[X86] Add calculation for elements in structures in getting uniform base for the Gather/Scatter intrinsic.

[pengfei]

Summary

Add calculation for elements in structures in getting uniform base for the Gather/Scatter intrinsic.

[pengfei]

Updated to only check the final type.

[pengfei]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4358

I think for X86, we can leverage the displacement to handle indices aren't 0.

[craig.topper]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4369

The normal way to do this is to use gep_type_iterator in the loop as well and increment it as you go through the loop.

[pengfei]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4358

Oh, displacement is already used for base. Sorry for the noise.

[pengfei]

Using gep_type_iterator to traverse.

[pengfei]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4369

Great, thanks!

[craig.topper]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4373

What is IndexVal isn't a constant?

4373

What if IndexVal isn't a constant?

[pengfei]

We don't need to check if constant is defined.

[pengfei]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4373

Makes sence.

[LuoYuanke]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4373

Maybe we can calculate the offset for struct if the index is constant.

APInt Offset(DL.getIndexTypeSizeInBits(GEP->getType()), 0);
if (!GEP->accumulateConstantOffset(DL, Offset))
  return false;
Index = DAG.getTargetConstant(Offset, SDB->getCurSDLoc(), TLI.getPointerTy(DL));
Scale = DAG.getTargetConstant(1, SDB->getCurSDLoc(), TLI.getPointerTy(DL));

Also recognize zeroinitializer in accumulateConstantOffset().

• a/llvm/lib/IR/Operator.cpp

+++ b/llvm/lib/IR/Operator.cpp
@@ -39,6 +39,9 @@ bool GEPOperator::accumulateConstantOffset(const DataLayout &DL,

for (gep_type_iterator GTI = gep_type_begin(this), GTE = gep_type_end(this);
     GTI != GTE; ++GTI) {

+
+ if (isa<ConstantAggregateZero>(GTI.getOperand()))
+ continue;

ConstantInt *OpC = dyn_cast<ConstantInt>(GTI.getOperand());
if (!OpC)
  return false;

[craig.topper]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4374

Does the test cover this change? I was just pointing out a mistake in the FIXME. I wasn't trying to ask for a code change unless its relevant to fixing the bug the test is for.

[pengfei]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4373

Index inside a structure is always constant.
I think your suggestions is reasonable. We can calculate the offset for a given constant index.
Besides, the intermediate non zero constant index still can be calculated as your suggestions.
But it looks like more optimization than bug fix. It's better to be implemented in another patch.

Why we need recognize zeroinitializer?

4374

Not currently. But there's a bit relationship. Without this change, I need to *make* the same constant somewhere to get the excepted result.

%4 = icmp eq <8 x i32> %trigger, <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7>

I think I can remove the constant to cover this change. I can also keep the FIXME.
Which one do you think it's better?

[LuoYuanke]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4373

Yes. I agree it is better to be implemented in another patch.

<<<Why we need recognize zeroinitializer?
Because in GEPOperator::accumulateConstantOffset() it only support scalar index, not support splat vector index. Checking zeroinitializer is to take the index as scalar zero.

[pengfei]

Optimized as Yuanke's suggestion.

[xbolva00]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4387

just 'cast'?

CI must be nonnull anyway, otherwise your assert fires.

[xbolva00]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4383

Same here

[pengfei]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4387

I wonder if there's a real case that index of a structure is not a constant? I guess not, but I'm not so sure. So I put an assert followed.

[pengfei]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4387

From LangRef "When indexing into a (optionally packed) structure, only i32 integer constants are allowed (when using a vector of indices they must all be the same i32 integer constant)."
I think I can remove the assert.

[pengfei]

Removed unnecessary assertions. Merged some common code.

[pengfei]

Removed unnecessary else.

[pengfei]

Removed unnecessary loop condition.
Added a vectorized index test.
Don't check if constant exist in DAG.

[craig.topper]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4364

Can we just handle the FinalIndex cases after the loop instead of trying to special case the last iteration of the loop? It might require a tiny bit of duplication, but I think it would be more understandable.

[pengfei]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4364

Sure!

[pengfei]

Address review comments.

[pengfei]

Ping.

[LuoYuanke]

In D71442#1789052, @pengfei wrote:

Ping.

LGTM

Hi Craig,
How about you?

[craig.topper]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

4374

use isa instead of dyn_cast if you don't need the casted pointer.

4374

Isn't GEP->getOperand(FinalIndex) already in IndexVal?

4384

Can this just be cast<Constant>(IndexVal)?

[pengfei]

Address review comments.

[craig.topper]

LGTM. Thanks!

[gribozavr2]

Hi,

I believe this commit broke the expensive checks bot. Could you take a look?

http://lab.llvm.org:8011/builders/llvm-clang-x86_64-expensive-checks-debian/builds/631

The previous build is green: http://lab.llvm.org:8011/builders/llvm-clang-x86_64-expensive-checks-debian/builds/630

[craig.topper]

In D71442#1789808, @gribozavr2 wrote:

Hi,

I believe this commit broke the expensive checks bot. Could you take a look?

http://lab.llvm.org:8011/builders/llvm-clang-x86_64-expensive-checks-debian/builds/631

The previous build is green: http://lab.llvm.org:8011/builders/llvm-clang-x86_64-expensive-checks-debian/builds/630

I should have a fix for this in a few minutes.

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

static bool getUniformBase(const Value *&Ptr, SDValue &Base, SDValue &Index,
   const Value *GEPPtr = GEP->getPointerOperand();
   if (!GEPPtr->getType()->isVectorTy())
     Ptr = GEPPtr;
   else if (!(Ptr = getSplatValue(GEPPtr)))
     return false;
 
   unsigned FinalIndex = GEP->getNumOperands() - 1;
   Value *IndexVal = GEP->getOperand(FinalIndex);
+  gep_type_iterator GTI = gep_type_begin(*GEP);
 
   // Ensure all the other indices are 0.
-  for (unsigned i = 1; i < FinalIndex; ++i) {
+  for (unsigned i = 1; i < FinalIndex; ++i, ++GTI) {
     auto *C = dyn_cast<Constant>(GEP->getOperand(i));
     if (!C)
       return false;
     if (isa<VectorType>(C->getType()))
       C = C->getSplatValue();
     auto *CI = dyn_cast_or_null<ConstantInt>(C);
     if (!CI || !CI->isZero())
       return false;
   }
 
   // The operands of the GEP may be defined in another basic block.
   // In this case we'll not find nodes for the operands.
-  if (!SDB->findValue(Ptr) || !SDB->findValue(IndexVal))
+  if (!SDB->findValue(Ptr))
+    return false;
+  Constant *C = dyn_cast<Constant>(IndexVal);
+  if (!C && !SDB->findValue(IndexVal))
     return false;
 
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
   const DataLayout &DL = DAG.getDataLayout();
-  Scale = DAG.getTargetConstant(DL.getTypeAllocSize(GEP->getResultElementType()),
-                                SDB->getCurSDLoc(), TLI.getPointerTy(DL));
-  Base = SDB->getValue(Ptr);
+  StructType *STy = GTI.getStructTypeOrNull();
+
+  if (STy) {
+    const StructLayout *SL = DL.getStructLayout(STy);
+    if (isa<VectorType>(C->getType())) {
+      C = C->getSplatValue();
+      // FIXME: If getSplatValue may return nullptr for a structure?
+      // If not, the following check can be removed.
+      if (!C)
+        return false;
+    }
+    auto *CI = cast<ConstantInt>(C);
+    Scale = DAG.getTargetConstant(1, SDB->getCurSDLoc(), TLI.getPointerTy(DL));
+    Index = DAG.getTargetConstant(SL->getElementOffset(CI->getZExtValue()),
+                                  SDB->getCurSDLoc(), TLI.getPointerTy(DL));
+  } else {
+    Scale = DAG.getTargetConstant(
+                DL.getTypeAllocSize(GEP->getResultElementType()),
     Index = SDB->getValue(IndexVal);
+  }
+  Base = SDB->getValue(Ptr);
   IndexType = ISD::SIGNED_SCALED;
 
-  if (!Index.getValueType().isVector()) {
+  if (STy || !Index.getValueType().isVector()) {
     unsigned GEPWidth = GEP->getType()->getVectorNumElements();
     EVT VT = EVT::getVectorVT(Context, Index.getValueType(), GEPWidth);
     Index = DAG.getSplatBuildVector(VT, SDLoc(Index), Index);
   }
   return true;
 }
 
 void SelectionDAGBuilder::visitMaskedScatter(const CallInst &I) {

llvm/test/CodeGen/X86/masked_gather.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc -mtriple=x86_64-unknown-linux-gnu  -mattr=+sse4.2 < %s | FileCheck %s --check-prefixes=ALL,SSE,SSE42
 ; RUN: llc -mtriple=x86_64-unknown-linux-gnu  -mattr=+avx < %s | FileCheck %s --check-prefixes=ALL,AVX,AVX1
 ; RUN: llc -mtriple=x86_64-unknown-linux-gnu  -mattr=+avx2 < %s | FileCheck %s --check-prefixes=ALL,AVX,AVX2
 ; RUN: llc -mtriple=x86_64-unknown-linux-gnu  -mattr=+avx512f < %s | FileCheck %s --check-prefixes=ALL,AVX512
 
 ;
 ; vXf32
 ;
 
+%struct.a = type { [4 x i32], [4 x i8], %struct.b, i32 }
+%struct.b = type { i32, i32 }
+@c = external dso_local global %struct.a, align 4
+
 define <4 x float> @gather_v4f32_ptr_v4i32(<4 x float*> %ptr, <4 x i32> %trigger, <4 x float> %passthru) {
 ; SSE-LABEL: gather_v4f32_ptr_v4i32:
 ; SSE:       # %bb.0:
 ; SSE-NEXT:    pxor %xmm4, %xmm4
 ; SSE-NEXT:    pcmpeqd %xmm2, %xmm4
 ; SSE-NEXT:    movmskps %xmm4, %eax
 ; SSE-NEXT:    testb $1, %al
 ; SSE-NEXT:    jne .LBB0_1
; AVX512-NEXT:    retq
   %vptr1 = shufflevector <16 x i8*> %vptr0, <16 x i8*> undef, <16 x i32> zeroinitializer
   %vptr2 = getelementptr i8, <16 x i8*> %vptr1, <16 x i32> %idx
 
   %mask = icmp eq <16 x i8> %trigger, zeroinitializer
   %res = call <16 x i8> @llvm.masked.gather.v16i8.v16p0i8(<16 x i8*> %vptr2, i32 4, <16 x i1> %mask, <16 x i8> %passthru)
   ret <16 x i8> %res
 }
 
+define <8 x i32> @gather_v8i32_v8i32(<8 x i32> %trigger) {
+; SSE-LABEL: gather_v8i32_v8i32:
+; SSE:       # %bb.0:
+; SSE-NEXT:    movdqa %xmm1, %xmm3
+; SSE-NEXT:    movdqa %xmm0, %xmm2
+; SSE-NEXT:    movl $c, %eax
+; SSE-NEXT:    movq %rax, %xmm0
+; SSE-NEXT:    pshufd {{.*#+}} xmm4 = xmm0[0,1,0,1]
+; SSE-NEXT:    movdqa {{.*#+}} xmm5 = [12,12]
+; SSE-NEXT:    paddq %xmm4, %xmm5
+; SSE-NEXT:    pxor %xmm0, %xmm0
+; SSE-NEXT:    pcmpeqd %xmm0, %xmm1
+; SSE-NEXT:    pcmpeqd %xmm2, %xmm0
+; SSE-NEXT:    packssdw %xmm1, %xmm0
+; SSE-NEXT:    packsswb %xmm0, %xmm0
+; SSE-NEXT:    pmovmskb %xmm0, %eax
+; SSE-NEXT:    testb $1, %al
+; SSE-NEXT:    je .LBB4_1
+; SSE-NEXT:  # %bb.2: # %cond.load
+; SSE-NEXT:    movq %xmm5, %rcx
+; SSE-NEXT:    movd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; SSE-NEXT:    testb $2, %al
+; SSE-NEXT:    jne .LBB4_4
+; SSE-NEXT:    jmp .LBB4_5
+; SSE-NEXT:  .LBB4_1:
+; SSE-NEXT:    # implicit-def: $xmm0
+; SSE-NEXT:    testb $2, %al
+; SSE-NEXT:    je .LBB4_5
+; SSE-NEXT:  .LBB4_4: # %cond.load1
+; SSE-NEXT:    pextrq $1, %xmm5, %rcx
+; SSE-NEXT:    pinsrd $1, (%rcx), %xmm0
+; SSE-NEXT:  .LBB4_5: # %else2
+; SSE-NEXT:    testb $4, %al
+; SSE-NEXT:    jne .LBB4_6
+; SSE-NEXT:  # %bb.7: # %else5
+; SSE-NEXT:    testb $8, %al
+; SSE-NEXT:    jne .LBB4_8
+; SSE-NEXT:  .LBB4_9: # %else8
+; SSE-NEXT:    testb $16, %al
+; SSE-NEXT:    je .LBB4_10
+; SSE-NEXT:  .LBB4_11: # %cond.load10
+; SSE-NEXT:    movq %xmm5, %rcx
+; SSE-NEXT:    pinsrd $0, (%rcx), %xmm1
+; SSE-NEXT:    testb $32, %al
+; SSE-NEXT:    jne .LBB4_13
+; SSE-NEXT:    jmp .LBB4_14
+; SSE-NEXT:  .LBB4_6: # %cond.load4
+; SSE-NEXT:    movq %xmm5, %rcx
+; SSE-NEXT:    pinsrd $2, (%rcx), %xmm0
+; SSE-NEXT:    testb $8, %al
+; SSE-NEXT:    je .LBB4_9
+; SSE-NEXT:  .LBB4_8: # %cond.load7
+; SSE-NEXT:    pextrq $1, %xmm5, %rcx
+; SSE-NEXT:    pinsrd $3, (%rcx), %xmm0
+; SSE-NEXT:    testb $16, %al
+; SSE-NEXT:    jne .LBB4_11
+; SSE-NEXT:  .LBB4_10:
+; SSE-NEXT:    # implicit-def: $xmm1
+; SSE-NEXT:    testb $32, %al
+; SSE-NEXT:    je .LBB4_14
+; SSE-NEXT:  .LBB4_13: # %cond.load13
+; SSE-NEXT:    pextrq $1, %xmm5, %rcx
+; SSE-NEXT:    pinsrd $1, (%rcx), %xmm1
+; SSE-NEXT:  .LBB4_14: # %else14
+; SSE-NEXT:    testb $64, %al
+; SSE-NEXT:    jne .LBB4_15
+; SSE-NEXT:  # %bb.16: # %else17
+; SSE-NEXT:    testb $-128, %al
+; SSE-NEXT:    je .LBB4_18
+; SSE-NEXT:  .LBB4_17: # %cond.load19
+; SSE-NEXT:    pextrq $1, %xmm5, %rax
+; SSE-NEXT:    pinsrd $3, (%rax), %xmm1
+; SSE-NEXT:  .LBB4_18: # %else20
+; SSE-NEXT:    pxor %xmm5, %xmm5
+; SSE-NEXT:    paddq {{.*}}(%rip), %xmm4
+; SSE-NEXT:    movdqa %xmm2, %xmm6
+; SSE-NEXT:    pcmpeqd %xmm5, %xmm6
+; SSE-NEXT:    pcmpeqd %xmm3, %xmm5
+; SSE-NEXT:    packssdw %xmm5, %xmm6
+; SSE-NEXT:    packsswb %xmm0, %xmm6
+; SSE-NEXT:    pmovmskb %xmm6, %eax
+; SSE-NEXT:    testb $1, %al
+; SSE-NEXT:    je .LBB4_19
+; SSE-NEXT:  # %bb.20: # %cond.load23
+; SSE-NEXT:    movq %xmm4, %rcx
+; SSE-NEXT:    movd {{.*#+}} xmm5 = mem[0],zero,zero,zero
+; SSE-NEXT:    testb $2, %al
+; SSE-NEXT:    jne .LBB4_22
+; SSE-NEXT:    jmp .LBB4_23
+; SSE-NEXT:  .LBB4_15: # %cond.load16
+; SSE-NEXT:    movq %xmm5, %rcx
+; SSE-NEXT:    pinsrd $2, (%rcx), %xmm1
+; SSE-NEXT:    testb $-128, %al
+; SSE-NEXT:    jne .LBB4_17
+; SSE-NEXT:    jmp .LBB4_18
+; SSE-NEXT:  .LBB4_19:
+; SSE-NEXT:    # implicit-def: $xmm5
+; SSE-NEXT:    testb $2, %al
+; SSE-NEXT:    je .LBB4_23
+; SSE-NEXT:  .LBB4_22: # %cond.load29
+; SSE-NEXT:    pextrq $1, %xmm4, %rcx
+; SSE-NEXT:    pinsrd $1, (%rcx), %xmm5
+; SSE-NEXT:  .LBB4_23: # %else33
+; SSE-NEXT:    testb $4, %al
+; SSE-NEXT:    jne .LBB4_24
+; SSE-NEXT:  # %bb.25: # %else39
+; SSE-NEXT:    testb $8, %al
+; SSE-NEXT:    jne .LBB4_26
+; SSE-NEXT:  .LBB4_27: # %else45
+; SSE-NEXT:    testb $16, %al
+; SSE-NEXT:    je .LBB4_28
+; SSE-NEXT:  .LBB4_29: # %cond.load47
+; SSE-NEXT:    movq %xmm4, %rcx
+; SSE-NEXT:    pinsrd $0, (%rcx), %xmm6
+; SSE-NEXT:    testb $32, %al
+; SSE-NEXT:    jne .LBB4_31
+; SSE-NEXT:    jmp .LBB4_32
+; SSE-NEXT:  .LBB4_24: # %cond.load35
+; SSE-NEXT:    movq %xmm4, %rcx
+; SSE-NEXT:    pinsrd $2, (%rcx), %xmm5
+; SSE-NEXT:    testb $8, %al
+; SSE-NEXT:    je .LBB4_27
+; SSE-NEXT:  .LBB4_26: # %cond.load41
+; SSE-NEXT:    pextrq $1, %xmm4, %rcx
+; SSE-NEXT:    pinsrd $3, (%rcx), %xmm5
+; SSE-NEXT:    testb $16, %al
+; SSE-NEXT:    jne .LBB4_29
+; SSE-NEXT:  .LBB4_28:
+; SSE-NEXT:    # implicit-def: $xmm6
+; SSE-NEXT:    testb $32, %al
+; SSE-NEXT:    je .LBB4_32
+; SSE-NEXT:  .LBB4_31: # %cond.load53
+; SSE-NEXT:    pextrq $1, %xmm4, %rcx
+; SSE-NEXT:    pinsrd $1, (%rcx), %xmm6
+; SSE-NEXT:  .LBB4_32: # %else57
+; SSE-NEXT:    testb $64, %al
+; SSE-NEXT:    jne .LBB4_33
+; SSE-NEXT:  # %bb.34: # %else63
+; SSE-NEXT:    testb $-128, %al
+; SSE-NEXT:    je .LBB4_36
+; SSE-NEXT:  .LBB4_35: # %cond.load65
+; SSE-NEXT:    pextrq $1, %xmm4, %rax
+; SSE-NEXT:    pinsrd $3, (%rax), %xmm6
+; SSE-NEXT:  .LBB4_36: # %else69
+; SSE-NEXT:    pxor %xmm7, %xmm7
+; SSE-NEXT:    pcmpeqd %xmm7, %xmm2
+; SSE-NEXT:    pcmpeqd %xmm7, %xmm3
+; SSE-NEXT:    packssdw %xmm3, %xmm2
+; SSE-NEXT:    packsswb %xmm0, %xmm2
+; SSE-NEXT:    pmovmskb %xmm2, %eax
+; SSE-NEXT:    testb $1, %al
+; SSE-NEXT:    je .LBB4_37
+; SSE-NEXT:  # %bb.38: # %cond.load72
+; SSE-NEXT:    movq %xmm4, %rcx
+; SSE-NEXT:    movd {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; SSE-NEXT:    testb $2, %al
+; SSE-NEXT:    jne .LBB4_40
+; SSE-NEXT:    jmp .LBB4_41
+; SSE-NEXT:  .LBB4_33: # %cond.load59
+; SSE-NEXT:    movq %xmm4, %rcx
+; SSE-NEXT:    pinsrd $2, (%rcx), %xmm6
+; SSE-NEXT:    testb $-128, %al
+; SSE-NEXT:    jne .LBB4_35
+; SSE-NEXT:    jmp .LBB4_36
+; SSE-NEXT:  .LBB4_37:
+; SSE-NEXT:    # implicit-def: $xmm2
+; SSE-NEXT:    testb $2, %al
+; SSE-NEXT:    je .LBB4_41
+; SSE-NEXT:  .LBB4_40: # %cond.load78
+; SSE-NEXT:    pextrq $1, %xmm4, %rcx
+; SSE-NEXT:    pinsrd $1, (%rcx), %xmm2
+; SSE-NEXT:  .LBB4_41: # %else82
+; SSE-NEXT:    testb $4, %al
+; SSE-NEXT:    jne .LBB4_42
+; SSE-NEXT:  # %bb.43: # %else88
+; SSE-NEXT:    testb $8, %al
+; SSE-NEXT:    jne .LBB4_44
+; SSE-NEXT:  .LBB4_45: # %else94
+; SSE-NEXT:    testb $16, %al
+; SSE-NEXT:    je .LBB4_46
+; SSE-NEXT:  .LBB4_47: # %cond.load96
+; SSE-NEXT:    movq %xmm4, %rcx
+; SSE-NEXT:    pinsrd $0, (%rcx), %xmm3
+; SSE-NEXT:    testb $32, %al
+; SSE-NEXT:    jne .LBB4_49
+; SSE-NEXT:    jmp .LBB4_50
+; SSE-NEXT:  .LBB4_42: # %cond.load84
+; SSE-NEXT:    movq %xmm4, %rcx
+; SSE-NEXT:    pinsrd $2, (%rcx), %xmm2
+; SSE-NEXT:    testb $8, %al
+; SSE-NEXT:    je .LBB4_45
+; SSE-NEXT:  .LBB4_44: # %cond.load90
+; SSE-NEXT:    pextrq $1, %xmm4, %rcx
+; SSE-NEXT:    pinsrd $3, (%rcx), %xmm2
+; SSE-NEXT:    testb $16, %al
+; SSE-NEXT:    jne .LBB4_47
+; SSE-NEXT:  .LBB4_46:
+; SSE-NEXT:    # implicit-def: $xmm3
+; SSE-NEXT:    testb $32, %al
+; SSE-NEXT:    je .LBB4_50
+; SSE-NEXT:  .LBB4_49: # %cond.load102
+; SSE-NEXT:    pextrq $1, %xmm4, %rcx
+; SSE-NEXT:    pinsrd $1, (%rcx), %xmm3
+; SSE-NEXT:  .LBB4_50: # %else106
+; SSE-NEXT:    testb $64, %al
+; SSE-NEXT:    je .LBB4_52
+; SSE-NEXT:  # %bb.51: # %cond.load108
+; SSE-NEXT:    movq %xmm4, %rcx
+; SSE-NEXT:    pinsrd $2, (%rcx), %xmm3
+; SSE-NEXT:  .LBB4_52: # %else112
+; SSE-NEXT:    paddd %xmm5, %xmm0
+; SSE-NEXT:    paddd %xmm6, %xmm1
+; SSE-NEXT:    testb $-128, %al
+; SSE-NEXT:    je .LBB4_54
+; SSE-NEXT:  # %bb.53: # %cond.load114
+; SSE-NEXT:    pextrq $1, %xmm4, %rax
+; SSE-NEXT:    pinsrd $3, (%rax), %xmm3
+; SSE-NEXT:  .LBB4_54: # %else118
+; SSE-NEXT:    paddd %xmm3, %xmm1
+; SSE-NEXT:    paddd %xmm2, %xmm0
+; SSE-NEXT:    retq
+;
+; AVX1-LABEL: gather_v8i32_v8i32:
+; AVX1:       # %bb.0:
+; AVX1-NEXT:    movl $c, %eax
+; AVX1-NEXT:    vmovq %rax, %xmm1
+; AVX1-NEXT:    vpshufd {{.*#+}} xmm3 = xmm1[0,1,0,1]
+; AVX1-NEXT:    vpaddq {{.*}}(%rip), %xmm3, %xmm1
+; AVX1-NEXT:    vinsertf128 $1, %xmm1, %ymm1, %ymm4
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm9
+; AVX1-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm9, %xmm5
+; AVX1-NEXT:    vpcmpeqd %xmm1, %xmm0, %xmm1
+; AVX1-NEXT:    vinsertf128 $1, %xmm5, %ymm1, %ymm1
+; AVX1-NEXT:    vmovmskps %ymm1, %eax
+; AVX1-NEXT:    testb $1, %al
+; AVX1-NEXT:    # implicit-def: $ymm1
+; AVX1-NEXT:    je .LBB4_2
+; AVX1-NEXT:  # %bb.1: # %cond.load
+; AVX1-NEXT:    vmovq %xmm4, %rcx
+; AVX1-NEXT:    vmovd {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; AVX1-NEXT:  .LBB4_2: # %else
+; AVX1-NEXT:    testb $2, %al
+; AVX1-NEXT:    je .LBB4_4
+; AVX1-NEXT:  # %bb.3: # %cond.load1
+; AVX1-NEXT:    vpextrq $1, %xmm4, %rcx
+; AVX1-NEXT:    vpinsrd $1, (%rcx), %xmm1, %xmm5
+; AVX1-NEXT:    vblendps {{.*#+}} ymm1 = ymm5[0,1,2,3],ymm1[4,5,6,7]
+; AVX1-NEXT:  .LBB4_4: # %else2
+; AVX1-NEXT:    testb $4, %al
+; AVX1-NEXT:    vextractf128 $1, %ymm4, %xmm6
+; AVX1-NEXT:    jne .LBB4_5
+; AVX1-NEXT:  # %bb.6: # %else5
+; AVX1-NEXT:    testb $8, %al
+; AVX1-NEXT:    jne .LBB4_7
+; AVX1-NEXT:  .LBB4_8: # %else8
+; AVX1-NEXT:    testb $16, %al
+; AVX1-NEXT:    jne .LBB4_9
+; AVX1-NEXT:  .LBB4_10: # %else11
+; AVX1-NEXT:    testb $32, %al
+; AVX1-NEXT:    jne .LBB4_11
+; AVX1-NEXT:  .LBB4_12: # %else14
+; AVX1-NEXT:    testb $64, %al
+; AVX1-NEXT:    je .LBB4_14
+; AVX1-NEXT:  .LBB4_13: # %cond.load16
+; AVX1-NEXT:    vmovq %xmm6, %rcx
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm4
+; AVX1-NEXT:    vpinsrd $2, (%rcx), %xmm4, %xmm4
+; AVX1-NEXT:    vinsertf128 $1, %xmm4, %ymm1, %ymm1
+; AVX1-NEXT:  .LBB4_14: # %else17
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm3, %ymm5
+; AVX1-NEXT:    testb $-128, %al
+; AVX1-NEXT:    je .LBB4_16
+; AVX1-NEXT:  # %bb.15: # %cond.load19
+; AVX1-NEXT:    vpextrq $1, %xmm6, %rax
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm3
+; AVX1-NEXT:    vpinsrd $3, (%rax), %xmm3, %xmm3
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm1, %ymm1
+; AVX1-NEXT:  .LBB4_16: # %else20
+; AVX1-NEXT:    vextractf128 $1, %ymm5, %xmm3
+; AVX1-NEXT:    vmovdqa {{.*#+}} xmm10 = [28,28]
+; AVX1-NEXT:    vpaddq %xmm3, %xmm10, %xmm3
+; AVX1-NEXT:    vpaddq %xmm5, %xmm10, %xmm4
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm4, %ymm7
+; AVX1-NEXT:    vpxor %xmm4, %xmm4, %xmm4
+; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm9, %xmm2
+; AVX1-NEXT:    vpcmpeqd %xmm4, %xmm0, %xmm4
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm4, %ymm2
+; AVX1-NEXT:    vmovmskps %ymm2, %eax
+; AVX1-NEXT:    testb $1, %al
+; AVX1-NEXT:    # implicit-def: $ymm4
+; AVX1-NEXT:    je .LBB4_18
+; AVX1-NEXT:  # %bb.17: # %cond.load23
+; AVX1-NEXT:    vmovq %xmm7, %rcx
+; AVX1-NEXT:    vmovd {{.*#+}} xmm4 = mem[0],zero,zero,zero
+; AVX1-NEXT:  .LBB4_18: # %else27
+; AVX1-NEXT:    testb $2, %al
+; AVX1-NEXT:    je .LBB4_20
+; AVX1-NEXT:  # %bb.19: # %cond.load29
+; AVX1-NEXT:    vpextrq $1, %xmm7, %rcx
+; AVX1-NEXT:    vpinsrd $1, (%rcx), %xmm4, %xmm2
+; AVX1-NEXT:    vblendps {{.*#+}} ymm4 = ymm2[0,1,2,3],ymm4[4,5,6,7]
+; AVX1-NEXT:  .LBB4_20: # %else33
+; AVX1-NEXT:    vpaddq %xmm5, %xmm10, %xmm8
+; AVX1-NEXT:    testb $4, %al
+; AVX1-NEXT:    vextractf128 $1, %ymm7, %xmm7
+; AVX1-NEXT:    je .LBB4_22
+; AVX1-NEXT:  # %bb.21: # %cond.load35
+; AVX1-NEXT:    vmovq %xmm7, %rcx
+; AVX1-NEXT:    vpinsrd $2, (%rcx), %xmm4, %xmm2
+; AVX1-NEXT:    vblendps {{.*#+}} ymm4 = ymm2[0,1,2,3],ymm4[4,5,6,7]
+; AVX1-NEXT:  .LBB4_22: # %else39
+; AVX1-NEXT:    testb $8, %al
+; AVX1-NEXT:    je .LBB4_24
+; AVX1-NEXT:  # %bb.23: # %cond.load41
+; AVX1-NEXT:    vpextrq $1, %xmm7, %rcx
+; AVX1-NEXT:    vpinsrd $3, (%rcx), %xmm4, %xmm2
+; AVX1-NEXT:    vblendps {{.*#+}} ymm4 = ymm2[0,1,2,3],ymm4[4,5,6,7]
+; AVX1-NEXT:  .LBB4_24: # %else45
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm8, %ymm7
+; AVX1-NEXT:    testb $16, %al
+; AVX1-NEXT:    je .LBB4_26
+; AVX1-NEXT:  # %bb.25: # %cond.load47
+; AVX1-NEXT:    vmovq %xmm7, %rcx
+; AVX1-NEXT:    vextractf128 $1, %ymm4, %xmm2
+; AVX1-NEXT:    vpinsrd $0, (%rcx), %xmm2, %xmm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm4, %ymm4
+; AVX1-NEXT:  .LBB4_26: # %else51
+; AVX1-NEXT:    testb $32, %al
+; AVX1-NEXT:    je .LBB4_28
+; AVX1-NEXT:  # %bb.27: # %cond.load53
+; AVX1-NEXT:    vpextrq $1, %xmm7, %rcx
+; AVX1-NEXT:    vextractf128 $1, %ymm4, %xmm2
+; AVX1-NEXT:    vpinsrd $1, (%rcx), %xmm2, %xmm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm4, %ymm4
+; AVX1-NEXT:  .LBB4_28: # %else57
+; AVX1-NEXT:    testb $64, %al
+; AVX1-NEXT:    vextractf128 $1, %ymm7, %xmm7
+; AVX1-NEXT:    je .LBB4_30
+; AVX1-NEXT:  # %bb.29: # %cond.load59
+; AVX1-NEXT:    vmovq %xmm7, %rcx
+; AVX1-NEXT:    vextractf128 $1, %ymm4, %xmm2
+; AVX1-NEXT:    vpinsrd $2, (%rcx), %xmm2, %xmm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm4, %ymm4
+; AVX1-NEXT:  .LBB4_30: # %else63
+; AVX1-NEXT:    testb $-128, %al
+; AVX1-NEXT:    je .LBB4_32
+; AVX1-NEXT:  # %bb.31: # %cond.load65
+; AVX1-NEXT:    vpextrq $1, %xmm7, %rax
+; AVX1-NEXT:    vextractf128 $1, %ymm4, %xmm2
+; AVX1-NEXT:    vpinsrd $3, (%rax), %xmm2, %xmm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm4, %ymm4
+; AVX1-NEXT:  .LBB4_32: # %else69
+; AVX1-NEXT:    vpaddq %xmm5, %xmm10, %xmm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm7
+; AVX1-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm9, %xmm6
+; AVX1-NEXT:    vpcmpeqd %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm6, %ymm0, %ymm0
+; AVX1-NEXT:    vmovmskps %ymm0, %eax
+; AVX1-NEXT:    testb $1, %al
+; AVX1-NEXT:    # implicit-def: $ymm0
+; AVX1-NEXT:    je .LBB4_34
+; AVX1-NEXT:  # %bb.33: # %cond.load72
+; AVX1-NEXT:    vmovq %xmm7, %rcx
+; AVX1-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; AVX1-NEXT:  .LBB4_34: # %else76
+; AVX1-NEXT:    testb $2, %al
+; AVX1-NEXT:    je .LBB4_36
+; AVX1-NEXT:  # %bb.35: # %cond.load78
+; AVX1-NEXT:    vpextrq $1, %xmm7, %rcx
+; AVX1-NEXT:    vpinsrd $1, (%rcx), %xmm0, %xmm2
+; AVX1-NEXT:    vblendps {{.*#+}} ymm0 = ymm2[0,1,2,3],ymm0[4,5,6,7]
+; AVX1-NEXT:  .LBB4_36: # %else82
+; AVX1-NEXT:    vpaddq %xmm5, %xmm10, %xmm2
+; AVX1-NEXT:    testb $4, %al
+; AVX1-NEXT:    vextractf128 $1, %ymm7, %xmm5
+; AVX1-NEXT:    je .LBB4_38
+; AVX1-NEXT:  # %bb.37: # %cond.load84
+; AVX1-NEXT:    vmovq %xmm5, %rcx
+; AVX1-NEXT:    vpinsrd $2, (%rcx), %xmm0, %xmm6
+; AVX1-NEXT:    vblendps {{.*#+}} ymm0 = ymm6[0,1,2,3],ymm0[4,5,6,7]
+; AVX1-NEXT:  .LBB4_38: # %else88
+; AVX1-NEXT:    testb $8, %al
+; AVX1-NEXT:    je .LBB4_40
+; AVX1-NEXT:  # %bb.39: # %cond.load90
+; AVX1-NEXT:    vpextrq $1, %xmm5, %rcx
+; AVX1-NEXT:    vpinsrd $3, (%rcx), %xmm0, %xmm5
+; AVX1-NEXT:    vblendps {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm0[4,5,6,7]
+; AVX1-NEXT:  .LBB4_40: # %else94
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm2, %ymm2
+; AVX1-NEXT:    testb $16, %al
+; AVX1-NEXT:    je .LBB4_42
+; AVX1-NEXT:  # %bb.41: # %cond.load96
+; AVX1-NEXT:    vmovq %xmm2, %rcx
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm3
+; AVX1-NEXT:    vpinsrd $0, (%rcx), %xmm3, %xmm3
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm0
+; AVX1-NEXT:  .LBB4_42: # %else100
+; AVX1-NEXT:    vextractf128 $1, %ymm4, %xmm3
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm5
+; AVX1-NEXT:    vpaddd %xmm4, %xmm1, %xmm1
+; AVX1-NEXT:    testb $32, %al
+; AVX1-NEXT:    je .LBB4_44
+; AVX1-NEXT:  # %bb.43: # %cond.load102
+; AVX1-NEXT:    vpextrq $1, %xmm2, %rcx
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm4
+; AVX1-NEXT:    vpinsrd $1, (%rcx), %xmm4, %xmm4
+; AVX1-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm0
+; AVX1-NEXT:  .LBB4_44: # %else106
+; AVX1-NEXT:    vpaddd %xmm3, %xmm5, %xmm3
+; AVX1-NEXT:    testb $64, %al
+; AVX1-NEXT:    vextractf128 $1, %ymm2, %xmm2
+; AVX1-NEXT:    je .LBB4_46
+; AVX1-NEXT:  # %bb.45: # %cond.load108
+; AVX1-NEXT:    vmovq %xmm2, %rcx
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm4
+; AVX1-NEXT:    vpinsrd $2, (%rcx), %xmm4, %xmm4
+; AVX1-NEXT:    vinsertf128 $1, %xmm4, %ymm0, %ymm0
+; AVX1-NEXT:  .LBB4_46: # %else112
+; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm1, %ymm1
+; AVX1-NEXT:    testb $-128, %al
+; AVX1-NEXT:    je .LBB4_48
+; AVX1-NEXT:  # %bb.47: # %cond.load114
+; AVX1-NEXT:    vpextrq $1, %xmm2, %rax
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
+; AVX1-NEXT:    vpinsrd $3, (%rax), %xmm2, %xmm2
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
+; AVX1-NEXT:  .LBB4_48: # %else118
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm3
+; AVX1-NEXT:    vpaddd %xmm2, %xmm3, %xmm2
+; AVX1-NEXT:    vpaddd %xmm0, %xmm1, %xmm0
+; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm0, %ymm0
+; AVX1-NEXT:    retq
+; AVX1-NEXT:  .LBB4_5: # %cond.load4
+; AVX1-NEXT:    vmovq %xmm6, %rcx
+; AVX1-NEXT:    vpinsrd $2, (%rcx), %xmm1, %xmm5
+; AVX1-NEXT:    vblendps {{.*#+}} ymm1 = ymm5[0,1,2,3],ymm1[4,5,6,7]
+; AVX1-NEXT:    testb $8, %al
+; AVX1-NEXT:    je .LBB4_8
+; AVX1-NEXT:  .LBB4_7: # %cond.load7
+; AVX1-NEXT:    vpextrq $1, %xmm6, %rcx
+; AVX1-NEXT:    vpinsrd $3, (%rcx), %xmm1, %xmm5
+; AVX1-NEXT:    vblendps {{.*#+}} ymm1 = ymm5[0,1,2,3],ymm1[4,5,6,7]
+; AVX1-NEXT:    testb $16, %al
+; AVX1-NEXT:    je .LBB4_10
+; AVX1-NEXT:  .LBB4_9: # %cond.load10
+; AVX1-NEXT:    vmovq %xmm4, %rcx
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm5
+; AVX1-NEXT:    vpinsrd $0, (%rcx), %xmm5, %xmm5
+; AVX1-NEXT:    vinsertf128 $1, %xmm5, %ymm1, %ymm1
+; AVX1-NEXT:    testb $32, %al
+; AVX1-NEXT:    je .LBB4_12
+; AVX1-NEXT:  .LBB4_11: # %cond.load13
+; AVX1-NEXT:    vpextrq $1, %xmm4, %rcx
+; AVX1-NEXT:    vextractf128 $1, %ymm1, %xmm4
+; AVX1-NEXT:    vpinsrd $1, (%rcx), %xmm4, %xmm4
+; AVX1-NEXT:    vinsertf128 $1, %xmm4, %ymm1, %ymm1
+; AVX1-NEXT:    testb $64, %al
+; AVX1-NEXT:    jne .LBB4_13
+; AVX1-NEXT:    jmp .LBB4_14
+;
+; AVX2-LABEL: gather_v8i32_v8i32:
+; AVX2:       # %bb.0:
+; AVX2-NEXT:    movl $c, %eax
+; AVX2-NEXT:    vmovq %rax, %xmm1
+; AVX2-NEXT:    vpbroadcastq %xmm1, %ymm2
+; AVX2-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [12,12,12,12]
+; AVX2-NEXT:    vpaddq %ymm1, %ymm2, %ymm3
+; AVX2-NEXT:    vpxor %xmm1, %xmm1, %xmm1
+; AVX2-NEXT:    vpcmpeqd %ymm1, %ymm0, %ymm1
+; AVX2-NEXT:    vmovmskps %ymm1, %eax
+; AVX2-NEXT:    testb $1, %al
+; AVX2-NEXT:    # implicit-def: $ymm1
+; AVX2-NEXT:    je .LBB4_2
+; AVX2-NEXT:  # %bb.1: # %cond.load
+; AVX2-NEXT:    vmovq %xmm3, %rcx
+; AVX2-NEXT:    vmovd {{.*#+}} xmm1 = mem[0],zero,zero,zero
+; AVX2-NEXT:  .LBB4_2: # %else
+; AVX2-NEXT:    testb $2, %al
+; AVX2-NEXT:    je .LBB4_4
+; AVX2-NEXT:  # %bb.3: # %cond.load1
+; AVX2-NEXT:    vpextrq $1, %xmm3, %rcx
+; AVX2-NEXT:    vpinsrd $1, (%rcx), %xmm1, %xmm4
+; AVX2-NEXT:    vpblendd {{.*#+}} ymm1 = ymm4[0,1,2,3],ymm1[4,5,6,7]
+; AVX2-NEXT:  .LBB4_4: # %else2
+; AVX2-NEXT:    testb $4, %al
+; AVX2-NEXT:    vextracti128 $1, %ymm3, %xmm4
+; AVX2-NEXT:    jne .LBB4_5
+; AVX2-NEXT:  # %bb.6: # %else5
+; AVX2-NEXT:    testb $8, %al
+; AVX2-NEXT:    jne .LBB4_7
+; AVX2-NEXT:  .LBB4_8: # %else8
+; AVX2-NEXT:    testb $16, %al
+; AVX2-NEXT:    jne .LBB4_9
+; AVX2-NEXT:  .LBB4_10: # %else11
+; AVX2-NEXT:    testb $32, %al
+; AVX2-NEXT:    jne .LBB4_11
+; AVX2-NEXT:  .LBB4_12: # %else14
+; AVX2-NEXT:    testb $64, %al
+; AVX2-NEXT:    jne .LBB4_13
+; AVX2-NEXT:  .LBB4_14: # %else17
+; AVX2-NEXT:    testb $-128, %al
+; AVX2-NEXT:    je .LBB4_16
+; AVX2-NEXT:  .LBB4_15: # %cond.load19
+; AVX2-NEXT:    vpextrq $1, %xmm4, %rax
+; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm3
+; AVX2-NEXT:    vpinsrd $3, (%rax), %xmm3, %xmm3
+; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm1, %ymm1
+; AVX2-NEXT:  .LBB4_16: # %else20
+; AVX2-NEXT:    vpxor %xmm4, %xmm4, %xmm4
+; AVX2-NEXT:    vpbroadcastq {{.*#+}} ymm3 = [28,28,28,28]
+; AVX2-NEXT:    vpaddq %ymm3, %ymm2, %ymm3
+; AVX2-NEXT:    vpcmpeqd %ymm4, %ymm0, %ymm2
+; AVX2-NEXT:    vmovmskps %ymm2, %eax
+; AVX2-NEXT:    testb $1, %al
+; AVX2-NEXT:    # implicit-def: $ymm2
+; AVX2-NEXT:    je .LBB4_18
+; AVX2-NEXT:  # %bb.17: # %cond.load23
+; AVX2-NEXT:    vmovq %xmm3, %rcx
+; AVX2-NEXT:    vmovd {{.*#+}} xmm2 = mem[0],zero,zero,zero
+; AVX2-NEXT:  .LBB4_18: # %else27
+; AVX2-NEXT:    testb $2, %al
+; AVX2-NEXT:    je .LBB4_20
+; AVX2-NEXT:  # %bb.19: # %cond.load29
+; AVX2-NEXT:    vpextrq $1, %xmm3, %rcx
+; AVX2-NEXT:    vpinsrd $1, (%rcx), %xmm2, %xmm4
+; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm4[0,1,2,3],ymm2[4,5,6,7]
+; AVX2-NEXT:  .LBB4_20: # %else33
+; AVX2-NEXT:    testb $4, %al
+; AVX2-NEXT:    vextracti128 $1, %ymm3, %xmm4
+; AVX2-NEXT:    jne .LBB4_21
+; AVX2-NEXT:  # %bb.22: # %else39
+; AVX2-NEXT:    testb $8, %al
+; AVX2-NEXT:    jne .LBB4_23
+; AVX2-NEXT:  .LBB4_24: # %else45
+; AVX2-NEXT:    testb $16, %al
+; AVX2-NEXT:    jne .LBB4_25
+; AVX2-NEXT:  .LBB4_26: # %else51
+; AVX2-NEXT:    testb $32, %al
+; AVX2-NEXT:    jne .LBB4_27
+; AVX2-NEXT:  .LBB4_28: # %else57
+; AVX2-NEXT:    testb $64, %al
+; AVX2-NEXT:    jne .LBB4_29
+; AVX2-NEXT:  .LBB4_30: # %else63
+; AVX2-NEXT:    testb $-128, %al
+; AVX2-NEXT:    je .LBB4_32
+; AVX2-NEXT:  .LBB4_31: # %cond.load65
+; AVX2-NEXT:    vpextrq $1, %xmm4, %rax
+; AVX2-NEXT:    vextracti128 $1, %ymm2, %xmm5
+; AVX2-NEXT:    vpinsrd $3, (%rax), %xmm5, %xmm5
+; AVX2-NEXT:    vinserti128 $1, %xmm5, %ymm2, %ymm2
+; AVX2-NEXT:  .LBB4_32: # %else69
+; AVX2-NEXT:    vpxor %xmm5, %xmm5, %xmm5
+; AVX2-NEXT:    vpcmpeqd %ymm5, %ymm0, %ymm0
+; AVX2-NEXT:    vmovmskps %ymm0, %eax
+; AVX2-NEXT:    testb $1, %al
+; AVX2-NEXT:    # implicit-def: $ymm0
+; AVX2-NEXT:    jne .LBB4_33
+; AVX2-NEXT:  # %bb.34: # %else76
+; AVX2-NEXT:    testb $2, %al
+; AVX2-NEXT:    jne .LBB4_35
+; AVX2-NEXT:  .LBB4_36: # %else82
+; AVX2-NEXT:    testb $4, %al
+; AVX2-NEXT:    jne .LBB4_37
+; AVX2-NEXT:  .LBB4_38: # %else88
+; AVX2-NEXT:    testb $8, %al
+; AVX2-NEXT:    jne .LBB4_39
+; AVX2-NEXT:  .LBB4_40: # %else94
+; AVX2-NEXT:    testb $16, %al
+; AVX2-NEXT:    jne .LBB4_41
+; AVX2-NEXT:  .LBB4_42: # %else100
+; AVX2-NEXT:    testb $32, %al
+; AVX2-NEXT:    jne .LBB4_43
+; AVX2-NEXT:  .LBB4_44: # %else106
+; AVX2-NEXT:    testb $64, %al
+; AVX2-NEXT:    je .LBB4_46
+; AVX2-NEXT:  .LBB4_45: # %cond.load108
+; AVX2-NEXT:    vmovq %xmm4, %rcx
+; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm3
+; AVX2-NEXT:    vpinsrd $2, (%rcx), %xmm3, %xmm3
+; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm0
+; AVX2-NEXT:  .LBB4_46: # %else112
+; AVX2-NEXT:    vpaddd %ymm2, %ymm1, %ymm1
+; AVX2-NEXT:    testb $-128, %al
+; AVX2-NEXT:    je .LBB4_48
+; AVX2-NEXT:  # %bb.47: # %cond.load114
+; AVX2-NEXT:    vpextrq $1, %xmm4, %rax
+; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm2
+; AVX2-NEXT:    vpinsrd $3, (%rax), %xmm2, %xmm2
+; AVX2-NEXT:    vinserti128 $1, %xmm2, %ymm0, %ymm0
+; AVX2-NEXT:  .LBB4_48: # %else118
+; AVX2-NEXT:    vpaddd %ymm0, %ymm1, %ymm0
+; AVX2-NEXT:    retq
+; AVX2-NEXT:  .LBB4_5: # %cond.load4
+; AVX2-NEXT:    vmovq %xmm4, %rcx
+; AVX2-NEXT:    vpinsrd $2, (%rcx), %xmm1, %xmm5
+; AVX2-NEXT:    vpblendd {{.*#+}} ymm1 = ymm5[0,1,2,3],ymm1[4,5,6,7]
+; AVX2-NEXT:    testb $8, %al
+; AVX2-NEXT:    je .LBB4_8
+; AVX2-NEXT:  .LBB4_7: # %cond.load7
+; AVX2-NEXT:    vpextrq $1, %xmm4, %rcx
+; AVX2-NEXT:    vpinsrd $3, (%rcx), %xmm1, %xmm5
+; AVX2-NEXT:    vpblendd {{.*#+}} ymm1 = ymm5[0,1,2,3],ymm1[4,5,6,7]
+; AVX2-NEXT:    testb $16, %al
+; AVX2-NEXT:    je .LBB4_10
+; AVX2-NEXT:  .LBB4_9: # %cond.load10
+; AVX2-NEXT:    vmovq %xmm3, %rcx
+; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm5
+; AVX2-NEXT:    vpinsrd $0, (%rcx), %xmm5, %xmm5
+; AVX2-NEXT:    vinserti128 $1, %xmm5, %ymm1, %ymm1
+; AVX2-NEXT:    testb $32, %al
+; AVX2-NEXT:    je .LBB4_12
+; AVX2-NEXT:  .LBB4_11: # %cond.load13
+; AVX2-NEXT:    vpextrq $1, %xmm3, %rcx
+; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm3
+; AVX2-NEXT:    vpinsrd $1, (%rcx), %xmm3, %xmm3
+; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm1, %ymm1
+; AVX2-NEXT:    testb $64, %al
+; AVX2-NEXT:    je .LBB4_14
+; AVX2-NEXT:  .LBB4_13: # %cond.load16
+; AVX2-NEXT:    vmovq %xmm4, %rcx
+; AVX2-NEXT:    vextracti128 $1, %ymm1, %xmm3
+; AVX2-NEXT:    vpinsrd $2, (%rcx), %xmm3, %xmm3
+; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm1, %ymm1
+; AVX2-NEXT:    testb $-128, %al
+; AVX2-NEXT:    jne .LBB4_15
+; AVX2-NEXT:    jmp .LBB4_16
+; AVX2-NEXT:  .LBB4_21: # %cond.load35
+; AVX2-NEXT:    vmovq %xmm4, %rcx
+; AVX2-NEXT:    vpinsrd $2, (%rcx), %xmm2, %xmm5
+; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
+; AVX2-NEXT:    testb $8, %al
+; AVX2-NEXT:    je .LBB4_24
+; AVX2-NEXT:  .LBB4_23: # %cond.load41
+; AVX2-NEXT:    vpextrq $1, %xmm4, %rcx
+; AVX2-NEXT:    vpinsrd $3, (%rcx), %xmm2, %xmm5
+; AVX2-NEXT:    vpblendd {{.*#+}} ymm2 = ymm5[0,1,2,3],ymm2[4,5,6,7]
+; AVX2-NEXT:    testb $16, %al
+; AVX2-NEXT:    je .LBB4_26
+; AVX2-NEXT:  .LBB4_25: # %cond.load47
+; AVX2-NEXT:    vmovq %xmm3, %rcx
+; AVX2-NEXT:    vextracti128 $1, %ymm2, %xmm5
+; AVX2-NEXT:    vpinsrd $0, (%rcx), %xmm5, %xmm5
+; AVX2-NEXT:    vinserti128 $1, %xmm5, %ymm2, %ymm2
+; AVX2-NEXT:    testb $32, %al
+; AVX2-NEXT:    je .LBB4_28
+; AVX2-NEXT:  .LBB4_27: # %cond.load53
+; AVX2-NEXT:    vpextrq $1, %xmm3, %rcx
+; AVX2-NEXT:    vextracti128 $1, %ymm2, %xmm5
+; AVX2-NEXT:    vpinsrd $1, (%rcx), %xmm5, %xmm5
+; AVX2-NEXT:    vinserti128 $1, %xmm5, %ymm2, %ymm2
+; AVX2-NEXT:    testb $64, %al
+; AVX2-NEXT:    je .LBB4_30
+; AVX2-NEXT:  .LBB4_29: # %cond.load59
+; AVX2-NEXT:    vmovq %xmm4, %rcx
+; AVX2-NEXT:    vextracti128 $1, %ymm2, %xmm5
+; AVX2-NEXT:    vpinsrd $2, (%rcx), %xmm5, %xmm5
+; AVX2-NEXT:    vinserti128 $1, %xmm5, %ymm2, %ymm2
+; AVX2-NEXT:    testb $-128, %al
+; AVX2-NEXT:    jne .LBB4_31
+; AVX2-NEXT:    jmp .LBB4_32
+; AVX2-NEXT:  .LBB4_33: # %cond.load72
+; AVX2-NEXT:    vmovq %xmm3, %rcx
+; AVX2-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
+; AVX2-NEXT:    testb $2, %al
+; AVX2-NEXT:    je .LBB4_36
+; AVX2-NEXT:  .LBB4_35: # %cond.load78
+; AVX2-NEXT:    vpextrq $1, %xmm3, %rcx
+; AVX2-NEXT:    vpinsrd $1, (%rcx), %xmm0, %xmm5
+; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm0[4,5,6,7]
+; AVX2-NEXT:    testb $4, %al
+; AVX2-NEXT:    je .LBB4_38
+; AVX2-NEXT:  .LBB4_37: # %cond.load84
+; AVX2-NEXT:    vmovq %xmm4, %rcx
+; AVX2-NEXT:    vpinsrd $2, (%rcx), %xmm0, %xmm5
+; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm0[4,5,6,7]
+; AVX2-NEXT:    testb $8, %al
+; AVX2-NEXT:    je .LBB4_40
+; AVX2-NEXT:  .LBB4_39: # %cond.load90
+; AVX2-NEXT:    vpextrq $1, %xmm4, %rcx
+; AVX2-NEXT:    vpinsrd $3, (%rcx), %xmm0, %xmm5
+; AVX2-NEXT:    vpblendd {{.*#+}} ymm0 = ymm5[0,1,2,3],ymm0[4,5,6,7]
+; AVX2-NEXT:    testb $16, %al
+; AVX2-NEXT:    je .LBB4_42
+; AVX2-NEXT:  .LBB4_41: # %cond.load96
+; AVX2-NEXT:    vmovq %xmm3, %rcx
+; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm5
+; AVX2-NEXT:    vpinsrd $0, (%rcx), %xmm5, %xmm5
+; AVX2-NEXT:    vinserti128 $1, %xmm5, %ymm0, %ymm0
+; AVX2-NEXT:    testb $32, %al
+; AVX2-NEXT:    je .LBB4_44
+; AVX2-NEXT:  .LBB4_43: # %cond.load102
+; AVX2-NEXT:    vpextrq $1, %xmm3, %rcx
+; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm3
+; AVX2-NEXT:    vpinsrd $1, (%rcx), %xmm3, %xmm3
+; AVX2-NEXT:    vinserti128 $1, %xmm3, %ymm0, %ymm0
+; AVX2-NEXT:    testb $64, %al
+; AVX2-NEXT:    jne .LBB4_45
+; AVX2-NEXT:    jmp .LBB4_46
+;
+; AVX512-LABEL: gather_v8i32_v8i32:
+; AVX512:       # %bb.0:
+; AVX512-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
+; AVX512-NEXT:    vptestnmd %zmm0, %zmm0, %k1
+; AVX512-NEXT:    kshiftlw $8, %k1, %k0
+; AVX512-NEXT:    kshiftrw $8, %k0, %k2
+; AVX512-NEXT:    vpbroadcastd {{.*#+}} zmm0 = [3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3]
+; AVX512-NEXT:    vpgatherdd c(,%zmm0,4), %zmm1 {%k2}
+; AVX512-NEXT:    vpbroadcastq $28, %zmm0
+; AVX512-NEXT:    vpgatherqd c(,%zmm0), %ymm2 {%k1}
+; AVX512-NEXT:    vpaddd %ymm2, %ymm2, %ymm0
+; AVX512-NEXT:    vpaddd %ymm0, %ymm1, %ymm0
+; AVX512-NEXT:    retq
+  %1 = insertelement <8 x %struct.a*> undef, %struct.a* @c, i32 0
+  %2 = shufflevector <8 x %struct.a*> %1, <8 x %struct.a*> undef, <8 x i32> zeroinitializer
+  %3 = getelementptr %struct.a, <8 x %struct.a*> %2, <8 x i32> zeroinitializer, i32 0, i32 3
+  %4 = icmp eq <8 x i32> %trigger, zeroinitializer
+  %5 = call <8 x i32> @llvm.masked.gather.v8i32.v8p0i32(<8 x i32*> %3, i32 4, <8 x i1> %4, <8 x i32> undef)
+  %6 = getelementptr %struct.a, <8 x %struct.a*> %2, <8 x i32> zeroinitializer, i32 3
+  %7 = call <8 x i32> @llvm.masked.gather.v8i32.v8p0i32(<8 x i32*> %6, i32 4, <8 x i1> %4, <8 x i32> undef)
+  %8 = add <8 x i32> %5, %7
+  %9 = getelementptr %struct.a, <8 x %struct.a*> %2, i32 0, <8 x i32> <i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3, i32 3>
+  %10 = call <8 x i32> @llvm.masked.gather.v8i32.v8p0i32(<8 x i32*> %9, i32 4, <8 x i1> %4, <8 x i32> undef)
+  %11 = add <8 x i32> %8, %10
+  ret <8 x i32> %11
+}
+
 declare <2 x double> @llvm.masked.gather.v2f64.v2p0f64(<2 x double*>, i32, <2 x i1>, <2 x double>)
 declare <4 x double> @llvm.masked.gather.v4f64.v4p0f64(<4 x double*>, i32, <4 x i1>, <4 x double>)
 
 declare <4 x float> @llvm.masked.gather.v4f32.v4p0f32(<4 x float*>, i32, <4 x i1>, <4 x float>)
 declare <8 x float> @llvm.masked.gather.v8f32.v8p0f32(<8 x float*>, i32, <8 x i1>, <8 x float>)
 
 declare <16 x i8> @llvm.masked.gather.v16i8.v16p0i8(<16 x i8*>, i32, <16 x i1>, <16 x i8>)
+
+declare <8 x i32> @llvm.masked.gather.v8i32.v8p0i32(<8 x i32*>, i32, <8 x i1>, <8 x i32>)