[SelectionDAG][X86] Explicitly store the scale in the gather/scatter ISD nodes

[craig.topper]

Summary

Currently we infer the scale at isel time by analyzing whether the base is a constant 0 or not. If it is we assume scale is 1, else we take it from the element size of the pass thru or stored value. This seems a little weird and I think it makes more sense to make it explicit in the DAG rather than doing tricky things in the backend.

Most of this patch is just making sure we copy the scale around everywhere.

The test changes are interesting because I believe the current codegen is actually a bug. We seem to be promote v2i32 gathers to v2i64 and double the scale. But really the address calculation shouldn't have been affected. This is a consquence of getting the scale from the data size. With it stored explicitly it comes from the original GEP. I don't think we should be promoting v2i32 gathers to v2i64, the IR we started with only said it was legal to read 32 bits at each index. If one of the elements is at the end of an array near a page boundary the extra data read might cause a fault. I'll look into a fixing this as a separate patch.

[delena]

test/CodeGen/X86/masked_gather_scatter.ll

1275 ↗

(On Diff #122931)

gatherQQ is a bug, we should read dwords. And scale 4 does not help. I've fixed this bug in some cases, probably not in all.
As far as I remember, the community voted against the scale - they said that I'm trying to push X86 specific things into the common nodes.
So I waited until instruction selection in push the scale. And I created X86MaskedGather to cover all other things, mainly <2 x i32> types.

[craig.topper]

Yes I'm going to fix the QQ issue as well.

The scale would have been part of the handling of the last index of the GEP if we hadn't split it. So from that perspective its not X86 specific. Its part of the GEP calculation. Do you recall who objected? Maybe we can revisit.

[craig.topper]

Ping

[craig.topper]

Alternatively, I wonder if it would be better to have a target hook to directly turn gather into a target specific ISD node. The generic type legalization code does have some trouble with the index argument as it is. We're currently doing various custom things to avoid certain cases.

[craig.topper]

@hfinkel do you have any opinions on this?

[reames]

This looks like the right approach to me. I think the previously expressed concern was about making the IR representation overly specific to what current hardware supports. This looks largely orthogonal to me and much less confusing as well.

[hfinkel]

In D40055#970447, @reames wrote:

This looks like the right approach to me. I think the previously expressed concern was about making the IR representation overly specific to what current hardware supports. This looks largely orthogonal to me and much less confusing as well.

I agree.

[RKSimon]

lib/CodeGen/SelectionDAG/SelectionDAG.cpp

6234 ↗

(On Diff #122931)

Should we assert that Scale op is a positive (pow2?) constant?

[craig.topper]

Rebase and add the asserts Simon requested

[zvi]

lib/Target/X86/X86ISelDAGToDAG.cpp

1533 ↗

(On Diff #129151)

Remove the line above?

[craig.topper]

Remove unneeded line that snuck in during rebase.

[zvi]

There are some occurrences of calls to getMaskedGather in DAGCombine.cpp which i do not see being addressed by this patch. I guess they are not being covered by tests?

[craig.topper]

Fix the calls in DAG combine and add test cases for it.

llvm/trunk/include/llvm/CodeGen/SelectionDAGNodes.h

public:
   friend class SelectionDAG;
 
   MaskedGatherScatterSDNode(ISD::NodeType NodeTy, unsigned Order,
                             const DebugLoc &dl, SDVTList VTs, EVT MemVT,
                             MachineMemOperand *MMO)
       : MemSDNode(NodeTy, Order, dl, VTs, MemVT, MMO) {}
 
   // In the both nodes address is Op1, mask is Op2:
-  // MaskedGatherSDNode  (Chain, src0, mask, base, index), src0 is a passthru value
-  // MaskedScatterSDNode (Chain, value, mask, base, index)
+  // MaskedGatherSDNode  (Chain, passthru, mask, base, index, scale)
+  // MaskedScatterSDNode (Chain, value, mask, base, index, scale)
   // Mask is a vector of i1 elements
   const SDValue &getBasePtr() const { return getOperand(3); }
   const SDValue &getIndex()   const { return getOperand(4); }
   const SDValue &getMask()    const { return getOperand(2); }
   const SDValue &getValue()   const { return getOperand(1); }
+  const SDValue &getScale()   const { return getOperand(5); }
 
   static bool classof(const SDNode *N) {
     return N->getOpcode() == ISD::MGATHER ||
            N->getOpcode() == ISD::MSCATTER;
   }
 };
 
 /// This class is used to represent an MGATHER node

llvm/trunk/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

SDValue DAGCombiner::visitMSCATTER(SDNode *N) {
   unsigned Alignment = MSC->getOriginalAlignment();
 
   EVT LoMemVT, HiMemVT;
   std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
 
   SDValue DataLo, DataHi;
   std::tie(DataLo, DataHi) = DAG.SplitVector(Data, DL);
 
+  SDValue Scale = MSC->getScale();
   SDValue BasePtr = MSC->getBasePtr();
   SDValue IndexLo, IndexHi;
   std::tie(IndexLo, IndexHi) = DAG.SplitVector(MSC->getIndex(), DL);
 
   MachineMemOperand *MMO = DAG.getMachineFunction().
     getMachineMemOperand(MSC->getPointerInfo(),
                           MachineMemOperand::MOStore,  LoMemVT.getStoreSize(),
                           Alignment, MSC->getAAInfo(), MSC->getRanges());
 
-  SDValue OpsLo[] = { Chain, DataLo, MaskLo, BasePtr, IndexLo };
+  SDValue OpsLo[] = { Chain, DataLo, MaskLo, BasePtr, IndexLo, Scale };
   Lo = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), DataLo.getValueType(),
                             DL, OpsLo, MMO);
 
-  SDValue OpsHi[] = {Chain, DataHi, MaskHi, BasePtr, IndexHi};
+  SDValue OpsHi[] = { Chain, DataHi, MaskHi, BasePtr, IndexHi, Scale };
   Hi = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), DataHi.getValueType(),
                             DL, OpsHi, MMO);
 
   AddToWorklist(Lo.getNode());
   AddToWorklist(Hi.getNode());
 
   return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo, Hi);
 }
SDValue DAGCombiner::visitMGATHER(SDNode *N) {
 
   SDValue Chain = MGT->getChain();
   EVT MemoryVT = MGT->getMemoryVT();
   unsigned Alignment = MGT->getOriginalAlignment();
 
   EVT LoMemVT, HiMemVT;
   std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
 
+  SDValue Scale = MGT->getScale();
   SDValue BasePtr = MGT->getBasePtr();
   SDValue Index = MGT->getIndex();
   SDValue IndexLo, IndexHi;
   std::tie(IndexLo, IndexHi) = DAG.SplitVector(Index, DL);
 
   MachineMemOperand *MMO = DAG.getMachineFunction().
     getMachineMemOperand(MGT->getPointerInfo(),
                           MachineMemOperand::MOLoad,  LoMemVT.getStoreSize(),
                           Alignment, MGT->getAAInfo(), MGT->getRanges());
 
-  SDValue OpsLo[] = { Chain, Src0Lo, MaskLo, BasePtr, IndexLo };
+  SDValue OpsLo[] = { Chain, Src0Lo, MaskLo, BasePtr, IndexLo, Scale };
   Lo = DAG.getMaskedGather(DAG.getVTList(LoVT, MVT::Other), LoVT, DL, OpsLo,
-                            MMO);
 
-  SDValue OpsHi[] = {Chain, Src0Hi, MaskHi, BasePtr, IndexHi};
+  SDValue OpsHi[] = { Chain, Src0Hi, MaskHi, BasePtr, IndexHi, Scale };
   Hi = DAG.getMaskedGather(DAG.getVTList(HiVT, MVT::Other), HiVT, DL, OpsHi,
-                            MMO);
 
   AddToWorklist(Lo.getNode());
   AddToWorklist(Hi.getNode());
 
   // Build a factor node to remember that this load is independent of the
   // other one.
   Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, Lo.getValue(1),
                       Hi.getValue(1));

llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp

 SDValue DAGTypeLegalizer::PromoteIntRes_MGATHER(MaskedGatherSDNode *N) {
   EVT NVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
   SDValue ExtSrc0 = GetPromotedInteger(N->getValue());
   assert(NVT == ExtSrc0.getValueType() &&
       "Gather result type and the passThru agrument type should be the same");
 
   SDLoc dl(N);
   SDValue Ops[] = {N->getChain(), ExtSrc0, N->getMask(), N->getBasePtr(),
-                   N->getIndex()};
+                   N->getIndex(), N->getScale() };
   SDValue Res = DAG.getMaskedGather(DAG.getVTList(NVT, MVT::Other),
                                     N->getMemoryVT(), dl, Ops,
                                     N->getMemOperand());
   // Legalize the chain result - switch anything that used the old chain to
   // use the new one.
   ReplaceValueWith(SDValue(N, 1), Res.getValue(1));
   return Res;
 }

llvm/trunk/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

void DAGTypeLegalizer::SplitVecRes_MGATHER(MaskedGatherSDNode *MGT,
   SDLoc dl(MGT);
   std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MGT->getValueType(0));
 
   SDValue Ch = MGT->getChain();
   SDValue Ptr = MGT->getBasePtr();
   SDValue Mask = MGT->getMask();
   SDValue Src0 = MGT->getValue();
   SDValue Index = MGT->getIndex();
+  SDValue Scale = MGT->getScale();
   unsigned Alignment = MGT->getOriginalAlignment();
 
   // Split Mask operand
   SDValue MaskLo, MaskHi;
   if (getTypeAction(Mask.getValueType()) == TargetLowering::TypeSplitVector)
     GetSplitVector(Mask, MaskLo, MaskHi);
   else
     std::tie(MaskLo, MaskHi) = DAG.SplitVector(Mask, dl);
void DAGTypeLegalizer::SplitVecRes_MGATHER(MaskedGatherSDNode *MGT,
   else
     std::tie(IndexLo, IndexHi) = DAG.SplitVector(Index, dl);
 
   MachineMemOperand *MMO = DAG.getMachineFunction().
     getMachineMemOperand(MGT->getPointerInfo(),
                          MachineMemOperand::MOLoad,  LoMemVT.getStoreSize(),
                          Alignment, MGT->getAAInfo(), MGT->getRanges());
 
-  SDValue OpsLo[] = {Ch, Src0Lo, MaskLo, Ptr, IndexLo};
+  SDValue OpsLo[] = {Ch, Src0Lo, MaskLo, Ptr, IndexLo, Scale};
   Lo = DAG.getMaskedGather(DAG.getVTList(LoVT, MVT::Other), LoVT, dl, OpsLo,
                            MMO);
 
-  SDValue OpsHi[] = {Ch, Src0Hi, MaskHi, Ptr, IndexHi};
+  SDValue OpsHi[] = {Ch, Src0Hi, MaskHi, Ptr, IndexHi, Scale};
   Hi = DAG.getMaskedGather(DAG.getVTList(HiVT, MVT::Other), HiVT, dl, OpsHi,
                            MMO);
 
   // Build a factor node to remember that this load is independent of the
   // other one.
   Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
                    Hi.getValue(1));
 
SDValue DAGTypeLegalizer::SplitVecOp_MGATHER(MaskedGatherSDNode *MGT,
                                              unsigned OpNo) {
   EVT LoVT, HiVT;
   SDLoc dl(MGT);
   std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MGT->getValueType(0));
 
   SDValue Ch = MGT->getChain();
   SDValue Ptr = MGT->getBasePtr();
   SDValue Index = MGT->getIndex();
+  SDValue Scale = MGT->getScale();
   SDValue Mask = MGT->getMask();
   SDValue Src0 = MGT->getValue();
   unsigned Alignment = MGT->getOriginalAlignment();
 
   SDValue MaskLo, MaskHi;
   if (getTypeAction(Mask.getValueType()) == TargetLowering::TypeSplitVector)
     // Split Mask operand
     GetSplitVector(Mask, MaskLo, MaskHi);
SDValue DAGTypeLegalizer::SplitVecOp_MGATHER(MaskedGatherSDNode *MGT,
   else
     std::tie(IndexLo, IndexHi) = DAG.SplitVector(Index, dl);
 
   MachineMemOperand *MMO = DAG.getMachineFunction().
     getMachineMemOperand(MGT->getPointerInfo(),
                          MachineMemOperand::MOLoad,  LoMemVT.getStoreSize(),
                          Alignment, MGT->getAAInfo(), MGT->getRanges());
 
-  SDValue OpsLo[] = {Ch, Src0Lo, MaskLo, Ptr, IndexLo};
+  SDValue OpsLo[] = {Ch, Src0Lo, MaskLo, Ptr, IndexLo, Scale};
   SDValue Lo = DAG.getMaskedGather(DAG.getVTList(LoVT, MVT::Other), LoVT, dl,
                                    OpsLo, MMO);
 
   MMO = DAG.getMachineFunction().
     getMachineMemOperand(MGT->getPointerInfo(),
                          MachineMemOperand::MOLoad,  HiMemVT.getStoreSize(),
                          Alignment, MGT->getAAInfo(),
                          MGT->getRanges());
 
-  SDValue OpsHi[] = {Ch, Src0Hi, MaskHi, Ptr, IndexHi};
+  SDValue OpsHi[] = {Ch, Src0Hi, MaskHi, Ptr, IndexHi, Scale};
   SDValue Hi = DAG.getMaskedGather(DAG.getVTList(HiVT, MVT::Other), HiVT, dl,
                                    OpsHi, MMO);
 
   // Build a factor node to remember that this load is independent of the
   // other one.
   Ch = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Lo.getValue(1),
                    Hi.getValue(1));
 

 }
 
 SDValue DAGTypeLegalizer::SplitVecOp_MSCATTER(MaskedScatterSDNode *N,
                                               unsigned OpNo) {
   SDValue Ch  = N->getChain();
   SDValue Ptr = N->getBasePtr();
   SDValue Mask = N->getMask();
   SDValue Index = N->getIndex();
+  SDValue Scale = N->getScale();
   SDValue Data = N->getValue();
   EVT MemoryVT = N->getMemoryVT();
   unsigned Alignment = N->getOriginalAlignment();
   SDLoc DL(N);
 
   // Split all operands
   EVT LoMemVT, HiMemVT;
   std::tie(LoMemVT, HiMemVT) = DAG.GetSplitDestVTs(MemoryVT);
else
     std::tie(IndexLo, IndexHi) = DAG.SplitVector(Index, DL);
 
   SDValue Lo;
   MachineMemOperand *MMO = DAG.getMachineFunction().
     getMachineMemOperand(N->getPointerInfo(),
                          MachineMemOperand::MOStore, LoMemVT.getStoreSize(),
                          Alignment, N->getAAInfo(), N->getRanges());
 
-  SDValue OpsLo[] = {Ch, DataLo, MaskLo, Ptr, IndexLo};
+  SDValue OpsLo[] = {Ch, DataLo, MaskLo, Ptr, IndexLo, Scale};
   Lo = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), DataLo.getValueType(),
                             DL, OpsLo, MMO);
 
   MMO = DAG.getMachineFunction().
     getMachineMemOperand(N->getPointerInfo(),
                          MachineMemOperand::MOStore,  HiMemVT.getStoreSize(),
                          Alignment, N->getAAInfo(), N->getRanges());
 
   // The order of the Scatter operation after split is well defined. The "Hi"
   // part comes after the "Lo". So these two operations should be chained one
   // after another.
-  SDValue OpsHi[] = {Lo, DataHi, MaskHi, Ptr, IndexHi};
+  SDValue OpsHi[] = {Lo, DataHi, MaskHi, Ptr, IndexHi, Scale};
   return DAG.getMaskedScatter(DAG.getVTList(MVT::Other), DataHi.getValueType(),
                               DL, OpsHi, MMO);
 }
 
 SDValue DAGTypeLegalizer::SplitVecOp_STORE(StoreSDNode *N, unsigned OpNo) {
   assert(N->isUnindexed() && "Indexed store of vector?");
   assert(OpNo == 1 && "Can only split the stored value");
   SDLoc DL(N);

 }
 
 SDValue DAGTypeLegalizer::WidenVecRes_MGATHER(MaskedGatherSDNode *N) {
 
   EVT WideVT = TLI.getTypeToTransformTo(*DAG.getContext(), N->getValueType(0));
   SDValue Mask = N->getMask();
   EVT MaskVT = Mask.getValueType();
   SDValue Src0 = GetWidenedVector(N->getValue());
+  SDValue Scale = N->getScale();
   unsigned NumElts = WideVT.getVectorNumElements();
   SDLoc dl(N);
 
   // The mask should be widened as well
   EVT WideMaskVT = EVT::getVectorVT(*DAG.getContext(),
                                     MaskVT.getVectorElementType(),
                                     WideVT.getVectorNumElements());
   Mask = ModifyToType(Mask, WideMaskVT, true);
 
   // Widen the Index operand
   SDValue Index = N->getIndex();
   EVT WideIndexVT = EVT::getVectorVT(*DAG.getContext(),
                                      Index.getValueType().getScalarType(),
                                      NumElts);
   Index = ModifyToType(Index, WideIndexVT);
-  SDValue Ops[] = { N->getChain(), Src0, Mask, N->getBasePtr(), Index };
+  SDValue Ops[] = { N->getChain(), Src0, Mask, N->getBasePtr(), Index, Scale };
   SDValue Res = DAG.getMaskedGather(DAG.getVTList(WideVT, MVT::Other),
                                     N->getMemoryVT(), dl, Ops,
                                     N->getMemOperand());
 
   // Legalize the chain result - switch anything that used the old chain to
   // use the new one.
   ReplaceValueWith(SDValue(N, 1), Res.getValue(1));
   return Res;

 }
 
 SDValue DAGTypeLegalizer::WidenVecOp_MSCATTER(SDNode *N, unsigned OpNo) {
   assert(OpNo == 1 && "Can widen only data operand of mscatter");
   MaskedScatterSDNode *MSC = cast<MaskedScatterSDNode>(N);
   SDValue DataOp = MSC->getValue();
   SDValue Mask = MSC->getMask();
   EVT MaskVT = Mask.getValueType();
+  SDValue Scale = MSC->getScale();
 
   // Widen the value.
   SDValue WideVal = GetWidenedVector(DataOp);
   EVT WideVT = WideVal.getValueType();
   unsigned NumElts = WideVT.getVectorNumElements();
   SDLoc dl(N);
 
   // The mask should be widened as well.
   EVT WideMaskVT = EVT::getVectorVT(*DAG.getContext(),
                                     MaskVT.getVectorElementType(), NumElts);
   Mask = ModifyToType(Mask, WideMaskVT, true);
 
   // Widen index.
   SDValue Index = MSC->getIndex();
   EVT WideIndexVT = EVT::getVectorVT(*DAG.getContext(),
                                      Index.getValueType().getScalarType(),
                                      NumElts);
   Index = ModifyToType(Index, WideIndexVT);
 
-  SDValue Ops[] = {MSC->getChain(), WideVal, Mask, MSC->getBasePtr(), Index};
+  SDValue Ops[] = {MSC->getChain(), WideVal, Mask, MSC->getBasePtr(), Index,
+                   Scale};
   return DAG.getMaskedScatter(DAG.getVTList(MVT::Other),
                               MSC->getMemoryVT(), dl, Ops,
                               MSC->getMemOperand());
 }
 
 SDValue DAGTypeLegalizer::WidenVecOp_SETCC(SDNode *N) {
   SDValue InOp0 = GetWidenedVector(N->getOperand(0));
   SDValue InOp1 = GetWidenedVector(N->getOperand(1));

llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

SDValue SelectionDAG::getMaskedStore(SDValue Chain, const SDLoc &dl,
   SDValue V(N, 0);
   NewSDValueDbgMsg(V, "Creating new node: ", this);
   return V;
 }
 
 SDValue SelectionDAG::getMaskedGather(SDVTList VTs, EVT VT, const SDLoc &dl,
                                       ArrayRef<SDValue> Ops,
                                       MachineMemOperand *MMO) {
-  assert(Ops.size() == 5 && "Incompatible number of operands");
+  assert(Ops.size() == 6 && "Incompatible number of operands");
 
   FoldingSetNodeID ID;
   AddNodeIDNode(ID, ISD::MGATHER, VTs, Ops);
   ID.AddInteger(VT.getRawBits());
   ID.AddInteger(getSyntheticNodeSubclassData<MaskedGatherSDNode>(
       dl.getIROrder(), VTs, VT, MMO));
   ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
   void *IP = nullptr;
SDValue SelectionDAG::getMaskedGather(SDVTList VTs, EVT VT, const SDLoc &dl,
   assert(N->getValue().getValueType() == N->getValueType(0) &&
          "Incompatible type of the PassThru value in MaskedGatherSDNode");
   assert(N->getMask().getValueType().getVectorNumElements() ==
              N->getValueType(0).getVectorNumElements() &&
          "Vector width mismatch between mask and data");
   assert(N->getIndex().getValueType().getVectorNumElements() ==
              N->getValueType(0).getVectorNumElements() &&
          "Vector width mismatch between index and data");
+  assert(isa<ConstantSDNode>(N->getScale()) &&
+         cast<ConstantSDNode>(N->getScale())->getAPIntValue().isPowerOf2() &&
+         "Scale should be a constant power of 2");
 
   CSEMap.InsertNode(N, IP);
   InsertNode(N);
   SDValue V(N, 0);
   NewSDValueDbgMsg(V, "Creating new node: ", this);
   return V;
 }
 
 SDValue SelectionDAG::getMaskedScatter(SDVTList VTs, EVT VT, const SDLoc &dl,
                                        ArrayRef<SDValue> Ops,
                                        MachineMemOperand *MMO) {
-  assert(Ops.size() == 5 && "Incompatible number of operands");
+  assert(Ops.size() == 6 && "Incompatible number of operands");
 
   FoldingSetNodeID ID;
   AddNodeIDNode(ID, ISD::MSCATTER, VTs, Ops);
   ID.AddInteger(VT.getRawBits());
   ID.AddInteger(getSyntheticNodeSubclassData<MaskedScatterSDNode>(
       dl.getIROrder(), VTs, VT, MMO));
   ID.AddInteger(MMO->getPointerInfo().getAddrSpace());
   void *IP = nullptr;
   if (SDNode *E = FindNodeOrInsertPos(ID, dl, IP)) {
     cast<MaskedScatterSDNode>(E)->refineAlignment(MMO);
     return SDValue(E, 0);
   }
   auto *N = newSDNode<MaskedScatterSDNode>(dl.getIROrder(), dl.getDebugLoc(),
                                            VTs, VT, MMO);
   createOperands(N, Ops);
 
   assert(N->getMask().getValueType().getVectorNumElements() ==
              N->getValue().getValueType().getVectorNumElements() &&
          "Vector width mismatch between mask and data");
   assert(N->getIndex().getValueType().getVectorNumElements() ==
              N->getValue().getValueType().getVectorNumElements() &&
          "Vector width mismatch between index and data");
+  assert(isa<ConstantSDNode>(N->getScale()) &&
+         cast<ConstantSDNode>(N->getScale())->getAPIntValue().isPowerOf2() &&
+         "Scale should be a constant power of 2");
 
   CSEMap.InsertNode(N, IP);
   InsertNode(N);
   SDValue V(N, 0);
   NewSDValueDbgMsg(V, "Creating new node: ", this);
   return V;
 }
 

llvm/trunk/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

 //   %gep.ptr = getelementptr i32, i32* %ptr,        <8 x i32> %ind
 // %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,
-                           SelectionDAGBuilder* SDB) {
+                           SDValue &Scale, SelectionDAGBuilder* SDB) {
   SelectionDAG& DAG = SDB->DAG;
   LLVMContext &Context = *DAG.getContext();
 
   assert(Ptr->getType()->isVectorTy() && "Uexpected pointer type");
   const GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr);
   if (!GEP)
     return false;
 
if (!C || !C->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))
     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);
   Index = SDB->getValue(IndexVal);
 
   if (!Index.getValueType().isVector()) {
     unsigned GEPWidth = GEP->getType()->getVectorNumElements();
     EVT VT = EVT::getVectorVT(Context, Index.getValueType(), GEPWidth);
     Index = DAG.getSplatBuildVector(VT, SDLoc(Index), Index);
   }
if (!Alignment)
     Alignment = DAG.getEVTAlignment(VT);
   const TargetLowering &TLI = DAG.getTargetLoweringInfo();
 
   AAMDNodes AAInfo;
   I.getAAMetadata(AAInfo);
 
   SDValue Base;
   SDValue Index;
+  SDValue Scale;
   const Value *BasePtr = Ptr;
-  bool UniformBase = getUniformBase(BasePtr, Base, Index, this);
+  bool UniformBase = getUniformBase(BasePtr, Base, Index, Scale, this);
 
   const Value *MemOpBasePtr = UniformBase ? BasePtr : nullptr;
   MachineMemOperand *MMO = DAG.getMachineFunction().
     getMachineMemOperand(MachinePointerInfo(MemOpBasePtr),
                          MachineMemOperand::MOStore,  VT.getStoreSize(),
                          Alignment, AAInfo);
   if (!UniformBase) {
-    Base = DAG.getTargetConstant(0, sdl, TLI.getPointerTy(DAG.getDataLayout()));
+    Base = DAG.getConstant(0, sdl, TLI.getPointerTy(DAG.getDataLayout()));
     Index = getValue(Ptr);
+    Scale = DAG.getTargetConstant(1, sdl, TLI.getPointerTy(DAG.getDataLayout()));
   }
-  SDValue Ops[] = { getRoot(), Src0, Mask, Base, Index };
+  SDValue Ops[] = { getRoot(), Src0, Mask, Base, Index, Scale };
   SDValue Scatter = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), VT, sdl,
                                          Ops, MMO);
   DAG.setRoot(Scatter);
   setValue(&I, Scatter);
 }
 
 void SelectionDAGBuilder::visitMaskedLoad(const CallInst &I, bool IsExpanding) {
   SDLoc sdl = getCurSDLoc();
void SelectionDAGBuilder::visitMaskedGather(const CallInst &I) {
 
   AAMDNodes AAInfo;
   I.getAAMetadata(AAInfo);
   const MDNode *Ranges = I.getMetadata(LLVMContext::MD_range);
 
   SDValue Root = DAG.getRoot();
   SDValue Base;
   SDValue Index;
+  SDValue Scale;
   const Value *BasePtr = Ptr;
-  bool UniformBase = getUniformBase(BasePtr, Base, Index, this);
+  bool UniformBase = getUniformBase(BasePtr, Base, Index, Scale, this);
   bool ConstantMemory = false;
   if (UniformBase &&
       AA && AA->pointsToConstantMemory(MemoryLocation(
           BasePtr, DAG.getDataLayout().getTypeStoreSize(I.getType()),
           AAInfo))) {
     // Do not serialize (non-volatile) loads of constant memory with anything.
     Root = DAG.getEntryNode();
     ConstantMemory = true;
   }
 
   MachineMemOperand *MMO =
     DAG.getMachineFunction().
     getMachineMemOperand(MachinePointerInfo(UniformBase ? BasePtr : nullptr),
                          MachineMemOperand::MOLoad,  VT.getStoreSize(),
                          Alignment, AAInfo, Ranges);
 
   if (!UniformBase) {
-    Base = DAG.getTargetConstant(0, sdl, TLI.getPointerTy(DAG.getDataLayout()));
+    Base = DAG.getConstant(0, sdl, TLI.getPointerTy(DAG.getDataLayout()));
     Index = getValue(Ptr);
+    Scale = DAG.getTargetConstant(1, sdl, TLI.getPointerTy(DAG.getDataLayout()));
   }
-  SDValue Ops[] = { Root, Src0, Mask, Base, Index };
+  SDValue Ops[] = { Root, Src0, Mask, Base, Index, Scale };
   SDValue Gather = DAG.getMaskedGather(DAG.getVTList(VT, MVT::Other), VT, sdl,
                                        Ops, MMO);
 
   SDValue OutChain = Gather.getValue(1);
   if (!ConstantMemory)
     PendingLoads.push_back(OutChain);
   setValue(&I, Gather);
 }

llvm/trunk/lib/Target/X86/X86ISelDAGToDAG.cpp

 }
 
 /// Helper for selectVectorAddr. Handles things that can be folded into a
 /// gather scatter address. The index register and scale should have already
 /// been handled.
 bool X86DAGToDAGISel::matchVectorAddress(SDValue N, X86ISelAddressMode &AM) {
   // TODO: Support other operations.
   switch (N.getOpcode()) {
+  case ISD::Constant: {
+    uint64_t Val = cast<ConstantSDNode>(N)->getSExtValue();
+    if (!foldOffsetIntoAddress(Val, AM))
+      return false;
+    break;
+  }
   case X86ISD::Wrapper:
     if (!matchWrapper(N, AM))
       return false;
     break;
   }
 
   return matchAddressBase(N, AM);
 }
 
 bool X86DAGToDAGISel::selectVectorAddr(SDNode *Parent, SDValue N, SDValue &Base,
                                        SDValue &Scale, SDValue &Index,
                                        SDValue &Disp, SDValue &Segment) {
   X86ISelAddressMode AM;
   auto *Mgs = cast<X86MaskedGatherScatterSDNode>(Parent);
   AM.IndexReg = Mgs->getIndex();
-  AM.Scale = Mgs->getValue().getScalarValueSizeInBits() / 8;
+  AM.Scale = cast<ConstantSDNode>(Mgs->getScale())->getZExtValue();
 
   unsigned AddrSpace = cast<MemSDNode>(Parent)->getPointerInfo().getAddrSpace();
   // AddrSpace 256 -> GS, 257 -> FS, 258 -> SS.
   if (AddrSpace == 256)
     AM.Segment = CurDAG->getRegister(X86::GS, MVT::i16);
   if (AddrSpace == 257)
     AM.Segment = CurDAG->getRegister(X86::FS, MVT::i16);
   if (AddrSpace == 258)
     AM.Segment = CurDAG->getRegister(X86::SS, MVT::i16);
 
-  // If Base is 0, the whole address is in index and the Scale is 1
-  if (isa<ConstantSDNode>(N)) {
-    assert(cast<ConstantSDNode>(N)->isNullValue() &&
-           "Unexpected base in gather/scatter");
-    AM.Scale = 1;
-  }
-  // Otherwise, try to match into the base and displacement fields.
-  else if (matchVectorAddress(N, AM))
+  // Try to match into the base and displacement fields.
+  if (matchVectorAddress(N, AM))
     return false;
 
   MVT VT = N.getSimpleValueType();
   if (AM.BaseType == X86ISelAddressMode::RegBase) {
     if (!AM.Base_Reg.getNode())
       AM.Base_Reg = CurDAG->getRegister(0, VT);
   }
 

llvm/trunk/lib/Target/X86/X86ISelLowering.h

X86MaskedGatherScatterSDNode(unsigned Opc, unsigned Order,
                                  const DebugLoc &dl, SDVTList VTs, EVT MemVT,
                                  MachineMemOperand *MMO)
         : MemSDNode(Opc, Order, dl, VTs, MemVT, MMO) {}
 
     const SDValue &getBasePtr() const { return getOperand(3); }
     const SDValue &getIndex()   const { return getOperand(4); }
     const SDValue &getMask()    const { return getOperand(2); }
     const SDValue &getValue()   const { return getOperand(1); }
+    const SDValue &getScale()   const { return getOperand(5); }
 
     static bool classof(const SDNode *N) {
       return N->getOpcode() == X86ISD::MGATHER ||
              N->getOpcode() == X86ISD::MSCATTER;
     }
   };
 
   class X86MaskedGatherSDNode : public X86MaskedGatherScatterSDNode {

llvm/trunk/lib/Target/X86/X86ISelLowering.cpp

assert(Subtarget.hasAVX512() &&
          "MGATHER/MSCATTER are supported on AVX-512 arch only");
 
   MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
   SDValue Src = N->getValue();
   MVT VT = Src.getSimpleValueType();
   assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
   SDLoc dl(Op);
 
+  SDValue Scale = N->getScale();
   SDValue Index = N->getIndex();
   SDValue Mask = N->getMask();
   SDValue Chain = N->getChain();
   SDValue BasePtr = N->getBasePtr();
   MVT MemVT = N->getMemoryVT().getSimpleVT();
   MVT IndexVT = Index.getSimpleValueType();
   MVT MaskVT = Mask.getSimpleValueType();
 
else {
       // The value that should be stored
       MVT NewVT = MVT::getVectorVT(VT.getScalarType(), NumElts);
       Src = ExtendToType(Src, NewVT, DAG);
     }
   }
 
   // The mask is killed by scatter, add it to the values
   SDVTList VTs = DAG.getVTList(Mask.getValueType(), MVT::Other);
-  SDValue Ops[] = {Chain, Src, Mask, BasePtr, Index};
+  SDValue Ops[] = {Chain, Src, Mask, BasePtr, Index, Scale};
   SDValue NewScatter = DAG.getTargetMemSDNode<X86MaskedScatterSDNode>(
       VTs, Ops, dl, N->getMemoryVT(), N->getMemOperand());
   DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
   return SDValue(NewScatter.getNode(), 1);
 }
 
 static SDValue LowerMLOAD(SDValue Op, const X86Subtarget &Subtarget,
                           SelectionDAG &DAG) {

 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget &Subtarget,
                             SelectionDAG &DAG) {
   assert(Subtarget.hasAVX2() &&
          "MGATHER/MSCATTER are supported on AVX-512/AVX-2 arch only");
 
   MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
   SDLoc dl(Op);
   MVT VT = Op.getSimpleValueType();
+  SDValue Scale = N->getScale();
   SDValue Index = N->getIndex();
   SDValue Mask = N->getMask();
   SDValue Src0 = N->getValue();
   MVT IndexVT = Index.getSimpleValueType();
   MVT MaskVT = Mask.getSimpleValueType();
 
   unsigned NumElts = VT.getVectorNumElements();
   assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
 
   // If the index is v2i32, we're being called by type legalization.
   if (IndexVT == MVT::v2i32)
     return SDValue();
 
   if (Subtarget.hasAVX512() && !Subtarget.hasVLX() && !VT.is512BitVector() &&
       !Index.getSimpleValueType().is512BitVector()) {
     // AVX512F supports only 512-bit vectors. Or data or index should
     // be 512 bit wide. If now the both index and data are 256-bit, but
     // the vector contains 8 elements, we just sign-extend the index
     if (NumElts == 8) {
       Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
-      SDValue Ops[] = { N->getChain(), Src0, Mask, N->getBasePtr(), Index };
+      SDValue Ops[] = { N->getChain(), Src0, Mask, N->getBasePtr(), Index,
+                        Scale };
       SDValue NewGather = DAG.getTargetMemSDNode<X86MaskedGatherSDNode>(
           DAG.getVTList(VT, MaskVT, MVT::Other), Ops, dl, N->getMemoryVT(),
           N->getMemOperand());
       return DAG.getMergeValues({NewGather, NewGather.getValue(2)}, dl);
     }
 
     // Minimal number of elements in Gather
     NumElts = 8;
     // Index
     MVT NewIndexVT = MVT::getVectorVT(IndexVT.getScalarType(), NumElts);
     Index = ExtendToType(Index, NewIndexVT, DAG);
     if (IndexVT.getScalarType() == MVT::i32)
       Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
 
     // Mask
     assert(MaskVT.getScalarType() == MVT::i1 && "unexpected mask type");
     MaskVT = MVT::getVectorVT(MVT::i1, NumElts);
     Mask = ExtendToType(Mask, MaskVT, DAG, true);
 
     // The pass-through value
     MVT NewVT = MVT::getVectorVT(VT.getScalarType(), NumElts);
     Src0 = ExtendToType(Src0, NewVT, DAG);
 
-    SDValue Ops[] = { N->getChain(), Src0, Mask, N->getBasePtr(), Index };
+    SDValue Ops[] = { N->getChain(), Src0, Mask, N->getBasePtr(), Index, Scale };
     SDValue NewGather = DAG.getTargetMemSDNode<X86MaskedGatherSDNode>(
         DAG.getVTList(NewVT, MaskVT, MVT::Other), Ops, dl, N->getMemoryVT(),
         N->getMemOperand());
     SDValue Extract = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
                                   NewGather.getValue(0),
                                   DAG.getIntPtrConstant(0, dl));
     SDValue RetOps[] = {Extract, NewGather.getValue(2)};
     return DAG.getMergeValues(RetOps, dl);
   }
 
-  SDValue Ops[] = { N->getChain(), Src0, Mask, N->getBasePtr(), Index };
+  SDValue Ops[] = { N->getChain(), Src0, Mask, N->getBasePtr(), Index, Scale };
   SDValue NewGather = DAG.getTargetMemSDNode<X86MaskedGatherSDNode>(
       DAG.getVTList(VT, MaskVT, MVT::Other), Ops, dl, N->getMemoryVT(),
       N->getMemOperand());
   return DAG.getMergeValues({NewGather, NewGather.getValue(2)}, dl);
 }
 
 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
                                                     SelectionDAG &DAG) const {
if (VT == MVT::v2f32 && (Subtarget.hasVLX() || !Subtarget.hasAVX512())) {
       if (!Subtarget.hasVLX()) {
         // We need to widen the mask, but the instruction will only use 2
         // of its elements. So we can use undef.
         Mask = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i1, Mask,
                            DAG.getUNDEF(MVT::v2i1));
         Mask = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Mask);
       }
       SDValue Ops[] = { Gather->getChain(), Src0, Mask, Gather->getBasePtr(),
-                        Index };
+                        Index, Gather->getScale() };
       SDValue Res = DAG.getTargetMemSDNode<X86MaskedGatherSDNode>(
         DAG.getVTList(MVT::v4f32, Mask.getValueType(), MVT::Other), Ops, dl,
         Gather->getMemoryVT(), Gather->getMemOperand());
       Results.push_back(Res);
       Results.push_back(Res.getValue(2));
       return;
     }
     if (VT == MVT::v2i32) {
if (VT == MVT::v2i32) {
         if (!Subtarget.hasVLX()) {
           // We need to widen the mask, but the instruction will only use 2
           // of its elements. So we can use undef.
           Mask = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i1, Mask,
                              DAG.getUNDEF(MVT::v2i1));
           Mask = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i32, Mask);
         }
         SDValue Ops[] = { Gather->getChain(), Src0, Mask, Gather->getBasePtr(),
-                          Index };
+                          Index, Gather->getScale() };
         SDValue Res = DAG.getTargetMemSDNode<X86MaskedGatherSDNode>(
           DAG.getVTList(MVT::v4i32, Mask.getValueType(), MVT::Other), Ops, dl,
           Gather->getMemoryVT(), Gather->getMemOperand());
         SDValue Chain = Res.getValue(2);
         if (!ExperimentalVectorWideningLegalization)
           Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v2i32, Res,
                             DAG.getIntPtrConstant(0, dl));
         Results.push_back(Res);
         Results.push_back(Chain);
         return;
       }
       EVT IndexVT = Index.getValueType();
       EVT NewIndexVT = EVT::getVectorVT(*DAG.getContext(),
                                         IndexVT.getScalarType(), 4);
       // Otherwise we need to custom widen everything to avoid promotion.
       Index = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewIndexVT, Index,
                           DAG.getUNDEF(IndexVT));
       Mask = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i1, Mask,
                          DAG.getConstant(0, dl, MVT::v2i1));
       SDValue Ops[] = { Gather->getChain(), Src0, Mask, Gather->getBasePtr(),
-                        Index };
+                        Index, Gather->getScale() };
       SDValue Res = DAG.getMaskedGather(DAG.getVTList(MVT::v4i32, MVT::Other),
                                         Gather->getMemoryVT(), dl, Ops,
                                         Gather->getMemOperand());
       SDValue Chain = Res.getValue(1);
       if (!ExperimentalVectorWideningLegalization)
         Res = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v2i32, Res,
                           DAG.getIntPtrConstant(0, dl));
       Results.push_back(Res);

llvm/trunk/test/CodeGen/X86/masked_gather_scatter.ll

 ; SKX_32-NEXT:    retl
   %ind_masked = and <16 x i32> %ind, <i32 15, i32 15, i32 15, i32 15, i32 15, i32 15, i32 15, i32 15, i32 15, i32 15, i32 15, i32 15, i32 15, i32 15, i32 15, i32 15>
   %sext_ind = zext <16 x i32> %ind_masked to <16 x i64>
   %gep.random = getelementptr float, float *%base, <16 x i64> %sext_ind
 
   %res = call <16 x float> @llvm.masked.gather.v16f32.v16p0f32(<16 x float*> %gep.random, i32 4, <16 x i1> <i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true>, <16 x float> undef)
   ret <16 x float>%res
 }
+
+define <16 x double> @test_gather_setcc_split(double* %base, <16 x i32> %ind, <16 x i32> %cmp, <16 x double> %passthru) {
+; KNL_64-LABEL: test_gather_setcc_split:
+; KNL_64:       # %bb.0:
+; KNL_64-NEXT:    vextractf64x4 $1, %zmm0, %ymm4
+; KNL_64-NEXT:    vpxor %xmm5, %xmm5, %xmm5
+; KNL_64-NEXT:    vextracti64x4 $1, %zmm1, %ymm6
+; KNL_64-NEXT:    vpcmpeqd %zmm5, %zmm6, %k1
+; KNL_64-NEXT:    vpcmpeqd %zmm5, %zmm1, %k2
+; KNL_64-NEXT:    vgatherdpd (%rdi,%ymm0,8), %zmm2 {%k2}
+; KNL_64-NEXT:    vgatherdpd (%rdi,%ymm4,8), %zmm3 {%k1}
+; KNL_64-NEXT:    vmovapd %zmm2, %zmm0
+; KNL_64-NEXT:    vmovapd %zmm3, %zmm1
+; KNL_64-NEXT:    retq
+;
+; KNL_32-LABEL: test_gather_setcc_split:
+; KNL_32:       # %bb.0:
+; KNL_32-NEXT:    pushl %ebp
+; KNL_32-NEXT:    .cfi_def_cfa_offset 8
+; KNL_32-NEXT:    .cfi_offset %ebp, -8
+; KNL_32-NEXT:    movl %esp, %ebp
+; KNL_32-NEXT:    .cfi_def_cfa_register %ebp
+; KNL_32-NEXT:    andl $-64, %esp
+; KNL_32-NEXT:    subl $64, %esp
+; KNL_32-NEXT:    vmovapd 72(%ebp), %zmm3
+; KNL_32-NEXT:    movl 8(%ebp), %eax
+; KNL_32-NEXT:    vextractf64x4 $1, %zmm0, %ymm4
+; KNL_32-NEXT:    vpxor %xmm5, %xmm5, %xmm5
+; KNL_32-NEXT:    vextracti64x4 $1, %zmm1, %ymm6
+; KNL_32-NEXT:    vpcmpeqd %zmm5, %zmm6, %k1
+; KNL_32-NEXT:    vpcmpeqd %zmm5, %zmm1, %k2
+; KNL_32-NEXT:    vgatherdpd (%eax,%ymm0,8), %zmm2 {%k2}
+; KNL_32-NEXT:    vgatherdpd (%eax,%ymm4,8), %zmm3 {%k1}
+; KNL_32-NEXT:    vmovapd %zmm2, %zmm0
+; KNL_32-NEXT:    vmovapd %zmm3, %zmm1
+; KNL_32-NEXT:    movl %ebp, %esp
+; KNL_32-NEXT:    popl %ebp
+; KNL_32-NEXT:    retl
+;
+; SKX-LABEL: test_gather_setcc_split:
+; SKX:       # %bb.0:
+; SKX-NEXT:    vextractf64x4 $1, %zmm0, %ymm4
+; SKX-NEXT:    vextracti64x4 $1, %zmm1, %ymm5
+; SKX-NEXT:    vpxor %xmm6, %xmm6, %xmm6
+; SKX-NEXT:    vpcmpeqd %ymm6, %ymm5, %k1
+; SKX-NEXT:    vpcmpeqd %ymm6, %ymm1, %k2
+; SKX-NEXT:    vgatherdpd (%rdi,%ymm0,8), %zmm2 {%k2}
+; SKX-NEXT:    vgatherdpd (%rdi,%ymm4,8), %zmm3 {%k1}
+; SKX-NEXT:    vmovapd %zmm2, %zmm0
+; SKX-NEXT:    vmovapd %zmm3, %zmm1
+; SKX-NEXT:    retq
+;
+; SKX_32-LABEL: test_gather_setcc_split:
+; SKX_32:       # %bb.0:
+; SKX_32-NEXT:    pushl %ebp
+; SKX_32-NEXT:    .cfi_def_cfa_offset 8
+; SKX_32-NEXT:    .cfi_offset %ebp, -8
+; SKX_32-NEXT:    movl %esp, %ebp
+; SKX_32-NEXT:    .cfi_def_cfa_register %ebp
+; SKX_32-NEXT:    andl $-64, %esp
+; SKX_32-NEXT:    subl $64, %esp
+; SKX_32-NEXT:    vmovapd 72(%ebp), %zmm3
+; SKX_32-NEXT:    movl 8(%ebp), %eax
+; SKX_32-NEXT:    vextractf64x4 $1, %zmm0, %ymm4
+; SKX_32-NEXT:    vextracti64x4 $1, %zmm1, %ymm5
+; SKX_32-NEXT:    vpxor %xmm6, %xmm6, %xmm6
+; SKX_32-NEXT:    vpcmpeqd %ymm6, %ymm5, %k1
+; SKX_32-NEXT:    vpcmpeqd %ymm6, %ymm1, %k2
+; SKX_32-NEXT:    vgatherdpd (%eax,%ymm0,8), %zmm2 {%k2}
+; SKX_32-NEXT:    vgatherdpd (%eax,%ymm4,8), %zmm3 {%k1}
+; SKX_32-NEXT:    vmovapd %zmm2, %zmm0
+; SKX_32-NEXT:    vmovapd %zmm3, %zmm1
+; SKX_32-NEXT:    movl %ebp, %esp
+; SKX_32-NEXT:    popl %ebp
+; SKX_32-NEXT:    retl
+  %sext_ind = sext <16 x i32> %ind to <16 x i64>
+  %gep.random = getelementptr double, double *%base, <16 x i64> %sext_ind
+
+  %mask = icmp eq <16 x i32> %cmp, zeroinitializer
+  %res = call <16 x double> @llvm.masked.gather.v16f64.v16p0f64(<16 x double*> %gep.random, i32 4, <16 x i1> %mask, <16 x double> %passthru)
+  ret <16 x double>%res
+}
+
+define void @test_scatter_setcc_split(double* %base, <16 x i32> %ind, <16 x i32> %cmp, <16 x double> %src0)  {
+; KNL_64-LABEL: test_scatter_setcc_split:
+; KNL_64:       # %bb.0:
+; KNL_64-NEXT:    vextractf64x4 $1, %zmm0, %ymm4
+; KNL_64-NEXT:    vpxor %xmm5, %xmm5, %xmm5
+; KNL_64-NEXT:    vpcmpeqd %zmm5, %zmm1, %k1
+; KNL_64-NEXT:    vextracti64x4 $1, %zmm1, %ymm1
+; KNL_64-NEXT:    vpcmpeqd %zmm5, %zmm1, %k2
+; KNL_64-NEXT:    vscatterdpd %zmm3, (%rdi,%ymm4,8) {%k2}
+; KNL_64-NEXT:    vscatterdpd %zmm2, (%rdi,%ymm0,8) {%k1}
+; KNL_64-NEXT:    vzeroupper
+; KNL_64-NEXT:    retq
+;
+; KNL_32-LABEL: test_scatter_setcc_split:
+; KNL_32:       # %bb.0:
+; KNL_32-NEXT:    pushl %ebp
+; KNL_32-NEXT:    .cfi_def_cfa_offset 8
+; KNL_32-NEXT:    .cfi_offset %ebp, -8
+; KNL_32-NEXT:    movl %esp, %ebp
+; KNL_32-NEXT:    .cfi_def_cfa_register %ebp
+; KNL_32-NEXT:    andl $-64, %esp
+; KNL_32-NEXT:    subl $64, %esp
+; KNL_32-NEXT:    vmovapd 72(%ebp), %zmm3
+; KNL_32-NEXT:    movl 8(%ebp), %eax
+; KNL_32-NEXT:    vextractf64x4 $1, %zmm0, %ymm4
+; KNL_32-NEXT:    vpxor %xmm5, %xmm5, %xmm5
+; KNL_32-NEXT:    vpcmpeqd %zmm5, %zmm1, %k1
+; KNL_32-NEXT:    vextracti64x4 $1, %zmm1, %ymm1
+; KNL_32-NEXT:    vpcmpeqd %zmm5, %zmm1, %k2
+; KNL_32-NEXT:    vscatterdpd %zmm3, (%eax,%ymm4,8) {%k2}
+; KNL_32-NEXT:    vscatterdpd %zmm2, (%eax,%ymm0,8) {%k1}
+; KNL_32-NEXT:    movl %ebp, %esp
+; KNL_32-NEXT:    popl %ebp
+; KNL_32-NEXT:    vzeroupper
+; KNL_32-NEXT:    retl
+;
+; SKX-LABEL: test_scatter_setcc_split:
+; SKX:       # %bb.0:
+; SKX-NEXT:    vextractf64x4 $1, %zmm0, %ymm4
+; SKX-NEXT:    vpxor %xmm5, %xmm5, %xmm5
+; SKX-NEXT:    vpcmpeqd %ymm5, %ymm1, %k1
+; SKX-NEXT:    vextracti64x4 $1, %zmm1, %ymm1
+; SKX-NEXT:    vpcmpeqd %ymm5, %ymm1, %k2
+; SKX-NEXT:    vscatterdpd %zmm3, (%rdi,%ymm4,8) {%k2}
+; SKX-NEXT:    vscatterdpd %zmm2, (%rdi,%ymm0,8) {%k1}
+; SKX-NEXT:    vzeroupper
+; SKX-NEXT:    retq
+;
+; SKX_32-LABEL: test_scatter_setcc_split:
+; SKX_32:       # %bb.0:
+; SKX_32-NEXT:    pushl %ebp
+; SKX_32-NEXT:    .cfi_def_cfa_offset 8
+; SKX_32-NEXT:    .cfi_offset %ebp, -8
+; SKX_32-NEXT:    movl %esp, %ebp
+; SKX_32-NEXT:    .cfi_def_cfa_register %ebp
+; SKX_32-NEXT:    andl $-64, %esp
+; SKX_32-NEXT:    subl $64, %esp
+; SKX_32-NEXT:    vmovapd 72(%ebp), %zmm3
+; SKX_32-NEXT:    movl 8(%ebp), %eax
+; SKX_32-NEXT:    vextractf64x4 $1, %zmm0, %ymm4
+; SKX_32-NEXT:    vpxor %xmm5, %xmm5, %xmm5
+; SKX_32-NEXT:    vpcmpeqd %ymm5, %ymm1, %k1
+; SKX_32-NEXT:    vextracti64x4 $1, %zmm1, %ymm1
+; SKX_32-NEXT:    vpcmpeqd %ymm5, %ymm1, %k2
+; SKX_32-NEXT:    vscatterdpd %zmm3, (%eax,%ymm4,8) {%k2}
+; SKX_32-NEXT:    vscatterdpd %zmm2, (%eax,%ymm0,8) {%k1}
+; SKX_32-NEXT:    movl %ebp, %esp
+; SKX_32-NEXT:    popl %ebp
+; SKX_32-NEXT:    vzeroupper
+; SKX_32-NEXT:    retl
+  %sext_ind = sext <16 x i32> %ind to <16 x i64>
+  %gep.random = getelementptr double, double *%base, <16 x i64> %sext_ind
+
+  %mask = icmp eq <16 x i32> %cmp, zeroinitializer
+  call void @llvm.masked.scatter.v16f64.v16p0f64(<16 x double> %src0, <16 x double*> %gep.random, i32 4, <16 x i1> %mask)
+  ret void
+}