[Scalarizer] Improve extractelement handling when returned type is vector pointer (PR54469)

[yurai007]

Summary

This patch learn ScalarizerVisitor how to scalarize cases like
load(..., extractelement(...)) and store(..., extractelement(...)) where extractelement returns vector pointer.
We avoid "Inconsistent vector sizes" failure by careful choice of instructions stored in
Scattered and Gathered collections when one of above cases is detected.

[yurai007]

TODO: one more test proving that only loads and stores are relevant?

[Benoit]

Is this related to https://github.com/llvm/llvm-project/issues/54469 ?

[yurai007]

In D123684#3448870, @Benoit wrote:

Is this related to https://github.com/llvm/llvm-project/issues/54469 ?

Yes, this change should fix the crash :)

[yurai007]

More tests fixing TODO.

[yurai007]

isExtractElementReturningVectorPointer update with cosmetic changes.

[lebedev.ri]

This review may be stuck/dead, consider abandoning if no longer relevant.
Removing myself as reviewer in attempt to clean dashboard.

llvm/lib/Transforms/Scalar/Scalarizer.cpp

for (unsigned I = 0; I < NumElems; ++I) {
                                     IEI.getName() + ".i" + Twine(I));
     }
   }
 
   gather(&IEI, Res);
   return true;
 }
 
+static bool isExtractElementReturningVectorPointer(const Value *V,
+                                                   bool Const = false) {
+  const Instruction *VI = dyn_cast<Instruction>(V);
+  if (VI && VI->getOpcode() == Instruction::ExtractElement) {
+    VectorType *VT = dyn_cast<VectorType>(VI->getOperand(0)->getType());
+    Value *ExtIdx = VI->getOperand(1);
+    PointerType *PtrTy = dyn_cast<PointerType>(VT->getElementType());
+    if (PtrTy &&
+        dyn_cast_or_null<FixedVectorType>(PtrTy->getPointerElementType())) {
+      return Const != !dyn_cast<ConstantInt>(ExtIdx);
+    }
+  }
+  return false;
+}
+
 bool ScalarizerVisitor::visitExtractElementInst(ExtractElementInst &EEI) {
   VectorType *VT = dyn_cast<VectorType>(EEI.getOperand(0)->getType());
   if (!VT)
     return false;
 
   unsigned NumSrcElems = cast<FixedVectorType>(VT)->getNumElements();
   IRBuilder<> Builder(&EEI);
   Scatterer Op0 = scatter(&EEI, EEI.getOperand(0));
   Value *ExtIdx = EEI.getOperand(1);
 
   if (auto *CI = dyn_cast<ConstantInt>(ExtIdx)) {
+    // Consider scenario when constant extractelement produce vector pointer.
+    // In such case we cannot scalarize extractelement so
+    // don't bother scattering and gathering it.
+    if (!isExtractElementReturningVectorPointer(&EEI, true)) {
       Value *Res = Op0[CI->getValue().getZExtValue()];
       gather(&EEI, {Res});
+    }
     return true;
   }
 
   if (!ScalarizeVariableInsertExtract)
     return false;
 
   Value *Res = UndefValue::get(VT->getElementType());
   for (unsigned I = 0; I < NumSrcElems; ++I) {
bool ScalarizerVisitor::visitLoadInst(LoadInst &LI) {
 
   Optional<VectorLayout> Layout = getVectorLayout(
       LI.getType(), LI.getAlign(), LI.getModule()->getDataLayout());
   if (!Layout)
     return false;
 
   unsigned NumElems = cast<FixedVectorType>(Layout->VecTy)->getNumElements();
   IRBuilder<> Builder(&LI);
-  Scatterer Ptr = scatter(&LI, LI.getPointerOperand(), LI.getType());
+  Value *Op = LI.getPointerOperand();
+  // Consider scenario when vector load consume vector pointer produced by
+  // variable extractelement. In such case we already scalarized extractelement,
+  // so proper input for scatterer is first item from extractelement's scattered
+  // form.
+  if (isExtractElementReturningVectorPointer(Op))
+    Op = Scattered[Op].front();
+  Scatterer Ptr = scatter(&LI, Op, LI.getType());
   ValueVector Res;
   Res.resize(NumElems);
 
   for (unsigned I = 0; I < NumElems; ++I)
     Res[I] = Builder.CreateAlignedLoad(Layout->VecTy->getElementType(), Ptr[I],
                                        Align(Layout->getElemAlign(I)),
                                        LI.getName() + ".i" + Twine(I));
   gather(&LI, Res);
bool ScalarizerVisitor::visitStoreInst(StoreInst &SI) {
   Value *FullValue = SI.getValueOperand();
   Optional<VectorLayout> Layout = getVectorLayout(
       FullValue->getType(), SI.getAlign(), SI.getModule()->getDataLayout());
   if (!Layout)
     return false;
 
   unsigned NumElems = cast<FixedVectorType>(Layout->VecTy)->getNumElements();
   IRBuilder<> Builder(&SI);
-  Scatterer VPtr = scatter(&SI, SI.getPointerOperand(), FullValue->getType());
+  Value *Op = SI.getPointerOperand();
+  // Consider scenario when vector store consume vector pointer produced by
+  // variable extractelement. In such case we already scalarized extractelement,
+  // so proper input for scatterer is first item from extractelement's scattered
+  // form.
+  if (isExtractElementReturningVectorPointer(Op))
+    Op = Scattered[Op].front();
+  Scatterer VPtr = scatter(&SI, Op, FullValue->getType());
   Scatterer VVal = scatter(&SI, FullValue);
 
   ValueVector Stores;
   Stores.resize(NumElems);
   for (unsigned I = 0; I < NumElems; ++I) {
     Value *Val = VVal[I];
     Value *Ptr = VPtr[I];
     Stores[I] = Builder.CreateAlignedStore(Val, Ptr, Layout->getElemAlign(I));

llvm/test/Transforms/Scalarizer/constant-extractelement.ll

-; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals
 ; RUN: opt %s -passes='function(scalarizer,dce)' -scalarize-load-store -S | FileCheck --check-prefixes=ALL %s
 
 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
 
 ; Test that constant extracts are nicely scalarized
-define i32 @f1(<4 x i32> *%src, i32 %index) {
+define i32 @f1(<4 x i32> *%src) {
 ; ALL-LABEL: @f1(
 ; ALL-NEXT:    [[SRC_I0:%.*]] = bitcast <4 x i32>* [[SRC:%.*]] to i32*
 ; ALL-NEXT:    [[SRC_I3:%.*]] = getelementptr i32, i32* [[SRC_I0]], i32 3
 ; ALL-NEXT:    [[VAL0_I3:%.*]] = load i32, i32* [[SRC_I3]], align 4
-; ALL-NEXT:    [[VAL2:%.*]] = shl i32 4, [[VAL0_I3]]
-; ALL-NEXT:    ret i32 [[VAL2]]
+; ALL-NEXT:    [[VAL1_I3:%.*]] = shl i32 4, [[VAL0_I3]]
+; ALL-NEXT:    ret i32 [[VAL1_I3]]
 ;
   %val0 = load <4 x i32> , <4 x i32> *%src
   %val1 = shl <4 x i32> <i32 1, i32 2, i32 3, i32 4>, %val0
   %val2 = extractelement <4 x i32> %val1, i32 3
   ret i32 %val2
 }
+
+; Test that constant extractelement returning vector pointer is properly scalarized without crash
+define <64 x i32> @f2(<16 x <64 x i32>*> %src) {
+; ALL-LABEL: @f2(
+; ALL-NEXT:    [[TMP1:%.*]] = extractelement <16 x <64 x i32>*> [[SRC:%.*]], i32 3
+; ALL-NEXT:    [[DOTI0:%.*]] = bitcast <64 x i32>* [[TMP1]] to i32*
+; ALL-NEXT:    [[DOTI01:%.*]] = load i32, i32* [[DOTI0]], align 4
+; ALL-NEXT:    [[DOTI1:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 1
+; ALL-NEXT:    [[DOTI12:%.*]] = load i32, i32* [[DOTI1]], align 4
+; ALL-NEXT:    [[DOTI2:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 2
+; ALL-NEXT:    [[DOTI23:%.*]] = load i32, i32* [[DOTI2]], align 4
+; ALL-NEXT:    [[DOTI3:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 3
+; ALL-NEXT:    [[DOTI34:%.*]] = load i32, i32* [[DOTI3]], align 4
+; ALL-NEXT:    [[DOTI4:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 4
+; ALL-NEXT:    [[DOTI45:%.*]] = load i32, i32* [[DOTI4]], align 4
+; ALL-NEXT:    [[DOTI5:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 5
+; ALL-NEXT:    [[DOTI56:%.*]] = load i32, i32* [[DOTI5]], align 4
+; ALL-NEXT:    [[DOTI6:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 6
+; ALL-NEXT:    [[DOTI67:%.*]] = load i32, i32* [[DOTI6]], align 4
+; ALL-NEXT:    [[DOTI7:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 7
+; ALL-NEXT:    [[DOTI78:%.*]] = load i32, i32* [[DOTI7]], align 4
+; ALL-NEXT:    [[DOTI8:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 8
+; ALL-NEXT:    [[DOTI89:%.*]] = load i32, i32* [[DOTI8]], align 4
+; ALL-NEXT:    [[DOTI9:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 9
+; ALL-NEXT:    [[DOTI910:%.*]] = load i32, i32* [[DOTI9]], align 4
+; ALL-NEXT:    [[DOTI10:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 10
+; ALL-NEXT:    [[DOTI1011:%.*]] = load i32, i32* [[DOTI10]], align 4
+; ALL-NEXT:    [[DOTI11:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 11
+; ALL-NEXT:    [[DOTI1112:%.*]] = load i32, i32* [[DOTI11]], align 4
+; ALL-NEXT:    [[DOTI1213:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 12
+; ALL-NEXT:    [[DOTI1214:%.*]] = load i32, i32* [[DOTI1213]], align 4
+; ALL-NEXT:    [[DOTI13:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 13
+; ALL-NEXT:    [[DOTI1315:%.*]] = load i32, i32* [[DOTI13]], align 4
+; ALL-NEXT:    [[DOTI14:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 14
+; ALL-NEXT:    [[DOTI1416:%.*]] = load i32, i32* [[DOTI14]], align 4
+; ALL-NEXT:    [[DOTI15:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 15
+; ALL-NEXT:    [[DOTI1517:%.*]] = load i32, i32* [[DOTI15]], align 4
+; ALL-NEXT:    [[DOTI16:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 16
+; ALL-NEXT:    [[DOTI1618:%.*]] = load i32, i32* [[DOTI16]], align 4
+; ALL-NEXT:    [[DOTI17:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 17
+; ALL-NEXT:    [[DOTI1719:%.*]] = load i32, i32* [[DOTI17]], align 4
+; ALL-NEXT:    [[DOTI18:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 18
+; ALL-NEXT:    [[DOTI1820:%.*]] = load i32, i32* [[DOTI18]], align 4
+; ALL-NEXT:    [[DOTI19:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 19
+; ALL-NEXT:    [[DOTI1921:%.*]] = load i32, i32* [[DOTI19]], align 4
+; ALL-NEXT:    [[DOTI20:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 20
+; ALL-NEXT:    [[DOTI2022:%.*]] = load i32, i32* [[DOTI20]], align 4
+; ALL-NEXT:    [[DOTI21:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 21
+; ALL-NEXT:    [[DOTI2123:%.*]] = load i32, i32* [[DOTI21]], align 4
+; ALL-NEXT:    [[DOTI22:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 22
+; ALL-NEXT:    [[DOTI2224:%.*]] = load i32, i32* [[DOTI22]], align 4
+; ALL-NEXT:    [[DOTI2325:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 23
+; ALL-NEXT:    [[DOTI2326:%.*]] = load i32, i32* [[DOTI2325]], align 4
+; ALL-NEXT:    [[DOTI24:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 24
+; ALL-NEXT:    [[DOTI2427:%.*]] = load i32, i32* [[DOTI24]], align 4
+; ALL-NEXT:    [[DOTI25:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 25
+; ALL-NEXT:    [[DOTI2528:%.*]] = load i32, i32* [[DOTI25]], align 4
+; ALL-NEXT:    [[DOTI26:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 26
+; ALL-NEXT:    [[DOTI2629:%.*]] = load i32, i32* [[DOTI26]], align 4
+; ALL-NEXT:    [[DOTI27:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 27
+; ALL-NEXT:    [[DOTI2730:%.*]] = load i32, i32* [[DOTI27]], align 4
+; ALL-NEXT:    [[DOTI28:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 28
+; ALL-NEXT:    [[DOTI2831:%.*]] = load i32, i32* [[DOTI28]], align 4
+; ALL-NEXT:    [[DOTI29:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 29
+; ALL-NEXT:    [[DOTI2932:%.*]] = load i32, i32* [[DOTI29]], align 4
+; ALL-NEXT:    [[DOTI30:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 30
+; ALL-NEXT:    [[DOTI3033:%.*]] = load i32, i32* [[DOTI30]], align 4
+; ALL-NEXT:    [[DOTI31:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 31
+; ALL-NEXT:    [[DOTI3134:%.*]] = load i32, i32* [[DOTI31]], align 4
+; ALL-NEXT:    [[DOTI32:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 32
+; ALL-NEXT:    [[DOTI3235:%.*]] = load i32, i32* [[DOTI32]], align 4
+; ALL-NEXT:    [[DOTI33:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 33
+; ALL-NEXT:    [[DOTI3336:%.*]] = load i32, i32* [[DOTI33]], align 4
+; ALL-NEXT:    [[DOTI3437:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 34
+; ALL-NEXT:    [[DOTI3438:%.*]] = load i32, i32* [[DOTI3437]], align 4
+; ALL-NEXT:    [[DOTI35:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 35
+; ALL-NEXT:    [[DOTI3539:%.*]] = load i32, i32* [[DOTI35]], align 4
+; ALL-NEXT:    [[DOTI36:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 36
+; ALL-NEXT:    [[DOTI3640:%.*]] = load i32, i32* [[DOTI36]], align 4
+; ALL-NEXT:    [[DOTI37:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 37
+; ALL-NEXT:    [[DOTI3741:%.*]] = load i32, i32* [[DOTI37]], align 4
+; ALL-NEXT:    [[DOTI38:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 38
+; ALL-NEXT:    [[DOTI3842:%.*]] = load i32, i32* [[DOTI38]], align 4
+; ALL-NEXT:    [[DOTI39:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 39
+; ALL-NEXT:    [[DOTI3943:%.*]] = load i32, i32* [[DOTI39]], align 4
+; ALL-NEXT:    [[DOTI40:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 40
+; ALL-NEXT:    [[DOTI4044:%.*]] = load i32, i32* [[DOTI40]], align 4
+; ALL-NEXT:    [[DOTI41:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 41
+; ALL-NEXT:    [[DOTI4145:%.*]] = load i32, i32* [[DOTI41]], align 4
+; ALL-NEXT:    [[DOTI42:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 42
+; ALL-NEXT:    [[DOTI4246:%.*]] = load i32, i32* [[DOTI42]], align 4
+; ALL-NEXT:    [[DOTI43:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 43
+; ALL-NEXT:    [[DOTI4347:%.*]] = load i32, i32* [[DOTI43]], align 4
+; ALL-NEXT:    [[DOTI44:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 44
+; ALL-NEXT:    [[DOTI4448:%.*]] = load i32, i32* [[DOTI44]], align 4
+; ALL-NEXT:    [[DOTI4549:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 45
+; ALL-NEXT:    [[DOTI4550:%.*]] = load i32, i32* [[DOTI4549]], align 4
+; ALL-NEXT:    [[DOTI46:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 46
+; ALL-NEXT:    [[DOTI4651:%.*]] = load i32, i32* [[DOTI46]], align 4
+; ALL-NEXT:    [[DOTI47:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 47
+; ALL-NEXT:    [[DOTI4752:%.*]] = load i32, i32* [[DOTI47]], align 4
+; ALL-NEXT:    [[DOTI48:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 48
+; ALL-NEXT:    [[DOTI4853:%.*]] = load i32, i32* [[DOTI48]], align 4
+; ALL-NEXT:    [[DOTI49:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 49
+; ALL-NEXT:    [[DOTI4954:%.*]] = load i32, i32* [[DOTI49]], align 4
+; ALL-NEXT:    [[DOTI50:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 50
+; ALL-NEXT:    [[DOTI5055:%.*]] = load i32, i32* [[DOTI50]], align 4
+; ALL-NEXT:    [[DOTI51:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 51
+; ALL-NEXT:    [[DOTI5156:%.*]] = load i32, i32* [[DOTI51]], align 4
+; ALL-NEXT:    [[DOTI52:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 52
+; ALL-NEXT:    [[DOTI5257:%.*]] = load i32, i32* [[DOTI52]], align 4
+; ALL-NEXT:    [[DOTI53:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 53
+; ALL-NEXT:    [[DOTI5358:%.*]] = load i32, i32* [[DOTI53]], align 4
+; ALL-NEXT:    [[DOTI54:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 54
+; ALL-NEXT:    [[DOTI5459:%.*]] = load i32, i32* [[DOTI54]], align 4
+; ALL-NEXT:    [[DOTI55:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 55
+; ALL-NEXT:    [[DOTI5560:%.*]] = load i32, i32* [[DOTI55]], align 4
+; ALL-NEXT:    [[DOTI5661:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 56
+; ALL-NEXT:    [[DOTI5662:%.*]] = load i32, i32* [[DOTI5661]], align 4
+; ALL-NEXT:    [[DOTI57:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 57
+; ALL-NEXT:    [[DOTI5763:%.*]] = load i32, i32* [[DOTI57]], align 4
+; ALL-NEXT:    [[DOTI58:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 58
+; ALL-NEXT:    [[DOTI5864:%.*]] = load i32, i32* [[DOTI58]], align 4
+; ALL-NEXT:    [[DOTI59:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 59
+; ALL-NEXT:    [[DOTI5965:%.*]] = load i32, i32* [[DOTI59]], align 4
+; ALL-NEXT:    [[DOTI60:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 60
+; ALL-NEXT:    [[DOTI6066:%.*]] = load i32, i32* [[DOTI60]], align 4
+; ALL-NEXT:    [[DOTI61:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 61
+; ALL-NEXT:    [[DOTI6167:%.*]] = load i32, i32* [[DOTI61]], align 4
+; ALL-NEXT:    [[DOTI62:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 62
+; ALL-NEXT:    [[DOTI6268:%.*]] = load i32, i32* [[DOTI62]], align 4
+; ALL-NEXT:    [[DOTI63:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 63
+; ALL-NEXT:    [[DOTI6369:%.*]] = load i32, i32* [[DOTI63]], align 4
+; ALL-NEXT:    [[DOTUPTO0:%.*]] = insertelement <64 x i32> poison, i32 [[DOTI01]], i32 0
+; ALL-NEXT:    [[DOTUPTO1:%.*]] = insertelement <64 x i32> [[DOTUPTO0]], i32 [[DOTI12]], i32 1
+; ALL-NEXT:    [[DOTUPTO2:%.*]] = insertelement <64 x i32> [[DOTUPTO1]], i32 [[DOTI23]], i32 2
+; ALL-NEXT:    [[DOTUPTO3:%.*]] = insertelement <64 x i32> [[DOTUPTO2]], i32 [[DOTI34]], i32 3
+; ALL-NEXT:    [[DOTUPTO4:%.*]] = insertelement <64 x i32> [[DOTUPTO3]], i32 [[DOTI45]], i32 4
+; ALL-NEXT:    [[DOTUPTO5:%.*]] = insertelement <64 x i32> [[DOTUPTO4]], i32 [[DOTI56]], i32 5
+; ALL-NEXT:    [[DOTUPTO6:%.*]] = insertelement <64 x i32> [[DOTUPTO5]], i32 [[DOTI67]], i32 6
+; ALL-NEXT:    [[DOTUPTO7:%.*]] = insertelement <64 x i32> [[DOTUPTO6]], i32 [[DOTI78]], i32 7
+; ALL-NEXT:    [[DOTUPTO8:%.*]] = insertelement <64 x i32> [[DOTUPTO7]], i32 [[DOTI89]], i32 8
+; ALL-NEXT:    [[DOTUPTO9:%.*]] = insertelement <64 x i32> [[DOTUPTO8]], i32 [[DOTI910]], i32 9
+; ALL-NEXT:    [[DOTUPTO10:%.*]] = insertelement <64 x i32> [[DOTUPTO9]], i32 [[DOTI1011]], i32 10
+; ALL-NEXT:    [[DOTUPTO11:%.*]] = insertelement <64 x i32> [[DOTUPTO10]], i32 [[DOTI1112]], i32 11
+; ALL-NEXT:    [[DOTUPTO12:%.*]] = insertelement <64 x i32> [[DOTUPTO11]], i32 [[DOTI1214]], i32 12
+; ALL-NEXT:    [[DOTUPTO13:%.*]] = insertelement <64 x i32> [[DOTUPTO12]], i32 [[DOTI1315]], i32 13
+; ALL-NEXT:    [[DOTUPTO14:%.*]] = insertelement <64 x i32> [[DOTUPTO13]], i32 [[DOTI1416]], i32 14
+; ALL-NEXT:    [[DOTUPTO15:%.*]] = insertelement <64 x i32> [[DOTUPTO14]], i32 [[DOTI1517]], i32 15
+; ALL-NEXT:    [[DOTUPTO16:%.*]] = insertelement <64 x i32> [[DOTUPTO15]], i32 [[DOTI1618]], i32 16
+; ALL-NEXT:    [[DOTUPTO17:%.*]] = insertelement <64 x i32> [[DOTUPTO16]], i32 [[DOTI1719]], i32 17
+; ALL-NEXT:    [[DOTUPTO18:%.*]] = insertelement <64 x i32> [[DOTUPTO17]], i32 [[DOTI1820]], i32 18
+; ALL-NEXT:    [[DOTUPTO19:%.*]] = insertelement <64 x i32> [[DOTUPTO18]], i32 [[DOTI1921]], i32 19
+; ALL-NEXT:    [[DOTUPTO20:%.*]] = insertelement <64 x i32> [[DOTUPTO19]], i32 [[DOTI2022]], i32 20
+; ALL-NEXT:    [[DOTUPTO21:%.*]] = insertelement <64 x i32> [[DOTUPTO20]], i32 [[DOTI2123]], i32 21
+; ALL-NEXT:    [[DOTUPTO22:%.*]] = insertelement <64 x i32> [[DOTUPTO21]], i32 [[DOTI2224]], i32 22
+; ALL-NEXT:    [[DOTUPTO23:%.*]] = insertelement <64 x i32> [[DOTUPTO22]], i32 [[DOTI2326]], i32 23
+; ALL-NEXT:    [[DOTUPTO24:%.*]] = insertelement <64 x i32> [[DOTUPTO23]], i32 [[DOTI2427]], i32 24
+; ALL-NEXT:    [[DOTUPTO25:%.*]] = insertelement <64 x i32> [[DOTUPTO24]], i32 [[DOTI2528]], i32 25
+; ALL-NEXT:    [[DOTUPTO26:%.*]] = insertelement <64 x i32> [[DOTUPTO25]], i32 [[DOTI2629]], i32 26
+; ALL-NEXT:    [[DOTUPTO27:%.*]] = insertelement <64 x i32> [[DOTUPTO26]], i32 [[DOTI2730]], i32 27
+; ALL-NEXT:    [[DOTUPTO28:%.*]] = insertelement <64 x i32> [[DOTUPTO27]], i32 [[DOTI2831]], i32 28
+; ALL-NEXT:    [[DOTUPTO29:%.*]] = insertelement <64 x i32> [[DOTUPTO28]], i32 [[DOTI2932]], i32 29
+; ALL-NEXT:    [[DOTUPTO30:%.*]] = insertelement <64 x i32> [[DOTUPTO29]], i32 [[DOTI3033]], i32 30
+; ALL-NEXT:    [[DOTUPTO31:%.*]] = insertelement <64 x i32> [[DOTUPTO30]], i32 [[DOTI3134]], i32 31
+; ALL-NEXT:    [[DOTUPTO32:%.*]] = insertelement <64 x i32> [[DOTUPTO31]], i32 [[DOTI3235]], i32 32
+; ALL-NEXT:    [[DOTUPTO33:%.*]] = insertelement <64 x i32> [[DOTUPTO32]], i32 [[DOTI3336]], i32 33
+; ALL-NEXT:    [[DOTUPTO34:%.*]] = insertelement <64 x i32> [[DOTUPTO33]], i32 [[DOTI3438]], i32 34
+; ALL-NEXT:    [[DOTUPTO35:%.*]] = insertelement <64 x i32> [[DOTUPTO34]], i32 [[DOTI3539]], i32 35
+; ALL-NEXT:    [[DOTUPTO36:%.*]] = insertelement <64 x i32> [[DOTUPTO35]], i32 [[DOTI3640]], i32 36
+; ALL-NEXT:    [[DOTUPTO37:%.*]] = insertelement <64 x i32> [[DOTUPTO36]], i32 [[DOTI3741]], i32 37
+; ALL-NEXT:    [[DOTUPTO38:%.*]] = insertelement <64 x i32> [[DOTUPTO37]], i32 [[DOTI3842]], i32 38
+; ALL-NEXT:    [[DOTUPTO39:%.*]] = insertelement <64 x i32> [[DOTUPTO38]], i32 [[DOTI3943]], i32 39
+; ALL-NEXT:    [[DOTUPTO40:%.*]] = insertelement <64 x i32> [[DOTUPTO39]], i32 [[DOTI4044]], i32 40
+; ALL-NEXT:    [[DOTUPTO41:%.*]] = insertelement <64 x i32> [[DOTUPTO40]], i32 [[DOTI4145]], i32 41
+; ALL-NEXT:    [[DOTUPTO42:%.*]] = insertelement <64 x i32> [[DOTUPTO41]], i32 [[DOTI4246]], i32 42
+; ALL-NEXT:    [[DOTUPTO43:%.*]] = insertelement <64 x i32> [[DOTUPTO42]], i32 [[DOTI4347]], i32 43
+; ALL-NEXT:    [[DOTUPTO44:%.*]] = insertelement <64 x i32> [[DOTUPTO43]], i32 [[DOTI4448]], i32 44
+; ALL-NEXT:    [[DOTUPTO45:%.*]] = insertelement <64 x i32> [[DOTUPTO44]], i32 [[DOTI4550]], i32 45
+; ALL-NEXT:    [[DOTUPTO46:%.*]] = insertelement <64 x i32> [[DOTUPTO45]], i32 [[DOTI4651]], i32 46
+; ALL-NEXT:    [[DOTUPTO47:%.*]] = insertelement <64 x i32> [[DOTUPTO46]], i32 [[DOTI4752]], i32 47
+; ALL-NEXT:    [[DOTUPTO48:%.*]] = insertelement <64 x i32> [[DOTUPTO47]], i32 [[DOTI4853]], i32 48
+; ALL-NEXT:    [[DOTUPTO49:%.*]] = insertelement <64 x i32> [[DOTUPTO48]], i32 [[DOTI4954]], i32 49
+; ALL-NEXT:    [[DOTUPTO50:%.*]] = insertelement <64 x i32> [[DOTUPTO49]], i32 [[DOTI5055]], i32 50
+; ALL-NEXT:    [[DOTUPTO51:%.*]] = insertelement <64 x i32> [[DOTUPTO50]], i32 [[DOTI5156]], i32 51
+; ALL-NEXT:    [[DOTUPTO52:%.*]] = insertelement <64 x i32> [[DOTUPTO51]], i32 [[DOTI5257]], i32 52
+; ALL-NEXT:    [[DOTUPTO53:%.*]] = insertelement <64 x i32> [[DOTUPTO52]], i32 [[DOTI5358]], i32 53
+; ALL-NEXT:    [[DOTUPTO54:%.*]] = insertelement <64 x i32> [[DOTUPTO53]], i32 [[DOTI5459]], i32 54
+; ALL-NEXT:    [[DOTUPTO55:%.*]] = insertelement <64 x i32> [[DOTUPTO54]], i32 [[DOTI5560]], i32 55
+; ALL-NEXT:    [[DOTUPTO56:%.*]] = insertelement <64 x i32> [[DOTUPTO55]], i32 [[DOTI5662]], i32 56
+; ALL-NEXT:    [[DOTUPTO57:%.*]] = insertelement <64 x i32> [[DOTUPTO56]], i32 [[DOTI5763]], i32 57
+; ALL-NEXT:    [[DOTUPTO58:%.*]] = insertelement <64 x i32> [[DOTUPTO57]], i32 [[DOTI5864]], i32 58
+; ALL-NEXT:    [[DOTUPTO59:%.*]] = insertelement <64 x i32> [[DOTUPTO58]], i32 [[DOTI5965]], i32 59
+; ALL-NEXT:    [[DOTUPTO60:%.*]] = insertelement <64 x i32> [[DOTUPTO59]], i32 [[DOTI6066]], i32 60
+; ALL-NEXT:    [[DOTUPTO61:%.*]] = insertelement <64 x i32> [[DOTUPTO60]], i32 [[DOTI6167]], i32 61
+; ALL-NEXT:    [[DOTUPTO62:%.*]] = insertelement <64 x i32> [[DOTUPTO61]], i32 [[DOTI6268]], i32 62
+; ALL-NEXT:    [[TMP2:%.*]] = insertelement <64 x i32> [[DOTUPTO62]], i32 [[DOTI6369]], i32 63
+; ALL-NEXT:    ret <64 x i32> [[TMP2]]
+;
+  %1 = extractelement <16 x <64 x i32>*> %src, i32 3
+  %2 = load <64 x i32>, <64 x i32>* %1, align 4
+  ret <64 x i32> %2
+}
+
+; Test that constant extractelement returning vector pointer is properly scalarized without crash
+define <2 x i32> @f3(<4 x <2 x i32>*> %src, <2 x i32> %value) {
+; ALL-LABEL: @f3(
+; ALL-NEXT:    [[VALUE_I0:%.*]] = extractelement <2 x i32> [[VALUE:%.*]], i32 0
+; ALL-NEXT:    [[VALUE_I1:%.*]] = extractelement <2 x i32> [[VALUE]], i32 1
+; ALL-NEXT:    [[TMP1:%.*]] = extractelement <4 x <2 x i32>*> [[SRC:%.*]], i32 3
+; ALL-NEXT:    [[DOTI0:%.*]] = bitcast <2 x i32>* [[TMP1]] to i32*
+; ALL-NEXT:    store i32 [[VALUE_I0]], i32* [[DOTI0]], align 8
+; ALL-NEXT:    [[DOTI1:%.*]] = getelementptr i32, i32* [[DOTI0]], i32 1
+; ALL-NEXT:    store i32 [[VALUE_I1]], i32* [[DOTI1]], align 4
+; ALL-NEXT:    [[DOTI01:%.*]] = load i32, i32* [[DOTI0]], align 4
+; ALL-NEXT:    [[DOTI12:%.*]] = load i32, i32* [[DOTI1]], align 4
+; ALL-NEXT:    [[DOTUPTO0:%.*]] = insertelement <2 x i32> poison, i32 [[DOTI01]], i32 0
+; ALL-NEXT:    [[TMP2:%.*]] = insertelement <2 x i32> [[DOTUPTO0]], i32 [[DOTI12]], i32 1
+; ALL-NEXT:    ret <2 x i32> [[TMP2]]
+;
+  %1 = extractelement <4 x <2 x i32>*> %src, i32 3
+  store <2 x i32> %value, <2 x i32>* %1
+  %2 = load <2 x i32>, <2 x i32>* %1, align 4
+  ret <2 x i32> %2
+}
+
+declare void @external1(<2 x i32*> %v)
+declare void @external2(<4 x <2 x i32*>*> %v)
+
+; Test that vector pointer returned by constant extractelement is properly handled
+; in other instructions
+define <2 x i32*>* @f4(<4 x <2 x i32*>*> %src, <2 x i32*> %value) {
+; ALL-LABEL: @f4(
+; ALL-NEXT:    [[SRC_I1:%.*]] = extractelement <4 x <2 x i32*>*> [[SRC:%.*]], i32 1
+; ALL-NEXT:    [[SRC_I2:%.*]] = extractelement <4 x <2 x i32*>*> [[SRC]], i32 2
+; ALL-NEXT:    [[SRC_I3:%.*]] = extractelement <4 x <2 x i32*>*> [[SRC]], i32 3
+; ALL-NEXT:    [[VALUE_I0:%.*]] = extractelement <2 x i32*> [[VALUE:%.*]], i32 0
+; ALL-NEXT:    [[VALUE_I1:%.*]] = extractelement <2 x i32*> [[VALUE]], i32 1
+; ALL-NEXT:    [[TMP1:%.*]] = extractelement <4 x <2 x i32*>*> [[SRC]], i32 3
+; ALL-NEXT:    [[DOTI0:%.*]] = bitcast <2 x i32*>* [[TMP1]] to i32**
+; ALL-NEXT:    [[DOTI01:%.*]] = load i32*, i32** [[DOTI0]], align 4
+; ALL-NEXT:    [[DOTI1:%.*]] = getelementptr i32*, i32** [[DOTI0]], i32 1
+; ALL-NEXT:    [[DOTI12:%.*]] = load i32*, i32** [[DOTI1]], align 4
+; ALL-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTI12]], align 4
+; ALL-NEXT:    store i32 [[TMP2]], i32* [[DOTI12]], align 4
+; ALL-NEXT:    [[TMP3:%.*]] = getelementptr <2 x i32*>, <2 x i32*>* [[TMP1]], i32 2
+; ALL-NEXT:    [[DOTI03:%.*]] = bitcast <2 x i32*>* [[TMP3]] to i32**
+; ALL-NEXT:    [[DOTI14:%.*]] = getelementptr i32*, i32** [[DOTI03]], i32 1
+; ALL-NEXT:    [[TMP4:%.*]] = bitcast <2 x i32*>* [[TMP3]] to i32**
+; ALL-NEXT:    [[TMP5:%.*]] = load i32*, i32** [[TMP4]], align 8
+; ALL-NEXT:    store i32 [[TMP2]], i32* [[TMP5]], align 4
+; ALL-NEXT:    store i32* [[VALUE_I0]], i32** [[DOTI0]], align 16
+; ALL-NEXT:    store i32* [[VALUE_I1]], i32** [[DOTI1]], align 8
+; ALL-NEXT:    [[TMP6:%.*]] = bitcast <2 x i32*>* [[TMP1]] to i32**
+; ALL-NEXT:    store i32* [[VALUE_I0]], i32** [[DOTI03]], align 16
+; ALL-NEXT:    store i32* [[VALUE_I1]], i32** [[DOTI14]], align 8
+; ALL-NEXT:    [[TMP7:%.*]] = load i32*, i32** [[TMP6]], align 8
+; ALL-NEXT:    [[DOTUPTO05:%.*]] = insertelement <2 x i32*> poison, i32* [[DOTI01]], i32 0
+; ALL-NEXT:    [[TMP8:%.*]] = insertelement <2 x i32*> [[DOTUPTO05]], i32* [[TMP7]], i32 1
+; ALL-NEXT:    call void @external1(<2 x i32*> [[TMP8]])
+; ALL-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP2]], 0
+; ALL-NEXT:    [[TMP10:%.*]] = select i1 [[TMP9]], <2 x i32*>* [[TMP1]], <2 x i32*>* undef
+; ALL-NEXT:    [[DOTUPTO06:%.*]] = insertelement <4 x <2 x i32*>*> poison, <2 x i32*>* [[TMP10]], i32 0
+; ALL-NEXT:    [[DOTUPTO1:%.*]] = insertelement <4 x <2 x i32*>*> [[DOTUPTO06]], <2 x i32*>* [[SRC_I1]], i32 1
+; ALL-NEXT:    [[DOTUPTO2:%.*]] = insertelement <4 x <2 x i32*>*> [[DOTUPTO1]], <2 x i32*>* [[SRC_I2]], i32 2
+; ALL-NEXT:    [[TMP11:%.*]] = insertelement <4 x <2 x i32*>*> [[DOTUPTO2]], <2 x i32*>* [[SRC_I3]], i32 3
+; ALL-NEXT:    call void @external2(<4 x <2 x i32*>*> [[TMP11]])
+; ALL-NEXT:    ret <2 x i32*>* [[TMP1]]
+;
+  %1 = extractelement <4 x <2 x i32*>*> %src, i32 3
+  %2 = load <2 x i32*>, <2 x i32*>* %1, align 4
+  %3 = extractelement <2 x i32*> %2, i32 1
+  %4 = load i32, i32* %3, align 4
+  store i32 %4, i32* %3
+  %5 = getelementptr <2 x i32*>, <2 x i32*>* %1, i32 2
+  %6 = bitcast <2 x i32*>* %5 to i32**
+  %7 = load i32*, i32** %6
+  store i32 %4, i32* %7
+  store <2 x i32*> %value, <2 x i32*>* %1
+  %8 = bitcast <2 x i32*>* %1 to i32**
+  store <2 x i32*> %value, <2 x i32*>* %5
+  %9 = insertelement <4 x <2 x i32*>*> %src, <2 x i32*>* %1, i32 1
+  %10 = load i32*, i32** %8
+  %11 = insertelement <2 x i32*> %2, i32* %10, i32 1
+  call void @external1(<2 x i32*> %11)
+  %12 = icmp eq i32 %4, 0
+  %13 = select i1 %12, <2 x i32*>* %1, <2 x i32*>* undef
+  %14 = insertelement <4 x <2 x i32*>*> %src, <2 x i32*>* %13, i32 0
+  call void @external2(<4 x <2 x i32*>*> %14)
+  ret <2 x i32*>* %1
+}

llvm/test/Transforms/Scalarizer/variable-extractelement.ll

-; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --check-globals
 ; RUN: opt %s -passes='function(scalarizer,dce)' -S | FileCheck --check-prefix=DEFAULT %s
 ; RUN: opt %s -passes='function(scalarizer,dce)' -scalarize-variable-insert-extract=false -S | FileCheck --check-prefix=OFF %s
 ; RUN: opt %s -passes='function(scalarizer,dce)' -scalarize-variable-insert-extract=true -S | FileCheck --check-prefix=DEFAULT %s
 
 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
 
 ; Test that variable extracts scalarized.
 define i32 @f1(<4 x i32> %src, i32 %index) {
 ; DEFAULT-LABEL: @f1(
 ; DEFAULT-NEXT:    [[INDEX_IS_0:%.*]] = icmp eq i32 [[INDEX:%.*]], 0
 ; DEFAULT-NEXT:    [[SRC_I0:%.*]] = extractelement <4 x i32> [[SRC:%.*]], i32 0
 ; DEFAULT-NEXT:    [[RES_UPTO0:%.*]] = select i1 [[INDEX_IS_0]], i32 [[SRC_I0]], i32 undef
 ; DEFAULT-NEXT:    [[INDEX_IS_1:%.*]] = icmp eq i32 [[INDEX]], 1
 ; DEFAULT-NEXT:    [[SRC_I1:%.*]] = extractelement <4 x i32> [[SRC]], i32 1
 ; DEFAULT-NEXT:    [[RES_UPTO1:%.*]] = select i1 [[INDEX_IS_1]], i32 [[SRC_I1]], i32 [[RES_UPTO0]]
 ; DEFAULT-NEXT:    [[INDEX_IS_2:%.*]] = icmp eq i32 [[INDEX]], 2
 ; DEFAULT-NEXT:    [[SRC_I2:%.*]] = extractelement <4 x i32> [[SRC]], i32 2
 ; DEFAULT-NEXT:    [[RES_UPTO2:%.*]] = select i1 [[INDEX_IS_2]], i32 [[SRC_I2]], i32 [[RES_UPTO1]]
 ; DEFAULT-NEXT:    [[INDEX_IS_3:%.*]] = icmp eq i32 [[INDEX]], 3
 ; DEFAULT-NEXT:    [[SRC_I3:%.*]] = extractelement <4 x i32> [[SRC]], i32 3
-; DEFAULT-NEXT:    [[RES:%.*]] = select i1 [[INDEX_IS_3]], i32 [[SRC_I3]], i32 [[RES_UPTO2]]
-; DEFAULT-NEXT:    ret i32 [[RES]]
+; DEFAULT-NEXT:    [[RES_UPTO3:%.*]] = select i1 [[INDEX_IS_3]], i32 [[SRC_I3]], i32 [[RES_UPTO2]]
+; DEFAULT-NEXT:    ret i32 [[RES_UPTO3]]
 ;
 ; OFF-LABEL: @f1(
 ; OFF-NEXT:    [[RES:%.*]] = extractelement <4 x i32> [[SRC:%.*]], i32 [[INDEX:%.*]]
 ; OFF-NEXT:    ret i32 [[RES]]
 ;
   %res = extractelement <4 x i32> %src, i32 %index
   ret i32 %res
 }

 ; DEFAULT-NEXT:    [[VAL1_I3:%.*]] = shl i32 4, [[VAL0_I3]]
 ; DEFAULT-NEXT:    [[INDEX_IS_0:%.*]] = icmp eq i32 [[INDEX:%.*]], 0
 ; DEFAULT-NEXT:    [[VAL2_UPTO0:%.*]] = select i1 [[INDEX_IS_0]], i32 [[VAL1_I0]], i32 undef
 ; DEFAULT-NEXT:    [[INDEX_IS_1:%.*]] = icmp eq i32 [[INDEX]], 1
 ; DEFAULT-NEXT:    [[VAL2_UPTO1:%.*]] = select i1 [[INDEX_IS_1]], i32 [[VAL1_I1]], i32 [[VAL2_UPTO0]]
 ; DEFAULT-NEXT:    [[INDEX_IS_2:%.*]] = icmp eq i32 [[INDEX]], 2
 ; DEFAULT-NEXT:    [[VAL2_UPTO2:%.*]] = select i1 [[INDEX_IS_2]], i32 [[VAL1_I2]], i32 [[VAL2_UPTO1]]
 ; DEFAULT-NEXT:    [[INDEX_IS_3:%.*]] = icmp eq i32 [[INDEX]], 3
-; DEFAULT-NEXT:    [[VAL2:%.*]] = select i1 [[INDEX_IS_3]], i32 [[VAL1_I3]], i32 [[VAL2_UPTO2]]
-; DEFAULT-NEXT:    ret i32 [[VAL2]]
+; DEFAULT-NEXT:    [[VAL2_UPTO3:%.*]] = select i1 [[INDEX_IS_3]], i32 [[VAL1_I3]], i32 [[VAL2_UPTO2]]
+; DEFAULT-NEXT:    ret i32 [[VAL2_UPTO3]]
 ;
 ; OFF-LABEL: @f2(
 ; OFF-NEXT:    [[VAL0:%.*]] = load <4 x i32>, <4 x i32>* [[SRC:%.*]], align 16
 ; OFF-NEXT:    [[VAL0_I0:%.*]] = extractelement <4 x i32> [[VAL0]], i32 0
 ; OFF-NEXT:    [[VAL1_I0:%.*]] = shl i32 1, [[VAL0_I0]]
 ; OFF-NEXT:    [[VAL0_I1:%.*]] = extractelement <4 x i32> [[VAL0]], i32 1
 ; OFF-NEXT:    [[VAL1_I1:%.*]] = shl i32 2, [[VAL0_I1]]
 ; OFF-NEXT:    [[VAL0_I2:%.*]] = extractelement <4 x i32> [[VAL0]], i32 2
 ; OFF-NEXT:    [[VAL1_I2:%.*]] = shl i32 3, [[VAL0_I2]]
 ; OFF-NEXT:    [[VAL0_I3:%.*]] = extractelement <4 x i32> [[VAL0]], i32 3
 ; OFF-NEXT:    [[VAL1_I3:%.*]] = shl i32 4, [[VAL0_I3]]
 ; OFF-NEXT:    [[VAL1_UPTO0:%.*]] = insertelement <4 x i32> poison, i32 [[VAL1_I0]], i32 0
 ; OFF-NEXT:    [[VAL1_UPTO1:%.*]] = insertelement <4 x i32> [[VAL1_UPTO0]], i32 [[VAL1_I1]], i32 1
 ; OFF-NEXT:    [[VAL1_UPTO2:%.*]] = insertelement <4 x i32> [[VAL1_UPTO1]], i32 [[VAL1_I2]], i32 2
 ; OFF-NEXT:    [[VAL1:%.*]] = insertelement <4 x i32> [[VAL1_UPTO2]], i32 [[VAL1_I3]], i32 3
 ; OFF-NEXT:    [[VAL2:%.*]] = extractelement <4 x i32> [[VAL1]], i32 [[INDEX:%.*]]
 ; OFF-NEXT:    ret i32 [[VAL2]]
 ;
   %val0 = load <4 x i32> , <4 x i32> *%src
   %val1 = shl <4 x i32> <i32 1, i32 2, i32 3, i32 4>, %val0
   %val2 = extractelement <4 x i32> %val1, i32 %index
   ret i32 %val2
 }
+
+; PR54469; Test that variable extractelement returning vector pointer is properly scalarized without crash.
+define <64 x i32> @f3(<16 x <64 x i32>*> %src, i32 %index) {
+; DEFAULT-LABEL: @f3(
+; DEFAULT-NEXT:    [[INDEX_IS_0:%.*]] = icmp eq i32 [[INDEX:%.*]], 0
+; DEFAULT-NEXT:    [[SRC_I0:%.*]] = extractelement <16 x <64 x i32>*> [[SRC:%.*]], i32 0
+; DEFAULT-NEXT:    [[DOTUPTO0:%.*]] = select i1 [[INDEX_IS_0]], <64 x i32>* [[SRC_I0]], <64 x i32>* undef
+; DEFAULT-NEXT:    [[INDEX_IS_1:%.*]] = icmp eq i32 [[INDEX]], 1
+; DEFAULT-NEXT:    [[SRC_I1:%.*]] = extractelement <16 x <64 x i32>*> [[SRC]], i32 1
+; DEFAULT-NEXT:    [[DOTUPTO1:%.*]] = select i1 [[INDEX_IS_1]], <64 x i32>* [[SRC_I1]], <64 x i32>* [[DOTUPTO0]]
+; DEFAULT-NEXT:    [[INDEX_IS_2:%.*]] = icmp eq i32 [[INDEX]], 2
+; DEFAULT-NEXT:    [[SRC_I2:%.*]] = extractelement <16 x <64 x i32>*> [[SRC]], i32 2
+; DEFAULT-NEXT:    [[DOTUPTO2:%.*]] = select i1 [[INDEX_IS_2]], <64 x i32>* [[SRC_I2]], <64 x i32>* [[DOTUPTO1]]
+; DEFAULT-NEXT:    [[INDEX_IS_3:%.*]] = icmp eq i32 [[INDEX]], 3
+; DEFAULT-NEXT:    [[SRC_I3:%.*]] = extractelement <16 x <64 x i32>*> [[SRC]], i32 3
+; DEFAULT-NEXT:    [[DOTUPTO3:%.*]] = select i1 [[INDEX_IS_3]], <64 x i32>* [[SRC_I3]], <64 x i32>* [[DOTUPTO2]]
+; DEFAULT-NEXT:    [[INDEX_IS_4:%.*]] = icmp eq i32 [[INDEX]], 4
+; DEFAULT-NEXT:    [[SRC_I4:%.*]] = extractelement <16 x <64 x i32>*> [[SRC]], i32 4
+; DEFAULT-NEXT:    [[DOTUPTO4:%.*]] = select i1 [[INDEX_IS_4]], <64 x i32>* [[SRC_I4]], <64 x i32>* [[DOTUPTO3]]
+; DEFAULT-NEXT:    [[INDEX_IS_5:%.*]] = icmp eq i32 [[INDEX]], 5
+; DEFAULT-NEXT:    [[SRC_I5:%.*]] = extractelement <16 x <64 x i32>*> [[SRC]], i32 5
+; DEFAULT-NEXT:    [[DOTUPTO5:%.*]] = select i1 [[INDEX_IS_5]], <64 x i32>* [[SRC_I5]], <64 x i32>* [[DOTUPTO4]]
+; DEFAULT-NEXT:    [[INDEX_IS_6:%.*]] = icmp eq i32 [[INDEX]], 6
+; DEFAULT-NEXT:    [[SRC_I6:%.*]] = extractelement <16 x <64 x i32>*> [[SRC]], i32 6
+; DEFAULT-NEXT:    [[DOTUPTO6:%.*]] = select i1 [[INDEX_IS_6]], <64 x i32>* [[SRC_I6]], <64 x i32>* [[DOTUPTO5]]
+; DEFAULT-NEXT:    [[INDEX_IS_7:%.*]] = icmp eq i32 [[INDEX]], 7
+; DEFAULT-NEXT:    [[SRC_I7:%.*]] = extractelement <16 x <64 x i32>*> [[SRC]], i32 7
+; DEFAULT-NEXT:    [[DOTUPTO7:%.*]] = select i1 [[INDEX_IS_7]], <64 x i32>* [[SRC_I7]], <64 x i32>* [[DOTUPTO6]]
+; DEFAULT-NEXT:    [[INDEX_IS_8:%.*]] = icmp eq i32 [[INDEX]], 8
+; DEFAULT-NEXT:    [[SRC_I8:%.*]] = extractelement <16 x <64 x i32>*> [[SRC]], i32 8
+; DEFAULT-NEXT:    [[DOTUPTO8:%.*]] = select i1 [[INDEX_IS_8]], <64 x i32>* [[SRC_I8]], <64 x i32>* [[DOTUPTO7]]
+; DEFAULT-NEXT:    [[INDEX_IS_9:%.*]] = icmp eq i32 [[INDEX]], 9
+; DEFAULT-NEXT:    [[SRC_I9:%.*]] = extractelement <16 x <64 x i32>*> [[SRC]], i32 9
+; DEFAULT-NEXT:    [[DOTUPTO9:%.*]] = select i1 [[INDEX_IS_9]], <64 x i32>* [[SRC_I9]], <64 x i32>* [[DOTUPTO8]]
+; DEFAULT-NEXT:    [[INDEX_IS_10:%.*]] = icmp eq i32 [[INDEX]], 10
+; DEFAULT-NEXT:    [[SRC_I10:%.*]] = extractelement <16 x <64 x i32>*> [[SRC]], i32 10
+; DEFAULT-NEXT:    [[DOTUPTO10:%.*]] = select i1 [[INDEX_IS_10]], <64 x i32>* [[SRC_I10]], <64 x i32>* [[DOTUPTO9]]
+; DEFAULT-NEXT:    [[INDEX_IS_11:%.*]] = icmp eq i32 [[INDEX]], 11
+; DEFAULT-NEXT:    [[SRC_I11:%.*]] = extractelement <16 x <64 x i32>*> [[SRC]], i32 11
+; DEFAULT-NEXT:    [[DOTUPTO11:%.*]] = select i1 [[INDEX_IS_11]], <64 x i32>* [[SRC_I11]], <64 x i32>* [[DOTUPTO10]]
+; DEFAULT-NEXT:    [[INDEX_IS_12:%.*]] = icmp eq i32 [[INDEX]], 12
+; DEFAULT-NEXT:    [[SRC_I12:%.*]] = extractelement <16 x <64 x i32>*> [[SRC]], i32 12
+; DEFAULT-NEXT:    [[DOTUPTO12:%.*]] = select i1 [[INDEX_IS_12]], <64 x i32>* [[SRC_I12]], <64 x i32>* [[DOTUPTO11]]
+; DEFAULT-NEXT:    [[INDEX_IS_13:%.*]] = icmp eq i32 [[INDEX]], 13
+; DEFAULT-NEXT:    [[SRC_I13:%.*]] = extractelement <16 x <64 x i32>*> [[SRC]], i32 13
+; DEFAULT-NEXT:    [[DOTUPTO13:%.*]] = select i1 [[INDEX_IS_13]], <64 x i32>* [[SRC_I13]], <64 x i32>* [[DOTUPTO12]]
+; DEFAULT-NEXT:    [[INDEX_IS_14:%.*]] = icmp eq i32 [[INDEX]], 14
+; DEFAULT-NEXT:    [[SRC_I14:%.*]] = extractelement <16 x <64 x i32>*> [[SRC]], i32 14
+; DEFAULT-NEXT:    [[DOTUPTO14:%.*]] = select i1 [[INDEX_IS_14]], <64 x i32>* [[SRC_I14]], <64 x i32>* [[DOTUPTO13]]
+; DEFAULT-NEXT:    [[INDEX_IS_15:%.*]] = icmp eq i32 [[INDEX]], 15
+; DEFAULT-NEXT:    [[SRC_I15:%.*]] = extractelement <16 x <64 x i32>*> [[SRC]], i32 15
+; DEFAULT-NEXT:    [[DOTUPTO15:%.*]] = select i1 [[INDEX_IS_15]], <64 x i32>* [[SRC_I15]], <64 x i32>* [[DOTUPTO14]]
+; DEFAULT-NEXT:    [[TMP1:%.*]] = load <64 x i32>, <64 x i32>* [[DOTUPTO15]], align 4
+; DEFAULT-NEXT:    ret <64 x i32> [[TMP1]]
+;
+; OFF-LABEL: @f3(
+; OFF-NEXT:    [[TMP1:%.*]] = extractelement <16 x <64 x i32>*> [[SRC:%.*]], i32 [[INDEX:%.*]]
+; OFF-NEXT:    [[TMP2:%.*]] = load <64 x i32>, <64 x i32>* [[TMP1]], align 4
+; OFF-NEXT:    ret <64 x i32> [[TMP2]]
+;
+  %1 = extractelement <16 x <64 x i32>*> %src, i32 %index
+  %2 = load <64 x i32>, <64 x i32>* %1, align 4
+  ret <64 x i32> %2
+}
+
+; Test that variable extractelement returning vector pointer is properly scalarized without crash.
+define <2 x i32> @f4(<4 x <2 x i32>*> %src, <2 x i32> %value, i32 %index) {
+; DEFAULT-LABEL: @f4(
+; DEFAULT-NEXT:    [[INDEX_IS_0:%.*]] = icmp eq i32 [[INDEX:%.*]], 0
+; DEFAULT-NEXT:    [[SRC_I0:%.*]] = extractelement <4 x <2 x i32>*> [[SRC:%.*]], i32 0
+; DEFAULT-NEXT:    [[DOTUPTO0:%.*]] = select i1 [[INDEX_IS_0]], <2 x i32>* [[SRC_I0]], <2 x i32>* undef
+; DEFAULT-NEXT:    [[INDEX_IS_1:%.*]] = icmp eq i32 [[INDEX]], 1
+; DEFAULT-NEXT:    [[SRC_I1:%.*]] = extractelement <4 x <2 x i32>*> [[SRC]], i32 1
+; DEFAULT-NEXT:    [[DOTUPTO1:%.*]] = select i1 [[INDEX_IS_1]], <2 x i32>* [[SRC_I1]], <2 x i32>* [[DOTUPTO0]]
+; DEFAULT-NEXT:    [[INDEX_IS_2:%.*]] = icmp eq i32 [[INDEX]], 2
+; DEFAULT-NEXT:    [[SRC_I2:%.*]] = extractelement <4 x <2 x i32>*> [[SRC]], i32 2
+; DEFAULT-NEXT:    [[DOTUPTO2:%.*]] = select i1 [[INDEX_IS_2]], <2 x i32>* [[SRC_I2]], <2 x i32>* [[DOTUPTO1]]
+; DEFAULT-NEXT:    [[INDEX_IS_3:%.*]] = icmp eq i32 [[INDEX]], 3
+; DEFAULT-NEXT:    [[SRC_I3:%.*]] = extractelement <4 x <2 x i32>*> [[SRC]], i32 3
+; DEFAULT-NEXT:    [[DOTUPTO3:%.*]] = select i1 [[INDEX_IS_3]], <2 x i32>* [[SRC_I3]], <2 x i32>* [[DOTUPTO2]]
+; DEFAULT-NEXT:    store <2 x i32> [[VALUE:%.*]], <2 x i32>* [[DOTUPTO3]], align 8
+; DEFAULT-NEXT:    [[TMP1:%.*]] = load <2 x i32>, <2 x i32>* [[DOTUPTO3]], align 4
+; DEFAULT-NEXT:    ret <2 x i32> [[TMP1]]
+;
+; OFF-LABEL: @f4(
+; OFF-NEXT:    [[TMP1:%.*]] = extractelement <4 x <2 x i32>*> [[SRC:%.*]], i32 [[INDEX:%.*]]
+; OFF-NEXT:    store <2 x i32> [[VALUE:%.*]], <2 x i32>* [[TMP1]], align 8
+; OFF-NEXT:    [[TMP2:%.*]] = load <2 x i32>, <2 x i32>* [[TMP1]], align 4
+; OFF-NEXT:    ret <2 x i32> [[TMP2]]
+;
+  %1 = extractelement <4 x <2 x i32>*> %src, i32 %index
+  store <2 x i32> %value, <2 x i32>* %1
+  %2 = load <2 x i32>, <2 x i32>* %1, align 4
+  ret <2 x i32> %2
+}
+
+declare void @external1(<2 x i32*> %v)
+declare void @external2(<4 x <2 x i32*>*> %v)
+
+; Test that vector pointer returned by variable extractelement is properly handled
+; in other instructions.
+define <2 x i32*>* @f5(<4 x <2 x i32*>*> %src, <2 x i32*> %value, i32 %index) {
+; DEFAULT-LABEL: @f5(
+; DEFAULT-NEXT:    [[INDEX_IS_0:%.*]] = icmp eq i32 [[INDEX:%.*]], 0
+; DEFAULT-NEXT:    [[SRC_I0:%.*]] = extractelement <4 x <2 x i32*>*> [[SRC:%.*]], i32 0
+; DEFAULT-NEXT:    [[DOTUPTO0:%.*]] = select i1 [[INDEX_IS_0]], <2 x i32*>* [[SRC_I0]], <2 x i32*>* undef
+; DEFAULT-NEXT:    [[INDEX_IS_1:%.*]] = icmp eq i32 [[INDEX]], 1
+; DEFAULT-NEXT:    [[SRC_I1:%.*]] = extractelement <4 x <2 x i32*>*> [[SRC]], i32 1
+; DEFAULT-NEXT:    [[DOTUPTO1:%.*]] = select i1 [[INDEX_IS_1]], <2 x i32*>* [[SRC_I1]], <2 x i32*>* [[DOTUPTO0]]
+; DEFAULT-NEXT:    [[INDEX_IS_2:%.*]] = icmp eq i32 [[INDEX]], 2
+; DEFAULT-NEXT:    [[SRC_I2:%.*]] = extractelement <4 x <2 x i32*>*> [[SRC]], i32 2
+; DEFAULT-NEXT:    [[DOTUPTO2:%.*]] = select i1 [[INDEX_IS_2]], <2 x i32*>* [[SRC_I2]], <2 x i32*>* [[DOTUPTO1]]
+; DEFAULT-NEXT:    [[INDEX_IS_3:%.*]] = icmp eq i32 [[INDEX]], 3
+; DEFAULT-NEXT:    [[SRC_I3:%.*]] = extractelement <4 x <2 x i32*>*> [[SRC]], i32 3
+; DEFAULT-NEXT:    [[DOTUPTO3:%.*]] = select i1 [[INDEX_IS_3]], <2 x i32*>* [[SRC_I3]], <2 x i32*>* [[DOTUPTO2]]
+; DEFAULT-NEXT:    [[TMP1:%.*]] = load <2 x i32*>, <2 x i32*>* [[DOTUPTO3]], align 4
+; DEFAULT-NEXT:    [[DOTI0:%.*]] = extractelement <2 x i32*> [[TMP1]], i32 0
+; DEFAULT-NEXT:    [[DOTI1:%.*]] = extractelement <2 x i32*> [[TMP1]], i32 1
+; DEFAULT-NEXT:    [[TMP2:%.*]] = load i32, i32* [[DOTI1]], align 4
+; DEFAULT-NEXT:    store i32 [[TMP2]], i32* [[DOTI1]], align 4
+; DEFAULT-NEXT:    [[TMP3:%.*]] = getelementptr <2 x i32*>, <2 x i32*>* [[DOTUPTO3]], i32 2
+; DEFAULT-NEXT:    [[TMP4:%.*]] = bitcast <2 x i32*>* [[TMP3]] to i32**
+; DEFAULT-NEXT:    [[TMP5:%.*]] = load i32*, i32** [[TMP4]], align 8
+; DEFAULT-NEXT:    store i32 [[TMP2]], i32* [[TMP5]], align 4
+; DEFAULT-NEXT:    store <2 x i32*> [[VALUE:%.*]], <2 x i32*>* [[DOTUPTO3]], align 16
+; DEFAULT-NEXT:    [[TMP6:%.*]] = bitcast <2 x i32*>* [[DOTUPTO3]] to i32**
+; DEFAULT-NEXT:    store <2 x i32*> [[VALUE]], <2 x i32*>* [[TMP3]], align 16
+; DEFAULT-NEXT:    [[TMP7:%.*]] = load i32*, i32** [[TMP6]], align 8
+; DEFAULT-NEXT:    [[DOTUPTO01:%.*]] = insertelement <2 x i32*> poison, i32* [[DOTI0]], i32 0
+; DEFAULT-NEXT:    [[TMP8:%.*]] = insertelement <2 x i32*> [[DOTUPTO01]], i32* [[TMP7]], i32 1
+; DEFAULT-NEXT:    call void @external1(<2 x i32*> [[TMP8]])
+; DEFAULT-NEXT:    [[TMP9:%.*]] = icmp eq i32 [[TMP2]], 0
+; DEFAULT-NEXT:    [[TMP10:%.*]] = select i1 [[TMP9]], <2 x i32*>* [[DOTUPTO3]], <2 x i32*>* undef
+; DEFAULT-NEXT:    [[DOTUPTO03:%.*]] = insertelement <4 x <2 x i32*>*> poison, <2 x i32*>* [[TMP10]], i32 0
+; DEFAULT-NEXT:    [[DOTUPTO14:%.*]] = insertelement <4 x <2 x i32*>*> [[DOTUPTO03]], <2 x i32*>* [[SRC_I1]], i32 1
+; DEFAULT-NEXT:    [[DOTUPTO25:%.*]] = insertelement <4 x <2 x i32*>*> [[DOTUPTO14]], <2 x i32*>* [[SRC_I2]], i32 2
+; DEFAULT-NEXT:    [[TMP11:%.*]] = insertelement <4 x <2 x i32*>*> [[DOTUPTO25]], <2 x i32*>* [[SRC_I3]], i32 3
+; DEFAULT-NEXT:    call void @external2(<4 x <2 x i32*>*> [[TMP11]])
+; DEFAULT-NEXT:    ret <2 x i32*>* [[DOTUPTO3]]
+;
+; OFF-LABEL: @f5(
+; OFF-NEXT:    [[SRC_I1:%.*]] = extractelement <4 x <2 x i32*>*> [[SRC:%.*]], i32 1
+; OFF-NEXT:    [[SRC_I2:%.*]] = extractelement <4 x <2 x i32*>*> [[SRC]], i32 2
+; OFF-NEXT:    [[SRC_I3:%.*]] = extractelement <4 x <2 x i32*>*> [[SRC]], i32 3
+; OFF-NEXT:    [[TMP1:%.*]] = extractelement <4 x <2 x i32*>*> [[SRC]], i32 [[INDEX:%.*]]
+; OFF-NEXT:    [[TMP2:%.*]] = load <2 x i32*>, <2 x i32*>* [[TMP1]], align 4
+; OFF-NEXT:    [[DOTI0:%.*]] = extractelement <2 x i32*> [[TMP2]], i32 0
+; OFF-NEXT:    [[DOTI1:%.*]] = extractelement <2 x i32*> [[TMP2]], i32 1
+; OFF-NEXT:    [[TMP3:%.*]] = load i32, i32* [[DOTI1]], align 4
+; OFF-NEXT:    store i32 [[TMP3]], i32* [[DOTI1]], align 4
+; OFF-NEXT:    [[TMP4:%.*]] = getelementptr <2 x i32*>, <2 x i32*>* [[TMP1]], i32 2
+; OFF-NEXT:    [[TMP5:%.*]] = bitcast <2 x i32*>* [[TMP4]] to i32**
+; OFF-NEXT:    [[TMP6:%.*]] = load i32*, i32** [[TMP5]], align 8
+; OFF-NEXT:    store i32 [[TMP3]], i32* [[TMP6]], align 4
+; OFF-NEXT:    store <2 x i32*> [[VALUE:%.*]], <2 x i32*>* [[TMP1]], align 16
+; OFF-NEXT:    [[TMP7:%.*]] = bitcast <2 x i32*>* [[TMP1]] to i32**
+; OFF-NEXT:    store <2 x i32*> [[VALUE]], <2 x i32*>* [[TMP4]], align 16
+; OFF-NEXT:    [[TMP8:%.*]] = load i32*, i32** [[TMP7]], align 8
+; OFF-NEXT:    [[DOTUPTO0:%.*]] = insertelement <2 x i32*> poison, i32* [[DOTI0]], i32 0
+; OFF-NEXT:    [[TMP9:%.*]] = insertelement <2 x i32*> [[DOTUPTO0]], i32* [[TMP8]], i32 1
+; OFF-NEXT:    call void @external1(<2 x i32*> [[TMP9]])
+; OFF-NEXT:    [[TMP10:%.*]] = icmp eq i32 [[TMP3]], 0
+; OFF-NEXT:    [[TMP11:%.*]] = select i1 [[TMP10]], <2 x i32*>* [[TMP1]], <2 x i32*>* undef
+; OFF-NEXT:    [[DOTUPTO01:%.*]] = insertelement <4 x <2 x i32*>*> poison, <2 x i32*>* [[TMP11]], i32 0
+; OFF-NEXT:    [[DOTUPTO1:%.*]] = insertelement <4 x <2 x i32*>*> [[DOTUPTO01]], <2 x i32*>* [[SRC_I1]], i32 1
+; OFF-NEXT:    [[DOTUPTO2:%.*]] = insertelement <4 x <2 x i32*>*> [[DOTUPTO1]], <2 x i32*>* [[SRC_I2]], i32 2
+; OFF-NEXT:    [[TMP12:%.*]] = insertelement <4 x <2 x i32*>*> [[DOTUPTO2]], <2 x i32*>* [[SRC_I3]], i32 3
+; OFF-NEXT:    call void @external2(<4 x <2 x i32*>*> [[TMP12]])
+; OFF-NEXT:    ret <2 x i32*>* [[TMP1]]
+;
+  %1 = extractelement <4 x <2 x i32*>*> %src, i32 %index
+  %2 = load <2 x i32*>, <2 x i32*>* %1, align 4
+  %3 = extractelement <2 x i32*> %2, i32 1
+  %4 = load i32, i32* %3, align 4
+  store i32 %4, i32* %3
+  %5 = getelementptr <2 x i32*>, <2 x i32*>* %1, i32 2
+  %6 = bitcast <2 x i32*>* %5 to i32**
+  %7 = load i32*, i32** %6
+  store i32 %4, i32* %7
+  store <2 x i32*> %value, <2 x i32*>* %1
+  %8 = bitcast <2 x i32*>* %1 to i32**
+  store <2 x i32*> %value, <2 x i32*>* %5
+  %9 = insertelement <4 x <2 x i32*>*> %src, <2 x i32*>* %1, i32 1
+  %10 = load i32*, i32** %8
+  %11 = insertelement <2 x i32*> %2, i32* %10, i32 1
+  call void @external1(<2 x i32*> %11)
+  %12 = icmp eq i32 %4, 0
+  %13 = select i1 %12, <2 x i32*>* %1, <2 x i32*>* undef
+  %14 = insertelement <4 x <2 x i32*>*> %src, <2 x i32*>* %13, i32 0
+  call void @external2(<4 x <2 x i32*>*> %14)
+  ret <2 x i32*>* %1
+}