[SLP]Further improvement of the cost model for scalars used in buildvectors.

[ABataev]

Summary

Further improvement of the cost model for the scalars used in buildvectors sequences. The main functionality is outlined into a separate function.

The cost is calculated in the following way:

1. If the Base vector is not undef vector, resizing the very first mask to have common VF and perform action for 2 input vectors (including non-undef Base). Other shuffle masks are combined with the resulting after the 1 stage and processed as a shuffle of 2 elements.
2. If the Base is undef vector and have only 1 shuffle mask, perform the action only for 1 vector with the given mask, if it is not the identity mask.
3. If > 2 masks are used, perform serie of shuffle actions for 2 vectors, combing the masks properly between the steps.

The original implementation misses the very first analysis for the Base vector, so the cost might too optimistic in some cases. But it improves the cost for the insertelements which are part of the current SLP graph.

Part of D107966.

[ABataev]

Rebase

[RKSimon]

Anything we can do to simplify this patch would be great - there's a lot going on.

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

6216

Do you intend to use performExtractsShuffleAction more than once in the future? Otherwise some of these function_ref seem superfluous.

6360

Some of this is NFC - pre-commit to reduce the patch?

6363

IEBase is invariant to the inner do-while loop - should this still use Base?

6382

Pulling out Idx is just a NFC - precommit?

[ABataev]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

6216

Yes, this patch is part of D107966, which unifies cost estimation and code emission.

6363

I'll check

[RKSimon]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

6450

It might make it cleaner if we pull out the lambda instead of embedding inside the performExtractsShuffleAction call?

[ABataev]

Address comments + rebase.

[RKSimon]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

6470

Is this correct - not sure if you want braces here or just a newline

[ABataev]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

6470

Want to dump TreeEntries here, if possible, helps to get more info about the graph, rather than just an instruction.

[ABataev]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

6470

Ah, did not understand the question. Everything is ok here, just the first element might be a nullptr

[RKSimon]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

6470

OK - but maybe put TEs.back()->dump(); on a separate line so it doesn't look like it should be:

if (TEs.front()) { TEs.front()->dump(); TEs.back()->dump(); }

I'm very surprised clang-format didn't catch this

[ABataev]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

6470

Will enclose first dump into braces explicitly, clang-format wants to make them in one line.

[RKSimon]

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

6470

That's scary - I guess because this is all inside the LLVM_DEBUG macro...

[ABataev]

Rebase + formatting

[RKSimon]

LGTM

[mstorsjo]

This caused (or exposed?) failing asserts:

$ cat sqrtf.c
float *a;
float c;
float sqrtf(float);
void b() {
  float d, e, f, g;
  d = c * (0 + 2 * sqrtf(c) * 0);
  e = c * (0 - (c + 1) * 0);
  f = c * (0 - 2 * sqrtf(c) * 0);
  g = 2 * (0 + (c + 1) * 0);
  a[0] = d / 0;
  a[1] = e / 0;
  a[2] = f / 0;
  a[3] = g / 0;
}
$ clang -target x86_64-linux-gnu -c -O2 -fno-math-errno sqrtf.c
clang: ../lib/Transforms/Vectorize/SLPVectorizer.cpp:6592: llvm::InstructionCost llvm::slpvectorizer::BoUpSLP::getTreeCost(llvm::ArrayRef<llvm::Value*>): Assertion `Mask[InIdx] == UndefMaskElem && "InsertElementInstruction used already."' failed.

[ABataev]

In D115750#3499441, @mstorsjo wrote:

This caused (or exposed?) failing asserts:

$ cat sqrtf.c
float *a;
float c;
float sqrtf(float);
void b() {
  float d, e, f, g;
  d = c * (0 + 2 * sqrtf(c) * 0);
  e = c * (0 - (c + 1) * 0);
  f = c * (0 - 2 * sqrtf(c) * 0);
  g = 2 * (0 + (c + 1) * 0);
  a[0] = d / 0;
  a[1] = e / 0;
  a[2] = f / 0;
  a[3] = g / 0;
}
$ clang -target x86_64-linux-gnu -c -O2 -fno-math-errno sqrtf.c
clang: ../lib/Transforms/Vectorize/SLPVectorizer.cpp:6592: llvm::InstructionCost llvm::slpvectorizer::BoUpSLP::getTreeCost(llvm::ArrayRef<llvm::Value*>): Assertion `Mask[InIdx] == UndefMaskElem && "InsertElementInstruction used already."' failed.

Hi, thanks for the report, feel free to revert, I'll fix it tomorrow.

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

6470

Yes, I believe so.

[RKSimon]

repro:

; ModuleID = 'bugpoint-reduced-simplified.bc'
source_filename = "fuzz.ll"
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

define dso_local void @b() local_unnamed_addr #0 {
entry:
  %mul = fmul float undef, 2.000000e+00
  %i = tail call float @llvm.fmuladd.f32(float %mul, float 0.000000e+00, float 0.000000e+00)
  %mul2 = fmul float undef, %i
  %add = fadd float undef, 1.000000e+00
  %neg = fneg float %add
  %i1 = tail call float @llvm.fmuladd.f32(float %neg, float 0.000000e+00, float 0.000000e+00)
  %mul4 = fmul float undef, %i1
  %neg7 = fneg float %mul
  %i2 = tail call float @llvm.fmuladd.f32(float %neg7, float 0.000000e+00, float 0.000000e+00)
  %mul8 = fmul float undef, %i2
  %i3 = tail call float @llvm.fmuladd.f32(float %add, float 0.000000e+00, float 0.000000e+00)
  %mul11 = fmul float %i3, 2.000000e+00
  %div = fdiv float %mul2, 0.000000e+00
  store float %div, ptr undef, align 4
  %div12 = fdiv float %mul4, 0.000000e+00
  %arrayidx13 = getelementptr inbounds float, ptr undef, i64 1
  store float %div12, ptr %arrayidx13, align 4
  %div14 = fdiv float %mul8, 0.000000e+00
  %arrayidx15 = getelementptr inbounds float, ptr undef, i64 2
  store float %div14, ptr %arrayidx15, align 4
  %div16 = fdiv float %mul11, 0.000000e+00
  %arrayidx17 = getelementptr inbounds float, ptr undef, i64 3
  store float %div16, ptr %arrayidx17, align 4
  ret void
}

; Function Attrs: nocallback nofree nosync nounwind readnone speculatable willreturn
declare float @llvm.fmuladd.f32(float, float, float) #1

attributes #0 = { "tune-cpu"="generic" }
attributes #1 = { nocallback nofree nosync nounwind readnone speculatable willreturn }

[ronlieb]

i am seeing the same assert on the amdgpu target

the reproducer posted by previous two responders is way smaller than what i have

[ABataev]

Must be fixed in 9c3a75eabf577f0e0e372be95ec4861600a5acdb

[sberg]

In D115750#3500827, @ABataev wrote:

Must be fixed in 9c3a75eabf577f0e0e372be95ec4861600a5acdb

still fails for me, see https://github.com/llvm/llvm-project/issues/55359 "'llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp:6596: llvm::InstructionCost llvm::slpvectorizer::BoUpSLP::getTreeCost(ArrayRef<llvm::Value *>): Assertion `Mask[InIdx] == UndefMaskElem && "InsertElementInstruction used already."' failed.' since https://github.com/llvm/llvm-project/commit/99f31acfce338417fea3c14983d6f8fedc8ed043 '[SLP]Further improvement of the cost model for scalars used in buildvectors.'" (which I filed so that I could attach the non-reduced reproducer there)

[aeubanks]

another issue that repros at head

$ cat /tmp/a.ll
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

define internal void @0() {
  br label %1

1:                                                ; preds = %1, %0
  %2 = fadd float 0.000000e+00, 0.000000e+00
  %3 = fadd float 0.000000e+00, 0.000000e+00
  %4 = insertelement <2 x float> zeroinitializer, float %2, i64 0
  %5 = insertelement <2 x float> %4, float %3, i64 1
  %6 = insertelement <2 x float> %5, float %2, i64 0
  br label %1
}
$ ./build/rel/bin/opt -passes=slp-vectorizer -disable-output /tmp/a.ll
opt: ../../llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp:6500: llvm::InstructionCost llvm::slpvectorizer::BoUpSLP::getTreeCost(ArrayRef<llvm::Value *>): Assertion `Mask[InIdx] == UndefMaskElem && "InsertElementInstruction used already."' failed.

(I did see a different instance of this crash get fixed with some other commit but this one still remains)

[ABataev]

In D115750#3501807, @aeubanks wrote:

another issue that repros at head

$ cat /tmp/a.ll
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

define internal void @0() {
  br label %1

1:                                                ; preds = %1, %0
  %2 = fadd float 0.000000e+00, 0.000000e+00
  %3 = fadd float 0.000000e+00, 0.000000e+00
  %4 = insertelement <2 x float> zeroinitializer, float %2, i64 0
  %5 = insertelement <2 x float> %4, float %3, i64 1
  %6 = insertelement <2 x float> %5, float %2, i64 0
  br label %1
}
$ ./build/rel/bin/opt -passes=slp-vectorizer -disable-output /tmp/a.ll
opt: ../../llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp:6500: llvm::InstructionCost llvm::slpvectorizer::BoUpSLP::getTreeCost(ArrayRef<llvm::Value *>): Assertion `Mask[InIdx] == UndefMaskElem && "InsertElementInstruction used already."' failed.

(I did see a different instance of this crash get fixed with some other commit but this one still remains)

Yes, have a fix for this issue already, need to relax an assert, will commit in couple minutes

[ABataev]

Decided to revert this patch and several others to fix the bugs.

[aeubanks]

another crash even with https://reviews.llvm.org/rGcce80bd8b74d54deb82b1b6ae0cbec1ab53c1dbb

target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

define internal void @0() {
.lr.ph.i41:
  br label %.lr.ph.split.us.i

.lr.ph.split.us.i:                                ; preds = %.lr.ph.split.us.i, %.lr.ph.i41
  %0 = fadd float 0.000000e+00, 0.000000e+00
  %1 = fadd float 0.000000e+00, 0.000000e+00
  %2 = fadd float %0, 0.000000e+00
  %3 = fadd float %1, 0.000000e+00
  %.sroa.3.8.vec.insert.i.us.i = insertelement <2 x float> zeroinitializer, float %2, i64 0
  %.sroa.3.12.vec.insert.i.us.i = insertelement <2 x float> %.sroa.3.8.vec.insert.i.us.i, float %3, i64 1
  %.sroa.025.4.vec.insert.us.i = insertelement <2 x float> %.sroa.3.12.vec.insert.i.us.i, float %0, i64 0
  br label %.lr.ph.split.us.i
}

[ABataev]

In D115750#3501860, @aeubanks wrote:

another crash even with https://reviews.llvm.org/rGcce80bd8b74d54deb82b1b6ae0cbec1ab53c1dbb

target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

define internal void @0() {
.lr.ph.i41:
  br label %.lr.ph.split.us.i

.lr.ph.split.us.i:                                ; preds = %.lr.ph.split.us.i, %.lr.ph.i41
  %0 = fadd float 0.000000e+00, 0.000000e+00
  %1 = fadd float 0.000000e+00, 0.000000e+00
  %2 = fadd float %0, 0.000000e+00
  %3 = fadd float %1, 0.000000e+00
  %.sroa.3.8.vec.insert.i.us.i = insertelement <2 x float> zeroinitializer, float %2, i64 0
  %.sroa.3.12.vec.insert.i.us.i = insertelement <2 x float> %.sroa.3.8.vec.insert.i.us.i, float %3, i64 1
  %.sroa.025.4.vec.insert.us.i = insertelement <2 x float> %.sroa.3.12.vec.insert.i.us.i, float %0, i64 0
  br label %.lr.ph.split.us.i
}

Thanks, will add it to fixed version of the patch

[aeubanks]

(if you're interested in testing the recommit, consider running opt -O2 on the attached file, that's where I've been pulling the crashes from)

b.ll.txt2 MBDownload

[aeubanks]

Another crash caused by this:

$ cat /tmp/f.ll
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-android23"

%0 = type { float, float, float, float }

; Function Attrs: argmemonly nocallback nofree nosync nounwind willreturn
declare void @llvm.lifetime.start.p0(i64 immarg, ptr nocapture) #0

; Function Attrs: argmemonly nocallback nofree nosync nounwind willreturn
declare void @llvm.lifetime.end.p0(i64 immarg, ptr nocapture) #0

; Function Attrs: argmemonly nofree nounwind willreturn writeonly
declare void @llvm.memset.p0.i64(ptr nocapture writeonly, i8, i64, i1 immarg) #1

define internal void @0() {
  %1 = getelementptr inbounds %0, ptr undef, i64 0, i32 2
  %2 = load float, ptr %1, align 4
  %3 = load float, ptr undef, align 4
  %4 = fsub float %2, %3
  %5 = getelementptr inbounds %0, ptr undef, i64 0, i32 3
  %6 = load float, ptr %5, align 4
  %7 = getelementptr inbounds %0, ptr undef, i64 0, i32 1
  %8 = load float, ptr %7, align 4
  %9 = fsub float %6, %8
  %10 = fcmp olt float %9, %4
  %11 = insertelement <2 x float> undef, float %3, i64 0
  %12 = insertelement <2 x float> zeroinitializer, float 0.000000e+00, i64 0
  store <2 x float> zeroinitializer, ptr null, align 4
  %13 = insertelement <2 x float> %11, float %6, i64 0
  store <2 x float> zeroinitializer, ptr null, align 4
  ret void
}

attributes #0 = { argmemonly nocallback nofree nosync nounwind willreturn }
attributes #1 = { argmemonly nofree nounwind willreturn writeonly }
$ bin/opt -passes=slp-vectorizer -disable-output /tmp/f.ll
opt: ../../llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp:6657: T *performExtractsShuffleAction(MutableArrayRef<std::pair<T *, SmallVector<int>>>, llvm::Value *, function_ref<unsigned int (T *)>
, function_ref<std::pair<T *, bool> (T *, ArrayRef<int>)>, function_ref<T *(ArrayRef<int>, ArrayRef<T *>)>) [T = const llvm::slpvectorizer::BoUpSLP::TreeEntry]: Assertion `Mask[I] == UndefMaskEle
m && "Multiple uses of scalars."' failed.

[ABataev]

In D115750#3511967, @aeubanks wrote:

Another crash caused by this:

$ cat /tmp/f.ll
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-android23"

%0 = type { float, float, float, float }

; Function Attrs: argmemonly nocallback nofree nosync nounwind willreturn
declare void @llvm.lifetime.start.p0(i64 immarg, ptr nocapture) #0

; Function Attrs: argmemonly nocallback nofree nosync nounwind willreturn
declare void @llvm.lifetime.end.p0(i64 immarg, ptr nocapture) #0

; Function Attrs: argmemonly nofree nounwind willreturn writeonly
declare void @llvm.memset.p0.i64(ptr nocapture writeonly, i8, i64, i1 immarg) #1

define internal void @0() {
  %1 = getelementptr inbounds %0, ptr undef, i64 0, i32 2
  %2 = load float, ptr %1, align 4
  %3 = load float, ptr undef, align 4
  %4 = fsub float %2, %3
  %5 = getelementptr inbounds %0, ptr undef, i64 0, i32 3
  %6 = load float, ptr %5, align 4
  %7 = getelementptr inbounds %0, ptr undef, i64 0, i32 1
  %8 = load float, ptr %7, align 4
  %9 = fsub float %6, %8
  %10 = fcmp olt float %9, %4
  %11 = insertelement <2 x float> undef, float %3, i64 0
  %12 = insertelement <2 x float> zeroinitializer, float 0.000000e+00, i64 0
  store <2 x float> zeroinitializer, ptr null, align 4
  %13 = insertelement <2 x float> %11, float %6, i64 0
  store <2 x float> zeroinitializer, ptr null, align 4
  ret void
}

attributes #0 = { argmemonly nocallback nofree nosync nounwind willreturn }
attributes #1 = { argmemonly nofree nounwind willreturn writeonly }
$ bin/opt -passes=slp-vectorizer -disable-output /tmp/f.ll
opt: ../../llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp:6657: T *performExtractsShuffleAction(MutableArrayRef<std::pair<T *, SmallVector<int>>>, llvm::Value *, function_ref<unsigned int (T *)>
, function_ref<std::pair<T *, bool> (T *, ArrayRef<int>)>, function_ref<T *(ArrayRef<int>, ArrayRef<T *>)>) [T = const llvm::slpvectorizer::BoUpSLP::TreeEntry]: Assertion `Mask[I] == UndefMaskEle
m && "Multiple uses of scalars."' failed.

Another corner case of too strict assertion I was afraid of. Will fix it ASAP.

[ABataev]

In D115750#3511967, @aeubanks wrote:

Another crash caused by this:

$ cat /tmp/f.ll
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-android23"

%0 = type { float, float, float, float }

; Function Attrs: argmemonly nocallback nofree nosync nounwind willreturn
declare void @llvm.lifetime.start.p0(i64 immarg, ptr nocapture) #0

; Function Attrs: argmemonly nocallback nofree nosync nounwind willreturn
declare void @llvm.lifetime.end.p0(i64 immarg, ptr nocapture) #0

; Function Attrs: argmemonly nofree nounwind willreturn writeonly
declare void @llvm.memset.p0.i64(ptr nocapture writeonly, i8, i64, i1 immarg) #1

define internal void @0() {
  %1 = getelementptr inbounds %0, ptr undef, i64 0, i32 2
  %2 = load float, ptr %1, align 4
  %3 = load float, ptr undef, align 4
  %4 = fsub float %2, %3
  %5 = getelementptr inbounds %0, ptr undef, i64 0, i32 3
  %6 = load float, ptr %5, align 4
  %7 = getelementptr inbounds %0, ptr undef, i64 0, i32 1
  %8 = load float, ptr %7, align 4
  %9 = fsub float %6, %8
  %10 = fcmp olt float %9, %4
  %11 = insertelement <2 x float> undef, float %3, i64 0
  %12 = insertelement <2 x float> zeroinitializer, float 0.000000e+00, i64 0
  store <2 x float> zeroinitializer, ptr null, align 4
  %13 = insertelement <2 x float> %11, float %6, i64 0
  store <2 x float> zeroinitializer, ptr null, align 4
  ret void
}

attributes #0 = { argmemonly nocallback nofree nosync nounwind willreturn }
attributes #1 = { argmemonly nofree nounwind willreturn writeonly }
$ bin/opt -passes=slp-vectorizer -disable-output /tmp/f.ll
opt: ../../llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp:6657: T *performExtractsShuffleAction(MutableArrayRef<std::pair<T *, SmallVector<int>>>, llvm::Value *, function_ref<unsigned int (T *)>
, function_ref<std::pair<T *, bool> (T *, ArrayRef<int>)>, function_ref<T *(ArrayRef<int>, ArrayRef<T *>)>) [T = const llvm::slpvectorizer::BoUpSLP::TreeEntry]: Assertion `Mask[I] == UndefMaskEle
m && "Multiple uses of scalars."' failed.

Fixed 85f6b15ee50f

[aeubanks]

another one :)

target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-android23"

%0 = type { float, float, float, float }

define internal void @0() {
  %1 = getelementptr inbounds %0, ptr undef, i64 0, i32 2
  %2 = load float, ptr %1, align 4
  %3 = load float, ptr undef, align 4
  %4 = fsub float %2, %3
  %5 = getelementptr inbounds %0, ptr undef, i64 0, i32 3
  %6 = load float, ptr %5, align 4
  %7 = getelementptr inbounds %0, ptr undef, i64 0, i32 1
  %8 = load float, ptr %7, align 4
  %9 = fsub float %6, %8
  %10 = fcmp olt float %9, %4
  %.sroa.0.0.vec.insert.i5.i10 = insertelement <2 x float> undef, float %3, i64 0
  %.sroa.0.4.vec.insert.i10.i13 = insertelement <2 x float> %.sroa.0.0.vec.insert.i5.i10, float %8, i64 1
  store <2 x float> %.sroa.0.4.vec.insert.i10.i13, ptr null, align 4
  %.sroa.0.4.vec.insert.i10.i13.2 = insertelement <2 x float> %.sroa.0.0.vec.insert.i5.i10, float %6, i64 1
  store <2 x float> %.sroa.0.4.vec.insert.i10.i13.2, ptr null, align 4
  ret void
}

[ABataev]

In D115750#3513649, @aeubanks wrote:

another one :)

target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-android23"

%0 = type { float, float, float, float }

define internal void @0() {
  %1 = getelementptr inbounds %0, ptr undef, i64 0, i32 2
  %2 = load float, ptr %1, align 4
  %3 = load float, ptr undef, align 4
  %4 = fsub float %2, %3
  %5 = getelementptr inbounds %0, ptr undef, i64 0, i32 3
  %6 = load float, ptr %5, align 4
  %7 = getelementptr inbounds %0, ptr undef, i64 0, i32 1
  %8 = load float, ptr %7, align 4
  %9 = fsub float %6, %8
  %10 = fcmp olt float %9, %4
  %.sroa.0.0.vec.insert.i5.i10 = insertelement <2 x float> undef, float %3, i64 0
  %.sroa.0.4.vec.insert.i10.i13 = insertelement <2 x float> %.sroa.0.0.vec.insert.i5.i10, float %8, i64 1
  store <2 x float> %.sroa.0.4.vec.insert.i10.i13, ptr null, align 4
  %.sroa.0.4.vec.insert.i10.i13.2 = insertelement <2 x float> %.sroa.0.0.vec.insert.i5.i10, float %6, i64 1
  store <2 x float> %.sroa.0.4.vec.insert.i10.i13.2, ptr null, align 4
  ret void
}

Fixed in 152072801e24fb1e5cd962b0cb089230bc27b6b9

llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp

if (IE2) {
         IE2 = nullptr;
       else
         IE2 = dyn_cast<InsertElementInst>(IE2->getOperand(0));
     }
   } while (IE1 || IE2);
   return false;
 }
 
+/// Checks if the \p IE1 instructions is followed by \p IE2 instruction in the
+/// buildvector sequence.
+static bool isFirstInsertElement(const InsertElementInst *IE1,
+                                 const InsertElementInst *IE2) {
+  const auto *I1 = IE1;
+  const auto *I2 = IE2;
+  do {
+    if (I2 == IE1)
+      return true;
+    if (I1 == IE2)
+      return false;
+    if (I1)
+      I1 = dyn_cast<InsertElementInst>(I1->getOperand(0));
+    if (I2)
+      I2 = dyn_cast<InsertElementInst>(I2->getOperand(0));
+  } while (I1 || I2);
+  llvm_unreachable("Two different buildvectors not expected.");
+}
+
+/// Does the analysis of the provided shuffle masks and performs the requested
+/// actions on the vectors with the given shuffle masks. It tries to do it in
+/// several steps.
+/// 1. If the Base vector is not undef vector, resizing the very first mask to
+/// have common VF and perform action for 2 input vectors (including non-undef
+/// Base). Other shuffle masks are combined with the resulting after the 1 stage
+/// and processed as a shuffle of 2 elements.
+/// 2. If the Base is undef vector and have only 1 shuffle mask, perform the
+/// action only for 1 vector with the given mask, if it is not the identity
+/// mask.
+/// 3. If > 2 masks are used, perform the remaining shuffle actions for 2
+/// vectors, combing the masks properly between the steps.
+template <typename T>
+static T *performExtractsShuffleAction(
+    MutableArrayRef<std::pair<T *, SmallVector<int>>> ShuffleMask, Value *Base,
+    function_ref<unsigned(T *)> GetVF,
+    function_ref<std::pair<T *, bool>(T *, ArrayRef<int>)> ResizeAction,
+    function_ref<T *(ArrayRef<int>, ArrayRef<T *>)> Action) {
+  assert(!ShuffleMask.empty() && "Empty list of shuffles for inserts.");
+  SmallVector<int> Mask(ShuffleMask.begin()->second);
+  auto VMIt = std::next(ShuffleMask.begin());
+  T *Prev = nullptr;
+  bool IsBaseNotUndef = !isUndefVector(Base);
+  if (IsBaseNotUndef) {
+    // Base is not undef, need to combine it with the next subvectors.
+    std::pair<T *, bool> Res = ResizeAction(ShuffleMask.begin()->first, Mask);
+    for (unsigned Idx = 0, VF = Mask.size(); Idx < VF; ++Idx) {
+      if (Mask[Idx] == UndefMaskElem)
+        Mask[Idx] = Idx;
+      else
+        Mask[Idx] = (Res.second ? Idx : Mask[Idx]) + VF;
+    }
+    Prev = Action(Mask, {nullptr, Res.first});
+  } else if (ShuffleMask.size() == 1) {
+    // Base is undef and only 1 vector is shuffled - perform the action only for
+    // single vector, if the mask is not the identity mask.
+    std::pair<T *, bool> Res = ResizeAction(ShuffleMask.begin()->first, Mask);
+    if (Res.second)
+      // Identity mask is found.
+      Prev = Res.first;
+    else
+      Prev = Action(Mask, {ShuffleMask.begin()->first});
+  } else {
+    // Base is undef and at least 2 input vectors shuffled - perform 2 vectors
+    // shuffles step by step, combining shuffle between the steps.
+    unsigned Vec1VF = GetVF(ShuffleMask.begin()->first);
+    unsigned Vec2VF = GetVF(VMIt->first);
+    if (Vec1VF == Vec2VF) {
+      // No need to resize the input vectors since they are of the same size, we
+      // can shuffle them directly.
+      ArrayRef<int> SecMask = VMIt->second;
+      for (unsigned I = 0, VF = Mask.size(); I < VF; ++I) {
+        if (SecMask[I] != UndefMaskElem) {
+          assert(Mask[I] == UndefMaskElem && "Multiple uses of scalars.");
+          Mask[I] = SecMask[I] + Vec1VF;
+        }
+      }
+      Prev = Action(Mask, {ShuffleMask.begin()->first, VMIt->first});
+    } else {
+      // Vectors of different sizes - resize and reshuffle.
+      std::pair<T *, bool> Res1 =
+          ResizeAction(ShuffleMask.begin()->first, Mask);
+      std::pair<T *, bool> Res2 = ResizeAction(VMIt->first, VMIt->second);
+      ArrayRef<int> SecMask = VMIt->second;
+      for (unsigned I = 0, VF = Mask.size(); I < VF; ++I) {
+        if (Mask[I] != UndefMaskElem) {
+          assert(SecMask[I] == UndefMaskElem && "Multiple uses of scalars.");
+          if (Res1.second)
+            Mask[I] = I;
+        } else if (SecMask[I] != UndefMaskElem) {
+          assert(Mask[I] == UndefMaskElem && "Multiple uses of scalars.");
+          Mask[I] = (Res2.second ? I : SecMask[I]) + VF;
+        }
+      }
+      Prev = Action(Mask, {Res1.first, Res2.first});
+    }
+    VMIt = std::next(VMIt);
+  }
+  // Perform requested actions for the remaining masks/vectors.
+  for (auto E = ShuffleMask.end(); VMIt != E; ++VMIt) {
+    // Shuffle other input vectors, if any.
+    std::pair<T *, bool> Res = ResizeAction(VMIt->first, VMIt->second);
+    ArrayRef<int> SecMask = VMIt->second;
+    for (unsigned I = 0, VF = Mask.size(); I < VF; ++I) {
+      if (SecMask[I] != UndefMaskElem) {
+        assert((Mask[I] == UndefMaskElem || IsBaseNotUndef) &&
+               "Multiple uses of scalars.");
+        Mask[I] = (Res.second ? I : SecMask[I]) + VF;
+      } else if (Mask[I] != UndefMaskElem) {
+        Mask[I] = I;
+      }
+    }
+    Prev = Action(Mask, {Prev, Res.first});
+  }
+  return Prev;
+}
+
 InstructionCost BoUpSLP::getTreeCost(ArrayRef<Value *> VectorizedVals) {
   InstructionCost Cost = 0;
   LLVM_DEBUG(dbgs() << "SLP: Calculating cost for tree of size "
                     << VectorizableTree.size() << ".\n");
 
   unsigned BundleWidth = VectorizableTree[0]->Scalars.size();
 
   for (unsigned I = 0, E = VectorizableTree.size(); I < E; ++I) {
     TreeEntry &TE = *VectorizableTree[I];
 
     InstructionCost C = getEntryCost(&TE, VectorizedVals);
     Cost += C;
     LLVM_DEBUG(dbgs() << "SLP: Adding cost " << C
                       << " for bundle that starts with " << *TE.Scalars[0]
                       << ".\n"
                       << "SLP: Current total cost = " << Cost << "\n");
   }
 
   SmallPtrSet<Value *, 16> ExtractCostCalculated;
   InstructionCost ExtractCost = 0;
-  SmallVector<unsigned> VF;
-  SmallVector<SmallVector<int>> ShuffleMask;
-  SmallVector<Value *> FirstUsers;
+  SmallVector<MapVector<const TreeEntry *, SmallVector<int>>> ShuffleMasks;
+  SmallVector<std::pair<Value *, const TreeEntry *>> FirstUsers;
   SmallVector<APInt> DemandedElts;
   for (ExternalUser &EU : ExternalUses) {
     // We only add extract cost once for the same scalar.
     if (!isa_and_nonnull<InsertElementInst>(EU.User) &&
         !ExtractCostCalculated.insert(EU.Scalar).second)
       continue;
 
     // Uses by ephemeral values are free (because the ephemeral value will be
if (isa<ExtractElementInst>(EU.Scalar))
       continue;
 
     // If found user is an insertelement, do not calculate extract cost but try
     // to detect it as a final shuffled/identity match.
     if (auto *VU = dyn_cast_or_null<InsertElementInst>(EU.User)) {
       if (auto *FTy = dyn_cast<FixedVectorType>(VU->getType())) {
         Optional<unsigned> InsertIdx = getInsertIndex(VU);
         if (InsertIdx) {
-          auto *It = find_if(FirstUsers, [VU](Value *V) {
-            return areTwoInsertFromSameBuildVector(VU,
-                                                   cast<InsertElementInst>(V));
+          const TreeEntry *ScalarTE = getTreeEntry(EU.Scalar);
+          auto *It =
+              find_if(FirstUsers,
+                      [VU](const std::pair<Value *, const TreeEntry *> &Pair) {
+                        return areTwoInsertFromSameBuildVector(
+                            VU, cast<InsertElementInst>(Pair.first));
                       });
           int VecId = -1;
           if (It == FirstUsers.end()) {
-            VF.push_back(FTy->getNumElements());
-            ShuffleMask.emplace_back(VF.back(), UndefMaskElem);
+            (void)ShuffleMasks.emplace_back();
             // Find the insertvector, vectorized in tree, if any.
             Value *Base = VU;
-            while (isa<InsertElementInst>(Base)) {
+            while (auto *IEBase = dyn_cast<InsertElementInst>(Base)) {
               // Build the mask for the vectorized insertelement instructions.
-              if (const TreeEntry *E = getTreeEntry(Base)) {
-                VU = cast<InsertElementInst>(Base);
+              if (const TreeEntry *E = getTreeEntry(IEBase)) {
+                VU = IEBase;
                 do {
-                  int Idx = E->findLaneForValue(Base);
-                  ShuffleMask.back()[Idx] = Idx;
+                  int Idx = E->findLaneForValue(IEBase);
+                  SmallVectorImpl<int> &Mask = ShuffleMasks.back()[ScalarTE];
+                  if (Mask.empty())
+                    Mask.assign(FTy->getNumElements(), UndefMaskElem);
+                  Mask[Idx] = Idx;
                   Base = cast<InsertElementInst>(Base)->getOperand(0);
                 } while (E == getTreeEntry(Base));
                 break;
               }
               Base = cast<InsertElementInst>(Base)->getOperand(0);
             }
-            FirstUsers.push_back(VU);
-            DemandedElts.push_back(APInt::getZero(VF.back()));
+            FirstUsers.emplace_back(VU, ScalarTE);
+            DemandedElts.push_back(APInt::getZero(FTy->getNumElements()));
             VecId = FirstUsers.size() - 1;
           } else {
+            if (isFirstInsertElement(VU, cast<InsertElementInst>(It->first)))
+              It->first = VU;
             VecId = std::distance(FirstUsers.begin(), It);
           }
-          ShuffleMask[VecId][*InsertIdx] = EU.Lane;
-          DemandedElts[VecId].setBit(*InsertIdx);
+          int Idx = *InsertIdx;
+          SmallVectorImpl<int> &Mask = ShuffleMasks[VecId][ScalarTE];
+          if (Mask.empty())
+            Mask.assign(FTy->getNumElements(), UndefMaskElem);
+          assert(Mask[Idx] == UndefMaskElem &&
+                 "InsertElementInstruction used already.");
+          Mask[Idx] = EU.Lane;
+          DemandedElts[VecId].setBit(Idx);
           continue;
         }
       }
     }
 
     // If we plan to rewrite the tree in a smaller type, we will need to sign
     // extend the extracted value back to the original type. Here, we account
     // for the extract and the added cost of the sign extend if needed.
for (ExternalUser &EU : ExternalUses) {
     } else {
       ExtractCost +=
           TTI->getVectorInstrCost(Instruction::ExtractElement, VecTy, EU.Lane);
     }
   }
 
   InstructionCost SpillCost = getSpillCost();
   Cost += SpillCost + ExtractCost;
-  if (FirstUsers.size() == 1) {
-    int Limit = ShuffleMask.front().size() * 2;
-    if (!all_of(ShuffleMask.front(),
-                [Limit](int Idx) { return Idx < Limit; }) ||
-        !ShuffleVectorInst::isIdentityMask(ShuffleMask.front())) {
-      InstructionCost C = TTI->getShuffleCost(
+  auto &&ResizeToVF = [this, &Cost](const TreeEntry *TE, ArrayRef<int> Mask) {
+    InstructionCost C = 0;
+    unsigned VF = Mask.size();
+    unsigned VecVF = TE->getVectorFactor();
+    if (VF != VecVF &&
+        (any_of(Mask, [VF](int Idx) { return Idx >= static_cast<int>(VF); }) ||
+         (all_of(Mask,
+                 [VF](int Idx) { return Idx < 2 * static_cast<int>(VF); }) &&
+          !ShuffleVectorInst::isIdentityMask(Mask)))) {
+      SmallVector<int> OrigMask(VecVF, UndefMaskElem);
+      std::copy(Mask.begin(), std::next(Mask.begin(), std::min(VF, VecVF)),
+                OrigMask.begin());
+      C = TTI->getShuffleCost(
           TTI::SK_PermuteSingleSrc,
-          cast<FixedVectorType>(FirstUsers.front()->getType()),
-          ShuffleMask.front());
-      LLVM_DEBUG(dbgs() << "SLP: Adding cost " << C
-                        << " for final shuffle of insertelement external users "
-                        << *VectorizableTree.front()->Scalars.front() << ".\n"
-                        << "SLP: Current total cost = " << Cost << "\n");
+          FixedVectorType::get(TE->getMainOp()->getType(), VecVF), OrigMask);
+      LLVM_DEBUG(
+          dbgs() << "SLP: Adding cost " << C
+                 << " for final shuffle of insertelement external users.\n";
+          TE->dump(); dbgs() << "SLP: Current total cost = " << Cost << "\n");
       Cost += C;
+      return std::make_pair(TE, true);
     }
-    InstructionCost InsertCost = TTI->getScalarizationOverhead(
-        cast<FixedVectorType>(FirstUsers.front()->getType()),
-        DemandedElts.front(), /*Insert*/ true, /*Extract*/ false);
-    LLVM_DEBUG(dbgs() << "SLP: subtracting the cost " << InsertCost
-                      << " for insertelements gather.\n"
-                      << "SLP: Current total cost = " << Cost << "\n");
-    Cost -= InsertCost;
-  } else if (FirstUsers.size() >= 2) {
-    unsigned MaxVF = *std::max_element(VF.begin(), VF.end());
-    // Combined masks of the first 2 vectors.
-    SmallVector<int> CombinedMask(MaxVF, UndefMaskElem);
-    copy(ShuffleMask.front(), CombinedMask.begin());
-    APInt CombinedDemandedElts = DemandedElts.front().zextOrSelf(MaxVF);
-    auto *VecTy = FixedVectorType::get(
-        cast<VectorType>(FirstUsers.front()->getType())->getElementType(),
-        MaxVF);
-    for (int I = 0, E = ShuffleMask[1].size(); I < E; ++I) {
-      if (ShuffleMask[1][I] != UndefMaskElem) {
-        CombinedMask[I] = ShuffleMask[1][I] + MaxVF;
-        CombinedDemandedElts.setBit(I);
-      }
-    }
+    return std::make_pair(TE, false);
+  };
+  // Calculate the cost of the reshuffled vectors, if any.
+  for (int I = 0, E = FirstUsers.size(); I < E; ++I) {
+    Value *Base = cast<Instruction>(FirstUsers[I].first)->getOperand(0);
+    unsigned VF = ShuffleMasks[I].begin()->second.size();
+    auto *FTy = FixedVectorType::get(
+        cast<VectorType>(FirstUsers[I].first->getType())->getElementType(), VF);
+    auto Vector = ShuffleMasks[I].takeVector();
+    (void)performExtractsShuffleAction<const TreeEntry>(
+        makeMutableArrayRef(Vector.data(), Vector.size()), Base,
+        [](const TreeEntry *E) { return E->getVectorFactor(); }, ResizeToVF,
+        [this, FTy, &Cost](ArrayRef<int> Mask,
+                           ArrayRef<const TreeEntry *> TEs) {
+          assert((TEs.size() == 1 || TEs.size() == 2) &&
+                 "Expected exactly 1 or 2 tree entries.");
+          if (TEs.size() == 1) {
+            int Limit = 2 * Mask.size();
+            if (!all_of(Mask, [Limit](int Idx) { return Idx < Limit; }) ||
+                !ShuffleVectorInst::isIdentityMask(Mask)) {
               InstructionCost C =
-        TTI->getShuffleCost(TTI::SK_PermuteTwoSrc, VecTy, CombinedMask);
+                  TTI->getShuffleCost(TTI::SK_PermuteSingleSrc, FTy, Mask);
               LLVM_DEBUG(dbgs() << "SLP: Adding cost " << C
-                      << " for final shuffle of vector node and external "
-                         "insertelement users "
-                      << *VectorizableTree.front()->Scalars.front() << ".\n"
+                                << " for final shuffle of insertelement "
+                                   "external users.\n";
+                         TEs.front()->dump();
+                         dbgs()
                          << "SLP: Current total cost = " << Cost << "\n");
               Cost += C;
-    InstructionCost InsertCost = TTI->getScalarizationOverhead(
-        VecTy, CombinedDemandedElts, /*Insert*/ true, /*Extract*/ false);
-    LLVM_DEBUG(dbgs() << "SLP: subtracting the cost " << InsertCost
-                      << " for insertelements gather.\n"
-                      << "SLP: Current total cost = " << Cost << "\n");
-    Cost -= InsertCost;
-    for (int I = 2, E = FirstUsers.size(); I < E; ++I) {
-      if (ShuffleMask[I].empty())
-        continue;
-      // Other elements - permutation of 2 vectors (the initial one and the
-      // next Ith incoming vector).
-      unsigned VF = ShuffleMask[I].size();
-      for (unsigned Idx = 0; Idx < VF; ++Idx) {
-        int Mask = ShuffleMask[I][Idx];
-        if (Mask != UndefMaskElem)
-          CombinedMask[Idx] = MaxVF + Mask;
-        else if (CombinedMask[Idx] != UndefMaskElem)
-          CombinedMask[Idx] = Idx;
-      }
-      for (unsigned Idx = VF; Idx < MaxVF; ++Idx)
-        if (CombinedMask[Idx] != UndefMaskElem)
-          CombinedMask[Idx] = Idx;
+            }
+          } else {
             InstructionCost C =
-          TTI->getShuffleCost(TTI::SK_PermuteTwoSrc, VecTy, CombinedMask);
-      LLVM_DEBUG(dbgs() << "SLP: Adding cost " << C
+                TTI->getShuffleCost(TTI::SK_PermuteTwoSrc, FTy, Mask);
+            LLVM_DEBUG(dbgs()
+                           << "SLP: Adding cost " << C
                            << " for final shuffle of vector node and external "
-                           "insertelement users "
-                        << *VectorizableTree.front()->Scalars.front() << ".\n"
-                        << "SLP: Current total cost = " << Cost << "\n");
+                              "insertelement users.\n";
+                       if (TEs.front()) TEs.front()->dump(); TEs.back()->dump();
+                       dbgs() << "SLP: Current total cost = " << Cost << "\n");
             Cost += C;
-      InstructionCost InsertCost = TTI->getScalarizationOverhead(
-          cast<FixedVectorType>(FirstUsers[I]->getType()), DemandedElts[I],
-          /*Insert*/ true, /*Extract*/ false);
-      LLVM_DEBUG(dbgs() << "SLP: subtracting the cost " << InsertCost
-                        << " for insertelements gather.\n"
-                        << "SLP: Current total cost = " << Cost << "\n");
-      Cost -= InsertCost;
-    }
-  }
+          }
+          return TEs.back();
+        });
+        cast<FixedVectorType>(FirstUsers[I].first->getType()), DemandedElts[I],
 
 #ifndef NDEBUG
   SmallString<256> Str;
   {
     raw_svector_ostream OS(Str);
     OS << "SLP: Spill Cost = " << SpillCost << ".\n"
        << "SLP: Extract Cost = " << ExtractCost << ".\n"
        << "SLP: Total Cost = " << Cost << ".\n";

llvm/test/Transforms/SLPVectorizer/X86/crash_scheduling-inseltpoison.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt < %s -basic-aa -slp-vectorizer -S -mtriple=x86_64-apple-macosx10.8.0 -mcpu=corei7 | FileCheck %s
 
 target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-apple-darwin13.3.0"
 
 define void @_foo(double %p1, double %p2, double %p3) #0 {
 ; CHECK-LABEL: @_foo(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[TAB1:%.*]] = alloca [256 x i32], align 16
 ; CHECK-NEXT:    [[TAB2:%.*]] = alloca [256 x i32], align 16
 ; CHECK-NEXT:    br label [[BB1:%.*]]
 ; CHECK:       bb1:
-; CHECK-NEXT:    [[MUL19:%.*]] = fmul double [[P1:%.*]], 1.638400e+04
 ; CHECK-NEXT:    [[MUL20:%.*]] = fmul double [[P3:%.*]], 1.638400e+04
 ; CHECK-NEXT:    [[ADD:%.*]] = fadd double [[MUL20]], 8.192000e+03
-; CHECK-NEXT:    [[MUL21:%.*]] = fmul double [[P2:%.*]], 1.638400e+04
+; CHECK-NEXT:    [[TMP0:%.*]] = insertelement <2 x double> poison, double [[P1:%.*]], i32 0
+; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <2 x double> [[TMP0]], double [[P2:%.*]], i32 1
+; CHECK-NEXT:    [[TMP2:%.*]] = fmul <2 x double> [[TMP1]], <double 1.638400e+04, double 1.638400e+04>
+; CHECK-NEXT:    [[TMP3:%.*]] = insertelement <2 x double> <double 0.000000e+00, double poison>, double [[ADD]], i32 1
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
 ; CHECK-NEXT:    [[INDVARS_IV266:%.*]] = phi i64 [ 0, [[BB1]] ], [ [[INDVARS_IV_NEXT267:%.*]], [[FOR_BODY]] ]
-; CHECK-NEXT:    [[T_0259:%.*]] = phi double [ 0.000000e+00, [[BB1]] ], [ [[ADD27:%.*]], [[FOR_BODY]] ]
-; CHECK-NEXT:    [[P3_ADDR_0258:%.*]] = phi double [ [[ADD]], [[BB1]] ], [ [[ADD28:%.*]], [[FOR_BODY]] ]
-; CHECK-NEXT:    [[VECINIT_I_I237:%.*]] = insertelement <2 x double> poison, double [[T_0259]], i32 0
+; CHECK-NEXT:    [[TMP4:%.*]] = phi <2 x double> [ [[TMP3]], [[BB1]] ], [ [[TMP7:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[TMP5:%.*]] = extractelement <2 x double> [[TMP4]], i32 0
+; CHECK-NEXT:    [[VECINIT_I_I237:%.*]] = insertelement <2 x double> poison, double [[TMP5]], i32 0
 ; CHECK-NEXT:    [[X13:%.*]] = tail call i32 @_xfn(<2 x double> [[VECINIT_I_I237]])
 ; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [256 x i32], [256 x i32]* [[TAB1]], i64 0, i64 [[INDVARS_IV266]]
 ; CHECK-NEXT:    store i32 [[X13]], i32* [[ARRAYIDX]], align 4, !tbaa [[TBAA0:![0-9]+]]
-; CHECK-NEXT:    [[VECINIT_I_I:%.*]] = insertelement <2 x double> poison, double [[P3_ADDR_0258]], i32 0
+; CHECK-NEXT:    [[TMP6:%.*]] = extractelement <2 x double> [[TMP4]], i32 1
+; CHECK-NEXT:    [[VECINIT_I_I:%.*]] = insertelement <2 x double> poison, double [[TMP6]], i32 0
 ; CHECK-NEXT:    [[X14:%.*]] = tail call i32 @_xfn(<2 x double> [[VECINIT_I_I]])
 ; CHECK-NEXT:    [[ARRAYIDX26:%.*]] = getelementptr inbounds [256 x i32], [256 x i32]* [[TAB2]], i64 0, i64 [[INDVARS_IV266]]
 ; CHECK-NEXT:    store i32 [[X14]], i32* [[ARRAYIDX26]], align 4, !tbaa [[TBAA0]]
-; CHECK-NEXT:    [[ADD27]] = fadd double [[MUL19]], [[T_0259]]
-; CHECK-NEXT:    [[ADD28]] = fadd double [[MUL21]], [[P3_ADDR_0258]]
+; CHECK-NEXT:    [[TMP7]] = fadd <2 x double> [[TMP2]], [[TMP4]]
 ; CHECK-NEXT:    [[INDVARS_IV_NEXT267]] = add nuw nsw i64 [[INDVARS_IV266]], 1
 ; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT267]], 256
 ; CHECK-NEXT:    br i1 [[EXITCOND]], label [[RETURN:%.*]], label [[FOR_BODY]]
 ; CHECK:       return:
 ; CHECK-NEXT:    ret void
 ;
 entry:
   %tab1 = alloca [256 x i32], align 16

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

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt < %s -basic-aa -slp-vectorizer -S -mtriple=x86_64-apple-macosx10.8.0 -mcpu=corei7 | FileCheck %s
 
 target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-apple-darwin13.3.0"
 
 define void @_foo(double %p1, double %p2, double %p3) #0 {
 ; CHECK-LABEL: @_foo(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[TAB1:%.*]] = alloca [256 x i32], align 16
 ; CHECK-NEXT:    [[TAB2:%.*]] = alloca [256 x i32], align 16
 ; CHECK-NEXT:    br label [[BB1:%.*]]
 ; CHECK:       bb1:
-; CHECK-NEXT:    [[MUL19:%.*]] = fmul double [[P1:%.*]], 1.638400e+04
 ; CHECK-NEXT:    [[MUL20:%.*]] = fmul double [[P3:%.*]], 1.638400e+04
 ; CHECK-NEXT:    [[ADD:%.*]] = fadd double [[MUL20]], 8.192000e+03
-; CHECK-NEXT:    [[MUL21:%.*]] = fmul double [[P2:%.*]], 1.638400e+04
+; CHECK-NEXT:    [[TMP0:%.*]] = insertelement <2 x double> poison, double [[P1:%.*]], i32 0
+; CHECK-NEXT:    [[TMP1:%.*]] = insertelement <2 x double> [[TMP0]], double [[P2:%.*]], i32 1
+; CHECK-NEXT:    [[TMP2:%.*]] = fmul <2 x double> [[TMP1]], <double 1.638400e+04, double 1.638400e+04>
+; CHECK-NEXT:    [[TMP3:%.*]] = insertelement <2 x double> <double 0.000000e+00, double poison>, double [[ADD]], i32 1
 ; CHECK-NEXT:    br label [[FOR_BODY:%.*]]
 ; CHECK:       for.body:
 ; CHECK-NEXT:    [[INDVARS_IV266:%.*]] = phi i64 [ 0, [[BB1]] ], [ [[INDVARS_IV_NEXT267:%.*]], [[FOR_BODY]] ]
-; CHECK-NEXT:    [[T_0259:%.*]] = phi double [ 0.000000e+00, [[BB1]] ], [ [[ADD27:%.*]], [[FOR_BODY]] ]
-; CHECK-NEXT:    [[P3_ADDR_0258:%.*]] = phi double [ [[ADD]], [[BB1]] ], [ [[ADD28:%.*]], [[FOR_BODY]] ]
-; CHECK-NEXT:    [[VECINIT_I_I237:%.*]] = insertelement <2 x double> undef, double [[T_0259]], i32 0
+; CHECK-NEXT:    [[TMP4:%.*]] = phi <2 x double> [ [[TMP3]], [[BB1]] ], [ [[TMP7:%.*]], [[FOR_BODY]] ]
+; CHECK-NEXT:    [[TMP5:%.*]] = extractelement <2 x double> [[TMP4]], i32 0
+; CHECK-NEXT:    [[VECINIT_I_I237:%.*]] = insertelement <2 x double> undef, double [[TMP5]], i32 0
 ; CHECK-NEXT:    [[X13:%.*]] = tail call i32 @_xfn(<2 x double> [[VECINIT_I_I237]])
 ; CHECK-NEXT:    [[ARRAYIDX:%.*]] = getelementptr inbounds [256 x i32], [256 x i32]* [[TAB1]], i64 0, i64 [[INDVARS_IV266]]
 ; CHECK-NEXT:    store i32 [[X13]], i32* [[ARRAYIDX]], align 4, !tbaa [[TBAA0:![0-9]+]]
-; CHECK-NEXT:    [[VECINIT_I_I:%.*]] = insertelement <2 x double> undef, double [[P3_ADDR_0258]], i32 0
+; CHECK-NEXT:    [[TMP6:%.*]] = extractelement <2 x double> [[TMP4]], i32 1
+; CHECK-NEXT:    [[VECINIT_I_I:%.*]] = insertelement <2 x double> undef, double [[TMP6]], i32 0
 ; CHECK-NEXT:    [[X14:%.*]] = tail call i32 @_xfn(<2 x double> [[VECINIT_I_I]])
 ; CHECK-NEXT:    [[ARRAYIDX26:%.*]] = getelementptr inbounds [256 x i32], [256 x i32]* [[TAB2]], i64 0, i64 [[INDVARS_IV266]]
 ; CHECK-NEXT:    store i32 [[X14]], i32* [[ARRAYIDX26]], align 4, !tbaa [[TBAA0]]
-; CHECK-NEXT:    [[ADD27]] = fadd double [[MUL19]], [[T_0259]]
-; CHECK-NEXT:    [[ADD28]] = fadd double [[MUL21]], [[P3_ADDR_0258]]
+; CHECK-NEXT:    [[TMP7]] = fadd <2 x double> [[TMP2]], [[TMP4]]
 ; CHECK-NEXT:    [[INDVARS_IV_NEXT267]] = add nuw nsw i64 [[INDVARS_IV266]], 1
 ; CHECK-NEXT:    [[EXITCOND:%.*]] = icmp eq i64 [[INDVARS_IV_NEXT267]], 256
 ; CHECK-NEXT:    br i1 [[EXITCOND]], label [[RETURN:%.*]], label [[FOR_BODY]]
 ; CHECK:       return:
 ; CHECK-NEXT:    ret void
 ;
 entry:
   %tab1 = alloca [256 x i32], align 16

llvm/test/Transforms/SLPVectorizer/X86/extracts-with-undefs.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
 ; RUN: opt < %s -slp-vectorizer -S -mtriple=x86_64-unknown-linux-gnu | FileCheck %s
 
 define void @test() {
 ; CHECK-LABEL: @test(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    br label [[BODY:%.*]]
 ; CHECK:       body:
-; CHECK-NEXT:    [[TMP0:%.*]] = phi <2 x double> [ zeroinitializer, [[ENTRY:%.*]] ], [ zeroinitializer, [[BODY]] ]
-; CHECK-NEXT:    [[TMP1:%.*]] = extractelement <2 x double> [[TMP0]], i32 1
-; CHECK-NEXT:    [[TMP2:%.*]] = insertelement <2 x double> <double undef, double poison>, double [[TMP1]], i32 1
-; CHECK-NEXT:    [[TMP3:%.*]] = fmul fast <2 x double> [[TMP2]], zeroinitializer
-; CHECK-NEXT:    [[TMP4:%.*]] = extractelement <2 x double> [[TMP3]], i32 0
-; CHECK-NEXT:    [[TMP5:%.*]] = extractelement <2 x double> [[TMP3]], i32 1
-; CHECK-NEXT:    [[ADD8_I_I:%.*]] = fadd fast double [[TMP5]], [[TMP4]]
+; CHECK-NEXT:    [[PHI1:%.*]] = phi double [ 0.000000e+00, [[ENTRY:%.*]] ], [ 0.000000e+00, [[BODY]] ]
+; CHECK-NEXT:    [[PHI2:%.*]] = phi double [ 0.000000e+00, [[ENTRY]] ], [ 0.000000e+00, [[BODY]] ]
+; CHECK-NEXT:    [[MUL_I478_I:%.*]] = fmul fast double [[PHI1]], 0.000000e+00
+; CHECK-NEXT:    [[MUL7_I485_I:%.*]] = fmul fast double undef, 0.000000e+00
+; CHECK-NEXT:    [[ADD8_I_I:%.*]] = fadd fast double [[MUL_I478_I]], [[MUL7_I485_I]]
 ; CHECK-NEXT:    [[CMP42_I:%.*]] = fcmp fast ole double [[ADD8_I_I]], 0.000000e+00
 ; CHECK-NEXT:    br i1 false, label [[BODY]], label [[EXIT:%.*]]
 ; CHECK:       exit:
 ; CHECK-NEXT:    br i1 false, label [[IF_THEN135_I:%.*]], label [[IF_END209_I:%.*]]
 ; CHECK:       if.then135.i:
-; CHECK-NEXT:    [[TMP6:%.*]] = fcmp fast olt <2 x double> [[TMP0]], zeroinitializer
-; CHECK-NEXT:    [[TMP7:%.*]] = extractelement <2 x i1> [[TMP6]], i32 0
-; CHECK-NEXT:    [[TMP8:%.*]] = insertelement <2 x i1> <i1 poison, i1 false>, i1 [[TMP7]], i32 0
-; CHECK-NEXT:    [[TMP9:%.*]] = select <2 x i1> [[TMP8]], <2 x double> zeroinitializer, <2 x double> zeroinitializer
-; CHECK-NEXT:    [[TMP10:%.*]] = fmul fast <2 x double> zeroinitializer, [[TMP9]]
-; CHECK-NEXT:    [[TMP11:%.*]] = fmul fast <2 x double> [[TMP10]], zeroinitializer
-; CHECK-NEXT:    [[TMP12:%.*]] = fadd fast <2 x double> [[TMP11]], zeroinitializer
+; CHECK-NEXT:    [[CMP145_I:%.*]] = fcmp fast olt double [[PHI1]], 0.000000e+00
+; CHECK-NEXT:    [[CMP152_I:%.*]] = fcmp fast olt double [[PHI2]], 0.000000e+00
+; CHECK-NEXT:    [[TMP0:%.*]] = insertelement <2 x i1> <i1 poison, i1 false>, i1 [[CMP152_I]], i32 0
+; CHECK-NEXT:    [[TMP1:%.*]] = select <2 x i1> [[TMP0]], <2 x double> zeroinitializer, <2 x double> zeroinitializer
+; CHECK-NEXT:    [[TMP2:%.*]] = fmul fast <2 x double> zeroinitializer, [[TMP1]]
+; CHECK-NEXT:    [[TMP3:%.*]] = fmul fast <2 x double> [[TMP2]], zeroinitializer
+; CHECK-NEXT:    [[TMP4:%.*]] = fadd fast <2 x double> [[TMP3]], zeroinitializer
 ; CHECK-NEXT:    br label [[IF_END209_I]]
 ; CHECK:       if.end209.i:
-; CHECK-NEXT:    [[TMP13:%.*]] = phi <2 x double> [ [[TMP12]], [[IF_THEN135_I]] ], [ zeroinitializer, [[EXIT]] ]
+; CHECK-NEXT:    [[TMP5:%.*]] = phi <2 x double> [ [[TMP4]], [[IF_THEN135_I]] ], [ zeroinitializer, [[EXIT]] ]
 ; CHECK-NEXT:    ret void
 ;
 entry:
   br label %body
 
 body:
   %phi1 = phi double [ 0.000000e+00, %entry ], [ 0.000000e+00, %body ]
   %phi2 = phi double [ 0.000000e+00, %entry ], [ 0.000000e+00, %body ]