Z: Move off-heap zero length allocation OOL

Refactored off-heap 2D array allocation logic to improve clarity and
maintainability on IBM Z platform.
Previously, STOCG was used to store the data address only when the
length was non-zero, relying on condition codes to control flow. This
approach introduced hidden dependencies on allocator condition codes,
making the code harder to read and prone to future bugs.
Updated the implementation to allocate non-zero length off-heap arrays
inline, while handling zero-length arrays out of line. This removes
reliance on condition codes for allocation decisions, simplifies
control flow, and ensures consistent behavior across allocation paths.

signed-off-by: Ehsan Kiani Far <ehsan.kianifar@gmail.com>

Author: ehsankianifar

runtime/compiler/z/codegen/J9TreeEvaluator.cpp

@@ -4905,6 +4905,7 @@ static TR::Register * generateMultianewArrayWithInlineAllocators(TR::Node *node,
       }
 
    TR::LabelSymbol *zeroSecondDimLabel = generateLabelSymbol(cg);
+
    // Bypass dim2 size calculation if the size is zero.
    cursor = generateS390BranchInstruction(cg, TR::InstOpCode::BRC, TR::InstOpCode::COND_BZ, node, zeroSecondDimLabel, cursor);
    // Call helper if the length is negative or length * componentSize is larger that INT32_MAX (overflow).
@@ -5017,16 +5018,19 @@ static TR::Register * generateMultianewArrayWithInlineAllocators(TR::Node *node,
 
 #if defined(J9VM_GC_SPARSE_HEAP_ALLOCATION)
    bool isOffHeapAllocationEnabled = TR::Compiler->om.isOffHeapAllocationEnabled();
+   TR::LabelSymbol *zeroLengthSecondDimLabel = NULL;
    if (isOffHeapAllocationEnabled)
       {
       // Store first element's address in data address field.
       cursor = generateRXInstruction(cg, TR::InstOpCode::STG, node, sizeReg, generateS390MemoryReference(resultReg, fej9->getOffsetOfContiguousDataAddrField(), cg), cursor);
-      // Subtract the header size from the dim2 size, so we can move to the next leaf by adding this value to the address of the leaf's first element.
-      cursor = generateRILInstruction(cg, TR::InstOpCode::SLFI, node, dim2SizeReg, (int32_t)TR::Compiler->om.contiguousArrayHeaderSizeInBytes(), cursor);
+      // If the second dimension length is zero, allocate the leaf array OOL.
+      zeroLengthSecondDimLabel = generateLabelSymbol(cg);
       // The length of the second dimension is stored in the upper 32 bits of the scratch register.
-      // Comparing the full 64-bit scratch register with its lower 32 bits sets the condition code to COND_BNE if the second dimension is non-zero.
+      // Comparing the full 64-bit scratch register with its lower 32 bits sets the condition code to COND_BE if the second dimension is zero.
       cursor = generateRREInstruction(cg, TR::InstOpCode::CGFR, node, scratchReg, scratchReg, cursor);
-      // The condition code has already been set and must remain unchanged during the second dimension allocation loop in this code path.
+      cursor = generateS390BranchInstruction(cg, TR::InstOpCode::BRC, TR::InstOpCode::COND_BE, node, zeroLengthSecondDimLabel, cursor);
+      // Subtract the header size from the dim2 size, so we can move to the next leaf by adding this value to the address of the leaf's first element.
+      cursor = generateRILInstruction(cg, TR::InstOpCode::SLFI, node, dim2SizeReg, (int32_t)TR::Compiler->om.contiguousArrayHeaderSizeInBytes(), cursor);
       }
 #endif /* J9VM_GC_SPARSE_HEAP_ALLOCATION */
 
@@ -5040,59 +5044,75 @@ static TR::Register * generateMultianewArrayWithInlineAllocators(TR::Node *node,
    cursor = generateRXInstruction(cg, TR::InstOpCode::STFH, node, scratchReg, generateS390MemoryReference(dim1SizeReg,
       static_cast<int32_t>(TR::Compiler->om.offsetOfContiguousArraySizeField()), cg), cursor);
    int32_t shiftAmount = TR::Compiler->om.compressedReferenceShift();
-
-   if (shiftAmount == 0)
+   if (shiftAmount > 0)
+      {
+      // Calculate the compressed reference of leaf array in the higher 32bits of dim2SizeReg.
+      cursor = generateRIEInstruction(cg, TR::InstOpCode::RISBG, node, dim2SizeReg, dim1SizeReg, 0, 31, 32-shiftAmount, cursor);
+      // Store the compressed 32 bit leaf address in the relevant element of the first dim array.
+      cursor = generateRXInstruction(cg, TR::InstOpCode::STFH, node, dim2SizeReg, generateS390MemoryReference(sizeReg, 0, cg), cursor);
+      }
+   else
       {
-      // Store the leaf address in the relevant element of the first dim array.
+      // Store the 32 bit leaf address in the relevant element of the first dim array.
       cursor = generateRXInstruction(cg, (TR::Compiler->om.compressObjectReferences() ? TR::InstOpCode::ST : TR::InstOpCode::STG), node, dim1SizeReg,
          generateS390MemoryReference(sizeReg, 0, cg), cursor);
       }
 
 #if defined(J9VM_GC_SPARSE_HEAP_ALLOCATION)
    if (isOffHeapAllocationEnabled)
       {
-      if (shiftAmount > 0)
-         {
-         // Compress the leaf array reference and store it in the first element of the dimension array.
-         // Avoid using any instructions that could affect the branch condition in this code path.
-         cursor = generateRSInstruction(cg, TR::InstOpCode::SRLG, node, dim1SizeReg, dim1SizeReg, shiftAmount, cursor);
-         cursor = generateRXInstruction(cg, TR::InstOpCode::ST, node, dim1SizeReg, generateS390MemoryReference(sizeReg, 0, cg), cursor);
-         // Decompress the leaf array reference.
-         cursor = generateRSInstruction(cg, TR::InstOpCode::SLLG, node, dim1SizeReg, dim1SizeReg, shiftAmount, cursor);
-         }
       // Load the first element address.
       cursor = generateRXInstruction(cg, TR::InstOpCode::LA, node, dim1SizeReg,
          generateS390MemoryReference(dim1SizeReg, TR::Compiler->om.contiguousArrayHeaderSizeInBytes(), cg), cursor);
-      // Store the address of the first element in the data address field only if the second dimension length is non-zero.
-      cursor = generateRSInstruction(cg, TR::InstOpCode::STOCG, node, dim1SizeReg, getMaskForBranchCondition(TR::InstOpCode::COND_BNE),
-         generateS390MemoryReference(dim1SizeReg, (fej9->getOffsetOfContiguousDataAddrField() - TR::Compiler->om.contiguousArrayHeaderSizeInBytes()), cg), cursor);
-      // Load the next leaf address in dim1SizeReg.
-      cursor = generateRXInstruction(cg, TR::InstOpCode::LA, node, dim1SizeReg, generateS390MemoryReference(dim1SizeReg, dim2SizeReg, 0, cg), cursor);
+      // Store the first element address in data address field.
+      cursor = generateRXInstruction(cg, TR::InstOpCode::STG, node, dim1SizeReg, generateS390MemoryReference(dim1SizeReg,
+         (fej9->getOffsetOfContiguousDataAddrField() - TR::Compiler->om.contiguousArrayHeaderSizeInBytes()), cg), cursor);
       }
-   else
 #endif /* J9VM_GC_SPARSE_HEAP_ALLOCATION */
+
+   // Load the next leaf address in dim1SizeReg.
+   cursor = generateRREInstruction(cg, TR::InstOpCode::ALGFR, node, dim1SizeReg, dim2SizeReg, cursor);
+   // Load the next element address of the first dim array in sizeReg.
+   cursor = generateRXInstruction(cg, TR::InstOpCode::LA, node, sizeReg, generateS390MemoryReference(sizeReg, elementSize, cg), cursor);
+   cursor = generateS390BranchInstruction(cg, TR::InstOpCode::BRCT, node, scratchReg, secondDimLabel, cursor);
+   cursor = generateS390BranchInstruction(cg, TR::InstOpCode::BRC, TR::InstOpCode::COND_B, node, controlFlowEndLabel, cursor);
+
+   /********************************************* OOL zero length offheap leaf allocator *********************************************/
+#if defined(J9VM_GC_SPARSE_HEAP_ALLOCATION)
+   if (isOffHeapAllocationEnabled)
       {
+      TR_S390OutOfLineCodeSection *zeroLengthSecondDimAlloc = new (cg->trHeapMemory()) TR_S390OutOfLineCodeSection(zeroLengthSecondDimLabel, controlFlowEndLabel, cg);
+      cg->getS390OutOfLineCodeSectionList().push_front(zeroLengthSecondDimAlloc);
+      zeroLengthSecondDimAlloc->swapInstructionListsWithCompilation();
+      generateS390LabelInstruction(cg, TR::InstOpCode::label, node, zeroLengthSecondDimLabel);
+
+      // Store the class field.
+      generateRXInstruction(cg, (compressedObjectHeaders ? TR::InstOpCode::ST : TR::InstOpCode::STG), node, classReg,
+         generateS390MemoryReference(dim1SizeReg, static_cast<int32_t>(TR::Compiler->om.offsetOfObjectVftField()), cg));
+      // Store the array length.
       if (shiftAmount > 0)
          {
          // Calculate the compressed reference of leaf array in the higher 32bits of dim2SizeReg.
-         cursor = generateRIEInstruction(cg, TR::InstOpCode::RISBG, node, dim2SizeReg, dim1SizeReg, 0, 31, 32-shiftAmount, cursor);
+         generateRIEInstruction(cg, TR::InstOpCode::RISBG, node, dim2SizeReg, dim1SizeReg, 0, 31, 32-shiftAmount);
          // Store the compressed 32 bit leaf address in the relevant element of the first dim array.
-         cursor = generateRXInstruction(cg, TR::InstOpCode::STFH, node, dim2SizeReg, generateS390MemoryReference(sizeReg, 0, cg), cursor);
-         // Load the next leaf address in dim1SizeReg.
-         cursor = generateRREInstruction(cg, TR::InstOpCode::ALGFR, node, dim1SizeReg, dim2SizeReg, cursor);
+         generateRXInstruction(cg, TR::InstOpCode::STFH, node, dim2SizeReg, generateS390MemoryReference(sizeReg, 0, cg));
          }
       else
          {
-         // Load the next leaf address in dim1SizeReg.
-         cursor = generateRXInstruction(cg, TR::InstOpCode::LA, node, dim1SizeReg, generateS390MemoryReference(dim1SizeReg, dim2SizeReg, 0, cg), cursor);
+         // Store the 32 bit leaf address in the relevant element of the first dim array.
+         generateRXInstruction(cg, (TR::Compiler->om.compressObjectReferences() ? TR::InstOpCode::ST : TR::InstOpCode::STG), node, dim1SizeReg,
+            generateS390MemoryReference(sizeReg, 0, cg));
          }
-      }
+      // Load the next leaf address in dim1SizeReg.
+      generateRREInstruction(cg, TR::InstOpCode::ALGFR, node, dim1SizeReg, dim2SizeReg);
+      // Load the next element address of the first dim array in sizeReg.
+      generateRXInstruction(cg, TR::InstOpCode::LA, node, sizeReg, generateS390MemoryReference(sizeReg, elementSize, cg));
+      generateS390BranchInstruction(cg, TR::InstOpCode::BRCT, node, scratchReg, zeroLengthSecondDimLabel);
 
-   // Load the next element address of the first dim array in sizeReg.
-   cursor = generateRXInstruction(cg, TR::InstOpCode::LA, node, sizeReg, generateS390MemoryReference(sizeReg, elementSize, cg), cursor);
-   cursor = generateS390BranchInstruction(cg, TR::InstOpCode::BRCT, node, scratchReg, secondDimLabel, cursor);
-   cursor = generateS390BranchInstruction(cg, TR::InstOpCode::BRC, TR::InstOpCode::COND_B, node, controlFlowEndLabel, cursor);
-   iComment("Allocation done!");
+      generateS390BranchInstruction(cg, TR::InstOpCode::BRC, TR::InstOpCode::COND_BRC, node, controlFlowEndLabel);
+      zeroLengthSecondDimAlloc->swapInstructionListsWithCompilation();
+      }
+#endif /* J9VM_GC_SPARSE_HEAP_ALLOCATION */
 
    TR::LabelSymbol *slowPathLabel = generateLabelSymbol(cg);
    cursor = generateS390LabelInstruction(cg, TR::InstOpCode::label, node, inlineAllocFailLabel, cursor);