[Transforms] Let amdgcn take advantage of sin(-x) --> -sin(x)

[AtariDreams]

We do it for amdgcn_cos, and we should do it for amdgcn_sin as well.

[llvmbot]

@llvm/pr-subscribers-backend-amdgpu

@llvm/pr-subscribers-llvm-transforms

Author: AtariDreams (AtariDreams)

Changes

We do it for amdgcn_cos, and we should do it for amdgcn_sin as well.

---

Full diff: https://github.com/llvm/llvm-project/pull/79700.diff

1 Files Affected:

• (modified) llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp (+3-2)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
index a647be2d26c7613..ce54d49323dc6eb 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -2497,11 +2497,12 @@ Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) {
     }
     break;
   }
-  case Intrinsic::sin: {
+  case Intrinsic::sin:
+  case Intrinsic::amdgcn_sin: {
     Value *X;
     if (match(II->getArgOperand(0), m_OneUse(m_FNeg(m_Value(X))))) {
       // sin(-x) --> -sin(x)
-      Value *NewSin = Builder.CreateUnaryIntrinsic(Intrinsic::sin, X, II);
+      Value *NewSin = Builder.CreateUnaryIntrinsic(IID, X, II);
       Instruction *FNeg = UnaryOperator::CreateFNeg(NewSin);
       FNeg->copyFastMathFlags(II);
       return FNeg;

[arsenm]

Missing test

[AtariDreams]

Missing test

Fixed! @arsenm

[arsenm]

This is the opposite of what we want for codegen, but I believe we undo this already

[yashssh]

Out of curiosity what's the purpose of doing this? From one of the tests attached It just looks like we are moving fneg from before sin to after.

[arsenm]

Why close this?