[SelectionDAG] When splitting scatter nodes during DAGCombine, create a serial chain dependency.

[craig.topper]

Summary

Scatter could have multiple identical indices. We need to maintain sequential order. We get this right in LegalizeVectorTypes, but not in this code.

[craig.topper]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

7426

I removed the AddToWorklist calls because the worklist code in the DAGCombiner ensure all operands of visited nodes have been added to worklist. So we return Hi, which should get added to the worklist and then visited. Lo is an operand of Hi due to the chain so it should get added to the worklist then.

[delena]

Yes. The scatter operation should be in-order after split.

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

SDValue DAGCombiner::visitMSCATTER(SDNode *N) {
   // and enables future optimizations (e.g. min/max pattern matching on X86).
   if (Mask.getOpcode() != ISD::SETCC)
     return SDValue();
 
   // Check if any splitting is required.
   if (TLI.getTypeAction(*DAG.getContext(), Data.getValueType()) !=
       TargetLowering::TypeSplitVector)
     return SDValue();
-  SDValue MaskLo, MaskHi, Lo, Hi;
+  SDValue MaskLo, MaskHi;
   std::tie(MaskLo, MaskHi) = SplitVSETCC(Mask.getNode(), DAG);
 
   EVT LoVT, HiVT;
   std::tie(LoVT, HiVT) = DAG.GetSplitDestVTs(MSC->getValueType(0));
 
   SDValue Chain = MSC->getChain();
 
   EVT MemoryVT = MSC->getMemoryVT();
SDValue DAGCombiner::visitMSCATTER(SDNode *N) {
   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, Scale };
-  Lo = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), DataLo.getValueType(),
-                            DL, OpsLo, MMO);
+  SDValue Lo = DAG.getMaskedScatter(DAG.getVTList(MVT::Other),
+                                    DataLo.getValueType(), DL, OpsLo, MMO);
 
-  SDValue OpsHi[] = { Chain, DataHi, MaskHi, BasePtr, IndexHi, Scale };
-  Hi = DAG.getMaskedScatter(DAG.getVTList(MVT::Other), DataHi.getValueType(),
+  // 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, BasePtr, IndexHi, Scale };
+  return 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::visitMSTORE(SDNode *N) {
   if (Level >= AfterLegalizeTypes)
     return SDValue();
 
   MaskedStoreSDNode *MST = dyn_cast<MaskedStoreSDNode>(N);
   SDValue Mask = MST->getMask();

test/CodeGen/X86/masked_gather_scatter.ll

; SKX_32-NEXT:    retl
   %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:    vptestnmd %zmm1, %zmm1, %k1
-; KNL_64-NEXT:    vextracti64x4 $1, %zmm1, %ymm1
+; KNL_64-NEXT:    vextracti64x4 $1, %zmm1, %ymm5
+; KNL_64-NEXT:    vptestnmd %zmm5, %zmm5, %k1
 ; KNL_64-NEXT:    vptestnmd %zmm1, %zmm1, %k2
-; KNL_64-NEXT:    vscatterdpd %zmm3, (%rdi,%ymm4,8) {%k2}
-; KNL_64-NEXT:    vscatterdpd %zmm2, (%rdi,%ymm0,8) {%k1}
+; KNL_64-NEXT:    vscatterdpd %zmm2, (%rdi,%ymm0,8) {%k2}
+; KNL_64-NEXT:    vscatterdpd %zmm3, (%rdi,%ymm4,8) {%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:    vptestnmd %zmm1, %zmm1, %k1
-; KNL_32-NEXT:    vextracti64x4 $1, %zmm1, %ymm1
+; KNL_32-NEXT:    vextracti64x4 $1, %zmm1, %ymm5
+; KNL_32-NEXT:    vptestnmd %zmm5, %zmm5, %k1
 ; KNL_32-NEXT:    vptestnmd %zmm1, %zmm1, %k2
-; KNL_32-NEXT:    vscatterdpd %zmm3, (%eax,%ymm4,8) {%k2}
-; KNL_32-NEXT:    vscatterdpd %zmm2, (%eax,%ymm0,8) {%k1}
+; KNL_32-NEXT:    vscatterdpd %zmm2, (%eax,%ymm0,8) {%k2}
+; KNL_32-NEXT:    vscatterdpd %zmm3, (%eax,%ymm4,8) {%k1}
 ; KNL_32-NEXT:    movl %ebp, %esp
 ; KNL_32-NEXT:    popl %ebp
 ; KNL_32-NEXT:    .cfi_def_cfa %esp, 4
 ; 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:    vptestnmd %ymm1, %ymm1, %k1
-; SKX-NEXT:    vextracti64x4 $1, %zmm1, %ymm1
+; SKX-NEXT:    vextracti64x4 $1, %zmm1, %ymm5
+; SKX-NEXT:    vptestnmd %ymm5, %ymm5, %k1
 ; SKX-NEXT:    vptestnmd %ymm1, %ymm1, %k2
-; SKX-NEXT:    vscatterdpd %zmm3, (%rdi,%ymm4,8) {%k2}
-; SKX-NEXT:    vscatterdpd %zmm2, (%rdi,%ymm0,8) {%k1}
+; SKX-NEXT:    vscatterdpd %zmm2, (%rdi,%ymm0,8) {%k2}
+; SKX-NEXT:    vscatterdpd %zmm3, (%rdi,%ymm4,8) {%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:    vptestnmd %ymm1, %ymm1, %k1
-; SKX_32-NEXT:    vextracti64x4 $1, %zmm1, %ymm1
+; SKX_32-NEXT:    vextracti64x4 $1, %zmm1, %ymm5
+; SKX_32-NEXT:    vptestnmd %ymm5, %ymm5, %k1
 ; SKX_32-NEXT:    vptestnmd %ymm1, %ymm1, %k2
-; SKX_32-NEXT:    vscatterdpd %zmm3, (%eax,%ymm4,8) {%k2}
-; SKX_32-NEXT:    vscatterdpd %zmm2, (%eax,%ymm0,8) {%k1}
+; SKX_32-NEXT:    vscatterdpd %zmm2, (%eax,%ymm0,8) {%k2}
+; SKX_32-NEXT:    vscatterdpd %zmm3, (%eax,%ymm4,8) {%k1}
 ; SKX_32-NEXT:    movl %ebp, %esp
 ; SKX_32-NEXT:    popl %ebp
 ; SKX_32-NEXT:    .cfi_def_cfa %esp, 4
 ; 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