[AArch64][SVE][InstCombine] last{a,b} of a splat vector

[mnadeem]

Summary

Replace last{a,b}(splat(X)) with X, irrespective of the predicate.

[sdesmalen]

This seems like a nice improvement to me, thanks!

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

555

The predicate is indeed irrelevant here, if all lanes of the predicate are inactive, then it would extract the first/last element.

[mnadeem]

Thanks for accepting, please go ahead and merge this as I dont have commit access.

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

 }
 
 static Optional<Instruction *> instCombineSVELast(InstCombiner &IC,
                                                   IntrinsicInst &II) {
   Value *Pg = II.getArgOperand(0);
   Value *Vec = II.getArgOperand(1);
   bool IsAfter = II.getIntrinsicID() == Intrinsic::aarch64_sve_lasta;
 
+  // lastX(splat(X)) --> X
+  if (auto *SplatVal = getSplatValue(Vec))
+    return IC.replaceInstUsesWith(II, SplatVal);
+
   auto *C = dyn_cast<Constant>(Pg);
   if (IsAfter && C && C->isNullValue()) {
     // The intrinsic is extracting lane 0 so use an extract instead.
     auto *IdxTy = Type::getInt64Ty(II.getContext());
     auto *Extract = ExtractElementInst::Create(Vec, ConstantInt::get(IdxTy, 0));
     Extract->insertBefore(&II);
     Extract->takeName(&II);
     return IC.replaceInstUsesWith(II, Extract);

llvm/test/Transforms/InstCombine/AArch64/sve-intrinsic-opts-lasta-lastb.ll

 ; OPT-NEXT:    [[E:%.*]] = tail call i8 @llvm.aarch64.sve.lastb.nxv16i8(<vscale x 16 x i1> [[PG]], <vscale x 16 x i8> [[V:%.*]])
 ; OPT-NEXT:    ret i8 [[E]]
 ;
   %pg = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 15)
   %e = tail call i8 @llvm.aarch64.sve.lastb.nxv16i8(<vscale x 16 x i1> %pg, <vscale x 16 x i8> %v)
   ret i8 %e
 }
 
+; Return the splatted value irrespective of the predicate.
+define i8 @lasta_splat(<vscale x 16 x i1> %pg, i8 %a) #0 {
+; OPT-LABEL: @lasta_splat(
+; OPT-NEXT:    ret i8 %a
+  %splat_insert = insertelement <vscale x 16 x i8> poison, i8 %a, i32 0
+  %splat = shufflevector <vscale x 16 x i8> %splat_insert, <vscale x 16 x i8> poison, <vscale x 16 x i32> zeroinitializer
+  %last = tail call i8 @llvm.aarch64.sve.lasta.nxv16i8(<vscale x 16 x i1> %pg, <vscale x 16 x i8> %splat)
+  ret i8 %last
+}
+
+define i8 @lastb_splat(<vscale x 16 x i1> %pg, i8 %a) #0 {
+; OPT-LABEL: @lastb_splat(
+; OPT-NEXT:    ret i8 %a
+  %splat_insert = insertelement <vscale x 16 x i8> poison, i8 %a, i32 0
+  %splat = shufflevector <vscale x 16 x i8> %splat_insert, <vscale x 16 x i8> poison, <vscale x 16 x i32> zeroinitializer
+  %last = tail call i8 @llvm.aarch64.sve.lastb.nxv16i8(<vscale x 16 x i1> %pg, <vscale x 16 x i8> %splat)
+  ret i8 %last
+}
+
 declare <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32)
 declare i8 @llvm.aarch64.sve.lasta.nxv16i8(<vscale x 16 x i1>, <vscale x 16 x i8>)
 declare i8 @llvm.aarch64.sve.lastb.nxv16i8(<vscale x 16 x i1>, <vscale x 16 x i8>)
 
 attributes #0 = { "target-features"="+sve" }