[X86] Directly emit VBROADCAST_LOAD from constant pool in lowerBuildVectorAsBroadcast

[craig.topper]

Summary

Add a DAG combine to combine different sized broadcasts from
constant pool to avoid a regression.

[RKSimon]

LGTM

llvm/lib/Target/X86/X86ISelLowering.cpp

if (BVOp->isConstantSplat(SplatValue, Undef, SplatBitSize, HasUndef) &&
           // Load the constant and broadcast it.
           MVT CVT = MVT::getIntegerVT(SplatBitSize);
           Type *ScalarTy = Type::getIntNTy(*Ctx, SplatBitSize);
           Constant *C = Constant::getIntegerValue(ScalarTy, SplatValue);
           SDValue CP = DAG.getConstantPool(C, PVT);
           unsigned Repeat = VT.getSizeInBits() / SplatBitSize;
 
           unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
-          Ld = DAG.getLoad(
-              CVT, dl, DAG.getEntryNode(), CP,
-              MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
-              Alignment);
-          SDValue Brdcst = DAG.getNode(X86ISD::VBROADCAST, dl,
-                                       MVT::getVectorVT(CVT, Repeat), Ld);
+          SDVTList Tys =
+              DAG.getVTList(MVT::getVectorVT(CVT, Repeat), MVT::Other);
+          SDValue Ops[] = {DAG.getEntryNode(), CP};
+          MachinePointerInfo MPI =
+              MachinePointerInfo::getConstantPool(DAG.getMachineFunction());
+          SDValue Brdcst = DAG.getMemIntrinsicNode(
+              X86ISD::VBROADCAST_LOAD, dl, Tys, Ops, CVT, MPI, Alignment,
+              MachineMemOperand::MOLoad);
           return DAG.getBitcast(VT, Brdcst);
         } else if (SplatBitSize == 32 || SplatBitSize == 64) {
           // Splatted value can fit in one FLOAT constant in constant pool.
           // Load the constant and broadcast it.
           // AVX have support for 32 and 64 bit broadcast for floats only.
           // No 64bit integer in 32bit subtarget.
           MVT CVT = MVT::getFloatingPointVT(SplatBitSize);
           // Lower the splat via APFloat directly, to avoid any conversion.
           Constant *C =
               SplatBitSize == 32
                   ? ConstantFP::get(*Ctx,
                                     APFloat(APFloat::IEEEsingle(), SplatValue))
                   : ConstantFP::get(*Ctx,
                                     APFloat(APFloat::IEEEdouble(), SplatValue));
           SDValue CP = DAG.getConstantPool(C, PVT);
           unsigned Repeat = VT.getSizeInBits() / SplatBitSize;
 
           unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
-          Ld = DAG.getLoad(
-              CVT, dl, DAG.getEntryNode(), CP,
-              MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
-              Alignment);
-          SDValue Brdcst = DAG.getNode(X86ISD::VBROADCAST, dl,
-                                       MVT::getVectorVT(CVT, Repeat), Ld);
+          SDVTList Tys =
+              DAG.getVTList(MVT::getVectorVT(CVT, Repeat), MVT::Other);
+          SDValue Ops[] = {DAG.getEntryNode(), CP};
+          MachinePointerInfo MPI =
+              MachinePointerInfo::getConstantPool(DAG.getMachineFunction());
+          SDValue Brdcst = DAG.getMemIntrinsicNode(
+              X86ISD::VBROADCAST_LOAD, dl, Tys, Ops, CVT, MPI, Alignment,
+              MachineMemOperand::MOLoad);
           return DAG.getBitcast(VT, Brdcst);
         } else if (SplatBitSize > 64) {
           // Load the vector of constants and broadcast it.
           MVT CVT = VT.getScalarType();
           Constant *VecC = getConstantVector(VT, SplatValue, SplatBitSize,
                                              *Ctx);
           SDValue VCP = DAG.getConstantPool(VecC, PVT);
           unsigned NumElm = SplatBitSize / VT.getScalarSizeInBits();
if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
         C = CF->getConstantFPValue();
 
       assert(C && "Invalid constant type");
 
       const TargetLowering &TLI = DAG.getTargetLoweringInfo();
       SDValue CP =
           DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
       unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
-      Ld = DAG.getLoad(
-          CVT, dl, DAG.getEntryNode(), CP,
-          MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
-          Alignment);
 
-      return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
+      SDVTList Tys = DAG.getVTList(VT, MVT::Other);
+      SDValue Ops[] = {DAG.getEntryNode(), CP};
+      MachinePointerInfo MPI =
+          MachinePointerInfo::getConstantPool(DAG.getMachineFunction());
+      return DAG.getMemIntrinsicNode(X86ISD::VBROADCAST_LOAD, dl, Tys, Ops, CVT,
+                                     MPI, Alignment, MachineMemOperand::MOLoad);
     }
   }
 
   // Handle AVX2 in-register broadcasts.
   if (!IsLoad && Subtarget.hasInt256() &&
       (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
     return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
 
if (IsStrict) {
     }
     return DAG.getMergeValues({Cvt, Chain}, dl);
   }
 
   // Extend to the original VT if necessary.
   return DAG.getNode(ISD::FP_EXTEND, dl, VT, Cvt);
 }
 
+// Try to find a larger VBROADCAST_LOAD that we can extract from. Limit this to
+// cases where the loads have the same input chain and the output chains are
+// unused. This avoids any memory ordering issues.
+static SDValue combineVBROADCAST_LOAD(SDNode *N, SelectionDAG &DAG,
+                                      TargetLowering::DAGCombinerInfo &DCI) {
+  // Only do this if the chain result is unused.
+  if (N->hasAnyUseOfValue(1))
+    return SDValue();
+
+  auto *MemIntrin = cast<MemIntrinsicSDNode>(N);
+
+  SDValue Ptr = MemIntrin->getBasePtr();
+  SDValue Chain = MemIntrin->getChain();
+  EVT VT = N->getSimpleValueType(0);
+  EVT MemVT = MemIntrin->getMemoryVT();
+
+  // Look at other users of our base pointer and try to find a wider broadcast.
+  // The input chain and the size of the memory VT must match.
+  for (SDNode *User : Ptr->uses())
+    if (User != N && User->getOpcode() == X86ISD::VBROADCAST_LOAD &&
+        cast<MemIntrinsicSDNode>(User)->getBasePtr() == Ptr &&
+        cast<MemIntrinsicSDNode>(User)->getChain() == Chain &&
+        cast<MemIntrinsicSDNode>(User)->getMemoryVT().getSizeInBits() ==
+            MemVT.getSizeInBits() &&
+        !User->hasAnyUseOfValue(1) &&
+        User->getValueSizeInBits(0) > VT.getSizeInBits()) {
+      SDValue Extract = extractSubVector(SDValue(User, 0), 0, DAG, SDLoc(N),
+                                         VT.getSizeInBits());
+      Extract = DAG.getBitcast(VT, Extract);
+      return DCI.CombineTo(N, Extract, SDValue(User, 1));
+    }
+
+  return SDValue();
+}
+
 static SDValue combineFP_ROUND(SDNode *N, SelectionDAG &DAG,
                                const X86Subtarget &Subtarget) {
   if (!Subtarget.hasF16C() || Subtarget.useSoftFloat())
     return SDValue();
 
   EVT VT = N->getValueType(0);
   SDValue Src = N->getOperand(0);
   EVT SrcVT = Src.getValueType();
SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case X86ISD::PMULDQ:
   case X86ISD::PMULUDQ:     return combinePMULDQ(N, DAG, DCI, Subtarget);
   case X86ISD::KSHIFTL:
   case X86ISD::KSHIFTR:     return combineKSHIFT(N, DAG, DCI);
   case ISD::FP16_TO_FP:     return combineFP16_TO_FP(N, DAG, Subtarget);
   case ISD::STRICT_FP_EXTEND:
   case ISD::FP_EXTEND:      return combineFP_EXTEND(N, DAG, Subtarget);
   case ISD::FP_ROUND:       return combineFP_ROUND(N, DAG, Subtarget);
+  case X86ISD::VBROADCAST_LOAD: return combineVBROADCAST_LOAD(N, DAG, DCI);
   }
 
   return SDValue();
 }
 
 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
   if (!isTypeLegal(VT))
     return false;