From e661185fb3e5a94177404e474d6000f386594090 Mon Sep 17 00:00:00 2001
From: Matt Arsenault <Matthew.Arsenault@amd.com>
Date: Sun, 16 Oct 2022 22:49:30 -0700
Subject: [PATCH] InstCombine: Fold fdiv nnan x, 0 -> copysign(inf, x)

https://alive2.llvm.org/ce/z/gLBFKB
---
 llvm/lib/Transforms/InstCombine/InstCombineInternal.h    |  1 +
 llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp | 13 ++++++++++++-
 llvm/test/Transforms/InstCombine/fdiv.ll                 |  8 ++++----
 3 files changed, 17 insertions(+), 5 deletions(-)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineInternal.h b/llvm/lib/Transforms/InstCombine/InstCombineInternal.h
index 11aed77..cc1dedf 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineInternal.h
+++ b/llvm/lib/Transforms/InstCombine/InstCombineInternal.h
@@ -370,6 +370,7 @@ private:
   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)
diff --git a/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp b/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
index abc88e3..9627530 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineMulDivRem.cpp
@@ -1403,7 +1403,7 @@ Instruction *InstCombinerImpl::visitSDiv(BinaryOperator &I) {
 }
 
 /// 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;
@@ -1415,6 +1415,17 @@ static Instruction *foldFDivConstantDivisor(BinaryOperator &I) {
     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).
diff --git a/llvm/test/Transforms/InstCombine/fdiv.ll b/llvm/test/Transforms/InstCombine/fdiv.ll
index 38557c3..dd49f8b 100644
--- a/llvm/test/Transforms/InstCombine/fdiv.ll
+++ b/llvm/test/Transforms/InstCombine/fdiv.ll
@@ -950,7 +950,7 @@ define float @fdiv_zero_f32(float %x) {
 ; 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 float @fdiv_nnan_zero_f32(float %x) {
 
 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 <2 x float> @fdiv_nnan_zero_v2f32(<2 x float> %x) {
 
 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 float @fdiv_nnan_zero_f32_fmf(float %x) {
 
 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
-- 
2.7.4