[AArch64][SVE] Add the SVE dupq_lane intrinsic

[kmclaughlin]

Summary

Implements the @llvm.aarch64.sve.dupq.lane intrinsic.

As specified in the ACLE, the behaviour of:

svdupq_lane_u64(data, index)

...is identical to:

svtbl(data, svadd_x(svptrue_b64(),
                                 svand_x(svptrue_b64(), svindex_u64(0, 1), 1),
                                 index * 2))

If the index is in the range [0,3], the operation is equivalent
to a single DUP (.q) instruction.

[sdesmalen]

LGTM!

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

7495

nit: auto *CIdx

7496

nit: can you replace auto in these cases with SDValue? (which I think this is?)

7506

nit: please move Zero down to its use below.

[kmclaughlin]

Thanks for taking a look at this, @sdesmalen!

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

7496

Replaced other cases of auto here with SDValue or SDNode

llvm/include/llvm/IR/IntrinsicsAArch64.td

let TargetPrefix = "aarch64" in {  // All intrinsics start with "llvm.aarch64.".
 
   class AdvSIMD_SVE_DUP_Intrinsic
     : Intrinsic<[llvm_anyvector_ty],
                 [LLVMMatchType<0>,
                  LLVMScalarOrSameVectorWidth<0, llvm_i1_ty>,
                  LLVMVectorElementType<0>],
                 [IntrNoMem]>;
 
+  class AdvSIMD_SVE_DUPQ_Intrinsic
+    : Intrinsic<[llvm_anyvector_ty],
+                [LLVMMatchType<0>,
+                 llvm_i64_ty],
+                [IntrNoMem]>;
+
   class AdvSIMD_SVE_EXPA_Intrinsic
     : Intrinsic<[llvm_anyvector_ty],
                 [LLVMVectorOfBitcastsToInt<0>],
                 [IntrNoMem]>;
 
   class AdvSIMD_SVE_FCVT_Intrinsic
     : Intrinsic<[llvm_anyvector_ty],
                 [LLVMMatchType<0>,

 // Permutations and selection
 //
 
 def int_aarch64_sve_clasta    : AdvSIMD_Pred2VectorArg_Intrinsic;
 def int_aarch64_sve_clasta_n  : AdvSIMD_SVE_ReduceWithInit_Intrinsic;
 def int_aarch64_sve_clastb    : AdvSIMD_Pred2VectorArg_Intrinsic;
 def int_aarch64_sve_clastb_n  : AdvSIMD_SVE_ReduceWithInit_Intrinsic;
 def int_aarch64_sve_compact   : AdvSIMD_Pred1VectorArg_Intrinsic;
+def int_aarch64_sve_dupq_lane : AdvSIMD_SVE_DUPQ_Intrinsic;
 def int_aarch64_sve_ext       : AdvSIMD_2VectorArgIndexed_Intrinsic;
 def int_aarch64_sve_lasta     : AdvSIMD_SVE_Reduce_Intrinsic;
 def int_aarch64_sve_lastb     : AdvSIMD_SVE_Reduce_Intrinsic;
 def int_aarch64_sve_rev       : AdvSIMD_1VectorArg_Intrinsic;
 def int_aarch64_sve_splice    : AdvSIMD_Pred2VectorArg_Intrinsic;
 def int_aarch64_sve_sunpkhi   : AdvSIMD_SVE_Unpack_Intrinsic;
 def int_aarch64_sve_sunpklo   : AdvSIMD_SVE_Unpack_Intrinsic;
 def int_aarch64_sve_tbl       : AdvSIMD_SVE_TBL_Intrinsic;

llvm/lib/Target/AArch64/AArch64ISelLowering.h

private:
   SDValue LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerFLT_ROUNDS_(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerSPLAT_VECTOR(SDValue Op, SelectionDAG &DAG) const;
+  SDValue LowerDUPQLane(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerVectorSRA_SRL_SHL(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerShiftLeftParts(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerShiftRightParts(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerCTPOP(SDValue Op, SelectionDAG &DAG) const;
   SDValue LowerF128Call(SDValue Op, SelectionDAG &DAG,
                         RTLIB::Libcall Call) const;

llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

case Intrinsic::aarch64_sve_zip1:
     return DAG.getNode(AArch64ISD::ZIP1, dl, Op.getValueType(),
                        Op.getOperand(1), Op.getOperand(2));
   case Intrinsic::aarch64_sve_zip2:
     return DAG.getNode(AArch64ISD::ZIP2, dl, Op.getValueType(),
                        Op.getOperand(1), Op.getOperand(2));
   case Intrinsic::aarch64_sve_ptrue:
     return DAG.getNode(AArch64ISD::PTRUE, dl, Op.getValueType(),
                        Op.getOperand(1));
+  case Intrinsic::aarch64_sve_dupq_lane:
+    return LowerDUPQLane(Op, DAG);
 
   case Intrinsic::aarch64_sve_insr: {
     SDValue Scalar = Op.getOperand(2);
     EVT ScalarTy = Scalar.getValueType();
     if ((ScalarTy == MVT::i8) || (ScalarTy == MVT::i16))
       Scalar = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Scalar);
 
     return DAG.getNode(AArch64ISD::INSR, dl, Op.getValueType(),
SDValue AArch64TargetLowering::LowerSPLAT_VECTOR(SDValue Op,
   case MVT::f16:
   case MVT::f32:
   case MVT::f64:
   default:
     report_fatal_error("Unsupported SPLAT_VECTOR input operand type");
   }
 }
 
+SDValue AArch64TargetLowering::LowerDUPQLane(SDValue Op,
+                                             SelectionDAG &DAG) const {
+  SDLoc DL(Op);
+
+  EVT VT = Op.getValueType();
+  if (!isTypeLegal(VT) || !VT.isScalableVector())
+    return SDValue();
+
+  // Current lowering only supports the SVE-ACLE types.
+  if (VT.getSizeInBits().getKnownMinSize() != AArch64::SVEBitsPerBlock)
+    return SDValue();
+
+  // The DUPQ operation is indepedent of element type so normalise to i64s.
+  auto V = DAG.getNode(ISD::BITCAST, DL, MVT::nxv2i64, Op.getOperand(1));
+  auto Idx128 = Op.getOperand(2);
+
+  // DUPQ can be used when idx is in range.
+  auto CIdx = dyn_cast<ConstantSDNode>(Idx128);
+  if (CIdx && (CIdx->getZExtValue() <= 3)) {
+    auto CI = DAG.getTargetConstant(CIdx->getZExtValue(), DL, MVT::i64);
+    auto DUPQ = DAG.getMachineNode(AArch64::DUP_ZZI_Q, DL, MVT::nxv2i64, V, CI);
+    return DAG.getNode(ISD::BITCAST, DL, VT, SDValue(DUPQ, 0));
+  }
+
+  // The ACLE says this must produce the same result as:
+  //   svtbl(data, svadd_x(svptrue_b64(),
+  //                       svand_x(svptrue_b64(), svindex_u64(0, 1), 1),
+  //                       index * 2))
+  auto Zero = DAG.getConstant(0, DL, MVT::i64);
+  auto One = DAG.getConstant(1, DL, MVT::i64);
+  auto SplatOne = DAG.getNode(ISD::SPLAT_VECTOR, DL, MVT::nxv2i64, One);
+
+  // create the vector 0,1,0,1,...
+  auto SV = DAG.getNode(AArch64ISD::INDEX_VECTOR, DL, MVT::nxv2i64, Zero, One);
+  SV = DAG.getNode(ISD::AND, DL, MVT::nxv2i64, SV, SplatOne);
+
+  // create the vector idx64,idx64+1,idx64,idx64+1,...
+  auto Idx64 = DAG.getNode(ISD::ADD, DL, MVT::i64, Idx128, Idx128);
+  auto SplatIdx64 = DAG.getNode(ISD::SPLAT_VECTOR, DL, MVT::nxv2i64, Idx64);
+  auto ShuffleMask = DAG.getNode(ISD::ADD, DL, MVT::nxv2i64, SV, SplatIdx64);
+
+  // create the vector Val[idx64],Val[idx64+1],Val[idx64],Val[idx64+1],...
+  auto TBL = DAG.getNode(AArch64ISD::TBL, DL, MVT::nxv2i64, V, ShuffleMask);
+  return DAG.getNode(ISD::BITCAST, DL, VT, TBL);
+}
+
 static bool resolveBuildVector(BuildVectorSDNode *BVN, APInt &CnstBits,
                                APInt &UndefBits) {
   EVT VT = BVN->getValueType(0);
   APInt SplatBits, SplatUndef;
   unsigned SplatBitSize;
   bool HasAnyUndefs;
   if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
     unsigned NumSplats = VT.getSizeInBits() / SplatBitSize;

llvm/test/CodeGen/AArch64/sve-intrinsics-perm-select.ll

 ; CHECK-NEXT: ret
   %out = call double @llvm.aarch64.sve.clastb.n.nxv2f64(<vscale x 2 x i1> %pg,
                                                         double %a,
                                                         <vscale x 2 x double> %b)
   ret double %out
 }
 
 ;
+; DUPQ
+;
+
+define <vscale x 16 x i8> @dupq_i8(<vscale x 16 x i8> %a) {
+; CHECK-LABEL: dupq_i8:
+; CHECK: mov z0.q, q0
+; CHECK-NEXT: ret
+  %out = call <vscale x 16 x i8> @llvm.aarch64.sve.dupq.lane.nxv16i8(<vscale x 16 x i8> %a, i64 0)
+  ret <vscale x 16 x i8> %out
+}
+
+define <vscale x 8 x i16> @dupq_i16(<vscale x 8 x i16> %a) {
+; CHECK-LABEL: dupq_i16:
+; CHECK: mov z0.q, z0.q[1]
+; CHECK-NEXT: ret
+  %out = call <vscale x 8 x i16> @llvm.aarch64.sve.dupq.lane.nxv8i16(<vscale x 8 x i16> %a, i64 1)
+  ret <vscale x 8 x i16> %out
+}
+
+define <vscale x 4 x i32> @dupq_i32(<vscale x 4 x i32> %a) {
+; CHECK-LABEL: dupq_i32:
+; CHECK: mov z0.q, z0.q[2]
+; CHECK-NEXT: ret
+  %out = call <vscale x 4 x i32> @llvm.aarch64.sve.dupq.lane.nxv4i32(<vscale x 4 x i32> %a, i64 2)
+  ret <vscale x 4 x i32> %out
+}
+
+define <vscale x 2 x i64> @dupq_i64(<vscale x 2 x i64> %a) {
+; CHECK-LABEL: dupq_i64:
+; CHECK: mov z0.q, z0.q[3]
+; CHECK-NEXT: ret
+  %out = call <vscale x 2 x i64> @llvm.aarch64.sve.dupq.lane.nxv2i64(<vscale x 2 x i64> %a, i64 3)
+  ret <vscale x 2 x i64> %out
+}
+
+define <vscale x 8 x half> @dupq_f16(<vscale x 8 x half> %a) {
+; CHECK-LABEL: dupq_f16:
+; CHECK: mov z0.q, q0
+; CHECK-NEXT: ret
+  %out = call <vscale x 8 x half> @llvm.aarch64.sve.dupq.lane.nxv8f16(<vscale x 8 x half> %a, i64 0)
+  ret <vscale x 8 x half> %out
+}
+
+define <vscale x 4 x float> @dupq_f32(<vscale x 4 x float> %a) {
+; CHECK-LABEL: dupq_f32:
+; CHECK: mov z0.q, z0.q[1]
+; CHECK-NEXT: ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.dupq.lane.nxv4f32(<vscale x 4 x float> %a, i64 1)
+  ret <vscale x 4 x float> %out
+}
+
+define <vscale x 2 x double> @dupq_f64(<vscale x 2 x double> %a) {
+; CHECK-LABEL: dupq_f64:
+; CHECK: mov z0.q, z0.q[2]
+; CHECK-NEXT: ret
+  %out = call <vscale x 2 x double> @llvm.aarch64.sve.dupq.lane.nxv2f64(<vscale x 2 x double> %a, i64 2)
+  ret <vscale x 2 x double> %out
+}
+
+;
+; DUPQ_LANE
+;
+
+define <vscale x 16 x i8> @dupq_lane_i8(<vscale x 16 x i8> %a, i64 %idx) {
+; CHECK-LABEL: dupq_lane_i8:
+; CHECK-DAG:  index [[Z1:z[0-9]+]].d, #0, #1
+; CHECK-DAG:  and   [[Z2:z[0-9]+]].d, [[Z1]].d, #0x1
+; CHECK-DAG:  add   [[X1:x[0-9]+]], x0, x0
+; CHECK-DAG:  mov   [[Z3:z[0-9]+]].d, [[X1]]
+; CHECK:      add   [[Z4:z[0-9]+]].d, [[Z2]].d, [[Z3]].d
+; CHECK-NEXT: tbl   z0.d, { z0.d }, [[Z4]].d
+; CHECK-NEXT: ret
+  %out = call <vscale x 16 x i8> @llvm.aarch64.sve.dupq.lane.nxv16i8(<vscale x 16 x i8> %a, i64 %idx)
+  ret <vscale x 16 x i8> %out
+}
+
+; NOTE: Identical operation to dupq_lane_i8 (i.e. element type is irrelevant).
+define <vscale x 8 x i16> @dupq_lane_i16(<vscale x 8 x i16> %a, i64 %idx) {
+; CHECK-LABEL: dupq_lane_i16:
+; CHECK-DAG:  index [[Z1:z[0-9]+]].d, #0, #1
+; CHECK-DAG:  and   [[Z2:z[0-9]+]].d, [[Z1]].d, #0x1
+; CHECK-DAG:  add   [[X1:x[0-9]+]], x0, x0
+; CHECK-DAG:  mov   [[Z3:z[0-9]+]].d, [[X1]]
+; CHECK:      add   [[Z4:z[0-9]+]].d, [[Z2]].d, [[Z3]].d
+; CHECK: tbl z0.d, { z0.d }, [[Z4]].d
+; CHECK-NEXT: ret
+  %out = call <vscale x 8 x i16> @llvm.aarch64.sve.dupq.lane.nxv8i16(<vscale x 8 x i16> %a, i64 %idx)
+  ret <vscale x 8 x i16> %out
+}
+
+; NOTE: Identical operation to dupq_lane_i8 (i.e. element type is irrelevant).
+define <vscale x 4 x i32> @dupq_lane_i32(<vscale x 4 x i32> %a, i64 %idx) {
+; CHECK-LABEL: dupq_lane_i32:
+; CHECK-DAG:  index [[Z1:z[0-9]+]].d, #0, #1
+; CHECK-DAG:  and   [[Z2:z[0-9]+]].d, [[Z1]].d, #0x1
+; CHECK-DAG:  add   [[X1:x[0-9]+]], x0, x0
+; CHECK-DAG:  mov   [[Z3:z[0-9]+]].d, [[X1]]
+; CHECK:      add   [[Z4:z[0-9]+]].d, [[Z2]].d, [[Z3]].d
+; CHECK: tbl z0.d, { z0.d }, [[Z4]].d
+; CHECK-NEXT: ret
+  %out = call <vscale x 4 x i32> @llvm.aarch64.sve.dupq.lane.nxv4i32(<vscale x 4 x i32> %a, i64 %idx)
+  ret <vscale x 4 x i32> %out
+}
+
+; NOTE: Identical operation to dupq_lane_i8 (i.e. element type is irrelevant).
+define <vscale x 2 x i64> @dupq_lane_i64(<vscale x 2 x i64> %a, i64 %idx) {
+; CHECK-LABEL: dupq_lane_i64:
+; CHECK-DAG:  index [[Z1:z[0-9]+]].d, #0, #1
+; CHECK-DAG:  and   [[Z2:z[0-9]+]].d, [[Z1]].d, #0x1
+; CHECK-DAG:  add   [[X1:x[0-9]+]], x0, x0
+; CHECK-DAG:  mov   [[Z3:z[0-9]+]].d, [[X1]]
+; CHECK:      add   [[Z4:z[0-9]+]].d, [[Z2]].d, [[Z3]].d
+; CHECK: tbl z0.d, { z0.d }, [[Z4]].d
+; CHECK-NEXT: ret
+  %out = call <vscale x 2 x i64> @llvm.aarch64.sve.dupq.lane.nxv2i64(<vscale x 2 x i64> %a, i64 %idx)
+  ret <vscale x 2 x i64> %out
+}
+
+; NOTE: Identical operation to dupq_lane_i8 (i.e. element type is irrelevant).
+define <vscale x 8 x half> @dupq_lane_f16(<vscale x 8 x half> %a, i64 %idx) {
+; CHECK-LABEL: dupq_lane_f16:
+; CHECK-DAG:  index [[Z1:z[0-9]+]].d, #0, #1
+; CHECK-DAG:  and   [[Z2:z[0-9]+]].d, [[Z1]].d, #0x1
+; CHECK-DAG:  add   [[X1:x[0-9]+]], x0, x0
+; CHECK-DAG:  mov   [[Z3:z[0-9]+]].d, [[X1]]
+; CHECK:      add   [[Z4:z[0-9]+]].d, [[Z2]].d, [[Z3]].d
+; CHECK: tbl z0.d, { z0.d }, [[Z4]].d
+; CHECK-NEXT: ret
+  %out = call <vscale x 8 x half> @llvm.aarch64.sve.dupq.lane.nxv8f16(<vscale x 8 x half> %a, i64 %idx)
+  ret <vscale x 8 x half> %out
+}
+
+; NOTE: Identical operation to dupq_lane_i8 (i.e. element type is irrelevant).
+define <vscale x 4 x float> @dupq_lane_f32(<vscale x 4 x float> %a, i64 %idx) {
+; CHECK-LABEL: dupq_lane_f32:
+; CHECK-DAG:  index [[Z1:z[0-9]+]].d, #0, #1
+; CHECK-DAG:  and   [[Z2:z[0-9]+]].d, [[Z1]].d, #0x1
+; CHECK-DAG:  add   [[X1:x[0-9]+]], x0, x0
+; CHECK-DAG:  mov   [[Z3:z[0-9]+]].d, [[X1]]
+; CHECK:      add   [[Z4:z[0-9]+]].d, [[Z2]].d, [[Z3]].d
+; CHECK: tbl z0.d, { z0.d }, [[Z4]].d
+; CHECK-NEXT: ret
+  %out = call <vscale x 4 x float> @llvm.aarch64.sve.dupq.lane.nxv4f32(<vscale x 4 x float> %a, i64 %idx)
+  ret <vscale x 4 x float> %out
+}
+
+; NOTE: Identical operation to dupq_lane_i8 (i.e. element type is irrelevant).
+define <vscale x 2 x double> @dupq_lane_f64(<vscale x 2 x double> %a, i64 %idx) {
+; CHECK-LABEL: dupq_lane_f64:
+; CHECK-DAG:  index [[Z1:z[0-9]+]].d, #0, #1
+; CHECK-DAG:  and   [[Z2:z[0-9]+]].d, [[Z1]].d, #0x1
+; CHECK-DAG:  add   [[X1:x[0-9]+]], x0, x0
+; CHECK-DAG:  mov   [[Z3:z[0-9]+]].d, [[X1]]
+; CHECK:      add   [[Z4:z[0-9]+]].d, [[Z2]].d, [[Z3]].d
+; CHECK: tbl z0.d, { z0.d }, [[Z4]].d
+; CHECK-NEXT: ret
+  %out = call <vscale x 2 x double> @llvm.aarch64.sve.dupq.lane.nxv2f64(<vscale x 2 x double> %a, i64 %idx)
+  ret <vscale x 2 x double> %out
+}
+
+; NOTE: Index out of range (0-3)
+define <vscale x 2 x i64> @dupq_i64_range(<vscale x 2 x i64> %a) {
+; CHECK-LABEL: dupq_i64_range:
+; CHECK-DAG:  index [[Z1:z[0-9]+]].d, #0, #1
+; CHECK-DAG:  and   [[Z2:z[0-9]+]].d, [[Z1]].d, #0x1
+; CHECK-DAG:  add   [[Z3:z[0-9]+]].d, [[Z2]].d, #8
+; CHECK: tbl z0.d, { z0.d }, [[Z3]].d
+; CHECK-NEXT: ret
+  %out = call <vscale x 2 x i64> @llvm.aarch64.sve.dupq.lane.nxv2i64(<vscale x 2 x i64> %a, i64 4)
+  ret <vscale x 2 x i64> %out
+}
+
+;
 ; EXT
 ;
 
 define <vscale x 16 x i8> @ext_i8(<vscale x 16 x i8> %a, <vscale x 16 x i8> %b) {
 ; CHECK-LABEL: ext_i8:
 ; CHECK: ext z0.b, z0.b, z1.b, #255
 ; CHECK-NEXT: ret
   %out = call <vscale x 16 x i8> @llvm.aarch64.sve.ext.nxv16i8(<vscale x 16 x i8> %a,

 declare float @llvm.aarch64.sve.clastb.n.nxv4f32(<vscale x 4 x i1>, float, <vscale x 4 x float>)
 declare double @llvm.aarch64.sve.clastb.n.nxv2f64(<vscale x 2 x i1>, double, <vscale x 2 x double>)
 
 declare <vscale x 4 x i32> @llvm.aarch64.sve.compact.nxv4i32(<vscale x 4 x i1>, <vscale x 4 x i32>)
 declare <vscale x 2 x i64> @llvm.aarch64.sve.compact.nxv2i64(<vscale x 2 x i1>, <vscale x 2 x i64>)
 declare <vscale x 4 x float> @llvm.aarch64.sve.compact.nxv4f32(<vscale x 4 x i1>, <vscale x 4 x float>)
 declare <vscale x 2 x double> @llvm.aarch64.sve.compact.nxv2f64(<vscale x 2 x i1>, <vscale x 2 x double>)
 
+declare <vscale x 16 x i8> @llvm.aarch64.sve.dupq.lane.nxv16i8(<vscale x 16 x i8>, i64)
+declare <vscale x 8 x i16> @llvm.aarch64.sve.dupq.lane.nxv8i16(<vscale x 8 x i16>, i64)
+declare <vscale x 4 x i32> @llvm.aarch64.sve.dupq.lane.nxv4i32(<vscale x 4 x i32>, i64)
+declare <vscale x 2 x i64> @llvm.aarch64.sve.dupq.lane.nxv2i64(<vscale x 2 x i64>, i64)
+declare <vscale x 8 x half> @llvm.aarch64.sve.dupq.lane.nxv8f16(<vscale x 8 x half>, i64)
+declare <vscale x 4 x float> @llvm.aarch64.sve.dupq.lane.nxv4f32(<vscale x 4 x float>, i64)
+declare <vscale x 2 x double> @llvm.aarch64.sve.dupq.lane.nxv2f64(<vscale x 2 x double>, i64)
+
 declare <vscale x 16 x i8> @llvm.aarch64.sve.ext.nxv16i8(<vscale x 16 x i8>, <vscale x 16 x i8>, i32)
 declare <vscale x 8 x i16> @llvm.aarch64.sve.ext.nxv8i16(<vscale x 8 x i16>, <vscale x 8 x i16>, i32)
 declare <vscale x 4 x i32> @llvm.aarch64.sve.ext.nxv4i32(<vscale x 4 x i32>, <vscale x 4 x i32>, i32)
 declare <vscale x 2 x i64> @llvm.aarch64.sve.ext.nxv2i64(<vscale x 2 x i64>, <vscale x 2 x i64>, i32)
 declare <vscale x 8 x half> @llvm.aarch64.sve.ext.nxv8f16(<vscale x 8 x half>, <vscale x 8 x half>, i32)
 declare <vscale x 4 x float> @llvm.aarch64.sve.ext.nxv4f32(<vscale x 4 x float>, <vscale x 4 x float>, i32)
 declare <vscale x 2 x double> @llvm.aarch64.sve.ext.nxv2f64(<vscale x 2 x double>, <vscale x 2 x double>, i32)