[llvm] [SLP] Avoid seeding related affine loop address computations (PR #226220)
Alexey Bataev via llvm-commits
llvm-commits at lists.llvm.org
Sun Oct 11 09:48:52 PDT 2026
================
@@ -36569,13 +36575,22 @@ bool SLPVectorizerPass::vectorizeOnceUsedSeeds(BasicBlock *BB, BoUpSLP &R) {
SmallDenseMap<std::pair<size_t, size_t>, unsigned> KeyToGroup;
SmallDenseMap<Value *, unsigned> SeedGroups;
PoorThroughputOpCache PoorThroughputCache;
+ SmallPtrSet<Value *, 16> IndexedGEPsInBB;
+ if (auto It = IndexedGEPs.find(BB); It != IndexedGEPs.end())
+ for (Value *GEP : It->second)
+ if (GEP)
+ IndexedGEPsInBB.insert(GEP);
for (Instruction &I : make_filter_range(*BB, [&](Instruction &I) {
return !R.isDeleted(&I) && I.hasOneUse() && !R.isEphemeralValue(&I) &&
!R.isVectorized(&I) && !R.isAnalyzedScalar(&I) &&
isValidElementType(getValueType(&I, SLPReVec), SLPReVec) &&
isOnceUsedSeed(&I) && !isNonVectorizableInst(&I, TLI) &&
!R.hasResolvedUser(&I);
----------------
alexey-bataev wrote:
```suggestion
!R.hasResolvedUser(&I) && !SkippedInBB.contains(&I);
```
https://github.com/llvm/llvm-project/pull/226220
More information about the llvm-commits
mailing list