[PATCH] D136098: InstCombine: Fold fdiv nnan x, 0 -> copysign(inf, x)

[Matt Arsenault via Phabricator via llvm-commits]

arsenm updated this revision to Diff 468289.
arsenm edited the summary of this revision.
arsenm added a comment.

Rebase on baseline tests


CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D136098/new/

https://reviews.llvm.org/D136098

Files:
  llvm/lib/Transforms/InstCombine/InstCombineInternal.h
  llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
  llvm/test/Transforms/InstCombine/fdiv.ll


Index: llvm/test/Transforms/InstCombine/fdiv.ll
===================================================================
--- llvm/test/Transforms/InstCombine/fdiv.ll
+++ llvm/test/Transforms/InstCombine/fdiv.ll
@@ -950,7 +950,7 @@
 ; https://alive2.llvm.org/ce/z/gLBFKB
 define float @fdiv_nnan_zero_f32(float %x) {
 ; CHECK-LABEL: @fdiv_nnan_zero_f32(
-; CHECK-NEXT:    [[FDIV:%.*]] = fdiv nnan float [[X:%.*]], 0.000000e+00
+; CHECK-NEXT:    [[FDIV:%.*]] = call nnan float @llvm.copysign.f32(float 0x7FF0000000000000, float [[X:%.*]])
 ; CHECK-NEXT:    ret float [[FDIV]]
 ;
   %fdiv = fdiv nnan float %x, 0.0
@@ -959,7 +959,7 @@
 
 define <2 x float> @fdiv_nnan_zero_v2f32(<2 x float> %x) {
 ; CHECK-LABEL: @fdiv_nnan_zero_v2f32(
-; CHECK-NEXT:    [[FDIV:%.*]] = fdiv nnan <2 x float> [[X:%.*]], zeroinitializer
+; CHECK-NEXT:    [[FDIV:%.*]] = call nnan <2 x float> @llvm.copysign.v2f32(<2 x float> <float 0x7FF0000000000000, float 0x7FF0000000000000>, <2 x float> [[X:%.*]])
 ; CHECK-NEXT:    ret <2 x float> [[FDIV]]
 ;
   %fdiv = fdiv nnan <2 x float> %x, zeroinitializer
@@ -968,7 +968,7 @@
 
 define float @fdiv_nnan_zero_f32_fmf(float %x) {
 ; CHECK-LABEL: @fdiv_nnan_zero_f32_fmf(
-; CHECK-NEXT:    [[FDIV:%.*]] = fdiv nnan nsz float [[X:%.*]], 0.000000e+00
+; CHECK-NEXT:    [[FDIV:%.*]] = call nnan nsz float @llvm.copysign.f32(float 0x7FF0000000000000, float [[X:%.*]])
 ; CHECK-NEXT:    ret float [[FDIV]]
 ;
   %fdiv = fdiv nnan nsz float %x, 0.0
@@ -977,7 +977,7 @@
 
 define <2 x float> @fdiv_nnan_zero_v2f32_fmf(<2 x float> %x) {
 ; CHECK-LABEL: @fdiv_nnan_zero_v2f32_fmf(
-; CHECK-NEXT:    [[FDIV:%.*]] = fdiv nnan nsz <2 x float> [[X:%.*]], zeroinitializer
+; CHECK-NEXT:    [[FDIV:%.*]] = call nnan nsz <2 x float> @llvm.copysign.v2f32(<2 x float> <float 0x7FF0000000000000, float 0x7FF0000000000000>, <2 x float> [[X:%.*]])
 ; CHECK-NEXT:    ret <2 x float> [[FDIV]]
 ;
   %fdiv = fdiv nnan nsz <2 x float> %x, zeroinitializer
Index: llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
===================================================================
--- llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
+++ llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
@@ -1372,7 +1372,7 @@
 }
 
 /// Remove negation and try to convert division into multiplication.
-static Instruction *foldFDivConstantDivisor(BinaryOperator &I) {
+Instruction *InstCombinerImpl::foldFDivConstantDivisor(BinaryOperator &I) {
   Constant *C;
   if (!match(I.getOperand(1), m_Constant(C)))
     return nullptr;
@@ -1384,6 +1384,17 @@
     if (Constant *NegC = ConstantFoldUnaryOpOperand(Instruction::FNeg, C, DL))
       return BinaryOperator::CreateFDivFMF(X, NegC, &I);
 
+  // nnan X / +0.0 -> copysign(inf, X)
+  if (I.hasNoNaNs() && match(I.getOperand(1), m_Zero())) {
+    IRBuilder<> B(&I);
+    // TODO: nnan nsz X / -0.0 -> copysign(inf, X)
+    CallInst *CopySign = B.CreateIntrinsic(
+        Intrinsic::copysign, {C->getType()},
+        {ConstantFP::getInfinity(I.getType()), I.getOperand(0)}, &I);
+    CopySign->takeName(&I);
+    return replaceInstUsesWith(I, CopySign);
+  }
+
   // If the constant divisor has an exact inverse, this is always safe. If not,
   // then we can still create a reciprocal if fast-math-flags allow it and the
   // constant is a regular number (not zero, infinite, or denormal).
Index: llvm/lib/Transforms/InstCombine/InstCombineInternal.h
===================================================================
--- llvm/lib/Transforms/InstCombine/InstCombineInternal.h
+++ llvm/lib/Transforms/InstCombine/InstCombineInternal.h
@@ -369,6 +369,7 @@
   Instruction *foldExtractOfOverflowIntrinsic(ExtractValueInst &EV);
   Instruction *foldIntrinsicWithOverflowCommon(IntrinsicInst *II);
   Instruction *foldFPSignBitOps(BinaryOperator &I);
+  Instruction *foldFDivConstantDivisor(BinaryOperator &I);
 
   // Optimize one of these forms:
   //   and i1 Op, SI / select i1 Op, i1 SI, i1 false (if IsAnd = true)


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D136098.468289.patch
Type: text/x-patch
Size: 3962 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20221017/36f01c08/attachment.bin>