_warn_unused simdjson_really_inline error_code field_value() noexcept;
+  simdjson_warn_unused simdjson_inline error_code field_value() noexcept;
 
   /**
    * Find the next field with the given key.
@@ -24429,7 +23710,7 @@ class value_iterator {
    * unescape it. This works well for typical ASCII and UTF-8 keys (almost all of them), but may
    * fail to match some keys with escapes (\u, \n, etc.).
    */
-  simdjson_warn_unused simdjson_really_inline error_code find_field(const std::string_view key) noexcept;
+  simdjson_warn_unused simdjson_inline error_code find_field(const std::string_view key) noexcept;
 
   /**
    * Find the next field with the given key, *without* unescaping. This assumes object order: it
@@ -24448,7 +23729,7 @@ class value_iterator {
    * unescape it. This works well for typical ASCII and UTF-8 keys (almost all of them), but may
    * fail to match some keys with escapes (\u, \n, etc.).
    */
-  simdjson_warn_unused simdjson_really_inline simdjson_result<bool> find_field_raw(const std::string_view key) noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<bool> find_field_raw(const std::string_view key) noexcept;
 
   /**
    * Find the field with the given key without regard to order, and *without* unescaping.
@@ -24468,7 +23749,7 @@ class value_iterator {
    * unescape it. This works well for typical ASCII and UTF-8 keys (almost all of them), but may
    * fail to match some keys with escapes (\u, \n, etc.).
    */
-  simdjson_warn_unused simdjson_really_inline simdjson_result<bool> find_field_unordered_raw(const std::string_view key) noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<bool> find_field_unordered_raw(const std::string_view key) noexcept;
 
   /** @} */
 
@@ -24485,7 +23766,7 @@ class value_iterator {
    * @returns Whether the array had any elements (returns false for empty).
    * @error INCORRECT_TYPE If there is no [.
    */
-  simdjson_warn_unused simdjson_really_inline simdjson_result<bool> start_array() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<bool> start_array() noexcept;
   /**
    * Check for an opening [ and start an array iteration while at the root.
    *
@@ -24493,7 +23774,7 @@ class value_iterator {
    * @error INCORRECT_TYPE If there is no [.
    * @error TAPE_ERROR if there is no matching ] at end of document
    */
-  simdjson_warn_unused simdjson_really_inline simdjson_result<bool> start_root_array() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<bool> start_root_array() noexcept;
 
   /**
    * Start an array iteration, after the user has already checked and moved past the [.
@@ -24504,7 +23785,7 @@ class value_iterator {
    * @error INCOMPLETE_ARRAY_OR_OBJECT If there are no more tokens (implying the *parent*
    *        array or object is incomplete).
    */
-  simdjson_warn_unused simdjson_really_inline simdjson_result<bool> started_array() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<bool> started_array() noexcept;
   /**
    * Start an array iteration from the root, after the user has already checked and moved past the [.
    *
@@ -24514,7 +23795,7 @@ class value_iterator {
    * @error INCOMPLETE_ARRAY_OR_OBJECT If there are no more tokens (implying the *parent*
    *        array or object is incomplete).
    */
-  simdjson_warn_unused simdjson_really_inline simdjson_result<bool> started_root_array() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<bool> started_root_array() noexcept;
 
   /**
    * Moves to the next element in an array.
@@ -24525,12 +23806,12 @@ class value_iterator {
    * @return Whether there is another element in the array.
    * @error TAPE_ERROR If there is a comma missing between elements.
    */
-  simdjson_warn_unused simdjson_really_inline simdjson_result<bool> has_next_element() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<bool> has_next_element() noexcept;
 
   /**
    * Get a child value iterator.
    */
-  simdjson_warn_unused simdjson_really_inline value_iterator child() const noexcept;
+  simdjson_warn_unused simdjson_inline value_iterator child() const noexcept;
 
   /** @} */
 
@@ -24540,43 +23821,43 @@ class value_iterator {
    * @{
    */
 
-  simdjson_warn_unused simdjson_really_inline simdjson_result<std::string_view> get_string() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<raw_json_string> get_raw_json_string() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<uint64_t> get_uint64() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<uint64_t> get_uint64_in_string() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<int64_t> get_int64() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<int64_t> get_int64_in_string() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<double> get_double() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<double> get_double_in_string() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<bool> get_bool() noexcept;
-  simdjson_really_inline bool is_null() noexcept;
-  simdjson_warn_unused simdjson_really_inline bool is_negative() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<bool> is_integer() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<number_type> get_number_type() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<number> get_number() noexcept;
-
-  simdjson_warn_unused simdjson_really_inline simdjson_result<std::string_view> get_root_string() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<raw_json_string> get_root_raw_json_string() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<uint64_t> get_root_uint64() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<uint64_t> get_root_uint64_in_string() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<int64_t> get_root_int64() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<int64_t> get_root_int64_in_string() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<double> get_root_double() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<double> get_root_double_in_string() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<bool> get_root_bool() noexcept;
-  simdjson_warn_unused simdjson_really_inline bool is_root_negative() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<bool> is_root_integer() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<number_type> get_root_number_type() noexcept;
-  simdjson_warn_unused simdjson_really_inline simdjson_result<number> get_root_number() noexcept;
-  simdjson_really_inline bool is_root_null() noexcept;
-
-  simdjson_really_inline error_code error() const noexcept;
-  simdjson_really_inline uint8_t *&string_buf_loc() noexcept;
-  simdjson_really_inline const json_iterator &json_iter() const noexcept;
-  simdjson_really_inline json_iterator &json_iter() noexcept;
-
-  simdjson_really_inline void assert_is_valid() const noexcept;
-  simdjson_really_inline bool is_valid() const noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<std::string_view> get_string() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<raw_json_string> get_raw_json_string() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<uint64_t> get_uint64() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<uint64_t> get_uint64_in_string() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<int64_t> get_int64() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<int64_t> get_int64_in_string() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<double> get_double() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<double> get_double_in_string() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<bool> get_bool() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<bool> is_null() noexcept;
+  simdjson_warn_unused simdjson_inline bool is_negative() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<bool> is_integer() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<number_type> get_number_type() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<number> get_number() noexcept;
+
+  simdjson_warn_unused simdjson_inline simdjson_result<std::string_view> get_root_string() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<raw_json_string> get_root_raw_json_string() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<uint64_t> get_root_uint64() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<uint64_t> get_root_uint64_in_string() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<int64_t> get_root_int64() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<int64_t> get_root_int64_in_string() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<double> get_root_double() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<double> get_root_double_in_string() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<bool> get_root_bool() noexcept;
+  simdjson_warn_unused simdjson_inline bool is_root_negative() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<bool> is_root_integer() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<number_type> get_root_number_type() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<number> get_root_number() noexcept;
+  simdjson_inline bool is_root_null() noexcept;
+
+  simdjson_inline error_code error() const noexcept;
+  simdjson_inline uint8_t *&string_buf_loc() noexcept;
+  simdjson_inline const json_iterator &json_iter() const noexcept;
+  simdjson_inline json_iterator &json_iter() noexcept;
+
+  simdjson_inline void assert_is_valid() const noexcept;
+  simdjson_inline bool is_valid() const noexcept;
 
   /** @} */
 protected:
@@ -24584,12 +23865,12 @@ class value_iterator {
    * Restarts an array iteration.
    * @returns Whether the array has any elements (returns false for empty).
    */
-  simdjson_really_inline simdjson_result<bool> reset_array() noexcept;
+  simdjson_inline simdjson_result<bool> reset_array() noexcept;
   /**
    * Restarts an object iteration.
    * @returns Whether the object has any fields (returns false for empty).
    */
-  simdjson_really_inline simdjson_result<bool> reset_object() noexcept;
+  simdjson_inline simdjson_result<bool> reset_object() noexcept;
   /**
    * move_at_start(): moves us so that we are pointing at the beginning of
    * the container. It updates the index so that at_start() is true and it
@@ -24597,7 +23878,7 @@ class value_iterator {
    *
    * Usage: used with value::count_elements().
    **/
-  simdjson_really_inline void move_at_start() noexcept;
+  simdjson_inline void move_at_start() noexcept;
 
   /**
    * move_at_container_start(): moves us so that we are pointing at the beginning of
@@ -24605,15 +23886,15 @@ class value_iterator {
    *
    * Usage: used with reset_array() and reset_object().
    **/
-   simdjson_really_inline void move_at_container_start() noexcept;
+   simdjson_inline void move_at_container_start() noexcept;
   /* Useful for debugging and logging purposes. */
   inline std::string to_string() const noexcept;
-  simdjson_really_inline value_iterator(json_iterator *json_iter, depth_t depth, token_position start_index) noexcept;
+  simdjson_inline value_iterator(json_iterator *json_iter, depth_t depth, token_position start_index) noexcept;
 
-  simdjson_really_inline bool parse_null(const uint8_t *json) const noexcept;
-  simdjson_really_inline simdjson_result<bool> parse_bool(const uint8_t *json) const noexcept;
-  simdjson_really_inline const uint8_t *peek_start() const noexcept;
-  simdjson_really_inline uint32_t peek_start_length() const noexcept;
+  simdjson_inline simdjson_result<bool> parse_null(const uint8_t *json) const noexcept;
+  simdjson_inline simdjson_result<bool> parse_bool(const uint8_t *json) const noexcept;
+  simdjson_inline const uint8_t *peek_start() const noexcept;
+  simdjson_inline uint32_t peek_start_length() const noexcept;
 
   /**
    * The general idea of the advance_... methods and the peek_* methods
@@ -24644,17 +23925,17 @@ class value_iterator {
    * Unfortunately, it makes the code more verbose, longer and maybe more error prone.
    */
 
-  simdjson_really_inline void advance_scalar(const char *type) noexcept;
-  simdjson_really_inline void advance_root_scalar(const char *type) noexcept;
-  simdjson_really_inline void advance_non_root_scalar(const char *type) noexcept;
+  simdjson_inline void advance_scalar(const char *type) noexcept;
+  simdjson_inline void advance_root_scalar(const char *type) noexcept;
+  simdjson_inline void advance_non_root_scalar(const char *type) noexcept;
 
-  simdjson_really_inline const uint8_t *peek_scalar(const char *type) noexcept;
-  simdjson_really_inline const uint8_t *peek_root_scalar(const char *type) noexcept;
-  simdjson_really_inline const uint8_t *peek_non_root_scalar(const char *type) noexcept;
+  simdjson_inline const uint8_t *peek_scalar(const char *type) noexcept;
+  simdjson_inline const uint8_t *peek_root_scalar(const char *type) noexcept;
+  simdjson_inline const uint8_t *peek_non_root_scalar(const char *type) noexcept;
 
 
-  simdjson_really_inline error_code start_container(uint8_t start_char, const char *incorrect_type_message, const char *type) noexcept;
-  simdjson_really_inline error_code end_container() noexcept;
+  simdjson_inline error_code start_container(uint8_t start_char, const char *incorrect_type_message, const char *type) noexcept;
+  simdjson_inline error_code end_container() noexcept;
 
   /**
    * Advance to a place expecting a value (increasing depth).
@@ -24662,19 +23943,19 @@ class value_iterator {
    * @return The current token (the one left behind).
    * @error TAPE_ERROR If the document ended early.
    */
-  simdjson_really_inline simdjson_result<const uint8_t *> advance_to_value() noexcept;
+  simdjson_inline simdjson_result<const uint8_t *> advance_to_value() noexcept;
 
-  simdjson_really_inline error_code incorrect_type_error(const char *message) const noexcept;
-  simdjson_really_inline error_code error_unless_more_tokens(uint32_t tokens=1) const noexcept;
+  simdjson_inline error_code incorrect_type_error(const char *message) const noexcept;
+  simdjson_inline error_code error_unless_more_tokens(uint32_t tokens=1) const noexcept;
 
-  simdjson_really_inline bool is_at_start() const noexcept;
+  simdjson_inline bool is_at_start() const noexcept;
   /**
    * is_at_iterator_start() returns true on an array or object after it has just been
    * created, whether the instance is empty or not.
    *
    * Usage: used by array::begin() in debug mode (SIMDJSON_DEVELOPMENT_CHECKS)
    */
-  simdjson_really_inline bool is_at_iterator_start() const noexcept;
+  simdjson_inline bool is_at_iterator_start() const noexcept;
 
   /**
    * Assuming that we are within an object, this returns true if we
@@ -24683,7 +23964,7 @@ class value_iterator {
    * Usage: the skip_child() method should never be used while we are pointing
    * at a key inside an object.
    */
-  simdjson_really_inline bool is_at_key() const noexcept;
+  simdjson_inline bool is_at_key() const noexcept;
 
   inline void assert_at_start() const noexcept;
   inline void assert_at_container_start() const noexcept;
@@ -24693,16 +23974,16 @@ class value_iterator {
   inline void assert_at_non_root_start() const noexcept;
 
   /** Get the starting position of this value */
-  simdjson_really_inline token_position start_position() const noexcept;
+  simdjson_inline token_position start_position() const noexcept;
 
   /** @copydoc error_code json_iterator::position() const noexcept; */
-  simdjson_really_inline token_position position() const noexcept;
+  simdjson_inline token_position position() const noexcept;
   /** @copydoc error_code json_iterator::end_position() const noexcept; */
-  simdjson_really_inline token_position last_position() const noexcept;
+  simdjson_inline token_position last_position() const noexcept;
   /** @copydoc error_code json_iterator::end_position() const noexcept; */
-  simdjson_really_inline token_position end_position() const noexcept;
+  simdjson_inline token_position end_position() const noexcept;
   /** @copydoc error_code json_iterator::report_error(error_code error, const char *message) noexcept; */
-  simdjson_really_inline error_code report_error(error_code error, const char *message) noexcept;
+  simdjson_inline error_code report_error(error_code error, const char *message) noexcept;
 
   friend class document;
   friend class object;
@@ -24719,9 +24000,9 @@ namespace simdjson {
 template<>
 struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value_iterator> : public SIMDJSON_BUILTIN_IMPLEMENTATION::implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value_iterator> {
 public:
-  simdjson_really_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value_iterator &&value) noexcept; ///< @private
-  simdjson_really_inline simdjson_result(error_code error) noexcept; ///< @private
-  simdjson_really_inline simdjson_result() noexcept = default;
+  simdjson_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value_iterator &&value) noexcept; ///< @private
+  simdjson_inline simdjson_result(error_code error) noexcept; ///< @private
+  simdjson_inline simdjson_result() noexcept = default;
 };
 
 } // namespace simdjson
@@ -24747,7 +24028,7 @@ class document;
 class array_iterator {
 public:
   /** Create a new, invalid array iterator. */
-  simdjson_really_inline array_iterator() noexcept = default;
+  simdjson_inline array_iterator() noexcept = default;
 
   //
   // Iterator interface
@@ -24758,7 +24039,7 @@ class array_iterator {
    *
    * Part of the std::iterator interface.
    */
-  simdjson_really_inline simdjson_result<value> operator*() noexcept; // MUST ONLY BE CALLED ONCE PER ITERATION.
+  simdjson_inline simdjson_result<value> operator*() noexcept; // MUST ONLY BE CALLED ONCE PER ITERATION.
   /**
    * Check if we are at the end of the JSON.
    *
@@ -24766,7 +24047,7 @@ class array_iterator {
    *
    * @return true if there are no more elements in the JSON array.
    */
-  simdjson_really_inline bool operator==(const array_iterator &) const noexcept;
+  simdjson_inline bool operator==(const array_iterator &) const noexcept;
   /**
    * Check if there are more elements in the JSON array.
    *
@@ -24774,18 +24055,18 @@ class array_iterator {
    *
    * @return true if there are more elements in the JSON array.
    */
-  simdjson_really_inline bool operator!=(const array_iterator &) const noexcept;
+  simdjson_inline bool operator!=(const array_iterator &) const noexcept;
   /**
    * Move to the next element.
    *
    * Part of the std::iterator interface.
    */
-  simdjson_really_inline array_iterator &operator++() noexcept;
+  simdjson_inline array_iterator &operator++() noexcept;
 
 private:
   value_iterator iter{};
 
-  simdjson_really_inline array_iterator(const value_iterator &iter) noexcept;
+  simdjson_inline array_iterator(const value_iterator &iter) noexcept;
 
   friend class array;
   friend class value;
@@ -24801,18 +24082,18 @@ namespace simdjson {
 template<>
 struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> : public SIMDJSON_BUILTIN_IMPLEMENTATION::implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> {
 public:
-  simdjson_really_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator &&value) noexcept; ///< @private
-  simdjson_really_inline simdjson_result(error_code error) noexcept; ///< @private
-  simdjson_really_inline simdjson_result() noexcept = default;
+  simdjson_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator &&value) noexcept; ///< @private
+  simdjson_inline simdjson_result(error_code error) noexcept; ///< @private
+  simdjson_inline simdjson_result() noexcept = default;
 
   //
   // Iterator interface
   //
 
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> operator*() noexcept; // MUST ONLY BE CALLED ONCE PER ITERATION.
-  simdjson_really_inline bool operator==(const simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> &) const noexcept;
-  simdjson_really_inline bool operator!=(const simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> &) const noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> &operator++() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> operator*() noexcept; // MUST ONLY BE CALLED ONCE PER ITERATION.
+  simdjson_inline bool operator==(const simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> &) const noexcept;
+  simdjson_inline bool operator!=(const simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> &) const noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> &operator++() noexcept;
 };
 
 } // namespace simdjson
@@ -24832,7 +24113,7 @@ class object_iterator {
    *
    * Exists so you can declare a variable and later assign to it before use.
    */
-  simdjson_really_inline object_iterator() noexcept = default;
+  simdjson_inline object_iterator() noexcept = default;
 
   //
   // Iterator interface
@@ -24840,13 +24121,13 @@ class object_iterator {
 
   // Reads key and value, yielding them to the user.
   // MUST ONLY BE CALLED ONCE PER ITERATION.
-  simdjson_really_inline simdjson_result<field> operator*() noexcept;
+  simdjson_inline simdjson_result<field> operator*() noexcept;
   // Assumes it's being compared with the end. true if depth < iter->depth.
-  simdjson_really_inline bool operator==(const object_iterator &) const noexcept;
+  simdjson_inline bool operator==(const object_iterator &) const noexcept;
   // Assumes it's being compared with the end. true if depth >= iter->depth.
-  simdjson_really_inline bool operator!=(const object_iterator &) const noexcept;
+  simdjson_inline bool operator!=(const object_iterator &) const noexcept;
   // Checks for ']' and ','
-  simdjson_really_inline object_iterator &operator++() noexcept;
+  simdjson_inline object_iterator &operator++() noexcept;
 
 private:
   /**
@@ -24857,7 +24138,7 @@ class object_iterator {
    */
   value_iterator iter{};
 
-  simdjson_really_inline object_iterator(const value_iterator &iter) noexcept;
+  simdjson_inline object_iterator(const value_iterator &iter) noexcept;
   friend struct simdjson_result<object_iterator>;
   friend class object;
 };
@@ -24871,22 +24152,22 @@ namespace simdjson {
 template<>
 struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> : public SIMDJSON_BUILTIN_IMPLEMENTATION::implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> {
 public:
-  simdjson_really_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator &&value) noexcept; ///< @private
-  simdjson_really_inline simdjson_result(error_code error) noexcept; ///< @private
-  simdjson_really_inline simdjson_result() noexcept = default;
+  simdjson_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator &&value) noexcept; ///< @private
+  simdjson_inline simdjson_result(error_code error) noexcept; ///< @private
+  simdjson_inline simdjson_result() noexcept = default;
 
   //
   // Iterator interface
   //
 
   // Reads key and value, yielding them to the user.
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field> operator*() noexcept; // MUST ONLY BE CALLED ONCE PER ITERATION.
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field> operator*() noexcept; // MUST ONLY BE CALLED ONCE PER ITERATION.
   // Assumes it's being compared with the end. true if depth < iter->depth.
-  simdjson_really_inline bool operator==(const simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> &) const noexcept;
+  simdjson_inline bool operator==(const simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> &) const noexcept;
   // Assumes it's being compared with the end. true if depth >= iter->depth.
-  simdjson_really_inline bool operator!=(const simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> &) const noexcept;
+  simdjson_inline bool operator!=(const simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> &) const noexcept;
   // Checks for ']' and ','
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> &operator++() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> &operator++() noexcept;
 };
 
 } // namespace simdjson
@@ -24910,20 +24191,20 @@ class array {
    *
    * Exists so you can declare a variable and later assign to it before use.
    */
-  simdjson_really_inline array() noexcept = default;
+  simdjson_inline array() noexcept = default;
 
   /**
    * Begin array iteration.
    *
    * Part of the std::iterable interface.
    */
-  simdjson_really_inline simdjson_result<array_iterator> begin() noexcept;
+  simdjson_inline simdjson_result<array_iterator> begin() noexcept;
   /**
    * Sentinel representing the end of the array.
    *
    * Part of the std::iterable interface.
    */
-  simdjson_really_inline simdjson_result<array_iterator> end() noexcept;
+  simdjson_inline simdjson_result<array_iterator> end() noexcept;
   /**
    * This method scans the array and counts the number of elements.
    * The count_elements method should always be called before you have begun
@@ -24938,7 +24219,7 @@ class array {
    * To check that an array is empty, it is more performant to use
    * the is_empty() method.
    */
-  simdjson_really_inline simdjson_result<size_t> count_elements() & noexcept;
+  simdjson_inline simdjson_result<size_t> count_elements() & noexcept;
   /**
    * This method scans the beginning of the array and checks whether the
    * array is empty.
@@ -24948,7 +24229,7 @@ class array {
    * there is a missing comma), then an error is returned and it is no longer
    * safe to continue.
    */
-  simdjson_really_inline simdjson_result<bool> is_empty() & noexcept;
+  simdjson_inline simdjson_result<bool> is_empty() & noexcept;
   /**
    * Reset the iterator so that we are pointing back at the
    * beginning of the array. You should still consume values only once even if you
@@ -24993,7 +24274,7 @@ class array {
    * Consumes the array and returns a string_view instance corresponding to the
    * array as represented in JSON. It points inside the original document.
    */
-  simdjson_really_inline simdjson_result<std::string_view> raw_json() noexcept;
+  simdjson_inline simdjson_result<std::string_view> raw_json() noexcept;
 
   /**
    * Get the value at the given index. This function has linear-time complexity.
@@ -25002,12 +24283,12 @@ class array {
    * @return The value at the given index, or:
    *         - INDEX_OUT_OF_BOUNDS if the array index is larger than an array length
    */
-  simdjson_really_inline simdjson_result<value> at(size_t index) noexcept;
+  simdjson_inline simdjson_result<value> at(size_t index) noexcept;
 protected:
   /**
    * Go to the end of the array, no matter where you are right now.
    */
-  simdjson_really_inline error_code consume() noexcept;
+  simdjson_inline error_code consume() noexcept;
 
   /**
    * Begin array iteration.
@@ -25016,7 +24297,7 @@ class array {
    *        resulting array.
    * @error INCORRECT_TYPE if the iterator is not at [.
    */
-  static simdjson_really_inline simdjson_result<array> start(value_iterator &iter) noexcept;
+  static simdjson_inline simdjson_result<array> start(value_iterator &iter) noexcept;
   /**
    * Begin array iteration from the root.
    *
@@ -25025,7 +24306,7 @@ class array {
    * @error INCORRECT_TYPE if the iterator is not at [.
    * @error TAPE_ERROR if there is no closing ] at the end of the document.
    */
-  static simdjson_really_inline simdjson_result<array> start_root(value_iterator &iter) noexcept;
+  static simdjson_inline simdjson_result<array> start_root(value_iterator &iter) noexcept;
   /**
    * Begin array iteration.
    *
@@ -25034,7 +24315,7 @@ class array {
    *
    * @param iter The iterator. Must be after the initial [. Will be *moved* into the resulting array.
    */
-  static simdjson_really_inline simdjson_result<array> started(value_iterator &iter) noexcept;
+  static simdjson_inline simdjson_result<array> started(value_iterator &iter) noexcept;
 
   /**
    * Create an array at the given Internal array creation. Call array::start() or array::started() instead of this.
@@ -25043,7 +24324,7 @@ class array {
    *        == true, or past the [] with is_alive() == false if the array is empty. Will be *moved*
    *        into the resulting array.
    */
-  simdjson_really_inline array(const value_iterator &iter) noexcept;
+  simdjson_inline array(const value_iterator &iter) noexcept;
 
   /**
    * Iterator marking current position.
@@ -25068,17 +24349,17 @@ namespace simdjson {
 template<>
 struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array> : public SIMDJSON_BUILTIN_IMPLEMENTATION::implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array> {
 public:
-  simdjson_really_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array &&value) noexcept; ///< @private
-  simdjson_really_inline simdjson_result(error_code error) noexcept; ///< @private
-  simdjson_really_inline simdjson_result() noexcept = default;
+  simdjson_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array &&value) noexcept; ///< @private
+  simdjson_inline simdjson_result(error_code error) noexcept; ///< @private
+  simdjson_inline simdjson_result() noexcept = default;
 
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> begin() noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> end() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> begin() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> end() noexcept;
   inline simdjson_result<size_t> count_elements() & noexcept;
   inline simdjson_result<bool> is_empty() & noexcept;
   inline simdjson_result<bool> reset() & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> at(size_t index) noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> at_pointer(std::string_view json_pointer) noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> at(size_t index) noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> at_pointer(std::string_view json_pointer) noexcept;
 };
 
 } // namespace simdjson
@@ -25111,11 +24392,11 @@ class document {
    *
    * Exists so you can declare a variable and later assign to it before use.
    */
-  simdjson_really_inline document() noexcept = default;
-  simdjson_really_inline document(const document &other) noexcept = delete; // pass your documents by reference, not by copy
-  simdjson_really_inline document(document &&other) noexcept = default;
-  simdjson_really_inline document &operator=(const document &other) noexcept = delete;
-  simdjson_really_inline document &operator=(document &&other) noexcept = default;
+  simdjson_inline document() noexcept = default;
+  simdjson_inline document(const document &other) noexcept = delete; // pass your documents by reference, not by copy
+  simdjson_inline document(document &&other) noexcept = default;
+  simdjson_inline document &operator=(const document &other) noexcept = delete;
+  simdjson_inline document &operator=(document &&other) noexcept = default;
 
   /**
    * Cast this JSON value to an array.
@@ -25123,49 +24404,49 @@ class document {
    * @returns An object that can be used to iterate the array.
    * @returns INCORRECT_TYPE If the JSON value is not an array.
    */
-  simdjson_really_inline simdjson_result<array> get_array() & noexcept;
+  simdjson_inline simdjson_result<array> get_array() & noexcept;
   /**
    * Cast this JSON value to an object.
    *
    * @returns An object that can be used to look up or iterate fields.
    * @returns INCORRECT_TYPE If the JSON value is not an object.
    */
-  simdjson_really_inline simdjson_result<object> get_object() & noexcept;
+  simdjson_inline simdjson_result<object> get_object() & noexcept;
   /**
    * Cast this JSON value to an unsigned integer.
    *
    * @returns A signed 64-bit integer.
    * @returns INCORRECT_TYPE If the JSON value is not a 64-bit unsigned integer.
    */
-  simdjson_really_inline simdjson_result<uint64_t> get_uint64() noexcept;
+  simdjson_inline simdjson_result<uint64_t> get_uint64() noexcept;
   /**
    * Cast this JSON value (inside string) to an unsigned integer.
    *
    * @returns A signed 64-bit integer.
    * @returns INCORRECT_TYPE If the JSON value is not a 64-bit unsigned integer.
    */
-  simdjson_really_inline simdjson_result<uint64_t> get_uint64_in_string() noexcept;
+  simdjson_inline simdjson_result<uint64_t> get_uint64_in_string() noexcept;
   /**
    * Cast this JSON value to a signed integer.
    *
    * @returns A signed 64-bit integer.
    * @returns INCORRECT_TYPE If the JSON value is not a 64-bit integer.
    */
-  simdjson_really_inline simdjson_result<int64_t> get_int64() noexcept;
+  simdjson_inline simdjson_result<int64_t> get_int64() noexcept;
   /**
    * Cast this JSON value (inside string) to a signed integer.
    *
    * @returns A signed 64-bit integer.
    * @returns INCORRECT_TYPE If the JSON value is not a 64-bit integer.
    */
-  simdjson_really_inline simdjson_result<int64_t> get_int64_in_string() noexcept;
+  simdjson_inline simdjson_result<int64_t> get_int64_in_string() noexcept;
   /**
    * Cast this JSON value to a double.
    *
    * @returns A double.
    * @returns INCORRECT_TYPE If the JSON value is not a valid floating-point number.
    */
-  simdjson_really_inline simdjson_result<double> get_double() noexcept;
+  simdjson_inline simdjson_result<double> get_double() noexcept;
 
   /**
    * Cast this JSON value (inside string) to a double.
@@ -25173,7 +24454,7 @@ class document {
    * @returns A double.
    * @returns INCORRECT_TYPE If the JSON value is not a valid floating-point number.
    */
-  simdjson_really_inline simdjson_result<double> get_double_in_string() noexcept;
+  simdjson_inline simdjson_result<double> get_double_in_string() noexcept;
   /**
    * Cast this JSON value to a string.
    *
@@ -25185,7 +24466,7 @@ class document {
    *          time it parses a document or when it is destroyed.
    * @returns INCORRECT_TYPE if the JSON value is not a string.
    */
-  simdjson_really_inline simdjson_result<std::string_view> get_string() noexcept;
+  simdjson_inline simdjson_result<std::string_view> get_string() noexcept;
   /**
    * Cast this JSON value to a raw_json_string.
    *
@@ -25194,28 +24475,31 @@ class document {
    * @returns A pointer to the raw JSON for the given string.
    * @returns INCORRECT_TYPE if the JSON value is not a string.
    */
-  simdjson_really_inline simdjson_result<raw_json_string> get_raw_json_string() noexcept;
+  simdjson_inline simdjson_result<raw_json_string> get_raw_json_string() noexcept;
   /**
    * Cast this JSON value to a bool.
    *
    * @returns A bool value.
    * @returns INCORRECT_TYPE if the JSON value is not true or false.
    */
-  simdjson_really_inline simdjson_result<bool> get_bool() noexcept;
+  simdjson_inline simdjson_result<bool> get_bool() noexcept;
   /**
    * Cast this JSON value to a value when the document is an object or an array.
    *
    * @returns A value if a JSON array or object cannot be found.
    * @returns SCALAR_DOCUMENT_AS_VALUE error is the document is a scalar (see is_scalar() function).
    */
-  simdjson_really_inline simdjson_result<value> get_value() noexcept;
+  simdjson_inline simdjson_result<value> get_value() noexcept;
 
   /**
-   * Checks if this JSON value is null.
+   * Checks if this JSON value is null.  If and only if the value is
+   * null, then it is consumed (we advance). If we find a token that
+   * begins with 'n' but is not 'null', then an error is returned.
    *
    * @returns Whether the value is null.
+   * @returns INCORRECT_TYPE If the JSON value begins with 'n' and is not 'null'.
    */
-  simdjson_really_inline bool is_null() noexcept;
+  simdjson_inline simdjson_result<bool> is_null() noexcept;
 
   /**
    * Get this value as the given type.
@@ -25228,13 +24512,13 @@ class document {
    * @returns A value of the given type, parsed from the JSON.
    * @returns INCORRECT_TYPE If the JSON value is not the given type.
    */
-  template<typename T> simdjson_really_inline simdjson_result<T> get() & noexcept {
+  template<typename T> simdjson_inline simdjson_result<T> get() & noexcept {
     // Unless the simdjson library provides an inline implementation, calling this method should
     // immediately fail.
     static_assert(!sizeof(T), "The get method with given type is not implemented by the simdjson library.");
   }
   /** @overload template<typename T> simdjson_result<T> get() & noexcept */
-  template<typename T> simdjson_really_inline simdjson_result<T> get() && noexcept {
+  template<typename T> simdjson_inline simdjson_result<T> get() && noexcept {
     // Unless the simdjson library provides an inline implementation, calling this method should
     // immediately fail.
     static_assert(!sizeof(T), "The get method with given type is not implemented by the simdjson library.");
@@ -25251,9 +24535,9 @@ class document {
    * @returns INCORRECT_TYPE If the JSON value is not an object.
    * @returns SUCCESS If the parse succeeded and the out parameter was set to the value.
    */
-  template<typename T> simdjson_really_inline error_code get(T &out) & noexcept;
+  template<typename T> simdjson_inline error_code get(T &out) & noexcept;
   /** @overload template<typename T> error_code get(T &out) & noexcept */
-  template<typename T> simdjson_really_inline error_code get(T &out) && noexcept;
+  template<typename T> simdjson_inline error_code get(T &out) && noexcept;
 
 #if SIMDJSON_EXCEPTIONS
   /**
@@ -25262,35 +24546,35 @@ class document {
    * @returns An object that can be used to iterate the array.
    * @exception simdjson_error(INCORRECT_TYPE) If the JSON value is not an array.
    */
-  simdjson_really_inline operator array() & noexcept(false);
+  simdjson_inline operator array() & noexcept(false);
   /**
    * Cast this JSON value to an object.
    *
    * @returns An object that can be used to look up or iterate fields.
    * @exception simdjson_error(INCORRECT_TYPE) If the JSON value is not an object.
    */
-  simdjson_really_inline operator object() & noexcept(false);
+  simdjson_inline operator object() & noexcept(false);
   /**
    * Cast this JSON value to an unsigned integer.
    *
    * @returns A signed 64-bit integer.
    * @exception simdjson_error(INCORRECT_TYPE) If the JSON value is not a 64-bit unsigned integer.
    */
-  simdjson_really_inline operator uint64_t() noexcept(false);
+  simdjson_inline operator uint64_t() noexcept(false);
   /**
    * Cast this JSON value to a signed integer.
    *
    * @returns A signed 64-bit integer.
    * @exception simdjson_error(INCORRECT_TYPE) If the JSON value is not a 64-bit integer.
    */
-  simdjson_really_inline operator int64_t() noexcept(false);
+  simdjson_inline operator int64_t() noexcept(false);
   /**
    * Cast this JSON value to a double.
    *
    * @returns A double.
    * @exception simdjson_error(INCORRECT_TYPE) If the JSON value is not a valid floating-point number.
    */
-  simdjson_really_inline operator double() noexcept(false);
+  simdjson_inline operator double() noexcept(false);
   /**
    * Cast this JSON value to a string.
    *
@@ -25300,7 +24584,7 @@ class document {
    *          time it parses a document or when it is destroyed.
    * @exception simdjson_error(INCORRECT_TYPE) if the JSON value is not a string.
    */
-  simdjson_really_inline operator std::string_view() noexcept(false);
+  simdjson_inline operator std::string_view() noexcept(false);
   /**
    * Cast this JSON value to a raw_json_string.
    *
@@ -25309,21 +24593,21 @@ class document {
    * @returns A pointer to the raw JSON for the given string.
    * @exception simdjson_error(INCORRECT_TYPE) if the JSON value is not a string.
    */
-  simdjson_really_inline operator raw_json_string() noexcept(false);
+  simdjson_inline operator raw_json_string() noexcept(false);
   /**
    * Cast this JSON value to a bool.
    *
    * @returns A bool value.
    * @exception simdjson_error(INCORRECT_TYPE) if the JSON value is not true or false.
    */
-  simdjson_really_inline operator bool() noexcept(false);
+  simdjson_inline operator bool() noexcept(false);
   /**
    * Cast this JSON value to a value.
    *
    * @returns A value value.
    * @exception if a JSON value cannot be found
    */
-  simdjson_really_inline operator value() noexcept(false);
+  simdjson_inline operator value() noexcept(false);
 #endif
   /**
    * This method scans the array and counts the number of elements.
@@ -25336,7 +24620,7 @@ class document {
    * there is a missing comma), then an error is returned and it is no longer
    * safe to continue.
    */
-  simdjson_really_inline simdjson_result<size_t> count_elements() & noexcept;
+  simdjson_inline simdjson_result<size_t> count_elements() & noexcept;
    /**
    * This method scans the object and counts the number of key-value pairs.
    * The count_fields method should always be called before you have begun
@@ -25351,7 +24635,7 @@ class document {
    * To check that an object is empty, it is more performant to use
    * the is_empty() method.
    */
-  simdjson_really_inline simdjson_result<size_t> count_fields() & noexcept;
+  simdjson_inline simdjson_result<size_t> count_fields() & noexcept;
   /**
    * Get the value at the given index in the array. This function has linear-time complexity.
    * This function should only be called once on an array instance since the array iterator is not reset between each call.
@@ -25359,19 +24643,19 @@ class document {
    * @return The value at the given index, or:
    *         - INDEX_OUT_OF_BOUNDS if the array index is larger than an array length
    */
-  simdjson_really_inline simdjson_result<value> at(size_t index) & noexcept;
+  simdjson_inline simdjson_result<value> at(size_t index) & noexcept;
   /**
    * Begin array iteration.
    *
    * Part of the std::iterable interface.
    */
-  simdjson_really_inline simdjson_result<array_iterator> begin() & noexcept;
+  simdjson_inline simdjson_result<array_iterator> begin() & noexcept;
   /**
    * Sentinel representing the end of the array.
    *
    * Part of the std::iterable interface.
    */
-  simdjson_really_inline simdjson_result<array_iterator> end() & noexcept;
+  simdjson_inline simdjson_result<array_iterator> end() & noexcept;
 
   /**
    * Look up a field by name on an object (order-sensitive).
@@ -25405,9 +24689,9 @@ class document {
    * @param key The key to look up.
    * @returns The value of the field, or NO_SUCH_FIELD if the field is not in the object.
    */
-  simdjson_really_inline simdjson_result<value> find_field(std::string_view key) & noexcept;
-  /** @overload simdjson_really_inline simdjson_result<value> find_field(std::string_view key) & noexcept; */
-  simdjson_really_inline simdjson_result<value> find_field(const char *key) & noexcept;
+  simdjson_inline simdjson_result<value> find_field(std::string_view key) & noexcept;
+  /** @overload simdjson_inline simdjson_result<value> find_field(std::string_view key) & noexcept; */
+  simdjson_inline simdjson_result<value> find_field(const char *key) & noexcept;
 
   /**
    * Look up a field by name on an object, without regard to key order.
@@ -25439,16 +24723,18 @@ class document {
    * @param key The key to look up.
    * @returns The value of the field, or NO_SUCH_FIELD if the field is not in the object.
    */
-  simdjson_really_inline simdjson_result<value> find_field_unordered(std::string_view key) & noexcept;
-  /** @overload simdjson_really_inline simdjson_result<value> find_field_unordered(std::string_view key) & noexcept; */
-  simdjson_really_inline simdjson_result<value> find_field_unordered(const char *key) & noexcept;
-  /** @overload simdjson_really_inline simdjson_result<value> find_field_unordered(std::string_view key) & noexcept; */
-  simdjson_really_inline simdjson_result<value> operator[](std::string_view key) & noexcept;
-  /** @overload simdjson_really_inline simdjson_result<value> find_field_unordered(std::string_view key) & noexcept; */
-  simdjson_really_inline simdjson_result<value> operator[](const char *key) & noexcept;
+  simdjson_inline simdjson_result<value> find_field_unordered(std::string_view key) & noexcept;
+  /** @overload simdjson_inline simdjson_result<value> find_field_unordered(std::string_view key) & noexcept; */
+  simdjson_inline simdjson_result<value> find_field_unordered(const char *key) & noexcept;
+  /** @overload simdjson_inline simdjson_result<value> find_field_unordered(std::string_view key) & noexcept; */
+  simdjson_inline simdjson_result<value> operator[](std::string_view key) & noexcept;
+  /** @overload simdjson_inline simdjson_result<value> find_field_unordered(std::string_view key) & noexcept; */
+  simdjson_inline simdjson_result<value> operator[](const char *key) & noexcept;
 
   /**
-   * Get the type of this JSON value.
+   * Get the type of this JSON value. It does not validate or consume the value.
+   * E.g., you must still call "is_null()" to check that a value is null even if
+   * "type()" returns json_type::null.
    *
    * NOTE: If you're only expecting a value to be one type (a typical case), it's generally
    * better to just call .get_double, .get_string, etc. and check for INCORRECT_TYPE (or just
@@ -25456,7 +24742,7 @@ class document {
    *
    * @error TAPE_ERROR when the JSON value is a bad token like "}" "," or "alse".
    */
-  simdjson_really_inline simdjson_result<json_type> type() noexcept;
+  simdjson_inline simdjson_result<json_type> type() noexcept;
 
   /**
    * Checks whether the document is a scalar (string, number, null, Boolean).
@@ -25465,14 +24751,14 @@ class document {
    * @returns true if the type is string, number, null, Boolean
    * @error TAPE_ERROR when the JSON value is a bad token like "}" "," or "alse".
    */
-  simdjson_really_inline simdjson_result<bool> is_scalar() noexcept;
+  simdjson_inline simdjson_result<bool> is_scalar() noexcept;
 
   /**
    * Checks whether the document is a negative number.
    *
    * @returns true if the number if negative.
    */
-  simdjson_really_inline bool is_negative() noexcept;
+  simdjson_inline bool is_negative() noexcept;
   /**
    * Checks whether the document is an integer number. Note that
    * this requires to partially parse the number string. If
@@ -25482,7 +24768,7 @@ class document {
    *
    * @returns true if the number if negative.
    */
-  simdjson_really_inline simdjson_result<bool> is_integer() noexcept;
+  simdjson_inline simdjson_result<bool> is_integer() noexcept;
   /**
    * Determine the number type (integer or floating-point number) as quickly
    * as possible. This function does not fully validate the input. It is
@@ -25505,7 +24791,7 @@ class document {
    *
    * @returns the type of the number
    */
-  simdjson_really_inline simdjson_result<number_type> get_number_type() noexcept;
+  simdjson_inline simdjson_result<number_type> get_number_type() noexcept;
 
   /**
    * Attempt to parse an ondemand::number. An ondemand::number may
@@ -25533,7 +24819,7 @@ class document {
    * to call "get_int64()" when number.get_number_type() is not
    * number_type::signed_integer and when number.is_int64() is false.
    */
-  simdjson_warn_unused simdjson_really_inline simdjson_result<number> get_number() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<number> get_number() noexcept;
 
   /**
    * Get the raw JSON for this token.
@@ -25557,7 +24843,7 @@ class document {
    * - false
    * - null
    */
-  simdjson_really_inline simdjson_result<std::string_view> raw_json_token() noexcept;
+  simdjson_inline simdjson_result<std::string_view> raw_json_token() noexcept;
 
   /**
    * Reset the iterator inside the document instance so we are pointing back at the
@@ -25589,7 +24875,7 @@ class document {
    *  2 = , or } inside root array/object
    *  3 = key or value inside root array/object.
    */
-  simdjson_really_inline int32_t current_depth() const noexcept;
+  simdjson_inline int32_t current_depth() const noexcept;
 
   /**
    * Get the value associated with the given JSON pointer.  We use the RFC 6901
@@ -25622,26 +24908,26 @@ class document {
    *         - INVALID_JSON_POINTER if the JSON pointer is invalid and cannot be parsed
    *         - SCALAR_DOCUMENT_AS_VALUE if the json_pointer is empty and the document is not a scalar (see is_scalar() function).
    */
-  simdjson_really_inline simdjson_result<value> at_pointer(std::string_view json_pointer) noexcept;
+  simdjson_inline simdjson_result<value> at_pointer(std::string_view json_pointer) noexcept;
   /**
    * Consumes the document and returns a string_view instance corresponding to the
    * document as represented in JSON. It points inside the original byte array containing
    * the JSON document.
    */
-  simdjson_really_inline simdjson_result<std::string_view> raw_json() noexcept;
+  simdjson_inline simdjson_result<std::string_view> raw_json() noexcept;
 protected:
   /**
    * Consumes the document.
    */
-  simdjson_really_inline error_code consume() noexcept;
+  simdjson_inline error_code consume() noexcept;
 
-  simdjson_really_inline document(ondemand::json_iterator &&iter) noexcept;
-  simdjson_really_inline const uint8_t *text(uint32_t idx) const noexcept;
+  simdjson_inline document(ondemand::json_iterator &&iter) noexcept;
+  simdjson_inline const uint8_t *text(uint32_t idx) const noexcept;
 
-  simdjson_really_inline value_iterator resume_value_iterator() noexcept;
-  simdjson_really_inline value_iterator get_root_value_iterator() noexcept;
-  simdjson_really_inline simdjson_result<object> start_or_resume_object() noexcept;
-  static simdjson_really_inline document start(ondemand::json_iterator &&iter) noexcept;
+  simdjson_inline value_iterator resume_value_iterator() noexcept;
+  simdjson_inline value_iterator get_root_value_iterator() noexcept;
+  simdjson_inline simdjson_result<object> start_or_resume_object() noexcept;
+  static simdjson_inline document start(ondemand::json_iterator &&iter) noexcept;
 
   //
   // Fields
@@ -25665,59 +24951,59 @@ class document {
  */
 class document_reference {
 public:
-  simdjson_really_inline document_reference() noexcept;
-  simdjson_really_inline document_reference(document &d) noexcept;
-  simdjson_really_inline document_reference(const document_reference &other) noexcept = default;
-  simdjson_really_inline document_reference& operator=(const document_reference &other) noexcept = default;
-  simdjson_really_inline void rewind() noexcept;
-  simdjson_really_inline simdjson_result<array> get_array() & noexcept;
-  simdjson_really_inline simdjson_result<object> get_object() & noexcept;
-  simdjson_really_inline simdjson_result<uint64_t> get_uint64() noexcept;
-  simdjson_really_inline simdjson_result<int64_t> get_int64() noexcept;
-  simdjson_really_inline simdjson_result<double> get_double() noexcept;
-  simdjson_really_inline simdjson_result<std::string_view> get_string() noexcept;
-  simdjson_really_inline simdjson_result<raw_json_string> get_raw_json_string() noexcept;
-  simdjson_really_inline simdjson_result<bool> get_bool() noexcept;
-  simdjson_really_inline simdjson_result<value> get_value() noexcept;
-
-  simdjson_really_inline bool is_null() noexcept;
-  simdjson_really_inline simdjson_result<std::string_view> raw_json() noexcept;
-  simdjson_really_inline operator document&() const noexcept;
+  simdjson_inline document_reference() noexcept;
+  simdjson_inline document_reference(document &d) noexcept;
+  simdjson_inline document_reference(const document_reference &other) noexcept = default;
+  simdjson_inline document_reference& operator=(const document_reference &other) noexcept = default;
+  simdjson_inline void rewind() noexcept;
+  simdjson_inline simdjson_result<array> get_array() & noexcept;
+  simdjson_inline simdjson_result<object> get_object() & noexcept;
+  simdjson_inline simdjson_result<uint64_t> get_uint64() noexcept;
+  simdjson_inline simdjson_result<int64_t> get_int64() noexcept;
+  simdjson_inline simdjson_result<double> get_double() noexcept;
+  simdjson_inline simdjson_result<std::string_view> get_string() noexcept;
+  simdjson_inline simdjson_result<raw_json_string> get_raw_json_string() noexcept;
+  simdjson_inline simdjson_result<bool> get_bool() noexcept;
+  simdjson_inline simdjson_result<value> get_value() noexcept;
+
+  simdjson_inline simdjson_result<bool> is_null() noexcept;
+  simdjson_inline simdjson_result<std::string_view> raw_json() noexcept;
+  simdjson_inline operator document&() const noexcept;
 
 #if SIMDJSON_EXCEPTIONS
-  simdjson_really_inline operator array() & noexcept(false);
-  simdjson_really_inline operator object() & noexcept(false);
-  simdjson_really_inline operator uint64_t() noexcept(false);
-  simdjson_really_inline operator int64_t() noexcept(false);
-  simdjson_really_inline operator double() noexcept(false);
-  simdjson_really_inline operator std::string_view() noexcept(false);
-  simdjson_really_inline operator raw_json_string() noexcept(false);
-  simdjson_really_inline operator bool() noexcept(false);
-  simdjson_really_inline operator value() noexcept(false);
+  simdjson_inline operator array() & noexcept(false);
+  simdjson_inline operator object() & noexcept(false);
+  simdjson_inline operator uint64_t() noexcept(false);
+  simdjson_inline operator int64_t() noexcept(false);
+  simdjson_inline operator double() noexcept(false);
+  simdjson_inline operator std::string_view() noexcept(false);
+  simdjson_inline operator raw_json_string() noexcept(false);
+  simdjson_inline operator bool() noexcept(false);
+  simdjson_inline operator value() noexcept(false);
 #endif
-  simdjson_really_inline simdjson_result<size_t> count_elements() & noexcept;
-  simdjson_really_inline simdjson_result<size_t> count_fields() & noexcept;
-  simdjson_really_inline simdjson_result<value> at(size_t index) & noexcept;
-  simdjson_really_inline simdjson_result<array_iterator> begin() & noexcept;
-  simdjson_really_inline simdjson_result<array_iterator> end() & noexcept;
-  simdjson_really_inline simdjson_result<value> find_field(std::string_view key) & noexcept;
-  simdjson_really_inline simdjson_result<value> find_field(const char *key) & noexcept;
-  simdjson_really_inline simdjson_result<value> operator[](std::string_view key) & noexcept;
-  simdjson_really_inline simdjson_result<value> operator[](const char *key) & noexcept;
-  simdjson_really_inline simdjson_result<value> find_field_unordered(std::string_view key) & noexcept;
-  simdjson_really_inline simdjson_result<value> find_field_unordered(const char *key) & noexcept;
-
-  simdjson_really_inline simdjson_result<json_type> type() noexcept;
-  simdjson_really_inline simdjson_result<bool> is_scalar() noexcept;
-
-  simdjson_really_inline simdjson_result<const char *> current_location() noexcept;
-  simdjson_really_inline int32_t current_depth() const noexcept;
-  simdjson_really_inline bool is_negative() noexcept;
-  simdjson_really_inline simdjson_result<bool> is_integer() noexcept;
-  simdjson_really_inline simdjson_result<number_type> get_number_type() noexcept;
-  simdjson_really_inline simdjson_result<number> get_number() noexcept;
-  simdjson_really_inline simdjson_result<std::string_view> raw_json_token() noexcept;
-  simdjson_really_inline simdjson_result<value> at_pointer(std::string_view json_pointer) noexcept;
+  simdjson_inline simdjson_result<size_t> count_elements() & noexcept;
+  simdjson_inline simdjson_result<size_t> count_fields() & noexcept;
+  simdjson_inline simdjson_result<value> at(size_t index) & noexcept;
+  simdjson_inline simdjson_result<array_iterator> begin() & noexcept;
+  simdjson_inline simdjson_result<array_iterator> end() & noexcept;
+  simdjson_inline simdjson_result<value> find_field(std::string_view key) & noexcept;
+  simdjson_inline simdjson_result<value> find_field(const char *key) & noexcept;
+  simdjson_inline simdjson_result<value> operator[](std::string_view key) & noexcept;
+  simdjson_inline simdjson_result<value> operator[](const char *key) & noexcept;
+  simdjson_inline simdjson_result<value> find_field_unordered(std::string_view key) & noexcept;
+  simdjson_inline simdjson_result<value> find_field_unordered(const char *key) & noexcept;
+
+  simdjson_inline simdjson_result<json_type> type() noexcept;
+  simdjson_inline simdjson_result<bool> is_scalar() noexcept;
+
+  simdjson_inline simdjson_result<const char *> current_location() noexcept;
+  simdjson_inline int32_t current_depth() const noexcept;
+  simdjson_inline bool is_negative() noexcept;
+  simdjson_inline simdjson_result<bool> is_integer() noexcept;
+  simdjson_inline simdjson_result<number_type> get_number_type() noexcept;
+  simdjson_inline simdjson_result<number> get_number() noexcept;
+  simdjson_inline simdjson_result<std::string_view> raw_json_token() noexcept;
+  simdjson_inline simdjson_result<value> at_pointer(std::string_view json_pointer) noexcept;
 private:
   document *doc{nullptr};
 };
@@ -25730,63 +25016,63 @@ namespace simdjson {
 template<>
 struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document> : public SIMDJSON_BUILTIN_IMPLEMENTATION::implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document> {
 public:
-  simdjson_really_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document &&value) noexcept; ///< @private
-  simdjson_really_inline simdjson_result(error_code error) noexcept; ///< @private
-  simdjson_really_inline simdjson_result() noexcept = default;
-  simdjson_really_inline error_code rewind() noexcept;
-
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array> get_array() & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object> get_object() & noexcept;
-  simdjson_really_inline simdjson_result<uint64_t> get_uint64() noexcept;
-  simdjson_really_inline simdjson_result<int64_t> get_int64() noexcept;
-  simdjson_really_inline simdjson_result<double> get_double() noexcept;
-  simdjson_really_inline simdjson_result<double> get_double_from_string() noexcept;
-  simdjson_really_inline simdjson_result<std::string_view> get_string() noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string> get_raw_json_string() noexcept;
-  simdjson_really_inline simdjson_result<bool> get_bool() noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> get_value() noexcept;
-  simdjson_really_inline bool is_null() noexcept;
-
-  template<typename T> simdjson_really_inline simdjson_result<T> get() & noexcept;
-  template<typename T> simdjson_really_inline simdjson_result<T> get() && noexcept;
-
-  template<typename T> simdjson_really_inline error_code get(T &out) & noexcept;
-  template<typename T> simdjson_really_inline error_code get(T &out) && noexcept;
+  simdjson_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document &&value) noexcept; ///< @private
+  simdjson_inline simdjson_result(error_code error) noexcept; ///< @private
+  simdjson_inline simdjson_result() noexcept = default;
+  simdjson_inline error_code rewind() noexcept;
+
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array> get_array() & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object> get_object() & noexcept;
+  simdjson_inline simdjson_result<uint64_t> get_uint64() noexcept;
+  simdjson_inline simdjson_result<int64_t> get_int64() noexcept;
+  simdjson_inline simdjson_result<double> get_double() noexcept;
+  simdjson_inline simdjson_result<double> get_double_from_string() noexcept;
+  simdjson_inline simdjson_result<std::string_view> get_string() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string> get_raw_json_string() noexcept;
+  simdjson_inline simdjson_result<bool> get_bool() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> get_value() noexcept;
+  simdjson_inline simdjson_result<bool> is_null() noexcept;
+
+  template<typename T> simdjson_inline simdjson_result<T> get() & noexcept;
+  template<typename T> simdjson_inline simdjson_result<T> get() && noexcept;
+
+  template<typename T> simdjson_inline error_code get(T &out) & noexcept;
+  template<typename T> simdjson_inline error_code get(T &out) && noexcept;
 
 #if SIMDJSON_EXCEPTIONS
-  simdjson_really_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array() & noexcept(false);
-  simdjson_really_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object() & noexcept(false);
-  simdjson_really_inline operator uint64_t() noexcept(false);
-  simdjson_really_inline operator int64_t() noexcept(false);
-  simdjson_really_inline operator double() noexcept(false);
-  simdjson_really_inline operator std::string_view() noexcept(false);
-  simdjson_really_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string() noexcept(false);
-  simdjson_really_inline operator bool() noexcept(false);
-  simdjson_really_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value() noexcept(false);
+  simdjson_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array() & noexcept(false);
+  simdjson_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object() & noexcept(false);
+  simdjson_inline operator uint64_t() noexcept(false);
+  simdjson_inline operator int64_t() noexcept(false);
+  simdjson_inline operator double() noexcept(false);
+  simdjson_inline operator std::string_view() noexcept(false);
+  simdjson_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string() noexcept(false);
+  simdjson_inline operator bool() noexcept(false);
+  simdjson_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value() noexcept(false);
 #endif
-  simdjson_really_inline simdjson_result<size_t> count_elements() & noexcept;
-  simdjson_really_inline simdjson_result<size_t> count_fields() & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> at(size_t index) & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> begin() & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> end() & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field(std::string_view key) & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field(const char *key) & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> operator[](std::string_view key) & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> operator[](const char *key) & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(std::string_view key) & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(const char *key) & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type> type() noexcept;
-  simdjson_really_inline simdjson_result<bool> is_scalar() noexcept;
-  simdjson_really_inline simdjson_result<const char *> current_location() noexcept;
-  simdjson_really_inline int32_t current_depth() const noexcept;
-  simdjson_really_inline bool is_negative() noexcept;
-  simdjson_really_inline simdjson_result<bool> is_integer() noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number_type> get_number_type() noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number> get_number() noexcept;
-  /** @copydoc simdjson_really_inline std::string_view document::raw_json_token() const noexcept */
-  simdjson_really_inline simdjson_result<std::string_view> raw_json_token() noexcept;
-
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> at_pointer(std::string_view json_pointer) noexcept;
+  simdjson_inline simdjson_result<size_t> count_elements() & noexcept;
+  simdjson_inline simdjson_result<size_t> count_fields() & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> at(size_t index) & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> begin() & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> end() & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field(std::string_view key) & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field(const char *key) & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> operator[](std::string_view key) & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> operator[](const char *key) & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(std::string_view key) & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(const char *key) & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type> type() noexcept;
+  simdjson_inline simdjson_result<bool> is_scalar() noexcept;
+  simdjson_inline simdjson_result<const char *> current_location() noexcept;
+  simdjson_inline int32_t current_depth() const noexcept;
+  simdjson_inline bool is_negative() noexcept;
+  simdjson_inline simdjson_result<bool> is_integer() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number_type> get_number_type() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number> get_number() noexcept;
+  /** @copydoc simdjson_inline std::string_view document::raw_json_token() const noexcept */
+  simdjson_inline simdjson_result<std::string_view> raw_json_token() noexcept;
+
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> at_pointer(std::string_view json_pointer) noexcept;
 };
 
 
@@ -25799,55 +25085,55 @@ namespace simdjson {
 template<>
 struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference> : public SIMDJSON_BUILTIN_IMPLEMENTATION::implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference> {
 public:
-  simdjson_really_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference value, error_code error) noexcept;
-  simdjson_really_inline simdjson_result() noexcept = default;
-  simdjson_really_inline error_code rewind() noexcept;
-
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array> get_array() & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object> get_object() & noexcept;
-  simdjson_really_inline simdjson_result<uint64_t> get_uint64() noexcept;
-  simdjson_really_inline simdjson_result<int64_t> get_int64() noexcept;
-  simdjson_really_inline simdjson_result<double> get_double() noexcept;
-  simdjson_really_inline simdjson_result<std::string_view> get_string() noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string> get_raw_json_string() noexcept;
-  simdjson_really_inline simdjson_result<bool> get_bool() noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> get_value() noexcept;
-  simdjson_really_inline bool is_null() noexcept;
+  simdjson_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference value, error_code error) noexcept;
+  simdjson_inline simdjson_result() noexcept = default;
+  simdjson_inline error_code rewind() noexcept;
+
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array> get_array() & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object> get_object() & noexcept;
+  simdjson_inline simdjson_result<uint64_t> get_uint64() noexcept;
+  simdjson_inline simdjson_result<int64_t> get_int64() noexcept;
+  simdjson_inline simdjson_result<double> get_double() noexcept;
+  simdjson_inline simdjson_result<std::string_view> get_string() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string> get_raw_json_string() noexcept;
+  simdjson_inline simdjson_result<bool> get_bool() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> get_value() noexcept;
+  simdjson_inline simdjson_result<bool> is_null() noexcept;
 
 #if SIMDJSON_EXCEPTIONS
-  simdjson_really_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array() & noexcept(false);
-  simdjson_really_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object() & noexcept(false);
-  simdjson_really_inline operator uint64_t() noexcept(false);
-  simdjson_really_inline operator int64_t() noexcept(false);
-  simdjson_really_inline operator double() noexcept(false);
-  simdjson_really_inline operator std::string_view() noexcept(false);
-  simdjson_really_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string() noexcept(false);
-  simdjson_really_inline operator bool() noexcept(false);
-  simdjson_really_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value() noexcept(false);
+  simdjson_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array() & noexcept(false);
+  simdjson_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object() & noexcept(false);
+  simdjson_inline operator uint64_t() noexcept(false);
+  simdjson_inline operator int64_t() noexcept(false);
+  simdjson_inline operator double() noexcept(false);
+  simdjson_inline operator std::string_view() noexcept(false);
+  simdjson_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string() noexcept(false);
+  simdjson_inline operator bool() noexcept(false);
+  simdjson_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value() noexcept(false);
 #endif
-  simdjson_really_inline simdjson_result<size_t> count_elements() & noexcept;
-  simdjson_really_inline simdjson_result<size_t> count_fields() & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> at(size_t index) & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> begin() & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> end() & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field(std::string_view key) & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field(const char *key) & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> operator[](std::string_view key) & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> operator[](const char *key) & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(std::string_view key) & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(const char *key) & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type> type() noexcept;
-  simdjson_really_inline simdjson_result<bool> is_scalar() noexcept;
-  simdjson_really_inline simdjson_result<const char *> current_location() noexcept;
-  simdjson_really_inline int32_t current_depth() const noexcept;
-  simdjson_really_inline bool is_negative() noexcept;
-  simdjson_really_inline simdjson_result<bool> is_integer() noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number_type> get_number_type() noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number> get_number() noexcept;
-  /** @copydoc simdjson_really_inline std::string_view document_reference::raw_json_token() const noexcept */
-  simdjson_really_inline simdjson_result<std::string_view> raw_json_token() noexcept;
-
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> at_pointer(std::string_view json_pointer) noexcept;
+  simdjson_inline simdjson_result<size_t> count_elements() & noexcept;
+  simdjson_inline simdjson_result<size_t> count_fields() & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> at(size_t index) & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> begin() & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> end() & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field(std::string_view key) & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field(const char *key) & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> operator[](std::string_view key) & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> operator[](const char *key) & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(std::string_view key) & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(const char *key) & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type> type() noexcept;
+  simdjson_inline simdjson_result<bool> is_scalar() noexcept;
+  simdjson_inline simdjson_result<const char *> current_location() noexcept;
+  simdjson_inline simdjson_result<int32_t> current_depth() const noexcept;
+  simdjson_inline simdjson_result<bool> is_negative() noexcept;
+  simdjson_inline simdjson_result<bool> is_integer() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number_type> get_number_type() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number> get_number() noexcept;
+  /** @copydoc simdjson_inline std::string_view document_reference::raw_json_token() const noexcept */
+  simdjson_inline simdjson_result<std::string_view> raw_json_token() noexcept;
+
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> at_pointer(std::string_view json_pointer) noexcept;
 };
 
 
@@ -25866,7 +25152,8 @@ class object;
 class raw_json_string;
 
 /**
- * An ephemeral JSON value returned during iteration.
+ * An ephemeral JSON value returned during iteration. It is only valid for as long as you do
+ * not access more data in the JSON document.
  */
 class value {
 public:
@@ -25875,7 +25162,7 @@ class value {
    *
    * Exists so you can declare a variable and later assign to it before use.
    */
-  simdjson_really_inline value() noexcept = default;
+  simdjson_inline value() noexcept = default;
 
   /**
    * Get this value as the given type.
@@ -25888,7 +25175,7 @@ class value {
    * @returns A value of the given type, parsed from the JSON.
    * @returns INCORRECT_TYPE If the JSON value is not the given type.
    */
-  template<typename T> simdjson_really_inline simdjson_result<T> get() noexcept {
+  template<typename T> simdjson_inline simdjson_result<T> get() noexcept {
     // Unless the simdjson library provides an inline implementation, calling this method should
     // immediately fail.
     static_assert(!sizeof(T), "The get method with given type is not implemented by the simdjson library.");
@@ -25903,7 +25190,7 @@ class value {
    * @returns INCORRECT_TYPE If the JSON value is not an object.
    * @returns SUCCESS If the parse succeeded and the out parameter was set to the value.
    */
-  template<typename T> simdjson_really_inline error_code get(T &out) noexcept;
+  template<typename T> simdjson_inline error_code get(T &out) noexcept;
 
   /**
    * Cast this JSON value to an array.
@@ -25911,7 +25198,7 @@ class value {
    * @returns An object that can be used to iterate the array.
    * @returns INCORRECT_TYPE If the JSON value is not an array.
    */
-  simdjson_really_inline simdjson_result<array> get_array() noexcept;
+  simdjson_inline simdjson_result<array> get_array() noexcept;
 
   /**
    * Cast this JSON value to an object.
@@ -25919,7 +25206,7 @@ class value {
    * @returns An object that can be used to look up or iterate fields.
    * @returns INCORRECT_TYPE If the JSON value is not an object.
    */
-  simdjson_really_inline simdjson_result<object> get_object() noexcept;
+  simdjson_inline simdjson_result<object> get_object() noexcept;
 
   /**
    * Cast this JSON value to an unsigned integer.
@@ -25927,7 +25214,7 @@ class value {
    * @returns A unsigned 64-bit integer.
    * @returns INCORRECT_TYPE If the JSON value is not a 64-bit unsigned integer.
    */
-  simdjson_really_inline simdjson_result<uint64_t> get_uint64() noexcept;
+  simdjson_inline simdjson_result<uint64_t> get_uint64() noexcept;
 
   /**
    * Cast this JSON value (inside string) to a unsigned integer.
@@ -25935,7 +25222,7 @@ class value {
    * @returns A unsigned 64-bit integer.
    * @returns INCORRECT_TYPE If the JSON value is not a 64-bit unsigned integer.
    */
-  simdjson_really_inline simdjson_result<uint64_t> get_uint64_in_string() noexcept;
+  simdjson_inline simdjson_result<uint64_t> get_uint64_in_string() noexcept;
 
   /**
    * Cast this JSON value to a signed integer.
@@ -25943,7 +25230,7 @@ class value {
    * @returns A signed 64-bit integer.
    * @returns INCORRECT_TYPE If the JSON value is not a 64-bit integer.
    */
-  simdjson_really_inline simdjson_result<int64_t> get_int64() noexcept;
+  simdjson_inline simdjson_result<int64_t> get_int64() noexcept;
 
   /**
    * Cast this JSON value (inside string) to a signed integer.
@@ -25951,7 +25238,7 @@ class value {
    * @returns A signed 64-bit integer.
    * @returns INCORRECT_TYPE If the JSON value is not a 64-bit integer.
    */
-  simdjson_really_inline simdjson_result<int64_t> get_int64_in_string() noexcept;
+  simdjson_inline simdjson_result<int64_t> get_int64_in_string() noexcept;
 
   /**
    * Cast this JSON value to a double.
@@ -25959,7 +25246,7 @@ class value {
    * @returns A double.
    * @returns INCORRECT_TYPE If the JSON value is not a valid floating-point number.
    */
-  simdjson_really_inline simdjson_result<double> get_double() noexcept;
+  simdjson_inline simdjson_result<double> get_double() noexcept;
 
   /**
    * Cast this JSON value (inside string) to a double
@@ -25967,7 +25254,7 @@ class value {
    * @returns A double.
    * @returns INCORRECT_TYPE If the JSON value is not a valid floating-point number.
    */
-  simdjson_really_inline simdjson_result<double> get_double_in_string() noexcept;
+  simdjson_inline simdjson_result<double> get_double_in_string() noexcept;
 
   /**
    * Cast this JSON value to a string.
@@ -25983,7 +25270,7 @@ class value {
    *          time it parses a document or when it is destroyed.
    * @returns INCORRECT_TYPE if the JSON value is not a string.
    */
-  simdjson_really_inline simdjson_result<std::string_view> get_string() noexcept;
+  simdjson_inline simdjson_result<std::string_view> get_string() noexcept;
 
   /**
    * Cast this JSON value to a raw_json_string.
@@ -25993,7 +25280,7 @@ class value {
    * @returns A pointer to the raw JSON for the given string.
    * @returns INCORRECT_TYPE if the JSON value is not a string.
    */
-  simdjson_really_inline simdjson_result<raw_json_string> get_raw_json_string() noexcept;
+  simdjson_inline simdjson_result<raw_json_string> get_raw_json_string() noexcept;
 
   /**
    * Cast this JSON value to a bool.
@@ -26001,14 +25288,17 @@ class value {
    * @returns A bool value.
    * @returns INCORRECT_TYPE if the JSON value is not true or false.
    */
-  simdjson_really_inline simdjson_result<bool> get_bool() noexcept;
+  simdjson_inline simdjson_result<bool> get_bool() noexcept;
 
   /**
-   * Checks if this JSON value is null.
+   * Checks if this JSON value is null. If and only if the value is
+   * null, then it is consumed (we advance). If we find a token that
+   * begins with 'n' but is not 'null', then an error is returned.
    *
    * @returns Whether the value is null.
+   * @returns INCORRECT_TYPE If the JSON value begins with 'n' and is not 'null'.
    */
-  simdjson_really_inline bool is_null() noexcept;
+  simdjson_inline simdjson_result<bool> is_null() noexcept;
 
 #if SIMDJSON_EXCEPTIONS
   /**
@@ -26017,35 +25307,35 @@ class value {
    * @returns An object that can be used to iterate the array.
    * @exception simdjson_error(INCORRECT_TYPE) If the JSON value is not an array.
    */
-  simdjson_really_inline operator array() noexcept(false);
+  simdjson_inline operator array() noexcept(false);
   /**
    * Cast this JSON value to an object.
    *
    * @returns An object that can be used to look up or iterate fields.
    * @exception simdjson_error(INCORRECT_TYPE) If the JSON value is not an object.
    */
-  simdjson_really_inline operator object() noexcept(false);
+  simdjson_inline operator object() noexcept(false);
   /**
    * Cast this JSON value to an unsigned integer.
    *
    * @returns A signed 64-bit integer.
    * @exception simdjson_error(INCORRECT_TYPE) If the JSON value is not a 64-bit unsigned integer.
    */
-  simdjson_really_inline operator uint64_t() noexcept(false);
+  simdjson_inline operator uint64_t() noexcept(false);
   /**
    * Cast this JSON value to a signed integer.
    *
    * @returns A signed 64-bit integer.
    * @exception simdjson_error(INCORRECT_TYPE) If the JSON value is not a 64-bit integer.
    */
-  simdjson_really_inline operator int64_t() noexcept(false);
+  simdjson_inline operator int64_t() noexcept(false);
   /**
    * Cast this JSON value to a double.
    *
    * @returns A double.
    * @exception simdjson_error(INCORRECT_TYPE) If the JSON value is not a valid floating-point number.
    */
-  simdjson_really_inline operator double() noexcept(false);
+  simdjson_inline operator double() noexcept(false);
   /**
    * Cast this JSON value to a string.
    *
@@ -26057,7 +25347,7 @@ class value {
    *          time it parses a document or when it is destroyed.
    * @exception simdjson_error(INCORRECT_TYPE) if the JSON value is not a string.
    */
-  simdjson_really_inline operator std::string_view() noexcept(false);
+  simdjson_inline operator std::string_view() noexcept(false);
   /**
    * Cast this JSON value to a raw_json_string.
    *
@@ -26066,14 +25356,14 @@ class value {
    * @returns A pointer to the raw JSON for the given string.
    * @exception simdjson_error(INCORRECT_TYPE) if the JSON value is not a string.
    */
-  simdjson_really_inline operator raw_json_string() noexcept(false);
+  simdjson_inline operator raw_json_string() noexcept(false);
   /**
    * Cast this JSON value to a bool.
    *
    * @returns A bool value.
    * @exception simdjson_error(INCORRECT_TYPE) if the JSON value is not true or false.
    */
-  simdjson_really_inline operator bool() noexcept(false);
+  simdjson_inline operator bool() noexcept(false);
 #endif
 
   /**
@@ -26083,13 +25373,13 @@ class value {
    *
    * @returns INCORRECT_TYPE If the JSON value is not an array.
    */
-  simdjson_really_inline simdjson_result<array_iterator> begin() & noexcept;
+  simdjson_inline simdjson_result<array_iterator> begin() & noexcept;
   /**
    * Sentinel representing the end of the array.
    *
    * Part of the std::iterable interface.
    */
-  simdjson_really_inline simdjson_result<array_iterator> end() & noexcept;
+  simdjson_inline simdjson_result<array_iterator> end() & noexcept;
   /**
    * This method scans the array and counts the number of elements.
    * The count_elements method should always be called before you have begun
@@ -26100,8 +25390,11 @@ class value {
    * beginning as if it had never been accessed. If the JSON is malformed (e.g.,
    * there is a missing comma), then an error is returned and it is no longer
    * safe to continue.
+   *
+   * Performance hint: You should only call count_elements() as a last
+   * resort as it may require scanning the document twice or more.
    */
-  simdjson_really_inline simdjson_result<size_t> count_elements() & noexcept;
+  simdjson_inline simdjson_result<size_t> count_elements() & noexcept;
   /**
    * This method scans the object and counts the number of key-value pairs.
    * The count_fields method should always be called before you have begun
@@ -26115,8 +25408,11 @@ class value {
    *
    * To check that an object is empty, it is more performant to use
    * the is_empty() method on the object instance.
+   *
+   * Performance hint: You should only call count_fields() as a last
+   * resort as it may require scanning the document twice or more.
    */
-  simdjson_really_inline simdjson_result<size_t> count_fields() & noexcept;
+  simdjson_inline simdjson_result<size_t> count_fields() & noexcept;
   /**
    * Get the value at the given index in the array. This function has linear-time complexity.
    * This function should only be called once on an array instance since the array iterator is not reset between each call.
@@ -26124,7 +25420,7 @@ class value {
    * @return The value at the given index, or:
    *         - INDEX_OUT_OF_BOUNDS if the array index is larger than an array length
    */
-  simdjson_really_inline simdjson_result<value> at(size_t index) noexcept;
+  simdjson_inline simdjson_result<value> at(size_t index) noexcept;
   /**
    * Look up a field by name on an object (order-sensitive).
    *
@@ -26147,9 +25443,9 @@ class value {
    * @param key The key to look up.
    * @returns The value of the field, or NO_SUCH_FIELD if the field is not in the object.
    */
-  simdjson_really_inline simdjson_result<value> find_field(std::string_view key) noexcept;
-  /** @overload simdjson_really_inline simdjson_result<value> find_field(std::string_view key) noexcept; */
-  simdjson_really_inline simdjson_result<value> find_field(const char *key) noexcept;
+  simdjson_inline simdjson_result<value> find_field(std::string_view key) noexcept;
+  /** @overload simdjson_inline simdjson_result<value> find_field(std::string_view key) noexcept; */
+  simdjson_inline simdjson_result<value> find_field(const char *key) noexcept;
 
   /**
    * Look up a field by name on an object, without regard to key order.
@@ -26173,16 +25469,18 @@ class value {
    * @param key The key to look up.
    * @returns The value of the field, or NO_SUCH_FIELD if the field is not in the object.
    */
-  simdjson_really_inline simdjson_result<value> find_field_unordered(std::string_view key) noexcept;
-  /** @overload simdjson_really_inline simdjson_result<value> find_field_unordered(std::string_view key) noexcept; */
-  simdjson_really_inline simdjson_result<value> find_field_unordered(const char *key) noexcept;
-  /** @overload simdjson_really_inline simdjson_result<value> find_field_unordered(std::string_view key) noexcept; */
-  simdjson_really_inline simdjson_result<value> operator[](std::string_view key) noexcept;
-  /** @overload simdjson_really_inline simdjson_result<value> find_field_unordered(std::string_view key) noexcept; */
-  simdjson_really_inline simdjson_result<value> operator[](const char *key) noexcept;
+  simdjson_inline simdjson_result<value> find_field_unordered(std::string_view key) noexcept;
+  /** @overload simdjson_inline simdjson_result<value> find_field_unordered(std::string_view key) noexcept; */
+  simdjson_inline simdjson_result<value> find_field_unordered(const char *key) noexcept;
+  /** @overload simdjson_inline simdjson_result<value> find_field_unordered(std::string_view key) noexcept; */
+  simdjson_inline simdjson_result<value> operator[](std::string_view key) noexcept;
+  /** @overload simdjson_inline simdjson_result<value> find_field_unordered(std::string_view key) noexcept; */
+  simdjson_inline simdjson_result<value> operator[](const char *key) noexcept;
 
   /**
-   * Get the type of this JSON value.
+   * Get the type of this JSON value. It does not validate or consume the value.
+   * E.g., you must still call "is_null()" to check that a value is null even if
+   * "type()" returns json_type::null.
    *
    * NOTE: If you're only expecting a value to be one type (a typical case), it's generally
    * better to just call .get_double, .get_string, etc. and check for INCORRECT_TYPE (or just
@@ -26192,7 +25490,7 @@ class value {
    *     json_type::number, json_type::boolean, or json_type::null).
    * @error TAPE_ERROR when the JSON value is a bad token like "}" "," or "alse".
    */
-  simdjson_really_inline simdjson_result<json_type> type() noexcept;
+  simdjson_inline simdjson_result<json_type> type() noexcept;
 
   /**
    * Checks whether the value is a scalar (string, number, null, Boolean).
@@ -26201,14 +25499,14 @@ class value {
    * @returns true if the type is string, number, null, Boolean
    * @error TAPE_ERROR when the JSON value is a bad token like "}" "," or "alse".
    */
-  simdjson_really_inline simdjson_result<bool> is_scalar() noexcept;
+  simdjson_inline simdjson_result<bool> is_scalar() noexcept;
 
   /**
    * Checks whether the value is a negative number.
    *
    * @returns true if the number if negative.
    */
-  simdjson_really_inline bool is_negative() noexcept;
+  simdjson_inline bool is_negative() noexcept;
   /**
    * Checks whether the value is an integer number. Note that
    * this requires to partially parse the number string. If
@@ -26222,7 +25520,7 @@ class value {
    *
    * @returns true if the number if negative.
    */
-  simdjson_really_inline simdjson_result<bool> is_integer() noexcept;
+  simdjson_inline simdjson_result<bool> is_integer() noexcept;
   /**
    * Determine the number type (integer or floating-point number) as quickly
    * as possible. This function does not fully validate the input. It is
@@ -26245,7 +25543,7 @@ class value {
    *
    * @returns the type of the number
    */
-  simdjson_really_inline simdjson_result<number_type> get_number_type() noexcept;
+  simdjson_inline simdjson_result<number_type> get_number_type() noexcept;
 
   /**
    * Attempt to parse an ondemand::number. An ondemand::number may
@@ -26277,7 +25575,7 @@ class value {
    * calling 'get_number()', you scan the number string only once, determining
    * efficiently the type and storing it in an efficient manner.
    */
-  simdjson_warn_unused simdjson_really_inline simdjson_result<number> get_number() noexcept;
+  simdjson_warn_unused simdjson_inline simdjson_result<number> get_number() noexcept;
 
 
   /**
@@ -26303,12 +25601,12 @@ class value {
    * - false
    * - null
    */
-  simdjson_really_inline std::string_view raw_json_token() noexcept;
+  simdjson_inline std::string_view raw_json_token() noexcept;
 
   /**
    * Returns the current location in the document if in bounds.
    */
-  simdjson_really_inline simdjson_result<const char *> current_location() noexcept;
+  simdjson_inline simdjson_result<const char *> current_location() noexcept;
 
   /**
    * Returns the current depth in the document if in bounds.
@@ -26319,7 +25617,7 @@ class value {
    *  2 = , or } inside root array/object
    *  3 = key or value inside root array/object.
    */
-  simdjson_really_inline int32_t current_depth() const noexcept;
+  simdjson_inline int32_t current_depth() const noexcept;
 
   /**
    * Get the value associated with the given JSON pointer.  We use the RFC 6901
@@ -26363,38 +25661,38 @@ class value {
    *         - INCORRECT_TYPE if a non-integer is used to access an array
    *         - INVALID_JSON_POINTER if the JSON pointer is invalid and cannot be parsed
    */
-  simdjson_really_inline simdjson_result<value> at_pointer(std::string_view json_pointer) noexcept;
+  simdjson_inline simdjson_result<value> at_pointer(std::string_view json_pointer) noexcept;
 
 protected:
   /**
    * Create a value.
    */
-  simdjson_really_inline value(const value_iterator &iter) noexcept;
+  simdjson_inline value(const value_iterator &iter) noexcept;
 
   /**
    * Skip this value, allowing iteration to continue.
    */
-  simdjson_really_inline void skip() noexcept;
+  simdjson_inline void skip() noexcept;
 
   /**
    * Start a value at the current position.
    *
    * (It should already be started; this is just a self-documentation method.)
    */
-  static simdjson_really_inline value start(const value_iterator &iter) noexcept;
+  static simdjson_inline value start(const value_iterator &iter) noexcept;
 
   /**
    * Resume a value.
    */
-  static simdjson_really_inline value resume(const value_iterator &iter) noexcept;
+  static simdjson_inline value resume(const value_iterator &iter) noexcept;
 
   /**
    * Get the object, starting or resuming it as necessary
    */
-  simdjson_really_inline simdjson_result<object> start_or_resume_object() noexcept;
+  simdjson_inline simdjson_result<object> start_or_resume_object() noexcept;
 
-  // simdjson_really_inline void log_value(const char *type) const noexcept;
-  // simdjson_really_inline void log_error(const char *message) const noexcept;
+  // simdjson_inline void log_value(const char *type) const noexcept;
+  // simdjson_inline void log_error(const char *message) const noexcept;
 
   value_iterator iter{};
 
@@ -26415,43 +25713,43 @@ namespace simdjson {
 template<>
 struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> : public SIMDJSON_BUILTIN_IMPLEMENTATION::implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> {
 public:
-  simdjson_really_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value &&value) noexcept; ///< @private
-  simdjson_really_inline simdjson_result(error_code error) noexcept; ///< @private
-  simdjson_really_inline simdjson_result() noexcept = default;
+  simdjson_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value &&value) noexcept; ///< @private
+  simdjson_inline simdjson_result(error_code error) noexcept; ///< @private
+  simdjson_inline simdjson_result() noexcept = default;
 
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array> get_array() noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object> get_object() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array> get_array() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object> get_object() noexcept;
 
-  simdjson_really_inline simdjson_result<uint64_t> get_uint64() noexcept;
-  simdjson_really_inline simdjson_result<uint64_t> get_uint64_in_string() noexcept;
-  simdjson_really_inline simdjson_result<int64_t> get_int64() noexcept;
-  simdjson_really_inline simdjson_result<int64_t> get_int64_in_string() noexcept;
-  simdjson_really_inline simdjson_result<double> get_double() noexcept;
-  simdjson_really_inline simdjson_result<double> get_double_in_string() noexcept;
-  simdjson_really_inline simdjson_result<std::string_view> get_string() noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string> get_raw_json_string() noexcept;
-  simdjson_really_inline simdjson_result<bool> get_bool() noexcept;
-  simdjson_really_inline bool is_null() noexcept;
+  simdjson_inline simdjson_result<uint64_t> get_uint64() noexcept;
+  simdjson_inline simdjson_result<uint64_t> get_uint64_in_string() noexcept;
+  simdjson_inline simdjson_result<int64_t> get_int64() noexcept;
+  simdjson_inline simdjson_result<int64_t> get_int64_in_string() noexcept;
+  simdjson_inline simdjson_result<double> get_double() noexcept;
+  simdjson_inline simdjson_result<double> get_double_in_string() noexcept;
+  simdjson_inline simdjson_result<std::string_view> get_string() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string> get_raw_json_string() noexcept;
+  simdjson_inline simdjson_result<bool> get_bool() noexcept;
+  simdjson_inline simdjson_result<bool> is_null() noexcept;
 
-  template<typename T> simdjson_really_inline simdjson_result<T> get() noexcept;
+  template<typename T> simdjson_inline simdjson_result<T> get() noexcept;
 
-  template<typename T> simdjson_really_inline error_code get(T &out) noexcept;
+  template<typename T> simdjson_inline error_code get(T &out) noexcept;
 
 #if SIMDJSON_EXCEPTIONS
-  simdjson_really_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array() noexcept(false);
-  simdjson_really_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object() noexcept(false);
-  simdjson_really_inline operator uint64_t() noexcept(false);
-  simdjson_really_inline operator int64_t() noexcept(false);
-  simdjson_really_inline operator double() noexcept(false);
-  simdjson_really_inline operator std::string_view() noexcept(false);
-  simdjson_really_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string() noexcept(false);
-  simdjson_really_inline operator bool() noexcept(false);
+  simdjson_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array() noexcept(false);
+  simdjson_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object() noexcept(false);
+  simdjson_inline operator uint64_t() noexcept(false);
+  simdjson_inline operator int64_t() noexcept(false);
+  simdjson_inline operator double() noexcept(false);
+  simdjson_inline operator std::string_view() noexcept(false);
+  simdjson_inline operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string() noexcept(false);
+  simdjson_inline operator bool() noexcept(false);
 #endif
-  simdjson_really_inline simdjson_result<size_t> count_elements() & noexcept;
-  simdjson_really_inline simdjson_result<size_t> count_fields() & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> at(size_t index) noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> begin() & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> end() & noexcept;
+  simdjson_inline simdjson_result<size_t> count_elements() & noexcept;
+  simdjson_inline simdjson_result<size_t> count_fields() & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> at(size_t index) noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> begin() & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> end() & noexcept;
 
   /**
    * Look up a field by name on an object (order-sensitive).
@@ -26473,9 +25771,9 @@ struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> : publi
    * @param key The key to look up.
    * @returns The value of the field, or NO_SUCH_FIELD if the field is not in the object.
    */
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field(std::string_view key) noexcept;
-  /** @overload simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field(std::string_view key) noexcept; */
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field(const char *key) noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field(std::string_view key) noexcept;
+  /** @overload simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field(std::string_view key) noexcept; */
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field(const char *key) noexcept;
 
   /**
    * Look up a field by name on an object, without regard to key order.
@@ -26496,13 +25794,13 @@ struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> : publi
    * @param key The key to look up.
    * @returns The value of the field, or NO_SUCH_FIELD if the field is not in the object.
    */
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(std::string_view key) noexcept;
-  /** @overload simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(std::string_view key) noexcept; */
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(const char *key) noexcept;
-  /** @overload simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(std::string_view key) noexcept; */
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> operator[](std::string_view key) noexcept;
-  /** @overload simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(std::string_view key) noexcept; */
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> operator[](const char *key) noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(std::string_view key) noexcept;
+  /** @overload simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(std::string_view key) noexcept; */
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(const char *key) noexcept;
+  /** @overload simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(std::string_view key) noexcept; */
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> operator[](std::string_view key) noexcept;
+  /** @overload simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(std::string_view key) noexcept; */
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> operator[](const char *key) noexcept;
 
   /**
    * Get the type of this JSON value.
@@ -26511,21 +25809,21 @@ struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> : publi
    * better to just call .get_double, .get_string, etc. and check for INCORRECT_TYPE (or just
    * let it throw an exception).
    */
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type> type() noexcept;
-  simdjson_really_inline simdjson_result<bool> is_scalar() noexcept;
-  simdjson_really_inline simdjson_result<bool> is_negative() noexcept;
-  simdjson_really_inline simdjson_result<bool> is_integer() noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number_type> get_number_type() noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number> get_number() noexcept;
-
-  /** @copydoc simdjson_really_inline std::string_view value::raw_json_token() const noexcept */
-  simdjson_really_inline simdjson_result<std::string_view> raw_json_token() noexcept;
-
-  /** @copydoc simdjson_really_inline simdjson_result<const char *> current_location() noexcept */
-  simdjson_really_inline simdjson_result<const char *> current_location() noexcept;
-  /** @copydoc simdjson_really_inline int32_t current_depth() const noexcept */
-  simdjson_really_inline int32_t current_depth() const noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> at_pointer(std::string_view json_pointer) noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type> type() noexcept;
+  simdjson_inline simdjson_result<bool> is_scalar() noexcept;
+  simdjson_inline simdjson_result<bool> is_negative() noexcept;
+  simdjson_inline simdjson_result<bool> is_integer() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number_type> get_number_type() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number> get_number() noexcept;
+
+  /** @copydoc simdjson_inline std::string_view value::raw_json_token() const noexcept */
+  simdjson_inline simdjson_result<std::string_view> raw_json_token() noexcept;
+
+  /** @copydoc simdjson_inline simdjson_result<const char *> current_location() noexcept */
+  simdjson_inline simdjson_result<const char *> current_location() noexcept;
+  /** @copydoc simdjson_inline int32_t current_depth() const noexcept */
+  simdjson_inline simdjson_result<int32_t> current_depth() const noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> at_pointer(std::string_view json_pointer) noexcept;
 };
 
 } // namespace simdjson
@@ -26550,7 +25848,7 @@ class field : public std::pair<raw_json_string, value> {
    *
    * Exists so you can declare a variable and later assign to it before use.
    */
-  simdjson_really_inline field() noexcept;
+  simdjson_inline field() noexcept;
 
   /**
    * Get the key as a string_view (for higher speed, consider raw_key).
@@ -26560,25 +25858,25 @@ class field : public std::pair<raw_json_string, value> {
    * This consumes the key: once you have called unescaped_key(), you cannot
    * call it again nor can you call key().
    */
-  simdjson_really_inline simdjson_warn_unused simdjson_result<std::string_view> unescaped_key() noexcept;
+  simdjson_inline simdjson_warn_unused simdjson_result<std::string_view> unescaped_key() noexcept;
   /**
    * Get the key as a raw_json_string. Can be used for direct comparison with
    * an unescaped C string: e.g., key() == "test".
    */
-  simdjson_really_inline raw_json_string key() const noexcept;
+  simdjson_inline raw_json_string key() const noexcept;
   /**
    * Get the field value.
    */
-  simdjson_really_inline ondemand::value &value() & noexcept;
+  simdjson_inline ondemand::value &value() & noexcept;
   /**
    * @overload ondemand::value &ondemand::value() & noexcept
    */
-  simdjson_really_inline ondemand::value value() && noexcept;
+  simdjson_inline ondemand::value value() && noexcept;
 
 protected:
-  simdjson_really_inline field(raw_json_string key, ondemand::value &&value) noexcept;
-  static simdjson_really_inline simdjson_result<field> start(value_iterator &parent_iter) noexcept;
-  static simdjson_really_inline simdjson_result<field> start(const value_iterator &parent_iter, raw_json_string key) noexcept;
+  simdjson_inline field(raw_json_string key, ondemand::value &&value) noexcept;
+  static simdjson_inline simdjson_result<field> start(value_iterator &parent_iter) noexcept;
+  static simdjson_inline simdjson_result<field> start(const value_iterator &parent_iter, raw_json_string key) noexcept;
   friend struct simdjson_result<field>;
   friend class object_iterator;
 };
@@ -26592,13 +25890,13 @@ namespace simdjson {
 template<>
 struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field> : public SIMDJSON_BUILTIN_IMPLEMENTATION::implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field> {
 public:
-  simdjson_really_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field &&value) noexcept; ///< @private
-  simdjson_really_inline simdjson_result(error_code error) noexcept; ///< @private
-  simdjson_really_inline simdjson_result() noexcept = default;
+  simdjson_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field &&value) noexcept; ///< @private
+  simdjson_inline simdjson_result(error_code error) noexcept; ///< @private
+  simdjson_inline simdjson_result() noexcept = default;
 
-  simdjson_really_inline simdjson_result<std::string_view> unescaped_key() noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string> key() noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> value() noexcept;
+  simdjson_inline simdjson_result<std::string_view> unescaped_key() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string> key() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> value() noexcept;
 };
 
 } // namespace simdjson
@@ -26619,10 +25917,10 @@ class object {
    *
    * Exists so you can declare a variable and later assign to it before use.
    */
-  simdjson_really_inline object() noexcept = default;
+  simdjson_inline object() noexcept = default;
 
-  simdjson_really_inline simdjson_result<object_iterator> begin() noexcept;
-  simdjson_really_inline simdjson_result<object_iterator> end() noexcept;
+  simdjson_inline simdjson_result<object_iterator> begin() noexcept;
+  simdjson_inline simdjson_result<object_iterator> end() noexcept;
   /**
    * Look up a field by name on an object (order-sensitive).
    *
@@ -26643,7 +25941,8 @@ class object {
    * e.g. `object["a"]` will match `{ "a": 1 }`, but will *not* match `{ "\u0061": 1 }`.
    *
    * You must consume the fields on an object one at a time. A request for a new key
-   * invalidates previous field values: it makes them unsafe. E.g., the array
+   * invalidates previous field values: it makes them unsafe. The value instance you get
+   * from  `content["bids"]` becomes invalid when you call `content["asks"]`. The array
    * given by content["bids"].get_array() should not be accessed after you have called
    * content["asks"].get_array(). You can detect such mistakes by first compiling and running
    * the code in Debug mode (or with the macro `SIMDJSON_DEVELOPMENT_CHECKS` set to 1): an
@@ -26656,9 +25955,9 @@ class object {
    * @param key The key to look up.
    * @returns The value of the field, or NO_SUCH_FIELD if the field is not in the object.
    */
-  simdjson_really_inline simdjson_result<value> find_field(std::string_view key) & noexcept;
-  /** @overload simdjson_really_inline simdjson_result<value> find_field(std::string_view key) & noexcept; */
-  simdjson_really_inline simdjson_result<value> find_field(std::string_view key) && noexcept;
+  simdjson_inline simdjson_result<value> find_field(std::string_view key) & noexcept;
+  /** @overload simdjson_inline simdjson_result<value> find_field(std::string_view key) & noexcept; */
+  simdjson_inline simdjson_result<value> find_field(std::string_view key) && noexcept;
 
   /**
    * Look up a field by name on an object, without regard to key order.
@@ -26680,7 +25979,8 @@ class object {
    * that only one field is returned.
    *
    * You must consume the fields on an object one at a time. A request for a new key
-   * invalidates previous field values: it makes them unsafe. E.g., the array
+   * invalidates previous field values: it makes them unsafe. The value instance you get
+   * from  `content["bids"]` becomes invalid when you call `content["asks"]`. The array
    * given by content["bids"].get_array() should not be accessed after you have called
    * content["asks"].get_array(). You can detect such mistakes by first compiling and running
    * the code in Debug mode (or with the macro `SIMDJSON_DEVELOPMENT_CHECKS` set to 1): an
@@ -26692,13 +25992,13 @@ class object {
    * @param key The key to look up.
    * @returns The value of the field, or NO_SUCH_FIELD if the field is not in the object.
    */
-  simdjson_really_inline simdjson_result<value> find_field_unordered(std::string_view key) & noexcept;
-  /** @overload simdjson_really_inline simdjson_result<value> find_field_unordered(std::string_view key) & noexcept; */
-  simdjson_really_inline simdjson_result<value> find_field_unordered(std::string_view key) && noexcept;
-  /** @overload simdjson_really_inline simdjson_result<value> find_field_unordered(std::string_view key) & noexcept; */
-  simdjson_really_inline simdjson_result<value> operator[](std::string_view key) & noexcept;
-  /** @overload simdjson_really_inline simdjson_result<value> find_field_unordered(std::string_view key) & noexcept; */
-  simdjson_really_inline simdjson_result<value> operator[](std::string_view key) && noexcept;
+  simdjson_inline simdjson_result<value> find_field_unordered(std::string_view key) & noexcept;
+  /** @overload simdjson_inline simdjson_result<value> find_field_unordered(std::string_view key) & noexcept; */
+  simdjson_inline simdjson_result<value> find_field_unordered(std::string_view key) && noexcept;
+  /** @overload simdjson_inline simdjson_result<value> find_field_unordered(std::string_view key) & noexcept; */
+  simdjson_inline simdjson_result<value> operator[](std::string_view key) & noexcept;
+  /** @overload simdjson_inline simdjson_result<value> find_field_unordered(std::string_view key) & noexcept; */
+  simdjson_inline simdjson_result<value> operator[](std::string_view key) && noexcept;
 
   /**
    * Get the value associated with the given JSON pointer. We use the RFC 6901
@@ -26769,27 +26069,30 @@ class object {
    *
    * To check that an object is empty, it is more performant to use
    * the is_empty() method.
+   *
+   * Performance hint: You should only call count_fields() as a last
+   * resort as it may require scanning the document twice or more.
    */
-  simdjson_really_inline simdjson_result<size_t> count_fields() & noexcept;
+  simdjson_inline simdjson_result<size_t> count_fields() & noexcept;
   /**
    * Consumes the object and returns a string_view instance corresponding to the
    * object as represented in JSON. It points inside the original byte array containing
    * the JSON document.
    */
-  simdjson_really_inline simdjson_result<std::string_view> raw_json() noexcept;
+  simdjson_inline simdjson_result<std::string_view> raw_json() noexcept;
 
 protected:
   /**
    * Go to the end of the object, no matter where you are right now.
    */
-  simdjson_really_inline error_code consume() noexcept;
-  static simdjson_really_inline simdjson_result<object> start(value_iterator &iter) noexcept;
-  static simdjson_really_inline simdjson_result<object> start_root(value_iterator &iter) noexcept;
-  static simdjson_really_inline simdjson_result<object> started(value_iterator &iter) noexcept;
-  static simdjson_really_inline object resume(const value_iterator &iter) noexcept;
-  simdjson_really_inline object(const value_iterator &iter) noexcept;
+  simdjson_inline error_code consume() noexcept;
+  static simdjson_inline simdjson_result<object> start(value_iterator &iter) noexcept;
+  static simdjson_inline simdjson_result<object> start_root(value_iterator &iter) noexcept;
+  static simdjson_inline simdjson_result<object> started(value_iterator &iter) noexcept;
+  static simdjson_inline object resume(const value_iterator &iter) noexcept;
+  simdjson_inline object(const value_iterator &iter) noexcept;
 
-  simdjson_warn_unused simdjson_really_inline error_code find_field_raw(const std::string_view key) noexcept;
+  simdjson_warn_unused simdjson_inline error_code find_field_raw(const std::string_view key) noexcept;
 
   value_iterator iter{};
 
@@ -26807,19 +26110,19 @@ namespace simdjson {
 template<>
 struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object> : public SIMDJSON_BUILTIN_IMPLEMENTATION::implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object> {
 public:
-  simdjson_really_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object &&value) noexcept; ///< @private
-  simdjson_really_inline simdjson_result(error_code error) noexcept; ///< @private
-  simdjson_really_inline simdjson_result() noexcept = default;
-
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> begin() noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> end() noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field(std::string_view key) & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field(std::string_view key) && noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(std::string_view key) & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(std::string_view key) && noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> operator[](std::string_view key) & noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> operator[](std::string_view key) && noexcept;
-  simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> at_pointer(std::string_view json_pointer) noexcept;
+  simdjson_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object &&value) noexcept; ///< @private
+  simdjson_inline simdjson_result(error_code error) noexcept; ///< @private
+  simdjson_inline simdjson_result() noexcept = default;
+
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> begin() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> end() noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field(std::string_view key) & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field(std::string_view key) && noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(std::string_view key) & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> find_field_unordered(std::string_view key) && noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> operator[](std::string_view key) & noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> operator[](std::string_view key) && noexcept;
+  simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> at_pointer(std::string_view json_pointer) noexcept;
   inline simdjson_result<bool> reset() noexcept;
   inline simdjson_result<bool> is_empty() noexcept;
   inline simdjson_result<size_t> count_fields() & noexcept;
@@ -26871,9 +26174,9 @@ class parser {
   inline explicit parser(size_t max_capacity = SIMDJSON_MAXSIZE_BYTES) noexcept;
 
   inline parser(parser &&other) noexcept = default;
-  simdjson_really_inline parser(const parser &other) = delete;
-  simdjson_really_inline parser &operator=(const parser &other) = delete;
-  simdjson_really_inline parser &operator=(parser &&other) noexcept = default;
+  simdjson_inline parser(const parser &other) = delete;
+  simdjson_inline parser &operator=(const parser &other) = delete;
+  simdjson_inline parser &operator=(parser &&other) noexcept = default;
 
   /** Deallocate the JSON parser. */
   inline ~parser() noexcept = default;
@@ -27056,17 +26359,26 @@ class parser {
   simdjson_result<document_stream> iterate_many(const char *buf, size_t batch_size = DEFAULT_BATCH_SIZE) noexcept = delete;
 
   /** The capacity of this parser (the largest document it can process). */
-  simdjson_really_inline size_t capacity() const noexcept;
+  simdjson_inline size_t capacity() const noexcept;
   /** The maximum capacity of this parser (the largest document it is allowed to process). */
-  simdjson_really_inline size_t max_capacity() const noexcept;
-  simdjson_really_inline void set_max_capacity(size_t max_capacity) noexcept;
-  /** The maximum depth of this parser (the most deeply nested objects and arrays it can process). */
-  simdjson_really_inline size_t max_depth() const noexcept;
+  simdjson_inline size_t max_capacity() const noexcept;
+  simdjson_inline void set_max_capacity(size_t max_capacity) noexcept;
+  /**
+   * The maximum depth of this parser (the most deeply nested objects and arrays it can process).
+   * This parameter is only relevant when the macro SIMDJSON_DEVELOPMENT_CHECKS is set to true.
+   * The document's instance current_depth() method should be used to monitor the parsing
+   * depth and limit it if desired.
+   */
+  simdjson_inline size_t max_depth() const noexcept;
 
   /**
    * Ensure this parser has enough memory to process JSON documents up to `capacity` bytes in length
    * and `max_depth` depth.
    *
+   * The max_depth parameter is only relevant when the macro SIMDJSON_DEVELOPMENT_CHECKS is set to true.
+   * The document's instance current_depth() method should be used to monitor the parsing
+   * depth and limit it if desired.
+   *
    * @param capacity The new capacity.
    * @param max_depth The new max_depth. Defaults to DEFAULT_MAX_DEPTH.
    * @return The error, if there is one.
@@ -27082,6 +26394,27 @@ class parser {
   bool threaded{true};
   #endif
 
+  /**
+   * Unescape this JSON string, replacing \\ with \, \n with newline, etc. to a user-provided buffer.
+   * The provided pointer is advanced to the end of the string by reference, and a string_view instance
+   * is returned. You can ensure that your buffer is large enough by allocating a block of memory at least
+   * as large as the input JSON plus SIMDJSON_PADDING and then unescape all strings to this one buffer.
+   *
+   * This unescape function is a low-level function. If you want a more user-friendly approach, you should
+   * avoid raw_json_string instances (e.g., by calling unescaped_key() instead of key() or get_string()
+   * instead of get_raw_json_string()).
+   *
+   * ## IMPORTANT: string_view lifetime
+   *
+   * The string_view is only valid as long as the bytes in dst.
+   *
+   * @param raw_json_string input
+   * @param dst A pointer to a buffer at least large enough to write this string as well as
+   *            an additional SIMDJSON_PADDING bytes.
+   * @return A string_view pointing at the unescaped string in dst
+   * @error STRING_ERROR if escapes are incorrect.
+   */
+  simdjson_inline simdjson_result<std::string_view> unescape(raw_json_string in, uint8_t *&dst) const noexcept;
 private:
   /** @private [for benchmarking access] The implementation to use */
   std::unique_ptr<internal::dom_parser_implementation> implementation{};
@@ -27089,7 +26422,7 @@ class parser {
   size_t _max_capacity;
   size_t _max_depth{DEFAULT_MAX_DEPTH};
   std::unique_ptr<uint8_t[]> string_buf{};
-#ifdef SIMDJSON_DEVELOPMENT_CHECKS
+#if SIMDJSON_DEVELOPMENT_CHECKS
   std::unique_ptr<token_position[]> start_positions{};
 #endif
 
@@ -27106,9 +26439,9 @@ namespace simdjson {
 template<>
 struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::parser> : public SIMDJSON_BUILTIN_IMPLEMENTATION::implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::parser> {
 public:
-  simdjson_really_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::parser &&value) noexcept; ///< @private
-  simdjson_really_inline simdjson_result(error_code error) noexcept; ///< @private
-  simdjson_really_inline simdjson_result() noexcept = default;
+  simdjson_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::parser &&value) noexcept; ///< @private
+  simdjson_inline simdjson_result(error_code error) noexcept; ///< @private
+  simdjson_inline simdjson_result() noexcept = default;
 };
 
 } // namespace simdjson
@@ -27196,13 +26529,13 @@ class document_stream {
    *  auto error = parser.iterate_many(json).get(docs);
    *  ```
    */
-  simdjson_really_inline document_stream() noexcept;
+  simdjson_inline document_stream() noexcept;
   /** Move one document_stream to another. */
-  simdjson_really_inline document_stream(document_stream &&other) noexcept = default;
+  simdjson_inline document_stream(document_stream &&other) noexcept = default;
   /** Move one document_stream to another. */
-  simdjson_really_inline document_stream &operator=(document_stream &&other) noexcept = default;
+  simdjson_inline document_stream &operator=(document_stream &&other) noexcept = default;
 
-  simdjson_really_inline ~document_stream() noexcept;
+  simdjson_inline ~document_stream() noexcept;
 
   /**
    * Returns the input size in bytes.
@@ -27241,11 +26574,11 @@ class document_stream {
     /**
      * Default constructor.
      */
-    simdjson_really_inline iterator() noexcept;
+    simdjson_inline iterator() noexcept;
     /**
      * Get the current document (or error).
      */
-    simdjson_really_inline simdjson_result<ondemand::document_reference> operator*() noexcept;
+    simdjson_inline simdjson_result<ondemand::document_reference> operator*() noexcept;
     /**
      * Advance to the next document (prefix).
      */
@@ -27254,7 +26587,7 @@ class document_stream {
      * Check if we're at the end yet.
      * @param other the end iterator to compare to.
      */
-    simdjson_really_inline bool operator!=(const iterator &other) const noexcept;
+    simdjson_inline bool operator!=(const iterator &other) const noexcept;
     /**
      * @private
      *
@@ -27270,7 +26603,7 @@ class document_stream {
      * may change in future versions of simdjson: we find the API somewhat
      * awkward and we would like to offer something friendlier.
      */
-     simdjson_really_inline size_t current_index() const noexcept;
+     simdjson_inline size_t current_index() const noexcept;
 
      /**
      * @private
@@ -27291,7 +26624,7 @@ class document_stream {
      * awkward and we would like to offer something friendlier.
      *
      */
-     simdjson_really_inline std::string_view source() const noexcept;
+     simdjson_inline std::string_view source() const noexcept;
 
     /**
      * Returns error of the stream (if any).
@@ -27299,7 +26632,7 @@ class document_stream {
      inline error_code error() const noexcept;
 
   private:
-    simdjson_really_inline iterator(document_stream *s, bool finished) noexcept;
+    simdjson_inline iterator(document_stream *s, bool finished) noexcept;
     /** The document_stream we're iterating through. */
     document_stream* stream;
     /** Whether we're finished or not. */
@@ -27313,11 +26646,11 @@ class document_stream {
   /**
    * Start iterating the documents in the stream.
    */
-  simdjson_really_inline iterator begin() noexcept;
+  simdjson_inline iterator begin() noexcept;
   /**
    * The end of the stream, for iterator comparison purposes.
    */
-  simdjson_really_inline iterator end() noexcept;
+  simdjson_inline iterator end() noexcept;
 
 private:
 
@@ -27333,7 +26666,7 @@ class document_stream {
    * @param len is the length of the raw byte buffer in bytes
    * @param batch_size is the size of the windows (must be strictly greater or equal to the largest JSON document)
    */
-  simdjson_really_inline document_stream(
+  simdjson_inline document_stream(
     ondemand::parser &parser,
     const uint8_t *buf,
     size_t len,
@@ -27437,9 +26770,9 @@ namespace simdjson {
 template<>
 struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_stream> : public SIMDJSON_BUILTIN_IMPLEMENTATION::implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_stream> {
 public:
-  simdjson_really_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_stream &&value) noexcept; ///< @private
-  simdjson_really_inline simdjson_result(error_code error) noexcept; ///< @private
-  simdjson_really_inline simdjson_result() noexcept = default;
+  simdjson_inline simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_stream &&value) noexcept; ///< @private
+  simdjson_inline simdjson_result(error_code error) noexcept; ///< @private
+  simdjson_inline simdjson_result() noexcept = default;
 };
 
 } // namespace simdjson
@@ -27449,23 +26782,27 @@ struct simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_strea
 namespace simdjson {
 /**
  * Create a string-view instance out of a document instance. The string-view instance
- * contains JSON text that is suitable to be parsed as JSON again.
+ * contains JSON text that is suitable to be parsed as JSON again. It does not
+ * validate the content.
  */
 inline simdjson_result<std::string_view> to_json_string(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document& x) noexcept;
 /**
  * Create a string-view instance out of a value instance. The string-view instance
  * contains JSON text that is suitable to be parsed as JSON again. The value must
- * not have been accessed previously.
+ * not have been accessed previously. It does not
+ * validate the content.
  */
 inline simdjson_result<std::string_view> to_json_string(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value& x) noexcept;
 /**
  * Create a string-view instance out of an object instance. The string-view instance
- * contains JSON text that is suitable to be parsed as JSON again.
+ * contains JSON text that is suitable to be parsed as JSON again. It does not
+ * validate the content.
  */
 inline simdjson_result<std::string_view> to_json_string(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object& x) noexcept;
 /**
  * Create a string-view instance out of an array instance. The string-view instance
- * contains JSON text that is suitable to be parsed as JSON again.
+ * contains JSON text that is suitable to be parsed as JSON again. It does not
+ * validate the content.
  */
 inline simdjson_result<std::string_view> to_json_string(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array& x) noexcept;
 inline simdjson_result<std::string_view> to_json_string(simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document> x);
@@ -27484,7 +26821,8 @@ inline simdjson_result<std::string_view> to_json_string(simdjson_result<SIMDJSON
 namespace simdjson { namespace SIMDJSON_BUILTIN_IMPLEMENTATION { namespace ondemand {
 
 /**
- * Print JSON to an output stream.
+ * Print JSON to an output stream.  It does not
+ * validate the content.
  *
  * @param out The output stream.
  * @param value The element.
@@ -27495,7 +26833,8 @@ inline std::ostream& operator<<(std::ostream& out, simdjson::SIMDJSON_BUILTIN_IM
 inline std::ostream& operator<<(std::ostream& out, simdjson::simdjson_result<simdjson::SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> x);
 #endif
 /**
- * Print JSON to an output stream.
+ * Print JSON to an output stream. It does not
+ * validate the content.
  *
  * @param out The output stream.
  * @param value The array.
@@ -27506,7 +26845,8 @@ inline std::ostream& operator<<(std::ostream& out, simdjson::SIMDJSON_BUILTIN_IM
 inline std::ostream& operator<<(std::ostream& out, simdjson::simdjson_result<simdjson::SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array> x);
 #endif
 /**
- * Print JSON to an output stream.
+ * Print JSON to an output stream. It does not
+ * validate the content.
  *
  * @param out The output stream.
  * @param value The array.
@@ -27521,7 +26861,8 @@ inline std::ostream& operator<<(std::ostream& out, simdjson::SIMDJSON_BUILTIN_IM
 inline std::ostream& operator<<(std::ostream& out, simdjson::simdjson_result<simdjson::SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>&& x);
 #endif
 /**
- * Print JSON to an output stream.
+ * Print JSON to an output stream. It does not
+ * validate the content.
  *
  * @param out The output stream.
  * @param value The object.
@@ -27545,7 +26886,7 @@ namespace SIMDJSON_BUILTIN_IMPLEMENTATION {
 //
 
 template<typename T>
-simdjson_really_inline void implementation_simdjson_result_base<T>::tie(T &value, error_code &error) && noexcept {
+simdjson_inline void implementation_simdjson_result_base<T>::tie(T &value, error_code &error) && noexcept {
   error = this->second;
   if (!error) {
     value = std::forward<implementation_simdjson_result_base<T>>(*this).first;
@@ -27553,66 +26894,66 @@ simdjson_really_inline void implementation_simdjson_result_base<T>::tie(T &value
 }
 
 template<typename T>
-simdjson_warn_unused simdjson_really_inline error_code implementation_simdjson_result_base<T>::get(T &value) && noexcept {
+simdjson_warn_unused simdjson_inline error_code implementation_simdjson_result_base<T>::get(T &value) && noexcept {
   error_code error;
   std::forward<implementation_simdjson_result_base<T>>(*this).tie(value, error);
   return error;
 }
 
 template<typename T>
-simdjson_really_inline error_code implementation_simdjson_result_base<T>::error() const noexcept {
+simdjson_inline error_code implementation_simdjson_result_base<T>::error() const noexcept {
   return this->second;
 }
 
 #if SIMDJSON_EXCEPTIONS
 
 template<typename T>
-simdjson_really_inline T& implementation_simdjson_result_base<T>::value() & noexcept(false) {
+simdjson_inline T& implementation_simdjson_result_base<T>::value() & noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return this->first;
 }
 
 template<typename T>
-simdjson_really_inline T&& implementation_simdjson_result_base<T>::value() && noexcept(false) {
+simdjson_inline T&& implementation_simdjson_result_base<T>::value() && noexcept(false) {
   return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
 }
 
 template<typename T>
-simdjson_really_inline T&& implementation_simdjson_result_base<T>::take_value() && noexcept(false) {
+simdjson_inline T&& implementation_simdjson_result_base<T>::take_value() && noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return std::forward<T>(this->first);
 }
 
 template<typename T>
-simdjson_really_inline implementation_simdjson_result_base<T>::operator T&&() && noexcept(false) {
+simdjson_inline implementation_simdjson_result_base<T>::operator T&&() && noexcept(false) {
   return std::forward<implementation_simdjson_result_base<T>>(*this).take_value();
 }
 
 #endif // SIMDJSON_EXCEPTIONS
 
 template<typename T>
-simdjson_really_inline const T& implementation_simdjson_result_base<T>::value_unsafe() const& noexcept {
+simdjson_inline const T& implementation_simdjson_result_base<T>::value_unsafe() const& noexcept {
   return this->first;
 }
 
 template<typename T>
-simdjson_really_inline T& implementation_simdjson_result_base<T>::value_unsafe() & noexcept {
+simdjson_inline T& implementation_simdjson_result_base<T>::value_unsafe() & noexcept {
   return this->first;
 }
 
 template<typename T>
-simdjson_really_inline T&& implementation_simdjson_result_base<T>::value_unsafe() && noexcept {
+simdjson_inline T&& implementation_simdjson_result_base<T>::value_unsafe() && noexcept {
   return std::forward<T>(this->first);
 }
 
 template<typename T>
-simdjson_really_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value, error_code error) noexcept
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value, error_code error) noexcept
     : first{std::forward<T>(value)}, second{error} {}
 template<typename T>
-simdjson_really_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(error_code error) noexcept
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(error_code error) noexcept
     : implementation_simdjson_result_base(T{}, error) {}
 template<typename T>
-simdjson_really_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value) noexcept
+simdjson_inline implementation_simdjson_result_base<T>::implementation_simdjson_result_base(T &&value) noexcept
     : implementation_simdjson_result_base(std::forward<T>(value), SUCCESS) {}
 
 } // namespace SIMDJSON_BUILTIN_IMPLEMENTATION
@@ -27654,48 +26995,48 @@ inline std::ostream& operator<<(std::ostream& out, simdjson_result<json_type> &t
 
 
 
-simdjson_really_inline number_type number::get_number_type() const noexcept {
+simdjson_inline number_type number::get_number_type() const noexcept {
   return type;
 }
 
-simdjson_really_inline bool number::is_uint64() const noexcept {
+simdjson_inline bool number::is_uint64() const noexcept {
   return get_number_type() == number_type::unsigned_integer;
 }
 
-simdjson_really_inline uint64_t number::get_uint64() const noexcept {
+simdjson_inline uint64_t number::get_uint64() const noexcept {
   return payload.unsigned_integer;
 }
 
-simdjson_really_inline number::operator uint64_t() const noexcept {
+simdjson_inline number::operator uint64_t() const noexcept {
   return get_uint64();
 }
 
 
-simdjson_really_inline bool number::is_int64() const noexcept {
+simdjson_inline bool number::is_int64() const noexcept {
   return get_number_type() == number_type::signed_integer;
 }
 
-simdjson_really_inline int64_t number::get_int64() const noexcept {
+simdjson_inline int64_t number::get_int64() const noexcept {
   return payload.signed_integer;
 }
 
-simdjson_really_inline number::operator int64_t() const noexcept {
+simdjson_inline number::operator int64_t() const noexcept {
   return get_int64();
 }
 
-simdjson_really_inline bool number::is_double() const noexcept {
+simdjson_inline bool number::is_double() const noexcept {
     return get_number_type() == number_type::floating_point_number;
 }
 
-simdjson_really_inline double number::get_double() const noexcept {
+simdjson_inline double number::get_double() const noexcept {
   return payload.floating_point_number;
 }
 
-simdjson_really_inline number::operator double() const noexcept {
+simdjson_inline number::operator double() const noexcept {
   return get_double();
 }
 
-simdjson_really_inline double number::as_double() const noexcept {
+simdjson_inline double number::as_double() const noexcept {
   if(is_double()) {
     return payload.floating_point_number;
   }
@@ -27705,22 +27046,22 @@ simdjson_really_inline double number::as_double() const noexcept {
   return double(payload.unsigned_integer);
 }
 
-simdjson_really_inline void number::append_s64(int64_t value) noexcept {
+simdjson_inline void number::append_s64(int64_t value) noexcept {
   payload.signed_integer = value;
   type = number_type::signed_integer;
 }
 
-simdjson_really_inline void number::append_u64(uint64_t value) noexcept {
+simdjson_inline void number::append_u64(uint64_t value) noexcept {
   payload.unsigned_integer = value;
   type = number_type::unsigned_integer;
 }
 
-simdjson_really_inline void number::append_double(double value) noexcept {
+simdjson_inline void number::append_double(double value) noexcept {
   payload.floating_point_number = value;
   type = number_type::floating_point_number;
 }
 
-simdjson_really_inline void number::skip_double() noexcept {
+simdjson_inline void number::skip_double() noexcept {
   type = number_type::floating_point_number;
 }
 
@@ -27730,9 +27071,9 @@ simdjson_really_inline void number::skip_double() noexcept {
 
 namespace simdjson {
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type>::simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type &&value) noexcept
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type>::simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type &&value) noexcept
     : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type>(std::forward<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type>(value)) {}
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type>::simdjson_result(error_code error) noexcept
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type>::simdjson_result(error_code error) noexcept
     : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type>(error) {}
 
 } // namespace simdjson
@@ -27892,7 +27233,7 @@ inline void log_line(const json_iterator &iter, token_position index, depth_t de
       printf(" ");
     }
     // printf("| %5u ", *(index+1));
-    printf("| %5u ", depth);
+    printf("| %5i ", depth);
     printf("| %.*s ", int(detail.size()), detail.data());
     printf("|\n");
     fflush(stdout);
@@ -27910,18 +27251,12 @@ namespace simdjson {
 namespace SIMDJSON_BUILTIN_IMPLEMENTATION {
 namespace ondemand {
 
-simdjson_really_inline raw_json_string::raw_json_string(const uint8_t * _buf) noexcept : buf{_buf} {}
+simdjson_inline raw_json_string::raw_json_string(const uint8_t * _buf) noexcept : buf{_buf} {}
 
-simdjson_really_inline const char * raw_json_string::raw() const noexcept { return reinterpret_cast<const char *>(buf); }
-simdjson_really_inline simdjson_warn_unused simdjson_result<std::string_view> raw_json_string::unescape(uint8_t *&dst) const noexcept {
-  uint8_t *end = stringparsing::parse_string(buf, dst);
-  if (!end) { return STRING_ERROR; }
-  std::string_view result(reinterpret_cast<const char *>(dst), end-dst);
-  dst = end;
-  return result;
-}
+simdjson_inline const char * raw_json_string::raw() const noexcept { return reinterpret_cast<const char *>(buf); }
 
-simdjson_really_inline bool raw_json_string::is_free_from_unescaped_quote(std::string_view target) noexcept {
+
+simdjson_inline bool raw_json_string::is_free_from_unescaped_quote(std::string_view target) noexcept {
   size_t pos{0};
   // if the content has no escape character, just scan through it quickly!
   for(;pos < target.size() && target[pos] != '\\';pos++) {}
@@ -27939,7 +27274,7 @@ simdjson_really_inline bool raw_json_string::is_free_from_unescaped_quote(std::s
   return true;
 }
 
-simdjson_really_inline bool raw_json_string::is_free_from_unescaped_quote(const char* target) noexcept {
+simdjson_inline bool raw_json_string::is_free_from_unescaped_quote(const char* target) noexcept {
   size_t pos{0};
   // if the content has no escape character, just scan through it quickly!
   for(;target[pos] && target[pos] != '\\';pos++) {}
@@ -27958,12 +27293,12 @@ simdjson_really_inline bool raw_json_string::is_free_from_unescaped_quote(const
 }
 
 
-simdjson_really_inline bool raw_json_string::unsafe_is_equal(size_t length, std::string_view target) const noexcept {
+simdjson_inline bool raw_json_string::unsafe_is_equal(size_t length, std::string_view target) const noexcept {
   // If we are going to call memcmp, then we must know something about the length of the raw_json_string.
   return (length >= target.size()) && (raw()[target.size()] == '"') && !memcmp(raw(), target.data(), target.size());
 }
 
-simdjson_really_inline bool raw_json_string::unsafe_is_equal(std::string_view target) const noexcept {
+simdjson_inline bool raw_json_string::unsafe_is_equal(std::string_view target) const noexcept {
   // Assumptions: does not contain unescaped quote characters, and
   // the raw content is quote terminated within a valid JSON string.
   if(target.size() <= SIMDJSON_PADDING) {
@@ -27978,7 +27313,7 @@ simdjson_really_inline bool raw_json_string::unsafe_is_equal(std::string_view ta
   return true;
 }
 
-simdjson_really_inline bool raw_json_string::is_equal(std::string_view target) const noexcept {
+simdjson_inline bool raw_json_string::is_equal(std::string_view target) const noexcept {
   const char * r{raw()};
   size_t pos{0};
   bool escaping{false};
@@ -28002,7 +27337,7 @@ simdjson_really_inline bool raw_json_string::is_equal(std::string_view target) c
 }
 
 
-simdjson_really_inline bool raw_json_string::unsafe_is_equal(const char * target) const noexcept {
+simdjson_inline bool raw_json_string::unsafe_is_equal(const char * target) const noexcept {
   // Assumptions: 'target' does not contain unescaped quote characters, is null terminated and
   // the raw content is quote terminated within a valid JSON string.
   const char * r{raw()};
@@ -28014,7 +27349,7 @@ simdjson_really_inline bool raw_json_string::unsafe_is_equal(const char * target
   return true;
 }
 
-simdjson_really_inline bool raw_json_string::is_equal(const char* target) const noexcept {
+simdjson_inline bool raw_json_string::is_equal(const char* target) const noexcept {
   // Assumptions: does not contain unescaped quote characters, and
   // the raw content is quote terminated within a valid JSON string.
   const char * r{raw()};
@@ -28039,29 +27374,29 @@ simdjson_really_inline bool raw_json_string::is_equal(const char* target) const
   return true;
 }
 
-simdjson_unused simdjson_really_inline bool operator==(const raw_json_string &a, std::string_view c) noexcept {
+simdjson_unused simdjson_inline bool operator==(const raw_json_string &a, std::string_view c) noexcept {
   return a.unsafe_is_equal(c);
 }
 
-simdjson_unused simdjson_really_inline bool operator==(std::string_view c, const raw_json_string &a) noexcept {
+simdjson_unused simdjson_inline bool operator==(std::string_view c, const raw_json_string &a) noexcept {
   return a == c;
 }
 
-simdjson_unused simdjson_really_inline bool operator!=(const raw_json_string &a, std::string_view c) noexcept {
+simdjson_unused simdjson_inline bool operator!=(const raw_json_string &a, std::string_view c) noexcept {
   return !(a == c);
 }
 
-simdjson_unused simdjson_really_inline bool operator!=(std::string_view c, const raw_json_string &a) noexcept {
+simdjson_unused simdjson_inline bool operator!=(std::string_view c, const raw_json_string &a) noexcept {
   return !(a == c);
 }
 
 
-simdjson_really_inline simdjson_warn_unused simdjson_result<std::string_view> raw_json_string::unescape(json_iterator &iter) const noexcept {
-  return unescape(iter.string_buf_loc());
+simdjson_inline simdjson_warn_unused simdjson_result<std::string_view> raw_json_string::unescape(json_iterator &iter) const noexcept {
+  return iter.unescape(*this);
 }
 
 
-simdjson_unused simdjson_really_inline std::ostream &operator<<(std::ostream &out, const raw_json_string &str) noexcept {
+simdjson_unused simdjson_inline std::ostream &operator<<(std::ostream &out, const raw_json_string &str) noexcept {
   bool in_escape = false;
   const char *s = str.raw();
   while (true) {
@@ -28081,20 +27416,16 @@ simdjson_unused simdjson_really_inline std::ostream &operator<<(std::ostream &ou
 
 namespace simdjson {
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string>::simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string &&value) noexcept
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string>::simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string &&value) noexcept
     : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string>(std::forward<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string>(value)) {}
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string>::simdjson_result(error_code error) noexcept
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string>::simdjson_result(error_code error) noexcept
     : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string>(error) {}
 
-simdjson_really_inline simdjson_result<const char *> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string>::raw() const noexcept {
+simdjson_inline simdjson_result<const char *> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string>::raw() const noexcept {
   if (error()) { return error(); }
   return first.raw();
 }
-simdjson_really_inline simdjson_warn_unused simdjson_result<std::string_view> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string>::unescape(uint8_t *&dst) const noexcept {
-  if (error()) { return error(); }
-  return first.unescape(dst);
-}
-simdjson_really_inline simdjson_warn_unused simdjson_result<std::string_view> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string>::unescape(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_iterator &iter) const noexcept {
+simdjson_inline simdjson_warn_unused simdjson_result<std::string_view> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string>::unescape(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_iterator &iter) const noexcept {
   if (error()) { return error(); }
   return first.unescape(iter);
 }
@@ -28106,65 +27437,65 @@ namespace simdjson {
 namespace SIMDJSON_BUILTIN_IMPLEMENTATION {
 namespace ondemand {
 
-simdjson_really_inline token_iterator::token_iterator(
+simdjson_inline token_iterator::token_iterator(
   const uint8_t *_buf,
   token_position position
 ) noexcept : buf{_buf}, _position{position}
 {
 }
 
-simdjson_really_inline uint32_t token_iterator::current_offset() const noexcept {
+simdjson_inline uint32_t token_iterator::current_offset() const noexcept {
   return *(_position);
 }
 
 
-simdjson_really_inline const uint8_t *token_iterator::return_current_and_advance() noexcept {
+simdjson_inline const uint8_t *token_iterator::return_current_and_advance() noexcept {
   return &buf[*(_position++)];
 }
 
-simdjson_really_inline const uint8_t *token_iterator::peek(token_position position) const noexcept {
+simdjson_inline const uint8_t *token_iterator::peek(token_position position) const noexcept {
   return &buf[*position];
 }
-simdjson_really_inline uint32_t token_iterator::peek_index(token_position position) const noexcept {
+simdjson_inline uint32_t token_iterator::peek_index(token_position position) const noexcept {
   return *position;
 }
-simdjson_really_inline uint32_t token_iterator::peek_length(token_position position) const noexcept {
+simdjson_inline uint32_t token_iterator::peek_length(token_position position) const noexcept {
   return *(position+1) - *position;
 }
 
-simdjson_really_inline const uint8_t *token_iterator::peek(int32_t delta) const noexcept {
+simdjson_inline const uint8_t *token_iterator::peek(int32_t delta) const noexcept {
   return &buf[*(_position+delta)];
 }
-simdjson_really_inline uint32_t token_iterator::peek_index(int32_t delta) const noexcept {
+simdjson_inline uint32_t token_iterator::peek_index(int32_t delta) const noexcept {
   return *(_position+delta);
 }
-simdjson_really_inline uint32_t token_iterator::peek_length(int32_t delta) const noexcept {
+simdjson_inline uint32_t token_iterator::peek_length(int32_t delta) const noexcept {
   return *(_position+delta+1) - *(_position+delta);
 }
 
-simdjson_really_inline token_position token_iterator::position() const noexcept {
+simdjson_inline token_position token_iterator::position() const noexcept {
   return _position;
 }
-simdjson_really_inline void token_iterator::set_position(token_position target_position) noexcept {
+simdjson_inline void token_iterator::set_position(token_position target_position) noexcept {
   _position = target_position;
 }
 
-simdjson_really_inline bool token_iterator::operator==(const token_iterator &other) const noexcept {
+simdjson_inline bool token_iterator::operator==(const token_iterator &other) const noexcept {
   return _position == other._position;
 }
-simdjson_really_inline bool token_iterator::operator!=(const token_iterator &other) const noexcept {
+simdjson_inline bool token_iterator::operator!=(const token_iterator &other) const noexcept {
   return _position != other._position;
 }
-simdjson_really_inline bool token_iterator::operator>(const token_iterator &other) const noexcept {
+simdjson_inline bool token_iterator::operator>(const token_iterator &other) const noexcept {
   return _position > other._position;
 }
-simdjson_really_inline bool token_iterator::operator>=(const token_iterator &other) const noexcept {
+simdjson_inline bool token_iterator::operator>=(const token_iterator &other) const noexcept {
   return _position >= other._position;
 }
-simdjson_really_inline bool token_iterator::operator<(const token_iterator &other) const noexcept {
+simdjson_inline bool token_iterator::operator<(const token_iterator &other) const noexcept {
   return _position < other._position;
 }
-simdjson_really_inline bool token_iterator::operator<=(const token_iterator &other) const noexcept {
+simdjson_inline bool token_iterator::operator<=(const token_iterator &other) const noexcept {
   return _position <= other._position;
 }
 
@@ -28174,9 +27505,9 @@ simdjson_really_inline bool token_iterator::operator<=(const token_iterator &oth
 
 namespace simdjson {
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::token_iterator>::simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::token_iterator &&value) noexcept
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::token_iterator>::simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::token_iterator &&value) noexcept
     : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::token_iterator>(std::forward<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::token_iterator>(value)) {}
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::token_iterator>::simdjson_result(error_code error) noexcept
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::token_iterator>::simdjson_result(error_code error) noexcept
     : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::token_iterator>(error) {}
 
 } // namespace simdjson
@@ -28186,7 +27517,7 @@ namespace simdjson {
 namespace SIMDJSON_BUILTIN_IMPLEMENTATION {
 namespace ondemand {
 
-simdjson_really_inline json_iterator::json_iterator(json_iterator &&other) noexcept
+simdjson_inline json_iterator::json_iterator(json_iterator &&other) noexcept
   : token(std::forward<token_iterator>(other.token)),
     parser{other.parser},
     _string_buf_loc{other._string_buf_loc},
@@ -28197,7 +27528,7 @@ simdjson_really_inline json_iterator::json_iterator(json_iterator &&other) noexc
 {
   other.parser = nullptr;
 }
-simdjson_really_inline json_iterator &json_iterator::operator=(json_iterator &&other) noexcept {
+simdjson_inline json_iterator &json_iterator::operator=(json_iterator &&other) noexcept {
   token = other.token;
   parser = other.parser;
   _string_buf_loc = other._string_buf_loc;
@@ -28209,7 +27540,7 @@ simdjson_really_inline json_iterator &json_iterator::operator=(json_iterator &&o
   return *this;
 }
 
-simdjson_really_inline json_iterator::json_iterator(const uint8_t *buf, ondemand::parser *_parser) noexcept
+simdjson_inline json_iterator::json_iterator(const uint8_t *buf, ondemand::parser *_parser) noexcept
   : token(buf, &_parser->implementation->structural_indexes[0]),
     parser{_parser},
     _string_buf_loc{parser->string_buf.get()},
@@ -28231,12 +27562,33 @@ inline void json_iterator::rewind() noexcept {
   _depth = 1;
 }
 
+inline bool json_iterator::balanced() const noexcept {
+  token_iterator ti(token);
+  int32_t count{0};
+  ti.set_position( root_position() );
+  while(ti.peek() <= peek_last()) {
+    switch (*ti.return_current_and_advance())
+    {
+    case '[': case '{':
+      count++;
+      break;
+    case ']': case '}':
+      count--;
+      break;
+    default:
+      break;
+    }
+  }
+  return count == 0;
+}
+
+
 // GCC 7 warns when the first line of this function is inlined away into oblivion due to the caller
 // relating depth and parent_depth, which is a desired effect. The warning does not show up if the
 // skip_child() function is not marked inline).
 SIMDJSON_PUSH_DISABLE_WARNINGS
 SIMDJSON_DISABLE_STRICT_OVERFLOW_WARNING
-simdjson_warn_unused simdjson_really_inline error_code json_iterator::skip_child(depth_t parent_depth) noexcept {
+simdjson_warn_unused simdjson_inline error_code json_iterator::skip_child(depth_t parent_depth) noexcept {
   if (depth() <= parent_depth) { return SUCCESS; }
   switch (*return_current_and_advance()) {
     // TODO consider whether matching braces is a requirement: if non-matching braces indicates
@@ -28313,23 +27665,27 @@ simdjson_warn_unused simdjson_really_inline error_code json_iterator::skip_child
 
 SIMDJSON_POP_DISABLE_WARNINGS
 
-simdjson_really_inline bool json_iterator::at_root() const noexcept {
+simdjson_inline bool json_iterator::at_root() const noexcept {
   return position() == root_position();
 }
 
-simdjson_really_inline bool json_iterator::streaming() const noexcept {
+simdjson_inline bool json_iterator::is_single_token() const noexcept {
+  return parser->implementation->n_structural_indexes == 1;
+}
+
+simdjson_inline bool json_iterator::streaming() const noexcept {
   return _streaming;
 }
 
-simdjson_really_inline token_position json_iterator::root_position() const noexcept {
+simdjson_inline token_position json_iterator::root_position() const noexcept {
   return _root;
 }
 
-simdjson_really_inline void json_iterator::assert_at_document_depth() const noexcept {
+simdjson_inline void json_iterator::assert_at_document_depth() const noexcept {
   SIMDJSON_ASSUME( _depth == 1 );
 }
 
-simdjson_really_inline void json_iterator::assert_at_root() const noexcept {
+simdjson_inline void json_iterator::assert_at_root() const noexcept {
   SIMDJSON_ASSUME( _depth == 1 );
 #ifndef SIMDJSON_CLANG_VISUAL_STUDIO
   // Under Visual Studio, the next SIMDJSON_ASSUME fails with: the argument
@@ -28338,21 +27694,21 @@ simdjson_really_inline void json_iterator::assert_at_root() const noexcept {
 #endif
 }
 
-simdjson_really_inline void json_iterator::assert_more_tokens(uint32_t required_tokens) const noexcept {
+simdjson_inline void json_iterator::assert_more_tokens(uint32_t required_tokens) const noexcept {
   assert_valid_position(token._position + required_tokens - 1);
 }
 
-simdjson_really_inline void json_iterator::assert_valid_position(token_position position) const noexcept {
+simdjson_inline void json_iterator::assert_valid_position(token_position position) const noexcept {
 #ifndef SIMDJSON_CLANG_VISUAL_STUDIO
   SIMDJSON_ASSUME( position >= &parser->implementation->structural_indexes[0] );
   SIMDJSON_ASSUME( position < &parser->implementation->structural_indexes[parser->implementation->n_structural_indexes] );
 #endif
 }
 
-simdjson_really_inline bool json_iterator::at_end() const noexcept {
+simdjson_inline bool json_iterator::at_end() const noexcept {
   return position() == end_position();
 }
-simdjson_really_inline token_position json_iterator::end_position() const noexcept {
+simdjson_inline token_position json_iterator::end_position() const noexcept {
   uint32_t n_structural_indexes{parser->implementation->n_structural_indexes};
   return &parser->implementation->structural_indexes[n_structural_indexes];
 }
@@ -28381,42 +27737,42 @@ inline simdjson_result<const char *> json_iterator::current_location() noexcept
   return reinterpret_cast<const char *>(token.peek());
 }
 
-simdjson_really_inline bool json_iterator::is_alive() const noexcept {
+simdjson_inline bool json_iterator::is_alive() const noexcept {
   return parser;
 }
 
-simdjson_really_inline void json_iterator::abandon() noexcept {
+simdjson_inline void json_iterator::abandon() noexcept {
   parser = nullptr;
   _depth = 0;
 }
 
-simdjson_really_inline const uint8_t *json_iterator::return_current_and_advance() noexcept {
+simdjson_inline const uint8_t *json_iterator::return_current_and_advance() noexcept {
 #if SIMDJSON_CHECK_EOF
   assert_more_tokens();
 #endif // SIMDJSON_CHECK_EOF
   return token.return_current_and_advance();
 }
 
-simdjson_really_inline const uint8_t *json_iterator::unsafe_pointer() const noexcept {
+simdjson_inline const uint8_t *json_iterator::unsafe_pointer() const noexcept {
   // deliberately done without safety guard:
   return token.peek(0);
 }
 
-simdjson_really_inline const uint8_t *json_iterator::peek(int32_t delta) const noexcept {
+simdjson_inline const uint8_t *json_iterator::peek(int32_t delta) const noexcept {
 #if SIMDJSON_CHECK_EOF
   assert_more_tokens(delta+1);
 #endif // SIMDJSON_CHECK_EOF
   return token.peek(delta);
 }
 
-simdjson_really_inline uint32_t json_iterator::peek_length(int32_t delta) const noexcept {
+simdjson_inline uint32_t json_iterator::peek_length(int32_t delta) const noexcept {
 #if SIMDJSON_CHECK_EOF
   assert_more_tokens(delta+1);
 #endif // #if SIMDJSON_CHECK_EOF
   return token.peek_length(delta);
 }
 
-simdjson_really_inline const uint8_t *json_iterator::peek(token_position position) const noexcept {
+simdjson_inline const uint8_t *json_iterator::peek(token_position position) const noexcept {
   // todo: currently we require end-of-string buffering, but the following
   // assert_valid_position should be turned on if/when we lift that condition.
   // assert_valid_position(position);
@@ -28425,14 +27781,14 @@ simdjson_really_inline const uint8_t *json_iterator::peek(token_position positio
   return token.peek(position);
 }
 
-simdjson_really_inline uint32_t json_iterator::peek_length(token_position position) const noexcept {
+simdjson_inline uint32_t json_iterator::peek_length(token_position position) const noexcept {
 #if SIMDJSON_CHECK_EOF
   assert_valid_position(position);
 #endif // SIMDJSON_CHECK_EOF
   return token.peek_length(position);
 }
 
-simdjson_really_inline token_position json_iterator::last_position() const noexcept {
+simdjson_inline token_position json_iterator::last_position() const noexcept {
   // The following line fails under some compilers...
   // SIMDJSON_ASSUME(parser->implementation->n_structural_indexes > 0);
   // since it has side-effects.
@@ -28440,46 +27796,51 @@ simdjson_really_inline token_position json_iterator::last_position() const noexc
   SIMDJSON_ASSUME(n_structural_indexes > 0);
   return &parser->implementation->structural_indexes[n_structural_indexes - 1];
 }
-simdjson_really_inline const uint8_t *json_iterator::peek_last() const noexcept {
+simdjson_inline const uint8_t *json_iterator::peek_last() const noexcept {
   return token.peek(last_position());
 }
 
-simdjson_really_inline void json_iterator::ascend_to(depth_t parent_depth) noexcept {
+simdjson_inline void json_iterator::ascend_to(depth_t parent_depth) noexcept {
   SIMDJSON_ASSUME(parent_depth >= 0 && parent_depth < INT32_MAX - 1);
   SIMDJSON_ASSUME(_depth == parent_depth + 1);
   _depth = parent_depth;
 }
 
-simdjson_really_inline void json_iterator::descend_to(depth_t child_depth) noexcept {
+simdjson_inline void json_iterator::descend_to(depth_t child_depth) noexcept {
   SIMDJSON_ASSUME(child_depth >= 1 && child_depth < INT32_MAX);
   SIMDJSON_ASSUME(_depth == child_depth - 1);
   _depth = child_depth;
 }
 
-simdjson_really_inline depth_t json_iterator::depth() const noexcept {
+simdjson_inline depth_t json_iterator::depth() const noexcept {
   return _depth;
 }
 
-simdjson_really_inline uint8_t *&json_iterator::string_buf_loc() noexcept {
+simdjson_inline uint8_t *&json_iterator::string_buf_loc() noexcept {
   return _string_buf_loc;
 }
 
-simdjson_really_inline error_code json_iterator::report_error(error_code _error, const char *message) noexcept {
+simdjson_inline error_code json_iterator::report_error(error_code _error, const char *message) noexcept {
   SIMDJSON_ASSUME(_error != SUCCESS && _error != UNINITIALIZED && _error != INCORRECT_TYPE && _error != NO_SUCH_FIELD);
   logger::log_error(*this, message);
   error = _error;
   return error;
 }
 
-simdjson_really_inline token_position json_iterator::position() const noexcept {
+simdjson_inline token_position json_iterator::position() const noexcept {
   return token.position();
 }
 
-simdjson_really_inline void json_iterator::reenter_child(token_position position, depth_t child_depth) noexcept {
+simdjson_inline simdjson_result<std::string_view> json_iterator::unescape(raw_json_string in) noexcept {
+  return parser->unescape(in, _string_buf_loc);
+}
+
+simdjson_inline void json_iterator::reenter_child(token_position position, depth_t child_depth) noexcept {
   SIMDJSON_ASSUME(child_depth >= 1 && child_depth < INT32_MAX);
   SIMDJSON_ASSUME(_depth == child_depth - 1);
-#ifdef SIMDJSON_DEVELOPMENT_CHECKS
+#if SIMDJSON_DEVELOPMENT_CHECKS
 #ifndef SIMDJSON_CLANG_VISUAL_STUDIO
+  SIMDJSON_ASSUME(size_t(child_depth) < parser->max_depth());
   SIMDJSON_ASSUME(position >= parser->start_positions[child_depth]);
 #endif
 #endif
@@ -28487,27 +27848,29 @@ simdjson_really_inline void json_iterator::reenter_child(token_position position
   _depth = child_depth;
 }
 
-#ifdef SIMDJSON_DEVELOPMENT_CHECKS
+#if SIMDJSON_DEVELOPMENT_CHECKS
 
-simdjson_really_inline token_position json_iterator::start_position(depth_t depth) const noexcept {
-  return parser->start_positions[depth];
+simdjson_inline token_position json_iterator::start_position(depth_t depth) const noexcept {
+  SIMDJSON_ASSUME(size_t(depth) < parser->max_depth());
+  return size_t(depth) < parser->max_depth() ? parser->start_positions[depth] : 0;
 }
 
-simdjson_really_inline void json_iterator::set_start_position(depth_t depth, token_position position) noexcept {
-  parser->start_positions[depth] = position;
+simdjson_inline void json_iterator::set_start_position(depth_t depth, token_position position) noexcept {
+  SIMDJSON_ASSUME(size_t(depth) < parser->max_depth());
+  if(size_t(depth) < parser->max_depth()) { parser->start_positions[depth] = position; }
 }
 
 #endif
 
 
-simdjson_really_inline error_code json_iterator::optional_error(error_code _error, const char *message) noexcept {
+simdjson_inline error_code json_iterator::optional_error(error_code _error, const char *message) noexcept {
   SIMDJSON_ASSUME(_error == INCORRECT_TYPE || _error == NO_SUCH_FIELD);
   logger::log_error(*this, message);
   return _error;
 }
 
 template<int N>
-simdjson_warn_unused simdjson_really_inline bool json_iterator::copy_to_buffer(const uint8_t *json, uint32_t max_len, uint8_t (&tmpbuf)[N]) noexcept {
+simdjson_warn_unused simdjson_inline bool json_iterator::copy_to_buffer(const uint8_t *json, uint32_t max_len, uint8_t (&tmpbuf)[N]) noexcept {
   // Let us guard against silly cases:
   if((N < max_len) || (N == 0)) { return false; }
   // Truncate whitespace to fit the buffer.
@@ -28528,9 +27891,9 @@ simdjson_warn_unused simdjson_really_inline bool json_iterator::copy_to_buffer(c
 
 namespace simdjson {
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_iterator>::simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_iterator &&value) noexcept
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_iterator>::simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_iterator &&value) noexcept
     : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_iterator>(std::forward<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_iterator>(value)) {}
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_iterator>::simdjson_result(error_code error) noexcept
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_iterator>::simdjson_result(error_code error) noexcept
     : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_iterator>(error) {}
 
 } // namespace simdjson
@@ -28540,7 +27903,7 @@ namespace simdjson {
 namespace SIMDJSON_BUILTIN_IMPLEMENTATION {
 namespace ondemand {
 
-simdjson_really_inline value_iterator::value_iterator(
+simdjson_inline value_iterator::value_iterator(
   json_iterator *json_iter,
   depth_t depth,
   token_position start_position
@@ -28548,19 +27911,19 @@ simdjson_really_inline value_iterator::value_iterator(
 {
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator::start_object() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<bool> value_iterator::start_object() noexcept {
   SIMDJSON_TRY( start_container('{', "Not an object", "object") );
   return started_object();
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator::start_root_object() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<bool> value_iterator::start_root_object() noexcept {
   SIMDJSON_TRY( start_container('{', "Not an object", "object") );
   return started_root_object();
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator::started_object() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<bool> value_iterator::started_object() noexcept {
   assert_at_container_start();
-#ifdef SIMDJSON_DEVELOPMENT_CHECKS
+#if SIMDJSON_DEVELOPMENT_CHECKS
   _json_iter->set_start_position(_depth, start_position());
 #endif
   if (*_json_iter->peek() == '}') {
@@ -28572,19 +27935,32 @@ simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator
   return true;
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator::started_root_object() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<bool> value_iterator::started_root_object() noexcept {
   // When in streaming mode, we cannot expect peek_last() to be the last structural element of the
   // current document. It only works in the normal mode where we have indexed a single document.
   // Note that adding a check for 'streaming' is not expensive since we only have at most
   // one root element.
-  if (! _json_iter->streaming() && (*_json_iter->peek_last() != '}')) {
-    _json_iter->abandon();
-    return report_error(INCOMPLETE_ARRAY_OR_OBJECT, "missing } at end");
+  if ( ! _json_iter->streaming() ) {
+    if (*_json_iter->peek_last() != '}') {
+      _json_iter->abandon();
+      return report_error(INCOMPLETE_ARRAY_OR_OBJECT, "missing } at end");
+    }
+    // If the last character is } *and* the first gibberish character is also '}'
+    // then on-demand could accidentally go over. So we need additional checks.
+    // https://github.com/simdjson/simdjson/issues/1834
+    // Checking that the document is balanced requires a full scan which is potentially
+    // expensive, but it only happens in edge cases where the first padding character is
+    // a closing bracket.
+    if ((*_json_iter->peek(_json_iter->end_position()) == '}') && (!_json_iter->balanced())) {
+      _json_iter->abandon();
+      // The exact error would require more work. It will typically be an unclosed object.
+      return report_error(INCOMPLETE_ARRAY_OR_OBJECT, "the document is unbalanced");
+    }
   }
   return started_object();
 }
 
-simdjson_warn_unused simdjson_really_inline error_code value_iterator::end_container() noexcept {
+simdjson_warn_unused simdjson_inline error_code value_iterator::end_container() noexcept {
 #if SIMDJSON_CHECK_EOF
     if (depth() > 1 && at_end()) { return report_error(INCOMPLETE_ARRAY_OR_OBJECT, "missing parent ] or }"); }
     // if (depth() <= 1 && !at_end()) { return report_error(INCOMPLETE_ARRAY_OR_OBJECT, "missing [ or { at start"); }
@@ -28593,7 +27969,7 @@ simdjson_warn_unused simdjson_really_inline error_code value_iterator::end_conta
     return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator::has_next_field() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<bool> value_iterator::has_next_field() noexcept {
   assert_at_next();
 
   // It's illegal to call this unless there are more tokens: anything that ends in } or ] is
@@ -28610,7 +27986,7 @@ simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator
   }
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator::find_field_raw(const std::string_view key) noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<bool> value_iterator::find_field_raw(const std::string_view key) noexcept {
   error_code error;
   bool has_value;
   //
@@ -28636,7 +28012,7 @@ simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator
   //    ```
   //
   } else if (!is_open()) {
-#ifdef SIMDJSON_DEVELOPMENT_CHECKS
+#if SIMDJSON_DEVELOPMENT_CHECKS
     // If we're past the end of the object, we're being iterated out of order.
     // Note: this isn't perfect detection. It's possible the user is inside some other object; if so,
     // this object iterator will blithely scan that object for fields.
@@ -28661,7 +28037,7 @@ simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator
   } else {
     if ((error = skip_child() )) { abandon(); return error; }
     if ((error = has_next_field().get(has_value) )) { abandon(); return error; }
-#ifdef SIMDJSON_DEVELOPMENT_CHECKS
+#if SIMDJSON_DEVELOPMENT_CHECKS
     if (_json_iter->start_position(_depth) != start_position()) { return OUT_OF_ORDER_ITERATION; }
 #endif
   }
@@ -28705,7 +28081,9 @@ simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator
   return false;
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator::find_field_unordered_raw(const std::string_view key) noexcept {
+SIMDJSON_PUSH_DISABLE_WARNINGS
+SIMDJSON_DISABLE_STRICT_OVERFLOW_WARNING
+simdjson_warn_unused simdjson_inline simdjson_result<bool> value_iterator::find_field_unordered_raw(const std::string_view key) noexcept {
   /**
    * When find_field_unordered_raw is called, we can either be pointing at the
    * first key, pointing outside (at the closing brace) or if a key was matched
@@ -28746,7 +28124,7 @@ simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator
   //
   } else if (!is_open()) {
 
-#ifdef SIMDJSON_DEVELOPMENT_CHECKS
+#if SIMDJSON_DEVELOPMENT_CHECKS
     // If we're past the end of the object, we're being iterated out of order.
     // Note: this isn't perfect detection. It's possible the user is inside some other object; if so,
     // this object iterator will blithely scan that object for fields.
@@ -28775,7 +28153,7 @@ simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator
     if ((error = skip_child() )) { abandon(); return error; }
     search_start = _json_iter->position();
     if ((error = has_next_field().get(has_value) )) { abandon(); return error; }
-#ifdef SIMDJSON_DEVELOPMENT_CHECKS
+#if SIMDJSON_DEVELOPMENT_CHECKS
     if (_json_iter->start_position(_depth) != start_position()) { return OUT_OF_ORDER_ITERATION; }
 #endif
   }
@@ -28892,8 +28270,9 @@ simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator
   // never reach this point.
   return false;
 }
+SIMDJSON_POP_DISABLE_WARNINGS
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<raw_json_string> value_iterator::field_key() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<raw_json_string> value_iterator::field_key() noexcept {
   assert_at_next();
 
   const uint8_t *key = _json_iter->return_current_and_advance();
@@ -28901,7 +28280,7 @@ simdjson_warn_unused simdjson_really_inline simdjson_result<raw_json_string> val
   return raw_json_string(key);
 }
 
-simdjson_warn_unused simdjson_really_inline error_code value_iterator::field_value() noexcept {
+simdjson_warn_unused simdjson_inline error_code value_iterator::field_value() noexcept {
   assert_at_next();
 
   if (*_json_iter->return_current_and_advance() != ':') { return report_error(TAPE_ERROR, "Missing colon in object field"); }
@@ -28909,12 +28288,12 @@ simdjson_warn_unused simdjson_really_inline error_code value_iterator::field_val
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator::start_array() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<bool> value_iterator::start_array() noexcept {
   SIMDJSON_TRY( start_container('[', "Not an array", "array") );
   return started_array();
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator::start_root_array() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<bool> value_iterator::start_root_array() noexcept {
   SIMDJSON_TRY( start_container('[', "Not an array", "array") );
   return started_root_array();
 }
@@ -28926,7 +28305,7 @@ inline std::string value_iterator::to_string() const noexcept {
   return answer;
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator::started_array() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<bool> value_iterator::started_array() noexcept {
   assert_at_container_start();
   if (*_json_iter->peek() == ']') {
     logger::log_value(*_json_iter, "empty array");
@@ -28935,25 +28314,38 @@ simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator
     return false;
   }
   _json_iter->descend_to(depth()+1);
-#ifdef SIMDJSON_DEVELOPMENT_CHECKS
+#if SIMDJSON_DEVELOPMENT_CHECKS
   _json_iter->set_start_position(_depth, start_position());
 #endif
   return true;
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator::started_root_array() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<bool> value_iterator::started_root_array() noexcept {
   // When in streaming mode, we cannot expect peek_last() to be the last structural element of the
   // current document. It only works in the normal mode where we have indexed a single document.
   // Note that adding a check for 'streaming' is not expensive since we only have at most
   // one root element.
-  if ( ! _json_iter->streaming() && (*_json_iter->peek_last() != ']')) {
-    _json_iter->abandon();
-    return report_error(INCOMPLETE_ARRAY_OR_OBJECT, "missing ] at end");
+  if ( ! _json_iter->streaming() ) {
+    if (*_json_iter->peek_last() != ']') {
+      _json_iter->abandon();
+      return report_error(INCOMPLETE_ARRAY_OR_OBJECT, "missing ] at end");
+    }
+    // If the last character is ] *and* the first gibberish character is also ']'
+    // then on-demand could accidentally go over. So we need additional checks.
+    // https://github.com/simdjson/simdjson/issues/1834
+    // Checking that the document is balanced requires a full scan which is potentially
+    // expensive, but it only happens in edge cases where the first padding character is
+    // a closing bracket.
+    if ((*_json_iter->peek(_json_iter->end_position()) == ']') && (!_json_iter->balanced())) {
+      _json_iter->abandon();
+      // The exact error would require more work. It will typically be an unclosed array.
+      return report_error(INCOMPLETE_ARRAY_OR_OBJECT, "the document is unbalanced");
+    }
   }
   return started_array();
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator::has_next_element() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<bool> value_iterator::has_next_element() noexcept {
   assert_at_next();
 
   logger::log_event(*this, "has_next_element");
@@ -28970,96 +28362,106 @@ simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator
   }
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator::parse_bool(const uint8_t *json) const noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<bool> value_iterator::parse_bool(const uint8_t *json) const noexcept {
   auto not_true = atomparsing::str4ncmp(json, "true");
   auto not_false = atomparsing::str4ncmp(json, "fals") | (json[4] ^ 'e');
   bool error = (not_true && not_false) || jsoncharutils::is_not_structural_or_whitespace(json[not_true ? 5 : 4]);
   if (error) { return incorrect_type_error("Not a boolean"); }
   return simdjson_result<bool>(!not_true);
 }
-simdjson_really_inline bool value_iterator::parse_null(const uint8_t *json) const noexcept {
-  return !atomparsing::str4ncmp(json, "null") && jsoncharutils::is_structural_or_whitespace(json[4]);
+simdjson_warn_unused simdjson_inline simdjson_result<bool> value_iterator::parse_null(const uint8_t *json) const noexcept {
+  bool is_null_string = !atomparsing::str4ncmp(json, "null") && jsoncharutils::is_structural_or_whitespace(json[4]);
+  // if we start with 'n', we must be a null
+  if(!is_null_string && json[0]=='n') { return incorrect_type_error("Not a null but starts with n"); }
+  return is_null_string;
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<std::string_view> value_iterator::get_string() noexcept {
-  return get_raw_json_string().unescape(_json_iter->string_buf_loc());
+simdjson_warn_unused simdjson_inline simdjson_result<std::string_view> value_iterator::get_string() noexcept {
+  return get_raw_json_string().unescape(json_iter());
 }
-simdjson_warn_unused simdjson_really_inline simdjson_result<raw_json_string> value_iterator::get_raw_json_string() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<raw_json_string> value_iterator::get_raw_json_string() noexcept {
   auto json = peek_scalar("string");
   if (*json != '"') { return incorrect_type_error("Not a string"); }
   advance_scalar("string");
   return raw_json_string(json+1);
 }
-simdjson_warn_unused simdjson_really_inline simdjson_result<uint64_t> value_iterator::get_uint64() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<uint64_t> value_iterator::get_uint64() noexcept {
   auto result = numberparsing::parse_unsigned(peek_non_root_scalar("uint64"));
   if(result.error() == SUCCESS) { advance_non_root_scalar("uint64"); }
   return result;
 }
-simdjson_warn_unused simdjson_really_inline simdjson_result<uint64_t> value_iterator::get_uint64_in_string() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<uint64_t> value_iterator::get_uint64_in_string() noexcept {
   auto result = numberparsing::parse_unsigned_in_string(peek_non_root_scalar("uint64"));
   if(result.error() == SUCCESS) { advance_non_root_scalar("uint64"); }
   return result;
 }
-simdjson_warn_unused simdjson_really_inline simdjson_result<int64_t> value_iterator::get_int64() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<int64_t> value_iterator::get_int64() noexcept {
   auto result = numberparsing::parse_integer(peek_non_root_scalar("int64"));
   if(result.error() == SUCCESS) { advance_non_root_scalar("int64"); }
   return result;
 }
-simdjson_warn_unused simdjson_really_inline simdjson_result<int64_t> value_iterator::get_int64_in_string() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<int64_t> value_iterator::get_int64_in_string() noexcept {
   auto result = numberparsing::parse_integer_in_string(peek_non_root_scalar("int64"));
   if(result.error() == SUCCESS) { advance_non_root_scalar("int64"); }
   return result;
 }
-simdjson_warn_unused simdjson_really_inline simdjson_result<double> value_iterator::get_double() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<double> value_iterator::get_double() noexcept {
   auto result = numberparsing::parse_double(peek_non_root_scalar("double"));
   if(result.error() == SUCCESS) { advance_non_root_scalar("double"); }
   return result;
 }
-simdjson_warn_unused simdjson_really_inline simdjson_result<double> value_iterator::get_double_in_string() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<double> value_iterator::get_double_in_string() noexcept {
   auto result = numberparsing::parse_double_in_string(peek_non_root_scalar("double"));
   if(result.error() == SUCCESS) { advance_non_root_scalar("double"); }
   return result;
 }
-simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator::get_bool() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<bool> value_iterator::get_bool() noexcept {
   auto result = parse_bool(peek_non_root_scalar("bool"));
   if(result.error() == SUCCESS) { advance_non_root_scalar("bool"); }
   return result;
 }
-simdjson_really_inline bool value_iterator::is_null() noexcept {
-  auto result = parse_null(peek_non_root_scalar("null"));
-  if(result) { advance_non_root_scalar("null"); }
-  return result;
+simdjson_inline simdjson_result<bool> value_iterator::is_null() noexcept {
+  bool is_null_value;
+  SIMDJSON_TRY(parse_null(peek_non_root_scalar("null")).get(is_null_value));
+  if(is_null_value) { advance_non_root_scalar("null"); }
+  return is_null_value;
 }
-simdjson_really_inline bool value_iterator::is_negative() noexcept {
+simdjson_inline bool value_iterator::is_negative() noexcept {
   return numberparsing::is_negative(peek_non_root_scalar("numbersign"));
 }
-simdjson_really_inline bool value_iterator::is_root_negative() noexcept {
+simdjson_inline bool value_iterator::is_root_negative() noexcept {
   return numberparsing::is_negative(peek_root_scalar("numbersign"));
 }
-simdjson_really_inline simdjson_result<bool> value_iterator::is_integer() noexcept {
+simdjson_inline simdjson_result<bool> value_iterator::is_integer() noexcept {
   return numberparsing::is_integer(peek_non_root_scalar("integer"));
 }
-simdjson_really_inline simdjson_result<number_type> value_iterator::get_number_type() noexcept {
+simdjson_inline simdjson_result<number_type> value_iterator::get_number_type() noexcept {
   return numberparsing::get_number_type(peek_non_root_scalar("integer"));
 }
-simdjson_really_inline simdjson_result<number> value_iterator::get_number() noexcept {
+simdjson_inline simdjson_result<number> value_iterator::get_number() noexcept {
   number num;
   error_code error =  numberparsing::parse_number(peek_non_root_scalar("number"), num);
   if(error) { return error; }
   return num;
 }
 
-simdjson_really_inline simdjson_result<bool> value_iterator::is_root_integer() noexcept {
+simdjson_inline simdjson_result<bool> value_iterator::is_root_integer() noexcept {
   auto max_len = peek_start_length();
   auto json = peek_root_scalar("is_root_integer");
   uint8_t tmpbuf[20+1]; // <20 digits> is the longest possible unsigned integer
   if (!_json_iter->copy_to_buffer(json, max_len, tmpbuf)) {
     return false; // if there are more than 20 characters, it cannot be represented as an integer.
   }
-  return numberparsing::is_integer(tmpbuf);
+  auto answer = numberparsing::is_integer(tmpbuf);
+  // If the parsing was a success, we must still check that it is
+  // a single scalar. Note that we parse first because of cases like '[]' where
+  // getting TRAILING_CONTENT is wrong.
+  if((answer.error() == SUCCESS) && (!_json_iter->is_single_token())) { return TRAILING_CONTENT; }
+  return answer;
 }
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number_type> value_iterator::get_root_number_type() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number_type> value_iterator::get_root_number_type() noexcept {
+  if (!_json_iter->is_single_token()) { return TRAILING_CONTENT; }
   auto max_len = peek_start_length();
   auto json = peek_root_scalar("number");
   // Per https://www.exploringbinary.com/maximum-number-of-decimal-digits-in-binary-floating-point-numbers/,
@@ -29070,9 +28472,14 @@ simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand
     logger::log_error(*_json_iter, start_position(), depth(), "Root number more than 1082 characters");
     return NUMBER_ERROR;
   }
-  return numberparsing::get_number_type(tmpbuf);
+  // If the parsing was a success, we must still check that it is
+  // a single scalar. Note that we parse first because of cases like '[]' where
+  // getting TRAILING_CONTENT is wrong.
+  auto answer = numberparsing::get_number_type(tmpbuf);
+  if((answer.error() == SUCCESS) && (!_json_iter->is_single_token())) { return TRAILING_CONTENT; }
+  return answer;
 }
-simdjson_really_inline simdjson_result<number> value_iterator::get_root_number() noexcept {
+simdjson_inline simdjson_result<number> value_iterator::get_root_number() noexcept {
   auto max_len = peek_start_length();
   auto json = peek_root_scalar("number");
   // Per https://www.exploringbinary.com/maximum-number-of-decimal-digits-in-binary-floating-point-numbers/,
@@ -29086,17 +28493,18 @@ simdjson_really_inline simdjson_result<number> value_iterator::get_root_number()
   number num;
   error_code error =  numberparsing::parse_number(tmpbuf, num);
   if(error) { return error; }
+  if (!_json_iter->is_single_token()) { return TRAILING_CONTENT; }
   advance_root_scalar("number");
   return num;
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<std::string_view> value_iterator::get_root_string() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<std::string_view> value_iterator::get_root_string() noexcept {
   return get_string();
 }
-simdjson_warn_unused simdjson_really_inline simdjson_result<raw_json_string> value_iterator::get_root_raw_json_string() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<raw_json_string> value_iterator::get_root_raw_json_string() noexcept {
   return get_raw_json_string();
 }
-simdjson_warn_unused simdjson_really_inline simdjson_result<uint64_t> value_iterator::get_root_uint64() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<uint64_t> value_iterator::get_root_uint64() noexcept {
   auto max_len = peek_start_length();
   auto json = peek_root_scalar("uint64");
   uint8_t tmpbuf[20+1]; // <20 digits> is the longest possible unsigned integer
@@ -29105,10 +28513,13 @@ simdjson_warn_unused simdjson_really_inline simdjson_result<uint64_t> value_iter
     return NUMBER_ERROR;
   }
   auto result = numberparsing::parse_unsigned(tmpbuf);
-  if(result.error() == SUCCESS) { advance_root_scalar("uint64"); }
+  if(result.error() == SUCCESS) {
+    if (!_json_iter->is_single_token()) { return TRAILING_CONTENT; }
+    advance_root_scalar("uint64");
+  }
   return result;
 }
-simdjson_warn_unused simdjson_really_inline simdjson_result<uint64_t> value_iterator::get_root_uint64_in_string() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<uint64_t> value_iterator::get_root_uint64_in_string() noexcept {
   auto max_len = peek_start_length();
   auto json = peek_root_scalar("uint64");
   uint8_t tmpbuf[20+1]; // <20 digits> is the longest possible unsigned integer
@@ -29117,10 +28528,13 @@ simdjson_warn_unused simdjson_really_inline simdjson_result<uint64_t> value_iter
     return NUMBER_ERROR;
   }
   auto result = numberparsing::parse_unsigned_in_string(tmpbuf);
-  if(result.error() == SUCCESS) { advance_root_scalar("uint64"); }
+  if(result.error() == SUCCESS) {
+    if (!_json_iter->is_single_token()) { return TRAILING_CONTENT; }
+    advance_root_scalar("uint64");
+  }
   return result;
 }
-simdjson_warn_unused simdjson_really_inline simdjson_result<int64_t> value_iterator::get_root_int64() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<int64_t> value_iterator::get_root_int64() noexcept {
   auto max_len = peek_start_length();
   auto json = peek_root_scalar("int64");
   uint8_t tmpbuf[20+1]; // -<19 digits> is the longest possible integer
@@ -29130,10 +28544,13 @@ simdjson_warn_unused simdjson_really_inline simdjson_result<int64_t> value_itera
   }
 
   auto result = numberparsing::parse_integer(tmpbuf);
-  if(result.error() == SUCCESS) { advance_root_scalar("int64"); }
+  if(result.error() == SUCCESS) {
+    if (!_json_iter->is_single_token()) { return TRAILING_CONTENT; }
+    advance_root_scalar("int64");
+  }
   return result;
 }
-simdjson_warn_unused simdjson_really_inline simdjson_result<int64_t> value_iterator::get_root_int64_in_string() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<int64_t> value_iterator::get_root_int64_in_string() noexcept {
   auto max_len = peek_start_length();
   auto json = peek_root_scalar("int64");
   uint8_t tmpbuf[20+1]; // -<19 digits> is the longest possible integer
@@ -29143,10 +28560,13 @@ simdjson_warn_unused simdjson_really_inline simdjson_result<int64_t> value_itera
   }
 
   auto result = numberparsing::parse_integer_in_string(tmpbuf);
-  if(result.error() == SUCCESS) { advance_root_scalar("int64"); }
+  if(result.error() == SUCCESS) {
+    if (!_json_iter->is_single_token()) { return TRAILING_CONTENT; }
+    advance_root_scalar("int64");
+  }
   return result;
 }
-simdjson_warn_unused simdjson_really_inline simdjson_result<double> value_iterator::get_root_double() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<double> value_iterator::get_root_double() noexcept {
   auto max_len = peek_start_length();
   auto json = peek_root_scalar("double");
   // Per https://www.exploringbinary.com/maximum-number-of-decimal-digits-in-binary-floating-point-numbers/,
@@ -29158,11 +28578,14 @@ simdjson_warn_unused simdjson_really_inline simdjson_result<double> value_iterat
     return NUMBER_ERROR;
   }
   auto result = numberparsing::parse_double(tmpbuf);
-  if(result.error() == SUCCESS) { advance_root_scalar("double"); }
+  if(result.error() == SUCCESS) {
+    if (!_json_iter->is_single_token()) { return TRAILING_CONTENT; }
+    advance_root_scalar("double");
+  }
   return result;
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<double> value_iterator::get_root_double_in_string() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<double> value_iterator::get_root_double_in_string() noexcept {
   auto max_len = peek_start_length();
   auto json = peek_root_scalar("double");
   // Per https://www.exploringbinary.com/maximum-number-of-decimal-digits-in-binary-floating-point-numbers/,
@@ -29174,19 +28597,27 @@ simdjson_warn_unused simdjson_really_inline simdjson_result<double> value_iterat
     return NUMBER_ERROR;
   }
   auto result = numberparsing::parse_double_in_string(tmpbuf);
-  if(result.error() == SUCCESS) { advance_root_scalar("double"); }
+  if(result.error() == SUCCESS) {
+    if (!_json_iter->is_single_token()) { return TRAILING_CONTENT; }
+    advance_root_scalar("double");
+  }
   return result;
 }
-simdjson_warn_unused simdjson_really_inline simdjson_result<bool> value_iterator::get_root_bool() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<bool> value_iterator::get_root_bool() noexcept {
   auto max_len = peek_start_length();
   auto json = peek_root_scalar("bool");
   uint8_t tmpbuf[5+1];
   if (!_json_iter->copy_to_buffer(json, max_len, tmpbuf)) { return incorrect_type_error("Not a boolean"); }
   auto result = parse_bool(tmpbuf);
-  if(result.error() == SUCCESS) { advance_root_scalar("bool"); }
+  if(result.error() == SUCCESS) {
+    if (!_json_iter->is_single_token()) { return TRAILING_CONTENT; }
+    advance_root_scalar("bool");
+  }
   return result;
 }
-simdjson_really_inline bool value_iterator::is_root_null() noexcept {
+simdjson_inline bool value_iterator::is_root_null() noexcept {
+  // If there is trailing content, then the document is not null.
+  if (!_json_iter->is_single_token()) { return false; }
   auto max_len = peek_start_length();
   auto json = peek_root_scalar("null");
   bool result = (max_len >= 4 && !atomparsing::str4ncmp(json, "null") &&
@@ -29195,14 +28626,14 @@ simdjson_really_inline bool value_iterator::is_root_null() noexcept {
   return result;
 }
 
-simdjson_warn_unused simdjson_really_inline error_code value_iterator::skip_child() noexcept {
+simdjson_warn_unused simdjson_inline error_code value_iterator::skip_child() noexcept {
   SIMDJSON_ASSUME( _json_iter->token._position > _start_position );
   SIMDJSON_ASSUME( _json_iter->_depth >= _depth );
 
   return _json_iter->skip_child(depth());
 }
 
-simdjson_really_inline value_iterator value_iterator::child() const noexcept {
+simdjson_inline value_iterator value_iterator::child() const noexcept {
   assert_at_child();
   return { _json_iter, depth()+1, _json_iter->token.position() };
 }
@@ -29212,52 +28643,52 @@ simdjson_really_inline value_iterator value_iterator::child() const noexcept {
 // marked non-inline.
 SIMDJSON_PUSH_DISABLE_WARNINGS
 SIMDJSON_DISABLE_STRICT_OVERFLOW_WARNING
-simdjson_really_inline bool value_iterator::is_open() const noexcept {
+simdjson_inline bool value_iterator::is_open() const noexcept {
   return _json_iter->depth() >= depth();
 }
 SIMDJSON_POP_DISABLE_WARNINGS
 
-simdjson_really_inline bool value_iterator::at_end() const noexcept {
+simdjson_inline bool value_iterator::at_end() const noexcept {
   return _json_iter->at_end();
 }
 
-simdjson_really_inline bool value_iterator::at_start() const noexcept {
+simdjson_inline bool value_iterator::at_start() const noexcept {
   return _json_iter->token.position() == start_position();
 }
 
-simdjson_really_inline bool value_iterator::at_first_field() const noexcept {
+simdjson_inline bool value_iterator::at_first_field() const noexcept {
   SIMDJSON_ASSUME( _json_iter->token._position > _start_position );
   return _json_iter->token.position() == start_position() + 1;
 }
 
-simdjson_really_inline void value_iterator::abandon() noexcept {
+simdjson_inline void value_iterator::abandon() noexcept {
   _json_iter->abandon();
 }
 
-simdjson_warn_unused simdjson_really_inline depth_t value_iterator::depth() const noexcept {
+simdjson_warn_unused simdjson_inline depth_t value_iterator::depth() const noexcept {
   return _depth;
 }
-simdjson_warn_unused simdjson_really_inline error_code value_iterator::error() const noexcept {
+simdjson_warn_unused simdjson_inline error_code value_iterator::error() const noexcept {
   return _json_iter->error;
 }
-simdjson_warn_unused simdjson_really_inline uint8_t *&value_iterator::string_buf_loc() noexcept {
+simdjson_warn_unused simdjson_inline uint8_t *&value_iterator::string_buf_loc() noexcept {
   return _json_iter->string_buf_loc();
 }
-simdjson_warn_unused simdjson_really_inline const json_iterator &value_iterator::json_iter() const noexcept {
+simdjson_warn_unused simdjson_inline const json_iterator &value_iterator::json_iter() const noexcept {
   return *_json_iter;
 }
-simdjson_warn_unused simdjson_really_inline json_iterator &value_iterator::json_iter() noexcept {
+simdjson_warn_unused simdjson_inline json_iterator &value_iterator::json_iter() noexcept {
   return *_json_iter;
 }
 
-simdjson_really_inline const uint8_t *value_iterator::peek_start() const noexcept {
+simdjson_inline const uint8_t *value_iterator::peek_start() const noexcept {
   return _json_iter->peek(start_position());
 }
-simdjson_really_inline uint32_t value_iterator::peek_start_length() const noexcept {
+simdjson_inline uint32_t value_iterator::peek_start_length() const noexcept {
   return _json_iter->peek_length(start_position());
 }
 
-simdjson_really_inline const uint8_t *value_iterator::peek_scalar(const char *type) noexcept {
+simdjson_inline const uint8_t *value_iterator::peek_scalar(const char *type) noexcept {
   logger::log_value(*_json_iter, start_position(), depth(), type);
   // If we're not at the position anymore, we don't want to advance the cursor.
   if (!is_at_start()) { return peek_start(); }
@@ -29267,7 +28698,7 @@ simdjson_really_inline const uint8_t *value_iterator::peek_scalar(const char *ty
   return _json_iter->peek();
 }
 
-simdjson_really_inline void value_iterator::advance_scalar(const char *type) noexcept {
+simdjson_inline void value_iterator::advance_scalar(const char *type) noexcept {
   logger::log_value(*_json_iter, start_position(), depth(), type);
   // If we're not at the position anymore, we don't want to advance the cursor.
   if (!is_at_start()) { return; }
@@ -29278,12 +28709,12 @@ simdjson_really_inline void value_iterator::advance_scalar(const char *type) noe
   _json_iter->ascend_to(depth()-1);
 }
 
-simdjson_really_inline error_code value_iterator::start_container(uint8_t start_char, const char *incorrect_type_message, const char *type) noexcept {
+simdjson_inline error_code value_iterator::start_container(uint8_t start_char, const char *incorrect_type_message, const char *type) noexcept {
   logger::log_start_value(*_json_iter, start_position(), depth(), type);
   // If we're not at the position anymore, we don't want to advance the cursor.
   const uint8_t *json;
   if (!is_at_start()) {
-#ifdef SIMDJSON_DEVELOPMENT_CHECKS
+#if SIMDJSON_DEVELOPMENT_CHECKS
     if (!is_at_iterator_start()) { return OUT_OF_ORDER_ITERATION; }
 #endif
     json = peek_start();
@@ -29305,14 +28736,14 @@ simdjson_really_inline error_code value_iterator::start_container(uint8_t start_
 }
 
 
-simdjson_really_inline const uint8_t *value_iterator::peek_root_scalar(const char *type) noexcept {
+simdjson_inline const uint8_t *value_iterator::peek_root_scalar(const char *type) noexcept {
   logger::log_value(*_json_iter, start_position(), depth(), type);
   if (!is_at_start()) { return peek_start(); }
 
   assert_at_root();
   return _json_iter->peek();
 }
-simdjson_really_inline const uint8_t *value_iterator::peek_non_root_scalar(const char *type) noexcept {
+simdjson_inline const uint8_t *value_iterator::peek_non_root_scalar(const char *type) noexcept {
   logger::log_value(*_json_iter, start_position(), depth(), type);
   if (!is_at_start()) { return peek_start(); }
 
@@ -29320,7 +28751,7 @@ simdjson_really_inline const uint8_t *value_iterator::peek_non_root_scalar(const
   return _json_iter->peek();
 }
 
-simdjson_really_inline void value_iterator::advance_root_scalar(const char *type) noexcept {
+simdjson_inline void value_iterator::advance_root_scalar(const char *type) noexcept {
   logger::log_value(*_json_iter, start_position(), depth(), type);
   if (!is_at_start()) { return; }
 
@@ -29328,7 +28759,7 @@ simdjson_really_inline void value_iterator::advance_root_scalar(const char *type
   _json_iter->return_current_and_advance();
   _json_iter->ascend_to(depth()-1);
 }
-simdjson_really_inline void value_iterator::advance_non_root_scalar(const char *type) noexcept {
+simdjson_inline void value_iterator::advance_non_root_scalar(const char *type) noexcept {
   logger::log_value(*_json_iter, start_position(), depth(), type);
   if (!is_at_start()) { return; }
 
@@ -29337,23 +28768,23 @@ simdjson_really_inline void value_iterator::advance_non_root_scalar(const char *
   _json_iter->ascend_to(depth()-1);
 }
 
-simdjson_really_inline error_code value_iterator::incorrect_type_error(const char *message) const noexcept {
+simdjson_inline error_code value_iterator::incorrect_type_error(const char *message) const noexcept {
   logger::log_error(*_json_iter, start_position(), depth(), message);
   return INCORRECT_TYPE;
 }
 
-simdjson_really_inline bool value_iterator::is_at_start() const noexcept {
+simdjson_inline bool value_iterator::is_at_start() const noexcept {
   return position() == start_position();
 }
 
-simdjson_really_inline bool value_iterator::is_at_key() const noexcept {
+simdjson_inline bool value_iterator::is_at_key() const noexcept {
   // Keys are at the same depth as the object.
   // Note here that we could be safer and check that we are within an object,
   // but we do not.
   return _depth == _json_iter->_depth && *_json_iter->peek() == '"';
 }
 
-simdjson_really_inline bool value_iterator::is_at_iterator_start() const noexcept {
+simdjson_inline bool value_iterator::is_at_iterator_start() const noexcept {
   // We can legitimately be either at the first value ([1]), or after the array if it's empty ([]).
   auto delta = position() - start_position();
   return delta == 1 || delta == 2;
@@ -29377,22 +28808,22 @@ inline void value_iterator::assert_at_next() const noexcept {
   SIMDJSON_ASSUME( _depth > 0 );
 }
 
-simdjson_really_inline void value_iterator::move_at_start() noexcept {
+simdjson_inline void value_iterator::move_at_start() noexcept {
   _json_iter->_depth = _depth;
   _json_iter->token.set_position(_start_position);
 }
 
-simdjson_really_inline void value_iterator::move_at_container_start() noexcept {
+simdjson_inline void value_iterator::move_at_container_start() noexcept {
   _json_iter->_depth = _depth;
   _json_iter->token.set_position(_start_position + 1);
 }
 
-simdjson_really_inline simdjson_result<bool> value_iterator::reset_array() noexcept {
+simdjson_inline simdjson_result<bool> value_iterator::reset_array() noexcept {
   move_at_container_start();
   return started_array();
 }
 
-simdjson_really_inline simdjson_result<bool> value_iterator::reset_object() noexcept {
+simdjson_inline simdjson_result<bool> value_iterator::reset_object() noexcept {
   move_at_container_start();
   return started_object();
 }
@@ -29417,11 +28848,11 @@ inline void value_iterator::assert_is_valid() const noexcept {
   SIMDJSON_ASSUME( _json_iter != nullptr );
 }
 
-simdjson_really_inline bool value_iterator::is_valid() const noexcept {
+simdjson_inline bool value_iterator::is_valid() const noexcept {
   return _json_iter != nullptr;
 }
 
-simdjson_really_inline simdjson_result<json_type> value_iterator::type() const noexcept {
+simdjson_inline simdjson_result<json_type> value_iterator::type() const noexcept {
   switch (*peek_start()) {
     case '{':
       return json_type::object;
@@ -29442,23 +28873,23 @@ simdjson_really_inline simdjson_result<json_type> value_iterator::type() const n
   }
 }
 
-simdjson_really_inline token_position value_iterator::start_position() const noexcept {
+simdjson_inline token_position value_iterator::start_position() const noexcept {
   return _start_position;
 }
 
-simdjson_really_inline token_position value_iterator::position() const noexcept {
+simdjson_inline token_position value_iterator::position() const noexcept {
   return _json_iter->position();
 }
 
-simdjson_really_inline token_position value_iterator::end_position() const noexcept {
+simdjson_inline token_position value_iterator::end_position() const noexcept {
   return _json_iter->end_position();
 }
 
-simdjson_really_inline token_position value_iterator::last_position() const noexcept {
+simdjson_inline token_position value_iterator::last_position() const noexcept {
   return _json_iter->last_position();
 }
 
-simdjson_really_inline error_code value_iterator::report_error(error_code error, const char *message) noexcept {
+simdjson_inline error_code value_iterator::report_error(error_code error, const char *message) noexcept {
   return _json_iter->report_error(error, message);
 }
 
@@ -29468,9 +28899,9 @@ simdjson_really_inline error_code value_iterator::report_error(error_code error,
 
 namespace simdjson {
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value_iterator>::simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value_iterator &&value) noexcept
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value_iterator>::simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value_iterator &&value) noexcept
     : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value_iterator>(std::forward<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value_iterator>(value)) {}
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value_iterator>::simdjson_result(error_code error) noexcept
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value_iterator>::simdjson_result(error_code error) noexcept
     : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value_iterator>(error) {}
 
 } // namespace simdjson
@@ -29480,21 +28911,21 @@ namespace simdjson {
 namespace SIMDJSON_BUILTIN_IMPLEMENTATION {
 namespace ondemand {
 
-simdjson_really_inline array_iterator::array_iterator(const value_iterator &_iter) noexcept
+simdjson_inline array_iterator::array_iterator(const value_iterator &_iter) noexcept
   : iter{_iter}
 {}
 
-simdjson_really_inline simdjson_result<value> array_iterator::operator*() noexcept {
+simdjson_inline simdjson_result<value> array_iterator::operator*() noexcept {
   if (iter.error()) { iter.abandon(); return iter.error(); }
   return value(iter.child());
 }
-simdjson_really_inline bool array_iterator::operator==(const array_iterator &other) const noexcept {
+simdjson_inline bool array_iterator::operator==(const array_iterator &other) const noexcept {
   return !(*this != other);
 }
-simdjson_really_inline bool array_iterator::operator!=(const array_iterator &) const noexcept {
+simdjson_inline bool array_iterator::operator!=(const array_iterator &) const noexcept {
   return iter.is_open();
 }
-simdjson_really_inline array_iterator &array_iterator::operator++() noexcept {
+simdjson_inline array_iterator &array_iterator::operator++() noexcept {
   error_code error;
   // PERF NOTE this is a safety rail ... users should exit loops as soon as they receive an error, so we'll never get here.
   // However, it does not seem to make a perf difference, so we add it out of an abundance of caution.
@@ -29510,31 +28941,31 @@ simdjson_really_inline array_iterator &array_iterator::operator++() noexcept {
 
 namespace simdjson {
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator>::simdjson_result(
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator>::simdjson_result(
   SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator &&value
 ) noexcept
   : SIMDJSON_BUILTIN_IMPLEMENTATION::implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator>(std::forward<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator>(value))
 {
   first.iter.assert_is_valid();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator>::simdjson_result(error_code error) noexcept
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator>::simdjson_result(error_code error) noexcept
   : SIMDJSON_BUILTIN_IMPLEMENTATION::implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator>({}, error)
 {
 }
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator>::operator*() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator>::operator*() noexcept {
   if (error()) { return error(); }
   return *first;
 }
-simdjson_really_inline bool simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator>::operator==(const simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> &other) const noexcept {
+simdjson_inline bool simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator>::operator==(const simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> &other) const noexcept {
   if (!first.iter.is_valid()) { return !error(); }
   return first == other.first;
 }
-simdjson_really_inline bool simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator>::operator!=(const simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> &other) const noexcept {
+simdjson_inline bool simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator>::operator!=(const simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> &other) const noexcept {
   if (!first.iter.is_valid()) { return error(); }
   return first != other.first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> &simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator>::operator++() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> &simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator>::operator++() noexcept {
   // Clear the error if there is one, so we don't yield it twice
   if (error()) { second = SUCCESS; return *this; }
   ++(first);
@@ -29552,11 +28983,11 @@ namespace ondemand {
 // object_iterator
 //
 
-simdjson_really_inline object_iterator::object_iterator(const value_iterator &_iter) noexcept
+simdjson_inline object_iterator::object_iterator(const value_iterator &_iter) noexcept
   : iter{_iter}
 {}
 
-simdjson_really_inline simdjson_result<field> object_iterator::operator*() noexcept {
+simdjson_inline simdjson_result<field> object_iterator::operator*() noexcept {
   error_code error = iter.error();
   if (error) { iter.abandon(); return error; }
   auto result = field::start(iter);
@@ -29565,14 +28996,16 @@ simdjson_really_inline simdjson_result<field> object_iterator::operator*() noexc
   if (result.error()) { iter.abandon(); }
   return result;
 }
-simdjson_really_inline bool object_iterator::operator==(const object_iterator &other) const noexcept {
+simdjson_inline bool object_iterator::operator==(const object_iterator &other) const noexcept {
   return !(*this != other);
 }
-simdjson_really_inline bool object_iterator::operator!=(const object_iterator &) const noexcept {
+simdjson_inline bool object_iterator::operator!=(const object_iterator &) const noexcept {
   return iter.is_open();
 }
 
-simdjson_really_inline object_iterator &object_iterator::operator++() noexcept {
+SIMDJSON_PUSH_DISABLE_WARNINGS
+SIMDJSON_DISABLE_STRICT_OVERFLOW_WARNING
+simdjson_inline object_iterator &object_iterator::operator++() noexcept {
   // TODO this is a safety rail ... users should exit loops as soon as they receive an error.
   // Nonetheless, let's see if performance is OK with this if statement--the compiler may give it to us for free.
   if (!iter.is_open()) { return *this; } // Iterator will be released if there is an error
@@ -29584,6 +29017,7 @@ simdjson_really_inline object_iterator &object_iterator::operator++() noexcept {
   if ((error = iter.has_next_field().get(has_value) )) { return *this; };
   return *this;
 }
+SIMDJSON_POP_DISABLE_WARNINGS
 
 //
 // ### Live States
@@ -29632,34 +29066,34 @@ simdjson_really_inline object_iterator &object_iterator::operator++() noexcept {
 
 namespace simdjson {
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator>::simdjson_result(
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator>::simdjson_result(
   SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator &&value
 ) noexcept
   : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator>(std::forward<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator>(value))
 {
   first.iter.assert_is_valid();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator>::simdjson_result(error_code error) noexcept
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator>::simdjson_result(error_code error) noexcept
   : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator>({}, error)
 {
 }
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator>::operator*() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator>::operator*() noexcept {
   if (error()) { return error(); }
   return *first;
 }
 // If we're iterating and there is an error, return the error once.
-simdjson_really_inline bool simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator>::operator==(const simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> &other) const noexcept {
+simdjson_inline bool simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator>::operator==(const simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> &other) const noexcept {
   if (!first.iter.is_valid()) { return !error(); }
   return first == other.first;
 }
 // If we're iterating and there is an error, return the error once.
-simdjson_really_inline bool simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator>::operator!=(const simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> &other) const noexcept {
+simdjson_inline bool simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator>::operator!=(const simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> &other) const noexcept {
   if (!first.iter.is_valid()) { return error(); }
   return first != other.first;
 }
 // Checks for ']' and ','
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> &simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator>::operator++() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> &simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator>::operator++() noexcept {
   // Clear the error if there is one, so we don't yield it twice
   if (error()) { second = SUCCESS; return *this; }
   ++first;
@@ -29711,45 +29145,45 @@ namespace ondemand {
 //   error == SUCCESS.
 //
 
-simdjson_really_inline array::array(const value_iterator &_iter) noexcept
+simdjson_inline array::array(const value_iterator &_iter) noexcept
   : iter{_iter}
 {
 }
 
-simdjson_really_inline simdjson_result<array> array::start(value_iterator &iter) noexcept {
+simdjson_inline simdjson_result<array> array::start(value_iterator &iter) noexcept {
   // We don't need to know if the array is empty to start iteration, but we do want to know if there
   // is an error--thus `simdjson_unused`.
   simdjson_unused bool has_value;
   SIMDJSON_TRY( iter.start_array().get(has_value) );
   return array(iter);
 }
-simdjson_really_inline simdjson_result<array> array::start_root(value_iterator &iter) noexcept {
+simdjson_inline simdjson_result<array> array::start_root(value_iterator &iter) noexcept {
   simdjson_unused bool has_value;
   SIMDJSON_TRY( iter.start_root_array().get(has_value) );
   return array(iter);
 }
-simdjson_really_inline simdjson_result<array> array::started(value_iterator &iter) noexcept {
+simdjson_inline simdjson_result<array> array::started(value_iterator &iter) noexcept {
   bool has_value;
   SIMDJSON_TRY(iter.started_array().get(has_value));
   return array(iter);
 }
 
-simdjson_really_inline simdjson_result<array_iterator> array::begin() noexcept {
-#ifdef SIMDJSON_DEVELOPMENT_CHECKS
+simdjson_inline simdjson_result<array_iterator> array::begin() noexcept {
+#if SIMDJSON_DEVELOPMENT_CHECKS
   if (!iter.is_at_iterator_start()) { return OUT_OF_ORDER_ITERATION; }
 #endif
   return array_iterator(iter);
 }
-simdjson_really_inline simdjson_result<array_iterator> array::end() noexcept {
+simdjson_inline simdjson_result<array_iterator> array::end() noexcept {
   return array_iterator(iter);
 }
-simdjson_really_inline error_code array::consume() noexcept {
+simdjson_inline error_code array::consume() noexcept {
   auto error = iter.json_iter().skip_child(iter.depth()-1);
   if(error) { iter.abandon(); }
   return error;
 }
 
-simdjson_really_inline simdjson_result<std::string_view> array::raw_json() noexcept {
+simdjson_inline simdjson_result<std::string_view> array::raw_json() noexcept {
   const uint8_t * starting_point{iter.peek_start()};
   auto error = consume();
   if(error) { return error; }
@@ -29760,8 +29194,9 @@ simdjson_really_inline simdjson_result<std::string_view> array::raw_json() noexc
   return std::string_view(reinterpret_cast<const char*>(starting_point), size_t(final_point - starting_point));
 }
 
+SIMDJSON_PUSH_DISABLE_WARNINGS
 SIMDJSON_DISABLE_STRICT_OVERFLOW_WARNING
-simdjson_really_inline simdjson_result<size_t> array::count_elements() & noexcept {
+simdjson_inline simdjson_result<size_t> array::count_elements() & noexcept {
   size_t count{0};
   // Important: we do not consume any of the values.
   for(simdjson_unused auto v : *this) { count++; }
@@ -29772,8 +29207,9 @@ simdjson_really_inline simdjson_result<size_t> array::count_elements() & noexcep
   iter.reset_array();
   return count;
 }
+SIMDJSON_POP_DISABLE_WARNINGS
 
-simdjson_really_inline simdjson_result<bool> array::is_empty() & noexcept {
+simdjson_inline simdjson_result<bool> array::is_empty() & noexcept {
   bool is_not_empty;
   auto error = iter.reset_array().get(is_not_empty);
   if(error) { return error; }
@@ -29820,7 +29256,7 @@ inline simdjson_result<value> array::at_pointer(std::string_view json_pointer) n
   return child;
 }
 
-simdjson_really_inline simdjson_result<value> array::at(size_t index) noexcept {
+simdjson_inline simdjson_result<value> array::at(size_t index) noexcept {
   size_t i = 0;
   for (auto value : *this) {
     if (i == index) { return value; }
@@ -29835,7 +29271,7 @@ simdjson_really_inline simdjson_result<value> array::at(size_t index) noexcept {
 
 namespace simdjson {
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array>::simdjson_result(
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array>::simdjson_result(
   SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array &&value
 ) noexcept
   : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array>(
@@ -29843,34 +29279,34 @@ simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand
     )
 {
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array>::simdjson_result(
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array>::simdjson_result(
   error_code error
 ) noexcept
   : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array>(error)
 {
 }
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array>::begin() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array>::begin() noexcept {
   if (error()) { return error(); }
   return first.begin();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array>::end() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array>::end() noexcept {
   if (error()) { return error(); }
   return first.end();
 }
-simdjson_really_inline  simdjson_result<size_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array>::count_elements() & noexcept {
+simdjson_inline  simdjson_result<size_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array>::count_elements() & noexcept {
   if (error()) { return error(); }
   return first.count_elements();
 }
-simdjson_really_inline  simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array>::is_empty() & noexcept {
+simdjson_inline  simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array>::is_empty() & noexcept {
   if (error()) { return error(); }
   return first.is_empty();
 }
-simdjson_really_inline  simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array>::at(size_t index) noexcept {
+simdjson_inline  simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array>::at(size_t index) noexcept {
   if (error()) { return error(); }
   return first.at(index);
 }
-simdjson_really_inline  simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array>::at_pointer(std::string_view json_pointer) noexcept {
+simdjson_inline  simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array>::at_pointer(std::string_view json_pointer) noexcept {
   if (error()) { return error(); }
   return first.at_pointer(json_pointer);
 }
@@ -29881,13 +29317,13 @@ namespace simdjson {
 namespace SIMDJSON_BUILTIN_IMPLEMENTATION {
 namespace ondemand {
 
-simdjson_really_inline document::document(ondemand::json_iterator &&_iter) noexcept
+simdjson_inline document::document(ondemand::json_iterator &&_iter) noexcept
   : iter{std::forward<json_iterator>(_iter)}
 {
   logger::log_start_value(iter, "document");
 }
 
-simdjson_really_inline document document::start(json_iterator &&iter) noexcept {
+simdjson_inline document document::start(json_iterator &&iter) noexcept {
   return document(std::forward<json_iterator>(iter));
 }
 
@@ -29910,20 +29346,20 @@ inline int32_t document::current_depth() const noexcept {
 inline bool document::is_alive() noexcept {
   return iter.is_alive();
 }
-simdjson_really_inline value_iterator document::resume_value_iterator() noexcept {
+simdjson_inline value_iterator document::resume_value_iterator() noexcept {
   return value_iterator(&iter, 1, iter.root_position());
 }
-simdjson_really_inline value_iterator document::get_root_value_iterator() noexcept {
+simdjson_inline value_iterator document::get_root_value_iterator() noexcept {
   return resume_value_iterator();
 }
-simdjson_really_inline simdjson_result<object> document::start_or_resume_object() noexcept {
+simdjson_inline simdjson_result<object> document::start_or_resume_object() noexcept {
   if (iter.at_root()) {
     return get_object();
   } else {
     return object::resume(resume_value_iterator());
   }
 }
-simdjson_really_inline simdjson_result<value> document::get_value() noexcept {
+simdjson_inline simdjson_result<value> document::get_value() noexcept {
   // Make sure we start any arrays or objects before returning, so that start_root_<object/array>()
   // gets called.
   iter.assert_at_document_depth();
@@ -29938,139 +29374,133 @@ simdjson_really_inline simdjson_result<value> document::get_value() noexcept {
       // return value(get_root_value_iterator());
   }
 }
-simdjson_really_inline simdjson_result<array> document::get_array() & noexcept {
+simdjson_inline simdjson_result<array> document::get_array() & noexcept {
   auto value = get_root_value_iterator();
   return array::start_root(value);
 }
-simdjson_really_inline simdjson_result<object> document::get_object() & noexcept {
+simdjson_inline simdjson_result<object> document::get_object() & noexcept {
   auto value = get_root_value_iterator();
   return object::start_root(value);
 }
-simdjson_really_inline simdjson_result<uint64_t> document::get_uint64() noexcept {
+simdjson_inline simdjson_result<uint64_t> document::get_uint64() noexcept {
   return get_root_value_iterator().get_root_uint64();
 }
-simdjson_really_inline simdjson_result<uint64_t> document::get_uint64_in_string() noexcept {
+simdjson_inline simdjson_result<uint64_t> document::get_uint64_in_string() noexcept {
   return get_root_value_iterator().get_root_uint64_in_string();
 }
-simdjson_really_inline simdjson_result<int64_t> document::get_int64() noexcept {
+simdjson_inline simdjson_result<int64_t> document::get_int64() noexcept {
   return get_root_value_iterator().get_root_int64();
 }
-simdjson_really_inline simdjson_result<int64_t> document::get_int64_in_string() noexcept {
+simdjson_inline simdjson_result<int64_t> document::get_int64_in_string() noexcept {
   return get_root_value_iterator().get_root_int64_in_string();
 }
-simdjson_really_inline simdjson_result<double> document::get_double() noexcept {
+simdjson_inline simdjson_result<double> document::get_double() noexcept {
   return get_root_value_iterator().get_root_double();
 }
-simdjson_really_inline simdjson_result<double> document::get_double_in_string() noexcept {
+simdjson_inline simdjson_result<double> document::get_double_in_string() noexcept {
   return get_root_value_iterator().get_root_double_in_string();
 }
-simdjson_really_inline simdjson_result<std::string_view> document::get_string() noexcept {
+simdjson_inline simdjson_result<std::string_view> document::get_string() noexcept {
   return get_root_value_iterator().get_root_string();
 }
-simdjson_really_inline simdjson_result<raw_json_string> document::get_raw_json_string() noexcept {
+simdjson_inline simdjson_result<raw_json_string> document::get_raw_json_string() noexcept {
   return get_root_value_iterator().get_root_raw_json_string();
 }
-simdjson_really_inline simdjson_result<bool> document::get_bool() noexcept {
+simdjson_inline simdjson_result<bool> document::get_bool() noexcept {
   return get_root_value_iterator().get_root_bool();
 }
-simdjson_really_inline bool document::is_null() noexcept {
+simdjson_inline simdjson_result<bool> document::is_null() noexcept {
   return get_root_value_iterator().is_root_null();
 }
 
-template<> simdjson_really_inline simdjson_result<array> document::get() & noexcept { return get_array(); }
-template<> simdjson_really_inline simdjson_result<object> document::get() & noexcept { return get_object(); }
-template<> simdjson_really_inline simdjson_result<raw_json_string> document::get() & noexcept { return get_raw_json_string(); }
-template<> simdjson_really_inline simdjson_result<std::string_view> document::get() & noexcept { return get_string(); }
-template<> simdjson_really_inline simdjson_result<double> document::get() & noexcept { return get_double(); }
-template<> simdjson_really_inline simdjson_result<uint64_t> document::get() & noexcept { return get_uint64(); }
-template<> simdjson_really_inline simdjson_result<int64_t> document::get() & noexcept { return get_int64(); }
-template<> simdjson_really_inline simdjson_result<bool> document::get() & noexcept { return get_bool(); }
-template<> simdjson_really_inline simdjson_result<value> document::get() & noexcept { return get_value(); }
-
-template<> simdjson_really_inline simdjson_result<raw_json_string> document::get() && noexcept { return get_raw_json_string(); }
-template<> simdjson_really_inline simdjson_result<std::string_view> document::get() && noexcept { return get_string(); }
-template<> simdjson_really_inline simdjson_result<double> document::get() && noexcept { return std::forward<document>(*this).get_double(); }
-template<> simdjson_really_inline simdjson_result<uint64_t> document::get() && noexcept { return std::forward<document>(*this).get_uint64(); }
-template<> simdjson_really_inline simdjson_result<int64_t> document::get() && noexcept { return std::forward<document>(*this).get_int64(); }
-template<> simdjson_really_inline simdjson_result<bool> document::get() && noexcept { return std::forward<document>(*this).get_bool(); }
-template<> simdjson_really_inline simdjson_result<value> document::get() && noexcept { return get_value(); }
-
-template<typename T> simdjson_really_inline error_code document::get(T &out) & noexcept {
+template<> simdjson_inline simdjson_result<array> document::get() & noexcept { return get_array(); }
+template<> simdjson_inline simdjson_result<object> document::get() & noexcept { return get_object(); }
+template<> simdjson_inline simdjson_result<raw_json_string> document::get() & noexcept { return get_raw_json_string(); }
+template<> simdjson_inline simdjson_result<std::string_view> document::get() & noexcept { return get_string(); }
+template<> simdjson_inline simdjson_result<double> document::get() & noexcept { return get_double(); }
+template<> simdjson_inline simdjson_result<uint64_t> document::get() & noexcept { return get_uint64(); }
+template<> simdjson_inline simdjson_result<int64_t> document::get() & noexcept { return get_int64(); }
+template<> simdjson_inline simdjson_result<bool> document::get() & noexcept { return get_bool(); }
+template<> simdjson_inline simdjson_result<value> document::get() & noexcept { return get_value(); }
+
+template<> simdjson_inline simdjson_result<raw_json_string> document::get() && noexcept { return get_raw_json_string(); }
+template<> simdjson_inline simdjson_result<std::string_view> document::get() && noexcept { return get_string(); }
+template<> simdjson_inline simdjson_result<double> document::get() && noexcept { return std::forward<document>(*this).get_double(); }
+template<> simdjson_inline simdjson_result<uint64_t> document::get() && noexcept { return std::forward<document>(*this).get_uint64(); }
+template<> simdjson_inline simdjson_result<int64_t> document::get() && noexcept { return std::forward<document>(*this).get_int64(); }
+template<> simdjson_inline simdjson_result<bool> document::get() && noexcept { return std::forward<document>(*this).get_bool(); }
+template<> simdjson_inline simdjson_result<value> document::get() && noexcept { return get_value(); }
+
+template<typename T> simdjson_inline error_code document::get(T &out) & noexcept {
   return get<T>().get(out);
 }
-template<typename T> simdjson_really_inline error_code document::get(T &out) && noexcept {
+template<typename T> simdjson_inline error_code document::get(T &out) && noexcept {
   return std::forward<document>(*this).get<T>().get(out);
 }
 
 #if SIMDJSON_EXCEPTIONS
-simdjson_really_inline document::operator array() & noexcept(false) { return get_array(); }
-simdjson_really_inline document::operator object() & noexcept(false) { return get_object(); }
-simdjson_really_inline document::operator uint64_t() noexcept(false) { return get_uint64(); }
-simdjson_really_inline document::operator int64_t() noexcept(false) { return get_int64(); }
-simdjson_really_inline document::operator double() noexcept(false) { return get_double(); }
-simdjson_really_inline document::operator std::string_view() noexcept(false) { return get_string(); }
-simdjson_really_inline document::operator raw_json_string() noexcept(false) { return get_raw_json_string(); }
-simdjson_really_inline document::operator bool() noexcept(false) { return get_bool(); }
-simdjson_really_inline document::operator value() noexcept(false) { return get_value(); }
+simdjson_inline document::operator array() & noexcept(false) { return get_array(); }
+simdjson_inline document::operator object() & noexcept(false) { return get_object(); }
+simdjson_inline document::operator uint64_t() noexcept(false) { return get_uint64(); }
+simdjson_inline document::operator int64_t() noexcept(false) { return get_int64(); }
+simdjson_inline document::operator double() noexcept(false) { return get_double(); }
+simdjson_inline document::operator std::string_view() noexcept(false) { return get_string(); }
+simdjson_inline document::operator raw_json_string() noexcept(false) { return get_raw_json_string(); }
+simdjson_inline document::operator bool() noexcept(false) { return get_bool(); }
+simdjson_inline document::operator value() noexcept(false) { return get_value(); }
 
 #endif
-simdjson_really_inline simdjson_result<size_t> document::count_elements() & noexcept {
+simdjson_inline simdjson_result<size_t> document::count_elements() & noexcept {
   auto a = get_array();
   simdjson_result<size_t> answer = a.count_elements();
   /* If there was an array, we are now left pointing at its first element. */
-  if(answer.error() == SUCCESS) {
-    iter._depth = 1 ; /* undoing the increment so we go back at the doc depth.*/
-    iter.assert_at_document_depth();
-  }
+  if(answer.error() == SUCCESS) { rewind(); }
   return answer;
 }
-simdjson_really_inline simdjson_result<size_t> document::count_fields() & noexcept {
+simdjson_inline simdjson_result<size_t> document::count_fields() & noexcept {
   auto a = get_object();
   simdjson_result<size_t> answer = a.count_fields();
-  /* If there was an array, we are now left pointing at its first element. */
-  if(answer.error() == SUCCESS) {
-    iter._depth = 1 ; /* undoing the increment so we go back at the doc depth.*/
-    iter.assert_at_document_depth();
-  }
+  /* If there was an object, we are now left pointing at its first element. */
+  if(answer.error() == SUCCESS) { rewind(); }
   return answer;
 }
-simdjson_really_inline simdjson_result<value> document::at(size_t index) & noexcept {
+simdjson_inline simdjson_result<value> document::at(size_t index) & noexcept {
   auto a = get_array();
   return a.at(index);
 }
-simdjson_really_inline simdjson_result<array_iterator> document::begin() & noexcept {
+simdjson_inline simdjson_result<array_iterator> document::begin() & noexcept {
   return get_array().begin();
 }
-simdjson_really_inline simdjson_result<array_iterator> document::end() & noexcept {
+simdjson_inline simdjson_result<array_iterator> document::end() & noexcept {
   return {};
 }
 
-simdjson_really_inline simdjson_result<value> document::find_field(std::string_view key) & noexcept {
+simdjson_inline simdjson_result<value> document::find_field(std::string_view key) & noexcept {
   return start_or_resume_object().find_field(key);
 }
-simdjson_really_inline simdjson_result<value> document::find_field(const char *key) & noexcept {
+simdjson_inline simdjson_result<value> document::find_field(const char *key) & noexcept {
   return start_or_resume_object().find_field(key);
 }
-simdjson_really_inline simdjson_result<value> document::find_field_unordered(std::string_view key) & noexcept {
+simdjson_inline simdjson_result<value> document::find_field_unordered(std::string_view key) & noexcept {
   return start_or_resume_object().find_field_unordered(key);
 }
-simdjson_really_inline simdjson_result<value> document::find_field_unordered(const char *key) & noexcept {
+simdjson_inline simdjson_result<value> document::find_field_unordered(const char *key) & noexcept {
   return start_or_resume_object().find_field_unordered(key);
 }
-simdjson_really_inline simdjson_result<value> document::operator[](std::string_view key) & noexcept {
+simdjson_inline simdjson_result<value> document::operator[](std::string_view key) & noexcept {
   return start_or_resume_object()[key];
 }
-simdjson_really_inline simdjson_result<value> document::operator[](const char *key) & noexcept {
+simdjson_inline simdjson_result<value> document::operator[](const char *key) & noexcept {
   return start_or_resume_object()[key];
 }
 
-simdjson_really_inline error_code document::consume() noexcept {
+simdjson_inline error_code document::consume() noexcept {
   auto error = iter.skip_child(0);
   if(error) { iter.abandon(); }
   return error;
 }
 
-simdjson_really_inline simdjson_result<std::string_view> document::raw_json() noexcept {
+simdjson_inline simdjson_result<std::string_view> document::raw_json() noexcept {
   auto _iter = get_root_value_iterator();
   const uint8_t * starting_point{_iter.peek_start()};
   auto error = consume();
@@ -30082,40 +29512,40 @@ simdjson_really_inline simdjson_result<std::string_view> document::raw_json() no
   return std::string_view(reinterpret_cast<const char*>(starting_point), size_t(final_point - starting_point));
 }
 
-simdjson_really_inline simdjson_result<json_type> document::type() noexcept {
+simdjson_inline simdjson_result<json_type> document::type() noexcept {
   return get_root_value_iterator().type();
 }
 
-simdjson_really_inline simdjson_result<bool> document::is_scalar() noexcept {
+simdjson_inline simdjson_result<bool> document::is_scalar() noexcept {
   json_type this_type;
   auto error = type().get(this_type);
   if(error) { return error; }
   return ! ((this_type == json_type::array) || (this_type == json_type::object));
 }
 
-simdjson_really_inline bool document::is_negative() noexcept {
+simdjson_inline bool document::is_negative() noexcept {
   return get_root_value_iterator().is_root_negative();
 }
 
-simdjson_really_inline simdjson_result<bool> document::is_integer() noexcept {
+simdjson_inline simdjson_result<bool> document::is_integer() noexcept {
   return get_root_value_iterator().is_root_integer();
 }
 
-simdjson_really_inline simdjson_result<number_type> document::get_number_type() noexcept {
+simdjson_inline simdjson_result<number_type> document::get_number_type() noexcept {
   return get_root_value_iterator().get_root_number_type();
 }
 
-simdjson_really_inline simdjson_result<number> document::get_number() noexcept {
+simdjson_inline simdjson_result<number> document::get_number() noexcept {
   return get_root_value_iterator().get_root_number();
 }
 
 
-simdjson_really_inline simdjson_result<std::string_view> document::raw_json_token() noexcept {
+simdjson_inline simdjson_result<std::string_view> document::raw_json_token() noexcept {
   auto _iter = get_root_value_iterator();
   return std::string_view(reinterpret_cast<const char*>(_iter.peek_start()), _iter.peek_start_length());
 }
 
-simdjson_really_inline simdjson_result<value> document::at_pointer(std::string_view json_pointer) noexcept {
+simdjson_inline simdjson_result<value> document::at_pointer(std::string_view json_pointer) noexcept {
   rewind(); // Rewind the document each time at_pointer is called
   if (json_pointer.empty()) {
     return this->get_value();
@@ -30139,7 +29569,7 @@ simdjson_really_inline simdjson_result<value> document::at_pointer(std::string_v
 
 namespace simdjson {
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::simdjson_result(
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::simdjson_result(
   SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document &&value
 ) noexcept :
     implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>(
@@ -30147,7 +29577,7 @@ simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand
     )
 {
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::simdjson_result(
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::simdjson_result(
   error_code error
 ) noexcept :
     implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>(
@@ -30155,216 +29585,216 @@ simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand
     )
 {
 }
-simdjson_really_inline simdjson_result<size_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::count_elements() & noexcept {
+simdjson_inline simdjson_result<size_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::count_elements() & noexcept {
   if (error()) { return error(); }
   return first.count_elements();
 }
-simdjson_really_inline simdjson_result<size_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::count_fields() & noexcept {
+simdjson_inline simdjson_result<size_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::count_fields() & noexcept {
   if (error()) { return error(); }
   return first.count_fields();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::at(size_t index) & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::at(size_t index) & noexcept {
   if (error()) { return error(); }
   return first.at(index);
 }
-simdjson_really_inline error_code simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::rewind() noexcept {
+simdjson_inline error_code simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::rewind() noexcept {
   if (error()) { return error(); }
   first.rewind();
   return SUCCESS;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::begin() & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::begin() & noexcept {
   if (error()) { return error(); }
   return first.begin();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::end() & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::end() & noexcept {
   return {};
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::find_field_unordered(std::string_view key) & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::find_field_unordered(std::string_view key) & noexcept {
   if (error()) { return error(); }
   return first.find_field_unordered(key);
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::find_field_unordered(const char *key) & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::find_field_unordered(const char *key) & noexcept {
   if (error()) { return error(); }
   return first.find_field_unordered(key);
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator[](std::string_view key) & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator[](std::string_view key) & noexcept {
   if (error()) { return error(); }
   return first[key];
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator[](const char *key) & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator[](const char *key) & noexcept {
   if (error()) { return error(); }
   return first[key];
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::find_field(std::string_view key) & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::find_field(std::string_view key) & noexcept {
   if (error()) { return error(); }
   return first.find_field(key);
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::find_field(const char *key) & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::find_field(const char *key) & noexcept {
   if (error()) { return error(); }
   return first.find_field(key);
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_array() & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_array() & noexcept {
   if (error()) { return error(); }
   return first.get_array();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_object() & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_object() & noexcept {
   if (error()) { return error(); }
   return first.get_object();
 }
-simdjson_really_inline simdjson_result<uint64_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_uint64() noexcept {
+simdjson_inline simdjson_result<uint64_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_uint64() noexcept {
   if (error()) { return error(); }
   return first.get_uint64();
 }
-simdjson_really_inline simdjson_result<int64_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_int64() noexcept {
+simdjson_inline simdjson_result<int64_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_int64() noexcept {
   if (error()) { return error(); }
   return first.get_int64();
 }
-simdjson_really_inline simdjson_result<double> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_double() noexcept {
+simdjson_inline simdjson_result<double> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_double() noexcept {
   if (error()) { return error(); }
   return first.get_double();
 }
-simdjson_really_inline simdjson_result<std::string_view> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_string() noexcept {
+simdjson_inline simdjson_result<std::string_view> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_string() noexcept {
   if (error()) { return error(); }
   return first.get_string();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_raw_json_string() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_raw_json_string() noexcept {
   if (error()) { return error(); }
   return first.get_raw_json_string();
 }
-simdjson_really_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_bool() noexcept {
+simdjson_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_bool() noexcept {
   if (error()) { return error(); }
   return first.get_bool();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_value() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_value() noexcept {
   if (error()) { return error(); }
   return first.get_value();
 }
-simdjson_really_inline bool simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::is_null() noexcept {
+simdjson_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::is_null() noexcept {
   if (error()) { return error(); }
   return first.is_null();
 }
 
 template<typename T>
-simdjson_really_inline simdjson_result<T> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get() & noexcept {
+simdjson_inline simdjson_result<T> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get() & noexcept {
   if (error()) { return error(); }
   return first.get<T>();
 }
 template<typename T>
-simdjson_really_inline simdjson_result<T> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get() && noexcept {
+simdjson_inline simdjson_result<T> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get() && noexcept {
   if (error()) { return error(); }
   return std::forward<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>(first).get<T>();
 }
 template<typename T>
-simdjson_really_inline error_code simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get(T &out) & noexcept {
+simdjson_inline error_code simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get(T &out) & noexcept {
   if (error()) { return error(); }
   return first.get<T>(out);
 }
 template<typename T>
-simdjson_really_inline error_code simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get(T &out) && noexcept {
+simdjson_inline error_code simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get(T &out) && noexcept {
   if (error()) { return error(); }
   return std::forward<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>(first).get<T>(out);
 }
 
-template<> simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>() & noexcept = delete;
-template<> simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>() && noexcept {
+template<> simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>() & noexcept = delete;
+template<> simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>() && noexcept {
   if (error()) { return error(); }
   return std::forward<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>(first);
 }
-template<> simdjson_really_inline error_code simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document &out) & noexcept = delete;
-template<> simdjson_really_inline error_code simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document &out) && noexcept {
+template<> simdjson_inline error_code simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document &out) & noexcept = delete;
+template<> simdjson_inline error_code simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document &out) && noexcept {
   if (error()) { return error(); }
   out = std::forward<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>(first);
   return SUCCESS;
 }
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::type() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::type() noexcept {
   if (error()) { return error(); }
   return first.type();
 }
 
-simdjson_really_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::is_scalar() noexcept {
+simdjson_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::is_scalar() noexcept {
   if (error()) { return error(); }
   return first.is_scalar();
 }
 
 
-simdjson_really_inline bool simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::is_negative() noexcept {
+simdjson_inline bool simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::is_negative() noexcept {
   if (error()) { return error(); }
   return first.is_negative();
 }
 
-simdjson_really_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::is_integer() noexcept {
+simdjson_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::is_integer() noexcept {
   if (error()) { return error(); }
   return first.is_integer();
 }
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number_type> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_number_type() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number_type> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_number_type() noexcept {
   if (error()) { return error(); }
   return first.get_number_type();
 }
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_number() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::get_number() noexcept {
   if (error()) { return error(); }
   return first.get_number();
 }
 
 
 #if SIMDJSON_EXCEPTIONS
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array() & noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array() & noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object() & noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object() & noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator uint64_t() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator uint64_t() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator int64_t() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator int64_t() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator double() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator double() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator std::string_view() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator std::string_view() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator bool() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator bool() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
 #endif
 
 
-simdjson_really_inline simdjson_result<const char *> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::current_location() noexcept {
+simdjson_inline simdjson_result<const char *> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::current_location() noexcept {
   if (error()) { return error(); }
   return first.current_location();
 }
 
-simdjson_really_inline int32_t simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::current_depth() const noexcept {
+simdjson_inline int32_t simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::current_depth() const noexcept {
   if (error()) { return error(); }
   return first.current_depth();
 }
 
-simdjson_really_inline simdjson_result<std::string_view> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::raw_json_token() noexcept {
+simdjson_inline simdjson_result<std::string_view> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::raw_json_token() noexcept {
   if (error()) { return error(); }
   return first.raw_json_token();
 }
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::at_pointer(std::string_view json_pointer) noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document>::at_pointer(std::string_view json_pointer) noexcept {
   if (error()) { return error(); }
   return first.at_pointer(json_pointer);
 }
@@ -30377,54 +29807,54 @@ namespace simdjson {
 namespace SIMDJSON_BUILTIN_IMPLEMENTATION {
 namespace ondemand {
 
-simdjson_really_inline document_reference::document_reference() noexcept : doc{nullptr} {}
-simdjson_really_inline document_reference::document_reference(document &d) noexcept : doc(&d) {}
-simdjson_really_inline void document_reference::rewind() noexcept { doc->rewind(); }
-simdjson_really_inline simdjson_result<array> document_reference::get_array() & noexcept { return doc->get_array(); }
-simdjson_really_inline simdjson_result<object> document_reference::get_object() & noexcept { return doc->get_object(); }
-simdjson_really_inline simdjson_result<uint64_t> document_reference::get_uint64() noexcept { return doc->get_uint64(); }
-simdjson_really_inline simdjson_result<int64_t> document_reference::get_int64() noexcept { return doc->get_int64(); }
-simdjson_really_inline simdjson_result<double> document_reference::get_double() noexcept { return doc->get_double(); }
-simdjson_really_inline simdjson_result<std::string_view> document_reference::get_string() noexcept { return doc->get_string(); }
-simdjson_really_inline simdjson_result<raw_json_string> document_reference::get_raw_json_string() noexcept { return doc->get_raw_json_string(); }
-simdjson_really_inline simdjson_result<bool> document_reference::get_bool() noexcept { return doc->get_bool(); }
-simdjson_really_inline simdjson_result<value> document_reference::get_value() noexcept { return doc->get_value(); }
-simdjson_really_inline bool document_reference::is_null() noexcept { return doc->is_null(); }
+simdjson_inline document_reference::document_reference() noexcept : doc{nullptr} {}
+simdjson_inline document_reference::document_reference(document &d) noexcept : doc(&d) {}
+simdjson_inline void document_reference::rewind() noexcept { doc->rewind(); }
+simdjson_inline simdjson_result<array> document_reference::get_array() & noexcept { return doc->get_array(); }
+simdjson_inline simdjson_result<object> document_reference::get_object() & noexcept { return doc->get_object(); }
+simdjson_inline simdjson_result<uint64_t> document_reference::get_uint64() noexcept { return doc->get_uint64(); }
+simdjson_inline simdjson_result<int64_t> document_reference::get_int64() noexcept { return doc->get_int64(); }
+simdjson_inline simdjson_result<double> document_reference::get_double() noexcept { return doc->get_double(); }
+simdjson_inline simdjson_result<std::string_view> document_reference::get_string() noexcept { return doc->get_string(); }
+simdjson_inline simdjson_result<raw_json_string> document_reference::get_raw_json_string() noexcept { return doc->get_raw_json_string(); }
+simdjson_inline simdjson_result<bool> document_reference::get_bool() noexcept { return doc->get_bool(); }
+simdjson_inline simdjson_result<value> document_reference::get_value() noexcept { return doc->get_value(); }
+simdjson_inline simdjson_result<bool> document_reference::is_null() noexcept { return doc->is_null(); }
 
 #if SIMDJSON_EXCEPTIONS
-simdjson_really_inline document_reference::operator array() & noexcept(false) { return array(*doc); }
-simdjson_really_inline document_reference::operator object() & noexcept(false) { return object(*doc); }
-simdjson_really_inline document_reference::operator uint64_t() noexcept(false) { return uint64_t(*doc); }
-simdjson_really_inline document_reference::operator int64_t() noexcept(false) { return int64_t(*doc); }
-simdjson_really_inline document_reference::operator double() noexcept(false) { return double(*doc); }
-simdjson_really_inline document_reference::operator std::string_view() noexcept(false) { return std::string_view(*doc); }
-simdjson_really_inline document_reference::operator raw_json_string() noexcept(false) { return raw_json_string(*doc); }
-simdjson_really_inline document_reference::operator bool() noexcept(false) { return bool(*doc); }
-simdjson_really_inline document_reference::operator value() noexcept(false) { return value(*doc); }
+simdjson_inline document_reference::operator array() & noexcept(false) { return array(*doc); }
+simdjson_inline document_reference::operator object() & noexcept(false) { return object(*doc); }
+simdjson_inline document_reference::operator uint64_t() noexcept(false) { return uint64_t(*doc); }
+simdjson_inline document_reference::operator int64_t() noexcept(false) { return int64_t(*doc); }
+simdjson_inline document_reference::operator double() noexcept(false) { return double(*doc); }
+simdjson_inline document_reference::operator std::string_view() noexcept(false) { return std::string_view(*doc); }
+simdjson_inline document_reference::operator raw_json_string() noexcept(false) { return raw_json_string(*doc); }
+simdjson_inline document_reference::operator bool() noexcept(false) { return bool(*doc); }
+simdjson_inline document_reference::operator value() noexcept(false) { return value(*doc); }
 #endif
-simdjson_really_inline simdjson_result<size_t> document_reference::count_elements() & noexcept { return doc->count_elements(); }
-simdjson_really_inline simdjson_result<size_t> document_reference::count_fields() & noexcept { return doc->count_fields(); }
-simdjson_really_inline simdjson_result<value> document_reference::at(size_t index) & noexcept { return doc->at(index); }
-simdjson_really_inline simdjson_result<array_iterator> document_reference::begin() & noexcept { return doc->begin(); }
-simdjson_really_inline simdjson_result<array_iterator> document_reference::end() & noexcept { return doc->end(); }
-simdjson_really_inline simdjson_result<value> document_reference::find_field(std::string_view key) & noexcept { return doc->find_field(key); }
-simdjson_really_inline simdjson_result<value> document_reference::find_field(const char *key) & noexcept { return doc->find_field(key); }
-simdjson_really_inline simdjson_result<value> document_reference::operator[](std::string_view key) & noexcept { return (*doc)[key]; }
-simdjson_really_inline simdjson_result<value> document_reference::operator[](const char *key) & noexcept { return (*doc)[key]; }
-simdjson_really_inline simdjson_result<value> document_reference::find_field_unordered(std::string_view key) & noexcept { return doc->find_field_unordered(key); }
-simdjson_really_inline simdjson_result<value> document_reference::find_field_unordered(const char *key) & noexcept { return doc->find_field_unordered(key); }
-simdjson_really_inline simdjson_result<json_type> document_reference::type() noexcept { return doc->type(); }
-simdjson_really_inline simdjson_result<bool> document_reference::is_scalar() noexcept { return doc->is_scalar(); }
-simdjson_really_inline simdjson_result<const char *> document_reference::current_location() noexcept { return doc->current_location(); }
-simdjson_really_inline int32_t document_reference::current_depth() const noexcept { return doc->current_depth(); }
-simdjson_really_inline bool document_reference::is_negative() noexcept { return doc->is_negative(); }
-simdjson_really_inline simdjson_result<bool> document_reference::is_integer() noexcept { return doc->is_integer(); }
-simdjson_really_inline simdjson_result<number_type> document_reference::get_number_type() noexcept { return doc->get_number_type(); }
-simdjson_really_inline simdjson_result<number> document_reference::get_number() noexcept { return doc->get_number(); }
-simdjson_really_inline simdjson_result<std::string_view> document_reference::raw_json_token() noexcept { return doc->raw_json_token(); }
-simdjson_really_inline simdjson_result<value> document_reference::at_pointer(std::string_view json_pointer) noexcept { return doc->at_pointer(json_pointer); }
-simdjson_really_inline simdjson_result<std::string_view> document_reference::raw_json() noexcept { return doc->raw_json();}
-simdjson_really_inline document_reference::operator document&() const noexcept { return *doc; }
+simdjson_inline simdjson_result<size_t> document_reference::count_elements() & noexcept { return doc->count_elements(); }
+simdjson_inline simdjson_result<size_t> document_reference::count_fields() & noexcept { return doc->count_fields(); }
+simdjson_inline simdjson_result<value> document_reference::at(size_t index) & noexcept { return doc->at(index); }
+simdjson_inline simdjson_result<array_iterator> document_reference::begin() & noexcept { return doc->begin(); }
+simdjson_inline simdjson_result<array_iterator> document_reference::end() & noexcept { return doc->end(); }
+simdjson_inline simdjson_result<value> document_reference::find_field(std::string_view key) & noexcept { return doc->find_field(key); }
+simdjson_inline simdjson_result<value> document_reference::find_field(const char *key) & noexcept { return doc->find_field(key); }
+simdjson_inline simdjson_result<value> document_reference::operator[](std::string_view key) & noexcept { return (*doc)[key]; }
+simdjson_inline simdjson_result<value> document_reference::operator[](const char *key) & noexcept { return (*doc)[key]; }
+simdjson_inline simdjson_result<value> document_reference::find_field_unordered(std::string_view key) & noexcept { return doc->find_field_unordered(key); }
+simdjson_inline simdjson_result<value> document_reference::find_field_unordered(const char *key) & noexcept { return doc->find_field_unordered(key); }
+simdjson_inline simdjson_result<json_type> document_reference::type() noexcept { return doc->type(); }
+simdjson_inline simdjson_result<bool> document_reference::is_scalar() noexcept { return doc->is_scalar(); }
+simdjson_inline simdjson_result<const char *> document_reference::current_location() noexcept { return doc->current_location(); }
+simdjson_inline int32_t document_reference::current_depth() const noexcept { return doc->current_depth(); }
+simdjson_inline bool document_reference::is_negative() noexcept { return doc->is_negative(); }
+simdjson_inline simdjson_result<bool> document_reference::is_integer() noexcept { return doc->is_integer(); }
+simdjson_inline simdjson_result<number_type> document_reference::get_number_type() noexcept { return doc->get_number_type(); }
+simdjson_inline simdjson_result<number> document_reference::get_number() noexcept { return doc->get_number(); }
+simdjson_inline simdjson_result<std::string_view> document_reference::raw_json_token() noexcept { return doc->raw_json_token(); }
+simdjson_inline simdjson_result<value> document_reference::at_pointer(std::string_view json_pointer) noexcept { return doc->at_pointer(json_pointer); }
+simdjson_inline simdjson_result<std::string_view> document_reference::raw_json() noexcept { return doc->raw_json();}
+simdjson_inline document_reference::operator document&() const noexcept { return *doc; }
 
 } // namespace ondemand
 } // namespace SIMDJSON_BUILTIN_IMPLEMENTATION
@@ -30433,172 +29863,172 @@ simdjson_really_inline document_reference::operator document&() const noexcept {
 
 
 namespace simdjson {
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference value, error_code error)
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference value, error_code error)
   noexcept : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>(std::forward<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>(value), error) {}
 
 
-simdjson_really_inline simdjson_result<size_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::count_elements() & noexcept {
+simdjson_inline simdjson_result<size_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::count_elements() & noexcept {
   if (error()) { return error(); }
   return first.count_elements();
 }
-simdjson_really_inline simdjson_result<size_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::count_fields() & noexcept {
+simdjson_inline simdjson_result<size_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::count_fields() & noexcept {
   if (error()) { return error(); }
   return first.count_fields();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::at(size_t index) & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::at(size_t index) & noexcept {
   if (error()) { return error(); }
   return first.at(index);
 }
-simdjson_really_inline error_code simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::rewind() noexcept {
+simdjson_inline error_code simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::rewind() noexcept {
   if (error()) { return error(); }
   first.rewind();
   return SUCCESS;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::begin() & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::begin() & noexcept {
   if (error()) { return error(); }
   return first.begin();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::end() & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::end() & noexcept {
   return {};
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::find_field_unordered(std::string_view key) & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::find_field_unordered(std::string_view key) & noexcept {
   if (error()) { return error(); }
   return first.find_field_unordered(key);
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::find_field_unordered(const char *key) & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::find_field_unordered(const char *key) & noexcept {
   if (error()) { return error(); }
   return first.find_field_unordered(key);
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator[](std::string_view key) & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator[](std::string_view key) & noexcept {
   if (error()) { return error(); }
   return first[key];
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator[](const char *key) & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator[](const char *key) & noexcept {
   if (error()) { return error(); }
   return first[key];
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::find_field(std::string_view key) & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::find_field(std::string_view key) & noexcept {
   if (error()) { return error(); }
   return first.find_field(key);
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::find_field(const char *key) & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::find_field(const char *key) & noexcept {
   if (error()) { return error(); }
   return first.find_field(key);
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_array() & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_array() & noexcept {
   if (error()) { return error(); }
   return first.get_array();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_object() & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_object() & noexcept {
   if (error()) { return error(); }
   return first.get_object();
 }
-simdjson_really_inline simdjson_result<uint64_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_uint64() noexcept {
+simdjson_inline simdjson_result<uint64_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_uint64() noexcept {
   if (error()) { return error(); }
   return first.get_uint64();
 }
-simdjson_really_inline simdjson_result<int64_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_int64() noexcept {
+simdjson_inline simdjson_result<int64_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_int64() noexcept {
   if (error()) { return error(); }
   return first.get_int64();
 }
-simdjson_really_inline simdjson_result<double> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_double() noexcept {
+simdjson_inline simdjson_result<double> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_double() noexcept {
   if (error()) { return error(); }
   return first.get_double();
 }
-simdjson_really_inline simdjson_result<std::string_view> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_string() noexcept {
+simdjson_inline simdjson_result<std::string_view> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_string() noexcept {
   if (error()) { return error(); }
   return first.get_string();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_raw_json_string() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_raw_json_string() noexcept {
   if (error()) { return error(); }
   return first.get_raw_json_string();
 }
-simdjson_really_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_bool() noexcept {
+simdjson_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_bool() noexcept {
   if (error()) { return error(); }
   return first.get_bool();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_value() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_value() noexcept {
   if (error()) { return error(); }
   return first.get_value();
 }
-simdjson_really_inline bool simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::is_null() noexcept {
+simdjson_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::is_null() noexcept {
   if (error()) { return error(); }
   return first.is_null();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::type() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::type() noexcept {
   if (error()) { return error(); }
   return first.type();
 }
-simdjson_really_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::is_scalar() noexcept {
+simdjson_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::is_scalar() noexcept {
   if (error()) { return error(); }
   return first.is_scalar();
 }
-simdjson_really_inline bool simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::is_negative() noexcept {
+simdjson_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::is_negative() noexcept {
   if (error()) { return error(); }
   return first.is_negative();
 }
-simdjson_really_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::is_integer() noexcept {
+simdjson_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::is_integer() noexcept {
   if (error()) { return error(); }
   return first.is_integer();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number_type> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_number_type() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number_type> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_number_type() noexcept {
   if (error()) { return error(); }
   return first.get_number_type();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_number() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::get_number() noexcept {
   if (error()) { return error(); }
   return first.get_number();
 }
 #if SIMDJSON_EXCEPTIONS
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array() & noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array() & noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object() & noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object() & noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator uint64_t() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator uint64_t() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator int64_t() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator int64_t() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator double() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator double() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator std::string_view() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator std::string_view() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator bool() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator bool() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
 #endif
 
-simdjson_really_inline simdjson_result<const char *> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::current_location() noexcept {
+simdjson_inline simdjson_result<const char *> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::current_location() noexcept {
   if (error()) { return error(); }
   return first.current_location();
 }
 
-simdjson_really_inline simdjson_result<std::string_view> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::raw_json_token() noexcept {
+simdjson_inline simdjson_result<std::string_view> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::raw_json_token() noexcept {
   if (error()) { return error(); }
   return first.raw_json_token();
 }
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::at_pointer(std::string_view json_pointer) noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_reference>::at_pointer(std::string_view json_pointer) noexcept {
   if (error()) { return error(); }
   return first.at_pointer(json_pointer);
 }
@@ -30611,24 +30041,24 @@ namespace simdjson {
 namespace SIMDJSON_BUILTIN_IMPLEMENTATION {
 namespace ondemand {
 
-simdjson_really_inline value::value(const value_iterator &_iter) noexcept
+simdjson_inline value::value(const value_iterator &_iter) noexcept
   : iter{_iter}
 {
 }
-simdjson_really_inline value value::start(const value_iterator &iter) noexcept {
+simdjson_inline value value::start(const value_iterator &iter) noexcept {
   return iter;
 }
-simdjson_really_inline value value::resume(const value_iterator &iter) noexcept {
+simdjson_inline value value::resume(const value_iterator &iter) noexcept {
   return iter;
 }
 
-simdjson_really_inline simdjson_result<array> value::get_array() noexcept {
+simdjson_inline simdjson_result<array> value::get_array() noexcept {
   return array::start(iter);
 }
-simdjson_really_inline simdjson_result<object> value::get_object() noexcept {
+simdjson_inline simdjson_result<object> value::get_object() noexcept {
   return object::start(iter);
 }
-simdjson_really_inline simdjson_result<object> value::start_or_resume_object() noexcept {
+simdjson_inline simdjson_result<object> value::start_or_resume_object() noexcept {
   if (iter.at_start()) {
     return get_object();
   } else {
@@ -30636,85 +30066,84 @@ simdjson_really_inline simdjson_result<object> value::start_or_resume_object() n
   }
 }
 
-simdjson_really_inline simdjson_result<raw_json_string> value::get_raw_json_string() noexcept {
+simdjson_inline simdjson_result<raw_json_string> value::get_raw_json_string() noexcept {
   return iter.get_raw_json_string();
 }
-simdjson_really_inline simdjson_result<std::string_view> value::get_string() noexcept {
+simdjson_inline simdjson_result<std::string_view> value::get_string() noexcept {
   return iter.get_string();
 }
-simdjson_really_inline simdjson_result<double> value::get_double() noexcept {
+simdjson_inline simdjson_result<double> value::get_double() noexcept {
   return iter.get_double();
 }
-simdjson_really_inline simdjson_result<double> value::get_double_in_string() noexcept {
+simdjson_inline simdjson_result<double> value::get_double_in_string() noexcept {
   return iter.get_double_in_string();
 }
-simdjson_really_inline simdjson_result<uint64_t> value::get_uint64() noexcept {
+simdjson_inline simdjson_result<uint64_t> value::get_uint64() noexcept {
   return iter.get_uint64();
 }
-simdjson_really_inline simdjson_result<uint64_t> value::get_uint64_in_string() noexcept {
+simdjson_inline simdjson_result<uint64_t> value::get_uint64_in_string() noexcept {
   return iter.get_uint64_in_string();
 }
-simdjson_really_inline simdjson_result<int64_t> value::get_int64() noexcept {
+simdjson_inline simdjson_result<int64_t> value::get_int64() noexcept {
   return iter.get_int64();
 }
-simdjson_really_inline simdjson_result<int64_t> value::get_int64_in_string() noexcept {
+simdjson_inline simdjson_result<int64_t> value::get_int64_in_string() noexcept {
   return iter.get_int64_in_string();
 }
-simdjson_really_inline simdjson_result<bool> value::get_bool() noexcept {
+simdjson_inline simdjson_result<bool> value::get_bool() noexcept {
   return iter.get_bool();
 }
-simdjson_really_inline bool value::is_null() noexcept {
+simdjson_inline simdjson_result<bool> value::is_null() noexcept {
   return iter.is_null();
 }
-
-template<> simdjson_really_inline simdjson_result<array> value::get() noexcept { return get_array(); }
-template<> simdjson_really_inline simdjson_result<object> value::get() noexcept { return get_object(); }
-template<> simdjson_really_inline simdjson_result<raw_json_string> value::get() noexcept { return get_raw_json_string(); }
-template<> simdjson_really_inline simdjson_result<std::string_view> value::get() noexcept { return get_string(); }
-template<> simdjson_really_inline simdjson_result<number> value::get() noexcept { return get_number(); }
-template<> simdjson_really_inline simdjson_result<double> value::get() noexcept { return get_double(); }
-template<> simdjson_really_inline simdjson_result<uint64_t> value::get() noexcept { return get_uint64(); }
-template<> simdjson_really_inline simdjson_result<int64_t> value::get() noexcept { return get_int64(); }
-template<> simdjson_really_inline simdjson_result<bool> value::get() noexcept { return get_bool(); }
-
-template<typename T> simdjson_really_inline error_code value::get(T &out) noexcept {
+template<> simdjson_inline simdjson_result<array> value::get() noexcept { return get_array(); }
+template<> simdjson_inline simdjson_result<object> value::get() noexcept { return get_object(); }
+template<> simdjson_inline simdjson_result<raw_json_string> value::get() noexcept { return get_raw_json_string(); }
+template<> simdjson_inline simdjson_result<std::string_view> value::get() noexcept { return get_string(); }
+template<> simdjson_inline simdjson_result<number> value::get() noexcept { return get_number(); }
+template<> simdjson_inline simdjson_result<double> value::get() noexcept { return get_double(); }
+template<> simdjson_inline simdjson_result<uint64_t> value::get() noexcept { return get_uint64(); }
+template<> simdjson_inline simdjson_result<int64_t> value::get() noexcept { return get_int64(); }
+template<> simdjson_inline simdjson_result<bool> value::get() noexcept { return get_bool(); }
+
+template<typename T> simdjson_inline error_code value::get(T &out) noexcept {
   return get<T>().get(out);
 }
 
 #if SIMDJSON_EXCEPTIONS
-simdjson_really_inline value::operator array() noexcept(false) {
+simdjson_inline value::operator array() noexcept(false) {
   return get_array();
 }
-simdjson_really_inline value::operator object() noexcept(false) {
+simdjson_inline value::operator object() noexcept(false) {
   return get_object();
 }
-simdjson_really_inline value::operator uint64_t() noexcept(false) {
+simdjson_inline value::operator uint64_t() noexcept(false) {
   return get_uint64();
 }
-simdjson_really_inline value::operator int64_t() noexcept(false) {
+simdjson_inline value::operator int64_t() noexcept(false) {
   return get_int64();
 }
-simdjson_really_inline value::operator double() noexcept(false) {
+simdjson_inline value::operator double() noexcept(false) {
   return get_double();
 }
-simdjson_really_inline value::operator std::string_view() noexcept(false) {
+simdjson_inline value::operator std::string_view() noexcept(false) {
   return get_string();
 }
-simdjson_really_inline value::operator raw_json_string() noexcept(false) {
+simdjson_inline value::operator raw_json_string() noexcept(false) {
   return get_raw_json_string();
 }
-simdjson_really_inline value::operator bool() noexcept(false) {
+simdjson_inline value::operator bool() noexcept(false) {
   return get_bool();
 }
 #endif
 
-simdjson_really_inline simdjson_result<array_iterator> value::begin() & noexcept {
+simdjson_inline simdjson_result<array_iterator> value::begin() & noexcept {
   return get_array().begin();
 }
-simdjson_really_inline simdjson_result<array_iterator> value::end() & noexcept {
+simdjson_inline simdjson_result<array_iterator> value::end() & noexcept {
   return {};
 }
-simdjson_really_inline simdjson_result<size_t> value::count_elements() & noexcept {
+simdjson_inline simdjson_result<size_t> value::count_elements() & noexcept {
   simdjson_result<size_t> answer;
   auto a = get_array();
   answer = a.count_elements();
@@ -30724,77 +30153,77 @@ simdjson_really_inline simdjson_result<size_t> value::count_elements() & noexcep
   iter.move_at_start();
   return answer;
 }
-simdjson_really_inline simdjson_result<size_t> value::count_fields() & noexcept {
+simdjson_inline simdjson_result<size_t> value::count_fields() & noexcept {
   simdjson_result<size_t> answer;
   auto a = get_object();
   answer = a.count_fields();
   iter.move_at_start();
   return answer;
 }
-simdjson_really_inline simdjson_result<value> value::at(size_t index) noexcept {
+simdjson_inline simdjson_result<value> value::at(size_t index) noexcept {
   auto a = get_array();
   return a.at(index);
 }
 
-simdjson_really_inline simdjson_result<value> value::find_field(std::string_view key) noexcept {
+simdjson_inline simdjson_result<value> value::find_field(std::string_view key) noexcept {
   return start_or_resume_object().find_field(key);
 }
-simdjson_really_inline simdjson_result<value> value::find_field(const char *key) noexcept {
+simdjson_inline simdjson_result<value> value::find_field(const char *key) noexcept {
   return start_or_resume_object().find_field(key);
 }
 
-simdjson_really_inline simdjson_result<value> value::find_field_unordered(std::string_view key) noexcept {
+simdjson_inline simdjson_result<value> value::find_field_unordered(std::string_view key) noexcept {
   return start_or_resume_object().find_field_unordered(key);
 }
-simdjson_really_inline simdjson_result<value> value::find_field_unordered(const char *key) noexcept {
+simdjson_inline simdjson_result<value> value::find_field_unordered(const char *key) noexcept {
   return start_or_resume_object().find_field_unordered(key);
 }
 
-simdjson_really_inline simdjson_result<value> value::operator[](std::string_view key) noexcept {
+simdjson_inline simdjson_result<value> value::operator[](std::string_view key) noexcept {
   return start_or_resume_object()[key];
 }
-simdjson_really_inline simdjson_result<value> value::operator[](const char *key) noexcept {
+simdjson_inline simdjson_result<value> value::operator[](const char *key) noexcept {
   return start_or_resume_object()[key];
 }
 
-simdjson_really_inline simdjson_result<json_type> value::type() noexcept {
+simdjson_inline simdjson_result<json_type> value::type() noexcept {
   return iter.type();
 }
 
-simdjson_really_inline simdjson_result<bool> value::is_scalar() noexcept {
+simdjson_inline simdjson_result<bool> value::is_scalar() noexcept {
   json_type this_type;
   auto error = type().get(this_type);
   if(error) { return error; }
   return ! ((this_type == json_type::array) || (this_type == json_type::object));
 }
 
-simdjson_really_inline bool value::is_negative() noexcept {
+simdjson_inline bool value::is_negative() noexcept {
   return iter.is_negative();
 }
 
-simdjson_really_inline simdjson_result<bool> value::is_integer() noexcept {
+simdjson_inline simdjson_result<bool> value::is_integer() noexcept {
   return iter.is_integer();
 }
-simdjson_warn_unused simdjson_really_inline simdjson_result<number_type> value::get_number_type() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<number_type> value::get_number_type() noexcept {
   return iter.get_number_type();
 }
-simdjson_warn_unused simdjson_really_inline simdjson_result<number> value::get_number() noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<number> value::get_number() noexcept {
   return iter.get_number();
 }
 
-simdjson_really_inline std::string_view value::raw_json_token() noexcept {
+simdjson_inline std::string_view value::raw_json_token() noexcept {
   return std::string_view(reinterpret_cast<const char*>(iter.peek_start()), iter.peek_start_length());
 }
 
-simdjson_really_inline simdjson_result<const char *> value::current_location() noexcept {
+simdjson_inline simdjson_result<const char *> value::current_location() noexcept {
   return iter.json_iter().current_location();
 }
 
-simdjson_really_inline int32_t value::current_depth() const noexcept{
+simdjson_inline int32_t value::current_depth() const noexcept{
   return iter.json_iter().depth();
 }
 
-simdjson_really_inline simdjson_result<value> value::at_pointer(std::string_view json_pointer) noexcept {
+simdjson_inline simdjson_result<value> value::at_pointer(std::string_view json_pointer) noexcept {
   json_type t;
   SIMDJSON_TRY(type().get(t));
   switch (t)
@@ -30814,7 +30243,7 @@ simdjson_really_inline simdjson_result<value> value::at_pointer(std::string_view
 
 namespace simdjson {
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::simdjson_result(
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::simdjson_result(
   SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value &&value
 ) noexcept :
     implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>(
@@ -30822,203 +30251,203 @@ simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand
     )
 {
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::simdjson_result(
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::simdjson_result(
   error_code error
 ) noexcept :
     implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>(error)
 {
 }
-simdjson_really_inline simdjson_result<size_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::count_elements() & noexcept {
+simdjson_inline simdjson_result<size_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::count_elements() & noexcept {
   if (error()) { return error(); }
   return first.count_elements();
 }
-simdjson_really_inline simdjson_result<size_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::count_fields() & noexcept {
+simdjson_inline simdjson_result<size_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::count_fields() & noexcept {
   if (error()) { return error(); }
   return first.count_fields();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::at(size_t index) noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::at(size_t index) noexcept {
   if (error()) { return error(); }
   return first.at(index);
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::begin() & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::begin() & noexcept {
   if (error()) { return error(); }
   return first.begin();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::end() & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::end() & noexcept {
   if (error()) { return error(); }
   return {};
 }
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::find_field(std::string_view key) noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::find_field(std::string_view key) noexcept {
   if (error()) { return error(); }
   return first.find_field(key);
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::find_field(const char *key) noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::find_field(const char *key) noexcept {
   if (error()) { return error(); }
   return first.find_field(key);
 }
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::find_field_unordered(std::string_view key) noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::find_field_unordered(std::string_view key) noexcept {
   if (error()) { return error(); }
   return first.find_field_unordered(key);
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::find_field_unordered(const char *key) noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::find_field_unordered(const char *key) noexcept {
   if (error()) { return error(); }
   return first.find_field_unordered(key);
 }
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator[](std::string_view key) noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator[](std::string_view key) noexcept {
   if (error()) { return error(); }
   return first[key];
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator[](const char *key) noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator[](const char *key) noexcept {
   if (error()) { return error(); }
   return first[key];
 }
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_array() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_array() noexcept {
   if (error()) { return error(); }
   return first.get_array();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_object() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_object() noexcept {
   if (error()) { return error(); }
   return first.get_object();
 }
-simdjson_really_inline simdjson_result<uint64_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_uint64() noexcept {
+simdjson_inline simdjson_result<uint64_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_uint64() noexcept {
   if (error()) { return error(); }
   return first.get_uint64();
 }
-simdjson_really_inline simdjson_result<uint64_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_uint64_in_string() noexcept {
+simdjson_inline simdjson_result<uint64_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_uint64_in_string() noexcept {
   if (error()) { return error(); }
   return first.get_uint64_in_string();
 }
-simdjson_really_inline simdjson_result<int64_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_int64() noexcept {
+simdjson_inline simdjson_result<int64_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_int64() noexcept {
   if (error()) { return error(); }
   return first.get_int64();
 }
-simdjson_really_inline simdjson_result<int64_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_int64_in_string() noexcept {
+simdjson_inline simdjson_result<int64_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_int64_in_string() noexcept {
   if (error()) { return error(); }
   return first.get_int64_in_string();
 }
-simdjson_really_inline simdjson_result<double> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_double() noexcept {
+simdjson_inline simdjson_result<double> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_double() noexcept {
   if (error()) { return error(); }
   return first.get_double();
 }
-simdjson_really_inline simdjson_result<double> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_double_in_string() noexcept {
+simdjson_inline simdjson_result<double> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_double_in_string() noexcept {
   if (error()) { return error(); }
   return first.get_double_in_string();
 }
-simdjson_really_inline simdjson_result<std::string_view> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_string() noexcept {
+simdjson_inline simdjson_result<std::string_view> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_string() noexcept {
   if (error()) { return error(); }
   return first.get_string();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_raw_json_string() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_raw_json_string() noexcept {
   if (error()) { return error(); }
   return first.get_raw_json_string();
 }
-simdjson_really_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_bool() noexcept {
+simdjson_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_bool() noexcept {
   if (error()) { return error(); }
   return first.get_bool();
 }
-simdjson_really_inline bool simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::is_null() noexcept {
-  if (error()) { return false; }
+simdjson_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::is_null() noexcept {
+  if (error()) { return error(); }
   return first.is_null();
 }
 
-template<typename T> simdjson_really_inline simdjson_result<T> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get() noexcept {
+template<typename T> simdjson_inline simdjson_result<T> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get() noexcept {
   if (error()) { return error(); }
   return first.get<T>();
 }
-template<typename T> simdjson_really_inline error_code simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get(T &out) noexcept {
+template<typename T> simdjson_inline error_code simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get(T &out) noexcept {
   if (error()) { return error(); }
   return first.get<T>(out);
 }
 
-template<> simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>() noexcept  {
+template<> simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>() noexcept  {
   if (error()) { return error(); }
   return std::move(first);
 }
-template<> simdjson_really_inline error_code simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value &out) noexcept {
+template<> simdjson_inline error_code simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value &out) noexcept {
   if (error()) { return error(); }
   out = first;
   return SUCCESS;
 }
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::type() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::json_type> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::type() noexcept {
   if (error()) { return error(); }
   return first.type();
 }
-simdjson_really_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::is_scalar() noexcept {
+simdjson_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::is_scalar() noexcept {
   if (error()) { return error(); }
   return first.is_scalar();
 }
-simdjson_really_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::is_negative() noexcept {
+simdjson_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::is_negative() noexcept {
   if (error()) { return error(); }
   return first.is_negative();
 }
-simdjson_really_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::is_integer() noexcept {
+simdjson_inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::is_integer() noexcept {
   if (error()) { return error(); }
   return first.is_integer();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number_type> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_number_type() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number_type> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_number_type() noexcept {
   if (error()) { return error(); }
   return first.get_number_type();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_number() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::number> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::get_number() noexcept {
   if (error()) { return error(); }
   return first.get_number();
 }
 #if SIMDJSON_EXCEPTIONS
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::array() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator uint64_t() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator uint64_t() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator int64_t() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator int64_t() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator double() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator double() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator std::string_view() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator std::string_view() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator bool() noexcept(false) {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::operator bool() noexcept(false) {
   if (error()) { throw simdjson_error(error()); }
   return first;
 }
 #endif
 
-simdjson_really_inline simdjson_result<std::string_view> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::raw_json_token() noexcept {
+simdjson_inline simdjson_result<std::string_view> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::raw_json_token() noexcept {
   if (error()) { return error(); }
   return first.raw_json_token();
 }
 
-simdjson_really_inline simdjson_result<const char *> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::current_location() noexcept {
+simdjson_inline simdjson_result<const char *> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::current_location() noexcept {
   if (error()) { return error(); }
   return first.current_location();
 }
 
-simdjson_really_inline int32_t simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::current_depth() const noexcept {
+simdjson_inline simdjson_result<int32_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::current_depth() const noexcept {
   if (error()) { return error(); }
   return first.current_depth();
 }
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::at_pointer(std::string_view json_pointer) noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value>::at_pointer(std::string_view json_pointer) noexcept {
   if (error()) { return error(); }
   return first.at_pointer(json_pointer);
 }
@@ -31031,41 +30460,41 @@ namespace SIMDJSON_BUILTIN_IMPLEMENTATION {
 namespace ondemand {
 
 // clang 6 doesn't think the default constructor can be noexcept, so we make it explicit
-simdjson_really_inline field::field() noexcept : std::pair<raw_json_string, ondemand::value>() {}
+simdjson_inline field::field() noexcept : std::pair<raw_json_string, ondemand::value>() {}
 
-simdjson_really_inline field::field(raw_json_string key, ondemand::value &&value) noexcept
+simdjson_inline field::field(raw_json_string key, ondemand::value &&value) noexcept
   : std::pair<raw_json_string, ondemand::value>(key, std::forward<ondemand::value>(value))
 {
 }
 
-simdjson_really_inline simdjson_result<field> field::start(value_iterator &parent_iter) noexcept {
+simdjson_inline simdjson_result<field> field::start(value_iterator &parent_iter) noexcept {
   raw_json_string key;
   SIMDJSON_TRY( parent_iter.field_key().get(key) );
   SIMDJSON_TRY( parent_iter.field_value() );
   return field::start(parent_iter, key);
 }
 
-simdjson_really_inline simdjson_result<field> field::start(const value_iterator &parent_iter, raw_json_string key) noexcept {
+simdjson_inline simdjson_result<field> field::start(const value_iterator &parent_iter, raw_json_string key) noexcept {
     return field(key, parent_iter.child());
 }
 
-simdjson_really_inline simdjson_warn_unused simdjson_result<std::string_view> field::unescaped_key() noexcept {
+simdjson_inline simdjson_warn_unused simdjson_result<std::string_view> field::unescaped_key() noexcept {
   SIMDJSON_ASSUME(first.buf != nullptr); // We would like to call .alive() but Visual Studio won't let us.
-  simdjson_result<std::string_view> answer = first.unescape(second.iter.string_buf_loc());
+  simdjson_result<std::string_view> answer = first.unescape(second.iter.json_iter());
   first.consume();
   return answer;
 }
 
-simdjson_really_inline raw_json_string field::key() const noexcept {
+simdjson_inline raw_json_string field::key() const noexcept {
   SIMDJSON_ASSUME(first.buf != nullptr); // We would like to call .alive() by Visual Studio won't let us.
   return first;
 }
 
-simdjson_really_inline value &field::value() & noexcept {
+simdjson_inline value &field::value() & noexcept {
   return second;
 }
 
-simdjson_really_inline value field::value() && noexcept {
+simdjson_inline value field::value() && noexcept {
   return std::forward<field>(*this).second;
 }
 
@@ -31075,7 +30504,7 @@ simdjson_really_inline value field::value() && noexcept {
 
 namespace simdjson {
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field>::simdjson_result(
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field>::simdjson_result(
   SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field &&value
 ) noexcept :
     implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field>(
@@ -31083,22 +30512,22 @@ simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand
     )
 {
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field>::simdjson_result(
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field>::simdjson_result(
   error_code error
 ) noexcept :
     implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field>(error)
 {
 }
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field>::key() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::raw_json_string> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field>::key() noexcept {
   if (error()) { return error(); }
   return first.key();
 }
-simdjson_really_inline simdjson_result<std::string_view> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field>::unescaped_key() noexcept {
+simdjson_inline simdjson_result<std::string_view> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field>::unescaped_key() noexcept {
   if (error()) { return error(); }
   return first.unescaped_key();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field>::value() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::field>::value() noexcept {
   if (error()) { return error(); }
   return std::move(first.value());
 }
@@ -31110,46 +30539,46 @@ namespace simdjson {
 namespace SIMDJSON_BUILTIN_IMPLEMENTATION {
 namespace ondemand {
 
-simdjson_really_inline simdjson_result<value> object::find_field_unordered(const std::string_view key) & noexcept {
+simdjson_inline simdjson_result<value> object::find_field_unordered(const std::string_view key) & noexcept {
   bool has_value;
   SIMDJSON_TRY( iter.find_field_unordered_raw(key).get(has_value) );
   if (!has_value) { return NO_SUCH_FIELD; }
   return value(iter.child());
 }
-simdjson_really_inline simdjson_result<value> object::find_field_unordered(const std::string_view key) && noexcept {
+simdjson_inline simdjson_result<value> object::find_field_unordered(const std::string_view key) && noexcept {
   bool has_value;
   SIMDJSON_TRY( iter.find_field_unordered_raw(key).get(has_value) );
   if (!has_value) { return NO_SUCH_FIELD; }
   return value(iter.child());
 }
-simdjson_really_inline simdjson_result<value> object::operator[](const std::string_view key) & noexcept {
+simdjson_inline simdjson_result<value> object::operator[](const std::string_view key) & noexcept {
   return find_field_unordered(key);
 }
-simdjson_really_inline simdjson_result<value> object::operator[](const std::string_view key) && noexcept {
+simdjson_inline simdjson_result<value> object::operator[](const std::string_view key) && noexcept {
   return std::forward<object>(*this).find_field_unordered(key);
 }
-simdjson_really_inline simdjson_result<value> object::find_field(const std::string_view key) & noexcept {
+simdjson_inline simdjson_result<value> object::find_field(const std::string_view key) & noexcept {
   bool has_value;
   SIMDJSON_TRY( iter.find_field_raw(key).get(has_value) );
   if (!has_value) { return NO_SUCH_FIELD; }
   return value(iter.child());
 }
-simdjson_really_inline simdjson_result<value> object::find_field(const std::string_view key) && noexcept {
+simdjson_inline simdjson_result<value> object::find_field(const std::string_view key) && noexcept {
   bool has_value;
   SIMDJSON_TRY( iter.find_field_raw(key).get(has_value) );
   if (!has_value) { return NO_SUCH_FIELD; }
   return value(iter.child());
 }
 
-simdjson_really_inline simdjson_result<object> object::start(value_iterator &iter) noexcept {
+simdjson_inline simdjson_result<object> object::start(value_iterator &iter) noexcept {
   SIMDJSON_TRY( iter.start_object().error() );
   return object(iter);
 }
-simdjson_really_inline simdjson_result<object> object::start_root(value_iterator &iter) noexcept {
+simdjson_inline simdjson_result<object> object::start_root(value_iterator &iter) noexcept {
   SIMDJSON_TRY( iter.start_root_object().error() );
   return object(iter);
 }
-simdjson_really_inline error_code object::consume() noexcept {
+simdjson_inline error_code object::consume() noexcept {
   if(iter.is_at_key()) {
     /**
      * whenever you are pointing at a key, calling skip_child() is
@@ -31175,7 +30604,7 @@ simdjson_really_inline error_code object::consume() noexcept {
   return error_skip;
 }
 
-simdjson_really_inline simdjson_result<std::string_view> object::raw_json() noexcept {
+simdjson_inline simdjson_result<std::string_view> object::raw_json() noexcept {
   const uint8_t * starting_point{iter.peek_start()};
   auto error = consume();
   if(error) { return error; }
@@ -31183,27 +30612,27 @@ simdjson_really_inline simdjson_result<std::string_view> object::raw_json() noex
   return std::string_view(reinterpret_cast<const char*>(starting_point), size_t(final_point - starting_point));
 }
 
-simdjson_really_inline simdjson_result<object> object::started(value_iterator &iter) noexcept {
+simdjson_inline simdjson_result<object> object::started(value_iterator &iter) noexcept {
   SIMDJSON_TRY( iter.started_object().error() );
   return object(iter);
 }
 
-simdjson_really_inline object object::resume(const value_iterator &iter) noexcept {
+simdjson_inline object object::resume(const value_iterator &iter) noexcept {
   return iter;
 }
 
-simdjson_really_inline object::object(const value_iterator &_iter) noexcept
+simdjson_inline object::object(const value_iterator &_iter) noexcept
   : iter{_iter}
 {
 }
 
-simdjson_really_inline simdjson_result<object_iterator> object::begin() noexcept {
-#ifdef SIMDJSON_DEVELOPMENT_CHECKS
+simdjson_inline simdjson_result<object_iterator> object::begin() noexcept {
+#if SIMDJSON_DEVELOPMENT_CHECKS
   if (!iter.is_at_iterator_start()) { return OUT_OF_ORDER_ITERATION; }
 #endif
   return object_iterator(iter);
 }
-simdjson_really_inline simdjson_result<object_iterator> object::end() noexcept {
+simdjson_inline simdjson_result<object_iterator> object::end() noexcept {
   return object_iterator(iter);
 }
 
@@ -31247,7 +30676,7 @@ inline simdjson_result<value> object::at_pointer(std::string_view json_pointer)
   return child;
 }
 
-simdjson_really_inline simdjson_result<size_t> object::count_fields() & noexcept {
+simdjson_inline simdjson_result<size_t> object::count_fields() & noexcept {
   size_t count{0};
   // Important: we do not consume any of the values.
   for(simdjson_unused auto v : *this) { count++; }
@@ -31259,14 +30688,14 @@ simdjson_really_inline simdjson_result<size_t> object::count_fields() & noexcept
   return count;
 }
 
-simdjson_really_inline simdjson_result<bool> object::is_empty() & noexcept {
+simdjson_inline simdjson_result<bool> object::is_empty() & noexcept {
   bool is_not_empty;
   auto error = iter.reset_object().get(is_not_empty);
   if(error) { return error; }
   return !is_not_empty;
 }
 
-simdjson_really_inline simdjson_result<bool> object::reset() & noexcept {
+simdjson_inline simdjson_result<bool> object::reset() & noexcept {
   return iter.reset_object();
 }
 
@@ -31276,45 +30705,45 @@ simdjson_really_inline simdjson_result<bool> object::reset() & noexcept {
 
 namespace simdjson {
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object &&value) noexcept
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object &&value) noexcept
     : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>(std::forward<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>(value)) {}
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::simdjson_result(error_code error) noexcept
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::simdjson_result(error_code error) noexcept
     : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>(error) {}
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::begin() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::begin() noexcept {
   if (error()) { return error(); }
   return first.begin();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::end() noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object_iterator> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::end() noexcept {
   if (error()) { return error(); }
   return first.end();
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::find_field_unordered(std::string_view key) & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::find_field_unordered(std::string_view key) & noexcept {
   if (error()) { return error(); }
   return first.find_field_unordered(key);
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::find_field_unordered(std::string_view key) && noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::find_field_unordered(std::string_view key) && noexcept {
   if (error()) { return error(); }
   return std::forward<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>(first).find_field_unordered(key);
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::operator[](std::string_view key) & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::operator[](std::string_view key) & noexcept {
   if (error()) { return error(); }
   return first[key];
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::operator[](std::string_view key) && noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::operator[](std::string_view key) && noexcept {
   if (error()) { return error(); }
   return std::forward<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>(first)[key];
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::find_field(std::string_view key) & noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::find_field(std::string_view key) & noexcept {
   if (error()) { return error(); }
   return first.find_field(key);
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::find_field(std::string_view key) && noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::find_field(std::string_view key) && noexcept {
   if (error()) { return error(); }
   return std::forward<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>(first).find_field(key);
 }
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::at_pointer(std::string_view json_pointer) noexcept {
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::value> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::at_pointer(std::string_view json_pointer) noexcept {
   if (error()) { return error(); }
   return first.at_pointer(json_pointer);
 }
@@ -31329,7 +30758,7 @@ inline simdjson_result<bool> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::on
   return first.is_empty();
 }
 
-simdjson_really_inline  simdjson_result<size_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::count_fields() & noexcept {
+simdjson_inline  simdjson_result<size_t> simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::object>::count_fields() & noexcept {
   if (error()) { return error(); }
   return first.count_fields();
 }
@@ -31341,11 +30770,11 @@ namespace simdjson {
 namespace SIMDJSON_BUILTIN_IMPLEMENTATION {
 namespace ondemand {
 
-simdjson_really_inline parser::parser(size_t max_capacity) noexcept
+simdjson_inline parser::parser(size_t max_capacity) noexcept
   : _max_capacity{max_capacity} {
 }
 
-simdjson_warn_unused simdjson_really_inline error_code parser::allocate(size_t new_capacity, size_t new_max_depth) noexcept {
+simdjson_warn_unused simdjson_inline error_code parser::allocate(size_t new_capacity, size_t new_max_depth) noexcept {
   if (new_capacity > max_capacity()) { return CAPACITY; }
   if (string_buf && new_capacity == capacity() && new_max_depth == max_depth()) { return SUCCESS; }
 
@@ -31353,7 +30782,7 @@ simdjson_warn_unused simdjson_really_inline error_code parser::allocate(size_t n
   _capacity = 0;
   size_t string_capacity = SIMDJSON_ROUNDUP_N(5 * new_capacity / 3 + SIMDJSON_PADDING, 64);
   string_buf.reset(new (std::nothrow) uint8_t[string_capacity]);
-#ifdef SIMDJSON_DEVELOPMENT_CHECKS
+#if SIMDJSON_DEVELOPMENT_CHECKS
   start_positions.reset(new (std::nothrow) token_position[new_max_depth]);
 #endif
   if (implementation) {
@@ -31367,7 +30796,7 @@ simdjson_warn_unused simdjson_really_inline error_code parser::allocate(size_t n
   return SUCCESS;
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<document> parser::iterate(padded_string_view json) & noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<document> parser::iterate(padded_string_view json) & noexcept {
   if (json.padding() < SIMDJSON_PADDING) { return INSUFFICIENT_PADDING; }
 
   // Allocate if needed
@@ -31380,37 +30809,37 @@ simdjson_warn_unused simdjson_really_inline simdjson_result<document> parser::it
   return document::start({ reinterpret_cast<const uint8_t *>(json.data()), this });
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<document> parser::iterate(const char *json, size_t len, size_t allocated) & noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<document> parser::iterate(const char *json, size_t len, size_t allocated) & noexcept {
   return iterate(padded_string_view(json, len, allocated));
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<document> parser::iterate(const uint8_t *json, size_t len, size_t allocated) & noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<document> parser::iterate(const uint8_t *json, size_t len, size_t allocated) & noexcept {
   return iterate(padded_string_view(json, len, allocated));
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<document> parser::iterate(std::string_view json, size_t allocated) & noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<document> parser::iterate(std::string_view json, size_t allocated) & noexcept {
   return iterate(padded_string_view(json, allocated));
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<document> parser::iterate(const std::string &json) & noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<document> parser::iterate(const std::string &json) & noexcept {
   return iterate(padded_string_view(json));
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<document> parser::iterate(const simdjson_result<padded_string_view> &result) & noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<document> parser::iterate(const simdjson_result<padded_string_view> &result) & noexcept {
   // We don't presently have a way to temporarily get a const T& from a simdjson_result<T> without throwing an exception
   SIMDJSON_TRY( result.error() );
   padded_string_view json = result.value_unsafe();
   return iterate(json);
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<document> parser::iterate(const simdjson_result<padded_string> &result) & noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<document> parser::iterate(const simdjson_result<padded_string> &result) & noexcept {
   // We don't presently have a way to temporarily get a const T& from a simdjson_result<T> without throwing an exception
   SIMDJSON_TRY( result.error() );
   const padded_string &json = result.value_unsafe();
   return iterate(json);
 }
 
-simdjson_warn_unused simdjson_really_inline simdjson_result<json_iterator> parser::iterate_raw(padded_string_view json) & noexcept {
+simdjson_warn_unused simdjson_inline simdjson_result<json_iterator> parser::iterate_raw(padded_string_view json) & noexcept {
   if (json.padding() < SIMDJSON_PADDING) { return INSUFFICIENT_PADDING; }
 
   // Allocate if needed
@@ -31437,34 +30866,41 @@ inline simdjson_result<document_stream> parser::iterate_many(const padded_string
   return iterate_many(s.data(), s.length(), batch_size);
 }
 
-simdjson_really_inline size_t parser::capacity() const noexcept {
+simdjson_inline size_t parser::capacity() const noexcept {
   return _capacity;
 }
-simdjson_really_inline size_t parser::max_capacity() const noexcept {
+simdjson_inline size_t parser::max_capacity() const noexcept {
   return _max_capacity;
 }
-simdjson_really_inline size_t parser::max_depth() const noexcept {
+simdjson_inline size_t parser::max_depth() const noexcept {
   return _max_depth;
 }
 
-simdjson_really_inline void parser::set_max_capacity(size_t max_capacity) noexcept {
-  size_t MINIMAL_DOCUMENT_CAPACITY = 32;
-  if(max_capacity < MINIMAL_DOCUMENT_CAPACITY) {
+simdjson_inline void parser::set_max_capacity(size_t max_capacity) noexcept {
+  if(max_capacity < dom::MINIMAL_DOCUMENT_CAPACITY) {
     _max_capacity = max_capacity;
   } else {
-    _max_capacity = MINIMAL_DOCUMENT_CAPACITY;
+    _max_capacity = dom::MINIMAL_DOCUMENT_CAPACITY;
   }
 }
 
+simdjson_inline simdjson_warn_unused simdjson_result<std::string_view> parser::unescape(raw_json_string in, uint8_t *&dst) const noexcept {
+  uint8_t *end = implementation->parse_string(in.buf, dst);
+  if (!end) { return STRING_ERROR; }
+  std::string_view result(reinterpret_cast<const char *>(dst), end-dst);
+  dst = end;
+  return result;
+}
+
 } // namespace ondemand
 } // namespace SIMDJSON_BUILTIN_IMPLEMENTATION
 } // namespace simdjson
 
 namespace simdjson {
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::parser>::simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::parser &&value) noexcept
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::parser>::simdjson_result(SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::parser &&value) noexcept
     : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::parser>(std::forward<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::parser>(value)) {}
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::parser>::simdjson_result(error_code error) noexcept
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::parser>::simdjson_result(error_code error) noexcept
     : implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::parser>(error) {}
 
 } // namespace simdjson
@@ -31552,7 +30988,7 @@ inline void stage1_worker::run(document_stream * ds, parser * stage1, size_t nex
 
 #endif  // SIMDJSON_THREADS_ENABLED
 
-simdjson_really_inline document_stream::document_stream(
+simdjson_inline document_stream::document_stream(
   ondemand::parser &_parser,
   const uint8_t *_buf,
   size_t _len,
@@ -31574,7 +31010,7 @@ simdjson_really_inline document_stream::document_stream(
 #endif
 }
 
-simdjson_really_inline document_stream::document_stream() noexcept
+simdjson_inline document_stream::document_stream() noexcept
   : parser{nullptr},
     buf{nullptr},
     len{0},
@@ -31586,7 +31022,7 @@ simdjson_really_inline document_stream::document_stream() noexcept
 {
 }
 
-simdjson_really_inline document_stream::~document_stream() noexcept
+simdjson_inline document_stream::~document_stream() noexcept
 {
   #ifdef SIMDJSON_THREADS_ENABLED
   worker.reset();
@@ -31602,20 +31038,20 @@ inline size_t document_stream::truncated_bytes() const noexcept {
   return parser->implementation->structural_indexes[parser->implementation->n_structural_indexes] - parser->implementation->structural_indexes[parser->implementation->n_structural_indexes + 1];
 }
 
-simdjson_really_inline document_stream::iterator::iterator() noexcept
+simdjson_inline document_stream::iterator::iterator() noexcept
   : stream{nullptr}, finished{true} {
 }
 
-simdjson_really_inline document_stream::iterator::iterator(document_stream* _stream, bool is_end) noexcept
+simdjson_inline document_stream::iterator::iterator(document_stream* _stream, bool is_end) noexcept
   : stream{_stream}, finished{is_end} {
 }
 
-simdjson_really_inline simdjson_result<ondemand::document_reference> document_stream::iterator::operator*() noexcept {
+simdjson_inline simdjson_result<ondemand::document_reference> document_stream::iterator::operator*() noexcept {
   //if(stream->error) { return stream->error; }
   return simdjson_result<ondemand::document_reference>(stream->doc, stream->error);
 }
 
-simdjson_really_inline document_stream::iterator& document_stream::iterator::operator++() noexcept {
+simdjson_inline document_stream::iterator& document_stream::iterator::operator++() noexcept {
   // If there is an error, then we want the iterator
   // to be finished, no matter what. (E.g., we do not
   // keep generating documents with errors, or go beyond
@@ -31642,17 +31078,17 @@ simdjson_really_inline document_stream::iterator& document_stream::iterator::ope
   return *this;
 }
 
-simdjson_really_inline bool document_stream::iterator::operator!=(const document_stream::iterator &other) const noexcept {
+simdjson_inline bool document_stream::iterator::operator!=(const document_stream::iterator &other) const noexcept {
   return finished != other.finished;
 }
 
-simdjson_really_inline document_stream::iterator document_stream::begin() noexcept {
+simdjson_inline document_stream::iterator document_stream::begin() noexcept {
   start();
   // If there are no documents, we're finished.
   return iterator(this, error == EMPTY);
 }
 
-simdjson_really_inline document_stream::iterator document_stream::end() noexcept {
+simdjson_inline document_stream::iterator document_stream::end() noexcept {
   return iterator(this, true);
 }
 
@@ -31782,11 +31218,11 @@ inline error_code document_stream::run_stage1(ondemand::parser &p, size_t _batch
   }
 }
 
-simdjson_really_inline size_t document_stream::iterator::current_index() const noexcept {
+simdjson_inline size_t document_stream::iterator::current_index() const noexcept {
   return stream->doc_index;
 }
 
-simdjson_really_inline std::string_view document_stream::iterator::source() const noexcept {
+simdjson_inline std::string_view document_stream::iterator::source() const noexcept {
   auto depth = stream->doc.iter.depth();
   auto cur_struct_index = stream->doc.iter._root - stream->parser->implementation->structural_indexes.get();
 
@@ -31861,13 +31297,13 @@ inline void document_stream::start_stage1_thread() noexcept {
 
 namespace simdjson {
 
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_stream>::simdjson_result(
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_stream>::simdjson_result(
   error_code error
 ) noexcept :
     implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_stream>(error)
 {
 }
-simdjson_really_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_stream>::simdjson_result(
+simdjson_inline simdjson_result<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_stream>::simdjson_result(
   SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_stream &&value
 ) noexcept :
     implementation_simdjson_result_base<SIMDJSON_BUILTIN_IMPLEMENTATION::ondemand::document_stream>(
diff --git a/tests/conftest.py b/tests/conftest.py
index a3d14a0..c598996 100644
--- a/tests/conftest.py
+++ b/tests/conftest.py
@@ -42,6 +42,8 @@ def doc(parser):
         "int64": -1,
         "uint64": 18446744073709551615,
         "double": 1.1,
+        "double_zero": 0.0,
+        "double_whole": 1.0,
         "string": "test",
         "bool": true,
         "null_value": null
diff --git a/tests/test_float.py b/tests/test_float.py
new file mode 100644
index 0000000..ca5e427
--- /dev/null
+++ b/tests/test_float.py
@@ -0,0 +1,9 @@
+
+def test_parser_understands_float(parser):
+    doc = parser.parse('1.0')
+    assert type(doc) is float
+
+
+def test_mini_does_not_drop_zero(parser):
+    doc = parser.parse(b'[0.0, 0.5, 1.0]')
+    assert doc.mini == b'[0.0,0.5,1.0]'