[ARM] Add support for MVE vaddv

[samtebbs]

Summary

This patch adds vecreduce_add and the relevant instruction selection for vaddv.

[dmgreen]

Context please!

[samtebbs]

In D66085#1625150, @dmgreen wrote:

Context please!

Added!

[dmgreen]

Thanks. I like the way this keeps the vecreduce_add as legal.

llvm/lib/Target/ARM/ARMInstrMVE.td

553

MQPR, over QPR.

llvm/test/CodeGen/ARM/vaddv.ll

1 ↗

(On Diff #215101)

Can you move this test to test/Codegen/Thumb2/mve-vaddv.ll, to be with the others.

[dmgreen]

LGTM, thanks.

llvm/include/llvm/Target/TargetSelectionDAG.td

def SDTVecShuffle : SDTypeProfile<1, 2, [
   SDTCisSameAs<0, 1>, SDTCisSameAs<1, 2>
 ]>;
 def SDTVecExtract : SDTypeProfile<1, 2, [   // vector extract
   SDTCisEltOfVec<0, 1>, SDTCisPtrTy<2>
 ]>;
 def SDTVecInsert : SDTypeProfile<1, 3, [    // vector insert
   SDTCisEltOfVec<2, 1>, SDTCisSameAs<0, 1>, SDTCisPtrTy<3>
 ]>;
+def SDTVecReduce : SDTypeProfile<1, 1, [    // vector reduction
+  SDTCisInt<0>, SDTCisVec<1>
+]>;
 
 def SDTSubVecExtract : SDTypeProfile<1, 2, [// subvector extract
   SDTCisSubVecOfVec<0,1>, SDTCisInt<2>
 ]>;
 def SDTSubVecInsert : SDTypeProfile<1, 3, [ // subvector insert
   SDTCisSubVecOfVec<2, 1>, SDTCisSameAs<0,1>, SDTCisInt<3>
 ]>;
 

 def zext       : SDNode<"ISD::ZERO_EXTEND", SDTIntExtendOp>;
 def anyext     : SDNode<"ISD::ANY_EXTEND" , SDTIntExtendOp>;
 def trunc      : SDNode<"ISD::TRUNCATE"   , SDTIntTruncOp>;
 def bitconvert : SDNode<"ISD::BITCAST"    , SDTUnaryOp>;
 def addrspacecast : SDNode<"ISD::ADDRSPACECAST", SDTUnaryOp>;
 def extractelt : SDNode<"ISD::EXTRACT_VECTOR_ELT", SDTVecExtract>;
 def insertelt  : SDNode<"ISD::INSERT_VECTOR_ELT", SDTVecInsert>;
 
+def vecreduce_add  : SDNode<"ISD::VECREDUCE_ADD", SDTVecReduce>;
+
 def fadd       : SDNode<"ISD::FADD"       , SDTFPBinOp, [SDNPCommutative]>;
 def fsub       : SDNode<"ISD::FSUB"       , SDTFPBinOp>;
 def fmul       : SDNode<"ISD::FMUL"       , SDTFPBinOp, [SDNPCommutative]>;
 def fdiv       : SDNode<"ISD::FDIV"       , SDTFPBinOp>;
 def frem       : SDNode<"ISD::FREM"       , SDTFPBinOp>;
 def fma        : SDNode<"ISD::FMA"        , SDTFPTernaryOp>;
 def fmad       : SDNode<"ISD::FMAD"       , SDTFPTernaryOp>;
 def fabs       : SDNode<"ISD::FABS"       , SDTFPUnaryOp>;

llvm/lib/Target/ARM/ARMISelLowering.cpp

for (auto VT : IntTypes) {
 
     // No native support for these.
     setOperationAction(ISD::UDIV, VT, Expand);
     setOperationAction(ISD::SDIV, VT, Expand);
     setOperationAction(ISD::UREM, VT, Expand);
     setOperationAction(ISD::SREM, VT, Expand);
     setOperationAction(ISD::CTPOP, VT, Expand);
 
+    // Vector reductions
+    setOperationAction(ISD::VECREDUCE_ADD, VT, Legal);
+
     if (!HasMVEFP) {
       setOperationAction(ISD::SINT_TO_FP, VT, Expand);
       setOperationAction(ISD::UINT_TO_FP, VT, Expand);
       setOperationAction(ISD::FP_TO_SINT, VT, Expand);
       setOperationAction(ISD::FP_TO_UINT, VT, Expand);
     }
 
     // Pre and Post inc are supported on loads and stores

llvm/lib/Target/ARM/ARMInstrMVE.td

 
 defm MVE_VADDVs8  : MVE_VADDV_A<"s8",  0b0, 0b00>;
 defm MVE_VADDVs16 : MVE_VADDV_A<"s16", 0b0, 0b01>;
 defm MVE_VADDVs32 : MVE_VADDV_A<"s32", 0b0, 0b10>;
 defm MVE_VADDVu8  : MVE_VADDV_A<"u8",  0b1, 0b00>;
 defm MVE_VADDVu16 : MVE_VADDV_A<"u16", 0b1, 0b01>;
 defm MVE_VADDVu32 : MVE_VADDV_A<"u32", 0b1, 0b10>;
 
+let Predicates = [HasMVEInt] in {
+  def : Pat<(i32 (vecreduce_add (v4i32 MQPR:$src))), (i32 (MVE_VADDVu32no_acc $src))>;
+  def : Pat<(i32 (vecreduce_add (v8i16 MQPR:$src))), (i32 (MVE_VADDVu16no_acc $src))>;
+  def : Pat<(i32 (vecreduce_add (v16i8 MQPR:$src))), (i32 (MVE_VADDVu8no_acc $src))>;
+}
+
 class MVE_VADDLV<string iname, string suffix, dag iops, string cstr,
                bit A, bit U, list<dag> pattern=[]>
   : MVE_rDest<(outs tGPREven:$RdaLo, tGPROdd:$RdaHi), iops, NoItinerary, iname,
               suffix, "$RdaLo, $RdaHi, $Qm", cstr, pattern> {
   bits<3> Qm;
   bits<4> RdaLo;
   bits<4> RdaHi;
 

llvm/lib/Target/ARM/ARMTargetTransformInfo.h

public:
   unsigned getMaxInterleaveFactor(unsigned VF) {
     return ST->getMaxInterleaveFactor();
   }
 
   int getMemcpyCost(const Instruction *I);
 
   int getShuffleCost(TTI::ShuffleKind Kind, Type *Tp, int Index, Type *SubTp);
 
+  bool useReductionIntrinsic(unsigned Opcode, Type *Ty,
+                             TTI::ReductionFlags Flags) const;
+
+  bool shouldExpandReduction(const IntrinsicInst *II) const {
+    return false;
+  }
+
   int getCastInstrCost(unsigned Opcode, Type *Dst, Type *Src,
                        const Instruction *I = nullptr);
 
   int getCmpSelInstrCost(unsigned Opcode, Type *ValTy, Type *CondTy,
                          const Instruction *I = nullptr);
 
   int getVectorInstrCost(unsigned Opcode, Type *Val, unsigned Index);
 

llvm/lib/Target/ARM/ARMTargetTransformInfo.cpp

void ARMTTIImpl::getUnrollingPreferences(Loop *L, ScalarEvolution &SE,
   UP.UnrollAndJam = true;
   UP.UnrollAndJamInnerLoopThreshold = 60;
 
   // Force unrolling small loops can be very useful because of the branch
   // taken cost of the backedge.
   if (Cost < 12)
     UP.Force = true;
 }
+
+bool ARMTTIImpl::useReductionIntrinsic(unsigned Opcode, Type *Ty,
+                                       TTI::ReductionFlags Flags) const {
+  assert(isa<VectorType>(Ty) && "Expected Ty to be a vector type");
+  unsigned ScalarBits = Ty->getScalarSizeInBits();
+  if (!ST->hasMVEIntegerOps())
+    return false;
+
+  switch (Opcode) {
+  case Instruction::FAdd:
+  case Instruction::FMul:
+  case Instruction::And:
+  case Instruction::Or:
+  case Instruction::Xor:
+  case Instruction::Mul:
+  case Instruction::ICmp:
+  case Instruction::FCmp:
+    return false;
+  case Instruction::Add:
+    return ScalarBits * Ty->getVectorNumElements() == 128;
+  default:
+    llvm_unreachable("Unhandled reduction opcode");
+  }
+  return false;
+}

llvm/test/CodeGen/Thumb2/mve-vaddv.ll

+; RUN: llc -mtriple=thumbv8.1m.main-arm-none-eabi -mattr=+mve.fp %s -o - | FileCheck %s
+
+declare i32 @llvm.experimental.vector.reduce.add.i32.v4i32(<4 x i32>)
+define arm_aapcs_vfpcc i32 @vaddv_v4i32_i32(<4 x i32> %s1) {
+; CHECK-LABEL: vaddv_v4i32_i32:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    vaddv.u32 r0, q0
+; CHECK-NEXT:    bx lr
+entry:
+  %r = call i32 @llvm.experimental.vector.reduce.add.i32.v4i32(<4 x i32> %s1)
+  ret i32 %r
+}
+
+declare i16 @llvm.experimental.vector.reduce.add.i16.v8i16(<8 x i16>)
+define arm_aapcs_vfpcc i16 @vaddv_v16i16_i16(<8 x i16> %s1) {
+; CHECK-LABEL: vaddv_v16i16_i16:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    vaddv.u16 r0, q0
+; CHECK-NEXT:    bx lr
+entry:
+  %r = call i16 @llvm.experimental.vector.reduce.add.i16.v8i16(<8 x i16> %s1)
+  ret i16 %r
+}
+
+declare i8 @llvm.experimental.vector.reduce.add.i8.v16i8(<16 x i8>)
+define arm_aapcs_vfpcc i8 @vaddv_v16i8_i8(<16 x i8> %s1) {
+; CHECK-LABEL: vaddv_v16i8_i8:
+; CHECK:       @ %bb.0: @ %entry
+; CHECK-NEXT:    vaddv.u8 r0, q0
+; CHECK-NEXT:    bx lr
+entry:
+  %r = call i8 @llvm.experimental.vector.reduce.add.i8.v16i8(<16 x i8> %s1)
+  ret i8 %r
+}