[DAGCombiner] Relax nsz constraint for FP optimizations (#165011)

guy-david · web-flow · commit 29611f4cbea0 · 2025-12-09T12:07:46.000+02:00
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 guaranteed to overwrite the sign-bit or
simply ignore it (comparisons, etc.).

The optimizations affected:
- fadd x, +0.0 -&gt; x
- fsub x, -0.0 -&gt; x
- fsub +0.0, x -&gt; fneg x
- fdiv(x, sqrt(x)) -&gt; sqrt(x)
- frem lowering with power-of-2 divisors
diff --git a/llvm/include/llvm/CodeGen/SelectionDAG.h b/llvm/include/llvm/CodeGen/SelectionDAG.h
@@ -2334,6 +2334,12 @@ class SelectionDAG {
   /// +nan are considered positive, -0.0, -inf and -nan are not.
   LLVM_ABI bool cannotBeOrderedNegativeFP(SDValue Op) const;
 
+  /// Check if a use of a float value is insensitive to signed zeros.
+  LLVM_ABI bool canIgnoreSignBitOfZero(const SDUse &Use) const;
+
+  /// Check if at most two uses of a value are insensitive to signed zeros.
+  LLVM_ABI bool canIgnoreSignBitOfZero(SDValue Op) const;
+
   /// Test whether two SDValues are known to compare equal. This
   /// is true if they are the same value, or if one is negative zero and the
   /// other positive zero.
diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -17920,7 +17920,8 @@ 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.canIgnoreSignBitOfZero(SDValue(N, 0)))
       return N0;
 
   if (SDValue NewSel = foldBinOpIntoSelect(N))
@@ -18132,7 +18133,8 @@ SDValue DAGCombiner::visitFSUB(SDNode *N) {
 
   // (fsub A, 0) -> A
   if (N1CFP && N1CFP->isZero()) {
-    if (!N1CFP->isNegative() || Flags.hasNoSignedZeros()) {
+    if (!N1CFP->isNegative() || Flags.hasNoSignedZeros() ||
+        DAG.canIgnoreSignBitOfZero(SDValue(N, 0))) {
       return N0;
     }
   }
@@ -18145,7 +18147,8 @@ 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.canIgnoreSignBitOfZero(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
@@ -18795,7 +18798,8 @@ SDValue DAGCombiner::visitFDIV(SDNode *N) {
   }
 
   // Fold X/Sqrt(X) -> Sqrt(X)
-  if (Flags.hasNoSignedZeros() && Flags.hasAllowReassociation())
+  if ((Flags.hasNoSignedZeros() || DAG.canIgnoreSignBitOfZero(SDValue(N, 0))) &&
+      Flags.hasAllowReassociation())
     if (N1.getOpcode() == ISD::FSQRT && N0 == N1.getOperand(0))
       return N1;
 
@@ -18846,8 +18850,9 @@ 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.canIgnoreSignBitOfZero(SDValue(N, 0)) &&
+                         !DAG.cannotBeOrderedNegativeFP(N0);
     SDValue Div = DAG.getNode(ISD::FDIV, DL, VT, N0, N1);
     SDValue Rnd = DAG.getNode(ISD::FTRUNC, DL, VT, Div);
     SDValue MLA;
diff --git a/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp
@@ -6160,6 +6160,46 @@ bool SelectionDAG::cannotBeOrderedNegativeFP(SDValue Op) const {
   llvm_unreachable("covered opcode switch");
 }
 
+bool SelectionDAG::canIgnoreSignBitOfZero(const SDUse &Use) const {
+  assert(Use.getValueType().isFloatingPoint());
+  const SDNode *User = Use.getUser();
+  unsigned OperandNo = Use.getOperandNo();
+  // Check if this use is insensitive to the sign of zero
+  switch (User->getOpcode()) {
+  case ISD::SETCC:
+    // Comparisons: IEEE-754 specifies +0.0 == -0.0.
+  case ISD::FABS:
+    // fabs always produces +0.0.
+    return true;
+  case ISD::FCOPYSIGN:
+    // copysign overwrites the sign bit of the first operand.
+    return OperandNo == 0;
+  case ISD::FADD:
+  case ISD::FSUB: {
+    // Arithmetic with non-zero constants fixes the uncertainty around the
+    // sign bit.
+    SDValue Other = User->getOperand(1 - OperandNo);
+    return isKnownNeverZeroFloat(Other);
+  }
+  case ISD::FP_TO_SINT:
+  case ISD::FP_TO_UINT:
+    // fp-to-int conversions normalize signed zeros.
+    return true;
+  default:
+    return false;
+  }
+}
+
+bool SelectionDAG::canIgnoreSignBitOfZero(SDValue Op) const {
+  // FIXME: Limit the amount of checked uses to not introduce a compile-time
+  // regression. Ideally, this should be implemented as a demanded-bits
+  // optimization that stems from the users.
+  if (Op->use_size() > 2)
+    return false;
+  return all_of(Op->uses(),
+                [&](const SDUse &Use) { return canIgnoreSignBitOfZero(Use); });
+}
+
 bool SelectionDAG::isEqualTo(SDValue A, SDValue B) const {
   // Check the obvious case.
   if (A == B) return true;
diff --git a/llvm/test/CodeGen/AArch64/ignore-signed-zero.ll b/llvm/test/CodeGen/AArch64/ignore-signed-zero.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)
diff --git a/llvm/test/CodeGen/AMDGPU/fcanonicalize-elimination.ll b/llvm/test/CodeGen/AMDGPU/fcanonicalize-elimination.ll
@@ -371,7 +371,7 @@ define amdgpu_kernel void @test_fold_canonicalize_fabs_value_f32(ptr addrspace(1
   %id = tail call i32 @llvm.amdgcn.workitem.id.x()
   %gep = getelementptr inbounds float, ptr addrspace(1) %arg, i32 %id
   %load = load float, ptr addrspace(1) %gep, align 4
-  %v0 = fadd float %load, 0.0
+  %v0 = fadd float %load, 1.0
   %v = tail call float @llvm.fabs.f32(float %v0)
   %canonicalized = tail call float @llvm.canonicalize.f32(float %v)
   store float %canonicalized, ptr addrspace(1) %gep, align 4
diff --git a/llvm/test/CodeGen/AMDGPU/swdev380865.ll b/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
