[SLP] Fixes bug in ShuffleBuilder when duplicate entry already exists.

[vporpo]

Summary

For more details please check out the lit test: X86/shufflebuilder-bug.ll.

@ABataev I am not sure this is the correct solution, but it seems to be fixing the bug.
Could you please take a look?

[ABataev]

Yeah, I doubt it is the right fix, most probably the bug is somewhere in the shuffles analysis for shuffleinstruction build. I'm on vacation, will be able to do more deep analysis in ~2 weeks. Or try to debug yourself, if you have time.

[vporpo]

@abatev could you take a look when you have time?

[ABataev]

Should be fixed already in 5d2cc8e242004161221f8b313279b83f18e1756c

[vporpo]

Ah nice, thanks!

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

if (GatherShuffle) {
           if (isa<PoisonValue>(V)) {
             Mask[I] = PoisonMaskElem;
             continue;
           }
           if (Mask[I] == PoisonMaskElem)
             Mask[I] = Entries.front()->findLaneForValue(V);
         }
         ShuffleBuilder.add(Entries.front()->VectorizedValue, Mask);
-        Res = ShuffleBuilder.finalize(E->ReuseShuffleIndices);
+        Res = ShuffleBuilder.finalize(std::nullopt);
         return Res;
       }
       if (!Resized) {
         unsigned VF1 = Entries.front()->getVectorFactor();
         unsigned VF2 = Entries.back()->getVectorFactor();
         if ((VF == VF1 || VF == VF2) && GatheredScalars.size() != VF)
           GatheredScalars.append(VF - GatheredScalars.size(),
                                  PoisonValue::get(ScalarTy));

llvm/test/Transforms/SLPVectorizer/X86/shufflebuilder-bug.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2
 ; RUN: opt -S -p slp-vectorizer -mtriple=x86_64-- %s | FileCheck %s
 
 define void @foo(<4 x float> %vec, float %val, ptr %ptr) {
 ; CHECK-LABEL: define void @foo
 ; CHECK-SAME: (<4 x float> [[VEC:%.*]], float [[VAL:%.*]], ptr [[PTR:%.*]]) {
 ; CHECK-NEXT:    [[GEP0:%.*]] = getelementptr inbounds float, ptr [[PTR]], i64 0
 ; CHECK-NEXT:    [[TMP1:%.*]] = load <4 x float>, ptr [[GEP0]], align 8
 ; CHECK-NEXT:    [[TMP2:%.*]] = shufflevector <4 x float> [[VEC]], <4 x float> poison, <2 x i32> <i32 3, i32 poison>
 ; CHECK-NEXT:    [[TMP3:%.*]] = insertelement <2 x float> [[TMP2]], float [[VAL]], i32 1
 ; CHECK-NEXT:    [[TMP4:%.*]] = shufflevector <2 x float> [[TMP3]], <2 x float> poison, <4 x i32> <i32 0, i32 0, i32 1, i32 1>
 ; CHECK-NEXT:    [[TMP5:%.*]] = fadd <4 x float> [[TMP1]], [[TMP4]]
-; CHECK-NEXT:    [[TMP6:%.*]] = shufflevector <2 x float> [[TMP3]], <2 x float> poison, <4 x i32> zeroinitializer
-; CHECK-NEXT:    [[TMP7:%.*]] = fmul <4 x float> [[TMP5]], [[TMP6]]
-; CHECK-NEXT:    store <4 x float> [[TMP7]], ptr [[GEP0]], align 4
+; CHECK-NEXT:    [[TMP6:%.*]] = fmul <4 x float> [[TMP5]], [[TMP4]]
+; CHECK-NEXT:    store <4 x float> [[TMP6]], ptr [[GEP0]], align 4
 ; CHECK-NEXT:    ret void
 ;
   %vec_3 = extractelement <4 x float> %vec, i32 3
 
   %gep0 = getelementptr inbounds float, ptr %ptr, i64 0
   %gep1 = getelementptr inbounds float, ptr %ptr, i64 1
   %gep2 = getelementptr inbounds float, ptr %ptr, i64 2
   %gep3 = getelementptr inbounds float, ptr %ptr, i64 3