[AArch64][SVE] Implement vector tuple intrinsics

[c-rhodes]

Summary

This patch adds the following intrinsics for creating two-tuple,
three-tuple and four-tuple scalable vectors:

• llvm.aarch64.sve.tuple.create2
• llvm.aarch64.sve.tuple.create3
• llvm.aarch64.sve.tuple.create4

As well as:

• llvm.aarch64.sve.tuple.get
• llvm.aarch64.sve.tuple.set

For extracting and inserting scalable vectors from vector tuples. These
intrinsics are intended to be used by the ACLE functions svcreate<n>,
svget and svset.

This patch also includes calling convention support for passing and
returning tuples of scalable vectors to/from functions.

[c-rhodes]

Fix warnings reported by Harbor.

[c-rhodes]

Added full context.

[efriedma]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

7038

This assertion seems ambitious; I'm not sure how you plan to ensure that the sve.tuple.get intrinsic is in the same basic block as the intrinsic that produces the value.

Is there some reason you're putting this handling into the target-independent SelectionDAGBuilder, as opposed to target-specific code?

[fpetrogalli]

I think that in this patch we are missing tests that check the output code in situations in which the tuple.get operates on tuple that are passed to the definition of the function, as in the following example:

define <vscale x 16 x i8> @foo(<vscale x 32 x i8> %res) {
; CHECK-LABEL: ld2b_i8_1:
; CHECK: mov z0.d, z1.d
; CHECK-NEXT: ret
  %v2 = call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv32i8(<vscale x 32 x i8> %res, i32 1)
  ret <vscale x 16 x i8> %v2
}

@sdesmalen - should we add them? Am I missing something?

Reference comment: https://reviews.llvm.org/D75751#inline-697705

Francesco

[c-rhodes]

Changes:

• Rebased on top of D80139 implementing copyToReg/copyFromReg support for illegal scalable vectors. This enables copying vector tuple types between BBs.
• Removed assert in the lowering of sve_tuple_get/sve_tuple_set intrinsics checking the tuple is CONCAT_VECTOR.
• Vector tuple elements are now extracted with EXTRACT_SUBVECTOR.
• Removed target specific calling conv methods from AArch64ISelLowering.
• Updated test/CodeGen/AArch64/sve-intrinsics-create-tuple.ll to test extract in different BB.

[c-rhodes]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

7038

This assertion seems ambitious; I'm not sure how you plan to ensure that the sve.tuple.get intrinsic is in the same basic block as the intrinsic that produces the value.

Sander's patch D80139 fixes copying between BBs and I've updated this patch to use EXTRACT_SUBVECTOR to get the vector from the tuple type, rather than looking through the CONCAT_VECTOR, so this assert has been removed.

Is there some reason you're putting this handling into the target-independent SelectionDAGBuilder, as opposed to target-specific code?

Tuple types for structured loads/stores were originally implemented with sizeless structs downstream but adding sizeless structs to the C standard isn't going to happen, so we chose to represent them as multiples of 128-bit vectors, e.g. svint32x2_t -> <n x 8 x i32>. For vector tuples containing 2 and 4 vectors (LD2/LD4, ST2/ST4) this was fine as the types are powers of 2 and can be broken down by the existing mechanisms, but for LD3/ST3 the types are odd sizes:

• svint8x3_t -> <vscale x 48 x i8>
• svint16x3_t -> <vscale x 24 x i16>
• svint32x3_t -> <vscale x 12 x i32>
• svint64x2_t -> <vscale x 6 x i64>

Which are problematic for legalisation when it comes to widening/splitting. To do this lowering in target-specific AArch64ISelLowering MVTs would have to be defined for these types which aren't legal. By lowering here we avoid these issues as these types don't reach type legalization.

[efriedma]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

7038

If you need to run before legalization, you can put code in AArch64TargetLowering::PerformDAGCombine.

Alternatively, additional MVTs wouldn't be that terrible.

[c-rhodes]

Changes:

• Moved lowering of vector tuple intrinsics to AArch64TargetLowering::PerformDAGCombine.
• Renamed function args in test/CodeGen/AArch64/sve-intrinsics-create-tuple.ll from x<n> to z<n>.
• Fixed bug in lowering of sve.tuple.set where the index (of element to update) was being used to extract subvectors from the tuple vector causing the wrong vectors to be copied. Added tests to catch this.

[c-rhodes]

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

7038

If you need to run before legalization, you can put code in AArch64TargetLowering::PerformDAGCombine.

That's a much better idea :) I've implemented your suggestion, cheers!

Alternatively, additional MVTs wouldn't be that terrible.

I'm a little confused about this, there's a comment at the top of MachineValueType.h:

// This file defines the set of machine-level target independent types which
// legal values in the code generator use.

which to me implies MVTs should only be implemented for legal types which tuple vectors aren't, although I've also seen v3i32 which seems to be used in the AMDGPU backend but I don't know if the ISA specifies a register for that. I don't think I want to implements MVTs in this patch but I'm curious if it's ok to define them for illegal types?

[efriedma]

LGTM

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

7038

We treat the "legal values in the code generator" bit a little loosely; really, it's any type that useful to define as an MVT for some backend.

And yes, if I recall that discussion correctly, I think AMDGPU actually does have operations on v3i32 registers.

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

13529

No point to explicitly checking isBeforeLegalizeOps(); the intrinsics should be gone after that.

[c-rhodes]

Changes:

• Test extract of tuple vector passed into function.
• Remove explicit check to exit early if after legalization when lowering intrinsics.

[c-rhodes]

In D75674#1931136, @fpetrogalli wrote:

I think that in this patch we are missing tests that check the output code in situations in which the tuple.get operates on tuple that are passed to the definition of the function, as in the following example:

I've added a few tests to test/CodeGen/AArch64/sve-intrinsics-insert-extract-tuple.ll to test this.

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

13529

Doh, done!

[c-rhodes]

In D75674#2051313, @efriedma wrote:

LGTM

Thanks for reviewing!

I'll hold off on merging this until Sander lands D80139

[craig.topper]

llvm/include/llvm/IR/IntrinsicsAArch64.td

815

Sorry for digging up an old patch here. Why is this IntrArgMemOnly? And why are any of these IntrReadMem? We're working on similar intrinsics for RISCV vector support and were wondering if we were missing something.

[c-rhodes]

llvm/include/llvm/IR/IntrinsicsAArch64.td

815

Sorry for digging up an old patch here. Why is this IntrArgMemOnly? And why are any of these IntrReadMem? We're working on similar intrinsics for RISCV vector support and were wondering if we were missing something.

Apologies for the delayed response.

When these intrinsics were first implemented CopyToReg/CopyFromReg didn't support some of the more unusual tuple types like nxv6i64 which could be created by an LD3 for example. This prevented copying between basic blocks since the compiler would crash. To get around this a temporary constraint of these intrinsics was they must immediately be followed by an extract to prevent the backend from ever seeing such types. I think at the time this was ok since these intrinsics were only used the the Arm C Language Extension (ACLE) intrinsics so we could apply this constraint, although it required marking the intrinsics as touching memory to prevent the mid end from mucking around with them. The CopyToReg/CopyFromReg issue has since been fixed however, so I can see no reason why these intrinsics can't be updated to mark them as not touching memory, I'll make a note to address that, cheers.

llvm/include/llvm/IR/IntrinsicsAArch64.td

 
 let TargetPrefix = "aarch64" in {  // All intrinsics start with "llvm.aarch64.".
 
+  class AdvSIMD_SVE_Create_2Vector_Tuple
+    : Intrinsic<[llvm_anyvector_ty],
+                [llvm_anyvector_ty, LLVMMatchType<1>],
+                [IntrReadMem]>;
+
+  class AdvSIMD_SVE_Create_3Vector_Tuple
+    : Intrinsic<[llvm_anyvector_ty],
+                [llvm_anyvector_ty, LLVMMatchType<1>, LLVMMatchType<1>],
+                [IntrReadMem]>;
+
+  class AdvSIMD_SVE_Create_4Vector_Tuple
+    : Intrinsic<[llvm_anyvector_ty],
+                [llvm_anyvector_ty, LLVMMatchType<1>, LLVMMatchType<1>,
+                 LLVMMatchType<1>],
+                [IntrReadMem]>;
+
+  class AdvSIMD_SVE_Set_Vector_Tuple
+    : Intrinsic<[llvm_anyvector_ty],
+                [LLVMMatchType<0>, llvm_i32_ty, llvm_anyvector_ty],
+                [IntrReadMem, ImmArg<1>]>;
+
+  class AdvSIMD_SVE_Get_Vector_Tuple
+    : Intrinsic<[llvm_anyvector_ty], [llvm_anyvector_ty, llvm_i32_ty],
+                [IntrReadMem, IntrArgMemOnly, ImmArg<1>]>;
+
   class AdvSIMD_1Vec_PredLoad_Intrinsic
     : Intrinsic<[llvm_anyvector_ty],
                 [LLVMScalarOrSameVectorWidth<0, llvm_i1_ty>,

                [IntrWriteMem, IntrArgMemOnly]>;
 
 //
+// Vector tuple creation intrinsics (ACLE)
+//
+
+def int_aarch64_sve_tuple_create2 : AdvSIMD_SVE_Create_2Vector_Tuple;
+def int_aarch64_sve_tuple_create3 : AdvSIMD_SVE_Create_3Vector_Tuple;
+def int_aarch64_sve_tuple_create4 : AdvSIMD_SVE_Create_4Vector_Tuple;
+
+//
+// Vector tuple insertion/extraction intrinsics (ACLE)
+//
+
+def int_aarch64_sve_tuple_get : AdvSIMD_SVE_Get_Vector_Tuple;
+def int_aarch64_sve_tuple_set : AdvSIMD_SVE_Set_Vector_Tuple;
+
+//
 // Loads
 //
 

llvm/lib/CodeGen/SelectionDAG/SelectionDAGBuilder.cpp

   // Convert the vector to the appropriate type if necessary.
   unsigned DestVectorNoElts = NumIntermediates * IntermediateNumElts;
 
-  EVT BuiltVectorTy = EVT::getVectorVT(
-      *DAG.getContext(), IntermediateVT.getScalarType(), DestVectorNoElts);
+  EVT BuiltVectorTy =
+      EVT::getVectorVT(*DAG.getContext(), IntermediateVT.getScalarType(),
+                       DestVectorNoElts, ValueVT.isScalableVector());
   if (ValueVT != BuiltVectorTy) {
     if (SDValue Widened = widenVectorToPartType(DAG, Val, DL, BuiltVectorTy))
       Val = Widened;

                              DAG.getZExtOrTrunc(Const, getCurSDLoc(), DestVT)));
     return;
   }
+  case Intrinsic::aarch64_sve_tuple_get: {
+    SDValue Src1 = getValue(I.getOperand(0));
+    SDValue Idx = getValue(I.getOperand(1));
+
+    assert(Src1.getOpcode() == ISD::CONCAT_VECTORS &&
+           "Unexpected operand for sve_tuple_get");
+
+    uint64_t IdxConst = cast<ConstantSDNode>(Idx)->getZExtValue();
+    if (IdxConst > Src1->getNumOperands() - 1)
+      report_fatal_error("index larger than expected");
+    setValue(&I, SDValue(Src1.getOperand(IdxConst)));
+    return;
+  }
+  case Intrinsic::aarch64_sve_tuple_set: {
+    SDValue Tuple = getValue(I.getOperand(0));
+    SDValue Idx = getValue(I.getOperand(1));
+    SDValue Vec = getValue(I.getOperand(2));
+
+    assert(Tuple.getOpcode() == ISD::CONCAT_VECTORS &&
+           "Unexpected operand for sve_tuple_set");
+
+    uint64_t IdxConst = cast<ConstantSDNode>(Idx)->getZExtValue();
+
+    SmallVector<SDValue, 4> Opnds;
+    for (unsigned J = 0; J < Tuple->getNumOperands(); ++J)
+      Opnds.push_back(J == IdxConst ? Vec : Tuple->getOperand(J));
+    SDValue Result =
+        DAG.getNode(ISD::CONCAT_VECTORS, sdl, Tuple.getValueType(), Opnds);
+    setValue(&I, Result);
+    return;
+  }
+  case Intrinsic::aarch64_sve_tuple_create2:
+  case Intrinsic::aarch64_sve_tuple_create3:
+  case Intrinsic::aarch64_sve_tuple_create4: {
+    unsigned N = I.getNumArgOperands();
+    SmallVector<SDValue, 4> Opnds;
+    for (const auto &Arg : I.args())
+      Opnds.push_back(getValue(Arg.get()));
+
+    EVT VT = Opnds[0].getValueType();
+    EVT EltVT = VT.getVectorElementType();
+    EVT DestVT =
+        EVT::getVectorVT(*Context, EltVT, VT.getVectorElementCount() * N);
+    SDValue Result = DAG.getNode(ISD::CONCAT_VECTORS, sdl, DestVT, Opnds);
+    setValue(&I, Result);
+    return;
+  }
   }
 }
 

llvm/lib/Target/AArch64/AArch64ISelLowering.h

                           MachineFunction &MF,
                           unsigned Intrinsic) const override;
 
+  unsigned getVectorTypeBreakdownForCallingConv(LLVMContext &Context,
+                                                CallingConv::ID CC, EVT VT,
+                                                EVT &IntermediateVT,
+                                                unsigned &NumIntermediates,
+                                                MVT &RegisterVT) const override;
+
+  MVT getRegisterTypeForCallingConv(LLVMContext &Context, CallingConv::ID CC,
+                                    EVT VT) const override;
+  unsigned getNumRegistersForCallingConv(LLVMContext &Context,
+                                         CallingConv::ID CC,
+                                         EVT VT) const override;
+
   bool shouldReduceLoadWidth(SDNode *Load, ISD::LoadExtType ExtTy,
                              EVT NewVT) const override;
 

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

   return false;
 }
 
+MVT AArch64TargetLowering::getRegisterTypeForCallingConv(LLVMContext &Context,
+                                                         CallingConv::ID CC,
+                                                         EVT VT) const {
+  if (!VT.isScalableVector() ||
+      VT.getSizeInBits().getKnownMinSize() <= AArch64::SVEBitsPerBlock)
+    return TargetLowering::getRegisterTypeForCallingConv(Context, CC, VT);
+
+  switch (VT.getVectorElementType().getSimpleVT().SimpleTy) {
+  case MVT::i8:
+    return MVT::nxv16i8;
+  case MVT::i16:
+    return MVT::nxv8i16;
+  case MVT::i32:
+    return MVT::nxv4i32;
+  case MVT::i64:
+    return MVT::nxv2i64;
+  case MVT::f16:
+    return MVT::nxv8f16;
+  case MVT::f32:
+    return MVT::nxv4f32;
+  case MVT::f64:
+    return MVT::nxv2f64;
+  default:
+    llvm_unreachable("Unsupported type for SVE vectors");
+  }
+}
+
+unsigned AArch64TargetLowering::getVectorTypeBreakdownForCallingConv(
+    LLVMContext &Context, CallingConv::ID CC, EVT VT, EVT &IntermediateVT,
+    unsigned &NumIntermediates, MVT &RegisterVT) const {
+  if (!VT.isScalableVector() ||
+      VT.getSizeInBits().getKnownMinSize() <= AArch64::SVEBitsPerBlock)
+    return TargetLowering::getVectorTypeBreakdownForCallingConv(
+        Context, CC, VT, IntermediateVT, NumIntermediates, RegisterVT);
+
+  IntermediateVT = RegisterVT = getRegisterTypeForCallingConv(Context, CC, VT);
+  NumIntermediates = getNumRegistersForCallingConv(Context, CC, VT);
+  return NumIntermediates;
+}
+
+unsigned AArch64TargetLowering::getNumRegistersForCallingConv(
+    LLVMContext &Context, CallingConv::ID CC, EVT VT) const {
+  if (!VT.isScalableVector() ||
+      VT.getSizeInBits().getKnownMinSize() <= AArch64::SVEBitsPerBlock)
+    return TargetLowering::getNumRegistersForCallingConv(Context, CC, VT);
+
+  unsigned NumVectors =
+      VT.getSizeInBits().getKnownMinSize() / AArch64::SVEBitsPerBlock;
+  assert(NumVectors * AArch64::SVEBitsPerBlock ==
+             VT.getSizeInBits().getKnownMinSize() &&
+         "Not a multiple of a full SVE vector");
+  return NumVectors;
+}
+
 bool AArch64TargetLowering::shouldReduceLoadWidth(SDNode *Load,
                                                   ISD::LoadExtType ExtTy,
                                                   EVT NewVT) const {

llvm/test/CodeGen/AArch64/sve-calling-convention-tuple-types.ll

+; RUN: llc -mtriple aarch64 -mattr=+sve -asm-verbose=0 < %s | FileCheck %s
+
+;
+; svint8x2_t
+;
+
+define <vscale x 32 x i8> @ret_svint8x2_t(<vscale x 16 x i8> %unused_z0, <vscale x 16 x i8> %z1, <vscale x 16 x i8> %z2) #0 {
+; CHECK-LABEL: ret_svint8x2_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 32 x i8> @llvm.aarch64.sve.tuple.create2.nxv32i8.nxv16i8(<vscale x 16 x i8> %z1, <vscale x 16 x i8> %z2)
+  ret <vscale x 32 x i8> %tuple
+}
+
+define void @call_svint8x2_t(<vscale x 16 x i8> %dummy_z0, <vscale x 16 x i8> %z1, <vscale x 16 x i8> %dummy_z2, <vscale x 16 x i8> %z3) #0 {
+; CHECK-LABEL: call_svint8x2_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z3.d
+; CHECK-NEXT: bl callee_svint8x2_t
+  %tuple = tail call <vscale x 32 x i8> @llvm.aarch64.sve.tuple.create2.nxv32i8.nxv16i8(<vscale x 16 x i8> %z1, <vscale x 16 x i8> %z3)
+  call void @callee_svint8x2_t(<vscale x 32 x i8> %tuple)
+  ret void
+}
+
+;
+; svint16x2_t
+;
+
+define <vscale x 16 x i16> @ret_svint16x2_t(<vscale x 8 x i16> %unused_z0, <vscale x 8 x i16> %z1, <vscale x 8 x i16> %z2) #0 {
+; CHECK-LABEL: ret_svint16x2_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 16 x i16> @llvm.aarch64.sve.tuple.create2.nxv16i16.nxv8i16(<vscale x 8 x i16> %z1, <vscale x 8 x i16> %z2)
+  ret <vscale x 16 x i16> %tuple
+}
+
+define void @call_svint16x2_t(<vscale x 8 x i16> %dummy_z0, <vscale x 8 x i16> %z1, <vscale x 8 x i16> %dummy_z2, <vscale x 8 x i16> %z3) #0 {
+; CHECK-LABEL: call_svint16x2_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z3.d
+; CHECK-NEXT: bl callee_svint16x2_t
+  %tuple = tail call <vscale x 16 x i16> @llvm.aarch64.sve.tuple.create2.nxv16i16.nxv8i16(<vscale x 8 x i16> %z1, <vscale x 8 x i16> %z3)
+  call void @callee_svint16x2_t(<vscale x 16 x i16> %tuple)
+  ret void
+}
+
+;
+; svint32x2_t
+;
+
+define <vscale x 8 x i32> @ret_svint32x2_t(<vscale x 4 x i32> %unused_z0, <vscale x 4 x i32> %z1, <vscale x 4 x i32> %z2) #0 {
+; CHECK-LABEL: ret_svint32x2_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 8 x i32> @llvm.aarch64.sve.tuple.create2.nxv8i32.nxv4i32(<vscale x 4 x i32> %z1, <vscale x 4 x i32> %z2)
+  ret <vscale x 8 x i32> %tuple
+}
+
+define void @call_svint32x2_t(<vscale x 4 x i32> %dummy_z0, <vscale x 4 x i32> %z1, <vscale x 4 x i32> %dummy_z2, <vscale x 4 x i32> %z3) #0 {
+; CHECK-LABEL: call_svint32x2_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z3.d
+; CHECK-NEXT: bl callee_svint32x2_t
+  %tuple = tail call <vscale x 8 x i32> @llvm.aarch64.sve.tuple.create2.nxv8i32.nxv4i32(<vscale x 4 x i32> %z1, <vscale x 4 x i32> %z3)
+  call void @callee_svint32x2_t(<vscale x 8 x i32> %tuple)
+  ret void
+}
+
+;
+; svint64x2_t
+;
+
+define <vscale x 4 x i64> @ret_svint64x2_t(<vscale x 2 x i64> %unused_z0, <vscale x 2 x i64> %z1, <vscale x 2 x i64> %z2) #0 {
+; CHECK-LABEL: ret_svint64x2_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 4 x i64> @llvm.aarch64.sve.tuple.create2.nxv4i64.nxv2i64(<vscale x 2 x i64> %z1, <vscale x 2 x i64> %z2)
+  ret <vscale x 4 x i64> %tuple
+}
+
+define void @call_svint64x2_t(<vscale x 2 x i64> %dummy_z0, <vscale x 2 x i64> %z1, <vscale x 2 x i64> %dummy_z2, <vscale x 2 x i64> %z3) #0 {
+; CHECK-LABEL: call_svint64x2_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z3.d
+; CHECK-NEXT: bl callee_svint64x2_t
+  %tuple = tail call <vscale x 4 x i64> @llvm.aarch64.sve.tuple.create2.nxv4i64.nxv2i64(<vscale x 2 x i64> %z1, <vscale x 2 x i64> %z3)
+  call void @callee_svint64x2_t(<vscale x 4 x i64> %tuple)
+  ret void
+}
+
+;
+; svfloatx2_t
+;
+
+define <vscale x 8 x float> @ret_svfloatx2_t(<vscale x 4 x float> %unused_z0, <vscale x 4 x float> %z1, <vscale x 4 x float> %z2) #0 {
+; CHECK-LABEL: ret_svfloatx2_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 8 x float> @llvm.aarch64.sve.tuple.create2.nxv8f32.nxv4f32(<vscale x 4 x float> %z1, <vscale x 4 x float> %z2)
+  ret <vscale x 8 x float> %tuple
+}
+
+define void @call_svfloatx2_t(<vscale x 4 x float> %dummy_z0, <vscale x 4 x float> %z1, <vscale x 4 x float> %dummy_z2, <vscale x 4 x float> %z3) #0 {
+; CHECK-LABEL: call_svfloatx2_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z3.d
+; CHECK-NEXT: bl callee_svfloatx2_t
+  %tuple = tail call <vscale x 8 x float> @llvm.aarch64.sve.tuple.create2.nxv8f32.nxv4f32(<vscale x 4 x float> %z1, <vscale x 4 x float> %z3)
+  call void @callee_svfloatx2_t(<vscale x 8 x float> %tuple)
+  ret void
+}
+
+;
+; svdoublex2_t
+;
+
+define <vscale x 4 x double> @ret_svdoublex2_t(<vscale x 2 x double> %unused_z0, <vscale x 2 x double> %z1, <vscale x 2 x double> %z2) #0 {
+; CHECK-LABEL: ret_svdoublex2_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 4 x double> @llvm.aarch64.sve.tuple.create2.nxv4f64.nxv2f64(<vscale x 2 x double> %z1, <vscale x 2 x double> %z2)
+  ret <vscale x 4 x double> %tuple
+}
+
+define void @call_svdoublex2_t(<vscale x 2 x double> %dummy_z0, <vscale x 2 x double> %z1, <vscale x 2 x double> %dummy_z2, <vscale x 2 x double> %z3) #0 {
+; CHECK-LABEL: call_svdoublex2_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z3.d
+; CHECK-NEXT: bl callee_svdoublex2_t
+  %tuple = tail call <vscale x 4 x double> @llvm.aarch64.sve.tuple.create2.nxv4f64.nxv2f64(<vscale x 2 x double> %z1, <vscale x 2 x double> %z3)
+  call void @callee_svdoublex2_t(<vscale x 4 x double> %tuple)
+  ret void
+}
+
+;
+; svint8x3_t
+;
+
+define <vscale x 48 x i8> @ret_svint8x3_t(<vscale x 16 x i8> %unused_z0, <vscale x 16 x i8> %z1, <vscale x 16 x i8> %z2, <vscale x 16 x i8> %z3) #0 {
+; CHECK-LABEL: ret_svint8x3_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z3.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 48 x i8> @llvm.aarch64.sve.tuple.create3.nxv48i8.nxv16i8(<vscale x 16 x i8> %z1, <vscale x 16 x i8> %z2, <vscale x 16 x i8> %z3)
+  ret <vscale x 48 x i8> %tuple
+}
+
+define void @call_svint8x3_t(<vscale x 16 x i8> %dummy_z0, <vscale x 16 x i8> %z1, <vscale x 16 x i8> %z2, <vscale x 16 x i8> %dummy_z3, <vscale x 16 x i8> %z4) #0 {
+; CHECK-LABEL: call_svint8x3_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z4.d
+; CHECK-NEXT: bl callee_svint8x3_t
+  %tuple = tail call <vscale x 48 x i8> @llvm.aarch64.sve.tuple.create3.nxv48i8.nxv16i8(<vscale x 16 x i8> %z1, <vscale x 16 x i8> %z2, <vscale x 16 x i8> %z4)
+  call void @callee_svint8x3_t(<vscale x 48 x i8> %tuple)
+  ret void
+}
+
+;
+; svint16x3_t
+;
+
+define <vscale x 24 x i16> @ret_svint16x3_t(<vscale x 8 x i16> %unused_z0, <vscale x 8 x i16> %z1, <vscale x 8 x i16> %z2, <vscale x 8 x i16> %z3) #0 {
+; CHECK-LABEL: ret_svint16x3_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z3.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 24 x i16> @llvm.aarch64.sve.tuple.create3.nxv24i16.nxv8i16(<vscale x 8 x i16> %z1, <vscale x 8 x i16> %z2, <vscale x 8 x i16> %z3)
+  ret <vscale x 24 x i16> %tuple
+}
+
+define void @call_svint16x3_t(<vscale x 8 x i16> %dummy_z0, <vscale x 8 x i16> %z1, <vscale x 8 x i16> %z2, <vscale x 8 x i16> %dummy_z3, <vscale x 8 x i16> %z4) #0 {
+; CHECK-LABEL: call_svint16x3_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z4.d
+; CHECK-NEXT: bl callee_svint16x3_t
+  %tuple = tail call <vscale x 24 x i16> @llvm.aarch64.sve.tuple.create3.nxv24i16.nxv8i16(<vscale x 8 x i16> %z1, <vscale x 8 x i16> %z2, <vscale x 8 x i16> %z4)
+  call void @callee_svint16x3_t(<vscale x 24 x i16> %tuple)
+  ret void
+}
+
+;
+; svint32x3_t
+;
+
+define <vscale x 12 x i32> @ret_svint32x3_t(<vscale x 4 x i32> %unused_z0, <vscale x 4 x i32> %z1, <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z3) #0 {
+; CHECK-LABEL: ret_svint32x3_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z3.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 12 x i32> @llvm.aarch64.sve.tuple.create3.nxv12i32.nxv4i32(<vscale x 4 x i32> %z1, <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z3)
+  ret <vscale x 12 x i32> %tuple
+}
+
+define void @call_svint32x3_t(<vscale x 4 x i32> %dummy_z0, <vscale x 4 x i32> %z1, <vscale x 4 x i32> %z2, <vscale x 4 x i32> %dummy_z3, <vscale x 4 x i32> %z4) #0 {
+; CHECK-LABEL: call_svint32x3_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z4.d
+; CHECK-NEXT: bl callee_svint32x3_t
+  %tuple = tail call <vscale x 12 x i32> @llvm.aarch64.sve.tuple.create3.nxv12i32.nxv4i32(<vscale x 4 x i32> %z1, <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z4)
+  call void @callee_svint32x3_t(<vscale x 12 x i32> %tuple)
+  ret void
+}
+
+;
+; svint64x3_t
+;
+
+define <vscale x 6 x i64> @ret_svint64x3_t(<vscale x 2 x i64> %unused_z0, <vscale x 2 x i64> %z1, <vscale x 2 x i64> %z2, <vscale x 2 x i64> %z3) #0 {
+; CHECK-LABEL: ret_svint64x3_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z3.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 6 x i64> @llvm.aarch64.sve.tuple.create3.nxv6i64.nxv2i64(<vscale x 2 x i64> %z1, <vscale x 2 x i64> %z2, <vscale x 2 x i64> %z3)
+  ret <vscale x 6 x i64> %tuple
+}
+
+define void @call_svint64x3_t(<vscale x 2 x i64> %dummy_z0, <vscale x 2 x i64> %z1, <vscale x 2 x i64> %z2, <vscale x 2 x i64> %dummy_z3, <vscale x 2 x i64> %z4) #0 {
+; CHECK-LABEL: call_svint64x3_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z4.d
+; CHECK-NEXT: bl callee_svint64x3_t
+  %tuple = tail call <vscale x 6 x i64> @llvm.aarch64.sve.tuple.create3.nxv6i64.nxv2i64(<vscale x 2 x i64> %z1, <vscale x 2 x i64> %z2, <vscale x 2 x i64> %z4)
+  call void @callee_svint64x3_t(<vscale x 6 x i64> %tuple)
+  ret void
+}
+
+;
+; svfloatx3_t
+;
+
+define <vscale x 12 x float> @ret_svfloatx3_t(<vscale x 4 x float> %unused_z0, <vscale x 4 x float> %z1, <vscale x 4 x float> %z2, <vscale x 4 x float> %z3) #0 {
+; CHECK-LABEL: ret_svfloatx3_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z3.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 12 x float> @llvm.aarch64.sve.tuple.create3.nxv12f32.nxv4f32(<vscale x 4 x float> %z1, <vscale x 4 x float> %z2, <vscale x 4 x float> %z3)
+  ret <vscale x 12 x float> %tuple
+}
+
+define void @call_svfloatx3_t(<vscale x 4 x float> %dummy_z0, <vscale x 4 x float> %z1, <vscale x 4 x float> %z2, <vscale x 4 x float> %dummy_z3, <vscale x 4 x float> %z4) #0 {
+; CHECK-LABEL: call_svfloatx3_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z4.d
+; CHECK-NEXT: bl callee_svfloatx3_t
+  %tuple = tail call <vscale x 12 x float> @llvm.aarch64.sve.tuple.create3.nxv12f32.nxv4f32(<vscale x 4 x float> %z1, <vscale x 4 x float> %z2, <vscale x 4 x float> %z4)
+  call void @callee_svfloatx3_t(<vscale x 12 x float> %tuple)
+  ret void
+}
+
+;
+; svdoublex3_t
+;
+
+define <vscale x 6 x double> @ret_svdoublex3_t(<vscale x 2 x double> %unused_z0, <vscale x 2 x double> %z1, <vscale x 2 x double> %z2, <vscale x 2 x double> %z3) #0 {
+; CHECK-LABEL: ret_svdoublex3_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z3.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 6 x double> @llvm.aarch64.sve.tuple.create3.nxv6f64.nxv2f64(<vscale x 2 x double> %z1, <vscale x 2 x double> %z2, <vscale x 2 x double> %z3)
+  ret <vscale x 6 x double> %tuple
+}
+
+define void @call_svdoublex3_t(<vscale x 2 x double> %dummy_z0, <vscale x 2 x double> %z1, <vscale x 2 x double> %z2, <vscale x 2 x double> %dummy_z3, <vscale x 2 x double> %z4) #0 {
+; CHECK-LABEL: call_svdoublex3_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z4.d
+; CHECK-NEXT: bl callee_svdoublex3_t
+  %tuple = tail call <vscale x 6 x double> @llvm.aarch64.sve.tuple.create3.nxv6f64.nxv2f64(<vscale x 2 x double> %z1, <vscale x 2 x double> %z2, <vscale x 2 x double> %z4)
+  call void @callee_svdoublex3_t(<vscale x 6 x double> %tuple)
+  ret void
+}
+
+;
+; svint8x4_t
+;
+
+define <vscale x 64 x i8> @ret_svint8x4_t(<vscale x 16 x i8> %unused_z0, <vscale x 16 x i8> %z1, <vscale x 16 x i8> %z2, <vscale x 16 x i8> %z3, <vscale x 16 x i8> %z4) #0 {
+; CHECK-LABEL: ret_svint8x4_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z3.d
+; CHECK-NEXT: mov z3.d, z4.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 64 x i8> @llvm.aarch64.sve.tuple.create4.nxv64i8.nxv16i8(<vscale x 16 x i8> %z1, <vscale x 16 x i8> %z2, <vscale x 16 x i8> %z3, <vscale x 16 x i8> %z4)
+  ret <vscale x 64 x i8> %tuple
+}
+
+define void @call_svint8x4_t(<vscale x 16 x i8> %dummy_z0, <vscale x 16 x i8> %z1, <vscale x 16 x i8> %z2, <vscale x 16 x i8> %dummy_z3, <vscale x 16 x i8> %z4, <vscale x 16 x i8> %z5) #0 {
+; CHECK-LABEL: call_svint8x4_t
+; CHECK:      mov z3.d, z5.d
+; CHECK-NEXT: mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z4.d
+; CHECK-NEXT: bl callee_svint8x4_t
+  %tuple = tail call <vscale x 64 x i8> @llvm.aarch64.sve.tuple.create4.nxv64i8.nxv16i8(<vscale x 16 x i8> %z1, <vscale x 16 x i8> %z2, <vscale x 16 x i8> %z4, <vscale x 16 x i8> %z5)
+  call void @callee_svint8x4_t(<vscale x 64 x i8> %tuple)
+  ret void
+}
+
+;
+; svint16x4_t
+;
+
+define <vscale x 32 x i16> @ret_svint16x4_t(<vscale x 8 x i16> %unused_z0, <vscale x 8 x i16> %z1, <vscale x 8 x i16> %z2, <vscale x 8 x i16> %z3, <vscale x 8 x i16> %z4) #0 {
+; CHECK-LABEL: ret_svint16x4_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z3.d
+; CHECK-NEXT: mov z3.d, z4.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 32 x i16> @llvm.aarch64.sve.tuple.create4.nxv32i16.nxv8i16(<vscale x 8 x i16> %z1, <vscale x 8 x i16> %z2, <vscale x 8 x i16> %z3, <vscale x 8 x i16> %z4)
+  ret <vscale x 32 x i16> %tuple
+}
+
+define void @call_svint16x4_t(<vscale x 8 x i16> %dummy_z0, <vscale x 8 x i16> %z1, <vscale x 8 x i16> %z2, <vscale x 8 x i16> %dummy_z3, <vscale x 8 x i16> %z4, <vscale x 8 x i16> %z5) #0 {
+; CHECK-LABEL: call_svint16x4_t
+; CHECK:      mov z3.d, z5.d
+; CHECK-NEXT: mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z4.d
+; CHECK-NEXT: bl callee_svint16x4_t
+  %tuple = tail call <vscale x 32 x i16> @llvm.aarch64.sve.tuple.create4.nxv32i16.nxv8i16(<vscale x 8 x i16> %z1, <vscale x 8 x i16> %z2, <vscale x 8 x i16> %z4, <vscale x 8 x i16> %z5)
+  call void @callee_svint16x4_t(<vscale x 32 x i16> %tuple)
+  ret void
+}
+
+;
+; svint32x4_t
+;
+
+define <vscale x 16 x i32> @ret_svint32x4_t(<vscale x 4 x i32> %unused_z0, <vscale x 4 x i32> %z1, <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z3, <vscale x 4 x i32> %z4) #0 {
+; CHECK-LABEL: ret_svint32x4_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z3.d
+; CHECK-NEXT: mov z3.d, z4.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 16 x i32> @llvm.aarch64.sve.tuple.create4.nxv16i32.nxv4i32(<vscale x 4 x i32> %z1, <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z3, <vscale x 4 x i32> %z4)
+  ret <vscale x 16 x i32> %tuple
+}
+
+define void @call_svint32x4_t(<vscale x 4 x i32> %dummy_z0, <vscale x 4 x i32> %z1, <vscale x 4 x i32> %z2, <vscale x 4 x i32> %dummy_z3, <vscale x 4 x i32> %z4, <vscale x 4 x i32> %z5) #0 {
+; CHECK-LABEL: call_svint32x4_t
+; CHECK:      mov z3.d, z5.d
+; CHECK-NEXT: mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z4.d
+; CHECK-NEXT: bl callee_svint32x4_t
+  %tuple = tail call <vscale x 16 x i32> @llvm.aarch64.sve.tuple.create4.nxv16i32.nxv4i32(<vscale x 4 x i32> %z1, <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z4, <vscale x 4 x i32> %z5)
+  call void @callee_svint32x4_t(<vscale x 16 x i32> %tuple)
+  ret void
+}
+
+;
+; svint64x4_t
+;
+
+define <vscale x 8 x i64> @ret_svint64x4_t(<vscale x 2 x i64> %unused_z0, <vscale x 2 x i64> %z1, <vscale x 2 x i64> %z2, <vscale x 2 x i64> %z3, <vscale x 2 x i64> %z4) #0 {
+; CHECK-LABEL: ret_svint64x4_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z3.d
+; CHECK-NEXT: mov z3.d, z4.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 8 x i64> @llvm.aarch64.sve.tuple.create4.nxv8i64.nxv2i64(<vscale x 2 x i64> %z1, <vscale x 2 x i64> %z2, <vscale x 2 x i64> %z3, <vscale x 2 x i64> %z4)
+  ret <vscale x 8 x i64> %tuple
+}
+
+define void @call_svint64x4_t(<vscale x 2 x i64> %dummy_z0, <vscale x 2 x i64> %z1, <vscale x 2 x i64> %z2, <vscale x 2 x i64> %dummy_z3, <vscale x 2 x i64> %z4, <vscale x 2 x i64> %z5) #0 {
+; CHECK-LABEL: call_svint64x4_t
+; CHECK:      mov z3.d, z5.d
+; CHECK-NEXT: mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z4.d
+; CHECK-NEXT: bl callee_svint64x4_t
+  %tuple = tail call <vscale x 8 x i64> @llvm.aarch64.sve.tuple.create4.nxv8i64.nxv2i64(<vscale x 2 x i64> %z1, <vscale x 2 x i64> %z2, <vscale x 2 x i64> %z4, <vscale x 2 x i64> %z5)
+  call void @callee_svint64x4_t(<vscale x 8 x i64> %tuple)
+  ret void
+}
+
+;
+; svfloatx4_t
+;
+
+define <vscale x 16 x float> @ret_svfloatx4_t(<vscale x 4 x float> %unused_z0, <vscale x 4 x float> %z1, <vscale x 4 x float> %z2, <vscale x 4 x float> %z3, <vscale x 4 x float> %z4) #0 {
+; CHECK-LABEL: ret_svfloatx4_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z3.d
+; CHECK-NEXT: mov z3.d, z4.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 16 x float> @llvm.aarch64.sve.tuple.create4.nxv16f32.nxv4f32(<vscale x 4 x float> %z1, <vscale x 4 x float> %z2, <vscale x 4 x float> %z3, <vscale x 4 x float> %z4)
+  ret <vscale x 16 x float> %tuple
+}
+
+define void @call_svfloatx4_t(<vscale x 4 x float> %dummy_z0, <vscale x 4 x float> %z1, <vscale x 4 x float> %z2, <vscale x 4 x float> %dummy_z3, <vscale x 4 x float> %z4, <vscale x 4 x float> %z5) #0 {
+; CHECK-LABEL: call_svfloatx4_t
+; CHECK:      mov z3.d, z5.d
+; CHECK-NEXT: mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z4.d
+; CHECK-NEXT: bl callee_svfloatx4_t
+  %tuple = tail call <vscale x 16 x float> @llvm.aarch64.sve.tuple.create4.nxv16f32.nxv4f32(<vscale x 4 x float> %z1, <vscale x 4 x float> %z2, <vscale x 4 x float> %z4, <vscale x 4 x float> %z5)
+  call void @callee_svfloatx4_t(<vscale x 16 x float> %tuple)
+  ret void
+}
+
+;
+; svdoublex4_t
+;
+
+define <vscale x 8 x double> @ret_svdoublex4_t(<vscale x 2 x double> %unused_z0, <vscale x 2 x double> %z1, <vscale x 2 x double> %z2, <vscale x 2 x double> %z3, <vscale x 2 x double> %z4) #0 {
+; CHECK-LABEL: ret_svdoublex4_t
+; CHECK:      mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z3.d
+; CHECK-NEXT: mov z3.d, z4.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 8 x double> @llvm.aarch64.sve.tuple.create4.nxv8f64.nxv2f64(<vscale x 2 x double> %z1, <vscale x 2 x double> %z2, <vscale x 2 x double> %z3, <vscale x 2 x double> %z4)
+  ret <vscale x 8 x double> %tuple
+}
+
+define void @call_svdoublex4_t(<vscale x 2 x double> %dummy_z0, <vscale x 2 x double> %z1, <vscale x 2 x double> %z2, <vscale x 2 x double> %dummy_z3, <vscale x 2 x double> %z4, <vscale x 2 x double> %z5) #0 {
+; CHECK-LABEL: call_svdoublex4_t
+; CHECK:      mov z3.d, z5.d
+; CHECK-NEXT: mov z0.d, z1.d
+; CHECK-NEXT: mov z1.d, z2.d
+; CHECK-NEXT: mov z2.d, z4.d
+; CHECK-NEXT: bl callee_svdoublex4_t
+  %tuple = tail call <vscale x 8 x double> @llvm.aarch64.sve.tuple.create4.nxv8f64.nxv2f64(<vscale x 2 x double> %z1, <vscale x 2 x double> %z2, <vscale x 2 x double> %z4, <vscale x 2 x double> %z5)
+  call void @callee_svdoublex4_t(<vscale x 8 x double> %tuple)
+  ret void
+}
+
+attributes #0 = { nounwind "target-features"="+sve" }
+
+declare void @callee_svint8x2_t(<vscale x 32 x i8>)
+declare void @callee_svint16x2_t(<vscale x 16 x i16>)
+declare void @callee_svint32x2_t(<vscale x 8 x i32>)
+declare void @callee_svint64x2_t(<vscale x 4 x i64>)
+declare void @callee_svfloatx2_t(<vscale x 8 x float>)
+declare void @callee_svdoublex2_t(<vscale x 4 x double>)
+
+declare void @callee_svint8x3_t(<vscale x 48 x i8>)
+declare void @callee_svint16x3_t(<vscale x 24 x i16>)
+declare void @callee_svint32x3_t(<vscale x 12 x i32>)
+declare void @callee_svint64x3_t(<vscale x 6 x i64>)
+declare void @callee_svfloatx3_t(<vscale x 12 x float>)
+declare void @callee_svdoublex3_t(<vscale x 6 x double>)
+
+declare void @callee_svint8x4_t(<vscale x 64 x i8>)
+declare void @callee_svint16x4_t(<vscale x 32 x i16>)
+declare void @callee_svint32x4_t(<vscale x 16 x i32>)
+declare void @callee_svint64x4_t(<vscale x 8 x i64>)
+declare void @callee_svfloatx4_t(<vscale x 16 x float>)
+declare void @callee_svdoublex4_t(<vscale x 8 x double>)
+
+
+; x2
+declare <vscale x 32 x i8> @llvm.aarch64.sve.tuple.create2.nxv32i8.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i8>)
+declare <vscale x 16 x i16> @llvm.aarch64.sve.tuple.create2.nxv16i16.nxv8i16(<vscale x 8 x i16>, <vscale x 8 x i16>)
+declare <vscale x 8 x i32> @llvm.aarch64.sve.tuple.create2.nxv8i32.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 4 x i64> @llvm.aarch64.sve.tuple.create2.nxv4i64.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>)
+declare <vscale x 8 x float> @llvm.aarch64.sve.tuple.create2.nxv8f32.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>)
+declare <vscale x 4 x double> @llvm.aarch64.sve.tuple.create2.nxv4f64.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>)
+
+; x3
+declare <vscale x 48 x i8> @llvm.aarch64.sve.tuple.create3.nxv48i8.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>)
+declare <vscale x 24 x i16> @llvm.aarch64.sve.tuple.create3.nxv24i16.nxv8i16(<vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)
+declare <vscale x 12 x i32> @llvm.aarch64.sve.tuple.create3.nxv12i32.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 6 x i64> @llvm.aarch64.sve.tuple.create3.nxv6i64.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>)
+declare <vscale x 12 x float> @llvm.aarch64.sve.tuple.create3.nxv12f32.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>)
+declare <vscale x 6 x double> @llvm.aarch64.sve.tuple.create3.nxv6f64.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>)
+
+; x4
+declare <vscale x 64 x i8> @llvm.aarch64.sve.tuple.create4.nxv64i8.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>)
+declare <vscale x 32 x i16> @llvm.aarch64.sve.tuple.create4.nxv32i16.nxv8i16(<vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)
+declare <vscale x 16 x i32> @llvm.aarch64.sve.tuple.create4.nxv16i32.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 8 x i64> @llvm.aarch64.sve.tuple.create4.nxv8i64.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>)
+declare <vscale x 16 x float> @llvm.aarch64.sve.tuple.create4.nxv16f32.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>)
+declare <vscale x 8 x double> @llvm.aarch64.sve.tuple.create4.nxv8f64.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>)

llvm/test/CodeGen/AArch64/sve-intrinsics-create-tuple.ll

+; RUN: llc -mtriple aarch64 -mattr=+sve -asm-verbose=0 < %s | FileCheck %s
+
+;
+; SVCREATE2 (i8)
+;
+
+define <vscale x 16 x i8> @test_svcreate2_s8_vec0(<vscale x 16 x i8> %x0, <vscale x 16 x i8> %x1) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate2_s8_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 32 x i8> @llvm.aarch64.sve.tuple.create2.nxv32i8.nxv16i8(<vscale x 16 x i8> %x0, <vscale x 16 x i8> %x1)
+  %extract = tail call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv32i8(<vscale x 32 x i8> %tuple, i32 0)
+  ret <vscale x 16 x i8> %extract
+}
+
+define <vscale x 16 x i8> @test_svcreate2_s8_vec1(<vscale x 16 x i8> %x0, <vscale x 16 x i8> %x1) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate2_s8_vec1:
+; CHECK-NEXT: mov z0.d, z1.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 32 x i8> @llvm.aarch64.sve.tuple.create2.nxv32i8.nxv16i8(<vscale x 16 x i8> %x0, <vscale x 16 x i8> %x1)
+  %extract = tail call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv32i8(<vscale x 32 x i8> %tuple, i32 1)
+  ret <vscale x 16 x i8> %extract
+}
+
+;
+; SVCREATE2 (i16)
+;
+
+define <vscale x 8 x i16> @test_svcreate2_s16_vec0(<vscale x 8 x i16> %x0, <vscale x 8 x i16> %x1) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate2_s16_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 16 x i16> @llvm.aarch64.sve.tuple.create2.nxv16i16.nxv8i16(<vscale x 8 x i16> %x0, <vscale x 8 x i16> %x1)
+  %extract = tail call <vscale x 8 x i16> @llvm.aarch64.sve.tuple.get.nxv8i16.nxv16i16(<vscale x 16 x i16> %tuple, i32 0)
+  ret <vscale x 8 x i16> %extract
+}
+
+define <vscale x 8 x i16> @test_svcreate2_s16_vec1(<vscale x 8 x i16> %x0, <vscale x 8 x i16> %x1) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate2_s16_vec1:
+; CHECK-NEXT: mov z0.d, z1.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 16 x i16> @llvm.aarch64.sve.tuple.create2.nxv16i16.nxv8i16(<vscale x 8 x i16> %x0, <vscale x 8 x i16> %x1)
+  %extract = tail call <vscale x 8 x i16> @llvm.aarch64.sve.tuple.get.nxv8i16.nxv16i16(<vscale x 16 x i16> %tuple, i32 1)
+  ret <vscale x 8 x i16> %extract
+}
+
+;
+; SVCREATE2 (half)
+;
+
+define <vscale x 8 x half> @test_svcreate2_f16_vec0(<vscale x 8 x half> %x0, <vscale x 8 x half> %x1) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate2_f16_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 16 x half> @llvm.aarch64.sve.tuple.create2.nxv16f16.nxv8f16(<vscale x 8 x half> %x0, <vscale x 8 x half> %x1)
+  %extract = tail call <vscale x 8 x half> @llvm.aarch64.sve.tuple.get.nxv8f16.nxv16f16(<vscale x 16 x half> %tuple, i32 0)
+  ret <vscale x 8 x half> %extract
+}
+
+define <vscale x 8 x half> @test_svcreate2_f16_vec1(<vscale x 8 x half> %x0, <vscale x 8 x half> %x1) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate2_f16_vec1:
+; CHECK-NEXT: mov z0.d, z1.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 16 x half> @llvm.aarch64.sve.tuple.create2.nxv16f16.nxv8f16(<vscale x 8 x half> %x0, <vscale x 8 x half> %x1)
+  %extract = tail call <vscale x 8 x half> @llvm.aarch64.sve.tuple.get.nxv8f16.nxv16f16(<vscale x 16 x half> %tuple, i32 1)
+  ret <vscale x 8 x half> %extract
+}
+
+;
+; SVCREATE2 (i32)
+;
+
+define <vscale x 4 x i32> @test_svcreate2_s32_vec0(<vscale x 4 x i32> %x0, <vscale x 4 x i32> %x1) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate2_s32_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 8 x i32> @llvm.aarch64.sve.tuple.create2.nxv8i32.nxv4i32(<vscale x 4 x i32> %x0, <vscale x 4 x i32> %x1)
+  %extract = tail call <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv4i32.nxv8i32(<vscale x 8 x i32> %tuple, i32 0)
+  ret <vscale x 4 x i32> %extract
+}
+
+define <vscale x 4 x i32> @test_svcreate2_s32_vec1(<vscale x 4 x i32> %x0, <vscale x 4 x i32> %x1) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate2_s32_vec1:
+; CHECK-NEXT: mov z0.d, z1.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 8 x i32> @llvm.aarch64.sve.tuple.create2.nxv8i32.nxv4i32(<vscale x 4 x i32> %x0, <vscale x 4 x i32> %x1)
+  %extract = tail call <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv4i32.nxv8i32(<vscale x 8 x i32> %tuple, i32 1)
+  ret <vscale x 4 x i32> %extract
+}
+
+;
+; SVCREATE2 (float)
+;
+
+define <vscale x 4 x float> @test_svcreate2_f32_vec0(<vscale x 4 x float> %x0, <vscale x 4 x float> %x1) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate2_f32_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 8 x float> @llvm.aarch64.sve.tuple.create2.nxv8f32.nxv4f32(<vscale x 4 x float> %x0, <vscale x 4 x float> %x1)
+  %extract = tail call <vscale x 4 x float> @llvm.aarch64.sve.tuple.get.nxv4f32.nxv8f32(<vscale x 8 x float> %tuple, i32 0)
+  ret <vscale x 4 x float> %extract
+}
+
+define <vscale x 4 x float> @test_svcreate2_f32_vec1(<vscale x 4 x float> %x0, <vscale x 4 x float> %x1) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate2_f32_vec1:
+; CHECK-NEXT: mov z0.d, z1.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 8 x float> @llvm.aarch64.sve.tuple.create2.nxv8f32.nxv4f32(<vscale x 4 x float> %x0, <vscale x 4 x float> %x1)
+  %extract = tail call <vscale x 4 x float> @llvm.aarch64.sve.tuple.get.nxv4f32.nxv8f32(<vscale x 8 x float> %tuple, i32 1)
+  ret <vscale x 4 x float> %extract
+}
+
+;
+; SVCREATE2 (i64)
+;
+
+define <vscale x 2 x i64> @test_svcreate2_s64_vec0(<vscale x 2 x i64> %x0, <vscale x 2 x i64> %x1) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate2_s64_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 4 x i64> @llvm.aarch64.sve.tuple.create2.nxv4i64.nxv2i64(<vscale x 2 x i64> %x0, <vscale x 2 x i64> %x1)
+  %extract = tail call <vscale x 2 x i64> @llvm.aarch64.sve.tuple.get.nxv2i64.nxv4i64(<vscale x 4 x i64> %tuple, i32 0)
+  ret <vscale x 2 x i64> %extract
+}
+
+define <vscale x 2 x i64> @test_svcreate2_s64_vec1(<vscale x 2 x i64> %x0, <vscale x 2 x i64> %x1) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate2_s64_vec1:
+; CHECK-NEXT: mov z0.d, z1.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 4 x i64> @llvm.aarch64.sve.tuple.create2.nxv4i64.nxv2i64(<vscale x 2 x i64> %x0, <vscale x 2 x i64> %x1)
+  %extract = tail call <vscale x 2 x i64> @llvm.aarch64.sve.tuple.get.nxv2i64.nxv4i64(<vscale x 4 x i64> %tuple, i32 1)
+  ret <vscale x 2 x i64> %extract
+}
+
+;
+; SVCREATE2 (double)
+;
+
+define <vscale x 2 x double> @test_svcreate2_f64_vec0(<vscale x 2 x double> %x0, <vscale x 2 x double> %x1) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate2_f64_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 4 x double> @llvm.aarch64.sve.tuple.create2.nxv4f64.nxv2f64(<vscale x 2 x double> %x0, <vscale x 2 x double> %x1)
+  %extract = tail call <vscale x 2 x double> @llvm.aarch64.sve.tuple.get.nxv2f64.nxv4f64(<vscale x 4 x double> %tuple, i32 0)
+  ret <vscale x 2 x double> %extract
+}
+
+define <vscale x 2 x double> @test_svcreate2_f64_vec1(<vscale x 2 x double> %x0, <vscale x 2 x double> %x1) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate2_f64_vec1:
+; CHECK-NEXT: mov z0.d, z1.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 4 x double> @llvm.aarch64.sve.tuple.create2.nxv4f64.nxv2f64(<vscale x 2 x double> %x0, <vscale x 2 x double> %x1)
+  %extract = tail call <vscale x 2 x double> @llvm.aarch64.sve.tuple.get.nxv2f64.nxv4f64(<vscale x 4 x double> %tuple, i32 1)
+  ret <vscale x 2 x double> %extract
+}
+
+;
+; SVCREATE3 (i8)
+;
+
+define <vscale x 16 x i8> @test_svcreate3_s8_vec0(<vscale x 16 x i8> %x0, <vscale x 16 x i8> %x1, <vscale x 16 x i8> %x2) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate3_s8_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 48 x i8> @llvm.aarch64.sve.tuple.create3.nxv48i8.nxv16i8(<vscale x 16 x i8> %x0, <vscale x 16 x i8> %x1, <vscale x 16 x i8> %x2)
+  %extract = tail call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv48i8(<vscale x 48 x i8> %tuple, i32 0)
+  ret <vscale x 16 x i8> %extract
+}
+
+define <vscale x 16 x i8> @test_svcreate3_s8_vec2(<vscale x 16 x i8> %x0, <vscale x 16 x i8> %x1, <vscale x 16 x i8> %x2) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate3_s8_vec2:
+; CHECK-NEXT: mov z0.d, z2.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 48 x i8> @llvm.aarch64.sve.tuple.create3.nxv48i8.nxv16i8(<vscale x 16 x i8> %x0, <vscale x 16 x i8> %x1, <vscale x 16 x i8> %x2)
+  %extract = tail call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv48i8(<vscale x 48 x i8> %tuple, i32 2)
+  ret <vscale x 16 x i8> %extract
+}
+
+;
+; SVCREATE3 (i16)
+;
+
+define <vscale x 8 x i16> @test_svcreate3_s16_vec0(<vscale x 8 x i16> %x0, <vscale x 8 x i16> %x1, <vscale x 8 x i16> %x2) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate3_s16_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 24 x i16> @llvm.aarch64.sve.tuple.create3.nxv24i16.nxv8i16(<vscale x 8 x i16> %x0, <vscale x 8 x i16> %x1, <vscale x 8 x i16> %x2)
+  %extract = tail call <vscale x 8 x i16> @llvm.aarch64.sve.tuple.get.nxv8i16.nxv24i16(<vscale x 24 x i16> %tuple, i32 0)
+  ret <vscale x 8 x i16> %extract
+}
+
+define <vscale x 8 x i16> @test_svcreate3_s16_vec2(<vscale x 8 x i16> %x0, <vscale x 8 x i16> %x1, <vscale x 8 x i16> %x2) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate3_s16_vec2:
+; CHECK-NEXT: mov z0.d, z2.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 24 x i16> @llvm.aarch64.sve.tuple.create3.nxv24i16.nxv8i16(<vscale x 8 x i16> %x0, <vscale x 8 x i16> %x1, <vscale x 8 x i16> %x2)
+  %extract = tail call <vscale x 8 x i16> @llvm.aarch64.sve.tuple.get.nxv8i16.nxv24i16(<vscale x 24 x i16> %tuple, i32 2)
+  ret <vscale x 8 x i16> %extract
+}
+;
+; SVCREATE3 (half)
+;
+
+define <vscale x 8 x half> @test_svcreate3_f16_vec0(<vscale x 8 x half> %x0, <vscale x 8 x half> %x1, <vscale x 8 x half> %x2) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate3_f16_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 24 x half> @llvm.aarch64.sve.tuple.create3.nxv24f16.nxv8f16(<vscale x 8 x half> %x0, <vscale x 8 x half> %x1, <vscale x 8 x half> %x2)
+  %extract = tail call <vscale x 8 x half> @llvm.aarch64.sve.tuple.get.nxv8f16.nxv24f16(<vscale x 24 x half> %tuple, i32 0)
+  ret <vscale x 8 x half> %extract
+}
+
+define <vscale x 8 x half> @test_svcreate3_f16_vec2(<vscale x 8 x half> %x0, <vscale x 8 x half> %x1, <vscale x 8 x half> %x2) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate3_f16_vec2:
+; CHECK-NEXT: mov z0.d, z2.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 24 x half> @llvm.aarch64.sve.tuple.create3.nxv24f16.nxv8f16(<vscale x 8 x half> %x0, <vscale x 8 x half> %x1, <vscale x 8 x half> %x2)
+  %extract = tail call <vscale x 8 x half> @llvm.aarch64.sve.tuple.get.nxv8f16.nxv24f16(<vscale x 24 x half> %tuple, i32 2)
+  ret <vscale x 8 x half> %extract
+}
+
+
+;
+; SVCREATE3 (i32)
+;
+
+define <vscale x 4 x i32> @test_svcreate3_s32_vec0(<vscale x 4 x i32> %x0, <vscale x 4 x i32> %x1, <vscale x 4 x i32> %x2) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate3_s32_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 12 x i32> @llvm.aarch64.sve.tuple.create3.nxv12i32.nxv4i32(<vscale x 4 x i32> %x0, <vscale x 4 x i32> %x1, <vscale x 4 x i32> %x2)
+  %extract = tail call <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv4i32.nxv12i32(<vscale x 12 x i32> %tuple, i32 0)
+  ret <vscale x 4 x i32> %extract
+}
+
+define <vscale x 4 x i32> @test_svcreate3_s32_vec2(<vscale x 4 x i32> %x0, <vscale x 4 x i32> %x1, <vscale x 4 x i32> %x2) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate3_s32_vec2:
+; CHECK-NEXT: mov z0.d, z2.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 12 x i32> @llvm.aarch64.sve.tuple.create3.nxv12i32.nxv4i32(<vscale x 4 x i32> %x0, <vscale x 4 x i32> %x1, <vscale x 4 x i32> %x2)
+  %extract = tail call <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv4i32.nxv12i32(<vscale x 12 x i32> %tuple, i32 2)
+  ret <vscale x 4 x i32> %extract
+}
+
+;
+; SVCREATE3 (float)
+;
+
+define <vscale x 4 x float> @test_svcreate3_f32_vec0(<vscale x 4 x float> %x0, <vscale x 4 x float> %x1, <vscale x 4 x float> %x2) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate3_f32_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 12 x float> @llvm.aarch64.sve.tuple.create3.nxv12f32.nxv4f32(<vscale x 4 x float> %x0, <vscale x 4 x float> %x1, <vscale x 4 x float> %x2)
+  %extract = tail call <vscale x 4 x float> @llvm.aarch64.sve.tuple.get.nxv4f32.nxv12f32(<vscale x 12 x float> %tuple, i32 0)
+  ret <vscale x 4 x float> %extract
+}
+
+define <vscale x 4 x float> @test_svcreate3_f32_vec2(<vscale x 4 x float> %x0, <vscale x 4 x float> %x1, <vscale x 4 x float> %x2) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate3_f32_vec2:
+; CHECK-NEXT: mov z0.d, z2.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 12 x float> @llvm.aarch64.sve.tuple.create3.nxv12f32.nxv4f32(<vscale x 4 x float> %x0, <vscale x 4 x float> %x1, <vscale x 4 x float> %x2)
+  %extract = tail call <vscale x 4 x float> @llvm.aarch64.sve.tuple.get.nxv4f32.nxv12f32(<vscale x 12 x float> %tuple, i32 2)
+  ret <vscale x 4 x float> %extract
+}
+
+;
+; SVCREATE3 (i64)
+;
+
+define <vscale x 2 x i64> @test_svcreate3_s64_vec0(<vscale x 2 x i64> %x0, <vscale x 2 x i64> %x1, <vscale x 2 x i64> %x2) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate3_s64_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 6 x i64> @llvm.aarch64.sve.tuple.create3.nxv6i64.nxv2i64(<vscale x 2 x i64> %x0, <vscale x 2 x i64> %x1, <vscale x 2 x i64> %x2)
+  %extract = tail call <vscale x 2 x i64> @llvm.aarch64.sve.tuple.get.nxv2i64.nxv6i64(<vscale x 6 x i64> %tuple, i32 0)
+  ret <vscale x 2 x i64> %extract
+}
+
+define <vscale x 2 x i64> @test_svcreate3_s64_vec2(<vscale x 2 x i64> %x0, <vscale x 2 x i64> %x1, <vscale x 2 x i64> %x2) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate3_s64_vec2:
+; CHECK-NEXT: mov z0.d, z2.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 6 x i64> @llvm.aarch64.sve.tuple.create3.nxv6i64.nxv2i64(<vscale x 2 x i64> %x0, <vscale x 2 x i64> %x1, <vscale x 2 x i64> %x2)
+  %extract = tail call <vscale x 2 x i64> @llvm.aarch64.sve.tuple.get.nxv2i64.nxv6i64(<vscale x 6 x i64> %tuple, i32 2)
+  ret <vscale x 2 x i64> %extract
+}
+
+;
+; SVCREATE3 (double)
+;
+
+define <vscale x 2 x double> @test_svcreate3_f64_vec0(<vscale x 2 x double> %x0, <vscale x 2 x double> %x1, <vscale x 2 x double> %x2) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate3_f64_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 6 x double> @llvm.aarch64.sve.tuple.create3.nxv6f64.nxv2f64(<vscale x 2 x double> %x0, <vscale x 2 x double> %x1, <vscale x 2 x double> %x2)
+  %extract = tail call <vscale x 2 x double> @llvm.aarch64.sve.tuple.get.nxv2f64.nxv6f64(<vscale x 6 x double> %tuple, i32 0)
+  ret <vscale x 2 x double> %extract
+}
+
+define <vscale x 2 x double> @test_svcreate3_f64_vec2(<vscale x 2 x double> %x0, <vscale x 2 x double> %x1, <vscale x 2 x double> %x2) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate3_f64_vec2:
+; CHECK-NEXT: mov z0.d, z2.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 6 x double> @llvm.aarch64.sve.tuple.create3.nxv6f64.nxv2f64(<vscale x 2 x double> %x0, <vscale x 2 x double> %x1, <vscale x 2 x double> %x2)
+  %extract = tail call <vscale x 2 x double> @llvm.aarch64.sve.tuple.get.nxv2f64.nxv6f64(<vscale x 6 x double> %tuple, i32 2)
+  ret <vscale x 2 x double> %extract
+}
+
+;
+; SVCREATE4 (i8)
+;
+
+define <vscale x 16 x i8> @test_svcreate4_s8_vec0(<vscale x 16 x i8> %x0, <vscale x 16 x i8> %x1, <vscale x 16 x i8> %x2, <vscale x 16 x i8> %x3) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate4_s8_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 64 x i8> @llvm.aarch64.sve.tuple.create4.nxv64i8.nxv16i8(<vscale x 16 x i8> %x0, <vscale x 16 x i8> %x1, <vscale x 16 x i8> %x2, <vscale x 16 x i8> %x3)
+  %extract = tail call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv64i8(<vscale x 64 x i8> %tuple, i32 0)
+  ret <vscale x 16 x i8> %extract
+}
+
+define <vscale x 16 x i8> @test_svcreate4_s8_vec3(<vscale x 16 x i8> %x0, <vscale x 16 x i8> %x1, <vscale x 16 x i8> %x2, <vscale x 16 x i8> %x3) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate4_s8_vec3:
+; CHECK-NEXT: mov z0.d, z3.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 64 x i8> @llvm.aarch64.sve.tuple.create4.nxv64i8.nxv16i8(<vscale x 16 x i8> %x0, <vscale x 16 x i8> %x1, <vscale x 16 x i8> %x2, <vscale x 16 x i8> %x3)
+  %extract = tail call <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv64i8(<vscale x 64 x i8> %tuple, i32 3)
+  ret <vscale x 16 x i8> %extract
+}
+
+;
+; SVCREATE4 (i16)
+;
+
+define <vscale x 8 x i16> @test_svcreate4_s16_vec0(<vscale x 8 x i16> %x0, <vscale x 8 x i16> %x1, <vscale x 8 x i16> %x2, <vscale x 8 x i16> %x3) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate4_s16_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 32 x i16> @llvm.aarch64.sve.tuple.create4.nxv32i16.nxv8i16(<vscale x 8 x i16> %x0, <vscale x 8 x i16> %x1, <vscale x 8 x i16> %x2, <vscale x 8 x i16> %x3)
+  %extract = tail call <vscale x 8 x i16> @llvm.aarch64.sve.tuple.get.nxv8i16.nxv32i16(<vscale x 32 x i16> %tuple, i32 0)
+  ret <vscale x 8 x i16> %extract
+}
+
+define <vscale x 8 x i16> @test_svcreate4_s16_vec3(<vscale x 8 x i16> %x0, <vscale x 8 x i16> %x1, <vscale x 8 x i16> %x2, <vscale x 8 x i16> %x3) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate4_s16_vec3:
+; CHECK-NEXT: mov z0.d, z3.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 32 x i16> @llvm.aarch64.sve.tuple.create4.nxv32i16.nxv8i16(<vscale x 8 x i16> %x0, <vscale x 8 x i16> %x1, <vscale x 8 x i16> %x2, <vscale x 8 x i16> %x3)
+  %extract = tail call <vscale x 8 x i16> @llvm.aarch64.sve.tuple.get.nxv8i16.nxv32i16(<vscale x 32 x i16> %tuple, i32 3)
+  ret <vscale x 8 x i16> %extract
+}
+
+;
+; SVCREATE4 (half)
+;
+
+define <vscale x 8 x half> @test_svcreate4_f16_vec0(<vscale x 8 x half> %x0, <vscale x 8 x half> %x1, <vscale x 8 x half> %x2, <vscale x 8 x half> %x3) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate4_f16_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 32 x half> @llvm.aarch64.sve.tuple.create4.nxv32f16.nxv8f16(<vscale x 8 x half> %x0, <vscale x 8 x half> %x1, <vscale x 8 x half> %x2, <vscale x 8 x half> %x3)
+  %extract = tail call <vscale x 8 x half> @llvm.aarch64.sve.tuple.get.nxv8f16.nxv32f16(<vscale x 32 x half> %tuple, i32 0)
+  ret <vscale x 8 x half> %extract
+}
+
+define <vscale x 8 x half> @test_svcreate4_f16_vec3(<vscale x 8 x half> %x0, <vscale x 8 x half> %x1, <vscale x 8 x half> %x2, <vscale x 8 x half> %x3) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate4_f16_vec3:
+; CHECK-NEXT: mov z0.d, z3.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 32 x half> @llvm.aarch64.sve.tuple.create4.nxv32f16.nxv8f16(<vscale x 8 x half> %x0, <vscale x 8 x half> %x1, <vscale x 8 x half> %x2, <vscale x 8 x half> %x3)
+  %extract = tail call <vscale x 8 x half> @llvm.aarch64.sve.tuple.get.nxv8f16.nxv32f16(<vscale x 32 x half> %tuple, i32 3)
+  ret <vscale x 8 x half> %extract
+}
+
+;
+; SVCREATE4 (i32)
+;
+
+define <vscale x 4 x i32> @test_svcreate4_s32_vec0(<vscale x 4 x i32> %x0, <vscale x 4 x i32> %x1, <vscale x 4 x i32> %x2, <vscale x 4 x i32> %x3) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate4_s32_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 16 x i32> @llvm.aarch64.sve.tuple.create4.nxv16i32.nxv4i32(<vscale x 4 x i32> %x0, <vscale x 4 x i32> %x1, <vscale x 4 x i32> %x2, <vscale x 4 x i32> %x3)
+  %extract = tail call <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv4i32.nxv16i32(<vscale x 16 x i32> %tuple, i32 0)
+  ret <vscale x 4 x i32> %extract
+}
+
+define <vscale x 4 x i32> @test_svcreate4_s32_vec3(<vscale x 4 x i32> %x0, <vscale x 4 x i32> %x1, <vscale x 4 x i32> %x2, <vscale x 4 x i32> %x3) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate4_s32_vec3:
+; CHECK-NEXT: mov z0.d, z3.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 16 x i32> @llvm.aarch64.sve.tuple.create4.nxv16i32.nxv4i32(<vscale x 4 x i32> %x0, <vscale x 4 x i32> %x1, <vscale x 4 x i32> %x2, <vscale x 4 x i32> %x3)
+  %extract = tail call <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv4i32.nxv16i32(<vscale x 16 x i32> %tuple, i32 3)
+  ret <vscale x 4 x i32> %extract
+}
+
+;
+; SVCREATE4 (float)
+;
+
+define <vscale x 4 x float> @test_svcreate4_f32_vec0(<vscale x 4 x float> %x0, <vscale x 4 x float> %x1, <vscale x 4 x float> %x2, <vscale x 4 x float> %x3) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate4_f32_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 16 x float> @llvm.aarch64.sve.tuple.create4.nxv16f32.nxv4f32(<vscale x 4 x float> %x0, <vscale x 4 x float> %x1, <vscale x 4 x float> %x2, <vscale x 4 x float> %x3)
+  %extract = tail call <vscale x 4 x float> @llvm.aarch64.sve.tuple.get.nxv4f32.nxv16f32(<vscale x 16 x float> %tuple, i32 0)
+  ret <vscale x 4 x float> %extract
+}
+
+define <vscale x 4 x float> @test_svcreate4_f32_vec3(<vscale x 4 x float> %x0, <vscale x 4 x float> %x1, <vscale x 4 x float> %x2, <vscale x 4 x float> %x3) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate4_f32_vec3:
+; CHECK-NEXT: mov z0.d, z3.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 16 x float> @llvm.aarch64.sve.tuple.create4.nxv16f32.nxv4f32(<vscale x 4 x float> %x0, <vscale x 4 x float> %x1, <vscale x 4 x float> %x2, <vscale x 4 x float> %x3)
+  %extract = tail call <vscale x 4 x float> @llvm.aarch64.sve.tuple.get.nxv4f32.nxv16f32(<vscale x 16 x float> %tuple, i32 3)
+  ret <vscale x 4 x float> %extract
+}
+
+;
+; SVCREATE4 (i64)
+;
+
+define <vscale x 2 x i64> @test_svcreate4_s64_vec0(<vscale x 2 x i64> %x0, <vscale x 2 x i64> %x1, <vscale x 2 x i64> %x2, <vscale x 2 x i64> %x3) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate4_s64_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 8 x i64> @llvm.aarch64.sve.tuple.create4.nxv8i64.nxv2i64(<vscale x 2 x i64> %x0, <vscale x 2 x i64> %x1, <vscale x 2 x i64> %x2, <vscale x 2 x i64> %x3)
+  %extract = tail call <vscale x 2 x i64> @llvm.aarch64.sve.tuple.get.nxv2i64.nxv8i64(<vscale x 8 x i64> %tuple, i32 0)
+  ret <vscale x 2 x i64> %extract
+}
+
+define <vscale x 2 x i64> @test_svcreate4_s64_vec3(<vscale x 2 x i64> %x0, <vscale x 2 x i64> %x1, <vscale x 2 x i64> %x2, <vscale x 2 x i64> %x3) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate4_s64_vec3:
+; CHECK-NEXT: mov z0.d, z3.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 8 x i64> @llvm.aarch64.sve.tuple.create4.nxv8i64.nxv2i64(<vscale x 2 x i64> %x0, <vscale x 2 x i64> %x1, <vscale x 2 x i64> %x2, <vscale x 2 x i64> %x3)
+  %extract = tail call <vscale x 2 x i64> @llvm.aarch64.sve.tuple.get.nxv2i64.nxv8i64(<vscale x 8 x i64> %tuple, i32 3)
+  ret <vscale x 2 x i64> %extract
+}
+
+;
+; SVCREATE4 (double)
+;
+
+define <vscale x 2 x double> @test_svcreate4_f64_vec0(<vscale x 2 x double> %x0, <vscale x 2 x double> %x1, <vscale x 2 x double> %x2, <vscale x 2 x double> %x3) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate4_f64_vec0:
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 8 x double> @llvm.aarch64.sve.tuple.create4.nxv8f64.nxv2f64(<vscale x 2 x double> %x0, <vscale x 2 x double> %x1, <vscale x 2 x double> %x2, <vscale x 2 x double> %x3)
+  %extract = tail call <vscale x 2 x double> @llvm.aarch64.sve.tuple.get.nxv2f64.nxv8f64(<vscale x 8 x double> %tuple, i32 0)
+  ret <vscale x 2 x double> %extract
+}
+
+define <vscale x 2 x double> @test_svcreate4_f64_vec3(<vscale x 2 x double> %x0, <vscale x 2 x double> %x1, <vscale x 2 x double> %x2, <vscale x 2 x double> %x3) local_unnamed_addr #0 {
+; CHECK-LABEL: test_svcreate4_f64_vec3:
+; CHECK-NEXT: mov z0.d, z3.d
+; CHECK-NEXT: ret
+  %tuple = tail call <vscale x 8 x double> @llvm.aarch64.sve.tuple.create4.nxv8f64.nxv2f64(<vscale x 2 x double> %x0, <vscale x 2 x double> %x1, <vscale x 2 x double> %x2, <vscale x 2 x double> %x3)
+  %extract = tail call <vscale x 2 x double> @llvm.aarch64.sve.tuple.get.nxv2f64.nxv8f64(<vscale x 8 x double> %tuple, i32 3)
+  ret <vscale x 2 x double> %extract
+}
+
+attributes #0 = { nounwind "target-features"="+sve" }
+
+declare <vscale x 4 x double>  @llvm.aarch64.sve.tuple.create2.nxv4f64.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>)
+declare <vscale x 8 x float>  @llvm.aarch64.sve.tuple.create2.nxv8f32.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>)
+declare <vscale x 16 x half>  @llvm.aarch64.sve.tuple.create2.nxv16f16.nxv8f16(<vscale x 8 x half>, <vscale x 8 x half>)
+declare <vscale x 4 x i64>  @llvm.aarch64.sve.tuple.create2.nxv4i64.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>)
+declare <vscale x 8 x i32>  @llvm.aarch64.sve.tuple.create2.nxv8i32.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 16 x i16> @llvm.aarch64.sve.tuple.create2.nxv16i16.nxv8i16(<vscale x 8 x i16>, <vscale x 8 x i16>)
+declare <vscale x 32 x i8>  @llvm.aarch64.sve.tuple.create2.nxv32i8.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i8>)
+
+declare <vscale x 6 x double>  @llvm.aarch64.sve.tuple.create3.nxv6f64.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>)
+declare <vscale x 12 x float> @llvm.aarch64.sve.tuple.create3.nxv12f32.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>)
+declare <vscale x 24 x half> @llvm.aarch64.sve.tuple.create3.nxv24f16.nxv8f16(<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
+declare <vscale x 6 x i64>  @llvm.aarch64.sve.tuple.create3.nxv6i64.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>)
+declare <vscale x 12 x i32> @llvm.aarch64.sve.tuple.create3.nxv12i32.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 24 x i16> @llvm.aarch64.sve.tuple.create3.nxv24i16.nxv8i16(<vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)
+declare <vscale x 48 x i8>  @llvm.aarch64.sve.tuple.create3.nxv48i8.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>)
+
+declare <vscale x 8 x double>  @llvm.aarch64.sve.tuple.create4.nxv8f64.nxv2f64 (<vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>, <vscale x 2 x double>)
+declare <vscale x 16 x float>  @llvm.aarch64.sve.tuple.create4.nxv16f32.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>, <vscale x 4 x float>)
+declare <vscale x 32 x half>  @llvm.aarch64.sve.tuple.create4.nxv32f16.nxv8f16(<vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>, <vscale x 8 x half>)
+declare <vscale x  8 x i64> @llvm.aarch64.sve.tuple.create4.nxv8i64.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>, <vscale x 2 x i64>)
+declare <vscale x 16 x i32> @llvm.aarch64.sve.tuple.create4.nxv16i32.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 32 x i16> @llvm.aarch64.sve.tuple.create4.nxv32i16.nxv8i16(<vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>, <vscale x 8 x i16>)
+declare <vscale x 64 x i8>  @llvm.aarch64.sve.tuple.create4.nxv64i8.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>, <vscale x 16 x i8>)
+
+declare <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv32i8(<vscale x 32 x i8>, i32 immarg)
+declare <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv48i8(<vscale x 48 x i8>, i32 immarg)
+declare <vscale x 16 x i8> @llvm.aarch64.sve.tuple.get.nxv16i8.nxv64i8(<vscale x 64 x i8>, i32 immarg)
+
+declare <vscale x 8 x i16> @llvm.aarch64.sve.tuple.get.nxv8i16.nxv16i16(<vscale x 16 x i16>, i32 immarg)
+declare <vscale x 8 x i16> @llvm.aarch64.sve.tuple.get.nxv8i16.nxv24i16(<vscale x 24 x i16>, i32 immarg)
+declare <vscale x 8 x i16> @llvm.aarch64.sve.tuple.get.nxv8i16.nxv32i16(<vscale x 32 x i16>, i32 immarg)
+
+declare <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv4i32.nxv8i32(<vscale x 8 x i32>, i32 immarg)
+declare <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv4i32.nxv12i32(<vscale x 12 x i32>, i32 immarg)
+declare <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv4i32.nxv16i32(<vscale x 16 x i32>, i32 immarg)
+
+declare <vscale x 2 x i64> @llvm.aarch64.sve.tuple.get.nxv2i64.nxv4i64(<vscale x 4 x i64>, i32 immarg)
+declare <vscale x 2 x i64> @llvm.aarch64.sve.tuple.get.nxv2i64.nxv6i64(<vscale x 6 x i64>, i32 immarg)
+declare <vscale x 2 x i64> @llvm.aarch64.sve.tuple.get.nxv2i64.nxv8i64(<vscale x 8 x i64>, i32 immarg)
+
+declare <vscale x 8 x half> @llvm.aarch64.sve.tuple.get.nxv8f16.nxv16f16(<vscale x 16 x half>, i32 immarg)
+declare <vscale x 8 x half> @llvm.aarch64.sve.tuple.get.nxv8f16.nxv24f16(<vscale x 24 x half>, i32 immarg)
+declare <vscale x 8 x half> @llvm.aarch64.sve.tuple.get.nxv8f16.nxv32f16(<vscale x 32 x half>, i32 immarg)
+
+declare <vscale x 4 x float> @llvm.aarch64.sve.tuple.get.nxv4f32.nxv8f32(<vscale x 8 x float>, i32 immarg)
+declare <vscale x 4 x float> @llvm.aarch64.sve.tuple.get.nxv4f32.nxv12f32(<vscale x 12 x float>, i32 immarg)
+declare <vscale x 4 x float> @llvm.aarch64.sve.tuple.get.nxv4f32.nxv16f32(<vscale x 16 x float>, i32 immarg)
+
+declare <vscale x 2 x double> @llvm.aarch64.sve.tuple.get.nxv2f64.nxv4f64(<vscale x 4 x double>, i32 immarg)
+declare <vscale x 2 x double> @llvm.aarch64.sve.tuple.get.nxv2f64.nxv6f64(<vscale x 6 x double>, i32 immarg)
+declare <vscale x 2 x double> @llvm.aarch64.sve.tuple.get.nxv2f64.nxv8f64(<vscale x 8 x double>, i32 immarg)

llvm/test/CodeGen/AArch64/sve-intrinsics-insert-extract-tuple.ll

+; RUN: llc -mtriple aarch64 -mattr=+sve -asm-verbose=0 < %s | FileCheck %s
+
+; All these tests create a vector tuple, insert z5 into one of the elements,
+; and finally extracts that element from the wide vector to return it.  These
+; checks ensure that z5 is always the value that is returned.
+
+;
+; Insert into two element tuples
+;
+
+; tuple:      { tuple2.res0, tuple2.res1 }
+; insert z5:  {     z5     , tuple2.res1 }
+; extract z5:       ^^
+define <vscale x 4 x i32> @set_tuple2_nxv8i32_elt0(<vscale x 4 x i32> %z0, <vscale x 4 x i32> %z1,
+                                                   <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z3,
+                                                   <vscale x 4 x i32> %z4, <vscale x 4 x i32> %z5) #0 {
+  ; CHECK-LABEL: set_tuple2_nxv8i32_elt0:
+  ; CHECK-NEXT:  mov     z0.d, z5.d
+  ; CHECK-NEXT:  ret
+  %tuple = call <vscale x 8 x i32> @llvm.aarch64.sve.tuple.create2.nxv8i32.nxv4i32(<vscale x 4 x i32> %z0, <vscale x 4 x i32> %z1)
+  %ins = call <vscale x 8 x i32> @llvm.aarch64.sve.tuple.set.nxv8i32.nxv4i32(<vscale x 8 x i32> %tuple, i32 0, <vscale x 4 x i32> %z5)
+  %ext = call <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv8i32(<vscale x 8 x i32> %ins, i32 0)
+  ret <vscale x 4 x i32> %ext
+}
+
+; tuple:       { tuple2.res0, tuple2.res1 }
+; insert z5:   { tuple2.res0,     z5      }
+; extract z5:                     ^^
+define <vscale x 4 x i32> @set_tuple2_nxv8i32_elt1(<vscale x 4 x i32> %z0, <vscale x 4 x i32> %z1,
+                                                   <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z3,
+                                                   <vscale x 4 x i32> %z4, <vscale x 4 x i32> %z5) #0 {
+  ; CHECK-LABEL: set_tuple2_nxv8i32_elt1:
+  ; CHECK-NEXT:  mov     z0.d, z5.d
+  ; CHECK-NEXT:  ret
+  %tuple = call <vscale x 8 x i32> @llvm.aarch64.sve.tuple.create2.nxv8i32.nxv4i32(<vscale x 4 x i32> %z0, <vscale x 4 x i32> %z1)
+  %ins = call <vscale x 8 x i32> @llvm.aarch64.sve.tuple.set.nxv8i32.nxv4i32(<vscale x 8 x i32> %tuple, i32 1, <vscale x 4 x i32> %z5)
+  %ext = call <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv8i32(<vscale x 8 x i32> %ins, i32 1)
+  ret <vscale x 4 x i32> %ext
+}
+
+
+;
+; Insert into three element tuples
+;
+
+; tuple:       { tuple3.res0, tuple3.res1, tuple3.res2 }
+; insert z5:   {     z5     , tuple3.res0, tuple3.res2 }
+; extract z5:        ^^
+define <vscale x 4 x i32> @set_tuple3_nxv12i32_elt0(<vscale x 4 x i32> %z0, <vscale x 4 x i32> %z1,
+                                                    <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z3,
+                                                    <vscale x 4 x i32> %z4, <vscale x 4 x i32> %z5) #0 {
+  ; CHECK-LABEL: set_tuple3_nxv12i32_elt0:
+  ; CHECK-NEXT:  mov     z0.d, z5.d
+  ; CHECK-NEXT:  ret
+  %tuple = call <vscale x 12 x i32> @llvm.aarch64.sve.tuple.create3.nxv12i32.nxv4i32(<vscale x 4 x i32> %z0, <vscale x 4 x i32> %z1, <vscale x 4 x i32> %z2)
+  %ins = call <vscale x 12 x i32> @llvm.aarch64.sve.tuple.set.nxv12i32.nxv4i32(<vscale x 12 x i32> %tuple, i32 0, <vscale x 4 x i32> %z5)
+  %ext = call <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv12i32(<vscale x 12 x i32> %ins, i32 0)
+  ret <vscale x 4 x i32> %ext
+}
+
+; tuple:       { tuple3.res0, tuple3.res1, tuple3.res2 }
+; insert z5:   { tuple3.res0,     z5     , tuple3.res2 }
+; extract z5:                     ^^
+define <vscale x 4 x i32> @set_tuple3_nxv12i32_elt1(<vscale x 4 x i32> %z0, <vscale x 4 x i32> %z1,
+                                                    <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z3,
+                                                    <vscale x 4 x i32> %z4, <vscale x 4 x i32> %z5) #0 {
+  ; CHECK-LABEL: set_tuple3_nxv12i32_elt1:
+  ; CHECK-NEXT:  mov     z0.d, z5.d
+  ; CHECK-NEXT:  ret
+  %tuple = call <vscale x 12 x i32> @llvm.aarch64.sve.tuple.create3.nxv12i32.nxv4i32(<vscale x 4 x i32> %z0, <vscale x 4 x i32> %z1, <vscale x 4 x i32> %z2)
+  %ins = call <vscale x 12 x i32> @llvm.aarch64.sve.tuple.set.nxv12i32.nxv4i32(<vscale x 12 x i32> %tuple, i32 1, <vscale x 4 x i32> %z5)
+  %ext = call <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv12i32(<vscale x 12 x i32> %ins, i32 1)
+  ret <vscale x 4 x i32> %ext
+}
+
+; tuple:       { tuple3.res0, tuple3.res1, tuple3.res2 }
+; insert z5:   { tuple3.res0, tuple3.res1,     z5      }
+; extract z5:                                  ^^
+define <vscale x 4 x i32> @set_tuple3_nxv12i32_elt2(<vscale x 4 x i32> %z0, <vscale x 4 x i32> %z1,
+                                                    <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z3,
+                                                    <vscale x 4 x i32> %z4, <vscale x 4 x i32> %z5) #0 {
+  ; CHECK-LABEL: set_tuple3_nxv12i32_elt2:
+  ; CHECK-NEXT:  mov     z0.d, z5.d
+  ; CHECK-NEXT:  ret
+  %tuple = call <vscale x 12 x i32> @llvm.aarch64.sve.tuple.create3.nxv12i32.nxv4i32(<vscale x 4 x i32> %z0, <vscale x 4 x i32> %z1, <vscale x 4 x i32> %z2)
+  %ins = call <vscale x 12 x i32> @llvm.aarch64.sve.tuple.set.nxv12i32.nxv4i32(<vscale x 12 x i32> %tuple, i32 2, <vscale x 4 x i32> %z5)
+  %ext = call <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv12i32(<vscale x 12 x i32> %ins, i32 2)
+  ret <vscale x 4 x i32> %ext
+}
+
+;
+; Insert into four element tuples
+;
+
+; tuple:       { tuple4.res0, tuple4.res1, tuple4.res2, tuple4.res3 }
+; insert z5:   {     z5     , tuple4.res1, tuple4.res2, tuple4.res3 }
+; extract z5:        ^^
+define <vscale x 4 x i32> @set_tuple4_nxv16i32_elt0(<vscale x 4 x i32> %z0, <vscale x 4 x i32> %z1,
+                                                    <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z3,
+                                                    <vscale x 4 x i32> %z4, <vscale x 4 x i32> %z5) #0 {
+  ; CHECK-LABEL: set_tuple4_nxv16i32_elt0:
+  ; CHECK-NEXT:  mov     z0.d, z5.d
+  ; CHECK-NEXT:  ret
+  %tuple = tail call <vscale x 16 x i32> @llvm.aarch64.sve.tuple.create4.nxv16i32.nxv4i32(<vscale x 4 x i32> %z0, <vscale x 4 x i32> %z1, <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z3)
+  %ins = call <vscale x 16 x i32> @llvm.aarch64.sve.tuple.set.nxv16i32.nxv4i32(<vscale x 16 x i32> %tuple, i32 0, <vscale x 4 x i32> %z5)
+  %ext = call <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv16i32(<vscale x 16 x i32> %ins, i32 0)
+  ret <vscale x 4 x i32> %ext
+}
+
+; tuple:       { tuple4.res0, tuple4.res1, tuple4.res2, tuple4.res3 }
+; insert z5:   { tuple4.res0,     z5     , tuple4.res2, tuple4.res3 }
+; extract z5:                     ^^
+define <vscale x 4 x i32> @set_tuple4_nxv16i32_elt1(<vscale x 4 x i32> %z0, <vscale x 4 x i32> %z1,
+                                                    <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z3,
+                                                    <vscale x 4 x i32> %z4, <vscale x 4 x i32> %z5) #0 {
+  ; CHECK-LABEL: set_tuple4_nxv16i32_elt1:
+  ; CHECK-NEXT:  mov     z0.d, z5.d
+  ; CHECK-NEXT:  ret
+  %tuple = tail call <vscale x 16 x i32> @llvm.aarch64.sve.tuple.create4.nxv16i32.nxv4i32(<vscale x 4 x i32> %z0, <vscale x 4 x i32> %z1, <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z3)
+  %ins = call <vscale x 16 x i32> @llvm.aarch64.sve.tuple.set.nxv16i32.nxv4i32(<vscale x 16 x i32> %tuple, i32 1, <vscale x 4 x i32> %z5)
+  %ext = call <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv16i32(<vscale x 16 x i32> %ins, i32 1)
+  ret <vscale x 4 x i32> %ext
+}
+
+; tuple:       { tuple4.res0, tuple4.res1, tuple4.res2, tuple4.res3 }
+; insert z5:   { tuple4.res0, tuple4.res1,     z5     , tuple4.res3 }
+; extract z5:                                  ^^
+define <vscale x 4 x i32> @set_tuple4_nxv16i32_elt2(<vscale x 4 x i32> %z0, <vscale x 4 x i32> %z1,
+                                                    <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z3,
+                                                    <vscale x 4 x i32> %z4, <vscale x 4 x i32> %z5) #0 {
+  ; CHECK-LABEL: set_tuple4_nxv16i32_elt2:
+  ; CHECK-NEXT:  mov     z0.d, z5.d
+  ; CHECK-NEXT:  ret
+  %tuple = tail call <vscale x 16 x i32> @llvm.aarch64.sve.tuple.create4.nxv16i32.nxv4i32(<vscale x 4 x i32> %z0, <vscale x 4 x i32> %z1, <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z3)
+  %ins = call <vscale x 16 x i32> @llvm.aarch64.sve.tuple.set.nxv16i32.nxv4i32(<vscale x 16 x i32> %tuple, i32 2, <vscale x 4 x i32> %z5)
+  %ext = call <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv16i32(<vscale x 16 x i32> %ins, i32 2)
+  ret <vscale x 4 x i32> %ext
+}
+
+; tuple:       { tuple4.res0, tuple4.res1, tuple4.res2, tuple4.res3 }
+; insert z5:   { tuple4.res0, tuple4.res1, tuple4.res2,     z5      }
+; extract z5:                                               ^^
+define <vscale x 4 x i32> @set_tuple4_nxv16i32_elt3(<vscale x 4 x i32> %z0, <vscale x 4 x i32> %z1,
+                                                    <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z3,
+                                                    <vscale x 4 x i32> %z4, <vscale x 4 x i32> %z5) #0 {
+  ; CHECK-LABEL: set_tuple4_nxv16i32_elt3:
+  ; CHECK-NEXT:  mov     z0.d, z5.d
+  ; CHECK-NEXT:  ret
+  %tuple = tail call <vscale x 16 x i32> @llvm.aarch64.sve.tuple.create4.nxv16i32.nxv4i32(<vscale x 4 x i32> %z0, <vscale x 4 x i32> %z1, <vscale x 4 x i32> %z2, <vscale x 4 x i32> %z3)
+  %ins = call <vscale x 16 x i32> @llvm.aarch64.sve.tuple.set.nxv16i32.nxv4i32(<vscale x 16 x i32> %tuple, i32 3, <vscale x 4 x i32> %z5)
+  %ext = call <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv16i32(<vscale x 16 x i32> %ins, i32 3)
+  ret <vscale x 4 x i32> %ext
+}
+
+attributes #0 = { nounwind "target-features"="+sve" }
+
+declare <vscale x 8 x i32>  @llvm.aarch64.sve.tuple.create2.nxv8i32.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 8 x i32> @llvm.aarch64.sve.tuple.set.nxv8i32.nxv4i32(<vscale x 8 x i32>, i32, <vscale x 4 x i32>)
+declare <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv8i32(<vscale x 8 x i32>, i32)
+
+declare <vscale x 12 x i32> @llvm.aarch64.sve.tuple.create3.nxv12i32.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 12 x i32> @llvm.aarch64.sve.tuple.set.nxv12i32.nxv4i32(<vscale x 12 x i32>, i32, <vscale x 4 x i32>)
+declare <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv12i32(<vscale x 12 x i32>, i32)
+
+declare <vscale x 16 x i32> @llvm.aarch64.sve.tuple.create4.nxv16i32.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>, <vscale x 4 x i32>)
+declare <vscale x 16 x i32> @llvm.aarch64.sve.tuple.set.nxv16i32.nxv4i32(<vscale x 16 x i32>, i32, <vscale x 4 x i32>)
+declare <vscale x 4 x i32> @llvm.aarch64.sve.tuple.get.nxv16i32(<vscale x 16 x i32>, i32)