[SLP]Improve splat detection.

[ABataev]

Summary

A bunch of scalars can be treated as a splat not only if all elements
are the same but also if some of them are undefvalues.

[vporpo]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

257–258

Please update the comment to mention that this also accepts undefs.

260–268

NIT: Perhaps renaming V to something like FirstNonUndef would make this more readable.

[ABataev]

Address comments

[RKSimon]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

261

for-range loop?

271

Should we accept all undef cases? The comment says 'some'.

return FirstNonUndef != nullptr;

[anton-afanasyev]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

2596

Change it or just remove?

[ABataev]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

2596

Why?

[ABataev]

Rebase + address comments.

[RKSimon]

LGTM - cheers

[anton-afanasyev]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

2596

Entry->Scalars[0] can be undef with this patch, for this case the output is uninformative.

[ABataev]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

2596

Hard to tell, this is still the very first element of the entry.

[anton-afanasyev]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

2596

Hmm, but why do we output the very first element specially? If Scalars = { undef, add .., add .., add .. }, we label this as <splat> undef. I would just remove this block at all, since "splat" now doesn't mean what it meant before.

[ABataev]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

2596

Better to output all the scalars here, I think

[anton-afanasyev]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

2596

That's exactly what I mean by removing this if-block which early returns otherwise.

[ABataev]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

2596

So, you suggest to remove << *Entry->Scalars[0]; and ёкуегкт Str;? We can keep splat` label here.

[anton-afanasyev]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

2596

Ok, keep "splat" label, but remove return Str; below.

[ABataev]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

2596

Ok, will do in a separate patch

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

 /// \returns True if all of the values in \p VL are constants (but not
 /// globals/constant expressions).
 static bool allConstant(ArrayRef<Value *> VL) {
   // Constant expressions and globals can't be vectorized like normal integer/FP
   // constants.
   return all_of(VL, isConstant);
 }
 
-/// \returns True if all of the values in \p VL are identical.
+/// \returns True if all of the values in \p VL are identical or some of them
+/// are UndefValue.
 static bool isSplat(ArrayRef<Value *> VL) {
-  for (unsigned i = 1, e = VL.size(); i < e; ++i)
-    if (VL[i] != VL[0])
+  Value *FirstNonUndef = nullptr;
+  for (unsigned I = 0, E = VL.size(); I < E; ++I) {
+    if (isa<UndefValue>(VL[I]))
+      continue;
+    if (!FirstNonUndef) {
+      FirstNonUndef = VL[I];
+      continue;
+    }
+    if (VL[I] != FirstNonUndef)
       return false;
+  }
   return true;
 }
 
 /// \returns True if \p I is commutative, handles CmpInst and BinaryOperator.
 static bool isCommutative(Instruction *I) {
   if (auto *Cmp = dyn_cast<CmpInst>(I))
     return Cmp->isCommutative();
   if (auto *BO = dyn_cast<BinaryOperator>(I))
     return BO->isCommutative();
template <> struct DOTGraphTraits<BoUpSLP *> : public DefaultDOTGraphTraits {
   using TreeEntry = BoUpSLP::TreeEntry;
 
   DOTGraphTraits(bool isSimple = false) : DefaultDOTGraphTraits(isSimple) {}
 
   std::string getNodeLabel(const TreeEntry *Entry, const BoUpSLP *R) {
     std::string Str;
     raw_string_ostream OS(Str);
     if (isSplat(Entry->Scalars)) {
       OS << "<splat> " << *Entry->Scalars[0];
       return Str;
     }
     for (auto V : Entry->Scalars) {
       OS << *V;
       if (llvm::any_of(R->ExternalUses, [&](const BoUpSLP::ExternalUser &EU) {
             return EU.Scalar == V;
           }))
         OS << " <extract>";

llvm/test/Transforms/SLPVectorizer/X86/pr35497.ll

for.body.lr.ph.i:
   ret void
 }
 
 ; Function Attrs: norecurse nounwind uwtable
 define void @pr35497() local_unnamed_addr #0 {
 ; SSE-LABEL: @pr35497(
 ; SSE-NEXT:  entry:
 ; SSE-NEXT:    [[TMP0:%.*]] = load i64, i64* undef, align 1
-; SSE-NEXT:    [[AND:%.*]] = shl i64 [[TMP0]], 2
-; SSE-NEXT:    [[SHL:%.*]] = and i64 [[AND]], 20
 ; SSE-NEXT:    [[ADD:%.*]] = add i64 undef, undef
 ; SSE-NEXT:    store i64 [[ADD]], i64* undef, align 1
 ; SSE-NEXT:    [[ARRAYIDX2_1:%.*]] = getelementptr inbounds [0 x i64], [0 x i64]* undef, i64 0, i64 5
-; SSE-NEXT:    [[AND_1:%.*]] = shl i64 undef, 2
-; SSE-NEXT:    [[SHL_1:%.*]] = and i64 [[AND_1]], 20
-; SSE-NEXT:    [[SHR_1:%.*]] = lshr i64 undef, 6
-; SSE-NEXT:    [[ADD_1:%.*]] = add nuw nsw i64 [[SHL]], [[SHR_1]]
+; SSE-NEXT:    [[TMP1:%.*]] = insertelement <2 x i64> <i64 undef, i64 poison>, i64 [[TMP0]], i32 1
+; SSE-NEXT:    [[TMP2:%.*]] = shl <2 x i64> [[TMP1]], <i64 2, i64 2>
+; SSE-NEXT:    [[TMP3:%.*]] = and <2 x i64> [[TMP2]], <i64 20, i64 20>
 ; SSE-NEXT:    [[ARRAYIDX2_2:%.*]] = getelementptr inbounds [0 x i64], [0 x i64]* undef, i64 0, i64 4
-; SSE-NEXT:    [[SHR_2:%.*]] = lshr i64 undef, 6
-; SSE-NEXT:    [[ADD_2:%.*]] = add nuw nsw i64 [[SHL_1]], [[SHR_2]]
-; SSE-NEXT:    [[AND_4:%.*]] = shl i64 [[ADD]], 2
-; SSE-NEXT:    [[SHL_4:%.*]] = and i64 [[AND_4]], 20
+; SSE-NEXT:    [[TMP4:%.*]] = add nuw nsw <2 x i64> [[TMP3]], zeroinitializer
 ; SSE-NEXT:    [[ARRAYIDX2_5:%.*]] = getelementptr inbounds [0 x i64], [0 x i64]* undef, i64 0, i64 1
-; SSE-NEXT:    store i64 [[ADD_1]], i64* [[ARRAYIDX2_5]], align 1
-; SSE-NEXT:    [[AND_5:%.*]] = shl nuw nsw i64 [[ADD_1]], 2
-; SSE-NEXT:    [[SHL_5:%.*]] = and i64 [[AND_5]], 20
-; SSE-NEXT:    [[SHR_5:%.*]] = lshr i64 [[ADD_1]], 6
-; SSE-NEXT:    [[ADD_5:%.*]] = add nuw nsw i64 [[SHL_4]], [[SHR_5]]
-; SSE-NEXT:    store i64 [[ADD_5]], i64* [[ARRAYIDX2_1]], align 1
+; SSE-NEXT:    [[TMP5:%.*]] = extractelement <2 x i64> [[TMP4]], i32 1
+; SSE-NEXT:    [[TMP6:%.*]] = insertelement <2 x i64> poison, i64 [[TMP5]], i32 0
+; SSE-NEXT:    [[TMP7:%.*]] = insertelement <2 x i64> [[TMP6]], i64 [[ADD]], i32 1
+; SSE-NEXT:    [[TMP8:%.*]] = shl <2 x i64> [[TMP7]], <i64 2, i64 2>
+; SSE-NEXT:    [[TMP9:%.*]] = and <2 x i64> [[TMP8]], <i64 20, i64 20>
 ; SSE-NEXT:    [[ARRAYIDX2_6:%.*]] = getelementptr inbounds [0 x i64], [0 x i64]* undef, i64 0, i64 0
-; SSE-NEXT:    store i64 [[ADD_2]], i64* [[ARRAYIDX2_6]], align 1
-; SSE-NEXT:    [[SHR_6:%.*]] = lshr i64 [[ADD_2]], 6
-; SSE-NEXT:    [[ADD_6:%.*]] = add nuw nsw i64 [[SHL_5]], [[SHR_6]]
-; SSE-NEXT:    store i64 [[ADD_6]], i64* [[ARRAYIDX2_2]], align 1
+; SSE-NEXT:    [[TMP10:%.*]] = bitcast i64* [[ARRAYIDX2_6]] to <2 x i64>*
+; SSE-NEXT:    store <2 x i64> [[TMP4]], <2 x i64>* [[TMP10]], align 1
+; SSE-NEXT:    [[TMP11:%.*]] = lshr <2 x i64> [[TMP4]], <i64 6, i64 6>
+; SSE-NEXT:    [[TMP12:%.*]] = add nuw nsw <2 x i64> [[TMP9]], [[TMP11]]
+; SSE-NEXT:    [[TMP13:%.*]] = bitcast i64* [[ARRAYIDX2_2]] to <2 x i64>*
+; SSE-NEXT:    store <2 x i64> [[TMP12]], <2 x i64>* [[TMP13]], align 1
 ; SSE-NEXT:    ret void
 ;
 ; AVX-LABEL: @pr35497(
 ; AVX-NEXT:  entry:
 ; AVX-NEXT:    [[TMP0:%.*]] = load i64, i64* undef, align 1
 ; AVX-NEXT:    [[ADD:%.*]] = add i64 undef, undef
 ; AVX-NEXT:    store i64 [[ADD]], i64* undef, align 1
 ; AVX-NEXT:    [[ARRAYIDX2_1:%.*]] = getelementptr inbounds [0 x i64], [0 x i64]* undef, i64 0, i64 5