Change from adding ISD::PARTIAL_REDUCE_S/UADD to adding

ISD::PARTIAL_REDUCE_S/UMLA

This makes the lowering function easier as you do not need to worry
about whether the MUL is lowered or not. Instead its operands are
taken from it. If there is no MUL instruction and just one operand,
the other operand is a vector of ones (for value types eligible for
wide add lowering).

Author: JamesChesterman

llvm/include/llvm/CodeGen/ISDOpcodes.h

@@ -1455,8 +1455,8 @@ enum NodeType {
   // unsigned).
   // Operands: Accumulator, Input
   // Outputs: Output
-  PARTIAL_REDUCE_SADD,
-  PARTIAL_REDUCE_UADD,
+  PARTIAL_REDUCE_SMLA,
+  PARTIAL_REDUCE_UMLA,
 
   // The `llvm.experimental.stackmap` intrinsic.
   // Operands: input chain, glue, <id>, <numShadowBytes>, [live0[, live1...]]

llvm/include/llvm/CodeGen/SelectionDAG.h

@@ -1602,7 +1602,7 @@ class SelectionDAG {
   /// the target's desired shift amount type.
   SDValue getShiftAmountOperand(EVT LHSTy, SDValue Op);
 
-  /// Expands PARTIAL_REDUCE_S/UADD nodes.
+  /// Expands PARTIAL_REDUCE_S/UMLA nodes.
   /// \p Op1 Accumulator for where the result is stored for the partial
   /// reduction operation.
   /// \p Op2 Input for the partial reduction operation.

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

@@ -8142,7 +8142,7 @@ void SelectionDAGBuilder::visitIntrinsicCall(const CallInst &I,
 
     if (!TLI.shouldExpandPartialReductionIntrinsic(cast<IntrinsicInst>(&I))) {
       setValue(&I,
-               DAG.getNode(ISD::PARTIAL_REDUCE_UADD, dl, AccVT, Acc, Input));
+               DAG.getNode(ISD::PARTIAL_REDUCE_UMLA, dl, AccVT, Acc, Input));
       return;
     }
     setValue(&I, DAG.expandPartialReduceAdd(dl, Acc, Input));

llvm/lib/CodeGen/SelectionDAG/SelectionDAGDumper.cpp

@@ -567,10 +567,10 @@ std::string SDNode::getOperationName(const SelectionDAG *G) const {
   case ISD::EXPERIMENTAL_VECTOR_HISTOGRAM:
     return "histogram";
 
-  case ISD::PARTIAL_REDUCE_UADD:
-    return "partial_reduce_uadd";
-  case ISD::PARTIAL_REDUCE_SADD:
-    return "partial_reduce_sadd";
+  case ISD::PARTIAL_REDUCE_UMLA:
+    return "partial_reduce_umla";
+  case ISD::PARTIAL_REDUCE_SMLA:
+    return "partial_reduce_smla";
 
     // Vector Predication
 #define BEGIN_REGISTER_VP_SDNODE(SDID, LEGALARG, NAME, ...)                    \

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -1124,7 +1124,7 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
   setTargetDAGCombine(
       {ISD::MGATHER, ISD::MSCATTER, ISD::EXPERIMENTAL_VECTOR_HISTOGRAM});
 
-  setTargetDAGCombine({ISD::PARTIAL_REDUCE_SADD, ISD::PARTIAL_REDUCE_UADD});
+  setTargetDAGCombine({ISD::PARTIAL_REDUCE_SMLA, ISD::PARTIAL_REDUCE_UMLA});
 
   setTargetDAGCombine(ISD::FP_EXTEND);
 
@@ -1842,14 +1842,14 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
     }
 
     for (auto VT : {MVT::nxv2i64, MVT::nxv4i32, MVT::nxv8i16}) {
-      setOperationAction(ISD::PARTIAL_REDUCE_UADD, VT, Custom);
-      setOperationAction(ISD::PARTIAL_REDUCE_SADD, VT, Custom);
+      setOperationAction(ISD::PARTIAL_REDUCE_UMLA, VT, Custom);
+      setOperationAction(ISD::PARTIAL_REDUCE_SMLA, VT, Custom);
     }
   }
 
   for (auto VT : {MVT::v4i64, MVT::v4i32, MVT::v2i32}) {
-    setOperationAction(ISD::PARTIAL_REDUCE_UADD, VT, Custom);
-    setOperationAction(ISD::PARTIAL_REDUCE_SADD, VT, Custom);
+    setOperationAction(ISD::PARTIAL_REDUCE_UMLA, VT, Custom);
+    setOperationAction(ISD::PARTIAL_REDUCE_SMLA, VT, Custom);
   }
 
   if (Subtarget->hasMOPS() && Subtarget->hasMTE()) {
@@ -7606,9 +7606,9 @@ SDValue AArch64TargetLowering::LowerOperation(SDValue Op,
     return LowerFLDEXP(Op, DAG);
   case ISD::EXPERIMENTAL_VECTOR_HISTOGRAM:
     return LowerVECTOR_HISTOGRAM(Op, DAG);
-  case ISD::PARTIAL_REDUCE_UADD:
-  case ISD::PARTIAL_REDUCE_SADD:
-    return LowerPARTIAL_REDUCE_ADD(Op, DAG);
+  case ISD::PARTIAL_REDUCE_UMLA:
+  case ISD::PARTIAL_REDUCE_SMLA:
+    return LowerPARTIAL_REDUCE_MLA(Op, DAG);
   }
 }
 
@@ -22070,9 +22070,8 @@ SDValue tryCombineToDotProduct(SDValue &Acc, SDValue &Input, SelectionDAG &DAG,
     return DAG.expandPartialReduceAdd(DL, Acc, Input);
 
   unsigned NewOpcode =
-      AIsSigned ? ISD::PARTIAL_REDUCE_SADD : ISD::PARTIAL_REDUCE_UADD;
-  auto NewMul = DAG.getNode(ISD::MUL, DL, A.getValueType(), A, B);
-  return DAG.getNode(NewOpcode, DL, AccVT, Acc, NewMul);
+      AIsSigned ? ISD::PARTIAL_REDUCE_SMLA : ISD::PARTIAL_REDUCE_UMLA;
+  return DAG.getNode(NewOpcode, DL, AccVT, Acc, A, B);
 }
 
 SDValue tryCombineToWideAdd(SDValue &Acc, SDValue &Input, SelectionDAG &DAG,
@@ -22094,9 +22093,10 @@ SDValue tryCombineToWideAdd(SDValue &Acc, SDValue &Input, SelectionDAG &DAG,
     return SDValue();
 
   unsigned NewOpcode = InputOpcode == ISD::SIGN_EXTEND
-                           ? ISD::PARTIAL_REDUCE_SADD
-                           : ISD::PARTIAL_REDUCE_UADD;
-  return DAG.getNode(NewOpcode, DL, AccVT, Acc, Input);
+                           ? ISD::PARTIAL_REDUCE_SMLA
+                           : ISD::PARTIAL_REDUCE_UMLA;
+  return DAG.getNode(NewOpcode, DL, AccVT, Acc, Input,
+                     DAG.getConstant(1, DL, InputVT));
 }
 
 SDValue performPartialReduceAddCombine(SDNode *N, SelectionDAG &DAG,
@@ -26412,8 +26412,8 @@ SDValue AArch64TargetLowering::PerformDAGCombine(SDNode *N,
   case ISD::MSCATTER:
   case ISD::EXPERIMENTAL_VECTOR_HISTOGRAM:
     return performMaskedGatherScatterCombine(N, DCI, DAG);
-  case ISD::PARTIAL_REDUCE_UADD:
-  case ISD::PARTIAL_REDUCE_SADD:
+  case ISD::PARTIAL_REDUCE_UMLA:
+  case ISD::PARTIAL_REDUCE_SMLA:
     return performPartialReduceAddCombine(N, DAG, Subtarget);
   case ISD::FP_EXTEND:
     return performFPExtendCombine(N, DAG, DCI, Subtarget);
@@ -29176,39 +29176,29 @@ SDValue AArch64TargetLowering::LowerVECTOR_HISTOGRAM(SDValue Op,
 }
 
 SDValue
-AArch64TargetLowering::LowerPARTIAL_REDUCE_ADD(SDValue Op,
+AArch64TargetLowering::LowerPARTIAL_REDUCE_MLA(SDValue Op,
                                                SelectionDAG &DAG) const {
   SDLoc DL(Op);
   SDValue Acc = Op.getOperand(0);
-  SDValue Input = Op.getOperand(1);
+  SDValue Input1 = Op.getOperand(1);
+  SDValue Input2 = Op.getOperand(2);
 
   EVT AccVT = Acc.getValueType();
-  EVT InputVT = Input.getValueType();
+  EVT InputVT = Input1.getValueType();
 
   unsigned Opcode = Op.getOpcode();
 
-  // If the following condition is true and the input opcode was not ISD::MUL
-  // during the DAG-combine, it is already expanded. So this condition means the
-  // input opcode must have been ISD::MUL.
   if (AccVT.getVectorElementCount() * 4 == InputVT.getVectorElementCount()) {
-    unsigned IndexAdd = 0;
-    // ISD::MUL may have already been lowered, meaning the operands would be in
-    // different positions.
-    if (Input.getOpcode() != ISD::MUL)
-      IndexAdd = 1;
-    auto A = Input.getOperand(IndexAdd);
-    auto B = Input.getOperand(IndexAdd + 1);
-
-    unsigned DotOpcode = Opcode == ISD::PARTIAL_REDUCE_SADD ? AArch64ISD::SDOT
+    unsigned DotOpcode = Opcode == ISD::PARTIAL_REDUCE_SMLA ? AArch64ISD::SDOT
                                                             : AArch64ISD::UDOT;
-    return DAG.getNode(DotOpcode, DL, AccVT, Acc, A, B);
+    return DAG.getNode(DotOpcode, DL, AccVT, Acc, Input1, Input2);
   }
-  bool InputIsSigned = Opcode == ISD::PARTIAL_REDUCE_SADD;
+  bool InputIsSigned = Opcode == ISD::PARTIAL_REDUCE_SMLA;
   unsigned BottomOpcode =
       InputIsSigned ? AArch64ISD::SADDWB : AArch64ISD::UADDWB;
   unsigned TopOpcode = InputIsSigned ? AArch64ISD::SADDWT : AArch64ISD::UADDWT;
-  auto BottomNode = DAG.getNode(BottomOpcode, DL, AccVT, Acc, Input);
-  return DAG.getNode(TopOpcode, DL, AccVT, BottomNode, Input);
+  auto BottomNode = DAG.getNode(BottomOpcode, DL, AccVT, Acc, Input1);
+  return DAG.getNode(TopOpcode, DL, AccVT, BottomNode, Input1);
 }
 
 SDValue

llvm/lib/Target/AArch64/AArch64ISelLowering.h

@@ -1184,7 +1184,7 @@ class AArch64TargetLowering : public TargetLowering {
   SDValue LowerVECTOR_DEINTERLEAVE(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerVECTOR_INTERLEAVE(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerVECTOR_HISTOGRAM(SDValue Op, SelectionDAG &DAG) const;
-  SDValue LowerPARTIAL_REDUCE_ADD(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerPARTIAL_REDUCE_MLA(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerDIV(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerVectorSRA_SRL_SHL(SDValue Op, SelectionDAG &DAG) const;