[SelectionDAGBuilder][CGP][X86] Move some of SDB's gather/scatter uniform base handling to CGP.

[craig.topper]

Summary

I've always found the "findValue" a little odd and
inconsistent with other things in SDB.

This simplies the code in SDB to just handle a splat constant
address or a 2 operand GEP in the same BB. This removes the
need for "findValue" since the operands to the GEP are
guaranteed to be available. The splat constant handling is
new, but was needed to avoid regressions due to constant
folding combining GEPs created in CGP.

CGP is now responsible for canonicalizing gather/scatters into
this form. The pattern I'm using for scalarizing, a scalar GEP
followed by a GEP with an all zeroes index, seems to be subject
to constant folding that the insertelement+shufflevector was not.

[craig.topper]

Ping

[spatel]

llvm/lib/CodeGen/CodeGenPrepare.cpp

5191–5192

I haven't looked at this before, so I'm not the best reviewer...
But it would be helpful to any first-time reader if this function had a documentation comment to describe the transform. An IR example in the comment and IR regression tests would also be educational.

5287–5304

Can this copied chunk be lifted to a helper function as a preliminary commit?

[craig.topper]

-Add IR tests
-Add comment to function.
-Simplify node deletion code. We don't have the iterator invalidation problem because we don't walk through phis.

[craig.topper]

Ping

[efriedma]

The general approach here makes sense.

llvm/lib/CodeGen/CodeGenPrepare.cpp

5240

Does it matter whether the value is specifically zero, as opposed to an arbitrary splat value?

5255

I think it would make sense to unify the handling where the last index is a scalar/splat, rather that splitting it based on whether GTI.isStruct() is true.

llvm/lib/IR/Verifier.cpp

2886 ↗

(On Diff #256134)

Not sure this is an improvement.

llvm/test/CodeGen/X86/masked_gather_scatter.ll

664

Orthogonal, but probably you could add a special-case here: vgather with an all-zero vector is equivalent to vbroadcast, I think?

[craig.topper]

llvm/lib/CodeGen/CodeGenPrepare.cpp

5240

It shouldn't matter, but I was just trying to avoid changing too much relative to the existing SelectionDAGBuilder code in this patch.

5255

Good point. I added the scalar/splat case later after facing some regressions.

llvm/lib/IR/Verifier.cpp

2886 ↗

(On Diff #256134)

I think this was from debugging something and I didn't mean to leave it in here. I'll remove it.

[craig.topper]

Remove accidental Verifier change.
Unify struct and splat last index handling.

[craig.topper]

llvm/test/CodeGen/X86/masked_gather_scatter.ll

664

Yep it is. Need to look at whether we need to do that in DAG combine with a new X86ISD opcode to carry the mask or if we can just pattern match it in isel.

[efriedma]

Maybe a weird question, but should we restrict this to masked gather/scatter operands? It seems like this should be profitable for any vector GEP to avoid vector operations where possible.

llvm/lib/CodeGen/CodeGenPrepare.cpp

5231

GTI is unused?

[craig.topper]

Removed GTI

[craig.topper]

In D76947#1985459, @efriedma wrote:

Maybe a weird question, but should we restrict this to masked gather/scatter operands? It seems like this should be profitable for any vector GEP to avoid vector operations where possible.

That's probably true. I've wondered if the GEP visitor in SelectionDAGBuilder should only splat when it finds the first vector index instead of for any scalar component of a vector GEP. Of course that means we need to scalarize splat indices earlier so that SelectionDAGBuilder sees them. Or we could split the GEPs in CGP or earlier like we're doing here. Splitting vector GEPs before LSR could be interesting to give LSR a chance to optimize induction variable uses in the scalar part.

[efriedma]

LGTM

llvm/lib/CodeGen/CodeGenPrepare.cpp

 //===- CodeGenPrepare.cpp - Prepare a function for code generation --------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
private:
     bool canMergeBlocks(const BasicBlock *BB, const BasicBlock *DestBB) const;
     void eliminateMostlyEmptyBlock(BasicBlock *BB);
     bool isMergingEmptyBlockProfitable(BasicBlock *BB, BasicBlock *DestBB,
                                        bool isPreheader);
     bool optimizeBlock(BasicBlock &BB, bool &ModifiedDT);
     bool optimizeInst(Instruction *I, bool &ModifiedDT);
     bool optimizeMemoryInst(Instruction *MemoryInst, Value *Addr,
                             Type *AccessTy, unsigned AddrSpace);
+    bool optimizeGatherScatterInst(Instruction *MemoryInst, Value *Ptr);
     bool optimizeInlineAsmInst(CallInst *CS);
     bool optimizeCallInst(CallInst *CI, bool &ModifiedDT);
     bool optimizeExt(Instruction *&I);
     bool optimizeExtUses(Instruction *I);
     bool optimizeLoadExt(LoadInst *Load);
     bool optimizeShiftInst(BinaryOperator *BO);
     bool optimizeSelectInst(SelectInst *SI);
     bool optimizeShuffleVectorInst(ShuffleVectorInst *SVI);
case Intrinsic::vscale: {
         auto Null = Constant::getNullValue(ScalableVectorTy->getPointerTo());
         auto One = ConstantInt::getSigned(II->getType(), 1);
         auto *CGep =
             ConstantExpr::getGetElementPtr(ScalableVectorTy, Null, One);
         II->replaceAllUsesWith(ConstantExpr::getPtrToInt(CGep, II->getType()));
         II->eraseFromParent();
         return true;
       }
+      break;
     }
+    case Intrinsic::masked_gather:
+      return optimizeGatherScatterInst(II, II->getArgOperand(0));
+    case Intrinsic::masked_scatter:
+      return optimizeGatherScatterInst(II, II->getArgOperand(1));
     }
 
     SmallVector<Value *, 2> PtrOps;
     Type *AccessTy;
     if (TLI->getAddrModeArguments(II, PtrOps, AccessTy))
       while (!PtrOps.empty()) {
         Value *PtrVal = PtrOps.pop_back_val();
         unsigned AS = PtrVal->getType()->getPointerAddressSpace();
if (IterHandle != CurValue) {
       CurInstIterator = BB->begin();
       SunkAddrs.clear();
     }
   }
   ++NumMemoryInsts;
   return true;
 }
 
+/// Rewrite GEP input to gather/scatter to enable SelectionDAGBuilder to find
+/// a uniform base to use for ISD::MGATHER/MSCATTER. SelectionDAGBuilder can
+/// only handle a 2 operand GEP in the same basic block or a splat constant
+/// vector. The 2 operands to the GEP must have a scalar pointer and a vector
+/// index.
+///
+/// If the existing GEP has a vector base pointer that is splat, we can look
+/// through the splat to find the scalar pointer. If we can't find a scalar
+/// pointer there's nothing we can do.
+///
+/// If we have a GEP with more than 2 indices where the middle indices are all
+/// zeroes, we can replace it with 2 GEPs where the second has 2 operands.
+///
+/// If the final index isn't a vector or is a splat, we can emit a scalar GEP
+/// followed by a GEP with an all zeroes vector index. This will enable
+/// SelectionDAGBuilder to use a the scalar GEP as the uniform base and have a
+/// zero index.
+bool CodeGenPrepare::optimizeGatherScatterInst(Instruction *MemoryInst,
+                                               Value *Ptr) {
+  const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
+  if (!GEP || !GEP->hasIndices())
+    return false;
+
+  // If the GEP and the gather/scatter aren't in the same BB, don't optimize.
+  // FIXME: We should support this by sinking the GEP.
+  if (MemoryInst->getParent() != GEP->getParent())
+    return false;
+
+  SmallVector<Value *, 2> Ops(GEP->op_begin(), GEP->op_end());
+
+  bool RewriteGEP = false;
+
+  if (Ops[0]->getType()->isVectorTy()) {
+    Ops[0] = const_cast<Value *>(getSplatValue(Ops[0]));
+    if (!Ops[0])
+      return false;
+    RewriteGEP = true;
+  }
+
+  unsigned FinalIndex = Ops.size() - 1;
+  gep_type_iterator GTI = gep_type_begin(*GEP);
+
+  // Ensure all but the last index is 0.
+  // FIXME: This isn't strictly required. All that's required is that they are
+  // all scalars or splats.
+  for (unsigned i = 1; i < FinalIndex; ++i, ++GTI) {
+    auto *C = dyn_cast<Constant>(Ops[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;
+    // Scalarize the index if needed.
+    Ops[i] = CI;
+  }
+
+  // Try to scalarize the final index.
+  if (Ops[FinalIndex]->getType()->isVectorTy()) {
+    if (Value *V = const_cast<Value *>(getSplatValue(Ops[FinalIndex]))) {
+      auto *C = dyn_cast<ConstantInt>(V);
+      // Don't scalarize all zeros vector.
+      if (!C || !C->isZero()) {
+        Ops[FinalIndex] = V;
+        RewriteGEP = true;
+      }
+    }
+  }
+
+  // If we made any changes or the we have extra operands, we need to generate
+  // new instructions.
+  if (!RewriteGEP && Ops.size() == 2)
+    return false;
+
+  unsigned NumElts = Ptr->getType()->getVectorNumElements();
+
+  IRBuilder<> Builder(MemoryInst);
+
+  Type *ScalarIndexTy = DL->getIndexType(Ops[0]->getType()->getScalarType());
+
+  Value *NewAddr;
+
+  // If the final index isn't a vector, emit a scalar GEP containing all ops
+  // and a vector GEP with all zeroes final index.
+  if (!Ops[FinalIndex]->getType()->isVectorTy()) {
+    NewAddr = Builder.CreateGEP(Ops[0], makeArrayRef(Ops).drop_front());
+    Type *IndexTy = VectorType::get(ScalarIndexTy, NumElts);
+    NewAddr = Builder.CreateGEP(NewAddr, Constant::getNullValue(IndexTy));
+  } else {
+    Value *Base = Ops[0];
+    Value *Index = Ops[FinalIndex];
+
+    // Create a scalar GEP if there are more than 2 operands.
+    if (Ops.size() != 2) {
+      // Replace the last index with 0.
+      Ops[FinalIndex] = Constant::getNullValue(ScalarIndexTy);
+      Base = Builder.CreateGEP(Base, makeArrayRef(Ops).drop_front());
+    }
+
+    // Now create the GEP with scalar pointer and vector index.
+    NewAddr = Builder.CreateGEP(Base, Index);
+  }
+
+  MemoryInst->replaceUsesOfWith(Ptr, NewAddr);
+
+  // If we have no uses, recursively delete the value and all dead instructions
+  // using it.
+  if (Ptr->use_empty())
+    RecursivelyDeleteTriviallyDeadInstructions(Ptr, TLInfo);
+
+  return true;
+}
+
 /// If there are any memory operands, use OptimizeMemoryInst to sink their
 /// address computing into the block when possible / profitable.
 bool CodeGenPrepare::optimizeInlineAsmInst(CallInst *CS) {
   bool MadeChange = false;
 
   const TargetRegisterInfo *TRI =
       TM->getSubtargetImpl(*CS->getFunction())->getRegisterInfo();
   TargetLowering::AsmOperandInfoVector TargetConstraints =

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.h

 //===- SelectionDAGBuilder.h - Selection-DAG building -----------*- C++ -*-===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
public:
   bool handleDebugValue(const Value *V, DILocalVariable *Var,
                         DIExpression *Expr, DebugLoc CurDL,
                         DebugLoc InstDL, unsigned Order);
 
   /// Evict any dangling debug information, attempting to salvage it first.
   void resolveOrClearDbgInfo();
 
   SDValue getValue(const Value *V);
-  bool findValue(const Value *V) const;
 
   /// Return the SDNode for the specified IR value if it exists.
   SDNode *getNodeForIRValue(const Value *V) {
     if (NodeMap.find(V) == NodeMap.end())
       return nullptr;
     return NodeMap[V].getNode();
   }
 

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

 //===- SelectionDAGBuilder.cpp - Selection-DAG building -------------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
SDValue SelectionDAGBuilder::getValue(const Value *V) {
 
   // Otherwise create a new SDValue and remember it.
   SDValue Val = getValueImpl(V);
   NodeMap[V] = Val;
   resolveDanglingDebugInfo(V, Val);
   return Val;
 }
 
-// Return true if SDValue exists for the given Value
-bool SelectionDAGBuilder::findValue(const Value *V) const {
-  return (NodeMap.find(V) != NodeMap.end()) ||
-    (FuncInfo.ValueMap.find(V) != FuncInfo.ValueMap.end());
-}
-
 /// getNonRegisterValue - Return an SDValue for the given Value, but
 /// don't look in FuncInfo.ValueMap for a virtual register.
 SDValue SelectionDAGBuilder::getNonRegisterValue(const Value *V) {
   // If we already have an SDValue for this value, use it.
   SDValue &N = NodeMap[V];
   if (N.getNode()) {
     if (isa<ConstantSDNode>(N) || isa<ConstantFPSDNode>(N)) {
       // Remove the debug location from the node as the node is about to be used

 // %res = call <8 x i32> @llvm.masked.gather.v8i32(<8 x i32*> %gep.ptr, ..
 //
 // When the first GEP operand is a single pointer - it is the uniform base we
 // are looking for. If first operand of the GEP is a splat vector - we
 // extract the splat value and use it as a uniform base.
 // In all other cases the function returns 'false'.
 static bool getUniformBase(const Value *Ptr, SDValue &Base, SDValue &Index,
                            ISD::MemIndexType &IndexType, SDValue &Scale,
-                           SelectionDAGBuilder *SDB) {
+                           SelectionDAGBuilder *SDB, const BasicBlock *CurBB) {
   SelectionDAG& DAG = SDB->DAG;
-  LLVMContext &Context = *DAG.getContext();
+  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+  const DataLayout &DL = DAG.getDataLayout();
 
   assert(Ptr->getType()->isVectorTy() && "Uexpected pointer type");
-  const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
-  if (!GEP)
-    return false;
 
-  const Value *BasePtr = GEP->getPointerOperand();
-  if (BasePtr->getType()->isVectorTy()) {
-    BasePtr = getSplatValue(BasePtr);
-    if (!BasePtr)
-      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, ++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())
+  // Handle splat constant pointer.
+  if (auto *C = dyn_cast<Constant>(Ptr)) {
+    if (!C)
       return false;
+
+    Base = SDB->getValue(C);
+
+    unsigned NumElts = Ptr->getType()->getVectorNumElements();
+    EVT VT = EVT::getVectorVT(*DAG.getContext(), TLI.getPointerTy(DL), NumElts);
+    Index = DAG.getConstant(0, SDB->getCurSDLoc(), VT);
+    IndexType = ISD::SIGNED_SCALED;
+    Scale = DAG.getTargetConstant(1, SDB->getCurSDLoc(), TLI.getPointerTy(DL));
+    return true;
   }
 
-  // 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(BasePtr))
+  const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
+  if (!GEP || GEP->getParent() != CurBB)
     return false;
-  Constant *C = dyn_cast<Constant>(IndexVal);
-  if (!C && !SDB->findValue(IndexVal))
+
+  if (GEP->getNumOperands() != 2)
     return false;
 
-  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
-  const DataLayout &DL = DAG.getDataLayout();
-  StructType *STy = GTI.getStructTypeOrNull();
+  const Value *BasePtr = GEP->getPointerOperand();
+  const Value *IndexVal = GEP->getOperand(GEP->getNumOperands() - 1);
 
-  if (STy) {
-    const StructLayout *SL = DL.getStructLayout(STy);
-    unsigned Field = cast<Constant>(IndexVal)->getUniqueInteger().getZExtValue();
-    Scale = DAG.getTargetConstant(1, SDB->getCurSDLoc(), TLI.getPointerTy(DL));
-    Index = DAG.getConstant(SL->getElementOffset(Field),
-                            SDB->getCurSDLoc(), TLI.getPointerTy(DL));
-  } else {
-    Scale = DAG.getTargetConstant(
-                DL.getTypeAllocSize(GEP->getResultElementType()),
-                SDB->getCurSDLoc(), TLI.getPointerTy(DL));
-    Index = SDB->getValue(IndexVal);
-  }
-  Base = SDB->getValue(BasePtr);
-  IndexType = ISD::SIGNED_SCALED;
-
-  if (STy || !Index.getValueType().isVector()) {
-    unsigned GEPWidth = cast<VectorType>(GEP->getType())->getNumElements();
-    EVT VT = EVT::getVectorVT(Context, Index.getValueType(), GEPWidth);
-    Index = DAG.getSplatBuildVector(VT, SDLoc(Index), Index);
-  }
+  // Make sure the base is scalar and the index is a vector.
+  if (BasePtr->getType()->isVectorTy() || !IndexVal->getType()->isVectorTy())
+    return false;
+
+  Base = SDB->getValue(BasePtr);
+  Index = SDB->getValue(IndexVal);
+  IndexType = ISD::SIGNED_SCALED;
   return true;
 }
 
 void SelectionDAGBuilder::visitMaskedScatter(const CallInst &I) {
   SDLoc sdl = getCurSDLoc();
 
   // llvm.masked.scatter.*(Src0, Ptrs, alignment, Mask)
   const Value *Ptr = I.getArgOperand(1);
   SDValue Src0 = getValue(I.getArgOperand(0));
   SDValue Mask = getValue(I.getArgOperand(3));
   EVT VT = Src0.getValueType();
   MaybeAlign Alignment(cast<ConstantInt>(I.getArgOperand(2))->getZExtValue());
   if (!Alignment)
     Alignment = DAG.getEVTAlign(VT);
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
 
   AAMDNodes AAInfo;
   I.getAAMetadata(AAInfo);
 
   SDValue Base;
   SDValue Index;
   ISD::MemIndexType IndexType;
   SDValue Scale;
-  bool UniformBase = getUniformBase(Ptr, Base, Index, IndexType, Scale, this);
+  bool UniformBase = getUniformBase(Ptr, Base, Index, IndexType, Scale, this,
+                                    I.getParent());
 
   unsigned AS = Ptr->getType()->getScalarType()->getPointerAddressSpace();
   MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
       MachinePointerInfo(AS), MachineMemOperand::MOStore,
       // TODO: Make MachineMemOperands aware of scalable
       // vectors.
       MemoryLocation::UnknownSize, *Alignment, AAInfo);
   if (!UniformBase) {
void SelectionDAGBuilder::visitMaskedGather(const CallInst &I) {
   I.getAAMetadata(AAInfo);
   const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
 
   SDValue Root = DAG.getRoot();
   SDValue Base;
   SDValue Index;
   ISD::MemIndexType IndexType;
   SDValue Scale;
-  bool UniformBase = getUniformBase(Ptr, Base, Index, IndexType, Scale, this);
+  bool UniformBase = getUniformBase(Ptr, Base, Index, IndexType, Scale, this,
+                                    I.getParent());
   unsigned AS = Ptr->getType()->getScalarType()->getPointerAddressSpace();
   MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
       MachinePointerInfo(AS), MachineMemOperand::MOLoad,
       // TODO: Make MachineMemOperands aware of scalable
       // vectors.
       MemoryLocation::UnknownSize, *Alignment, AAInfo, Ranges);
 
   if (!UniformBase) {

llvm/test/CodeGen/X86/masked_gather.ll

 ; 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, %k0
 ; AVX512-NEXT:    kshiftlw $8, %k0, %k0
 ; AVX512-NEXT:    kshiftrw $8, %k0, %k1
-; AVX512-NEXT:    vpbroadcastd {{.*#+}} zmm0 = [3,3,3,3,3,3,3,3,3,3,3,3,3,3,3,3]
+; AVX512-NEXT:    vpxor %xmm0, %xmm0, %xmm0
 ; AVX512-NEXT:    kmovw %k1, %k2
-; AVX512-NEXT:    vpgatherdd c(,%zmm0,4), %zmm1 {%k2}
-; AVX512-NEXT:    vpbroadcastd {{.*#+}} zmm0 = [28,28,28,28,28,28,28,28,28,28,28,28,28,28,28,28]
-; AVX512-NEXT:    vpgatherdd c(,%zmm0), %zmm2 {%k1}
+; AVX512-NEXT:    vpgatherdd c+12(,%zmm0), %zmm1 {%k2}
+; AVX512-NEXT:    vpgatherdd c+28(,%zmm0), %zmm2 {%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)

llvm/test/CodeGen/X86/masked_gather_scatter.ll

entry:
   %res = call <8 x i32 >  @llvm.masked.gather.v8i32.v8p0i32(<8 x i32*>%arrayidx, i32 4, <8 x i1> <i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true>, <8 x i32> undef)
   ret <8 x i32> %res
 }
 
 ; Splat index in GEP, requires broadcast
 define <16 x float> @test11(float* %base, i32 %ind) {
 ; KNL_64-LABEL: test11:
 ; KNL_64:       # %bb.0:
-; KNL_64-NEXT:    vpbroadcastd %esi, %zmm1
+; KNL_64-NEXT:    movslq %esi, %rax
+; KNL_64-NEXT:    leaq (%rdi,%rax,4), %rax
+; KNL_64-NEXT:    vxorps %xmm1, %xmm1, %xmm1
 ; KNL_64-NEXT:    kxnorw %k0, %k0, %k1
-; KNL_64-NEXT:    vgatherdps (%rdi,%zmm1,4), %zmm0 {%k1}
+; KNL_64-NEXT:    vgatherdps (%rax,%zmm1,4), %zmm0 {%k1}
 ; KNL_64-NEXT:    retq
 ;
 ; KNL_32-LABEL: test11:
 ; KNL_32:       # %bb.0:
 ; KNL_32-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; KNL_32-NEXT:    vbroadcastss {{[0-9]+}}(%esp), %zmm1
+; KNL_32-NEXT:    shll $2, %eax
+; KNL_32-NEXT:    addl {{[0-9]+}}(%esp), %eax
+; KNL_32-NEXT:    vxorps %xmm1, %xmm1, %xmm1
 ; KNL_32-NEXT:    kxnorw %k0, %k0, %k1
 ; KNL_32-NEXT:    vgatherdps (%eax,%zmm1,4), %zmm0 {%k1}
 ; KNL_32-NEXT:    retl
 ;
 ; SKX-LABEL: test11:
 ; SKX:       # %bb.0:
-; SKX-NEXT:    vpbroadcastd %esi, %zmm1
+; SKX-NEXT:    movslq %esi, %rax
+; SKX-NEXT:    leaq (%rdi,%rax,4), %rax
+; SKX-NEXT:    vxorps %xmm1, %xmm1, %xmm1
 ; SKX-NEXT:    kxnorw %k0, %k0, %k1
-; SKX-NEXT:    vgatherdps (%rdi,%zmm1,4), %zmm0 {%k1}
+; SKX-NEXT:    vgatherdps (%rax,%zmm1,4), %zmm0 {%k1}
 ; SKX-NEXT:    retq
 ;
 ; SKX_32-LABEL: test11:
 ; SKX_32:       # %bb.0:
 ; SKX_32-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; SKX_32-NEXT:    vbroadcastss {{[0-9]+}}(%esp), %zmm1
+; SKX_32-NEXT:    shll $2, %eax
+; SKX_32-NEXT:    addl {{[0-9]+}}(%esp), %eax
+; SKX_32-NEXT:    vxorps %xmm1, %xmm1, %xmm1
 ; SKX_32-NEXT:    kxnorw %k0, %k0, %k1
 ; SKX_32-NEXT:    vgatherdps (%eax,%zmm1,4), %zmm0 {%k1}
 ; SKX_32-NEXT:    retl
 
   %broadcast.splatinsert = insertelement <16 x float*> undef, float* %base, i32 0
   %broadcast.splat = shufflevector <16 x float*> %broadcast.splatinsert, <16 x float*> undef, <16 x i32> zeroinitializer
 
   %gep.random = getelementptr float, <16 x float*> %broadcast.splat, i32 %ind

llvm/test/CodeGen/X86/pr45067.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=x86_64-apple-darwin -mcpu=skylake | FileCheck %s
 
 @global = external global i32, align 4
 
 define void @foo(<8 x i32>* %x, <8 x i1> %y) {
 ; CHECK-LABEL: foo:
 ; CHECK:       ## %bb.0:
-; CHECK-NEXT:    vpcmpeqd %xmm1, %xmm1, %xmm1
-; CHECK-NEXT:    vpbroadcastq _global@{{.*}}(%rip), %ymm2
-; CHECK-NEXT:    vpgatherqd %xmm1, (,%ymm2), %xmm3
+; CHECK-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
+; CHECK-NEXT:    vpxor %xmm2, %xmm2, %xmm2
+; CHECK-NEXT:    movq _global@{{.*}}(%rip), %rax
+; CHECK-NEXT:    vpgatherdd %ymm1, (%rax,%ymm2), %ymm3
 ; CHECK-NEXT:    vpmovzxwd {{.*#+}} ymm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero,xmm0[4],zero,xmm0[5],zero,xmm0[6],zero,xmm0[7],zero
 ; CHECK-NEXT:    vpslld $31, %ymm0, %ymm0
-; CHECK-NEXT:    vinserti128 $1, %xmm3, %ymm3, %ymm1
-; CHECK-NEXT:    vpmaskmovd %ymm1, %ymm0, (%rdi)
+; CHECK-NEXT:    vpmaskmovd %ymm3, %ymm0, (%rdi)
 ; CHECK-NEXT:    ud2
   %tmp = call <8 x i32> @llvm.masked.gather.v8i32.v8p0i32(<8 x i32*> <i32* @global, i32* @global, i32* @global, i32* @global, i32* @global, i32* @global, i32* @global, i32* @global>, i32 4, <8 x i1> <i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true>, <8 x i32> undef)
   call void @llvm.masked.store.v8i32.p0v8i32(<8 x i32> %tmp, <8 x i32>* %x, i32 4, <8 x i1> %y)
   unreachable
 }
 
 declare <8 x i32> @llvm.masked.gather.v8i32.v8p0i32(<8 x i32*>, i32, <8 x i1>, <8 x i32>)
 declare void @llvm.masked.store.v8i32.p0v8i32(<8 x i32>, <8 x i32>*, i32, <8 x i1>)