[DAGCombiner] Relax nsz constraint for more FP optimizations

Some floating-point optimization don't trigger because they can produce
incorrect results around signed zeros, and rely on the existence of the
nsz flag which commonly appears when fast-math is enabled.
However, this flag is not a hard requirement when all of the users of
the combined value are either guranteed to overwrite the sign-bit or
simply ignore it (comparisons, etc.).

The optimizations affected:
- fadd x, -0.0 -> x
- fsub x, 0.0 -> x
- fsub -0.0, x -> fneg x
- fdiv x, sqrt(x) -> sqrt(x)
- frem lowering with power-of-2 divisors

Author: guy-david

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

@@ -17781,7 +17781,8 @@ SDValue DAGCombiner::visitFADD(SDNode *N) {
   // N0 + -0.0 --> N0 (also allowed with +0.0 and fast-math)
   ConstantFPSDNode *N1C = isConstOrConstSplatFP(N1, true);
   if (N1C && N1C->isZero())
-    if (N1C->isNegative() || Flags.hasNoSignedZeros())
+    if (N1C->isNegative() || Flags.hasNoSignedZeros() ||
+        DAG.allUsesSignedZeroInsensitive(SDValue(N, 0)))
       return N0;
 
   if (SDValue NewSel = foldBinOpIntoSelect(N))
@@ -17993,7 +17994,8 @@ SDValue DAGCombiner::visitFSUB(SDNode *N) {
 
   // (fsub A, 0) -> A
   if (N1CFP && N1CFP->isZero()) {
-    if (!N1CFP->isNegative() || Flags.hasNoSignedZeros()) {
+    if (!N1CFP->isNegative() || Flags.hasNoSignedZeros() ||
+        DAG.allUsesSignedZeroInsensitive(SDValue(N, 0))) {
       return N0;
     }
   }
@@ -18006,7 +18008,8 @@ SDValue DAGCombiner::visitFSUB(SDNode *N) {
 
   // (fsub -0.0, N1) -> -N1
   if (N0CFP && N0CFP->isZero()) {
-    if (N0CFP->isNegative() || Flags.hasNoSignedZeros()) {
+    if (N0CFP->isNegative() || Flags.hasNoSignedZeros() ||
+        DAG.allUsesSignedZeroInsensitive(SDValue(N, 0))) {
       // We cannot replace an FSUB(+-0.0,X) with FNEG(X) when denormals are
       // flushed to zero, unless all users treat denorms as zero (DAZ).
       // FIXME: This transform will change the sign of a NaN and the behavior
@@ -18654,7 +18657,9 @@ SDValue DAGCombiner::visitFDIV(SDNode *N) {
   }
 
   // Fold X/Sqrt(X) -> Sqrt(X)
-  if (Flags.hasNoSignedZeros() && Flags.hasAllowReassociation())
+  if ((Flags.hasNoSignedZeros() ||
+       DAG.allUsesSignedZeroInsensitive(SDValue(N, 0))) &&
+      Flags.hasAllowReassociation())
     if (N1.getOpcode() == ISD::FSQRT && N0 == N1.getOperand(0))
       return N1;
 
@@ -18705,8 +18710,9 @@ SDValue DAGCombiner::visitFREM(SDNode *N) {
       TLI.isOperationLegalOrCustom(ISD::FDIV, VT) &&
       TLI.isOperationLegalOrCustom(ISD::FTRUNC, VT) &&
       DAG.isKnownToBeAPowerOfTwoFP(N1)) {
-    bool NeedsCopySign =
-        !Flags.hasNoSignedZeros() && !DAG.cannotBeOrderedNegativeFP(N0);
+    bool NeedsCopySign = !Flags.hasNoSignedZeros() &&
+                         !DAG.cannotBeOrderedNegativeFP(N0) &&
+                         !DAG.allUsesSignedZeroInsensitive(SDValue(N, 0));
     SDValue Div = DAG.getNode(ISD::FDIV, DL, VT, N0, N1);
     SDValue Rnd = DAG.getNode(ISD::FTRUNC, DL, VT, Div);
     SDValue MLA;

llvm/test/CodeGen/AArch64/nsz-bypass.ll

@@ -0,0 +1,72 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=aarch64 | FileCheck %s
+
+; Test that nsz constraint can be bypassed when all uses are sign-insensitive.
+
+define i1 @test_fadd_neg_zero_fcmp(float %x) {
+; CHECK-LABEL: test_fadd_neg_zero_fcmp:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    fmov s1, #1.00000000
+; CHECK-NEXT:    fcmp s0, s1
+; CHECK-NEXT:    cset w0, eq
+; CHECK-NEXT:    ret
+  %add = fadd float %x, -0.0
+  %cmp = fcmp oeq float %add, 1.0
+  ret i1 %cmp
+}
+
+define float @test_fsub_zero_fabs(float %x) {
+; CHECK-LABEL: test_fsub_zero_fabs:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    fabs s0, s0
+; CHECK-NEXT:    ret
+  %sub = fsub float %x, 0.0
+  %abs = call float @llvm.fabs.f32(float %sub)
+  ret float %abs
+}
+
+define float @test_fsub_neg_zero_copysign(float %x, float %y) {
+; CHECK-LABEL: test_fsub_neg_zero_copysign:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    mvni v2.4s, #128, lsl #24
+; CHECK-NEXT:    // kill: def $s0 killed $s0 def $q0
+; CHECK-NEXT:    // kill: def $s1 killed $s1 def $q1
+; CHECK-NEXT:    bif v0.16b, v1.16b, v2.16b
+; CHECK-NEXT:    // kill: def $s0 killed $s0 killed $q0
+; CHECK-NEXT:    ret
+  %sub = fsub float -0.0, %x
+  %copysign = call float @llvm.copysign.f32(float %sub, float %y)
+  ret float %copysign
+}
+
+define i1 @test_div_sqrt_fcmp(float %x) {
+; CHECK-LABEL: test_div_sqrt_fcmp:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    fsqrt s0, s0
+; CHECK-NEXT:    fcmp s0, #0.0
+; CHECK-NEXT:    cset w0, gt
+; CHECK-NEXT:    ret
+  %sqrt = call float @llvm.sqrt.f32(float %x)
+  %div = fdiv reassoc float %x, %sqrt
+  %cmp = fcmp ogt float %div, 0.0
+  ret i1 %cmp
+}
+
+define float @test_frem_fabs(float %x) {
+; CHECK-LABEL: test_frem_fabs:
+; CHECK:       // %bb.0:
+; CHECK-NEXT:    fmov s1, #0.50000000
+; CHECK-NEXT:    fmov s2, #-2.00000000
+; CHECK-NEXT:    fmul s1, s0, s1
+; CHECK-NEXT:    frintz s1, s1
+; CHECK-NEXT:    fmadd s0, s1, s2, s0
+; CHECK-NEXT:    fabs s0, s0
+; CHECK-NEXT:    ret
+  %rem = frem float %x, 2.0
+  %abs = call float @llvm.fabs.f32(float %rem)
+  ret float %abs
+}
+
+declare float @llvm.fabs.f32(float)
+declare float @llvm.copysign.f32(float, float)
+declare float @llvm.sqrt.f32(float)

llvm/test/CodeGen/AMDGPU/swdev380865.ll

@@ -28,14 +28,13 @@ define amdgpu_kernel void @_Z6kernelILi4000ELi1EEvPd(ptr addrspace(1) %x.coerce)
 ; CHECK-NEXT:    v_mov_b32_e32 v1, s7
 ; CHECK-NEXT:  .LBB0_1: ; %for.cond4.preheader
 ; CHECK-NEXT:    ; =>This Inner Loop Header: Depth=1
-; CHECK-NEXT:    v_add_f64 v[0:1], v[0:1], 0
 ; CHECK-NEXT:    s_mov_b32 s6, 0
 ; CHECK-NEXT:    s_mov_b32 s7, 0x40140000
-; CHECK-NEXT:    s_add_i32 s1, s1, s0
-; CHECK-NEXT:    s_cmpk_lt_i32 s1, 0xa00
 ; CHECK-NEXT:    v_add_f64 v[0:1], v[0:1], s[6:7]
 ; CHECK-NEXT:    s_mov_b32 s6, 0
 ; CHECK-NEXT:    s_mov_b32 s7, 0x40180000
+; CHECK-NEXT:    s_add_i32 s1, s1, s0
+; CHECK-NEXT:    s_cmpk_lt_i32 s1, 0xa00
 ; CHECK-NEXT:    v_add_f64 v[0:1], v[0:1], s[6:7]
 ; CHECK-NEXT:    s_mov_b32 s6, 0
 ; CHECK-NEXT:    s_mov_b32 s7, 0x401c0000