[AArch64][SVE] Extend svdup->insertelement instcombine pattern to support ...

[peterwaller-arm]

Summary

ptrue patterns other other than VL1.

The current algorithm identifies svdup intrinsic calls where the
predicate is a ptrue with the pattern VL1. It is possible to also
perform this optimisation for ptrue patterns in the range VL1 to VL8
inclusive.

Suppose we have a svdup intrinsic call:

svdup (vec, (ptrue pat), elm)

If pat is VL1, we perform the optimisation as before. Else, if pat is in
the range VL2 to VL8 inclusive, we check that the only uses of this
svdup are other svdups that overwrite every vector element except for
one -- if so, we can replace the svdup call with an insertelement
instruction. In any other case, we bail out.

[joechrisellis]

Bugfix: check that UserPg's intrinsic ID is a ptrue, otherwise we crash in some circumstances.

[efriedma]

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

392

Instead of checking for this specific pattern, can we leverage simplifyDemandedVectorEltsIntrinsic? I guess that doesn't currently work with scalable vectors, though...

[joechrisellis]

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

392

I'm not sure -- that function doesn't seem to have an override for AArch64 at the moment. Perhaps something for a different patch. 🙂

[peterwaller-arm]

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

172

Where are the multiple users?

[joechrisellis]

Address review comments.

• @peterwaller-arm:
  • better name for existing test.
  • add a new test to defend against behaviour when there are multiple users of a DUP.

[peterwaller-arm]

Thanks for your work on this @joechrisellis! Much appreciated. Commandeering to abandon, superseded by D106265.

[peterwaller-arm]

Superseded by D106265.

llvm/lib/Target/AArch64/AArch64TargetTransformInfo.cpp

static Optional<Instruction *> instCombineConvertFromSVBool(InstCombiner &IC,
   if (!EarliestReplacement)
     return None;
 
   return IC.replaceInstUsesWith(II, EarliestReplacement);
 }
 
 static Optional<Instruction *> instCombineSVEDup(InstCombiner &IC,
                                                  IntrinsicInst &II) {
+  assert(II.getIntrinsicID() == Intrinsic::aarch64_sve_dup &&
+         "Expected SVE DUP!");
+
   IntrinsicInst *Pg = dyn_cast<IntrinsicInst>(II.getArgOperand(1));
   if (!Pg)
     return None;
 
   if (Pg->getIntrinsicID() != Intrinsic::aarch64_sve_ptrue)
     return None;
 
   const auto PTruePattern =
       cast<ConstantInt>(Pg->getOperand(0))->getZExtValue();
-  if (PTruePattern != AArch64SVEPredPattern::vl1)
+
+  // Do not consider ptrues whose pattern is not in the range VL1 to VL8
+  // inclusive.
+  if (!(PTruePattern >= AArch64SVEPredPattern::vl1 &&
+        PTruePattern <= AArch64SVEPredPattern::vl8))
+    return None;
+
+  // If the pattern is not VL1, it is necessary to check that the only uses of
+  // this DUP are other DUPs that overwrite every vector element except for
+  // one. If this condition is not true, bail out.
+  if (PTruePattern != AArch64SVEPredPattern::vl1) {
+    for (User *U : II.users()) {
+      IntrinsicInst *UserIntrInst = dyn_cast<IntrinsicInst>(U);
+      if (!UserIntrInst ||
+          UserIntrInst->getIntrinsicID() != Intrinsic::aarch64_sve_dup)
+        return None;
+
+      IntrinsicInst *UserPg =
+          dyn_cast<IntrinsicInst>(UserIntrInst->getArgOperand(1));
+      if (UserPg->getIntrinsicID() != Intrinsic::aarch64_sve_ptrue)
+        return None;
+
+      const auto UserPTruePattern =
+          cast<ConstantInt>(UserPg->getOperand(0))->getZExtValue();
+      if (UserPTruePattern != PTruePattern - 1)
         return None;
+    }
+  }
 
-  // The intrinsic is inserting into lane zero so use an insert instead.
+  // Replace with insertelement.
   auto *IdxTy = Type::getInt64Ty(II.getContext());
-  auto *Insert = InsertElementInst::Create(
-      II.getArgOperand(0), II.getArgOperand(2), ConstantInt::get(IdxTy, 0));
+  auto *Insert =
+      InsertElementInst::Create(II.getArgOperand(0), II.getArgOperand(2),
+                                ConstantInt::get(IdxTy, PTruePattern - 1));
   Insert->insertBefore(&II);
   Insert->takeName(&II);
 
   return IC.replaceInstUsesWith(II, Insert);
 }
 
 static Optional<Instruction *> instCombineSVELast(InstCombiner &IC,
                                                   IntrinsicInst &II) {

llvm/test/Transforms/InstCombine/AArch64/sve-intrinsic-opts-dup.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 16 x i8> @dup_insertelement_0(<vscale x 16 x i8> %v, i8 %s) #0 {
 ; CHECK-LABEL: @dup_insertelement_0(
-; CHECK: %insert = insertelement <vscale x 16 x i8> %v, i8 %s, i64 0
-; CHECK-NEXT: ret <vscale x 16 x i8> %insert
+; CHECK-NEXT:    [[INSERT:%.*]] = insertelement <vscale x 16 x i8> [[V:%.*]], i8 [[S:%.*]], i64 0
+; CHECK-NEXT:    ret <vscale x 16 x i8> [[INSERT]]
+;
   %pg = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 1)
   %insert = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %v, <vscale x 16 x i1> %pg, i8 %s)
   ret <vscale x 16 x i8> %insert
 }
 
 define <vscale x 16 x i8> @dup_insertelement_1(<vscale x 16 x i8> %v, i8 %s) #0 {
 ; CHECK-LABEL: @dup_insertelement_1(
-; CHECK: %pg = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 2)
-; CHECK-NEXT: %insert = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %v, <vscale x 16 x i1> %pg, i8 %s)
-; CHECK-NEXT: ret <vscale x 16 x i8> %insert
+; CHECK-NEXT:    [[PG:%.*]] = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 2)
+; CHECK-NEXT:    [[INSERT:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> [[V:%.*]], <vscale x 16 x i1> [[PG]], i8 [[S:%.*]])
+; CHECK-NEXT:    ret <vscale x 16 x i8> [[INSERT]]
+;
   %pg = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 2)
   %insert = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %v, <vscale x 16 x i1> %pg, i8 %s)
   ret <vscale x 16 x i8> %insert
 }
 
 define <vscale x 16 x i8> @dup_insertelement_x(<vscale x 16 x i8> %v, i8 %s, <vscale x 16 x i1> %pg) #0 {
 ; CHECK-LABEL: @dup_insertelement_x(
-; CHECK: %insert = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %v, <vscale x 16 x i1> %pg, i8 %s)
-; CHECK-NEXT: ret <vscale x 16 x i8> %insert
+; CHECK-NEXT:    [[INSERT:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> [[V:%.*]], <vscale x 16 x i1> [[PG:%.*]], i8 [[S:%.*]])
+; CHECK-NEXT:    ret <vscale x 16 x i8> [[INSERT]]
+;
   %insert = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %v, <vscale x 16 x i1> %pg, i8 %s)
   ret <vscale x 16 x i8> %insert
 }
 
 define <vscale x 8 x i16> @dup_insertelement_0_convert(<vscale x 8 x i16> %v, i16 %s) #0 {
 ; CHECK-LABEL: @dup_insertelement_0_convert(
-; CHECK: %insert = insertelement <vscale x 8 x i16> %v, i16 %s, i64 0
-; CHECK-NEXT: ret <vscale x 8 x i16> %insert
+; CHECK-NEXT:    [[INSERT:%.*]] = insertelement <vscale x 8 x i16> [[V:%.*]], i16 [[S:%.*]], i64 0
+; CHECK-NEXT:    ret <vscale x 8 x i16> [[INSERT]]
+;
   %pg = tail call <vscale x 8 x i1> @llvm.aarch64.sve.ptrue.nxv8i1(i32 1)
   %1 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.convert.to.svbool.nxv8i1(<vscale x 8 x i1> %pg)
   %2 = tail call <vscale x 8 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv8i1(<vscale x 16 x i1> %1)
   %insert = tail call <vscale x 8 x i16> @llvm.aarch64.sve.dup.nxv8i16(<vscale x 8 x i16> %v, <vscale x 8 x i1> %2, i16 %s)
   ret <vscale x 8 x i16> %insert
 }
 
+define <vscale x 16 x i8> @dup_insertelement_two(<vscale x 16 x i8> %v, i8 %a, i8 %b) #0 {
+; CHECK-LABEL: @dup_insertelement_two(
+; CHECK-NEXT:    [[DUP1:%.*]] = insertelement <vscale x 16 x i8> [[V:%.*]], i8 [[A:%.*]], i64 1
+; CHECK-NEXT:    [[DUP2:%.*]] = insertelement <vscale x 16 x i8> [[DUP1]], i8 [[B:%.*]], i64 0
+; CHECK-NEXT:    ret <vscale x 16 x i8> [[DUP2]]
+;
+  %pg1 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 2)
+  %dup1 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %v, <vscale x 16 x i1> %pg1, i8 %a)
+  %pg2 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 1)
+  %dup2 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup1, <vscale x 16 x i1> %pg2, i8 %b)
+  ret <vscale x 16 x i8> %dup2
+}
+
+define <vscale x 16 x i8> @dup_insertelement_four(<vscale x 16 x i8> %v, i8 %a, i8 %b) #0 {
+; CHECK-LABEL: @dup_insertelement_four(
+; CHECK-NEXT:    [[DUP1:%.*]] = insertelement <vscale x 16 x i8> [[V:%.*]], i8 [[A:%.*]], i64 3
+; CHECK-NEXT:    [[DUP2:%.*]] = insertelement <vscale x 16 x i8> [[DUP1]], i8 [[B:%.*]], i64 2
+; CHECK-NEXT:    [[DUP3:%.*]] = insertelement <vscale x 16 x i8> [[DUP2]], i8 [[A]], i64 1
+; CHECK-NEXT:    [[DUP4:%.*]] = insertelement <vscale x 16 x i8> [[DUP3]], i8 [[B]], i64 0
+; CHECK-NEXT:    ret <vscale x 16 x i8> [[DUP4]]
+;
+  %pg1 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 4)
+  %dup1 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %v, <vscale x 16 x i1> %pg1, i8 %a)
+  %pg2 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 3)
+  %dup2 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup1, <vscale x 16 x i1> %pg2, i8 %b)
+  %pg3 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 2)
+  %dup3 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup2, <vscale x 16 x i1> %pg3, i8 %a)
+  %pg4 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 1)
+  %dup4 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup3, <vscale x 16 x i1> %pg4, i8 %b)
+  ret <vscale x 16 x i8> %dup4
+}
+
+define <vscale x 16 x i8> @dup_insertelement_eight(<vscale x 16 x i8> %v, i8 %a, i8 %b) #0 {
+; CHECK-LABEL: @dup_insertelement_eight(
+; CHECK-NEXT:    [[DUP1:%.*]] = insertelement <vscale x 16 x i8> [[V:%.*]], i8 [[A:%.*]], i64 7
+; CHECK-NEXT:    [[DUP2:%.*]] = insertelement <vscale x 16 x i8> [[DUP1]], i8 [[B:%.*]], i64 6
+; CHECK-NEXT:    [[DUP3:%.*]] = insertelement <vscale x 16 x i8> [[DUP2]], i8 [[A]], i64 5
+; CHECK-NEXT:    [[DUP4:%.*]] = insertelement <vscale x 16 x i8> [[DUP3]], i8 [[B]], i64 4
+; CHECK-NEXT:    [[DUP5:%.*]] = insertelement <vscale x 16 x i8> [[DUP4]], i8 [[A]], i64 3
+; CHECK-NEXT:    [[DUP6:%.*]] = insertelement <vscale x 16 x i8> [[DUP5]], i8 [[B]], i64 2
+; CHECK-NEXT:    [[DUP7:%.*]] = insertelement <vscale x 16 x i8> [[DUP6]], i8 [[A]], i64 1
+; CHECK-NEXT:    [[DUP8:%.*]] = insertelement <vscale x 16 x i8> [[DUP7]], i8 [[B]], i64 0
+; CHECK-NEXT:    ret <vscale x 16 x i8> [[DUP8]]
+;
+  %pg1 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 8)
+  %dup1 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %v, <vscale x 16 x i1> %pg1, i8 %a)
+  %pg2 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 7)
+  %dup2 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup1, <vscale x 16 x i1> %pg2, i8 %b)
+  %pg3 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 6)
+  %dup3 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup2, <vscale x 16 x i1> %pg3, i8 %a)
+  %pg4 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 5)
+  %dup4 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup3, <vscale x 16 x i1> %pg4, i8 %b)
+  %pg5 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 4)
+  %dup5 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup4, <vscale x 16 x i1> %pg5, i8 %a)
+  %pg6 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 3)
+  %dup6 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup5, <vscale x 16 x i1> %pg6, i8 %b)
+  %pg7 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 2)
+  %dup7 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup6, <vscale x 16 x i1> %pg7, i8 %a)
+  %pg8 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 1)
+  %dup8 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup7, <vscale x 16 x i1> %pg8, i8 %b)
+  ret <vscale x 16 x i8> %dup8
+}
+
+define <vscale x 16 x i8> @dup_insertelement_partial_chain(<vscale x 16 x i8> %v, i8 %a, i8 %b) #0 {
+; CHECK-LABEL: @dup_insertelement_partial_chain(
+; CHECK-NEXT:    [[DUP1:%.*]] = insertelement <vscale x 16 x i8> [[V:%.*]], i8 [[A:%.*]], i64 3
+; CHECK-NEXT:    [[PG2:%.*]] = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 3)
+; CHECK-NEXT:    [[DUP2:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> [[DUP1]], <vscale x 16 x i1> [[PG2]], i8 [[B:%.*]])
+; CHECK-NEXT:    [[DUP4:%.*]] = insertelement <vscale x 16 x i8> [[DUP2]], i8 [[B]], i64 0
+; CHECK-NEXT:    ret <vscale x 16 x i8> [[DUP4]]
+;
+  %pg1 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 4)
+  %dup1 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %v, <vscale x 16 x i1> %pg1, i8 %a)
+  %pg2 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 3)
+  %dup2 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup1, <vscale x 16 x i1> %pg2, i8 %b)
+  %pg4 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 1)
+  %dup4 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup2, <vscale x 16 x i1> %pg4, i8 %b)
+  ret <vscale x 16 x i8> %dup4
+}
+
+define <vscale x 16 x i8> @dup_insertelement_non_vl_ptrue_pat_0(<vscale x 16 x i8> %v, i8 %a, i8 %b) #0 {
+; CHECK-LABEL: @dup_insertelement_non_vl_ptrue_pat_0(
+; CHECK-NEXT:    [[DUP1:%.*]] = insertelement <vscale x 16 x i8> [[V:%.*]], i8 [[A:%.*]], i64 0
+; CHECK-NEXT:    [[PG2:%.*]] = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 0)
+; CHECK-NEXT:    [[DUP2:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> [[DUP1]], <vscale x 16 x i1> [[PG2]], i8 [[B:%.*]])
+; CHECK-NEXT:    ret <vscale x 16 x i8> [[DUP2]]
+;
+  %pg1 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 1)
+  %dup1 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %v, <vscale x 16 x i1> %pg1, i8 %a)
+  %pg2 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 0) ; POW2 -- not in range.
+  %dup2 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup1, <vscale x 16 x i1> %pg2, i8 %b)
+  ret <vscale x 16 x i8> %dup2
+}
+
+define <vscale x 16 x i8> @dup_insertelement_non_vl_ptrue_pat_9(<vscale x 16 x i8> %v, i8 %a, i8 %b) #0 {
+; CHECK-LABEL: @dup_insertelement_non_vl_ptrue_pat_9(
+; CHECK-NEXT:    [[PG1:%.*]] = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 9)
+; CHECK-NEXT:    [[DUP1:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> [[V:%.*]], <vscale x 16 x i1> [[PG1]], i8 [[A:%.*]])
+; CHECK-NEXT:    [[PG2:%.*]] = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 8)
+; CHECK-NEXT:    [[DUP2:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> [[DUP1]], <vscale x 16 x i1> [[PG2]], i8 [[B:%.*]])
+; CHECK-NEXT:    ret <vscale x 16 x i8> [[DUP2]]
+;
+  %pg1 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 9) ; VL16 -- not in range.
+  %dup1 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %v, <vscale x 16 x i1> %pg1, i8 %a)
+  %pg2 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 8)
+  %dup2 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup1, <vscale x 16 x i1> %pg2, i8 %b)
+  ret <vscale x 16 x i8> %dup2
+}
+
+define <vscale x 16 x i8> @dup_insertelement_non_contiguous_ptrues(<vscale x 16 x i8> %v, i8 %a, i8 %b) #0 {
+; CHECK-LABEL: @dup_insertelement_non_contiguous_ptrues(
+; CHECK-NEXT:    [[PG1:%.*]] = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 4)
+; CHECK-NEXT:    [[DUP1:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> [[V:%.*]], <vscale x 16 x i1> [[PG1]], i8 [[A:%.*]])
+; CHECK-NEXT:    [[PG2:%.*]] = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 2)
+; CHECK-NEXT:    [[DUP2:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> [[DUP1]], <vscale x 16 x i1> [[PG2]], i8 [[B:%.*]])
+; CHECK-NEXT:    ret <vscale x 16 x i8> [[DUP2]]
+;
+  %pg1 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 4)
+  %dup1 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %v, <vscale x 16 x i1> %pg1, i8 %a)
+  %pg2 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 2)
+  %dup2 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup1, <vscale x 16 x i1> %pg2, i8 %b)
+  ret <vscale x 16 x i8> %dup2
+}
+
+define <vscale x 16 x i8> @dup_insertelement_ret_dup(<vscale x 16 x i8> %v, i8 %a, i8 %b) #0 {
+; CHECK-LABEL: @dup_insertelement_ret_dup(
+; CHECK-NEXT:    [[DUP1:%.*]] = insertelement <vscale x 16 x i8> [[V:%.*]], i8 [[A:%.*]], i64 3
+; CHECK-NEXT:    [[DUP2:%.*]] = insertelement <vscale x 16 x i8> [[DUP1]], i8 [[B:%.*]], i64 2
+; CHECK-NEXT:    [[PG3:%.*]] = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 2)
+; CHECK-NEXT:    [[DUP3:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> [[DUP2]], <vscale x 16 x i1> [[PG3]], i8 [[A]])
+; CHECK-NEXT:    ret <vscale x 16 x i8> [[DUP3]]
+;
+  %pg1 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 4)
+  %dup1 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %v, <vscale x 16 x i1> %pg1, i8 %a)
+  %pg2 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 3)
+  %dup2 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup1, <vscale x 16 x i1> %pg2, i8 %b)
+  %pg3 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 2)
+  %dup3 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup2, <vscale x 16 x i1> %pg3, i8 %a)
+  ret <vscale x 16 x i8> %dup3
+}
+
+define i8 @dup_insertelement_multiple_users(<vscale x 16 x i8> %v, i8 %a, i8 %b) #0 {
+; CHECK-LABEL: @dup_insertelement_multiple_users(
+; CHECK-NEXT:    [[DUP1:%.*]] = insertelement <vscale x 16 x i8> [[V:%.*]], i8 [[A:%.*]], i64 3
+; CHECK-NEXT:    [[PG2:%.*]] = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 3)
+; CHECK-NEXT:    [[DUP2:%.*]] = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> [[DUP1]], <vscale x 16 x i1> [[PG2]], i8 [[B:%.*]])
+; CHECK-NEXT:    [[X:%.*]] = extractelement <vscale x 16 x i8> [[DUP2]], i32 0
+; CHECK-NEXT:    [[Y:%.*]] = extractelement <vscale x 16 x i8> [[DUP2]], i32 1
+; CHECK-NEXT:    [[RES:%.*]] = add i8 [[X]], [[Y]]
+; CHECK-NEXT:    ret i8 [[RES]]
+;
+  %pg1 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 4)
+  %dup1 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %v, <vscale x 16 x i1> %pg1, i8 %a)
+  %pg2 = tail call <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32 3)
+  %dup2 = tail call <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8> %dup1, <vscale x 16 x i1> %pg2, i8 %b)
+  %x = extractelement <vscale x 16 x i8> %dup2, i32 0
+  %y = extractelement <vscale x 16 x i8> %dup2, i32 1
+  %res = add i8 %x, %y
+  ret i8 %res
+}
+
 declare <vscale x 16 x i8> @llvm.aarch64.sve.dup.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i1>, i8)
 declare <vscale x 8 x i16> @llvm.aarch64.sve.dup.nxv8i16(<vscale x 8 x i16>, <vscale x 8 x i1>, i16)
 
 declare <vscale x 16 x i1> @llvm.aarch64.sve.ptrue.nxv16i1(i32)
 declare <vscale x 8 x i1> @llvm.aarch64.sve.ptrue.nxv8i1(i32)
 
 declare <vscale x 16 x i1> @llvm.aarch64.sve.convert.to.svbool.nxv8i1(<vscale x 8 x i1>)
 declare <vscale x 8 x i1> @llvm.aarch64.sve.convert.from.svbool.nxv8i1(<vscale x 16 x i1>)
 
 attributes #0 = { "target-features"="+sve" }