[AArch64][SVE][InstCombine] Eliminate redundant chains of tuple get/set

[mnadeem]

Summary

Probably makes no difference in codegen but makes the IR simpler and easier to read.

[david-arm]

LGTM!

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

static Optional<Instruction *> instCombineSVETBL(InstCombiner &IC,
   auto *Extract = Builder.CreateExtractElement(OpVal, SplatValue);
   auto *VectorSplat =
       Builder.CreateVectorSplat(VTy->getElementCount(), Extract);
 
   VectorSplat->takeName(&II);
   return IC.replaceInstUsesWith(II, VectorSplat);
 }
 
+static Optional<Instruction *> instCombineSVETupleGet(InstCombiner &IC,
+                                                      IntrinsicInst &II) {
+  // Try to remove sequences of tuple get/set.
+  Value *SetTuple, *SetIndex, *SetValue;
+  auto *GetTuple = II.getArgOperand(0);
+  auto *GetIndex = II.getArgOperand(1);
+  // Check that we have tuple_get(GetTuple, GetIndex) where GetTuple is a
+  // call to tuple_set i.e. tuple_set(SetTuple, SetIndex, SetValue).
+  // Make sure that the types of the current intrinsic and SetValue match
+  // in order to safely remove the sequence.
+  if (!match(GetTuple,
+             m_Intrinsic<Intrinsic::aarch64_sve_tuple_set>(
+                 m_Value(SetTuple), m_Value(SetIndex), m_Value(SetValue))) ||
+      SetValue->getType() != II.getType())
+    return None;
+  // Case where we get the same index right after setting it.
+  // tuple_get(tuple_set(SetTuple, SetIndex, SetValue), GetIndex) --> SetValue
+  if (GetIndex == SetIndex)
+    return IC.replaceInstUsesWith(II, SetValue);
+  // If we are getting a different index than what was set in the tuple_set
+  // intrinsic. We can just set the input tuple to the one up in the chain.
+  // tuple_get(tuple_set(SetTuple, SetIndex, SetValue), GetIndex)
+  // --> tuple_get(SetTuple, GetIndex)
+  return IC.replaceOperand(II, 0, SetTuple);
+}
+
 static Optional<Instruction *> instCombineSVEZip(InstCombiner &IC,
                                                  IntrinsicInst &II) {
   // zip1(uzp1(A, B), uzp2(A, B)) --> A
   // zip2(uzp1(A, B), uzp2(A, B)) --> B
   Value *Uzp1Op0, *Uzp1Op1;
   if (match(II.getArgOperand(0), m_Intrinsic<Intrinsic::aarch64_sve_uzp1>(
                                      m_Value(Uzp1Op0), m_Value(Uzp1Op1))) &&
       match(II.getArgOperand(1), m_Intrinsic<Intrinsic::aarch64_sve_uzp2>(
case Intrinsic::aarch64_sve_fmul:
     return instCombineSVEVectorMul(IC, II);
   case Intrinsic::aarch64_sve_tbl:
     return instCombineSVETBL(IC, II);
   case Intrinsic::aarch64_sve_uunpkhi:
   case Intrinsic::aarch64_sve_uunpklo:
   case Intrinsic::aarch64_sve_sunpkhi:
   case Intrinsic::aarch64_sve_sunpklo:
     return instCombineSVEUnpack(IC, II);
+  case Intrinsic::aarch64_sve_tuple_get:
+    return instCombineSVETupleGet(IC, II);
   case Intrinsic::aarch64_sve_zip1:
   case Intrinsic::aarch64_sve_zip2:
     return instCombineSVEZip(IC, II);
   }
 
   return None;
 }
 

llvm/test/Transforms/InstCombine/AArch64/sve-intrinsic-opts-tuple-get.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"
+
+; This stores %a using st4 after reversing the 4 tuples. Check that the
+; redundant sequences of get/set are eliminated.
+define void @redundant_tuple_get_set(<vscale x 64 x i8> %a, i8* %ptr) #0 {
+; CHECK-LABEL: @redundant_tuple_get_set(
+; CHECK-NEXT:    [[TMP1:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv64i8(<vscale x 64 x i8> [[A:%.*]], i32 3)
+; CHECK-NEXT:    [[TMP2:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv64i8(<vscale x 64 x i8> [[A]], i32 0)
+; CHECK-NEXT:    [[TMP3:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv64i8(<vscale x 64 x i8> [[A]], i32 2)
+; CHECK-NEXT:    [[TMP4:%.*]] = call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv64i8(<vscale x 64 x i8> [[A]], i32 1)
+; CHECK-NEXT:    call void @llvm.aarch64.sve.st4.nxv16i8(<vscale x 16 x i8> [[TMP1]], <vscale x 16 x i8> [[TMP3]], <vscale x 16 x i8> [[TMP4]], <vscale x 16 x i8> [[TMP2]], <vscale x 16 x i1> shufflevector (<vscale x 16 x i1> insertelement (<vscale x 16 x i1> poison, i1 true, i32 0), <vscale x 16 x i1> poison, <vscale x 16 x i32> zeroinitializer), i8* [[PTR:%.*]])
+; CHECK-NEXT:    ret void
+;
+  %1 = call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv64i8(<vscale x 64 x i8> %a, i32 3)
+  %2 = call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv64i8(<vscale x 64 x i8> %a, i32 0)
+  %3 = call <vscale x 64 x i8> @llvm.aarch64.sve.tuple.set.nxv64i8.nxv16i8(<vscale x 64 x i8> %a, i32 3, <vscale x 16 x i8> %2)
+  %4 = call <vscale x 64 x i8> @llvm.aarch64.sve.tuple.set.nxv64i8.nxv16i8(<vscale x 64 x i8> %3, i32 0, <vscale x 16 x i8> %1)
+  %5 = call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv64i8(<vscale x 64 x i8> %4, i32 2)
+  %6 = call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv64i8(<vscale x 64 x i8> %4, i32 1)
+  %7 = call <vscale x 64 x i8> @llvm.aarch64.sve.tuple.set.nxv64i8.nxv16i8(<vscale x 64 x i8> %4, i32 2, <vscale x 16 x i8> %6)
+  %8 = call <vscale x 64 x i8> @llvm.aarch64.sve.tuple.set.nxv64i8.nxv16i8(<vscale x 64 x i8> %7, i32 1, <vscale x 16 x i8> %5)
+  %9 = call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv64i8(<vscale x 64 x i8> %8, i32 0)
+  %10 = call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv64i8(<vscale x 64 x i8> %8, i32 1)
+  %11 = call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv64i8(<vscale x 64 x i8> %8, i32 2)
+  %12 = call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv64i8(<vscale x 64 x i8> %8, i32 3)
+  call void @llvm.aarch64.sve.st4.nxv16i8(<vscale x 16 x i8> %9, <vscale x 16 x i8> %10, <vscale x 16 x i8> %11, <vscale x 16 x i8> %12, <vscale x 16 x i1> shufflevector (<vscale x 16 x i1> insertelement (<vscale x 16 x i1> poison, i1 true, i32 0), <vscale x 16 x i1> poison, <vscale x 16 x i32> zeroinitializer), i8* %ptr)
+  ret void
+}
+
+declare <vscale x 64 x i8> @llvm.aarch64.sve.tuple.set.nxv64i8.nxv16i8(<vscale x 64 x i8>, i32, <vscale x 16 x i8>)
+declare <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv64i8(<vscale x 64 x i8>, i32)
+declare void @llvm.aarch64.sve.st4.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i1>, i8*)
+
+attributes #0 = { "target-features"="+sve" }