xref: /freebsd-9.3-release/contrib/llvm/lib/Target/AArch64/AArch64ISelDAGToDAG.cpp

//===-- AArch64ISelDAGToDAG.cpp - A dag to dag inst selector for AArch64 --===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file defines an instruction selector for the AArch64 target.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "aarch64-isel"
#include "AArch64.h"
#include "AArch64InstrInfo.h"
#include "AArch64Subtarget.h"
#include "AArch64TargetMachine.h"
#include "Utils/AArch64BaseInfo.h"
#include "llvm/ADT/APSInt.h"
#include "llvm/CodeGen/SelectionDAGISel.h"
#include "llvm/IR/GlobalValue.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/raw_ostream.h"

using namespace llvm;

//===--------------------------------------------------------------------===//
/// AArch64 specific code to select AArch64 machine instructions for
/// SelectionDAG operations.
///
namespace {

class AArch64DAGToDAGISel : public SelectionDAGISel {
  AArch64TargetMachine &TM;

  /// Keep a pointer to the AArch64Subtarget around so that we can
  /// make the right decision when generating code for different targets.
  const AArch64Subtarget *Subtarget;

public:
  explicit AArch64DAGToDAGISel(AArch64TargetMachine &tm,
                               CodeGenOpt::Level OptLevel)
    : SelectionDAGISel(tm, OptLevel), TM(tm),
      Subtarget(&TM.getSubtarget<AArch64Subtarget>()) {
  }

  virtual const char *getPassName() const {
    return "AArch64 Instruction Selection";
  }

  // Include the pieces autogenerated from the target description.
#include "AArch64GenDAGISel.inc"

  template<unsigned MemSize>
  bool SelectOffsetUImm12(SDValue N, SDValue &UImm12) {
    const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N);
    if (!CN || CN->getZExtValue() % MemSize != 0
        || CN->getZExtValue() / MemSize > 0xfff)
      return false;

    UImm12 =  CurDAG->getTargetConstant(CN->getZExtValue() / MemSize, MVT::i64);
    return true;
  }

  template<unsigned RegWidth>
  bool SelectCVTFixedPosOperand(SDValue N, SDValue &FixedPos) {
    return SelectCVTFixedPosOperand(N, FixedPos, RegWidth);
  }

  /// Used for pre-lowered address-reference nodes, so we already know
  /// the fields match. This operand's job is simply to add an
  /// appropriate shift operand to the MOVZ/MOVK instruction.
  template<unsigned LogShift>
  bool SelectMOVWAddressRef(SDValue N, SDValue &Imm, SDValue &Shift) {
    Imm = N;
    Shift = CurDAG->getTargetConstant(LogShift, MVT::i32);
    return true;
  }

  bool SelectFPZeroOperand(SDValue N, SDValue &Dummy);

  bool SelectCVTFixedPosOperand(SDValue N, SDValue &FixedPos,
                                unsigned RegWidth);

  bool SelectInlineAsmMemoryOperand(const SDValue &Op,
                                    char ConstraintCode,
                                    std::vector<SDValue> &OutOps);

  bool SelectLogicalImm(SDValue N, SDValue &Imm);

  template<unsigned RegWidth>
  bool SelectTSTBOperand(SDValue N, SDValue &FixedPos) {
    return SelectTSTBOperand(N, FixedPos, RegWidth);
  }

  bool SelectTSTBOperand(SDValue N, SDValue &FixedPos, unsigned RegWidth);

  SDNode *SelectAtomic(SDNode *N, unsigned Op8, unsigned Op16, unsigned Op32,
                       unsigned Op64);

  /// Put the given constant into a pool and return a DAG which will give its
  /// address.
  SDValue getConstantPoolItemAddress(SDLoc DL, const Constant *CV);

  SDNode *TrySelectToMoveImm(SDNode *N);
  SDNode *LowerToFPLitPool(SDNode *Node);
  SDNode *SelectToLitPool(SDNode *N);

  SDNode* Select(SDNode*);
private:
  /// Get the opcode for table lookup instruction
  unsigned getTBLOpc(bool IsExt, bool Is64Bit, unsigned NumOfVec);

  /// Select NEON table lookup intrinsics.  NumVecs should be 1, 2, 3 or 4.
  /// IsExt is to indicate if the result will be extended with an argument.
  SDNode *SelectVTBL(SDNode *N, unsigned NumVecs, bool IsExt);

  /// Select NEON load intrinsics.  NumVecs should be 1, 2, 3 or 4.
  SDNode *SelectVLD(SDNode *N, bool isUpdating, unsigned NumVecs,
                    const uint16_t *Opcode);

  /// Select NEON store intrinsics.  NumVecs should be 1, 2, 3 or 4.
  SDNode *SelectVST(SDNode *N, bool isUpdating, unsigned NumVecs,
                    const uint16_t *Opcodes);

  /// Form sequences of consecutive 64/128-bit registers for use in NEON
  /// instructions making use of a vector-list (e.g. ldN, tbl). Vecs must have
  /// between 1 and 4 elements. If it contains a single element that is returned
  /// unchanged; otherwise a REG_SEQUENCE value is returned.
  SDValue createDTuple(ArrayRef<SDValue> Vecs);
  SDValue createQTuple(ArrayRef<SDValue> Vecs);

  /// Generic helper for the createDTuple/createQTuple
  /// functions. Those should almost always be called instead.
  SDValue createTuple(ArrayRef<SDValue> Vecs, unsigned RegClassIDs[],
                      unsigned SubRegs[]);

  /// Select NEON load-duplicate intrinsics.  NumVecs should be 2, 3 or 4.
  /// The opcode array specifies the instructions used for load.
  SDNode *SelectVLDDup(SDNode *N, bool isUpdating, unsigned NumVecs,
                       const uint16_t *Opcodes);

  /// Select NEON load/store lane intrinsics.  NumVecs should be 2, 3 or 4.
  /// The opcode arrays specify the instructions used for load/store.
  SDNode *SelectVLDSTLane(SDNode *N, bool IsLoad, bool isUpdating,
                          unsigned NumVecs, const uint16_t *Opcodes);

  SDValue getTargetSubregToReg(int SRIdx, SDLoc DL, EVT VT, EVT VTD,
                               SDValue Operand);
};
}

bool
AArch64DAGToDAGISel::SelectCVTFixedPosOperand(SDValue N, SDValue &FixedPos,
                                              unsigned RegWidth) {
  const ConstantFPSDNode *CN = dyn_cast<ConstantFPSDNode>(N);
  if (!CN) return false;

  // An FCVT[SU] instruction performs: convertToInt(Val * 2^fbits) where fbits
  // is between 1 and 32 for a destination w-register, or 1 and 64 for an
  // x-register.
  //
  // By this stage, we've detected (fp_to_[su]int (fmul Val, THIS_NODE)) so we
  // want THIS_NODE to be 2^fbits. This is much easier to deal with using
  // integers.
  bool IsExact;

  // fbits is between 1 and 64 in the worst-case, which means the fmul
  // could have 2^64 as an actual operand. Need 65 bits of precision.
  APSInt IntVal(65, true);
  CN->getValueAPF().convertToInteger(IntVal, APFloat::rmTowardZero, &IsExact);

  // N.b. isPowerOf2 also checks for > 0.
  if (!IsExact || !IntVal.isPowerOf2()) return false;
  unsigned FBits = IntVal.logBase2();

  // Checks above should have guaranteed that we haven't lost information in
  // finding FBits, but it must still be in range.
  if (FBits == 0 || FBits > RegWidth) return false;

  FixedPos = CurDAG->getTargetConstant(64 - FBits, MVT::i32);
  return true;
}

bool
AArch64DAGToDAGISel::SelectInlineAsmMemoryOperand(const SDValue &Op,
                                                 char ConstraintCode,
                                                 std::vector<SDValue> &OutOps) {
  switch (ConstraintCode) {
  default: llvm_unreachable("Unrecognised AArch64 memory constraint");
  case 'm':
    // FIXME: more freedom is actually permitted for 'm'. We can go
    // hunting for a base and an offset if we want. Of course, since
    // we don't really know how the operand is going to be used we're
    // probably restricted to the load/store pair's simm7 as an offset
    // range anyway.
  case 'Q':
    OutOps.push_back(Op);
  }

  return false;
}

bool
AArch64DAGToDAGISel::SelectFPZeroOperand(SDValue N, SDValue &Dummy) {
  ConstantFPSDNode *Imm = dyn_cast<ConstantFPSDNode>(N);
  if (!Imm || !Imm->getValueAPF().isPosZero())
    return false;

  // Doesn't actually carry any information, but keeps TableGen quiet.
  Dummy = CurDAG->getTargetConstant(0, MVT::i32);
  return true;
}

bool AArch64DAGToDAGISel::SelectLogicalImm(SDValue N, SDValue &Imm) {
  uint32_t Bits;
  uint32_t RegWidth = N.getValueType().getSizeInBits();

  ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N);
  if (!CN) return false;

  if (!A64Imms::isLogicalImm(RegWidth, CN->getZExtValue(), Bits))
    return false;

  Imm = CurDAG->getTargetConstant(Bits, MVT::i32);
  return true;
}

SDNode *AArch64DAGToDAGISel::TrySelectToMoveImm(SDNode *Node) {
  SDNode *ResNode;
  SDLoc dl(Node);
  EVT DestType = Node->getValueType(0);
  unsigned DestWidth = DestType.getSizeInBits();

  unsigned MOVOpcode;
  EVT MOVType;
  int UImm16, Shift;
  uint32_t LogicalBits;

  uint64_t BitPat = cast<ConstantSDNode>(Node)->getZExtValue();
  if (A64Imms::isMOVZImm(DestWidth, BitPat, UImm16, Shift)) {
    MOVType = DestType;
    MOVOpcode = DestWidth == 64 ? AArch64::MOVZxii : AArch64::MOVZwii;
  } else if (A64Imms::isMOVNImm(DestWidth, BitPat, UImm16, Shift)) {
    MOVType = DestType;
    MOVOpcode = DestWidth == 64 ? AArch64::MOVNxii : AArch64::MOVNwii;
  } else if (DestWidth == 64 && A64Imms::isMOVNImm(32, BitPat, UImm16, Shift)) {
    // To get something like 0x0000_0000_ffff_1234 into a 64-bit register we can
    // use a 32-bit instruction: "movn w0, 0xedbc".
    MOVType = MVT::i32;
    MOVOpcode = AArch64::MOVNwii;
  } else if (A64Imms::isLogicalImm(DestWidth, BitPat, LogicalBits))  {
    MOVOpcode = DestWidth == 64 ? AArch64::ORRxxi : AArch64::ORRwwi;
    uint16_t ZR = DestWidth == 64 ? AArch64::XZR : AArch64::WZR;

    return CurDAG->getMachineNode(MOVOpcode, dl, DestType,
                              CurDAG->getRegister(ZR, DestType),
                              CurDAG->getTargetConstant(LogicalBits, MVT::i32));
  } else {
    // Can't handle it in one instruction. There's scope for permitting two (or
    // more) instructions, but that'll need more thought.
    return NULL;
  }

  ResNode = CurDAG->getMachineNode(MOVOpcode, dl, MOVType,
                                   CurDAG->getTargetConstant(UImm16, MVT::i32),
                                   CurDAG->getTargetConstant(Shift, MVT::i32));

  if (MOVType != DestType) {
    ResNode = CurDAG->getMachineNode(TargetOpcode::SUBREG_TO_REG, dl,
                          MVT::i64, MVT::i32, MVT::Other,
                          CurDAG->getTargetConstant(0, MVT::i64),
                          SDValue(ResNode, 0),
                          CurDAG->getTargetConstant(AArch64::sub_32, MVT::i32));
  }

  return ResNode;
}

SDValue
AArch64DAGToDAGISel::getConstantPoolItemAddress(SDLoc DL,
                                                const Constant *CV) {
  EVT PtrVT = getTargetLowering()->getPointerTy();

  switch (getTargetLowering()->getTargetMachine().getCodeModel()) {
  case CodeModel::Small: {
    unsigned Alignment =
      getTargetLowering()->getDataLayout()->getABITypeAlignment(CV->getType());
    return CurDAG->getNode(
        AArch64ISD::WrapperSmall, DL, PtrVT,
        CurDAG->getTargetConstantPool(CV, PtrVT, 0, 0, AArch64II::MO_NO_FLAG),
        CurDAG->getTargetConstantPool(CV, PtrVT, 0, 0, AArch64II::MO_LO12),
        CurDAG->getConstant(Alignment, MVT::i32));
  }
  case CodeModel::Large: {
    SDNode *LitAddr;
    LitAddr = CurDAG->getMachineNode(
        AArch64::MOVZxii, DL, PtrVT,
        CurDAG->getTargetConstantPool(CV, PtrVT, 0, 0, AArch64II::MO_ABS_G3),
        CurDAG->getTargetConstant(3, MVT::i32));
    LitAddr = CurDAG->getMachineNode(
        AArch64::MOVKxii, DL, PtrVT, SDValue(LitAddr, 0),
        CurDAG->getTargetConstantPool(CV, PtrVT, 0, 0, AArch64II::MO_ABS_G2_NC),
        CurDAG->getTargetConstant(2, MVT::i32));
    LitAddr = CurDAG->getMachineNode(
        AArch64::MOVKxii, DL, PtrVT, SDValue(LitAddr, 0),
        CurDAG->getTargetConstantPool(CV, PtrVT, 0, 0, AArch64II::MO_ABS_G1_NC),
        CurDAG->getTargetConstant(1, MVT::i32));
    LitAddr = CurDAG->getMachineNode(
        AArch64::MOVKxii, DL, PtrVT, SDValue(LitAddr, 0),
        CurDAG->getTargetConstantPool(CV, PtrVT, 0, 0, AArch64II::MO_ABS_G0_NC),
        CurDAG->getTargetConstant(0, MVT::i32));
    return SDValue(LitAddr, 0);
  }
  default:
    llvm_unreachable("Only small and large code models supported now");
  }
}

SDNode *AArch64DAGToDAGISel::SelectToLitPool(SDNode *Node) {
  SDLoc DL(Node);
  uint64_t UnsignedVal = cast<ConstantSDNode>(Node)->getZExtValue();
  int64_t SignedVal = cast<ConstantSDNode>(Node)->getSExtValue();
  EVT DestType = Node->getValueType(0);

  // Since we may end up loading a 64-bit constant from a 32-bit entry the
  // constant in the pool may have a different type to the eventual node.
  ISD::LoadExtType Extension;
  EVT MemType;

  assert((DestType == MVT::i64 || DestType == MVT::i32)
         && "Only expect integer constants at the moment");

  if (DestType == MVT::i32) {
    Extension = ISD::NON_EXTLOAD;
    MemType = MVT::i32;
  } else if (UnsignedVal <= UINT32_MAX) {
    Extension = ISD::ZEXTLOAD;
    MemType = MVT::i32;
  } else if (SignedVal >= INT32_MIN && SignedVal <= INT32_MAX) {
    Extension = ISD::SEXTLOAD;
    MemType = MVT::i32;
  } else {
    Extension = ISD::NON_EXTLOAD;
    MemType = MVT::i64;
  }

  Constant *CV = ConstantInt::get(Type::getIntNTy(*CurDAG->getContext(),
                                                  MemType.getSizeInBits()),
                                  UnsignedVal);
  SDValue PoolAddr = getConstantPoolItemAddress(DL, CV);
  unsigned Alignment =
    getTargetLowering()->getDataLayout()->getABITypeAlignment(CV->getType());

  return CurDAG->getExtLoad(Extension, DL, DestType, CurDAG->getEntryNode(),
                            PoolAddr,
                            MachinePointerInfo::getConstantPool(), MemType,
                            /* isVolatile = */ false,
                            /* isNonTemporal = */ false,
                            Alignment).getNode();
}

SDNode *AArch64DAGToDAGISel::LowerToFPLitPool(SDNode *Node) {
  SDLoc DL(Node);
  const ConstantFP *FV = cast<ConstantFPSDNode>(Node)->getConstantFPValue();
  EVT DestType = Node->getValueType(0);

  unsigned Alignment =
    getTargetLowering()->getDataLayout()->getABITypeAlignment(FV->getType());
  SDValue PoolAddr = getConstantPoolItemAddress(DL, FV);

  return CurDAG->getLoad(DestType, DL, CurDAG->getEntryNode(), PoolAddr,
                         MachinePointerInfo::getConstantPool(),
                         /* isVolatile = */ false,
                         /* isNonTemporal = */ false,
                         /* isInvariant = */ true,
                         Alignment).getNode();
}

bool
AArch64DAGToDAGISel::SelectTSTBOperand(SDValue N, SDValue &FixedPos,
                                       unsigned RegWidth) {
  const ConstantSDNode *CN = dyn_cast<ConstantSDNode>(N);
  if (!CN) return false;

  uint64_t Val = CN->getZExtValue();

  if (!isPowerOf2_64(Val)) return false;

  unsigned TestedBit = Log2_64(Val);
  // Checks above should have guaranteed that we haven't lost information in
  // finding TestedBit, but it must still be in range.
  if (TestedBit >= RegWidth) return false;

  FixedPos = CurDAG->getTargetConstant(TestedBit, MVT::i64);
  return true;
}

SDNode *AArch64DAGToDAGISel::SelectAtomic(SDNode *Node, unsigned Op8,
                                          unsigned Op16,unsigned Op32,
                                          unsigned Op64) {
  // Mostly direct translation to the given operations, except that we preserve
  // the AtomicOrdering for use later on.
  AtomicSDNode *AN = cast<AtomicSDNode>(Node);
  EVT VT = AN->getMemoryVT();

  unsigned Op;
  if (VT == MVT::i8)
    Op = Op8;
  else if (VT == MVT::i16)
    Op = Op16;
  else if (VT == MVT::i32)
    Op = Op32;
  else if (VT == MVT::i64)
    Op = Op64;
  else
    llvm_unreachable("Unexpected atomic operation");

  SmallVector<SDValue, 4> Ops;
  for (unsigned i = 1; i < AN->getNumOperands(); ++i)
      Ops.push_back(AN->getOperand(i));

  Ops.push_back(CurDAG->getTargetConstant(AN->getOrdering(), MVT::i32));
  Ops.push_back(AN->getOperand(0)); // Chain moves to the end

  return CurDAG->SelectNodeTo(Node, Op,
                              AN->getValueType(0), MVT::Other,
                              &Ops[0], Ops.size());
}

SDValue AArch64DAGToDAGISel::createDTuple(ArrayRef<SDValue> Regs) {
  static unsigned RegClassIDs[] = { AArch64::DPairRegClassID,
                                    AArch64::DTripleRegClassID,
                                    AArch64::DQuadRegClassID };
  static unsigned SubRegs[] = { AArch64::dsub_0, AArch64::dsub_1,
                                AArch64::dsub_2, AArch64::dsub_3 };

  return createTuple(Regs, RegClassIDs, SubRegs);
}

SDValue AArch64DAGToDAGISel::createQTuple(ArrayRef<SDValue> Regs) {
  static unsigned RegClassIDs[] = { AArch64::QPairRegClassID,
                                    AArch64::QTripleRegClassID,
                                    AArch64::QQuadRegClassID };
  static unsigned SubRegs[] = { AArch64::qsub_0, AArch64::qsub_1,
                                AArch64::qsub_2, AArch64::qsub_3 };

  return createTuple(Regs, RegClassIDs, SubRegs);
}

SDValue AArch64DAGToDAGISel::createTuple(ArrayRef<SDValue> Regs,
                                         unsigned RegClassIDs[],
                                         unsigned SubRegs[]) {
  // There's no special register-class for a vector-list of 1 element: it's just
  // a vector.
  if (Regs.size() == 1)
    return Regs[0];

  assert(Regs.size() >= 2 && Regs.size() <= 4);

  SDLoc DL(Regs[0].getNode());

  SmallVector<SDValue, 4> Ops;

  // First operand of REG_SEQUENCE is the desired RegClass.
  Ops.push_back(
      CurDAG->getTargetConstant(RegClassIDs[Regs.size() - 2], MVT::i32));

  // Then we get pairs of source & subregister-position for the components.
  for (unsigned i = 0; i < Regs.size(); ++i) {
    Ops.push_back(Regs[i]);
    Ops.push_back(CurDAG->getTargetConstant(SubRegs[i], MVT::i32));
  }

  SDNode *N =
      CurDAG->getMachineNode(TargetOpcode::REG_SEQUENCE, DL, MVT::Untyped, Ops);
  return SDValue(N, 0);
}


// Get the register stride update opcode of a VLD/VST instruction that
// is otherwise equivalent to the given fixed stride updating instruction.
static unsigned getVLDSTRegisterUpdateOpcode(unsigned Opc) {
  switch (Opc) {
  default: break;
  case AArch64::LD1WB_8B_fixed: return AArch64::LD1WB_8B_register;
  case AArch64::LD1WB_4H_fixed: return AArch64::LD1WB_4H_register;
  case AArch64::LD1WB_2S_fixed: return AArch64::LD1WB_2S_register;
  case AArch64::LD1WB_1D_fixed: return AArch64::LD1WB_1D_register;
  case AArch64::LD1WB_16B_fixed: return AArch64::LD1WB_16B_register;
  case AArch64::LD1WB_8H_fixed: return AArch64::LD1WB_8H_register;
  case AArch64::LD1WB_4S_fixed: return AArch64::LD1WB_4S_register;
  case AArch64::LD1WB_2D_fixed: return AArch64::LD1WB_2D_register;

  case AArch64::LD2WB_8B_fixed: return AArch64::LD2WB_8B_register;
  case AArch64::LD2WB_4H_fixed: return AArch64::LD2WB_4H_register;
  case AArch64::LD2WB_2S_fixed: return AArch64::LD2WB_2S_register;
  case AArch64::LD2WB_16B_fixed: return AArch64::LD2WB_16B_register;
  case AArch64::LD2WB_8H_fixed: return AArch64::LD2WB_8H_register;
  case AArch64::LD2WB_4S_fixed: return AArch64::LD2WB_4S_register;
  case AArch64::LD2WB_2D_fixed: return AArch64::LD2WB_2D_register;

  case AArch64::LD3WB_8B_fixed: return AArch64::LD3WB_8B_register;
  case AArch64::LD3WB_4H_fixed: return AArch64::LD3WB_4H_register;
  case AArch64::LD3WB_2S_fixed: return AArch64::LD3WB_2S_register;
  case AArch64::LD3WB_16B_fixed: return AArch64::LD3WB_16B_register;
  case AArch64::LD3WB_8H_fixed: return AArch64::LD3WB_8H_register;
  case AArch64::LD3WB_4S_fixed: return AArch64::LD3WB_4S_register;
  case AArch64::LD3WB_2D_fixed: return AArch64::LD3WB_2D_register;

  case AArch64::LD4WB_8B_fixed: return AArch64::LD4WB_8B_register;
  case AArch64::LD4WB_4H_fixed: return AArch64::LD4WB_4H_register;
  case AArch64::LD4WB_2S_fixed: return AArch64::LD4WB_2S_register;
  case AArch64::LD4WB_16B_fixed: return AArch64::LD4WB_16B_register;
  case AArch64::LD4WB_8H_fixed: return AArch64::LD4WB_8H_register;
  case AArch64::LD4WB_4S_fixed: return AArch64::LD4WB_4S_register;
  case AArch64::LD4WB_2D_fixed: return AArch64::LD4WB_2D_register;

  case AArch64::LD1x2WB_8B_fixed: return AArch64::LD1x2WB_8B_register;
  case AArch64::LD1x2WB_4H_fixed: return AArch64::LD1x2WB_4H_register;
  case AArch64::LD1x2WB_2S_fixed: return AArch64::LD1x2WB_2S_register;
  case AArch64::LD1x2WB_1D_fixed: return AArch64::LD1x2WB_1D_register;
  case AArch64::LD1x2WB_16B_fixed: return AArch64::LD1x2WB_16B_register;
  case AArch64::LD1x2WB_8H_fixed: return AArch64::LD1x2WB_8H_register;
  case AArch64::LD1x2WB_4S_fixed: return AArch64::LD1x2WB_4S_register;
  case AArch64::LD1x2WB_2D_fixed: return AArch64::LD1x2WB_2D_register;

  case AArch64::LD1x3WB_8B_fixed: return AArch64::LD1x3WB_8B_register;
  case AArch64::LD1x3WB_4H_fixed: return AArch64::LD1x3WB_4H_register;
  case AArch64::LD1x3WB_2S_fixed: return AArch64::LD1x3WB_2S_register;
  case AArch64::LD1x3WB_1D_fixed: return AArch64::LD1x3WB_1D_register;
  case AArch64::LD1x3WB_16B_fixed: return AArch64::LD1x3WB_16B_register;
  case AArch64::LD1x3WB_8H_fixed: return AArch64::LD1x3WB_8H_register;
  case AArch64::LD1x3WB_4S_fixed: return AArch64::LD1x3WB_4S_register;
  case AArch64::LD1x3WB_2D_fixed: return AArch64::LD1x3WB_2D_register;

  case AArch64::LD1x4WB_8B_fixed: return AArch64::LD1x4WB_8B_register;
  case AArch64::LD1x4WB_4H_fixed: return AArch64::LD1x4WB_4H_register;
  case AArch64::LD1x4WB_2S_fixed: return AArch64::LD1x4WB_2S_register;
  case AArch64::LD1x4WB_1D_fixed: return AArch64::LD1x4WB_1D_register;
  case AArch64::LD1x4WB_16B_fixed: return AArch64::LD1x4WB_16B_register;
  case AArch64::LD1x4WB_8H_fixed: return AArch64::LD1x4WB_8H_register;
  case AArch64::LD1x4WB_4S_fixed: return AArch64::LD1x4WB_4S_register;
  case AArch64::LD1x4WB_2D_fixed: return AArch64::LD1x4WB_2D_register;

  case AArch64::ST1WB_8B_fixed: return AArch64::ST1WB_8B_register;
  case AArch64::ST1WB_4H_fixed: return AArch64::ST1WB_4H_register;
  case AArch64::ST1WB_2S_fixed: return AArch64::ST1WB_2S_register;
  case AArch64::ST1WB_1D_fixed: return AArch64::ST1WB_1D_register;
  case AArch64::ST1WB_16B_fixed: return AArch64::ST1WB_16B_register;
  case AArch64::ST1WB_8H_fixed: return AArch64::ST1WB_8H_register;
  case AArch64::ST1WB_4S_fixed: return AArch64::ST1WB_4S_register;
  case AArch64::ST1WB_2D_fixed: return AArch64::ST1WB_2D_register;

  case AArch64::ST2WB_8B_fixed: return AArch64::ST2WB_8B_register;
  case AArch64::ST2WB_4H_fixed: return AArch64::ST2WB_4H_register;
  case AArch64::ST2WB_2S_fixed: return AArch64::ST2WB_2S_register;
  case AArch64::ST2WB_16B_fixed: return AArch64::ST2WB_16B_register;
  case AArch64::ST2WB_8H_fixed: return AArch64::ST2WB_8H_register;
  case AArch64::ST2WB_4S_fixed: return AArch64::ST2WB_4S_register;
  case AArch64::ST2WB_2D_fixed: return AArch64::ST2WB_2D_register;

  case AArch64::ST3WB_8B_fixed: return AArch64::ST3WB_8B_register;
  case AArch64::ST3WB_4H_fixed: return AArch64::ST3WB_4H_register;
  case AArch64::ST3WB_2S_fixed: return AArch64::ST3WB_2S_register;
  case AArch64::ST3WB_16B_fixed: return AArch64::ST3WB_16B_register;
  case AArch64::ST3WB_8H_fixed: return AArch64::ST3WB_8H_register;
  case AArch64::ST3WB_4S_fixed: return AArch64::ST3WB_4S_register;
  case AArch64::ST3WB_2D_fixed: return AArch64::ST3WB_2D_register;

  case AArch64::ST4WB_8B_fixed: return AArch64::ST4WB_8B_register;
  case AArch64::ST4WB_4H_fixed: return AArch64::ST4WB_4H_register;
  case AArch64::ST4WB_2S_fixed: return AArch64::ST4WB_2S_register;
  case AArch64::ST4WB_16B_fixed: return AArch64::ST4WB_16B_register;
  case AArch64::ST4WB_8H_fixed: return AArch64::ST4WB_8H_register;
  case AArch64::ST4WB_4S_fixed: return AArch64::ST4WB_4S_register;
  case AArch64::ST4WB_2D_fixed: return AArch64::ST4WB_2D_register;

  case AArch64::ST1x2WB_8B_fixed: return AArch64::ST1x2WB_8B_register;
  case AArch64::ST1x2WB_4H_fixed: return AArch64::ST1x2WB_4H_register;
  case AArch64::ST1x2WB_2S_fixed: return AArch64::ST1x2WB_2S_register;
  case AArch64::ST1x2WB_1D_fixed: return AArch64::ST1x2WB_1D_register;
  case AArch64::ST1x2WB_16B_fixed: return AArch64::ST1x2WB_16B_register;
  case AArch64::ST1x2WB_8H_fixed: return AArch64::ST1x2WB_8H_register;
  case AArch64::ST1x2WB_4S_fixed: return AArch64::ST1x2WB_4S_register;
  case AArch64::ST1x2WB_2D_fixed: return AArch64::ST1x2WB_2D_register;

  case AArch64::ST1x3WB_8B_fixed: return AArch64::ST1x3WB_8B_register;
  case AArch64::ST1x3WB_4H_fixed: return AArch64::ST1x3WB_4H_register;
  case AArch64::ST1x3WB_2S_fixed: return AArch64::ST1x3WB_2S_register;
  case AArch64::ST1x3WB_1D_fixed: return AArch64::ST1x3WB_1D_register;
  case AArch64::ST1x3WB_16B_fixed: return AArch64::ST1x3WB_16B_register;
  case AArch64::ST1x3WB_8H_fixed: return AArch64::ST1x3WB_8H_register;
  case AArch64::ST1x3WB_4S_fixed: return AArch64::ST1x3WB_4S_register;
  case AArch64::ST1x3WB_2D_fixed: return AArch64::ST1x3WB_2D_register;

  case AArch64::ST1x4WB_8B_fixed: return AArch64::ST1x4WB_8B_register;
  case AArch64::ST1x4WB_4H_fixed: return AArch64::ST1x4WB_4H_register;
  case AArch64::ST1x4WB_2S_fixed: return AArch64::ST1x4WB_2S_register;
  case AArch64::ST1x4WB_1D_fixed: return AArch64::ST1x4WB_1D_register;
  case AArch64::ST1x4WB_16B_fixed: return AArch64::ST1x4WB_16B_register;
  case AArch64::ST1x4WB_8H_fixed: return AArch64::ST1x4WB_8H_register;
  case AArch64::ST1x4WB_4S_fixed: return AArch64::ST1x4WB_4S_register;
  case AArch64::ST1x4WB_2D_fixed: return AArch64::ST1x4WB_2D_register;

  // Post-index of duplicate loads
  case AArch64::LD2R_WB_8B_fixed: return AArch64::LD2R_WB_8B_register;
  case AArch64::LD2R_WB_4H_fixed: return AArch64::LD2R_WB_4H_register;
  case AArch64::LD2R_WB_2S_fixed: return AArch64::LD2R_WB_2S_register;
  case AArch64::LD2R_WB_1D_fixed: return AArch64::LD2R_WB_1D_register;
  case AArch64::LD2R_WB_16B_fixed: return AArch64::LD2R_WB_16B_register;
  case AArch64::LD2R_WB_8H_fixed: return AArch64::LD2R_WB_8H_register;
  case AArch64::LD2R_WB_4S_fixed: return AArch64::LD2R_WB_4S_register;
  case AArch64::LD2R_WB_2D_fixed: return AArch64::LD2R_WB_2D_register;

  case AArch64::LD3R_WB_8B_fixed: return AArch64::LD3R_WB_8B_register;
  case AArch64::LD3R_WB_4H_fixed: return AArch64::LD3R_WB_4H_register;
  case AArch64::LD3R_WB_2S_fixed: return AArch64::LD3R_WB_2S_register;
  case AArch64::LD3R_WB_1D_fixed: return AArch64::LD3R_WB_1D_register;
  case AArch64::LD3R_WB_16B_fixed: return AArch64::LD3R_WB_16B_register;
  case AArch64::LD3R_WB_8H_fixed: return AArch64::LD3R_WB_8H_register;
  case AArch64::LD3R_WB_4S_fixed: return AArch64::LD3R_WB_4S_register;
  case AArch64::LD3R_WB_2D_fixed: return AArch64::LD3R_WB_2D_register;

  case AArch64::LD4R_WB_8B_fixed: return AArch64::LD4R_WB_8B_register;
  case AArch64::LD4R_WB_4H_fixed: return AArch64::LD4R_WB_4H_register;
  case AArch64::LD4R_WB_2S_fixed: return AArch64::LD4R_WB_2S_register;
  case AArch64::LD4R_WB_1D_fixed: return AArch64::LD4R_WB_1D_register;
  case AArch64::LD4R_WB_16B_fixed: return AArch64::LD4R_WB_16B_register;
  case AArch64::LD4R_WB_8H_fixed: return AArch64::LD4R_WB_8H_register;
  case AArch64::LD4R_WB_4S_fixed: return AArch64::LD4R_WB_4S_register;
  case AArch64::LD4R_WB_2D_fixed: return AArch64::LD4R_WB_2D_register;

  // Post-index of lane loads
  case AArch64::LD2LN_WB_B_fixed: return AArch64::LD2LN_WB_B_register;
  case AArch64::LD2LN_WB_H_fixed: return AArch64::LD2LN_WB_H_register;
  case AArch64::LD2LN_WB_S_fixed: return AArch64::LD2LN_WB_S_register;
  case AArch64::LD2LN_WB_D_fixed: return AArch64::LD2LN_WB_D_register;

  case AArch64::LD3LN_WB_B_fixed: return AArch64::LD3LN_WB_B_register;
  case AArch64::LD3LN_WB_H_fixed: return AArch64::LD3LN_WB_H_register;
  case AArch64::LD3LN_WB_S_fixed: return AArch64::LD3LN_WB_S_register;
  case AArch64::LD3LN_WB_D_fixed: return AArch64::LD3LN_WB_D_register;

  case AArch64::LD4LN_WB_B_fixed: return AArch64::LD4LN_WB_B_register;
  case AArch64::LD4LN_WB_H_fixed: return AArch64::LD4LN_WB_H_register;
  case AArch64::LD4LN_WB_S_fixed: return AArch64::LD4LN_WB_S_register;
  case AArch64::LD4LN_WB_D_fixed: return AArch64::LD4LN_WB_D_register;

  // Post-index of lane stores
  case AArch64::ST2LN_WB_B_fixed: return AArch64::ST2LN_WB_B_register;
  case AArch64::ST2LN_WB_H_fixed: return AArch64::ST2LN_WB_H_register;
  case AArch64::ST2LN_WB_S_fixed: return AArch64::ST2LN_WB_S_register;
  case AArch64::ST2LN_WB_D_fixed: return AArch64::ST2LN_WB_D_register;

  case AArch64::ST3LN_WB_B_fixed: return AArch64::ST3LN_WB_B_register;
  case AArch64::ST3LN_WB_H_fixed: return AArch64::ST3LN_WB_H_register;
  case AArch64::ST3LN_WB_S_fixed: return AArch64::ST3LN_WB_S_register;
  case AArch64::ST3LN_WB_D_fixed: return AArch64::ST3LN_WB_D_register;

  case AArch64::ST4LN_WB_B_fixed: return AArch64::ST4LN_WB_B_register;
  case AArch64::ST4LN_WB_H_fixed: return AArch64::ST4LN_WB_H_register;
  case AArch64::ST4LN_WB_S_fixed: return AArch64::ST4LN_WB_S_register;
  case AArch64::ST4LN_WB_D_fixed: return AArch64::ST4LN_WB_D_register;
  }
  return Opc; // If not one we handle, return it unchanged.
}

SDNode *AArch64DAGToDAGISel::SelectVLD(SDNode *N, bool isUpdating,
                                       unsigned NumVecs,
                                       const uint16_t *Opcodes) {
  assert(NumVecs >= 1 && NumVecs <= 4 && "VLD NumVecs out-of-range");

  EVT VT = N->getValueType(0);
  unsigned OpcodeIndex;
  bool is64BitVector = VT.is64BitVector();
  switch (VT.getScalarType().getSizeInBits()) {
  case 8: OpcodeIndex = is64BitVector ? 0 : 4; break;
  case 16: OpcodeIndex = is64BitVector ? 1 : 5; break;
  case 32: OpcodeIndex = is64BitVector ? 2 : 6; break;
  case 64: OpcodeIndex = is64BitVector ? 3 : 7; break;
  default: llvm_unreachable("unhandled vector load type");
  }
  unsigned Opc = Opcodes[OpcodeIndex];

  SmallVector<SDValue, 2> Ops;
  unsigned AddrOpIdx = isUpdating ? 1 : 2;
  Ops.push_back(N->getOperand(AddrOpIdx)); // Push back the Memory Address

  if (isUpdating) {
    SDValue Inc = N->getOperand(AddrOpIdx + 1);
    if (!isa<ConstantSDNode>(Inc.getNode())) // Increment in Register
      Opc = getVLDSTRegisterUpdateOpcode(Opc);
    Ops.push_back(Inc);
  }

  Ops.push_back(N->getOperand(0)); // Push back the Chain

  SmallVector<EVT, 3> ResTys;
  // Push back the type of return super register
  if (NumVecs == 1)
    ResTys.push_back(VT);
  else if (NumVecs == 3)
    ResTys.push_back(MVT::Untyped);
  else {
    EVT ResTy = EVT::getVectorVT(*CurDAG->getContext(), MVT::i64,
                                 is64BitVector ? NumVecs : NumVecs * 2);
    ResTys.push_back(ResTy);
  }

  if (isUpdating)
    ResTys.push_back(MVT::i64); // Type of the updated register
  ResTys.push_back(MVT::Other); // Type of the Chain
  SDLoc dl(N);
  SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops);

  // Transfer memoperands.
  MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
  MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
  cast<MachineSDNode>(VLd)->setMemRefs(MemOp, MemOp + 1);

  if (NumVecs == 1)
    return VLd;

  // If NumVecs > 1, the return result is a super register containing 2-4
  // consecutive vector registers.
  SDValue SuperReg = SDValue(VLd, 0);

  unsigned Sub0 = is64BitVector ? AArch64::dsub_0 : AArch64::qsub_0;
  for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
    ReplaceUses(SDValue(N, Vec),
                CurDAG->getTargetExtractSubreg(Sub0 + Vec, dl, VT, SuperReg));
  // Update users of the Chain
  ReplaceUses(SDValue(N, NumVecs), SDValue(VLd, 1));
  if (isUpdating)
    ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLd, 2));

  return NULL;
}

SDNode *AArch64DAGToDAGISel::SelectVST(SDNode *N, bool isUpdating,
                                       unsigned NumVecs,
                                       const uint16_t *Opcodes) {
  assert(NumVecs >= 1 && NumVecs <= 4 && "VST NumVecs out-of-range");
  SDLoc dl(N);

  MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
  MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();

  unsigned AddrOpIdx = isUpdating ? 1 : 2;
  unsigned Vec0Idx = 3;
  EVT VT = N->getOperand(Vec0Idx).getValueType();
  unsigned OpcodeIndex;
  bool is64BitVector = VT.is64BitVector();
  switch (VT.getScalarType().getSizeInBits()) {
  case 8: OpcodeIndex = is64BitVector ? 0 : 4; break;
  case 16: OpcodeIndex = is64BitVector ? 1 : 5; break;
  case 32: OpcodeIndex = is64BitVector ? 2 : 6; break;
  case 64: OpcodeIndex = is64BitVector ? 3 : 7; break;
  default: llvm_unreachable("unhandled vector store type");
  }
  unsigned Opc = Opcodes[OpcodeIndex];

  SmallVector<EVT, 2> ResTys;
  if (isUpdating)
    ResTys.push_back(MVT::i64);
  ResTys.push_back(MVT::Other); // Type for the Chain

  SmallVector<SDValue, 6> Ops;
  Ops.push_back(N->getOperand(AddrOpIdx)); // Push back the Memory Address

  if (isUpdating) {
    SDValue Inc = N->getOperand(AddrOpIdx + 1);
    if (!isa<ConstantSDNode>(Inc.getNode())) // Increment in Register
      Opc = getVLDSTRegisterUpdateOpcode(Opc);
    Ops.push_back(Inc);
  }

  SmallVector<SDValue, 4> Regs(N->op_begin() + Vec0Idx,
                               N->op_begin() + Vec0Idx + NumVecs);
  SDValue SrcReg = is64BitVector ? createDTuple(Regs) : createQTuple(Regs);
  Ops.push_back(SrcReg);

  // Push back the Chain
  Ops.push_back(N->getOperand(0));

  // Transfer memoperands.
  SDNode *VSt = CurDAG->getMachineNode(Opc, dl, ResTys, Ops);
  cast<MachineSDNode>(VSt)->setMemRefs(MemOp, MemOp + 1);

  return VSt;
}

SDValue
AArch64DAGToDAGISel::getTargetSubregToReg(int SRIdx, SDLoc DL, EVT VT, EVT VTD,
                                          SDValue Operand) {
  SDNode *Reg = CurDAG->getMachineNode(TargetOpcode::SUBREG_TO_REG, DL,
                        VT, VTD, MVT::Other,
                        CurDAG->getTargetConstant(0, MVT::i64),
                        Operand,
                        CurDAG->getTargetConstant(AArch64::sub_64, MVT::i32));
  return SDValue(Reg, 0);
}

SDNode *AArch64DAGToDAGISel::SelectVLDDup(SDNode *N, bool isUpdating,
                                          unsigned NumVecs,
                                          const uint16_t *Opcodes) {
  assert(NumVecs >=2 && NumVecs <= 4 && "Load Dup NumVecs out-of-range");
  SDLoc dl(N);

  EVT VT = N->getValueType(0);
  unsigned OpcodeIndex;
  bool is64BitVector = VT.is64BitVector();
  switch (VT.getScalarType().getSizeInBits()) {
  case 8: OpcodeIndex = is64BitVector ? 0 : 4; break;
  case 16: OpcodeIndex = is64BitVector ? 1 : 5; break;
  case 32: OpcodeIndex = is64BitVector ? 2 : 6; break;
  case 64: OpcodeIndex = is64BitVector ? 3 : 7; break;
  default: llvm_unreachable("unhandled vector duplicate lane load type");
  }
  unsigned Opc = Opcodes[OpcodeIndex];

  SDValue SuperReg;
  SmallVector<SDValue, 6> Ops;
  Ops.push_back(N->getOperand(1)); // Push back the Memory Address
  if (isUpdating) {
    SDValue Inc = N->getOperand(2);
    if (!isa<ConstantSDNode>(Inc.getNode())) // Increment in Register
      Opc = getVLDSTRegisterUpdateOpcode(Opc);
    Ops.push_back(Inc);
  }
  Ops.push_back(N->getOperand(0)); // Push back the Chain

  SmallVector<EVT, 3> ResTys;
  // Push back the type of return super register
  if (NumVecs == 3)
    ResTys.push_back(MVT::Untyped);
  else {
    EVT ResTy = EVT::getVectorVT(*CurDAG->getContext(), MVT::i64,
                                 is64BitVector ? NumVecs : NumVecs * 2);
    ResTys.push_back(ResTy);
  }
  if (isUpdating)
    ResTys.push_back(MVT::i64); // Type of the updated register
  ResTys.push_back(MVT::Other); // Type of the Chain
  SDNode *VLdDup = CurDAG->getMachineNode(Opc, dl, ResTys, Ops);

  // Transfer memoperands.
  MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
  MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
  cast<MachineSDNode>(VLdDup)->setMemRefs(MemOp, MemOp + 1);

  SuperReg = SDValue(VLdDup, 0);
  unsigned Sub0 = is64BitVector ? AArch64::dsub_0 : AArch64::qsub_0;
  // Update uses of each registers in super register
  for (unsigned Vec = 0; Vec < NumVecs; ++Vec)
    ReplaceUses(SDValue(N, Vec),
                CurDAG->getTargetExtractSubreg(Sub0 + Vec, dl, VT, SuperReg));
  // Update uses of the Chain
  ReplaceUses(SDValue(N, NumVecs), SDValue(VLdDup, 1));
  if (isUpdating)
    ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLdDup, 2));
  return NULL;
}

// We only have 128-bit vector type of load/store lane instructions.
// If it is 64-bit vector, we also select it to the 128-bit instructions.
// Just use SUBREG_TO_REG to adapt the input to 128-bit vector and
// EXTRACT_SUBREG to get the 64-bit vector from the 128-bit vector output.
SDNode *AArch64DAGToDAGISel::SelectVLDSTLane(SDNode *N, bool IsLoad,
                                             bool isUpdating, unsigned NumVecs,
                                             const uint16_t *Opcodes) {
  assert(NumVecs >= 2 && NumVecs <= 4 && "VLDSTLane NumVecs out-of-range");
  SDLoc dl(N);
  unsigned AddrOpIdx = isUpdating ? 1 : 2;
  unsigned Vec0Idx = 3;

  SDValue Chain = N->getOperand(0);
  unsigned Lane =
      cast<ConstantSDNode>(N->getOperand(Vec0Idx + NumVecs))->getZExtValue();
  EVT VT = N->getOperand(Vec0Idx).getValueType();
  bool is64BitVector = VT.is64BitVector();
  EVT VT64; // 64-bit Vector Type

  if (is64BitVector) {
    VT64 = VT;
    VT = EVT::getVectorVT(*CurDAG->getContext(), VT.getVectorElementType(),
                          VT.getVectorNumElements() * 2);
  }

  unsigned OpcodeIndex;
  switch (VT.getScalarType().getSizeInBits()) {
  case 8: OpcodeIndex = 0; break;
  case 16: OpcodeIndex = 1; break;
  case 32: OpcodeIndex = 2; break;
  case 64: OpcodeIndex = 3; break;
  default: llvm_unreachable("unhandled vector lane load/store type");
  }
  unsigned Opc = Opcodes[OpcodeIndex];

  SmallVector<EVT, 3> ResTys;
  if (IsLoad) {
    // Push back the type of return super register
    if (NumVecs == 3)
      ResTys.push_back(MVT::Untyped);
    else {
      EVT ResTy = EVT::getVectorVT(*CurDAG->getContext(), MVT::i64,
                                   is64BitVector ? NumVecs : NumVecs * 2);
      ResTys.push_back(ResTy);
    }
  }
  if (isUpdating)
    ResTys.push_back(MVT::i64); // Type of the updated register
  ResTys.push_back(MVT::Other); // Type of Chain
  SmallVector<SDValue, 5> Ops;
  Ops.push_back(N->getOperand(AddrOpIdx)); // Push back the Memory Address
  if (isUpdating) {
    SDValue Inc = N->getOperand(AddrOpIdx + 1);
    if (!isa<ConstantSDNode>(Inc.getNode())) // Increment in Register
      Opc = getVLDSTRegisterUpdateOpcode(Opc);
    Ops.push_back(Inc);
  }

  SmallVector<SDValue, 4> Regs(N->op_begin() + Vec0Idx,
                               N->op_begin() + Vec0Idx + NumVecs);
  if (is64BitVector)
    for (unsigned i = 0; i < Regs.size(); i++)
      Regs[i] = getTargetSubregToReg(AArch64::sub_64, dl, VT, VT64, Regs[i]);
  SDValue SuperReg = createQTuple(Regs);

  Ops.push_back(SuperReg); // Source Reg
  SDValue LaneValue = CurDAG->getTargetConstant(Lane, MVT::i32);
  Ops.push_back(LaneValue);
  Ops.push_back(Chain); // Push back the Chain

  SDNode *VLdLn = CurDAG->getMachineNode(Opc, dl, ResTys, Ops);
  MachineSDNode::mmo_iterator MemOp = MF->allocateMemRefsArray(1);
  MemOp[0] = cast<MemIntrinsicSDNode>(N)->getMemOperand();
  cast<MachineSDNode>(VLdLn)->setMemRefs(MemOp, MemOp + 1);
  if (!IsLoad)
    return VLdLn;

  // Extract the subregisters.
  SuperReg = SDValue(VLdLn, 0);
  unsigned Sub0 = AArch64::qsub_0;
  // Update uses of each registers in super register
  for (unsigned Vec = 0; Vec < NumVecs; ++Vec) {
    SDValue SUB0 = CurDAG->getTargetExtractSubreg(Sub0 + Vec, dl, VT, SuperReg);
    if (is64BitVector) {
      SUB0 = CurDAG->getTargetExtractSubreg(AArch64::sub_64, dl, VT64, SUB0);
    }
    ReplaceUses(SDValue(N, Vec), SUB0);
  }
  ReplaceUses(SDValue(N, NumVecs), SDValue(VLdLn, 1));
  if (isUpdating)
    ReplaceUses(SDValue(N, NumVecs + 1), SDValue(VLdLn, 2));
  return NULL;
}

unsigned AArch64DAGToDAGISel::getTBLOpc(bool IsExt, bool Is64Bit,
                                        unsigned NumOfVec) {
  assert(NumOfVec >= 1 && NumOfVec <= 4 && "VST NumVecs out-of-range");

  unsigned Opc = 0;
  switch (NumOfVec) {
  default:
    break;
  case 1:
    if (IsExt)
      Opc = Is64Bit ? AArch64::TBX1_8b : AArch64::TBX1_16b;
    else
      Opc = Is64Bit ? AArch64::TBL1_8b : AArch64::TBL1_16b;
    break;
  case 2:
    if (IsExt)
      Opc = Is64Bit ? AArch64::TBX2_8b : AArch64::TBX2_16b;
    else
      Opc = Is64Bit ? AArch64::TBL2_8b : AArch64::TBL2_16b;
    break;
  case 3:
    if (IsExt)
      Opc = Is64Bit ? AArch64::TBX3_8b : AArch64::TBX3_16b;
    else
      Opc = Is64Bit ? AArch64::TBL3_8b : AArch64::TBL3_16b;
    break;
  case 4:
    if (IsExt)
      Opc = Is64Bit ? AArch64::TBX4_8b : AArch64::TBX4_16b;
    else
      Opc = Is64Bit ? AArch64::TBL4_8b : AArch64::TBL4_16b;
    break;
  }

  return Opc;
}

SDNode *AArch64DAGToDAGISel::SelectVTBL(SDNode *N, unsigned NumVecs,
                                        bool IsExt) {
  assert(NumVecs >= 1 && NumVecs <= 4 && "VST NumVecs out-of-range");
  SDLoc dl(N);

  // Check the element of look up table is 64-bit or not
  unsigned Vec0Idx = IsExt ? 2 : 1;
  assert(!N->getOperand(Vec0Idx + 0).getValueType().is64BitVector() &&
         "The element of lookup table for vtbl and vtbx must be 128-bit");

  // Check the return value type is 64-bit or not
  EVT ResVT = N->getValueType(0);
  bool is64BitRes = ResVT.is64BitVector();

  // Create new SDValue for vector list
  SmallVector<SDValue, 4> Regs(N->op_begin() + Vec0Idx,
                               N->op_begin() + Vec0Idx + NumVecs);
  SDValue TblReg = createQTuple(Regs);
  unsigned Opc = getTBLOpc(IsExt, is64BitRes, NumVecs);

  SmallVector<SDValue, 3> Ops;
  if (IsExt)
    Ops.push_back(N->getOperand(1));
  Ops.push_back(TblReg);
  Ops.push_back(N->getOperand(Vec0Idx + NumVecs));
  return CurDAG->getMachineNode(Opc, dl, ResVT, Ops);
}

SDNode *AArch64DAGToDAGISel::Select(SDNode *Node) {
  // Dump information about the Node being selected
  DEBUG(dbgs() << "Selecting: "; Node->dump(CurDAG); dbgs() << "\n");

  if (Node->isMachineOpcode()) {
    DEBUG(dbgs() << "== "; Node->dump(CurDAG); dbgs() << "\n");
    Node->setNodeId(-1);
    return NULL;
  }

  switch (Node->getOpcode()) {
  case ISD::ATOMIC_LOAD_ADD:
    return SelectAtomic(Node,
                        AArch64::ATOMIC_LOAD_ADD_I8,
                        AArch64::ATOMIC_LOAD_ADD_I16,
                        AArch64::ATOMIC_LOAD_ADD_I32,
                        AArch64::ATOMIC_LOAD_ADD_I64);
  case ISD::ATOMIC_LOAD_SUB:
    return SelectAtomic(Node,
                        AArch64::ATOMIC_LOAD_SUB_I8,
                        AArch64::ATOMIC_LOAD_SUB_I16,
                        AArch64::ATOMIC_LOAD_SUB_I32,
                        AArch64::ATOMIC_LOAD_SUB_I64);
  case ISD::ATOMIC_LOAD_AND:
    return SelectAtomic(Node,
                        AArch64::ATOMIC_LOAD_AND_I8,
                        AArch64::ATOMIC_LOAD_AND_I16,
                        AArch64::ATOMIC_LOAD_AND_I32,
                        AArch64::ATOMIC_LOAD_AND_I64);
  case ISD::ATOMIC_LOAD_OR:
    return SelectAtomic(Node,
                        AArch64::ATOMIC_LOAD_OR_I8,
                        AArch64::ATOMIC_LOAD_OR_I16,
                        AArch64::ATOMIC_LOAD_OR_I32,
                        AArch64::ATOMIC_LOAD_OR_I64);
  case ISD::ATOMIC_LOAD_XOR:
    return SelectAtomic(Node,
                        AArch64::ATOMIC_LOAD_XOR_I8,
                        AArch64::ATOMIC_LOAD_XOR_I16,
                        AArch64::ATOMIC_LOAD_XOR_I32,
                        AArch64::ATOMIC_LOAD_XOR_I64);
  case ISD::ATOMIC_LOAD_NAND:
    return SelectAtomic(Node,
                        AArch64::ATOMIC_LOAD_NAND_I8,
                        AArch64::ATOMIC_LOAD_NAND_I16,
                        AArch64::ATOMIC_LOAD_NAND_I32,
                        AArch64::ATOMIC_LOAD_NAND_I64);
  case ISD::ATOMIC_LOAD_MIN:
    return SelectAtomic(Node,
                        AArch64::ATOMIC_LOAD_MIN_I8,
                        AArch64::ATOMIC_LOAD_MIN_I16,
                        AArch64::ATOMIC_LOAD_MIN_I32,
                        AArch64::ATOMIC_LOAD_MIN_I64);
  case ISD::ATOMIC_LOAD_MAX:
    return SelectAtomic(Node,
                        AArch64::ATOMIC_LOAD_MAX_I8,
                        AArch64::ATOMIC_LOAD_MAX_I16,
                        AArch64::ATOMIC_LOAD_MAX_I32,
                        AArch64::ATOMIC_LOAD_MAX_I64);
  case ISD::ATOMIC_LOAD_UMIN:
    return SelectAtomic(Node,
                        AArch64::ATOMIC_LOAD_UMIN_I8,
                        AArch64::ATOMIC_LOAD_UMIN_I16,
                        AArch64::ATOMIC_LOAD_UMIN_I32,
                        AArch64::ATOMIC_LOAD_UMIN_I64);
  case ISD::ATOMIC_LOAD_UMAX:
    return SelectAtomic(Node,
                        AArch64::ATOMIC_LOAD_UMAX_I8,
                        AArch64::ATOMIC_LOAD_UMAX_I16,
                        AArch64::ATOMIC_LOAD_UMAX_I32,
                        AArch64::ATOMIC_LOAD_UMAX_I64);
  case ISD::ATOMIC_SWAP:
    return SelectAtomic(Node,
                        AArch64::ATOMIC_SWAP_I8,
                        AArch64::ATOMIC_SWAP_I16,
                        AArch64::ATOMIC_SWAP_I32,
                        AArch64::ATOMIC_SWAP_I64);
  case ISD::ATOMIC_CMP_SWAP:
    return SelectAtomic(Node,
                        AArch64::ATOMIC_CMP_SWAP_I8,
                        AArch64::ATOMIC_CMP_SWAP_I16,
                        AArch64::ATOMIC_CMP_SWAP_I32,
                        AArch64::ATOMIC_CMP_SWAP_I64);
  case ISD::FrameIndex: {
    int FI = cast<FrameIndexSDNode>(Node)->getIndex();
    EVT PtrTy = getTargetLowering()->getPointerTy();
    SDValue TFI = CurDAG->getTargetFrameIndex(FI, PtrTy);
    return CurDAG->SelectNodeTo(Node, AArch64::ADDxxi_lsl0_s, PtrTy,
                                TFI, CurDAG->getTargetConstant(0, PtrTy));
  }
  case ISD::ConstantPool: {
    // Constant pools are fine, just create a Target entry.
    ConstantPoolSDNode *CN = cast<ConstantPoolSDNode>(Node);
    const Constant *C = CN->getConstVal();
    SDValue CP = CurDAG->getTargetConstantPool(C, CN->getValueType(0));

    ReplaceUses(SDValue(Node, 0), CP);
    return NULL;
  }
  case ISD::Constant: {
    SDNode *ResNode = 0;
    if (cast<ConstantSDNode>(Node)->getZExtValue() == 0) {
      // XZR and WZR are probably even better than an actual move: most of the
      // time they can be folded into another instruction with *no* cost.

      EVT Ty = Node->getValueType(0);
      assert((Ty == MVT::i32 || Ty == MVT::i64) && "unexpected type");
      uint16_t Register = Ty == MVT::i32 ? AArch64::WZR : AArch64::XZR;
      ResNode = CurDAG->getCopyFromReg(CurDAG->getEntryNode(),
                                       SDLoc(Node),
                                       Register, Ty).getNode();
    }

    // Next best option is a move-immediate, see if we can do that.
    if (!ResNode) {
      ResNode = TrySelectToMoveImm(Node);
    }

    if (ResNode)
      return ResNode;

    // If even that fails we fall back to a lit-pool entry at the moment. Future
    // tuning may change this to a sequence of MOVZ/MOVN/MOVK instructions.
    ResNode = SelectToLitPool(Node);
    assert(ResNode && "We need *some* way to materialise a constant");

    // We want to continue selection at this point since the litpool access
    // generated used generic nodes for simplicity.
    ReplaceUses(SDValue(Node, 0), SDValue(ResNode, 0));
    Node = ResNode;
    break;
  }
  case ISD::ConstantFP: {
    if (A64Imms::isFPImm(cast<ConstantFPSDNode>(Node)->getValueAPF())) {
      // FMOV will take care of it from TableGen
      break;
    }

    SDNode *ResNode = LowerToFPLitPool(Node);
    ReplaceUses(SDValue(Node, 0), SDValue(ResNode, 0));

    // We want to continue selection at this point since the litpool access
    // generated used generic nodes for simplicity.
    Node = ResNode;
    break;
  }
  case AArch64ISD::NEON_LD1_UPD: {
    static const uint16_t Opcodes[] = {
      AArch64::LD1WB_8B_fixed,  AArch64::LD1WB_4H_fixed,
      AArch64::LD1WB_2S_fixed,  AArch64::LD1WB_1D_fixed,
      AArch64::LD1WB_16B_fixed, AArch64::LD1WB_8H_fixed,
      AArch64::LD1WB_4S_fixed,  AArch64::LD1WB_2D_fixed
    };
    return SelectVLD(Node, true, 1, Opcodes);
  }
  case AArch64ISD::NEON_LD2_UPD: {
    static const uint16_t Opcodes[] = {
      AArch64::LD2WB_8B_fixed,  AArch64::LD2WB_4H_fixed,
      AArch64::LD2WB_2S_fixed,  AArch64::LD1x2WB_1D_fixed,
      AArch64::LD2WB_16B_fixed, AArch64::LD2WB_8H_fixed,
      AArch64::LD2WB_4S_fixed,  AArch64::LD2WB_2D_fixed
    };
    return SelectVLD(Node, true, 2, Opcodes);
  }
  case AArch64ISD::NEON_LD3_UPD: {
    static const uint16_t Opcodes[] = {
      AArch64::LD3WB_8B_fixed,  AArch64::LD3WB_4H_fixed,
      AArch64::LD3WB_2S_fixed,  AArch64::LD1x3WB_1D_fixed,
      AArch64::LD3WB_16B_fixed, AArch64::LD3WB_8H_fixed,
      AArch64::LD3WB_4S_fixed,  AArch64::LD3WB_2D_fixed
    };
    return SelectVLD(Node, true, 3, Opcodes);
  }
  case AArch64ISD::NEON_LD4_UPD: {
    static const uint16_t Opcodes[] = {
      AArch64::LD4WB_8B_fixed,  AArch64::LD4WB_4H_fixed,
      AArch64::LD4WB_2S_fixed,  AArch64::LD1x4WB_1D_fixed,
      AArch64::LD4WB_16B_fixed, AArch64::LD4WB_8H_fixed,
      AArch64::LD4WB_4S_fixed,  AArch64::LD4WB_2D_fixed
    };
    return SelectVLD(Node, true, 4, Opcodes);
  }
  case AArch64ISD::NEON_LD1x2_UPD: {
    static const uint16_t Opcodes[] = {
      AArch64::LD1x2WB_8B_fixed,  AArch64::LD1x2WB_4H_fixed,
      AArch64::LD1x2WB_2S_fixed,  AArch64::LD1x2WB_1D_fixed,
      AArch64::LD1x2WB_16B_fixed, AArch64::LD1x2WB_8H_fixed,
      AArch64::LD1x2WB_4S_fixed,  AArch64::LD1x2WB_2D_fixed
    };
    return SelectVLD(Node, true, 2, Opcodes);
  }
  case AArch64ISD::NEON_LD1x3_UPD: {
    static const uint16_t Opcodes[] = {
      AArch64::LD1x3WB_8B_fixed,  AArch64::LD1x3WB_4H_fixed,
      AArch64::LD1x3WB_2S_fixed,  AArch64::LD1x3WB_1D_fixed,
      AArch64::LD1x3WB_16B_fixed, AArch64::LD1x3WB_8H_fixed,
      AArch64::LD1x3WB_4S_fixed,  AArch64::LD1x3WB_2D_fixed
    };
    return SelectVLD(Node, true, 3, Opcodes);
  }
  case AArch64ISD::NEON_LD1x4_UPD: {
    static const uint16_t Opcodes[] = {
      AArch64::LD1x4WB_8B_fixed,  AArch64::LD1x4WB_4H_fixed,
      AArch64::LD1x4WB_2S_fixed,  AArch64::LD1x4WB_1D_fixed,
      AArch64::LD1x4WB_16B_fixed, AArch64::LD1x4WB_8H_fixed,
      AArch64::LD1x4WB_4S_fixed,  AArch64::LD1x4WB_2D_fixed
    };
    return SelectVLD(Node, true, 4, Opcodes);
  }
  case AArch64ISD::NEON_ST1_UPD: {
    static const uint16_t Opcodes[] = {
      AArch64::ST1WB_8B_fixed,  AArch64::ST1WB_4H_fixed,
      AArch64::ST1WB_2S_fixed,  AArch64::ST1WB_1D_fixed,
      AArch64::ST1WB_16B_fixed, AArch64::ST1WB_8H_fixed,
      AArch64::ST1WB_4S_fixed,  AArch64::ST1WB_2D_fixed
    };
    return SelectVST(Node, true, 1, Opcodes);
  }
  case AArch64ISD::NEON_ST2_UPD: {
    static const uint16_t Opcodes[] = {
      AArch64::ST2WB_8B_fixed,  AArch64::ST2WB_4H_fixed,
      AArch64::ST2WB_2S_fixed,  AArch64::ST1x2WB_1D_fixed,
      AArch64::ST2WB_16B_fixed, AArch64::ST2WB_8H_fixed,
      AArch64::ST2WB_4S_fixed,  AArch64::ST2WB_2D_fixed
    };
    return SelectVST(Node, true, 2, Opcodes);
  }
  case AArch64ISD::NEON_ST3_UPD: {
    static const uint16_t Opcodes[] = {
      AArch64::ST3WB_8B_fixed,  AArch64::ST3WB_4H_fixed,
      AArch64::ST3WB_2S_fixed,  AArch64::ST1x3WB_1D_fixed,
      AArch64::ST3WB_16B_fixed, AArch64::ST3WB_8H_fixed,
      AArch64::ST3WB_4S_fixed,  AArch64::ST3WB_2D_fixed
    };
    return SelectVST(Node, true, 3, Opcodes);
  }
  case AArch64ISD::NEON_ST4_UPD: {
    static const uint16_t Opcodes[] = {
      AArch64::ST4WB_8B_fixed,  AArch64::ST4WB_4H_fixed,
      AArch64::ST4WB_2S_fixed,  AArch64::ST1x4WB_1D_fixed,
      AArch64::ST4WB_16B_fixed, AArch64::ST4WB_8H_fixed,
      AArch64::ST4WB_4S_fixed,  AArch64::ST4WB_2D_fixed
    };
    return SelectVST(Node, true, 4, Opcodes);
  }
  case AArch64ISD::NEON_LD2DUP: {
    static const uint16_t Opcodes[] = {
        AArch64::LD2R_8B, AArch64::LD2R_4H, AArch64::LD2R_2S,
        AArch64::LD2R_1D, AArch64::LD2R_16B, AArch64::LD2R_8H,
        AArch64::LD2R_4S, AArch64::LD2R_2D
    };
    return SelectVLDDup(Node, false, 2, Opcodes);
  }
  case AArch64ISD::NEON_LD3DUP: {
    static const uint16_t Opcodes[] = {
        AArch64::LD3R_8B, AArch64::LD3R_4H, AArch64::LD3R_2S,
        AArch64::LD3R_1D, AArch64::LD3R_16B, AArch64::LD3R_8H,
        AArch64::LD3R_4S, AArch64::LD3R_2D
    };
    return SelectVLDDup(Node, false, 3, Opcodes);
  }
  case AArch64ISD::NEON_LD4DUP: {
    static const uint16_t Opcodes[] = {
        AArch64::LD4R_8B, AArch64::LD4R_4H, AArch64::LD4R_2S,
        AArch64::LD4R_1D, AArch64::LD4R_16B, AArch64::LD4R_8H,
        AArch64::LD4R_4S, AArch64::LD4R_2D
    };
    return SelectVLDDup(Node, false, 4, Opcodes);
  }
  case AArch64ISD::NEON_LD2DUP_UPD: {
    static const uint16_t Opcodes[] = {
      AArch64::LD2R_WB_8B_fixed,  AArch64::LD2R_WB_4H_fixed,
      AArch64::LD2R_WB_2S_fixed,  AArch64::LD2R_WB_1D_fixed,
      AArch64::LD2R_WB_16B_fixed, AArch64::LD2R_WB_8H_fixed,
      AArch64::LD2R_WB_4S_fixed,  AArch64::LD2R_WB_2D_fixed
    };
    return SelectVLDDup(Node, true, 2, Opcodes);
  }
  case AArch64ISD::NEON_LD3DUP_UPD: {
    static const uint16_t Opcodes[] = {
      AArch64::LD3R_WB_8B_fixed,  AArch64::LD3R_WB_4H_fixed,
      AArch64::LD3R_WB_2S_fixed,  AArch64::LD3R_WB_1D_fixed,
      AArch64::LD3R_WB_16B_fixed, AArch64::LD3R_WB_8H_fixed,
      AArch64::LD3R_WB_4S_fixed,  AArch64::LD3R_WB_2D_fixed
    };
    return SelectVLDDup(Node, true, 3, Opcodes);
  }
  case AArch64ISD::NEON_LD4DUP_UPD: {
    static const uint16_t Opcodes[] = {
      AArch64::LD4R_WB_8B_fixed,  AArch64::LD4R_WB_4H_fixed,
      AArch64::LD4R_WB_2S_fixed,  AArch64::LD4R_WB_1D_fixed,
      AArch64::LD4R_WB_16B_fixed, AArch64::LD4R_WB_8H_fixed,
      AArch64::LD4R_WB_4S_fixed,  AArch64::LD4R_WB_2D_fixed
    };
    return SelectVLDDup(Node, true, 4, Opcodes);
  }
  case AArch64ISD::NEON_LD2LN_UPD: {
    static const uint16_t Opcodes[] = {
        AArch64::LD2LN_WB_B_fixed, AArch64::LD2LN_WB_H_fixed,
        AArch64::LD2LN_WB_S_fixed, AArch64::LD2LN_WB_D_fixed
    };
    return SelectVLDSTLane(Node, true, true, 2, Opcodes);
  }
  case AArch64ISD::NEON_LD3LN_UPD: {
    static const uint16_t Opcodes[] = {
        AArch64::LD3LN_WB_B_fixed, AArch64::LD3LN_WB_H_fixed,
        AArch64::LD3LN_WB_S_fixed, AArch64::LD3LN_WB_D_fixed
    };
    return SelectVLDSTLane(Node, true, true, 3, Opcodes);
  }
  case AArch64ISD::NEON_LD4LN_UPD: {
    static const uint16_t Opcodes[] = {
        AArch64::LD4LN_WB_B_fixed, AArch64::LD4LN_WB_H_fixed,
        AArch64::LD4LN_WB_S_fixed, AArch64::LD4LN_WB_D_fixed
    };
    return SelectVLDSTLane(Node, true, true, 4, Opcodes);
  }
  case AArch64ISD::NEON_ST2LN_UPD: {
    static const uint16_t Opcodes[] = {
        AArch64::ST2LN_WB_B_fixed, AArch64::ST2LN_WB_H_fixed,
        AArch64::ST2LN_WB_S_fixed, AArch64::ST2LN_WB_D_fixed
    };
    return SelectVLDSTLane(Node, false, true, 2, Opcodes);
  }
  case AArch64ISD::NEON_ST3LN_UPD: {
    static const uint16_t Opcodes[] = {
        AArch64::ST3LN_WB_B_fixed, AArch64::ST3LN_WB_H_fixed,
        AArch64::ST3LN_WB_S_fixed, AArch64::ST3LN_WB_D_fixed
    };
    return SelectVLDSTLane(Node, false, true, 3, Opcodes);
  }
  case AArch64ISD::NEON_ST4LN_UPD: {
    static const uint16_t Opcodes[] = {
        AArch64::ST4LN_WB_B_fixed, AArch64::ST4LN_WB_H_fixed,
        AArch64::ST4LN_WB_S_fixed, AArch64::ST4LN_WB_D_fixed
    };
    return SelectVLDSTLane(Node, false, true, 4, Opcodes);
  }
  case AArch64ISD::NEON_ST1x2_UPD: {
    static const uint16_t Opcodes[] = {
      AArch64::ST1x2WB_8B_fixed,  AArch64::ST1x2WB_4H_fixed,
      AArch64::ST1x2WB_2S_fixed,  AArch64::ST1x2WB_1D_fixed,
      AArch64::ST1x2WB_16B_fixed, AArch64::ST1x2WB_8H_fixed,
      AArch64::ST1x2WB_4S_fixed,  AArch64::ST1x2WB_2D_fixed
    };
    return SelectVST(Node, true, 2, Opcodes);
  }
  case AArch64ISD::NEON_ST1x3_UPD: {
    static const uint16_t Opcodes[] = {
      AArch64::ST1x3WB_8B_fixed,  AArch64::ST1x3WB_4H_fixed,
      AArch64::ST1x3WB_2S_fixed,  AArch64::ST1x3WB_1D_fixed,
      AArch64::ST1x3WB_16B_fixed, AArch64::ST1x3WB_8H_fixed,
      AArch64::ST1x3WB_4S_fixed,  AArch64::ST1x3WB_2D_fixed
    };
    return SelectVST(Node, true, 3, Opcodes);
  }
  case AArch64ISD::NEON_ST1x4_UPD: {
    static const uint16_t Opcodes[] = {
      AArch64::ST1x4WB_8B_fixed,  AArch64::ST1x4WB_4H_fixed,
      AArch64::ST1x4WB_2S_fixed,  AArch64::ST1x4WB_1D_fixed,
      AArch64::ST1x4WB_16B_fixed, AArch64::ST1x4WB_8H_fixed,
      AArch64::ST1x4WB_4S_fixed,  AArch64::ST1x4WB_2D_fixed
    };
    return SelectVST(Node, true, 4, Opcodes);
  }
  case ISD::INTRINSIC_WO_CHAIN: {
    unsigned IntNo = cast<ConstantSDNode>(Node->getOperand(0))->getZExtValue();
    bool IsExt = false;
    switch (IntNo) {
      default:
        break;
      case Intrinsic::aarch64_neon_vtbx1:
        IsExt = true;
      case Intrinsic::aarch64_neon_vtbl1:
        return SelectVTBL(Node, 1, IsExt);
      case Intrinsic::aarch64_neon_vtbx2:
        IsExt = true;
      case Intrinsic::aarch64_neon_vtbl2:
        return SelectVTBL(Node, 2, IsExt);
      case Intrinsic::aarch64_neon_vtbx3:
        IsExt = true;
      case Intrinsic::aarch64_neon_vtbl3:
        return SelectVTBL(Node, 3, IsExt);
      case Intrinsic::aarch64_neon_vtbx4:
        IsExt = true;
      case Intrinsic::aarch64_neon_vtbl4:
        return SelectVTBL(Node, 4, IsExt);
    }
    break;
  }
  case ISD::INTRINSIC_VOID:
  case ISD::INTRINSIC_W_CHAIN: {
    unsigned IntNo = cast<ConstantSDNode>(Node->getOperand(1))->getZExtValue();
    switch (IntNo) {
    default:
      break;
    case Intrinsic::arm_neon_vld1: {
      static const uint16_t Opcodes[] = {
          AArch64::LD1_8B,  AArch64::LD1_4H, AArch64::LD1_2S, AArch64::LD1_1D,
          AArch64::LD1_16B, AArch64::LD1_8H, AArch64::LD1_4S, AArch64::LD1_2D
      };
      return SelectVLD(Node, false, 1, Opcodes);
    }
    case Intrinsic::arm_neon_vld2: {
      static const uint16_t Opcodes[] = {
          AArch64::LD2_8B,  AArch64::LD2_4H, AArch64::LD2_2S, AArch64::LD1x2_1D,
          AArch64::LD2_16B, AArch64::LD2_8H, AArch64::LD2_4S, AArch64::LD2_2D
      };
      return SelectVLD(Node, false, 2, Opcodes);
    }
    case Intrinsic::arm_neon_vld3: {
      static const uint16_t Opcodes[] = {
          AArch64::LD3_8B,  AArch64::LD3_4H, AArch64::LD3_2S, AArch64::LD1x3_1D,
          AArch64::LD3_16B, AArch64::LD3_8H, AArch64::LD3_4S, AArch64::LD3_2D
      };
      return SelectVLD(Node, false, 3, Opcodes);
    }
    case Intrinsic::arm_neon_vld4: {
      static const uint16_t Opcodes[] = {
          AArch64::LD4_8B,  AArch64::LD4_4H, AArch64::LD4_2S, AArch64::LD1x4_1D,
          AArch64::LD4_16B, AArch64::LD4_8H, AArch64::LD4_4S, AArch64::LD4_2D
      };
      return SelectVLD(Node, false, 4, Opcodes);
    }
    case Intrinsic::aarch64_neon_vld1x2: {
      static const uint16_t Opcodes[] = {
          AArch64::LD1x2_8B, AArch64::LD1x2_4H,  AArch64::LD1x2_2S,
          AArch64::LD1x2_1D, AArch64::LD1x2_16B, AArch64::LD1x2_8H,
          AArch64::LD1x2_4S, AArch64::LD1x2_2D
      };
      return SelectVLD(Node, false, 2, Opcodes);
    }
    case Intrinsic::aarch64_neon_vld1x3: {
      static const uint16_t Opcodes[] = {
          AArch64::LD1x3_8B, AArch64::LD1x3_4H,  AArch64::LD1x3_2S,
          AArch64::LD1x3_1D, AArch64::LD1x3_16B, AArch64::LD1x3_8H,
          AArch64::LD1x3_4S, AArch64::LD1x3_2D
      };
      return SelectVLD(Node, false, 3, Opcodes);
    }
    case Intrinsic::aarch64_neon_vld1x4: {
      static const uint16_t Opcodes[] = {
          AArch64::LD1x4_8B, AArch64::LD1x4_4H,  AArch64::LD1x4_2S,
          AArch64::LD1x4_1D, AArch64::LD1x4_16B, AArch64::LD1x4_8H,
          AArch64::LD1x4_4S, AArch64::LD1x4_2D
      };
      return SelectVLD(Node, false, 4, Opcodes);
    }
    case Intrinsic::arm_neon_vst1: {
      static const uint16_t Opcodes[] = {
          AArch64::ST1_8B,  AArch64::ST1_4H, AArch64::ST1_2S, AArch64::ST1_1D,
          AArch64::ST1_16B, AArch64::ST1_8H, AArch64::ST1_4S, AArch64::ST1_2D
      };
      return SelectVST(Node, false, 1, Opcodes);
    }
    case Intrinsic::arm_neon_vst2: {
      static const uint16_t Opcodes[] = {
          AArch64::ST2_8B,  AArch64::ST2_4H, AArch64::ST2_2S, AArch64::ST1x2_1D,
          AArch64::ST2_16B, AArch64::ST2_8H, AArch64::ST2_4S, AArch64::ST2_2D
      };
      return SelectVST(Node, false, 2, Opcodes);
    }
    case Intrinsic::arm_neon_vst3: {
      static const uint16_t Opcodes[] = {
          AArch64::ST3_8B,  AArch64::ST3_4H, AArch64::ST3_2S, AArch64::ST1x3_1D,
          AArch64::ST3_16B, AArch64::ST3_8H, AArch64::ST3_4S, AArch64::ST3_2D
      };
      return SelectVST(Node, false, 3, Opcodes);
    }
    case Intrinsic::arm_neon_vst4: {
      static const uint16_t Opcodes[] = {
          AArch64::ST4_8B,  AArch64::ST4_4H, AArch64::ST4_2S, AArch64::ST1x4_1D,
          AArch64::ST4_16B, AArch64::ST4_8H, AArch64::ST4_4S, AArch64::ST4_2D
      };
      return SelectVST(Node, false, 4, Opcodes);
    }
    case Intrinsic::aarch64_neon_vst1x2: {
      static const uint16_t Opcodes[] = {
          AArch64::ST1x2_8B, AArch64::ST1x2_4H,  AArch64::ST1x2_2S,
          AArch64::ST1x2_1D, AArch64::ST1x2_16B, AArch64::ST1x2_8H,
          AArch64::ST1x2_4S, AArch64::ST1x2_2D
      };
      return SelectVST(Node, false, 2, Opcodes);
    }
    case Intrinsic::aarch64_neon_vst1x3: {
      static const uint16_t Opcodes[] = {
          AArch64::ST1x3_8B, AArch64::ST1x3_4H,  AArch64::ST1x3_2S,
          AArch64::ST1x3_1D, AArch64::ST1x3_16B, AArch64::ST1x3_8H,
          AArch64::ST1x3_4S, AArch64::ST1x3_2D
      };
      return SelectVST(Node, false, 3, Opcodes);
    }
    case Intrinsic::aarch64_neon_vst1x4: {
      static const uint16_t Opcodes[] = {
          AArch64::ST1x4_8B, AArch64::ST1x4_4H,  AArch64::ST1x4_2S,
          AArch64::ST1x4_1D, AArch64::ST1x4_16B, AArch64::ST1x4_8H,
          AArch64::ST1x4_4S, AArch64::ST1x4_2D
      };
      return SelectVST(Node, false, 4, Opcodes);
    }
    case Intrinsic::arm_neon_vld2lane: {
      static const uint16_t Opcodes[] = {
          AArch64::LD2LN_B, AArch64::LD2LN_H, AArch64::LD2LN_S, AArch64::LD2LN_D
      };
      return SelectVLDSTLane(Node, true, false, 2, Opcodes);
    }
    case Intrinsic::arm_neon_vld3lane: {
      static const uint16_t Opcodes[] = {
          AArch64::LD3LN_B, AArch64::LD3LN_H, AArch64::LD3LN_S, AArch64::LD3LN_D
      };
      return SelectVLDSTLane(Node, true, false, 3, Opcodes);
    }
    case Intrinsic::arm_neon_vld4lane: {
      static const uint16_t Opcodes[] = {
          AArch64::LD4LN_B, AArch64::LD4LN_H, AArch64::LD4LN_S, AArch64::LD4LN_D
      };
      return SelectVLDSTLane(Node, true, false, 4, Opcodes);
    }
    case Intrinsic::arm_neon_vst2lane: {
      static const uint16_t Opcodes[] = {
          AArch64::ST2LN_B, AArch64::ST2LN_H, AArch64::ST2LN_S, AArch64::ST2LN_D
      };
      return SelectVLDSTLane(Node, false, false, 2, Opcodes);
    }
    case Intrinsic::arm_neon_vst3lane: {
      static const uint16_t Opcodes[] = {
          AArch64::ST3LN_B, AArch64::ST3LN_H, AArch64::ST3LN_S, AArch64::ST3LN_D
      };
      return SelectVLDSTLane(Node, false, false, 3, Opcodes);
    }
    case Intrinsic::arm_neon_vst4lane: {
      static const uint16_t Opcodes[] = {
          AArch64::ST4LN_B, AArch64::ST4LN_H, AArch64::ST4LN_S, AArch64::ST4LN_D
      };
      return SelectVLDSTLane(Node, false, false, 4, Opcodes);
    }
    } // End of switch IntNo
    break;
  } // End of case ISD::INTRINSIC_VOID and :ISD::INTRINSIC_W_CHAIN
  default:
    break; // Let generic code handle it
  }

  SDNode *ResNode = SelectCode(Node);

  DEBUG(dbgs() << "=> ";
        if (ResNode == NULL || ResNode == Node)
          Node->dump(CurDAG);
        else
          ResNode->dump(CurDAG);
        dbgs() << "\n");

  return ResNode;
}

/// This pass converts a legalized DAG into a AArch64-specific DAG, ready for
/// instruction scheduling.
FunctionPass *llvm::createAArch64ISelDAG(AArch64TargetMachine &TM,
                                         CodeGenOpt::Level OptLevel) {
  return new AArch64DAGToDAGISel(TM, OptLevel);
}