llvm-project/llvm/lib/Target/ARM/ARMISelDAGToDAG.cpp

//===-- ARMISelDAGToDAG.cpp - A dag to dag inst selector for ARM ----------===//
//
//                     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 ARM target.
//
//===----------------------------------------------------------------------===//

#include "ARM.h"
#include "ARMAddressingModes.h"
#include "ARMConstantPoolValue.h"
#include "ARMISelLowering.h"
#include "ARMTargetMachine.h"
#include "llvm/CallingConv.h"
#include "llvm/Constants.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
#include "llvm/Intrinsics.h"
#include "llvm/LLVMContext.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/CodeGen/SelectionDAGISel.h"
#include "llvm/Target/TargetLowering.h"
#include "llvm/Target/TargetOptions.h"
#include "llvm/Support/Compiler.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"

using namespace llvm;

//===--------------------------------------------------------------------===//
/// ARMDAGToDAGISel - ARM specific code to select ARM machine
/// instructions for SelectionDAG operations.
///
namespace {
class ARMDAGToDAGISel : public SelectionDAGISel {
  ARMBaseTargetMachine &TM;

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

public:
  explicit ARMDAGToDAGISel(ARMBaseTargetMachine &tm,
                           CodeGenOpt::Level OptLevel)
    : SelectionDAGISel(tm, OptLevel), TM(tm),
    Subtarget(&TM.getSubtarget<ARMSubtarget>()) {
  }

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

  /// getI32Imm - Return a target constant of type i32 with the specified
  /// value.
  inline SDValue getI32Imm(unsigned Imm) {
    return CurDAG->getTargetConstant(Imm, MVT::i32);
  }

  SDNode *Select(SDValue Op);
  virtual void InstructionSelect();
  bool SelectShifterOperandReg(SDValue Op, SDValue N, SDValue &A,
                               SDValue &B, SDValue &C);
  bool SelectAddrMode2(SDValue Op, SDValue N, SDValue &Base,
                       SDValue &Offset, SDValue &Opc);
  bool SelectAddrMode2Offset(SDValue Op, SDValue N,
                             SDValue &Offset, SDValue &Opc);
  bool SelectAddrMode3(SDValue Op, SDValue N, SDValue &Base,
                       SDValue &Offset, SDValue &Opc);
  bool SelectAddrMode3Offset(SDValue Op, SDValue N,
                             SDValue &Offset, SDValue &Opc);
  bool SelectAddrMode4(SDValue Op, SDValue N, SDValue &Addr,
                       SDValue &Mode);
  bool SelectAddrMode5(SDValue Op, SDValue N, SDValue &Base,
                       SDValue &Offset);
  bool SelectAddrMode6(SDValue Op, SDValue N, SDValue &Addr, SDValue &Update,
                       SDValue &Opc);

  bool SelectAddrModePC(SDValue Op, SDValue N, SDValue &Offset,
                        SDValue &Label);

  bool SelectThumbAddrModeRR(SDValue Op, SDValue N, SDValue &Base,
                             SDValue &Offset);
  bool SelectThumbAddrModeRI5(SDValue Op, SDValue N, unsigned Scale,
                              SDValue &Base, SDValue &OffImm,
                              SDValue &Offset);
  bool SelectThumbAddrModeS1(SDValue Op, SDValue N, SDValue &Base,
                             SDValue &OffImm, SDValue &Offset);
  bool SelectThumbAddrModeS2(SDValue Op, SDValue N, SDValue &Base,
                             SDValue &OffImm, SDValue &Offset);
  bool SelectThumbAddrModeS4(SDValue Op, SDValue N, SDValue &Base,
                             SDValue &OffImm, SDValue &Offset);
  bool SelectThumbAddrModeSP(SDValue Op, SDValue N, SDValue &Base,
                             SDValue &OffImm);

  bool SelectT2ShifterOperandReg(SDValue Op, SDValue N,
                                 SDValue &BaseReg, SDValue &Opc);
  bool SelectT2AddrModeImm12(SDValue Op, SDValue N, SDValue &Base,
                             SDValue &OffImm);
  bool SelectT2AddrModeImm8(SDValue Op, SDValue N, SDValue &Base,
                            SDValue &OffImm);
  bool SelectT2AddrModeImm8Offset(SDValue Op, SDValue N,
                                 SDValue &OffImm);
  bool SelectT2AddrModeImm8s4(SDValue Op, SDValue N, SDValue &Base,
                              SDValue &OffImm);
  bool SelectT2AddrModeSoReg(SDValue Op, SDValue N, SDValue &Base,
                             SDValue &OffReg, SDValue &ShImm);

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

private:
  /// SelectARMIndexedLoad - Indexed (pre/post inc/dec) load matching code for
  /// ARM.
  SDNode *SelectARMIndexedLoad(SDValue Op);
  SDNode *SelectT2IndexedLoad(SDValue Op);

  /// SelectDYN_ALLOC - Select dynamic alloc for Thumb.
  SDNode *SelectDYN_ALLOC(SDValue Op);

  /// SelectV6T2BitfieldExtractOp - Select SBFX/UBFX instructions for ARM.
  SDNode *SelectV6T2BitfieldExtractOp(SDValue Op, unsigned Opc);

  /// SelectInlineAsmMemoryOperand - Implement addressing mode selection for
  /// inline asm expressions.
  virtual bool SelectInlineAsmMemoryOperand(const SDValue &Op,
                                            char ConstraintCode,
                                            std::vector<SDValue> &OutOps);

  /// PairDRegs - Insert a pair of double registers into an implicit def to
  /// form a quad register.
  SDNode *PairDRegs(EVT VT, SDValue V0, SDValue V1);
};
}

/// isInt32Immediate - This method tests to see if the node is a 32-bit constant
/// operand. If so Imm will receive the 32-bit value.
static bool isInt32Immediate(SDNode *N, unsigned &Imm) {
  if (N->getOpcode() == ISD::Constant && N->getValueType(0) == MVT::i32) {
    Imm = cast<ConstantSDNode>(N)->getZExtValue();
    return true;
  }
  return false;
}

// isInt32Immediate - This method tests to see if a constant operand.
// If so Imm will receive the 32 bit value.
static bool isInt32Immediate(SDValue N, unsigned &Imm) {
  return isInt32Immediate(N.getNode(), Imm);
}

// isOpcWithIntImmediate - This method tests to see if the node is a specific
// opcode and that it has a immediate integer right operand.
// If so Imm will receive the 32 bit value.
static bool isOpcWithIntImmediate(SDNode *N, unsigned Opc, unsigned& Imm) {
  return N->getOpcode() == Opc &&
         isInt32Immediate(N->getOperand(1).getNode(), Imm);
}


void ARMDAGToDAGISel::InstructionSelect() {
  DEBUG(BB->dump());

  SelectRoot(*CurDAG);
  CurDAG->RemoveDeadNodes();
}

bool ARMDAGToDAGISel::SelectShifterOperandReg(SDValue Op,
                                              SDValue N,
                                              SDValue &BaseReg,
                                              SDValue &ShReg,
                                              SDValue &Opc) {
  ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);

  // Don't match base register only case. That is matched to a separate
  // lower complexity pattern with explicit register operand.
  if (ShOpcVal == ARM_AM::no_shift) return false;

  BaseReg = N.getOperand(0);
  unsigned ShImmVal = 0;
  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
    ShReg = CurDAG->getRegister(0, MVT::i32);
    ShImmVal = RHS->getZExtValue() & 31;
  } else {
    ShReg = N.getOperand(1);
  }
  Opc = CurDAG->getTargetConstant(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal),
                                  MVT::i32);
  return true;
}

bool ARMDAGToDAGISel::SelectAddrMode2(SDValue Op, SDValue N,
                                      SDValue &Base, SDValue &Offset,
                                      SDValue &Opc) {
  if (N.getOpcode() == ISD::MUL) {
    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
      // X * [3,5,9] -> X + X * [2,4,8] etc.
      int RHSC = (int)RHS->getZExtValue();
      if (RHSC & 1) {
        RHSC = RHSC & ~1;
        ARM_AM::AddrOpc AddSub = ARM_AM::add;
        if (RHSC < 0) {
          AddSub = ARM_AM::sub;
          RHSC = - RHSC;
        }
        if (isPowerOf2_32(RHSC)) {
          unsigned ShAmt = Log2_32(RHSC);
          Base = Offset = N.getOperand(0);
          Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt,
                                                            ARM_AM::lsl),
                                          MVT::i32);
          return true;
        }
      }
    }
  }

  if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
    Base = N;
    if (N.getOpcode() == ISD::FrameIndex) {
      int FI = cast<FrameIndexSDNode>(N)->getIndex();
      Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
    } else if (N.getOpcode() == ARMISD::Wrapper) {
      Base = N.getOperand(0);
    }
    Offset = CurDAG->getRegister(0, MVT::i32);
    Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(ARM_AM::add, 0,
                                                      ARM_AM::no_shift),
                                    MVT::i32);
    return true;
  }

  // Match simple R +/- imm12 operands.
  if (N.getOpcode() == ISD::ADD)
    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
      int RHSC = (int)RHS->getZExtValue();
      if ((RHSC >= 0 && RHSC < 0x1000) ||
          (RHSC < 0 && RHSC > -0x1000)) { // 12 bits.
        Base = N.getOperand(0);
        if (Base.getOpcode() == ISD::FrameIndex) {
          int FI = cast<FrameIndexSDNode>(Base)->getIndex();
          Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
        }
        Offset = CurDAG->getRegister(0, MVT::i32);

        ARM_AM::AddrOpc AddSub = ARM_AM::add;
        if (RHSC < 0) {
          AddSub = ARM_AM::sub;
          RHSC = - RHSC;
        }
        Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, RHSC,
                                                          ARM_AM::no_shift),
                                        MVT::i32);
        return true;
      }
    }

  // Otherwise this is R +/- [possibly shifted] R
  ARM_AM::AddrOpc AddSub = N.getOpcode() == ISD::ADD ? ARM_AM::add:ARM_AM::sub;
  ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(1));
  unsigned ShAmt = 0;

  Base   = N.getOperand(0);
  Offset = N.getOperand(1);

  if (ShOpcVal != ARM_AM::no_shift) {
    // Check to see if the RHS of the shift is a constant, if not, we can't fold
    // it.
    if (ConstantSDNode *Sh =
           dyn_cast<ConstantSDNode>(N.getOperand(1).getOperand(1))) {
      ShAmt = Sh->getZExtValue();
      Offset = N.getOperand(1).getOperand(0);
    } else {
      ShOpcVal = ARM_AM::no_shift;
    }
  }

  // Try matching (R shl C) + (R).
  if (N.getOpcode() == ISD::ADD && ShOpcVal == ARM_AM::no_shift) {
    ShOpcVal = ARM_AM::getShiftOpcForNode(N.getOperand(0));
    if (ShOpcVal != ARM_AM::no_shift) {
      // Check to see if the RHS of the shift is a constant, if not, we can't
      // fold it.
      if (ConstantSDNode *Sh =
          dyn_cast<ConstantSDNode>(N.getOperand(0).getOperand(1))) {
        ShAmt = Sh->getZExtValue();
        Offset = N.getOperand(0).getOperand(0);
        Base = N.getOperand(1);
      } else {
        ShOpcVal = ARM_AM::no_shift;
      }
    }
  }

  Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
                                  MVT::i32);
  return true;
}

bool ARMDAGToDAGISel::SelectAddrMode2Offset(SDValue Op, SDValue N,
                                            SDValue &Offset, SDValue &Opc) {
  unsigned Opcode = Op.getOpcode();
  ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
    ? cast<LoadSDNode>(Op)->getAddressingMode()
    : cast<StoreSDNode>(Op)->getAddressingMode();
  ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
    ? ARM_AM::add : ARM_AM::sub;
  if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
    int Val = (int)C->getZExtValue();
    if (Val >= 0 && Val < 0x1000) { // 12 bits.
      Offset = CurDAG->getRegister(0, MVT::i32);
      Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, Val,
                                                        ARM_AM::no_shift),
                                      MVT::i32);
      return true;
    }
  }

  Offset = N;
  ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);
  unsigned ShAmt = 0;
  if (ShOpcVal != ARM_AM::no_shift) {
    // Check to see if the RHS of the shift is a constant, if not, we can't fold
    // it.
    if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
      ShAmt = Sh->getZExtValue();
      Offset = N.getOperand(0);
    } else {
      ShOpcVal = ARM_AM::no_shift;
    }
  }

  Opc = CurDAG->getTargetConstant(ARM_AM::getAM2Opc(AddSub, ShAmt, ShOpcVal),
                                  MVT::i32);
  return true;
}


bool ARMDAGToDAGISel::SelectAddrMode3(SDValue Op, SDValue N,
                                      SDValue &Base, SDValue &Offset,
                                      SDValue &Opc) {
  if (N.getOpcode() == ISD::SUB) {
    // X - C  is canonicalize to X + -C, no need to handle it here.
    Base = N.getOperand(0);
    Offset = N.getOperand(1);
    Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::sub, 0),MVT::i32);
    return true;
  }

  if (N.getOpcode() != ISD::ADD) {
    Base = N;
    if (N.getOpcode() == ISD::FrameIndex) {
      int FI = cast<FrameIndexSDNode>(N)->getIndex();
      Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
    }
    Offset = CurDAG->getRegister(0, MVT::i32);
    Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0),MVT::i32);
    return true;
  }

  // If the RHS is +/- imm8, fold into addr mode.
  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
    int RHSC = (int)RHS->getZExtValue();
    if ((RHSC >= 0 && RHSC < 256) ||
        (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
      Base = N.getOperand(0);
      if (Base.getOpcode() == ISD::FrameIndex) {
        int FI = cast<FrameIndexSDNode>(Base)->getIndex();
        Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
      }
      Offset = CurDAG->getRegister(0, MVT::i32);

      ARM_AM::AddrOpc AddSub = ARM_AM::add;
      if (RHSC < 0) {
        AddSub = ARM_AM::sub;
        RHSC = - RHSC;
      }
      Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, RHSC),MVT::i32);
      return true;
    }
  }

  Base = N.getOperand(0);
  Offset = N.getOperand(1);
  Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(ARM_AM::add, 0), MVT::i32);
  return true;
}

bool ARMDAGToDAGISel::SelectAddrMode3Offset(SDValue Op, SDValue N,
                                            SDValue &Offset, SDValue &Opc) {
  unsigned Opcode = Op.getOpcode();
  ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
    ? cast<LoadSDNode>(Op)->getAddressingMode()
    : cast<StoreSDNode>(Op)->getAddressingMode();
  ARM_AM::AddrOpc AddSub = (AM == ISD::PRE_INC || AM == ISD::POST_INC)
    ? ARM_AM::add : ARM_AM::sub;
  if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(N)) {
    int Val = (int)C->getZExtValue();
    if (Val >= 0 && Val < 256) {
      Offset = CurDAG->getRegister(0, MVT::i32);
      Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, Val), MVT::i32);
      return true;
    }
  }

  Offset = N;
  Opc = CurDAG->getTargetConstant(ARM_AM::getAM3Opc(AddSub, 0), MVT::i32);
  return true;
}

bool ARMDAGToDAGISel::SelectAddrMode4(SDValue Op, SDValue N,
                                      SDValue &Addr, SDValue &Mode) {
  Addr = N;
  Mode = CurDAG->getTargetConstant(0, MVT::i32);
  return true;
}

bool ARMDAGToDAGISel::SelectAddrMode5(SDValue Op, SDValue N,
                                      SDValue &Base, SDValue &Offset) {
  if (N.getOpcode() != ISD::ADD) {
    Base = N;
    if (N.getOpcode() == ISD::FrameIndex) {
      int FI = cast<FrameIndexSDNode>(N)->getIndex();
      Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
    } else if (N.getOpcode() == ARMISD::Wrapper) {
      Base = N.getOperand(0);
    }
    Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
                                       MVT::i32);
    return true;
  }

  // If the RHS is +/- imm8, fold into addr mode.
  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
    int RHSC = (int)RHS->getZExtValue();
    if ((RHSC & 3) == 0) {  // The constant is implicitly multiplied by 4.
      RHSC >>= 2;
      if ((RHSC >= 0 && RHSC < 256) ||
          (RHSC < 0 && RHSC > -256)) { // note -256 itself isn't allowed.
        Base = N.getOperand(0);
        if (Base.getOpcode() == ISD::FrameIndex) {
          int FI = cast<FrameIndexSDNode>(Base)->getIndex();
          Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
        }

        ARM_AM::AddrOpc AddSub = ARM_AM::add;
        if (RHSC < 0) {
          AddSub = ARM_AM::sub;
          RHSC = - RHSC;
        }
        Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(AddSub, RHSC),
                                           MVT::i32);
        return true;
      }
    }
  }

  Base = N;
  Offset = CurDAG->getTargetConstant(ARM_AM::getAM5Opc(ARM_AM::add, 0),
                                     MVT::i32);
  return true;
}

bool ARMDAGToDAGISel::SelectAddrMode6(SDValue Op, SDValue N,
                                      SDValue &Addr, SDValue &Update,
                                      SDValue &Opc) {
  Addr = N;
  // Default to no writeback.
  Update = CurDAG->getRegister(0, MVT::i32);
  Opc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(false), MVT::i32);
  return true;
}

bool ARMDAGToDAGISel::SelectAddrModePC(SDValue Op, SDValue N,
                                       SDValue &Offset, SDValue &Label) {
  if (N.getOpcode() == ARMISD::PIC_ADD && N.hasOneUse()) {
    Offset = N.getOperand(0);
    SDValue N1 = N.getOperand(1);
    Label  = CurDAG->getTargetConstant(cast<ConstantSDNode>(N1)->getZExtValue(),
                                       MVT::i32);
    return true;
  }
  return false;
}

bool ARMDAGToDAGISel::SelectThumbAddrModeRR(SDValue Op, SDValue N,
                                            SDValue &Base, SDValue &Offset){
  // FIXME dl should come from the parent load or store, not the address
  DebugLoc dl = Op.getDebugLoc();
  if (N.getOpcode() != ISD::ADD) {
    ConstantSDNode *NC = dyn_cast<ConstantSDNode>(N);
    if (!NC || NC->getZExtValue() != 0)
      return false;

    Base = Offset = N;
    return true;
  }

  Base = N.getOperand(0);
  Offset = N.getOperand(1);
  return true;
}

bool
ARMDAGToDAGISel::SelectThumbAddrModeRI5(SDValue Op, SDValue N,
                                        unsigned Scale, SDValue &Base,
                                        SDValue &OffImm, SDValue &Offset) {
  if (Scale == 4) {
    SDValue TmpBase, TmpOffImm;
    if (SelectThumbAddrModeSP(Op, N, TmpBase, TmpOffImm))
      return false;  // We want to select tLDRspi / tSTRspi instead.
    if (N.getOpcode() == ARMISD::Wrapper &&
        N.getOperand(0).getOpcode() == ISD::TargetConstantPool)
      return false;  // We want to select tLDRpci instead.
  }

  if (N.getOpcode() != ISD::ADD) {
    Base = (N.getOpcode() == ARMISD::Wrapper) ? N.getOperand(0) : N;
    Offset = CurDAG->getRegister(0, MVT::i32);
    OffImm = CurDAG->getTargetConstant(0, MVT::i32);
    return true;
  }

  // Thumb does not have [sp, r] address mode.
  RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
  RegisterSDNode *RHSR = dyn_cast<RegisterSDNode>(N.getOperand(1));
  if ((LHSR && LHSR->getReg() == ARM::SP) ||
      (RHSR && RHSR->getReg() == ARM::SP)) {
    Base = N;
    Offset = CurDAG->getRegister(0, MVT::i32);
    OffImm = CurDAG->getTargetConstant(0, MVT::i32);
    return true;
  }

  // If the RHS is + imm5 * scale, fold into addr mode.
  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
    int RHSC = (int)RHS->getZExtValue();
    if ((RHSC & (Scale-1)) == 0) {  // The constant is implicitly multiplied.
      RHSC /= Scale;
      if (RHSC >= 0 && RHSC < 32) {
        Base = N.getOperand(0);
        Offset = CurDAG->getRegister(0, MVT::i32);
        OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
        return true;
      }
    }
  }

  Base = N.getOperand(0);
  Offset = N.getOperand(1);
  OffImm = CurDAG->getTargetConstant(0, MVT::i32);
  return true;
}

bool ARMDAGToDAGISel::SelectThumbAddrModeS1(SDValue Op, SDValue N,
                                            SDValue &Base, SDValue &OffImm,
                                            SDValue &Offset) {
  return SelectThumbAddrModeRI5(Op, N, 1, Base, OffImm, Offset);
}

bool ARMDAGToDAGISel::SelectThumbAddrModeS2(SDValue Op, SDValue N,
                                            SDValue &Base, SDValue &OffImm,
                                            SDValue &Offset) {
  return SelectThumbAddrModeRI5(Op, N, 2, Base, OffImm, Offset);
}

bool ARMDAGToDAGISel::SelectThumbAddrModeS4(SDValue Op, SDValue N,
                                            SDValue &Base, SDValue &OffImm,
                                            SDValue &Offset) {
  return SelectThumbAddrModeRI5(Op, N, 4, Base, OffImm, Offset);
}

bool ARMDAGToDAGISel::SelectThumbAddrModeSP(SDValue Op, SDValue N,
                                           SDValue &Base, SDValue &OffImm) {
  if (N.getOpcode() == ISD::FrameIndex) {
    int FI = cast<FrameIndexSDNode>(N)->getIndex();
    Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
    OffImm = CurDAG->getTargetConstant(0, MVT::i32);
    return true;
  }

  if (N.getOpcode() != ISD::ADD)
    return false;

  RegisterSDNode *LHSR = dyn_cast<RegisterSDNode>(N.getOperand(0));
  if (N.getOperand(0).getOpcode() == ISD::FrameIndex ||
      (LHSR && LHSR->getReg() == ARM::SP)) {
    // If the RHS is + imm8 * scale, fold into addr mode.
    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
      int RHSC = (int)RHS->getZExtValue();
      if ((RHSC & 3) == 0) {  // The constant is implicitly multiplied.
        RHSC >>= 2;
        if (RHSC >= 0 && RHSC < 256) {
          Base = N.getOperand(0);
          if (Base.getOpcode() == ISD::FrameIndex) {
            int FI = cast<FrameIndexSDNode>(Base)->getIndex();
            Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
          }
          OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
          return true;
        }
      }
    }
  }

  return false;
}

bool ARMDAGToDAGISel::SelectT2ShifterOperandReg(SDValue Op, SDValue N,
                                                SDValue &BaseReg,
                                                SDValue &Opc) {
  ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(N);

  // Don't match base register only case. That is matched to a separate
  // lower complexity pattern with explicit register operand.
  if (ShOpcVal == ARM_AM::no_shift) return false;

  BaseReg = N.getOperand(0);
  unsigned ShImmVal = 0;
  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
    ShImmVal = RHS->getZExtValue() & 31;
    Opc = getI32Imm(ARM_AM::getSORegOpc(ShOpcVal, ShImmVal));
    return true;
  }

  return false;
}

bool ARMDAGToDAGISel::SelectT2AddrModeImm12(SDValue Op, SDValue N,
                                            SDValue &Base, SDValue &OffImm) {
  // Match simple R + imm12 operands.

  // Base only.
  if (N.getOpcode() != ISD::ADD && N.getOpcode() != ISD::SUB) {
    if (N.getOpcode() == ISD::FrameIndex) {
      // Match frame index...
      int FI = cast<FrameIndexSDNode>(N)->getIndex();
      Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
      OffImm  = CurDAG->getTargetConstant(0, MVT::i32);
      return true;
    } else if (N.getOpcode() == ARMISD::Wrapper) {
      Base = N.getOperand(0);
      if (Base.getOpcode() == ISD::TargetConstantPool)
        return false;  // We want to select t2LDRpci instead.
    } else
      Base = N;
    OffImm  = CurDAG->getTargetConstant(0, MVT::i32);
    return true;
  }

  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
    if (SelectT2AddrModeImm8(Op, N, Base, OffImm))
      // Let t2LDRi8 handle (R - imm8).
      return false;

    int RHSC = (int)RHS->getZExtValue();
    if (N.getOpcode() == ISD::SUB)
      RHSC = -RHSC;

    if (RHSC >= 0 && RHSC < 0x1000) { // 12 bits (unsigned)
      Base   = N.getOperand(0);
      if (Base.getOpcode() == ISD::FrameIndex) {
        int FI = cast<FrameIndexSDNode>(Base)->getIndex();
        Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
      }
      OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
      return true;
    }
  }

  // Base only.
  Base = N;
  OffImm  = CurDAG->getTargetConstant(0, MVT::i32);
  return true;
}

bool ARMDAGToDAGISel::SelectT2AddrModeImm8(SDValue Op, SDValue N,
                                           SDValue &Base, SDValue &OffImm) {
  // Match simple R - imm8 operands.
  if (N.getOpcode() == ISD::ADD || N.getOpcode() == ISD::SUB) {
    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
      int RHSC = (int)RHS->getSExtValue();
      if (N.getOpcode() == ISD::SUB)
        RHSC = -RHSC;

      if ((RHSC >= -255) && (RHSC < 0)) { // 8 bits (always negative)
        Base = N.getOperand(0);
        if (Base.getOpcode() == ISD::FrameIndex) {
          int FI = cast<FrameIndexSDNode>(Base)->getIndex();
          Base = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
        }
        OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
        return true;
      }
    }
  }

  return false;
}

bool ARMDAGToDAGISel::SelectT2AddrModeImm8Offset(SDValue Op, SDValue N,
                                                 SDValue &OffImm){
  unsigned Opcode = Op.getOpcode();
  ISD::MemIndexedMode AM = (Opcode == ISD::LOAD)
    ? cast<LoadSDNode>(Op)->getAddressingMode()
    : cast<StoreSDNode>(Op)->getAddressingMode();
  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N)) {
    int RHSC = (int)RHS->getZExtValue();
    if (RHSC >= 0 && RHSC < 0x100) { // 8 bits.
      OffImm = ((AM == ISD::PRE_INC) || (AM == ISD::POST_INC))
        ? CurDAG->getTargetConstant(RHSC, MVT::i32)
        : CurDAG->getTargetConstant(-RHSC, MVT::i32);
      return true;
    }
  }

  return false;
}

bool ARMDAGToDAGISel::SelectT2AddrModeImm8s4(SDValue Op, SDValue N,
                                             SDValue &Base, SDValue &OffImm) {
  if (N.getOpcode() == ISD::ADD) {
    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
      int RHSC = (int)RHS->getZExtValue();
      if (((RHSC & 0x3) == 0) &&
          ((RHSC >= 0 && RHSC < 0x400) || (RHSC < 0 && RHSC > -0x400))) { // 8 bits.
        Base   = N.getOperand(0);
        OffImm = CurDAG->getTargetConstant(RHSC, MVT::i32);
        return true;
      }
    }
  } else if (N.getOpcode() == ISD::SUB) {
    if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
      int RHSC = (int)RHS->getZExtValue();
      if (((RHSC & 0x3) == 0) && (RHSC >= 0 && RHSC < 0x400)) { // 8 bits.
        Base   = N.getOperand(0);
        OffImm = CurDAG->getTargetConstant(-RHSC, MVT::i32);
        return true;
      }
    }
  }

  return false;
}

bool ARMDAGToDAGISel::SelectT2AddrModeSoReg(SDValue Op, SDValue N,
                                            SDValue &Base,
                                            SDValue &OffReg, SDValue &ShImm) {
  // (R - imm8) should be handled by t2LDRi8. The rest are handled by t2LDRi12.
  if (N.getOpcode() != ISD::ADD)
    return false;

  // Leave (R + imm12) for t2LDRi12, (R - imm8) for t2LDRi8.
  if (ConstantSDNode *RHS = dyn_cast<ConstantSDNode>(N.getOperand(1))) {
    int RHSC = (int)RHS->getZExtValue();
    if (RHSC >= 0 && RHSC < 0x1000) // 12 bits (unsigned)
      return false;
    else if (RHSC < 0 && RHSC >= -255) // 8 bits
      return false;
  }

  // Look for (R + R) or (R + (R << [1,2,3])).
  unsigned ShAmt = 0;
  Base   = N.getOperand(0);
  OffReg = N.getOperand(1);

  // Swap if it is ((R << c) + R).
  ARM_AM::ShiftOpc ShOpcVal = ARM_AM::getShiftOpcForNode(OffReg);
  if (ShOpcVal != ARM_AM::lsl) {
    ShOpcVal = ARM_AM::getShiftOpcForNode(Base);
    if (ShOpcVal == ARM_AM::lsl)
      std::swap(Base, OffReg);
  }

  if (ShOpcVal == ARM_AM::lsl) {
    // Check to see if the RHS of the shift is a constant, if not, we can't fold
    // it.
    if (ConstantSDNode *Sh = dyn_cast<ConstantSDNode>(OffReg.getOperand(1))) {
      ShAmt = Sh->getZExtValue();
      if (ShAmt >= 4) {
        ShAmt = 0;
        ShOpcVal = ARM_AM::no_shift;
      } else
        OffReg = OffReg.getOperand(0);
    } else {
      ShOpcVal = ARM_AM::no_shift;
    }
  }

  ShImm = CurDAG->getTargetConstant(ShAmt, MVT::i32);

  return true;
}

//===--------------------------------------------------------------------===//

/// getAL - Returns a ARMCC::AL immediate node.
static inline SDValue getAL(SelectionDAG *CurDAG) {
  return CurDAG->getTargetConstant((uint64_t)ARMCC::AL, MVT::i32);
}

SDNode *ARMDAGToDAGISel::SelectARMIndexedLoad(SDValue Op) {
  LoadSDNode *LD = cast<LoadSDNode>(Op);
  ISD::MemIndexedMode AM = LD->getAddressingMode();
  if (AM == ISD::UNINDEXED)
    return NULL;

  EVT LoadedVT = LD->getMemoryVT();
  SDValue Offset, AMOpc;
  bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
  unsigned Opcode = 0;
  bool Match = false;
  if (LoadedVT == MVT::i32 &&
      SelectAddrMode2Offset(Op, LD->getOffset(), Offset, AMOpc)) {
    Opcode = isPre ? ARM::LDR_PRE : ARM::LDR_POST;
    Match = true;
  } else if (LoadedVT == MVT::i16 &&
             SelectAddrMode3Offset(Op, LD->getOffset(), Offset, AMOpc)) {
    Match = true;
    Opcode = (LD->getExtensionType() == ISD::SEXTLOAD)
      ? (isPre ? ARM::LDRSH_PRE : ARM::LDRSH_POST)
      : (isPre ? ARM::LDRH_PRE : ARM::LDRH_POST);
  } else if (LoadedVT == MVT::i8 || LoadedVT == MVT::i1) {
    if (LD->getExtensionType() == ISD::SEXTLOAD) {
      if (SelectAddrMode3Offset(Op, LD->getOffset(), Offset, AMOpc)) {
        Match = true;
        Opcode = isPre ? ARM::LDRSB_PRE : ARM::LDRSB_POST;
      }
    } else {
      if (SelectAddrMode2Offset(Op, LD->getOffset(), Offset, AMOpc)) {
        Match = true;
        Opcode = isPre ? ARM::LDRB_PRE : ARM::LDRB_POST;
      }
    }
  }

  if (Match) {
    SDValue Chain = LD->getChain();
    SDValue Base = LD->getBasePtr();
    SDValue Ops[]= { Base, Offset, AMOpc, getAL(CurDAG),
                     CurDAG->getRegister(0, MVT::i32), Chain };
    return CurDAG->getMachineNode(Opcode, Op.getDebugLoc(), MVT::i32, MVT::i32,
                                  MVT::Other, Ops, 6);
  }

  return NULL;
}

SDNode *ARMDAGToDAGISel::SelectT2IndexedLoad(SDValue Op) {
  LoadSDNode *LD = cast<LoadSDNode>(Op);
  ISD::MemIndexedMode AM = LD->getAddressingMode();
  if (AM == ISD::UNINDEXED)
    return NULL;

  EVT LoadedVT = LD->getMemoryVT();
  bool isSExtLd = LD->getExtensionType() == ISD::SEXTLOAD;
  SDValue Offset;
  bool isPre = (AM == ISD::PRE_INC) || (AM == ISD::PRE_DEC);
  unsigned Opcode = 0;
  bool Match = false;
  if (SelectT2AddrModeImm8Offset(Op, LD->getOffset(), Offset)) {
    switch (LoadedVT.getSimpleVT().SimpleTy) {
    case MVT::i32:
      Opcode = isPre ? ARM::t2LDR_PRE : ARM::t2LDR_POST;
      break;
    case MVT::i16:
      if (isSExtLd)
        Opcode = isPre ? ARM::t2LDRSH_PRE : ARM::t2LDRSH_POST;
      else
        Opcode = isPre ? ARM::t2LDRH_PRE : ARM::t2LDRH_POST;
      break;
    case MVT::i8:
    case MVT::i1:
      if (isSExtLd)
        Opcode = isPre ? ARM::t2LDRSB_PRE : ARM::t2LDRSB_POST;
      else
        Opcode = isPre ? ARM::t2LDRB_PRE : ARM::t2LDRB_POST;
      break;
    default:
      return NULL;
    }
    Match = true;
  }

  if (Match) {
    SDValue Chain = LD->getChain();
    SDValue Base = LD->getBasePtr();
    SDValue Ops[]= { Base, Offset, getAL(CurDAG),
                     CurDAG->getRegister(0, MVT::i32), Chain };
    return CurDAG->getMachineNode(Opcode, Op.getDebugLoc(), MVT::i32, MVT::i32,
                                  MVT::Other, Ops, 5);
  }

  return NULL;
}

SDNode *ARMDAGToDAGISel::SelectDYN_ALLOC(SDValue Op) {
  SDNode *N = Op.getNode();
  DebugLoc dl = N->getDebugLoc();
  EVT VT = Op.getValueType();
  SDValue Chain = Op.getOperand(0);
  SDValue Size = Op.getOperand(1);
  SDValue Align = Op.getOperand(2);
  SDValue SP = CurDAG->getRegister(ARM::SP, MVT::i32);
  int32_t AlignVal = cast<ConstantSDNode>(Align)->getSExtValue();
  if (AlignVal < 0)
    // We need to align the stack. Use Thumb1 tAND which is the only thumb
    // instruction that can read and write SP. This matches to a pseudo
    // instruction that has a chain to ensure the result is written back to
    // the stack pointer.
    SP = SDValue(CurDAG->getMachineNode(ARM::tANDsp, dl, VT, SP, Align), 0);

  bool isC = isa<ConstantSDNode>(Size);
  uint32_t C = isC ? cast<ConstantSDNode>(Size)->getZExtValue() : ~0UL;
  // Handle the most common case for both Thumb1 and Thumb2:
  // tSUBspi - immediate is between 0 ... 508 inclusive.
  if (C <= 508 && ((C & 3) == 0))
    // FIXME: tSUBspi encode scale 4 implicitly.
    return CurDAG->SelectNodeTo(N, ARM::tSUBspi_, VT, MVT::Other, SP,
                                CurDAG->getTargetConstant(C/4, MVT::i32),
                                Chain);

  if (Subtarget->isThumb1Only()) {
    // Use tADDspr since Thumb1 does not have a sub r, sp, r. ARMISelLowering
    // should have negated the size operand already. FIXME: We can't insert
    // new target independent node at this stage so we are forced to negate
    // it earlier. Is there a better solution?
    return CurDAG->SelectNodeTo(N, ARM::tADDspr_, VT, MVT::Other, SP, Size,
                                Chain);
  } else if (Subtarget->isThumb2()) {
    if (isC && Predicate_t2_so_imm(Size.getNode())) {
      // t2SUBrSPi
      SDValue Ops[] = { SP, CurDAG->getTargetConstant(C, MVT::i32), Chain };
      return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPi_, VT, MVT::Other, Ops, 3);
    } else if (isC && Predicate_imm0_4095(Size.getNode())) {
      // t2SUBrSPi12
      SDValue Ops[] = { SP, CurDAG->getTargetConstant(C, MVT::i32), Chain };
      return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPi12_, VT, MVT::Other, Ops, 3);
    } else {
      // t2SUBrSPs
      SDValue Ops[] = { SP, Size,
                        getI32Imm(ARM_AM::getSORegOpc(ARM_AM::lsl,0)), Chain };
      return CurDAG->SelectNodeTo(N, ARM::t2SUBrSPs_, VT, MVT::Other, Ops, 4);
    }
  }

  // FIXME: Add ADD / SUB sp instructions for ARM.
  return 0;
}

/// PairDRegs - Insert a pair of double registers into an implicit def to
/// form a quad register.
SDNode *ARMDAGToDAGISel::PairDRegs(EVT VT, SDValue V0, SDValue V1) {
  DebugLoc dl = V0.getNode()->getDebugLoc();
  SDValue Undef =
    SDValue(CurDAG->getMachineNode(TargetInstrInfo::IMPLICIT_DEF, dl, VT), 0);
  SDValue SubReg0 = CurDAG->getTargetConstant(ARM::DSUBREG_0, MVT::i32);
  SDValue SubReg1 = CurDAG->getTargetConstant(ARM::DSUBREG_1, MVT::i32);
  SDNode *Pair = CurDAG->getMachineNode(TargetInstrInfo::INSERT_SUBREG, dl,
                                        VT, Undef, V0, SubReg0);
  return CurDAG->getMachineNode(TargetInstrInfo::INSERT_SUBREG, dl,
                                VT, SDValue(Pair, 0), V1, SubReg1);
}

SDNode *ARMDAGToDAGISel::SelectV6T2BitfieldExtractOp(SDValue Op,
                                                     unsigned Opc) {
  if (!Subtarget->hasV6T2Ops())
    return NULL;
    
  unsigned Shl_imm = 0;
  if (isOpcWithIntImmediate(Op.getOperand(0).getNode(), ISD::SHL, Shl_imm)){
    assert(Shl_imm > 0 && Shl_imm < 32 && "bad amount in shift node!");
    unsigned Srl_imm = 0;
    if (isInt32Immediate(Op.getOperand(1), Srl_imm)) {
      assert(Srl_imm > 0 && Srl_imm < 32 && "bad amount in shift node!");
      unsigned Width = 32 - Srl_imm;
      int LSB = Srl_imm - Shl_imm;
      if ((LSB + Width) > 32)
        return NULL;
      SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
      SDValue Ops[] = { Op.getOperand(0).getOperand(0),
                        CurDAG->getTargetConstant(LSB, MVT::i32),
                        CurDAG->getTargetConstant(Width, MVT::i32),
                        getAL(CurDAG), Reg0 };
      return CurDAG->SelectNodeTo(Op.getNode(), Opc, MVT::i32, Ops, 5);
    }
  }
  return NULL;
}

SDNode *ARMDAGToDAGISel::Select(SDValue Op) {
  SDNode *N = Op.getNode();
  DebugLoc dl = N->getDebugLoc();

  if (N->isMachineOpcode())
    return NULL;   // Already selected.

  switch (N->getOpcode()) {
  default: break;
  case ISD::Constant: {
    unsigned Val = cast<ConstantSDNode>(N)->getZExtValue();
    bool UseCP = true;
    if (Subtarget->hasThumb2())
      // Thumb2-aware targets have the MOVT instruction, so all immediates can
      // be done with MOV + MOVT, at worst.
      UseCP = 0;
    else {
      if (Subtarget->isThumb()) {
        UseCP = (Val > 255 &&                          // MOV
                 ~Val > 255 &&                         // MOV + MVN
                 !ARM_AM::isThumbImmShiftedVal(Val));  // MOV + LSL
      } else
        UseCP = (ARM_AM::getSOImmVal(Val) == -1 &&     // MOV
                 ARM_AM::getSOImmVal(~Val) == -1 &&    // MVN
                 !ARM_AM::isSOImmTwoPartVal(Val));     // two instrs.
    }

    if (UseCP) {
      SDValue CPIdx =
        CurDAG->getTargetConstantPool(ConstantInt::get(
                                  Type::getInt32Ty(*CurDAG->getContext()), Val),
                                      TLI.getPointerTy());

      SDNode *ResNode;
      if (Subtarget->isThumb1Only()) {
        SDValue Pred = CurDAG->getTargetConstant(0xEULL, MVT::i32);
        SDValue PredReg = CurDAG->getRegister(0, MVT::i32);
        SDValue Ops[] = { CPIdx, Pred, PredReg, CurDAG->getEntryNode() };
        ResNode = CurDAG->getMachineNode(ARM::tLDRcp, dl, MVT::i32, MVT::Other,
                                         Ops, 4);
      } else {
        SDValue Ops[] = {
          CPIdx,
          CurDAG->getRegister(0, MVT::i32),
          CurDAG->getTargetConstant(0, MVT::i32),
          getAL(CurDAG),
          CurDAG->getRegister(0, MVT::i32),
          CurDAG->getEntryNode()
        };
        ResNode=CurDAG->getMachineNode(ARM::LDRcp, dl, MVT::i32, MVT::Other,
                                       Ops, 6);
      }
      ReplaceUses(Op, SDValue(ResNode, 0));
      return NULL;
    }

    // Other cases are autogenerated.
    break;
  }
  case ISD::FrameIndex: {
    // Selects to ADDri FI, 0 which in turn will become ADDri SP, imm.
    int FI = cast<FrameIndexSDNode>(N)->getIndex();
    SDValue TFI = CurDAG->getTargetFrameIndex(FI, TLI.getPointerTy());
    if (Subtarget->isThumb1Only()) {
      return CurDAG->SelectNodeTo(N, ARM::tADDrSPi, MVT::i32, TFI,
                                  CurDAG->getTargetConstant(0, MVT::i32));
    } else {
      unsigned Opc = ((Subtarget->isThumb() && Subtarget->hasThumb2()) ?
                      ARM::t2ADDri : ARM::ADDri);
      SDValue Ops[] = { TFI, CurDAG->getTargetConstant(0, MVT::i32),
                        getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
                        CurDAG->getRegister(0, MVT::i32) };
      return CurDAG->SelectNodeTo(N, Opc, MVT::i32, Ops, 5);
    }
  }
  case ARMISD::DYN_ALLOC:
    return SelectDYN_ALLOC(Op);
  case ISD::SRL:
    if (SDNode *I = SelectV6T2BitfieldExtractOp(Op,
                      Subtarget->isThumb() ? ARM::t2UBFX : ARM::UBFX))
      return I;
    break;
  case ISD::SRA:
    if (SDNode *I = SelectV6T2BitfieldExtractOp(Op,
                      Subtarget->isThumb() ? ARM::t2SBFX : ARM::SBFX))
      return I;
    break;
  case ISD::MUL:
    if (Subtarget->isThumb1Only())
      break;
    if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
      unsigned RHSV = C->getZExtValue();
      if (!RHSV) break;
      if (isPowerOf2_32(RHSV-1)) {  // 2^n+1?
        unsigned ShImm = Log2_32(RHSV-1);
        if (ShImm >= 32)
          break;
        SDValue V = Op.getOperand(0);
        ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
        SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
        SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
        if (Subtarget->isThumb()) {
          SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
          return CurDAG->SelectNodeTo(N, ARM::t2ADDrs, MVT::i32, Ops, 6);
        } else {
          SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
          return CurDAG->SelectNodeTo(N, ARM::ADDrs, MVT::i32, Ops, 7);
        }
      }
      if (isPowerOf2_32(RHSV+1)) {  // 2^n-1?
        unsigned ShImm = Log2_32(RHSV+1);
        if (ShImm >= 32)
          break;
        SDValue V = Op.getOperand(0);
        ShImm = ARM_AM::getSORegOpc(ARM_AM::lsl, ShImm);
        SDValue ShImmOp = CurDAG->getTargetConstant(ShImm, MVT::i32);
        SDValue Reg0 = CurDAG->getRegister(0, MVT::i32);
        if (Subtarget->isThumb()) {
          SDValue Ops[] = { V, V, ShImmOp, getAL(CurDAG), Reg0 };
          return CurDAG->SelectNodeTo(N, ARM::t2RSBrs, MVT::i32, Ops, 5);
        } else {
          SDValue Ops[] = { V, V, Reg0, ShImmOp, getAL(CurDAG), Reg0, Reg0 };
          return CurDAG->SelectNodeTo(N, ARM::RSBrs, MVT::i32, Ops, 7);
        }
      }
    }
    break;
  case ARMISD::FMRRD:
    return CurDAG->getMachineNode(ARM::FMRRD, dl, MVT::i32, MVT::i32,
                                  Op.getOperand(0), getAL(CurDAG),
                                  CurDAG->getRegister(0, MVT::i32));
  case ISD::UMUL_LOHI: {
    if (Subtarget->isThumb1Only())
      break;
    if (Subtarget->isThumb()) {
      SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
                        getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
                        CurDAG->getRegister(0, MVT::i32) };
      return CurDAG->getMachineNode(ARM::t2UMULL, dl, MVT::i32, MVT::i32, Ops,4);
    } else {
      SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
                        getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
                        CurDAG->getRegister(0, MVT::i32) };
      return CurDAG->getMachineNode(ARM::UMULL, dl, MVT::i32, MVT::i32, Ops, 5);
    }
  }
  case ISD::SMUL_LOHI: {
    if (Subtarget->isThumb1Only())
      break;
    if (Subtarget->isThumb()) {
      SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
                        getAL(CurDAG), CurDAG->getRegister(0, MVT::i32) };
      return CurDAG->getMachineNode(ARM::t2SMULL, dl, MVT::i32, MVT::i32, Ops,4);
    } else {
      SDValue Ops[] = { Op.getOperand(0), Op.getOperand(1),
                        getAL(CurDAG), CurDAG->getRegister(0, MVT::i32),
                        CurDAG->getRegister(0, MVT::i32) };
      return CurDAG->getMachineNode(ARM::SMULL, dl, MVT::i32, MVT::i32, Ops, 5);
    }
  }
  case ISD::LOAD: {
    SDNode *ResNode = 0;
    if (Subtarget->isThumb() && Subtarget->hasThumb2())
      ResNode = SelectT2IndexedLoad(Op);
    else
      ResNode = SelectARMIndexedLoad(Op);
    if (ResNode)
      return ResNode;
    // Other cases are autogenerated.
    break;
  }
  case ARMISD::BRCOND: {
    // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
    // Emits: (Bcc:void (bb:Other):$dst, (imm:i32):$cc)
    // Pattern complexity = 6  cost = 1  size = 0

    // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
    // Emits: (tBcc:void (bb:Other):$dst, (imm:i32):$cc)
    // Pattern complexity = 6  cost = 1  size = 0

    // Pattern: (ARMbrcond:void (bb:Other):$dst, (imm:i32):$cc)
    // Emits: (t2Bcc:void (bb:Other):$dst, (imm:i32):$cc)
    // Pattern complexity = 6  cost = 1  size = 0

    unsigned Opc = Subtarget->isThumb() ?
      ((Subtarget->hasThumb2()) ? ARM::t2Bcc : ARM::tBcc) : ARM::Bcc;
    SDValue Chain = Op.getOperand(0);
    SDValue N1 = Op.getOperand(1);
    SDValue N2 = Op.getOperand(2);
    SDValue N3 = Op.getOperand(3);
    SDValue InFlag = Op.getOperand(4);
    assert(N1.getOpcode() == ISD::BasicBlock);
    assert(N2.getOpcode() == ISD::Constant);
    assert(N3.getOpcode() == ISD::Register);

    SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
                               cast<ConstantSDNode>(N2)->getZExtValue()),
                               MVT::i32);
    SDValue Ops[] = { N1, Tmp2, N3, Chain, InFlag };
    SDNode *ResNode = CurDAG->getMachineNode(Opc, dl, MVT::Other,
                                             MVT::Flag, Ops, 5);
    Chain = SDValue(ResNode, 0);
    if (Op.getNode()->getNumValues() == 2) {
      InFlag = SDValue(ResNode, 1);
      ReplaceUses(SDValue(Op.getNode(), 1), InFlag);
    }
    ReplaceUses(SDValue(Op.getNode(), 0), SDValue(Chain.getNode(), Chain.getResNo()));
    return NULL;
  }
  case ARMISD::CMOV: {
    EVT VT = Op.getValueType();
    SDValue N0 = Op.getOperand(0);
    SDValue N1 = Op.getOperand(1);
    SDValue N2 = Op.getOperand(2);
    SDValue N3 = Op.getOperand(3);
    SDValue InFlag = Op.getOperand(4);
    assert(N2.getOpcode() == ISD::Constant);
    assert(N3.getOpcode() == ISD::Register);

    if (!Subtarget->isThumb1Only() && VT == MVT::i32) {
      // Pattern: (ARMcmov:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
      // Emits: (MOVCCs:i32 GPR:i32:$false, so_reg:i32:$true, (imm:i32):$cc)
      // Pattern complexity = 18  cost = 1  size = 0
      SDValue CPTmp0;
      SDValue CPTmp1;
      SDValue CPTmp2;
      if (Subtarget->isThumb()) {
        if (SelectT2ShifterOperandReg(Op, N1, CPTmp0, CPTmp1)) {
          unsigned SOVal = cast<ConstantSDNode>(CPTmp1)->getZExtValue();
          unsigned SOShOp = ARM_AM::getSORegShOp(SOVal);
          unsigned Opc = 0;
          switch (SOShOp) {
          case ARM_AM::lsl: Opc = ARM::t2MOVCClsl; break;
          case ARM_AM::lsr: Opc = ARM::t2MOVCClsr; break;
          case ARM_AM::asr: Opc = ARM::t2MOVCCasr; break;
          case ARM_AM::ror: Opc = ARM::t2MOVCCror; break;
          default:
            llvm_unreachable("Unknown so_reg opcode!");
            break;
          }
          SDValue SOShImm =
            CurDAG->getTargetConstant(ARM_AM::getSORegOffset(SOVal), MVT::i32);
          SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
                                   cast<ConstantSDNode>(N2)->getZExtValue()),
                                   MVT::i32);
          SDValue Ops[] = { N0, CPTmp0, SOShImm, Tmp2, N3, InFlag };
          return CurDAG->SelectNodeTo(Op.getNode(), Opc, MVT::i32,Ops, 6);
        }
      } else {
        if (SelectShifterOperandReg(Op, N1, CPTmp0, CPTmp1, CPTmp2)) {
          SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
                                   cast<ConstantSDNode>(N2)->getZExtValue()),
                                   MVT::i32);
          SDValue Ops[] = { N0, CPTmp0, CPTmp1, CPTmp2, Tmp2, N3, InFlag };
          return CurDAG->SelectNodeTo(Op.getNode(),
                                      ARM::MOVCCs, MVT::i32, Ops, 7);
        }
      }

      // Pattern: (ARMcmov:i32 GPR:i32:$false,
      //             (imm:i32)<<P:Predicate_so_imm>>:$true,
      //             (imm:i32):$cc)
      // Emits: (MOVCCi:i32 GPR:i32:$false,
      //           (so_imm:i32 (imm:i32):$true), (imm:i32):$cc)
      // Pattern complexity = 10  cost = 1  size = 0
      if (N3.getOpcode() == ISD::Constant) {
        if (Subtarget->isThumb()) {
          if (Predicate_t2_so_imm(N3.getNode())) {
            SDValue Tmp1 = CurDAG->getTargetConstant(((unsigned)
                                     cast<ConstantSDNode>(N1)->getZExtValue()),
                                     MVT::i32);
            SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
                                     cast<ConstantSDNode>(N2)->getZExtValue()),
                                     MVT::i32);
            SDValue Ops[] = { N0, Tmp1, Tmp2, N3, InFlag };
            return CurDAG->SelectNodeTo(Op.getNode(),
                                        ARM::t2MOVCCi, MVT::i32, Ops, 5);
          }
        } else {
          if (Predicate_so_imm(N3.getNode())) {
            SDValue Tmp1 = CurDAG->getTargetConstant(((unsigned)
                                     cast<ConstantSDNode>(N1)->getZExtValue()),
                                     MVT::i32);
            SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
                                     cast<ConstantSDNode>(N2)->getZExtValue()),
                                     MVT::i32);
            SDValue Ops[] = { N0, Tmp1, Tmp2, N3, InFlag };
            return CurDAG->SelectNodeTo(Op.getNode(),
                                        ARM::MOVCCi, MVT::i32, Ops, 5);
          }
        }
      }
    }

    // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
    // Emits: (MOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
    // Pattern complexity = 6  cost = 1  size = 0
    //
    // Pattern: (ARMcmov:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
    // Emits: (tMOVCCr:i32 GPR:i32:$false, GPR:i32:$true, (imm:i32):$cc)
    // Pattern complexity = 6  cost = 11  size = 0
    //
    // Also FCPYScc and FCPYDcc.
    SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
                               cast<ConstantSDNode>(N2)->getZExtValue()),
                               MVT::i32);
    SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
    unsigned Opc = 0;
    switch (VT.getSimpleVT().SimpleTy) {
    default: assert(false && "Illegal conditional move type!");
      break;
    case MVT::i32:
      Opc = Subtarget->isThumb()
        ? (Subtarget->hasThumb2() ? ARM::t2MOVCCr : ARM::tMOVCCr_pseudo)
        : ARM::MOVCCr;
      break;
    case MVT::f32:
      Opc = ARM::FCPYScc;
      break;
    case MVT::f64:
      Opc = ARM::FCPYDcc;
      break;
    }
    return CurDAG->SelectNodeTo(Op.getNode(), Opc, VT, Ops, 5);
  }
  case ARMISD::CNEG: {
    EVT VT = Op.getValueType();
    SDValue N0 = Op.getOperand(0);
    SDValue N1 = Op.getOperand(1);
    SDValue N2 = Op.getOperand(2);
    SDValue N3 = Op.getOperand(3);
    SDValue InFlag = Op.getOperand(4);
    assert(N2.getOpcode() == ISD::Constant);
    assert(N3.getOpcode() == ISD::Register);

    SDValue Tmp2 = CurDAG->getTargetConstant(((unsigned)
                               cast<ConstantSDNode>(N2)->getZExtValue()),
                               MVT::i32);
    SDValue Ops[] = { N0, N1, Tmp2, N3, InFlag };
    unsigned Opc = 0;
    switch (VT.getSimpleVT().SimpleTy) {
    default: assert(false && "Illegal conditional move type!");
      break;
    case MVT::f32:
      Opc = ARM::FNEGScc;
      break;
    case MVT::f64:
      Opc = ARM::FNEGDcc;
      break;
    }
    return CurDAG->SelectNodeTo(Op.getNode(), Opc, VT, Ops, 5);
  }

  case ARMISD::VZIP: {
    unsigned Opc = 0;
    EVT VT = N->getValueType(0);
    switch (VT.getSimpleVT().SimpleTy) {
    default: return NULL;
    case MVT::v8i8:  Opc = ARM::VZIPd8; break;
    case MVT::v4i16: Opc = ARM::VZIPd16; break;
    case MVT::v2f32:
    case MVT::v2i32: Opc = ARM::VZIPd32; break;
    case MVT::v16i8: Opc = ARM::VZIPq8; break;
    case MVT::v8i16: Opc = ARM::VZIPq16; break;
    case MVT::v4f32:
    case MVT::v4i32: Opc = ARM::VZIPq32; break;
    }
    return CurDAG->getMachineNode(Opc, dl, VT, VT,
                                  N->getOperand(0), N->getOperand(1));
  }
  case ARMISD::VUZP: {
    unsigned Opc = 0;
    EVT VT = N->getValueType(0);
    switch (VT.getSimpleVT().SimpleTy) {
    default: return NULL;
    case MVT::v8i8:  Opc = ARM::VUZPd8; break;
    case MVT::v4i16: Opc = ARM::VUZPd16; break;
    case MVT::v2f32:
    case MVT::v2i32: Opc = ARM::VUZPd32; break;
    case MVT::v16i8: Opc = ARM::VUZPq8; break;
    case MVT::v8i16: Opc = ARM::VUZPq16; break;
    case MVT::v4f32:
    case MVT::v4i32: Opc = ARM::VUZPq32; break;
    }
    return CurDAG->getMachineNode(Opc, dl, VT, VT,
                                  N->getOperand(0), N->getOperand(1));
  }
  case ARMISD::VTRN: {
    unsigned Opc = 0;
    EVT VT = N->getValueType(0);
    switch (VT.getSimpleVT().SimpleTy) {
    default: return NULL;
    case MVT::v8i8:  Opc = ARM::VTRNd8; break;
    case MVT::v4i16: Opc = ARM::VTRNd16; break;
    case MVT::v2f32:
    case MVT::v2i32: Opc = ARM::VTRNd32; break;
    case MVT::v16i8: Opc = ARM::VTRNq8; break;
    case MVT::v8i16: Opc = ARM::VTRNq16; break;
    case MVT::v4f32:
    case MVT::v4i32: Opc = ARM::VTRNq32; break;
    }
    return CurDAG->getMachineNode(Opc, dl, VT, VT,
                                  N->getOperand(0), N->getOperand(1));
  }

  case ISD::INTRINSIC_VOID:
  case ISD::INTRINSIC_W_CHAIN: {
    unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
    EVT VT = N->getValueType(0);
    unsigned Opc = 0;

    switch (IntNo) {
    default:
      break;

    case Intrinsic::arm_neon_vld2: {
      SDValue MemAddr, MemUpdate, MemOpc;
      if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
        return NULL;
      if (VT.is64BitVector()) {
        switch (VT.getSimpleVT().SimpleTy) {
        default: llvm_unreachable("unhandled vld2 type");
        case MVT::v8i8:  Opc = ARM::VLD2d8; break;
        case MVT::v4i16: Opc = ARM::VLD2d16; break;
        case MVT::v2f32:
        case MVT::v2i32: Opc = ARM::VLD2d32; break;
        case MVT::v1i64: Opc = ARM::VLD2d64; break;
        }
        SDValue Chain = N->getOperand(0);
        const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain };
        return CurDAG->getMachineNode(Opc, dl, VT, VT, MVT::Other, Ops, 4);
      }
      // Quad registers are loaded as pairs of double registers.
      EVT RegVT;
      switch (VT.getSimpleVT().SimpleTy) {
      default: llvm_unreachable("unhandled vld2 type");
      case MVT::v16i8: Opc = ARM::VLD2q8; RegVT = MVT::v8i8; break;
      case MVT::v8i16: Opc = ARM::VLD2q16; RegVT = MVT::v4i16; break;
      case MVT::v4f32: Opc = ARM::VLD2q32; RegVT = MVT::v2f32; break;
      case MVT::v4i32: Opc = ARM::VLD2q32; RegVT = MVT::v2i32; break;
      }
      SDValue Chain = N->getOperand(0);
      const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain };
      std::vector<EVT> ResTys(4, RegVT);
      ResTys.push_back(MVT::Other);
      SDNode *VLd = CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 4);
      SDNode *Q0 = PairDRegs(VT, SDValue(VLd, 0), SDValue(VLd, 1));
      SDNode *Q1 = PairDRegs(VT, SDValue(VLd, 2), SDValue(VLd, 3));
      ReplaceUses(SDValue(N, 0), SDValue(Q0, 0));
      ReplaceUses(SDValue(N, 1), SDValue(Q1, 0));
      ReplaceUses(SDValue(N, 2), SDValue(VLd, 4));
      return NULL;
    }

    case Intrinsic::arm_neon_vld3: {
      SDValue MemAddr, MemUpdate, MemOpc;
      if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
        return NULL;
      if (VT.is64BitVector()) {
        switch (VT.getSimpleVT().SimpleTy) {
        default: llvm_unreachable("unhandled vld3 type");
        case MVT::v8i8:  Opc = ARM::VLD3d8; break;
        case MVT::v4i16: Opc = ARM::VLD3d16; break;
        case MVT::v2f32:
        case MVT::v2i32: Opc = ARM::VLD3d32; break;
        case MVT::v1i64: Opc = ARM::VLD3d64; break;
        }
        SDValue Chain = N->getOperand(0);
        const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain };
        return CurDAG->getMachineNode(Opc, dl, VT, VT, VT, MVT::Other, Ops, 4);
      }
      // Quad registers are loaded with two separate instructions, where one
      // loads the even registers and the other loads the odd registers.
      EVT RegVT;
      unsigned Opc2 = 0;
      switch (VT.getSimpleVT().SimpleTy) {
      default: llvm_unreachable("unhandled vld3 type");
      case MVT::v16i8:
        Opc = ARM::VLD3q8a;  Opc2 = ARM::VLD3q8b;  RegVT = MVT::v8i8; break;
      case MVT::v8i16:
        Opc = ARM::VLD3q16a; Opc2 = ARM::VLD3q16b; RegVT = MVT::v4i16; break;
      case MVT::v4f32:
        Opc = ARM::VLD3q32a; Opc2 = ARM::VLD3q32b; RegVT = MVT::v2f32; break;
      case MVT::v4i32:
        Opc = ARM::VLD3q32a; Opc2 = ARM::VLD3q32b; RegVT = MVT::v2i32; break;
      }
      SDValue Chain = N->getOperand(0);
      // Enable writeback to the address register.
      MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32);

      std::vector<EVT> ResTys(3, RegVT);
      ResTys.push_back(MemAddr.getValueType());
      ResTys.push_back(MVT::Other);

      const SDValue OpsA[] = { MemAddr, MemUpdate, MemOpc, Chain };
      SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 4);
      Chain = SDValue(VLdA, 4);

      const SDValue OpsB[] = { SDValue(VLdA, 3), MemUpdate, MemOpc, Chain };
      SDNode *VLdB = CurDAG->getMachineNode(Opc2, dl, ResTys, OpsB, 4);
      Chain = SDValue(VLdB, 4);

      SDNode *Q0 = PairDRegs(VT, SDValue(VLdA, 0), SDValue(VLdB, 0));
      SDNode *Q1 = PairDRegs(VT, SDValue(VLdA, 1), SDValue(VLdB, 1));
      SDNode *Q2 = PairDRegs(VT, SDValue(VLdA, 2), SDValue(VLdB, 2));
      ReplaceUses(SDValue(N, 0), SDValue(Q0, 0));
      ReplaceUses(SDValue(N, 1), SDValue(Q1, 0));
      ReplaceUses(SDValue(N, 2), SDValue(Q2, 0));
      ReplaceUses(SDValue(N, 3), Chain);
      return NULL;
    }

    case Intrinsic::arm_neon_vld4: {
      SDValue MemAddr, MemUpdate, MemOpc;
      if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
        return NULL;
      if (VT.is64BitVector()) {
        switch (VT.getSimpleVT().SimpleTy) {
        default: llvm_unreachable("unhandled vld4 type");
        case MVT::v8i8:  Opc = ARM::VLD4d8; break;
        case MVT::v4i16: Opc = ARM::VLD4d16; break;
        case MVT::v2f32:
        case MVT::v2i32: Opc = ARM::VLD4d32; break;
        case MVT::v1i64: Opc = ARM::VLD4d64; break;
        }
        SDValue Chain = N->getOperand(0);
        const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, Chain };
        std::vector<EVT> ResTys(4, VT);
        ResTys.push_back(MVT::Other);
        return CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 4);
      }
      // Quad registers are loaded with two separate instructions, where one
      // loads the even registers and the other loads the odd registers.
      EVT RegVT;
      unsigned Opc2 = 0;
      switch (VT.getSimpleVT().SimpleTy) {
      default: llvm_unreachable("unhandled vld4 type");
      case MVT::v16i8:
        Opc = ARM::VLD4q8a;  Opc2 = ARM::VLD4q8b;  RegVT = MVT::v8i8; break;
      case MVT::v8i16:
        Opc = ARM::VLD4q16a; Opc2 = ARM::VLD4q16b; RegVT = MVT::v4i16; break;
      case MVT::v4f32:
        Opc = ARM::VLD4q32a; Opc2 = ARM::VLD4q32b; RegVT = MVT::v2f32; break;
      case MVT::v4i32:
        Opc = ARM::VLD4q32a; Opc2 = ARM::VLD4q32b; RegVT = MVT::v2i32; break;
      }
      SDValue Chain = N->getOperand(0);
      // Enable writeback to the address register.
      MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32);

      std::vector<EVT> ResTys(4, RegVT);
      ResTys.push_back(MemAddr.getValueType());
      ResTys.push_back(MVT::Other);

      const SDValue OpsA[] = { MemAddr, MemUpdate, MemOpc, Chain };
      SDNode *VLdA = CurDAG->getMachineNode(Opc, dl, ResTys, OpsA, 4);
      Chain = SDValue(VLdA, 5);

      const SDValue OpsB[] = { SDValue(VLdA, 4), MemUpdate, MemOpc, Chain };
      SDNode *VLdB = CurDAG->getMachineNode(Opc2, dl, ResTys, OpsB, 4);
      Chain = SDValue(VLdB, 5);

      SDNode *Q0 = PairDRegs(VT, SDValue(VLdA, 0), SDValue(VLdB, 0));
      SDNode *Q1 = PairDRegs(VT, SDValue(VLdA, 1), SDValue(VLdB, 1));
      SDNode *Q2 = PairDRegs(VT, SDValue(VLdA, 2), SDValue(VLdB, 2));
      SDNode *Q3 = PairDRegs(VT, SDValue(VLdA, 3), SDValue(VLdB, 3));
      ReplaceUses(SDValue(N, 0), SDValue(Q0, 0));
      ReplaceUses(SDValue(N, 1), SDValue(Q1, 0));
      ReplaceUses(SDValue(N, 2), SDValue(Q2, 0));
      ReplaceUses(SDValue(N, 3), SDValue(Q3, 0));
      ReplaceUses(SDValue(N, 4), Chain);
      return NULL;
    }

    case Intrinsic::arm_neon_vld2lane: {
      SDValue MemAddr, MemUpdate, MemOpc;
      if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
        return NULL;
      if (VT.is64BitVector()) {
        switch (VT.getSimpleVT().SimpleTy) {
        default: llvm_unreachable("unhandled vld2lane type");
        case MVT::v8i8:  Opc = ARM::VLD2LNd8; break;
        case MVT::v4i16: Opc = ARM::VLD2LNd16; break;
        case MVT::v2f32:
        case MVT::v2i32: Opc = ARM::VLD2LNd32; break;
        }
        SDValue Chain = N->getOperand(0);
        const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
                                N->getOperand(3), N->getOperand(4),
                                N->getOperand(5), Chain };
        return CurDAG->getMachineNode(Opc, dl, VT, VT, MVT::Other, Ops, 7);
      }
      // Quad registers are handled by extracting subregs, doing the load,
      // and then inserting the results as subregs.
      EVT RegVT;
      unsigned Opc2 = 0;
      switch (VT.getSimpleVT().SimpleTy) {
      default: llvm_unreachable("unhandled vld2lane type");
      case MVT::v8i16:
        Opc = ARM::VLD2LNq16a;
        Opc2 = ARM::VLD2LNq16b;
        RegVT = MVT::v4i16;
        break;
      case MVT::v4f32:
        Opc = ARM::VLD2LNq32a;
        Opc2 = ARM::VLD2LNq32b;
        RegVT = MVT::v2f32;
        break;
      case MVT::v4i32:
        Opc = ARM::VLD2LNq32a;
        Opc2 = ARM::VLD2LNq32b;
        RegVT = MVT::v2i32;
        break;
      }
      SDValue Chain = N->getOperand(0);
      unsigned Lane = cast<ConstantSDNode>(N->getOperand(5))->getZExtValue();
      unsigned NumElts = RegVT.getVectorNumElements();
      int SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;

      SDValue D0 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
                                                  N->getOperand(3));
      SDValue D1 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
                                                  N->getOperand(4));
      const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, D0, D1,
                              getI32Imm(Lane % NumElts), Chain };
      SDNode *VLdLn = CurDAG->getMachineNode((Lane < NumElts) ? Opc : Opc2,
                                             dl, RegVT, RegVT, MVT::Other,
                                             Ops, 7);
      SDValue Q0 = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
                                                 N->getOperand(3),
                                                 SDValue(VLdLn, 0));
      SDValue Q1 = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
                                                 N->getOperand(4),
                                                 SDValue(VLdLn, 1));
      Chain = SDValue(VLdLn, 2);
      ReplaceUses(SDValue(N, 0), Q0);
      ReplaceUses(SDValue(N, 1), Q1);
      ReplaceUses(SDValue(N, 2), Chain);
      return NULL;
    }

    case Intrinsic::arm_neon_vld3lane: {
      SDValue MemAddr, MemUpdate, MemOpc;
      if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
        return NULL;
      if (VT.is64BitVector()) {
        switch (VT.getSimpleVT().SimpleTy) {
        default: llvm_unreachable("unhandled vld3lane type");
        case MVT::v8i8:  Opc = ARM::VLD3LNd8; break;
        case MVT::v4i16: Opc = ARM::VLD3LNd16; break;
        case MVT::v2f32:
        case MVT::v2i32: Opc = ARM::VLD3LNd32; break;
        }
        SDValue Chain = N->getOperand(0);
        const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
                                N->getOperand(3), N->getOperand(4),
                                N->getOperand(5), N->getOperand(6), Chain };
        return CurDAG->getMachineNode(Opc, dl, VT, VT, VT, MVT::Other, Ops, 8);
      }
      // Quad registers are handled by extracting subregs, doing the load,
      // and then inserting the results as subregs.
      EVT RegVT;
      unsigned Opc2 = 0;
      switch (VT.getSimpleVT().SimpleTy) {
      default: llvm_unreachable("unhandled vld3lane type");
      case MVT::v8i16:
        Opc = ARM::VLD3LNq16a;
        Opc2 = ARM::VLD3LNq16b;
        RegVT = MVT::v4i16;
        break;
      case MVT::v4f32:
        Opc = ARM::VLD3LNq32a;
        Opc2 = ARM::VLD3LNq32b;
        RegVT = MVT::v2f32;
        break;
      case MVT::v4i32:
        Opc = ARM::VLD3LNq32a;
        Opc2 = ARM::VLD3LNq32b;
        RegVT = MVT::v2i32;
        break;
      }
      SDValue Chain = N->getOperand(0);
      unsigned Lane = cast<ConstantSDNode>(N->getOperand(6))->getZExtValue();
      unsigned NumElts = RegVT.getVectorNumElements();
      int SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;

      SDValue D0 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
                                                  N->getOperand(3));
      SDValue D1 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
                                                  N->getOperand(4));
      SDValue D2 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
                                                  N->getOperand(5));
      const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, D0, D1, D2,
                              getI32Imm(Lane % NumElts), Chain };
      SDNode *VLdLn = CurDAG->getMachineNode((Lane < NumElts) ? Opc : Opc2,
                                             dl, RegVT, RegVT, RegVT,
                                             MVT::Other, Ops, 8);
      SDValue Q0 = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
                                                 N->getOperand(3),
                                                 SDValue(VLdLn, 0));
      SDValue Q1 = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
                                                 N->getOperand(4),
                                                 SDValue(VLdLn, 1));
      SDValue Q2 = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
                                                 N->getOperand(5),
                                                 SDValue(VLdLn, 2));
      Chain = SDValue(VLdLn, 3);
      ReplaceUses(SDValue(N, 0), Q0);
      ReplaceUses(SDValue(N, 1), Q1);
      ReplaceUses(SDValue(N, 2), Q2);
      ReplaceUses(SDValue(N, 3), Chain);
      return NULL;
    }

    case Intrinsic::arm_neon_vld4lane: {
      SDValue MemAddr, MemUpdate, MemOpc;
      if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
        return NULL;
      if (VT.is64BitVector()) {
        switch (VT.getSimpleVT().SimpleTy) {
        default: llvm_unreachable("unhandled vld4lane type");
        case MVT::v8i8:  Opc = ARM::VLD4LNd8; break;
        case MVT::v4i16: Opc = ARM::VLD4LNd16; break;
        case MVT::v2f32:
        case MVT::v2i32: Opc = ARM::VLD4LNd32; break;
        }
        SDValue Chain = N->getOperand(0);
        const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
                                N->getOperand(3), N->getOperand(4),
                                N->getOperand(5), N->getOperand(6),
                                N->getOperand(7), Chain };
        std::vector<EVT> ResTys(4, VT);
        ResTys.push_back(MVT::Other);
        return CurDAG->getMachineNode(Opc, dl, ResTys, Ops, 9);
      }
      // Quad registers are handled by extracting subregs, doing the load,
      // and then inserting the results as subregs.
      EVT RegVT;
      unsigned Opc2 = 0;
      switch (VT.getSimpleVT().SimpleTy) {
      default: llvm_unreachable("unhandled vld4lane type");
      case MVT::v8i16:
        Opc = ARM::VLD4LNq16a;
        Opc2 = ARM::VLD4LNq16b;
        RegVT = MVT::v4i16;
        break;
      case MVT::v4f32:
        Opc = ARM::VLD4LNq32a;
        Opc2 = ARM::VLD4LNq32b;
        RegVT = MVT::v2f32;
        break;
      case MVT::v4i32:
        Opc = ARM::VLD4LNq32a;
        Opc2 = ARM::VLD4LNq32b;
        RegVT = MVT::v2i32;
        break;
      }
      SDValue Chain = N->getOperand(0);
      unsigned Lane = cast<ConstantSDNode>(N->getOperand(7))->getZExtValue();
      unsigned NumElts = RegVT.getVectorNumElements();
      int SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;

      SDValue D0 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
                                                  N->getOperand(3));
      SDValue D1 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
                                                  N->getOperand(4));
      SDValue D2 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
                                                  N->getOperand(5));
      SDValue D3 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
                                                  N->getOperand(6));
      const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, D0, D1, D2, D3,
                              getI32Imm(Lane % NumElts), Chain };
      std::vector<EVT> ResTys(4, RegVT);
      ResTys.push_back(MVT::Other);
      SDNode *VLdLn = CurDAG->getMachineNode((Lane < NumElts) ? Opc : Opc2,
                                             dl, ResTys, Ops, 9);
      SDValue Q0 = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
                                                 N->getOperand(3),
                                                 SDValue(VLdLn, 0));
      SDValue Q1 = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
                                                 N->getOperand(4),
                                                 SDValue(VLdLn, 1));
      SDValue Q2 = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
                                                 N->getOperand(5),
                                                 SDValue(VLdLn, 2));
      SDValue Q3 = CurDAG->getTargetInsertSubreg(SubregIdx, dl, VT,
                                                 N->getOperand(6),
                                                 SDValue(VLdLn, 3));
      Chain = SDValue(VLdLn, 4);
      ReplaceUses(SDValue(N, 0), Q0);
      ReplaceUses(SDValue(N, 1), Q1);
      ReplaceUses(SDValue(N, 2), Q2);
      ReplaceUses(SDValue(N, 3), Q3);
      ReplaceUses(SDValue(N, 4), Chain);
      return NULL;
    }

    case Intrinsic::arm_neon_vst2: {
      SDValue MemAddr, MemUpdate, MemOpc;
      if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
        return NULL;
      VT = N->getOperand(3).getValueType();
      if (VT.is64BitVector()) {
        switch (VT.getSimpleVT().SimpleTy) {
        default: llvm_unreachable("unhandled vst2 type");
        case MVT::v8i8:  Opc = ARM::VST2d8; break;
        case MVT::v4i16: Opc = ARM::VST2d16; break;
        case MVT::v2f32:
        case MVT::v2i32: Opc = ARM::VST2d32; break;
        case MVT::v1i64: Opc = ARM::VST2d64; break;
        }
        SDValue Chain = N->getOperand(0);
        const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
                                N->getOperand(3), N->getOperand(4), Chain };
        return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 6);
      }
      // Quad registers are stored as pairs of double registers.
      EVT RegVT;
      switch (VT.getSimpleVT().SimpleTy) {
      default: llvm_unreachable("unhandled vst2 type");
      case MVT::v16i8: Opc = ARM::VST2q8; RegVT = MVT::v8i8; break;
      case MVT::v8i16: Opc = ARM::VST2q16; RegVT = MVT::v4i16; break;
      case MVT::v4f32: Opc = ARM::VST2q32; RegVT = MVT::v2f32; break;
      case MVT::v4i32: Opc = ARM::VST2q32; RegVT = MVT::v2i32; break;
      }
      SDValue Chain = N->getOperand(0);
      SDValue D0 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
                                                  N->getOperand(3));
      SDValue D1 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
                                                  N->getOperand(3));
      SDValue D2 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
                                                  N->getOperand(4));
      SDValue D3 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
                                                  N->getOperand(4));
      const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
                              D0, D1, D2, D3, Chain };
      return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 8);
    }

    case Intrinsic::arm_neon_vst3: {
      SDValue MemAddr, MemUpdate, MemOpc;
      if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
        return NULL;
      VT = N->getOperand(3).getValueType();
      if (VT.is64BitVector()) {
        switch (VT.getSimpleVT().SimpleTy) {
        default: llvm_unreachable("unhandled vst3 type");
        case MVT::v8i8:  Opc = ARM::VST3d8; break;
        case MVT::v4i16: Opc = ARM::VST3d16; break;
        case MVT::v2f32:
        case MVT::v2i32: Opc = ARM::VST3d32; break;
        case MVT::v1i64: Opc = ARM::VST3d64; break;
        }
        SDValue Chain = N->getOperand(0);
        const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
                                N->getOperand(3), N->getOperand(4),
                                N->getOperand(5), Chain };
        return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 7);
      }
      // Quad registers are stored with two separate instructions, where one
      // stores the even registers and the other stores the odd registers.
      EVT RegVT;
      unsigned Opc2 = 0;
      switch (VT.getSimpleVT().SimpleTy) {
      default: llvm_unreachable("unhandled vst3 type");
      case MVT::v16i8:
        Opc = ARM::VST3q8a;  Opc2 = ARM::VST3q8b;  RegVT = MVT::v8i8; break;
      case MVT::v8i16:
        Opc = ARM::VST3q16a; Opc2 = ARM::VST3q16b; RegVT = MVT::v4i16; break;
      case MVT::v4f32:
        Opc = ARM::VST3q32a; Opc2 = ARM::VST3q32b; RegVT = MVT::v2f32; break;
      case MVT::v4i32:
        Opc = ARM::VST3q32a; Opc2 = ARM::VST3q32b; RegVT = MVT::v2i32; break;
      }
      SDValue Chain = N->getOperand(0);
      // Enable writeback to the address register.
      MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32);

      SDValue D0 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
                                                  N->getOperand(3));
      SDValue D2 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
                                                  N->getOperand(4));
      SDValue D4 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
                                                  N->getOperand(5));
      const SDValue OpsA[] = { MemAddr, MemUpdate, MemOpc, D0, D2, D4, Chain };
      SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
                                            MVT::Other, OpsA, 7);
      Chain = SDValue(VStA, 1);

      SDValue D1 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
                                                  N->getOperand(3));
      SDValue D3 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
                                                  N->getOperand(4));
      SDValue D5 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
                                                  N->getOperand(5));
      MemAddr = SDValue(VStA, 0);
      const SDValue OpsB[] = { MemAddr, MemUpdate, MemOpc, D1, D3, D5, Chain };
      SDNode *VStB = CurDAG->getMachineNode(Opc2, dl, MemAddr.getValueType(),
                                            MVT::Other, OpsB, 7);
      Chain = SDValue(VStB, 1);
      ReplaceUses(SDValue(N, 0), Chain);
      return NULL;
    }

    case Intrinsic::arm_neon_vst4: {
      SDValue MemAddr, MemUpdate, MemOpc;
      if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
        return NULL;
      VT = N->getOperand(3).getValueType();
      if (VT.is64BitVector()) {
        switch (VT.getSimpleVT().SimpleTy) {
        default: llvm_unreachable("unhandled vst4 type");
        case MVT::v8i8:  Opc = ARM::VST4d8; break;
        case MVT::v4i16: Opc = ARM::VST4d16; break;
        case MVT::v2f32:
        case MVT::v2i32: Opc = ARM::VST4d32; break;
        case MVT::v1i64: Opc = ARM::VST4d64; break;
        }
        SDValue Chain = N->getOperand(0);
        const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
                                N->getOperand(3), N->getOperand(4),
                                N->getOperand(5), N->getOperand(6), Chain };
        return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 8);
      }
      // Quad registers are stored with two separate instructions, where one
      // stores the even registers and the other stores the odd registers.
      EVT RegVT;
      unsigned Opc2 = 0;
      switch (VT.getSimpleVT().SimpleTy) {
      default: llvm_unreachable("unhandled vst4 type");
      case MVT::v16i8:
        Opc = ARM::VST4q8a;  Opc2 = ARM::VST4q8b;  RegVT = MVT::v8i8; break;
      case MVT::v8i16:
        Opc = ARM::VST4q16a; Opc2 = ARM::VST4q16b; RegVT = MVT::v4i16; break;
      case MVT::v4f32:
        Opc = ARM::VST4q32a; Opc2 = ARM::VST4q32b; RegVT = MVT::v2f32; break;
      case MVT::v4i32:
        Opc = ARM::VST4q32a; Opc2 = ARM::VST4q32b; RegVT = MVT::v2i32; break;
      }
      SDValue Chain = N->getOperand(0);
      // Enable writeback to the address register.
      MemOpc = CurDAG->getTargetConstant(ARM_AM::getAM6Opc(true), MVT::i32);

      SDValue D0 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
                                                  N->getOperand(3));
      SDValue D2 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
                                                  N->getOperand(4));
      SDValue D4 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
                                                  N->getOperand(5));
      SDValue D6 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_0, dl, RegVT,
                                                  N->getOperand(6));
      const SDValue OpsA[] = { MemAddr, MemUpdate, MemOpc,
                               D0, D2, D4, D6, Chain };
      SDNode *VStA = CurDAG->getMachineNode(Opc, dl, MemAddr.getValueType(),
                                            MVT::Other, OpsA, 8);
      Chain = SDValue(VStA, 1);

      SDValue D1 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
                                                  N->getOperand(3));
      SDValue D3 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
                                                  N->getOperand(4));
      SDValue D5 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
                                                  N->getOperand(5));
      SDValue D7 = CurDAG->getTargetExtractSubreg(ARM::DSUBREG_1, dl, RegVT,
                                                  N->getOperand(6));
      MemAddr = SDValue(VStA, 0);
      const SDValue OpsB[] = { MemAddr, MemUpdate, MemOpc,
                               D1, D3, D5, D7, Chain };
      SDNode *VStB = CurDAG->getMachineNode(Opc2, dl, MemAddr.getValueType(),
                                            MVT::Other, OpsB, 8);
      Chain = SDValue(VStB, 1);
      ReplaceUses(SDValue(N, 0), Chain);
      return NULL;
    }

    case Intrinsic::arm_neon_vst2lane: {
      SDValue MemAddr, MemUpdate, MemOpc;
      if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
        return NULL;
      VT = N->getOperand(3).getValueType();
      if (VT.is64BitVector()) {
        switch (VT.getSimpleVT().SimpleTy) {
        default: llvm_unreachable("unhandled vst2lane type");
        case MVT::v8i8:  Opc = ARM::VST2LNd8; break;
        case MVT::v4i16: Opc = ARM::VST2LNd16; break;
        case MVT::v2f32:
        case MVT::v2i32: Opc = ARM::VST2LNd32; break;
        }
        SDValue Chain = N->getOperand(0);
        const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
                                N->getOperand(3), N->getOperand(4),
                                N->getOperand(5), Chain };
        return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 7);
      }
      // Quad registers are handled by extracting subregs and then doing
      // the store.
      EVT RegVT;
      unsigned Opc2 = 0;
      switch (VT.getSimpleVT().SimpleTy) {
      default: llvm_unreachable("unhandled vst2lane type");
      case MVT::v8i16:
        Opc = ARM::VST2LNq16a;
        Opc2 = ARM::VST2LNq16b;
        RegVT = MVT::v4i16;
        break;
      case MVT::v4f32:
        Opc = ARM::VST2LNq32a;
        Opc2 = ARM::VST2LNq32b;
        RegVT = MVT::v2f32;
        break;
      case MVT::v4i32:
        Opc = ARM::VST2LNq32a;
        Opc2 = ARM::VST2LNq32b;
        RegVT = MVT::v2i32;
        break;
      }
      SDValue Chain = N->getOperand(0);
      unsigned Lane = cast<ConstantSDNode>(N->getOperand(5))->getZExtValue();
      unsigned NumElts = RegVT.getVectorNumElements();
      int SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;

      SDValue D0 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
                                                  N->getOperand(3));
      SDValue D1 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
                                                  N->getOperand(4));
      const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, D0, D1,
                              getI32Imm(Lane % NumElts), Chain };
      return CurDAG->getMachineNode((Lane < NumElts) ? Opc : Opc2,
                                    dl, MVT::Other, Ops, 7);
    }

    case Intrinsic::arm_neon_vst3lane: {
      SDValue MemAddr, MemUpdate, MemOpc;
      if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
        return NULL;
      VT = N->getOperand(3).getValueType();
      if (VT.is64BitVector()) {
        switch (VT.getSimpleVT().SimpleTy) {
        default: llvm_unreachable("unhandled vst3lane type");
        case MVT::v8i8:  Opc = ARM::VST3LNd8; break;
        case MVT::v4i16: Opc = ARM::VST3LNd16; break;
        case MVT::v2f32:
        case MVT::v2i32: Opc = ARM::VST3LNd32; break;
        }
        SDValue Chain = N->getOperand(0);
        const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
                                N->getOperand(3), N->getOperand(4),
                                N->getOperand(5), N->getOperand(6), Chain };
        return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 8);
      }
      // Quad registers are handled by extracting subregs and then doing
      // the store.
      EVT RegVT;
      unsigned Opc2 = 0;
      switch (VT.getSimpleVT().SimpleTy) {
      default: llvm_unreachable("unhandled vst3lane type");
      case MVT::v8i16:
        Opc = ARM::VST3LNq16a;
        Opc2 = ARM::VST3LNq16b;
        RegVT = MVT::v4i16;
        break;
      case MVT::v4f32:
        Opc = ARM::VST3LNq32a;
        Opc2 = ARM::VST3LNq32b;
        RegVT = MVT::v2f32;
        break;
      case MVT::v4i32:
        Opc = ARM::VST3LNq32a;
        Opc2 = ARM::VST3LNq32b;
        RegVT = MVT::v2i32;
        break;
      }
      SDValue Chain = N->getOperand(0);
      unsigned Lane = cast<ConstantSDNode>(N->getOperand(6))->getZExtValue();
      unsigned NumElts = RegVT.getVectorNumElements();
      int SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;

      SDValue D0 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
                                                  N->getOperand(3));
      SDValue D1 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
                                                  N->getOperand(4));
      SDValue D2 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
                                                  N->getOperand(5));
      const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, D0, D1, D2,
                              getI32Imm(Lane % NumElts), Chain };
      return CurDAG->getMachineNode((Lane < NumElts) ? Opc : Opc2,
                                    dl, MVT::Other, Ops, 8);
    }

    case Intrinsic::arm_neon_vst4lane: {
      SDValue MemAddr, MemUpdate, MemOpc;
      if (!SelectAddrMode6(Op, N->getOperand(2), MemAddr, MemUpdate, MemOpc))
        return NULL;
      VT = N->getOperand(3).getValueType();
      if (VT.is64BitVector()) {
        switch (VT.getSimpleVT().SimpleTy) {
        default: llvm_unreachable("unhandled vst4lane type");
        case MVT::v8i8:  Opc = ARM::VST4LNd8; break;
        case MVT::v4i16: Opc = ARM::VST4LNd16; break;
        case MVT::v2f32:
        case MVT::v2i32: Opc = ARM::VST4LNd32; break;
        }
        SDValue Chain = N->getOperand(0);
        const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc,
                                N->getOperand(3), N->getOperand(4),
                                N->getOperand(5), N->getOperand(6),
                                N->getOperand(7), Chain };
        return CurDAG->getMachineNode(Opc, dl, MVT::Other, Ops, 9);
      }
      // Quad registers are handled by extracting subregs and then doing
      // the store.
      EVT RegVT;
      unsigned Opc2 = 0;
      switch (VT.getSimpleVT().SimpleTy) {
      default: llvm_unreachable("unhandled vst4lane type");
      case MVT::v8i16:
        Opc = ARM::VST4LNq16a;
        Opc2 = ARM::VST4LNq16b;
        RegVT = MVT::v4i16;
        break;
      case MVT::v4f32:
        Opc = ARM::VST4LNq32a;
        Opc2 = ARM::VST4LNq32b;
        RegVT = MVT::v2f32;
        break;
      case MVT::v4i32:
        Opc = ARM::VST4LNq32a;
        Opc2 = ARM::VST4LNq32b;
        RegVT = MVT::v2i32;
        break;
      }
      SDValue Chain = N->getOperand(0);
      unsigned Lane = cast<ConstantSDNode>(N->getOperand(7))->getZExtValue();
      unsigned NumElts = RegVT.getVectorNumElements();
      int SubregIdx = (Lane < NumElts) ? ARM::DSUBREG_0 : ARM::DSUBREG_1;

      SDValue D0 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
                                                  N->getOperand(3));
      SDValue D1 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
                                                  N->getOperand(4));
      SDValue D2 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
                                                  N->getOperand(5));
      SDValue D3 = CurDAG->getTargetExtractSubreg(SubregIdx, dl, RegVT,
                                                  N->getOperand(6));
      const SDValue Ops[] = { MemAddr, MemUpdate, MemOpc, D0, D1, D2, D3,
                              getI32Imm(Lane % NumElts), Chain };
      return CurDAG->getMachineNode((Lane < NumElts) ? Opc : Opc2,
                                    dl, MVT::Other, Ops, 9);
    }
    }
  }
  }

  return SelectCode(Op);
}

bool ARMDAGToDAGISel::
SelectInlineAsmMemoryOperand(const SDValue &Op, char ConstraintCode,
                             std::vector<SDValue> &OutOps) {
  assert(ConstraintCode == 'm' && "unexpected asm memory constraint");

  SDValue Base, Offset, Opc;
  if (!SelectAddrMode2(Op, Op, Base, Offset, Opc))
    return true;

  OutOps.push_back(Base);
  OutOps.push_back(Offset);
  OutOps.push_back(Opc);
  return false;
}

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