[InstCombine] Resolve TODO: Remove one-time check if other logic operand (Y) is constant

[AtariDreams]

By using match(W, m_ImmConstant()), we do not need to worry about one-time use anymore.

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: AtariDreams (AtariDreams)

Changes

By using isa<Constant>(W), we do not need to worry about one-time use anymore.

---

Full diff: https://github.com/llvm/llvm-project/pull/77973.diff

1 Files Affected:

• (modified) llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp (+8-4)

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp b/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp
index b7958978c450c9..8614aad529eb4a 100644
--- a/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineShifts.cpp
@@ -368,10 +368,14 @@ static Instruction *foldShiftOfShiftedBinOp(BinaryOperator &I,
 
   // Find a matching one-use shift by constant. The fold is not valid if the sum
   // of the shift values equals or exceeds bitwidth.
-  // TODO: Remove the one-use check if the other logic operand (Y) is constant.
   Value *X, *Y;
-  auto matchFirstShift = [&](Value *V) {
+  auto matchFirstShift = [&](Value *V, Value *W) {
     APInt Threshold(Ty->getScalarSizeInBits(), Ty->getScalarSizeInBits());
+    if (isa<Constant>(W)) {
+      return match(V, (m_BinOp(ShiftOpcode, m_Value(X), m_Constant(C0)))) &&
+             match(ConstantExpr::getAdd(C0, C1),
+                   m_SpecificInt_ICMP(ICmpInst::ICMP_ULT, Threshold));
+    }
     return match(V,
                  m_OneUse(m_BinOp(ShiftOpcode, m_Value(X), m_Constant(C0)))) &&
            match(ConstantExpr::getAdd(C0, C1),
@@ -382,9 +386,9 @@ static Instruction *foldShiftOfShiftedBinOp(BinaryOperator &I,
   // is not so we cannot reoder if we match operand(1) and need to keep the
   // operands in their original positions.
   bool FirstShiftIsOp1 = false;
-  if (matchFirstShift(BinInst->getOperand(0)))
+  if (matchFirstShift(BinInst->getOperand(0), BinInst->getOperand(1)))
     Y = BinInst->getOperand(1);
-  else if (matchFirstShift(BinInst->getOperand(1))) {
+  else if (matchFirstShift(BinInst->getOperand(1), BinInst->getOperand(0))) {
     Y = BinInst->getOperand(0);
     FirstShiftIsOp1 = BinInst->getOpcode() == Instruction::Sub;
   } else

[nikic]

Tests?

[AtariDreams]

Tests?

Sure. I added a whole bunch.

[AtariDreams]

@nikic Alright. I think I got this right. Though I am concerned about the shl undef tests that seem to fold things into poison values. I do not know if this is what we want.

[AtariDreams]

@nikic Can I use the other tests that I removed but is affected by the transform then?

[nikic]

You need the simplest possible test (no vectors...) that does not fold before your changes and folds after your changes. Please find this test on your own.

[AtariDreams]

You need the simplest possible test (no vectors...) that does not fold before your changes and folds after your changes. Please find this test on your own.

Did. @nikic Is this good?

[jrtc27]

Uh, did update_llc_test_checks.py really add that comment?...

[AtariDreams]

It removed the comment, not added it @nikic

[AtariDreams]

It added the comment as it was in the original IR, but when it was transformed llc removed it.

[jrtc27]

I know, but why was it there in the first place. The previous commit added it, but that suggests the first commit is wrong.

And please do not resolve conversations off the bat like that.

[AtariDreams]

Because I put it there originally.

[jrtc27]

Well don't then?

And for the second time, stop resolving conversations like that.

[AtariDreams]

Alright @nikic Can we please merge this?

[AtariDreams]

@nikic Done!

[nikic]

LGTM

[LLVMCoCCommittee]

@AtariDreams We have reached out to request your feedback via the email attached to your GitHub account. Please respond to conduct@llvm.org