[AArch64][GlobalISel] Combine vecreduce(ext) to {U/S}ADDLV

llvm/lib/Target/AArch64/AArch64Combine.td

@@ -44,13 +44,22 @@ def ext_addv_to_udot_addv : GICombineRule<
 >;
 }
 
+def ext_uaddv_to_uaddlv_matchinfo : GIDefMatchData<"std::pair<Register, bool>">;
+def ext_uaddv_to_uaddlv : GICombineRule<
+  (defs root:$root, ext_uaddv_to_uaddlv_matchinfo:$matchinfo),
+  (match (wip_match_opcode G_VECREDUCE_ADD):$root,
+         [{ return matchExtUaddvToUaddlv(*${root}, MRI, ${matchinfo}); }]),
+  (apply [{ applyExtUaddvToUaddlv(*${root}, MRI, B, Observer, ${matchinfo}); }])
+>;
+
 def AArch64PreLegalizerCombiner: GICombiner<
   "AArch64PreLegalizerCombinerImpl", [all_combines,
                                       fconstant_to_constant,
                                       icmp_redundant_trunc,
                                       fold_global_offset,
                                       shuffle_to_extract,
-                                      ext_addv_to_udot_addv]> {
+                                      ext_addv_to_udot_addv,
+                                      ext_uaddv_to_uaddlv]> {
   let CombineAllMethodName = "tryCombineAllImpl";
 }
 

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

@@ -2464,6 +2464,7 @@ const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
     MAKE_CASE(AArch64ISD::SADDV)
     MAKE_CASE(AArch64ISD::UADDV)
     MAKE_CASE(AArch64ISD::UADDLV)
+    MAKE_CASE(AArch64ISD::SADDLV)
     MAKE_CASE(AArch64ISD::SDOT)
     MAKE_CASE(AArch64ISD::UDOT)
     MAKE_CASE(AArch64ISD::SMINV)

llvm/lib/Target/AArch64/AArch64ISelLowering.h

@@ -248,6 +248,7 @@ enum NodeType : unsigned {
 
   // Unsigned sum Long across Vector
   UADDLV,
+  SADDLV,
 
   // Add Pairwise of two vectors
   ADDP,

llvm/lib/Target/AArch64/AArch64InstrGISel.td

@@ -227,6 +227,18 @@ def G_SMULL : AArch64GenericInstruction {
   let hasSideEffects = 0;
 }
 
+def G_UADDLV : AArch64GenericInstruction {
+  let OutOperandList = (outs type0:$dst);
+  let InOperandList = (ins type0:$src1);
+  let hasSideEffects = 0;
+}
+
+def G_SADDLV : AArch64GenericInstruction {
+  let OutOperandList = (outs type0:$dst);
+  let InOperandList = (ins type0:$src1);
+  let hasSideEffects = 0;
+}
+
 def G_UDOT : AArch64GenericInstruction {
   let OutOperandList = (outs type0:$dst);
   let InOperandList = (ins type0:$src1, type0:$src2, type0:$src3);
@@ -282,6 +294,9 @@ def : GINodeEquiv<G_BSP, AArch64bsp>;
 def : GINodeEquiv<G_UMULL, AArch64umull>;
 def : GINodeEquiv<G_SMULL, AArch64smull>;
 
+def : GINodeEquiv<G_SADDLV, AArch64saddlv>;
+def : GINodeEquiv<G_UADDLV, AArch64uaddlv>;
+
 def : GINodeEquiv<G_UDOT, AArch64udot>;
 def : GINodeEquiv<G_SDOT, AArch64sdot>;
 

llvm/lib/Target/AArch64/AArch64InstrInfo.td

@@ -799,6 +799,7 @@ def AArch64uminv    : SDNode<"AArch64ISD::UMINV", SDT_AArch64UnaryVec>;
 def AArch64smaxv    : SDNode<"AArch64ISD::SMAXV", SDT_AArch64UnaryVec>;
 def AArch64umaxv    : SDNode<"AArch64ISD::UMAXV", SDT_AArch64UnaryVec>;
 def AArch64uaddlv   : SDNode<"AArch64ISD::UADDLV", SDT_AArch64uaddlp>;
+def AArch64saddlv   : SDNode<"AArch64ISD::SADDLV", SDT_AArch64uaddlp>;
 
 def AArch64uabd     : PatFrags<(ops node:$lhs, node:$rhs),
                                [(abdu node:$lhs, node:$rhs),
@@ -6680,17 +6681,25 @@ def : Pat<(v4i32 (AArch64uaddlv (v8i16 (AArch64uaddlp (v16i8 V128:$op))))),
 def : Pat<(v4i32 (AArch64uaddlv (v4i16 (AArch64uaddlp (v8i8 V64:$op))))),
           (v4i32 (SUBREG_TO_REG (i64 0), (UADDLVv8i8v V64:$op), hsub))>;
 
-def : Pat<(v4i32 (AArch64uaddlv (v8i8 V64:$Rn))),
-          (v4i32 (SUBREG_TO_REG (i64 0), (UADDLVv8i8v V64:$Rn), hsub))>;
+multiclass SIMDAcrossLaneLongReductionIntrinsic<string Opc, SDPatternOperator addlv> {
+  def : Pat<(v4i32 (addlv (v8i8 V64:$Rn))),
+            (v4i32 (SUBREG_TO_REG (i64 0), (!cast<Instruction>(Opc#"v8i8v") V64:$Rn), hsub))>;
 
-def : Pat<(v4i32 (AArch64uaddlv (v4i16 V64:$Rn))),
-          (v4i32 (SUBREG_TO_REG (i64 0), (UADDLVv4i16v V64:$Rn), ssub))>;
+  def : Pat<(v4i32 (addlv (v4i16 V64:$Rn))),
+            (v4i32 (SUBREG_TO_REG (i64 0), (!cast<Instruction>(Opc#"v4i16v") V64:$Rn), ssub))>;
 
-def : Pat<(v4i32 (AArch64uaddlv (v16i8 V128:$Rn))),
-          (v4i32 (SUBREG_TO_REG (i64 0), (UADDLVv16i8v V128:$Rn), hsub))>;
+  def : Pat<(v4i32 (addlv (v16i8 V128:$Rn))),
+            (v4i32 (SUBREG_TO_REG (i64 0), (!cast<Instruction>(Opc#"v16i8v") V128:$Rn), hsub))>;
 
-def : Pat<(v4i32 (AArch64uaddlv (v8i16 V128:$Rn))),
-          (v4i32 (SUBREG_TO_REG (i64 0), (UADDLVv8i16v V128:$Rn), ssub))>;
+  def : Pat<(v4i32 (addlv (v8i16 V128:$Rn))),
+            (v4i32 (SUBREG_TO_REG (i64 0), (!cast<Instruction>(Opc#"v8i16v") V128:$Rn), ssub))>;
+
+  def : Pat<(v2i64 (addlv (v4i32 V128:$Rn))),
+            (v2i64 (SUBREG_TO_REG (i64 0), (!cast<Instruction>(Opc#"v4i32v") V128:$Rn), dsub))>;
+}
+
+defm : SIMDAcrossLaneLongReductionIntrinsic<"UADDLV", AArch64uaddlv>;
+defm : SIMDAcrossLaneLongReductionIntrinsic<"SADDLV", AArch64saddlv>;
 
 // Patterns for across-vector intrinsics, that have a node equivalent, that
 // returns a vector (with only the low lane defined) instead of a scalar.

llvm/lib/Target/AArch64/GISel/AArch64PreLegalizerCombiner.cpp

@@ -410,6 +410,150 @@ void applyExtAddvToUdotAddv(MachineInstr &MI, MachineRegisterInfo &MRI,
   MI.eraseFromParent();
 }
 
+// Matches {U/S}ADDV(ext(x)) => {U/S}ADDLV(x)
+// Ensure that the type coming from the extend instruction is the right size
+bool matchExtUaddvToUaddlv(MachineInstr &MI, MachineRegisterInfo &MRI,
+                           std::pair<Register, bool> &MatchInfo) {
+  assert(MI.getOpcode() == TargetOpcode::G_VECREDUCE_ADD &&
+         "Expected G_VECREDUCE_ADD Opcode");
+
+  // Check if the last instruction is an extend
+  MachineInstr *ExtMI = getDefIgnoringCopies(MI.getOperand(1).getReg(), MRI);
+  auto ExtOpc = ExtMI->getOpcode();
+
+  if (ExtOpc == TargetOpcode::G_ZEXT)
+    std::get<1>(MatchInfo) = 0;
+  else if (ExtOpc == TargetOpcode::G_SEXT)
+    std::get<1>(MatchInfo) = 1;
+  else
+    return false;
+
+  // Check if the source register is a valid type
+  Register ExtSrcReg = ExtMI->getOperand(1).getReg();
+  LLT ExtSrcTy = MRI.getType(ExtSrcReg);
+  LLT DstTy = MRI.getType(MI.getOperand(0).getReg());
+  if ((DstTy.getScalarSizeInBits() == 16 &&
+       ExtSrcTy.getNumElements() % 8 == 0 && ExtSrcTy.getNumElements() < 256) ||
+      (DstTy.getScalarSizeInBits() == 32 &&
+       ExtSrcTy.getNumElements() % 4 == 0) ||
+      (DstTy.getScalarSizeInBits() == 64 &&
+       ExtSrcTy.getNumElements() % 4 == 0)) {
+    std::get<0>(MatchInfo) = ExtSrcReg;
+    return true;
+  }
+  return false;
+}
+
+void applyExtUaddvToUaddlv(MachineInstr &MI, MachineRegisterInfo &MRI,
+                           MachineIRBuilder &B, GISelChangeObserver &Observer,
+                           std::pair<Register, bool> &MatchInfo) {
+  assert(MI.getOpcode() == TargetOpcode::G_VECREDUCE_ADD &&
+         "Expected G_VECREDUCE_ADD Opcode");
+
+  unsigned Opc = std::get<1>(MatchInfo) ? AArch64::G_SADDLV : AArch64::G_UADDLV;
+  Register SrcReg = std::get<0>(MatchInfo);
+  Register DstReg = MI.getOperand(0).getReg();
+  LLT SrcTy = MRI.getType(SrcReg);
+  LLT DstTy = MRI.getType(DstReg);
+
+  // If SrcTy has more elements than expected, split them into multiple
+  // insructions and sum the results
+  LLT MainTy;
+  SmallVector<Register, 1> WorkingRegisters;
+  unsigned SrcScalSize = SrcTy.getScalarSizeInBits();
+  unsigned SrcNumElem = SrcTy.getNumElements();
+  if ((SrcScalSize == 8 && SrcNumElem > 16) ||
+      (SrcScalSize == 16 && SrcNumElem > 8) ||
+      (SrcScalSize == 32 && SrcNumElem > 4)) {
+
+    LLT LeftoverTy;
+    SmallVector<Register, 4> LeftoverRegs;
+    if (SrcScalSize == 8)
+      MainTy = LLT::fixed_vector(16, 8);
+    else if (SrcScalSize == 16)
+      MainTy = LLT::fixed_vector(8, 16);
+    else if (SrcScalSize == 32)
+      MainTy = LLT::fixed_vector(4, 32);
+    else
+      llvm_unreachable("Source's Scalar Size not supported");
+
+    // Extract the parts and put each extracted sources through U/SADDLV and put
+    // the values inside a small vec
+    extractParts(SrcReg, SrcTy, MainTy, LeftoverTy, WorkingRegisters,
+                 LeftoverRegs, B, MRI);
+    for (unsigned I = 0; I < LeftoverRegs.size(); I++) {
+      WorkingRegisters.push_back(LeftoverRegs[I]);
+    }
+  } else {
+    WorkingRegisters.push_back(SrcReg);
+    MainTy = SrcTy;
+  }
+
+  unsigned MidScalarSize = MainTy.getScalarSizeInBits() * 2;
+  LLT MidScalarLLT = LLT::scalar(MidScalarSize);
+  Register zeroReg = B.buildConstant(LLT::scalar(64), 0).getReg(0);
+  for (unsigned I = 0; I < WorkingRegisters.size(); I++) {
+    // If the number of elements is too small to build an instruction, extend
+    // its size before applying addlv
+    LLT WorkingRegTy = MRI.getType(WorkingRegisters[I]);
+    if ((WorkingRegTy.getScalarSizeInBits() == 8) &&
+        (WorkingRegTy.getNumElements() == 4)) {
+      WorkingRegisters[I] =
+          B.buildInstr(std::get<1>(MatchInfo) ? TargetOpcode::G_SEXT
+                                              : TargetOpcode::G_ZEXT,
+                       {LLT::fixed_vector(4, 16)}, {WorkingRegisters[I]})
+              .getReg(0);
+    }
+
+    // Generate the {U/S}ADDLV instruction, whose output is always double of the
+    // Src's Scalar size
+    LLT addlvTy = MidScalarSize <= 32 ? LLT::fixed_vector(4, 32)
+                                      : LLT::fixed_vector(2, 64);
+    Register addlvReg =
+        B.buildInstr(Opc, {addlvTy}, {WorkingRegisters[I]}).getReg(0);
+
+    // The output from {U/S}ADDLV gets placed in the lowest lane of a v4i32 or
+    // v2i64 register.
+    //     i16, i32 results uses v4i32 registers
+    //     i64      results uses v2i64 registers
+    // Therefore we have to extract/truncate the the value to the right type
+    if (MidScalarSize == 32 || MidScalarSize == 64) {
+      WorkingRegisters[I] = B.buildInstr(AArch64::G_EXTRACT_VECTOR_ELT,
+                                         {MidScalarLLT}, {addlvReg, zeroReg})
+                                .getReg(0);
+    } else {
+      Register extractReg = B.buildInstr(AArch64::G_EXTRACT_VECTOR_ELT,
+                                         {LLT::scalar(32)}, {addlvReg, zeroReg})
+                                .getReg(0);
+      WorkingRegisters[I] =
+          B.buildTrunc({MidScalarLLT}, {extractReg}).getReg(0);
+    }
+  }
+
+  Register outReg;
+  if (WorkingRegisters.size() > 1) {
+    outReg = B.buildAdd(MidScalarLLT, WorkingRegisters[0], WorkingRegisters[1])
+                 .getReg(0);
+    for (unsigned I = 2; I < WorkingRegisters.size(); I++) {
+      outReg = B.buildAdd(MidScalarLLT, outReg, WorkingRegisters[I]).getReg(0);
+    }
+  } else {
+    outReg = WorkingRegisters[0];
+  }
+
+  if (DstTy.getScalarSizeInBits() > MidScalarSize) {
+    // Handle the scalar value if the DstTy's Scalar Size is more than double
+    // Src's ScalarType
+    B.buildInstr(std::get<1>(MatchInfo) ? TargetOpcode::G_SEXT
+                                        : TargetOpcode::G_ZEXT,
+                 {DstReg}, {outReg});
+  } else {
+    B.buildCopy(DstReg, outReg);
+  }
+
+  MI.eraseFromParent();
+}
+
 bool tryToSimplifyUADDO(MachineInstr &MI, MachineIRBuilder &B,
                         CombinerHelper &Helper, GISelChangeObserver &Observer) {
   // Try simplify G_UADDO with 8 or 16 bit operands to wide G_ADD and TBNZ if