[AArch64][SVE] Instcombine SDIV to ASRD

[MattDevereau]

Summary

Instcombine SDIV to ASRD when the third operand of SDIV is a power of 2

[MattDevereau]

rename intr -> ASRD

[paulwalker-arm]

To match the equivalent functionality in D113281, can you also add support for signed divides by negative powers or two?

[peterwaller-arm]

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

1031

nit. I think the name 'instCombineSVESDiv' would be in keeping with the other functions in this file.

1044

nit. Extraneous blank.

1053

nit. Suggest extracting the constant as its own variable to make 'CreateIntrinsic' a bit tidier.

[MattDevereau]

Added negation of ASRD if the splat exponent is negative

[MattDevereau]

Force int32 type and add i64 tests

[peterwaller-arm]

LGTM with a nit.

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

1046

I think this would be better called Divisor.

[paulwalker-arm]

Just a couple of things to consider before committing.

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

1040–1041

FYI: If you use dyn_cast_or_null<ConstantInt>(SplatValue) you can remove this extra check.

1050

You shouldn't need this as the exponent cannot be negative and the intrinsic it's passed to (aarch64_sve_asrd) does not support this operand being negative either.

1058

As above.

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

if (match(Index, m_Intrinsic<Intrinsic::aarch64_sve_index>(
     (void)Builder.CreateMaskedStore(Val, Ptr, Alignment, Mask);
 
     return IC.eraseInstFromFunction(II);
   }
 
   return None;
 }
 
+static Optional<Instruction *> instCombineSVESDIV(InstCombiner &IC,
+                                                  IntrinsicInst &II) {
+  IRBuilder<> Builder(II.getContext());
+  Builder.SetInsertPoint(&II);
+  Type *Int32Ty = Builder.getInt32Ty();
+  Value *Pred = II.getOperand(0);
+  Value *Vec = II.getOperand(1);
+  Value *DivVec = II.getOperand(2);
+
+  if (!isSplatValue(DivVec))
+    return None;
+  Value *SplatValue = getSplatValue(DivVec);
+  ConstantInt *SplatConstantInt = dyn_cast<ConstantInt>(SplatValue);
+  if (!SplatConstantInt)
+    return None;
+  APInt Exponent = SplatConstantInt->getValue();
+
+  if (Exponent.isPowerOf2()) {
+    Constant *ExponentLog2 =
+        ConstantInt::get(Int32Ty, Exponent.logBase2(), true);
+    auto ASRD = Builder.CreateIntrinsic(
+        Intrinsic::aarch64_sve_asrd, {II.getType()}, {Pred, Vec, ExponentLog2});
+    return IC.replaceInstUsesWith(II, ASRD);
+  }
+  if (Exponent.isNegatedPowerOf2()) {
+    Exponent.negate();
+    Constant *ExponentLog2 =
+        ConstantInt::get(Int32Ty, Exponent.logBase2(), true);
+    auto ASRD = Builder.CreateIntrinsic(
+        Intrinsic::aarch64_sve_asrd, {II.getType()}, {Pred, Vec, ExponentLog2});
+    auto NEG = Builder.CreateIntrinsic(Intrinsic::aarch64_sve_neg,
+                                       {ASRD->getType()}, {ASRD, Pred, ASRD});
+    return IC.replaceInstUsesWith(II, NEG);
+  }
+
+  return None;
+}
+
 Optional<Instruction *>
 AArch64TTIImpl::instCombineIntrinsic(InstCombiner &IC,
                                      IntrinsicInst &II) const {
   Intrinsic::ID IID = II.getIntrinsicID();
   switch (IID) {
   default:
     break;
   case Intrinsic::aarch64_sve_convert_from_svbool:
AArch64TTIImpl::instCombineIntrinsic(InstCombiner &IC,
   case Intrinsic::aarch64_sve_ld1_gather_index:
     return instCombineLD1GatherIndex(IC, II);
   case Intrinsic::aarch64_sve_st1_scatter_index:
     return instCombineST1ScatterIndex(IC, II);
   case Intrinsic::aarch64_sve_ld1:
     return instCombineSVELD1(IC, II, DL);
   case Intrinsic::aarch64_sve_st1:
     return instCombineSVEST1(IC, II, DL);
+  case Intrinsic::aarch64_sve_sdiv:
+    return instCombineSVESDIV(IC, II);
   }
 
   return None;
 }
 
 bool AArch64TTIImpl::isWideningInstruction(Type *DstTy, unsigned Opcode,
                                            ArrayRef<const Value *> Args) {
 

llvm/test/Transforms/InstCombine/AArch64/sve-intrinsic-sdiv.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -instcombine < %s | FileCheck %s
+
+target triple = "aarch64-unknown-linux-gnu"
+
+define <vscale x 4 x i32> @sdiv_i32(<vscale x 4 x i32> %a, <vscale x 4 x i1> %pg) #0 {
+; CHECK-LABEL: @sdiv_i32(
+; CHECK-NEXT:    [[TMP1:%.*]] = call <vscale x 4 x i32> @llvm.aarch64.sve.asrd.nxv4i32(<vscale x 4 x i1> [[PG:%.*]], <vscale x 4 x i32> [[A:%.*]], i32 23)
+; CHECK-NEXT:    ret <vscale x 4 x i32> [[TMP1]]
+;
+  %out = call <vscale x 4 x i32> @llvm.aarch64.sve.sdiv.nxv4i32(<vscale x 4 x i1> %pg, <vscale x 4 x i32> %a, <vscale x 4 x i32> shufflevector (<vscale x 4 x i32> insertelement (<vscale x 4 x i32> poison, i32 8388608, i32 0), <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer))
+  ret <vscale x 4 x i32> %out
+}
+
+define <vscale x 4 x i32> @sdiv_i32_neg(<vscale x 4 x i32> %a, <vscale x 4 x i1> %pg) #0 {
+; CHECK-LABEL: @sdiv_i32_neg(
+; CHECK-NEXT:    [[TMP1:%.*]] = call <vscale x 4 x i32> @llvm.aarch64.sve.asrd.nxv4i32(<vscale x 4 x i1> [[PG:%.*]], <vscale x 4 x i32> [[A:%.*]], i32 23)
+; CHECK-NEXT:    [[TMP2:%.*]] = call <vscale x 4 x i32> @llvm.aarch64.sve.neg.nxv4i32(<vscale x 4 x i32> [[TMP1]], <vscale x 4 x i1> [[PG]], <vscale x 4 x i32> [[TMP1]])
+; CHECK-NEXT:    ret <vscale x 4 x i32> [[TMP2]]
+;
+  %out = call <vscale x 4 x i32> @llvm.aarch64.sve.sdiv.nxv4i32(<vscale x 4 x i1> %pg, <vscale x 4 x i32> %a, <vscale x 4 x i32> shufflevector (<vscale x 4 x i32> insertelement (<vscale x 4 x i32> poison, i32 -8388608, i32 0), <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer))
+  ret <vscale x 4 x i32> %out
+}
+
+define <vscale x 2 x i64> @sdiv_i64(<vscale x 2 x i64> %a, <vscale x 2 x i1> %pg) #0 {
+; CHECK-LABEL: @sdiv_i64(
+; CHECK-NEXT:    [[TMP1:%.*]] = call <vscale x 2 x i64> @llvm.aarch64.sve.asrd.nxv2i64(<vscale x 2 x i1> [[PG:%.*]], <vscale x 2 x i64> [[A:%.*]], i32 23)
+; CHECK-NEXT:    ret <vscale x 2 x i64> [[TMP1]]
+;
+  %out = call <vscale x 2 x i64> @llvm.aarch64.sve.sdiv.nxv2i64(<vscale x 2 x i1> %pg, <vscale x 2 x i64> %a, <vscale x 2 x i64> shufflevector (<vscale x 2 x i64> insertelement (<vscale x 2 x i64> poison, i64 8388608, i64 0), <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer))
+  ret <vscale x 2 x i64> %out
+}
+
+define <vscale x 2 x i64> @sdiv_i64_neg(<vscale x 2 x i64> %a, <vscale x 2 x i1> %pg) #0 {
+; CHECK-LABEL: @sdiv_i64_neg(
+; CHECK-NEXT:    [[TMP1:%.*]] = call <vscale x 2 x i64> @llvm.aarch64.sve.asrd.nxv2i64(<vscale x 2 x i1> [[PG:%.*]], <vscale x 2 x i64> [[A:%.*]], i32 23)
+; CHECK-NEXT:    [[TMP2:%.*]] = call <vscale x 2 x i64> @llvm.aarch64.sve.neg.nxv2i64(<vscale x 2 x i64> [[TMP1]], <vscale x 2 x i1> [[PG]], <vscale x 2 x i64> [[TMP1]])
+; CHECK-NEXT:    ret <vscale x 2 x i64> [[TMP2]]
+;
+  %out = call <vscale x 2 x i64> @llvm.aarch64.sve.sdiv.nxv2i64(<vscale x 2 x i1> %pg, <vscale x 2 x i64> %a, <vscale x 2 x i64> shufflevector (<vscale x 2 x i64> insertelement (<vscale x 2 x i64> poison, i64 -8388608, i64 0), <vscale x 2 x i64> poison, <vscale x 2 x i32> zeroinitializer))
+  ret <vscale x 2 x i64> %out
+}
+
+define <vscale x 4 x i32> @sdiv_i32_not_base2(<vscale x 4 x i32> %a, <vscale x 4 x i1> %pg) #0 {
+; CHECK-LABEL: @sdiv_i32_not_base2(
+; CHECK-NEXT:    [[OUT:%.*]] = call <vscale x 4 x i32> @llvm.aarch64.sve.sdiv.nxv4i32(<vscale x 4 x i1> [[PG:%.*]], <vscale x 4 x i32> [[A:%.*]], <vscale x 4 x i32> shufflevector (<vscale x 4 x i32> insertelement (<vscale x 4 x i32> poison, i32 8388607, i32 0), <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer))
+; CHECK-NEXT:    ret <vscale x 4 x i32> [[OUT]]
+;
+  %out = call <vscale x 4 x i32> @llvm.aarch64.sve.sdiv.nxv4i32(<vscale x 4 x i1> %pg, <vscale x 4 x i32> %a, <vscale x 4 x i32> shufflevector (<vscale x 4 x i32> insertelement (<vscale x 4 x i32> poison, i32 8388607, i32 0), <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer))
+  ret <vscale x 4 x i32> %out
+}
+
+define <vscale x 4 x i32> @sdiv_i32_not_base2_neg(<vscale x 4 x i32> %a, <vscale x 4 x i1> %pg) #0 {
+; CHECK-LABEL: @sdiv_i32_not_base2_neg(
+; CHECK-NEXT:    [[OUT:%.*]] = call <vscale x 4 x i32> @llvm.aarch64.sve.sdiv.nxv4i32(<vscale x 4 x i1> [[PG:%.*]], <vscale x 4 x i32> [[A:%.*]], <vscale x 4 x i32> shufflevector (<vscale x 4 x i32> insertelement (<vscale x 4 x i32> poison, i32 -8388607, i32 0), <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer))
+; CHECK-NEXT:    ret <vscale x 4 x i32> [[OUT]]
+;
+  %out = call <vscale x 4 x i32> @llvm.aarch64.sve.sdiv.nxv4i32(<vscale x 4 x i1> %pg, <vscale x 4 x i32> %a, <vscale x 4 x i32> shufflevector (<vscale x 4 x i32> insertelement (<vscale x 4 x i32> poison, i32 -8388607, i32 0), <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer))
+  ret <vscale x 4 x i32> %out
+}
+
+define <vscale x 4 x i32> @sdiv_i32_not_zero(<vscale x 4 x i32> %a, <vscale x 4 x i1> %pg) #0 {
+; CHECK-LABEL: @sdiv_i32_not_zero(
+; CHECK-NEXT:    [[OUT:%.*]] = call <vscale x 4 x i32> @llvm.aarch64.sve.sdiv.nxv4i32(<vscale x 4 x i1> [[PG:%.*]], <vscale x 4 x i32> [[A:%.*]], <vscale x 4 x i32> zeroinitializer)
+; CHECK-NEXT:    ret <vscale x 4 x i32> [[OUT]]
+;
+  %out = call <vscale x 4 x i32> @llvm.aarch64.sve.sdiv.nxv4i32(<vscale x 4 x i1> %pg, <vscale x 4 x i32> %a, <vscale x 4 x i32> shufflevector (<vscale x 4 x i32> insertelement (<vscale x 4 x i32> poison, i32 0, i32 0), <vscale x 4 x i32> poison, <vscale x 4 x i32> zeroinitializer))
+  ret <vscale x 4 x i32> %out
+}
+
+
+declare <vscale x 4 x i32> @llvm.aarch64.sve.sdiv.nxv4i32(<vscale x 4 x i1>, <vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 2 x i64> @llvm.aarch64.sve.sdiv.nxv2i64(<vscale x 2 x i1>, <vscale x 2 x i64>, <vscale x 2 x i64>)
+
+attributes #0 = { "target-features"="+sve" }