[AArch64][GlobalISel] TableGen Selection for G_VECREDUCE_ADD (llvm#70785)

Instruction Selection for G_VECREDUCE_ADD now uses TableGen

Author: chuongg3

llvm/include/llvm/Target/GlobalISel/SelectionDAGCompat.td

@@ -177,6 +177,7 @@ def : GINodeEquiv<G_VECREDUCE_UMIN, vecreduce_umin>;
 def : GINodeEquiv<G_VECREDUCE_UMAX, vecreduce_umax>;
 def : GINodeEquiv<G_VECREDUCE_SMIN, vecreduce_smin>;
 def : GINodeEquiv<G_VECREDUCE_SMAX, vecreduce_smax>;
+def : GINodeEquiv<G_VECREDUCE_ADD, vecreduce_add>;
 
 def : GINodeEquiv<G_STRICT_FADD, strict_fadd>;
 def : GINodeEquiv<G_STRICT_FSUB, strict_fsub>;

llvm/lib/Target/AArch64/AArch64InstrInfo.td

@@ -6676,6 +6676,22 @@ def : Pat<(i32 (and (i32 (vector_extract (opNode (v8i16 V128:$Rn)), (i64 0))),
           ssub))>;
 }
 
+// For vecreduce_add, used by GlobalISel not SDAG
+def : Pat<(i8 (vecreduce_add (v8i8 V64:$Rn))), 
+          (i8 (ADDVv8i8v V64:$Rn))>;
+def : Pat<(i8 (vecreduce_add (v16i8 V128:$Rn))), 
+          (i8 (ADDVv16i8v V128:$Rn))>;
+def : Pat<(i16 (vecreduce_add (v4i16 V64:$Rn))), 
+          (i16 (ADDVv4i16v V64:$Rn))>;
+def : Pat<(i16 (vecreduce_add (v8i16 V128:$Rn))), 
+          (i16 (ADDVv8i16v V128:$Rn))>;
+def : Pat<(i32 (vecreduce_add (v2i32 V64:$Rn))), 
+          (i32 (EXTRACT_SUBREG (ADDPv2i32 V64:$Rn, V64:$Rn), ssub))>;
+def : Pat<(i32 (vecreduce_add (v4i32 V128:$Rn))), 
+          (i32 (ADDVv4i32v V128:$Rn))>;
+def : Pat<(i64 (vecreduce_add (v2i64 V128:$Rn))), 
+          (i64 (ADDPv2i64p V128:$Rn))>;
+
 defm : SIMDAcrossLanesSignedIntrinsic<"ADDV",  AArch64saddv>;
 // vaddv_[su]32 is special; -> ADDP Vd.2S,Vn.2S,Vm.2S; return Vd.s[0];Vn==Vm
 def : Pat<(v2i32 (AArch64saddv (v2i32 V64:$Rn))),

llvm/lib/Target/AArch64/GISel/AArch64InstructionSelector.cpp

@@ -3558,8 +3558,6 @@ bool AArch64InstructionSelector::select(MachineInstr &I) {
     return selectConcatVectors(I, MRI);
   case TargetOpcode::G_JUMP_TABLE:
     return selectJumpTable(I, MRI);
-  case TargetOpcode::G_VECREDUCE_ADD:
-    return selectReduction(I, MRI);
   case TargetOpcode::G_MEMCPY:
   case TargetOpcode::G_MEMCPY_INLINE:
   case TargetOpcode::G_MEMMOVE:
@@ -3578,49 +3576,6 @@ bool AArch64InstructionSelector::selectAndRestoreState(MachineInstr &I) {
   return Success;
 }
 
-bool AArch64InstructionSelector::selectReduction(MachineInstr &I,
-                                                 MachineRegisterInfo &MRI) {
-  Register VecReg = I.getOperand(1).getReg();
-  LLT VecTy = MRI.getType(VecReg);
-  if (I.getOpcode() == TargetOpcode::G_VECREDUCE_ADD) {
-    // For <2 x i32> ADDPv2i32 generates an FPR64 value, so we need to emit
-    // a subregister copy afterwards.
-    if (VecTy == LLT::fixed_vector(2, 32)) {
-      Register DstReg = I.getOperand(0).getReg();
-      auto AddP = MIB.buildInstr(AArch64::ADDPv2i32, {&AArch64::FPR64RegClass},
-                                 {VecReg, VecReg});
-      auto Copy = MIB.buildInstr(TargetOpcode::COPY, {DstReg}, {})
-                      .addReg(AddP.getReg(0), 0, AArch64::ssub)
-                      .getReg(0);
-      RBI.constrainGenericRegister(Copy, AArch64::FPR32RegClass, MRI);
-      I.eraseFromParent();
-      return constrainSelectedInstRegOperands(*AddP, TII, TRI, RBI);
-    }
-
-    unsigned Opc = 0;
-    if (VecTy == LLT::fixed_vector(16, 8))
-      Opc = AArch64::ADDVv16i8v;
-    else if (VecTy == LLT::fixed_vector(8, 8))
-      Opc = AArch64::ADDVv8i8v;
-    else if (VecTy == LLT::fixed_vector(8, 16))
-      Opc = AArch64::ADDVv8i16v;
-    else if (VecTy == LLT::fixed_vector(4, 16))
-      Opc = AArch64::ADDVv4i16v;
-    else if (VecTy == LLT::fixed_vector(4, 32))
-      Opc = AArch64::ADDVv4i32v;
-    else if (VecTy == LLT::fixed_vector(2, 64))
-      Opc = AArch64::ADDPv2i64p;
-    else {
-      LLVM_DEBUG(dbgs() << "Unhandled type for add reduction");
-      return false;
-    }
-    I.setDesc(TII.get(Opc));
-    return constrainSelectedInstRegOperands(I, TII, TRI, RBI);
-  }
-
-  return false;
-}
-
 bool AArch64InstructionSelector::selectMOPS(MachineInstr &GI,
                                             MachineRegisterInfo &MRI) {
   unsigned Mopcode;