Skip to content

[VectorCombine] foldShuffleOfIntrinsics - support multiple uses of shuffled ops #173037

New issue
New issue
Closed
#173183
Closed
[VectorCombine] foldShuffleOfIntrinsics - support multiple uses of shuffled ops#173037
#173183
Assignees
Xylecrack
Labels
good first issuehttps://github.com/llvm/llvm-project/contributellvm::vectorcombineCost-based vector combine passmissed-optimization
@RKSimon

Description

@RKSimon
RKSimon
opened on Dec 19, 2025

The "shuffle (intrinsic), (intrinsic)" -> "intrinsic (shuffle), (shuffle)" fold is currently limited by a mOneUse check:

llvm-project/llvm/lib/Transforms/Vectorize/VectorCombine.cpp

Lines 3076 to 3081 in ba55ef3

bool VectorCombine::foldShuffleOfIntrinsics(Instruction &I) {
Value *V0, *V1;
ArrayRef<int> OldMask;
if (!match(&I, m_Shuffle(m_OneUse(m_Value(V0)), m_OneUse(m_Value(V1)),
m_Mask(OldMask))))
return false;

We should try to remove the m_OneUse limits and adjust the NewCost accordingly - similar to what happens in foldShuffleOfCastops

A large portion of this task will be creating test coverage.

Metadata

Metadata

Assignees

Labels

good first issuehttps://github.com/llvm/llvm-project/contributellvm::vectorcombineCost-based vector combine passmissed-optimization

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions