ARMISelLowering.cpp

Go to the documentation of this file.

//===-- ARMISelLowering.cpp - ARM DAG Lowering Implementation -------------===//
//
//                     The LLVM Compiler Infrastructure
//
// This file is distributed under the University of Illinois Open Source
// License. See LICENSE.TXT for details.
//
//===----------------------------------------------------------------------===//
//
// This file defines the interfaces that ARM uses to lower LLVM code into a
// selection DAG.
//
//===----------------------------------------------------------------------===//

#define DEBUG_TYPE "arm-isel"
#include "ARMISelLowering.h"
#include "ARM.h"
#include "ARMCallingConv.h"
#include "ARMConstantPoolValue.h"
#include "ARMMachineFunctionInfo.h"
#include "ARMPerfectShuffle.h"
#include "ARMSubtarget.h"
#include "ARMTargetMachine.h"
#include "ARMTargetObjectFile.h"
#include "MCTargetDesc/ARMAddressingModes.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/ADT/StringExtras.h"
#include "llvm/CodeGen/CallingConvLower.h"
#include "llvm/CodeGen/IntrinsicLowering.h"
#include "llvm/CodeGen/MachineBasicBlock.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/CodeGen/SelectionDAG.h"
#include "llvm/IR/CallingConv.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/GlobalValue.h"
#include "llvm/IR/Instruction.h"
#include "llvm/IR/Instructions.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/Type.h"
#include "llvm/MC/MCSectionMachO.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/MathExtras.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetOptions.h"
using namespace llvm;

STATISTIC(NumTailCalls, "Number of tail calls");
STATISTIC(NumMovwMovt, "Number of GAs materialized with movw + movt");
STATISTIC(NumLoopByVals, "Number of loops generated for byval arguments");

// This option should go away when tail calls fully work.
static cl::opt<bool>
EnableARMTailCalls("arm-tail-calls", cl::Hidden,
  cl::desc("Generate tail calls (TEMPORARY OPTION)."),
  cl::init(false));

cl::opt<bool>
EnableARMLongCalls("arm-long-calls", cl::Hidden,
  cl::desc("Generate calls via indirect call instructions"),
  cl::init(false));

static cl::opt<bool>
ARMInterworking("arm-interworking", cl::Hidden,
  cl::desc("Enable / disable ARM interworking (for debugging only)"),
  cl::init(true));

namespace {
  class ARMCCState : public CCState {
  public:
    ARMCCState(CallingConv::ID CC, bool isVarArg, MachineFunction &MF,
               const TargetMachine &TM, SmallVector<CCValAssign, 16> &locs,
               LLVMContext &C, ParmContext PC)
        : CCState(CC, isVarArg, MF, TM, locs, C) {
      assert(((PC == Call) || (PC == Prologue)) &&
             "ARMCCState users must specify whether their context is call"
             "or prologue generation.");
      CallOrPrologue = PC;
    }
  };
}

// The APCS parameter registers.
static const uint16_t GPRArgRegs[] = {
  ARM::R0, ARM::R1, ARM::R2, ARM::R3
};

void ARMTargetLowering::addTypeForNEON(MVT VT, MVT PromotedLdStVT,
                                       MVT PromotedBitwiseVT) {
  if (VT != PromotedLdStVT) {
    setOperationAction(ISD::LOAD, VT, Promote);
    AddPromotedToType (ISD::LOAD, VT, PromotedLdStVT);

    setOperationAction(ISD::STORE, VT, Promote);
    AddPromotedToType (ISD::STORE, VT, PromotedLdStVT);
  }

  MVT ElemTy = VT.getVectorElementType();
  if (ElemTy != MVT::i64 && ElemTy != MVT::f64)
    setOperationAction(ISD::SETCC, VT, Custom);
  setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
  setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
  if (ElemTy == MVT::i32) {
    setOperationAction(ISD::SINT_TO_FP, VT, Custom);
    setOperationAction(ISD::UINT_TO_FP, VT, Custom);
    setOperationAction(ISD::FP_TO_SINT, VT, Custom);
    setOperationAction(ISD::FP_TO_UINT, VT, Custom);
  } else {
    setOperationAction(ISD::SINT_TO_FP, VT, Expand);
    setOperationAction(ISD::UINT_TO_FP, VT, Expand);
    setOperationAction(ISD::FP_TO_SINT, VT, Expand);
    setOperationAction(ISD::FP_TO_UINT, VT, Expand);
  }
  setOperationAction(ISD::BUILD_VECTOR,      VT, Custom);
  setOperationAction(ISD::VECTOR_SHUFFLE,    VT, Custom);
  setOperationAction(ISD::CONCAT_VECTORS,    VT, Legal);
  setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
  setOperationAction(ISD::SELECT,            VT, Expand);
  setOperationAction(ISD::SELECT_CC,         VT, Expand);
  setOperationAction(ISD::VSELECT,           VT, Expand);
  setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);
  if (VT.isInteger()) {
    setOperationAction(ISD::SHL, VT, Custom);
    setOperationAction(ISD::SRA, VT, Custom);
    setOperationAction(ISD::SRL, VT, Custom);
  }

  // Promote all bit-wise operations.
  if (VT.isInteger() && VT != PromotedBitwiseVT) {
    setOperationAction(ISD::AND, VT, Promote);
    AddPromotedToType (ISD::AND, VT, PromotedBitwiseVT);
    setOperationAction(ISD::OR,  VT, Promote);
    AddPromotedToType (ISD::OR,  VT, PromotedBitwiseVT);
    setOperationAction(ISD::XOR, VT, Promote);
    AddPromotedToType (ISD::XOR, VT, PromotedBitwiseVT);
  }

  // Neon does not support vector divide/remainder operations.
  setOperationAction(ISD::SDIV, VT, Expand);
  setOperationAction(ISD::UDIV, VT, Expand);
  setOperationAction(ISD::FDIV, VT, Expand);
  setOperationAction(ISD::SREM, VT, Expand);
  setOperationAction(ISD::UREM, VT, Expand);
  setOperationAction(ISD::FREM, VT, Expand);
}

void ARMTargetLowering::addDRTypeForNEON(MVT VT) {
  addRegisterClass(VT, &ARM::DPRRegClass);
  addTypeForNEON(VT, MVT::f64, MVT::v2i32);
}

void ARMTargetLowering::addQRTypeForNEON(MVT VT) {
  addRegisterClass(VT, &ARM::QPRRegClass);
  addTypeForNEON(VT, MVT::v2f64, MVT::v4i32);
}

static TargetLoweringObjectFile *createTLOF(TargetMachine &TM) {
  if (TM.getSubtarget<ARMSubtarget>().isTargetDarwin())
    return new TargetLoweringObjectFileMachO();

  return new ARMElfTargetObjectFile();
}

ARMTargetLowering::ARMTargetLowering(TargetMachine &TM)
    : TargetLowering(TM, createTLOF(TM)) {
  Subtarget = &TM.getSubtarget<ARMSubtarget>();
  RegInfo = TM.getRegisterInfo();
  Itins = TM.getInstrItineraryData();

  setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);

  if (Subtarget->isTargetDarwin()) {
    // Uses VFP for Thumb libfuncs if available.
    if (Subtarget->isThumb() && Subtarget->hasVFP2()) {
      // Single-precision floating-point arithmetic.
      setLibcallName(RTLIB::ADD_F32, "__addsf3vfp");
      setLibcallName(RTLIB::SUB_F32, "__subsf3vfp");
      setLibcallName(RTLIB::MUL_F32, "__mulsf3vfp");
      setLibcallName(RTLIB::DIV_F32, "__divsf3vfp");

      // Double-precision floating-point arithmetic.
      setLibcallName(RTLIB::ADD_F64, "__adddf3vfp");
      setLibcallName(RTLIB::SUB_F64, "__subdf3vfp");
      setLibcallName(RTLIB::MUL_F64, "__muldf3vfp");
      setLibcallName(RTLIB::DIV_F64, "__divdf3vfp");

      // Single-precision comparisons.
      setLibcallName(RTLIB::OEQ_F32, "__eqsf2vfp");
      setLibcallName(RTLIB::UNE_F32, "__nesf2vfp");
      setLibcallName(RTLIB::OLT_F32, "__ltsf2vfp");
      setLibcallName(RTLIB::OLE_F32, "__lesf2vfp");
      setLibcallName(RTLIB::OGE_F32, "__gesf2vfp");
      setLibcallName(RTLIB::OGT_F32, "__gtsf2vfp");
      setLibcallName(RTLIB::UO_F32,  "__unordsf2vfp");
      setLibcallName(RTLIB::O_F32,   "__unordsf2vfp");

      setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
      setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETNE);
      setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
      setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
      setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
      setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
      setCmpLibcallCC(RTLIB::UO_F32,  ISD::SETNE);
      setCmpLibcallCC(RTLIB::O_F32,   ISD::SETEQ);

      // Double-precision comparisons.
      setLibcallName(RTLIB::OEQ_F64, "__eqdf2vfp");
      setLibcallName(RTLIB::UNE_F64, "__nedf2vfp");
      setLibcallName(RTLIB::OLT_F64, "__ltdf2vfp");
      setLibcallName(RTLIB::OLE_F64, "__ledf2vfp");
      setLibcallName(RTLIB::OGE_F64, "__gedf2vfp");
      setLibcallName(RTLIB::OGT_F64, "__gtdf2vfp");
      setLibcallName(RTLIB::UO_F64,  "__unorddf2vfp");
      setLibcallName(RTLIB::O_F64,   "__unorddf2vfp");

      setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
      setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETNE);
      setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
      setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
      setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
      setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
      setCmpLibcallCC(RTLIB::UO_F64,  ISD::SETNE);
      setCmpLibcallCC(RTLIB::O_F64,   ISD::SETEQ);

      // Floating-point to integer conversions.
      // i64 conversions are done via library routines even when generating VFP
      // instructions, so use the same ones.
      setLibcallName(RTLIB::FPTOSINT_F64_I32, "__fixdfsivfp");
      setLibcallName(RTLIB::FPTOUINT_F64_I32, "__fixunsdfsivfp");
      setLibcallName(RTLIB::FPTOSINT_F32_I32, "__fixsfsivfp");
      setLibcallName(RTLIB::FPTOUINT_F32_I32, "__fixunssfsivfp");

      // Conversions between floating types.
      setLibcallName(RTLIB::FPROUND_F64_F32, "__truncdfsf2vfp");
      setLibcallName(RTLIB::FPEXT_F32_F64,   "__extendsfdf2vfp");

      // Integer to floating-point conversions.
      // i64 conversions are done via library routines even when generating VFP
      // instructions, so use the same ones.
      // FIXME: There appears to be some naming inconsistency in ARM libgcc:
      // e.g., __floatunsidf vs. __floatunssidfvfp.
      setLibcallName(RTLIB::SINTTOFP_I32_F64, "__floatsidfvfp");
      setLibcallName(RTLIB::UINTTOFP_I32_F64, "__floatunssidfvfp");
      setLibcallName(RTLIB::SINTTOFP_I32_F32, "__floatsisfvfp");
      setLibcallName(RTLIB::UINTTOFP_I32_F32, "__floatunssisfvfp");
    }
  }

  // These libcalls are not available in 32-bit.
  setLibcallName(RTLIB::SHL_I128, 0);
  setLibcallName(RTLIB::SRL_I128, 0);
  setLibcallName(RTLIB::SRA_I128, 0);

  if (Subtarget->isAAPCS_ABI() && !Subtarget->isTargetDarwin()) {
    // Double-precision floating-point arithmetic helper functions
    // RTABI chapter 4.1.2, Table 2
    setLibcallName(RTLIB::ADD_F64, "__aeabi_dadd");
    setLibcallName(RTLIB::DIV_F64, "__aeabi_ddiv");
    setLibcallName(RTLIB::MUL_F64, "__aeabi_dmul");
    setLibcallName(RTLIB::SUB_F64, "__aeabi_dsub");
    setLibcallCallingConv(RTLIB::ADD_F64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::DIV_F64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::MUL_F64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::SUB_F64, CallingConv::ARM_AAPCS);

    // Double-precision floating-point comparison helper functions
    // RTABI chapter 4.1.2, Table 3
    setLibcallName(RTLIB::OEQ_F64, "__aeabi_dcmpeq");
    setCmpLibcallCC(RTLIB::OEQ_F64, ISD::SETNE);
    setLibcallName(RTLIB::UNE_F64, "__aeabi_dcmpeq");
    setCmpLibcallCC(RTLIB::UNE_F64, ISD::SETEQ);
    setLibcallName(RTLIB::OLT_F64, "__aeabi_dcmplt");
    setCmpLibcallCC(RTLIB::OLT_F64, ISD::SETNE);
    setLibcallName(RTLIB::OLE_F64, "__aeabi_dcmple");
    setCmpLibcallCC(RTLIB::OLE_F64, ISD::SETNE);
    setLibcallName(RTLIB::OGE_F64, "__aeabi_dcmpge");
    setCmpLibcallCC(RTLIB::OGE_F64, ISD::SETNE);
    setLibcallName(RTLIB::OGT_F64, "__aeabi_dcmpgt");
    setCmpLibcallCC(RTLIB::OGT_F64, ISD::SETNE);
    setLibcallName(RTLIB::UO_F64,  "__aeabi_dcmpun");
    setCmpLibcallCC(RTLIB::UO_F64,  ISD::SETNE);
    setLibcallName(RTLIB::O_F64,   "__aeabi_dcmpun");
    setCmpLibcallCC(RTLIB::O_F64,   ISD::SETEQ);
    setLibcallCallingConv(RTLIB::OEQ_F64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::UNE_F64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::OLT_F64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::OLE_F64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::OGE_F64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::OGT_F64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::UO_F64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::O_F64, CallingConv::ARM_AAPCS);

    // Single-precision floating-point arithmetic helper functions
    // RTABI chapter 4.1.2, Table 4
    setLibcallName(RTLIB::ADD_F32, "__aeabi_fadd");
    setLibcallName(RTLIB::DIV_F32, "__aeabi_fdiv");
    setLibcallName(RTLIB::MUL_F32, "__aeabi_fmul");
    setLibcallName(RTLIB::SUB_F32, "__aeabi_fsub");
    setLibcallCallingConv(RTLIB::ADD_F32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::DIV_F32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::MUL_F32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::SUB_F32, CallingConv::ARM_AAPCS);

    // Single-precision floating-point comparison helper functions
    // RTABI chapter 4.1.2, Table 5
    setLibcallName(RTLIB::OEQ_F32, "__aeabi_fcmpeq");
    setCmpLibcallCC(RTLIB::OEQ_F32, ISD::SETNE);
    setLibcallName(RTLIB::UNE_F32, "__aeabi_fcmpeq");
    setCmpLibcallCC(RTLIB::UNE_F32, ISD::SETEQ);
    setLibcallName(RTLIB::OLT_F32, "__aeabi_fcmplt");
    setCmpLibcallCC(RTLIB::OLT_F32, ISD::SETNE);
    setLibcallName(RTLIB::OLE_F32, "__aeabi_fcmple");
    setCmpLibcallCC(RTLIB::OLE_F32, ISD::SETNE);
    setLibcallName(RTLIB::OGE_F32, "__aeabi_fcmpge");
    setCmpLibcallCC(RTLIB::OGE_F32, ISD::SETNE);
    setLibcallName(RTLIB::OGT_F32, "__aeabi_fcmpgt");
    setCmpLibcallCC(RTLIB::OGT_F32, ISD::SETNE);
    setLibcallName(RTLIB::UO_F32,  "__aeabi_fcmpun");
    setCmpLibcallCC(RTLIB::UO_F32,  ISD::SETNE);
    setLibcallName(RTLIB::O_F32,   "__aeabi_fcmpun");
    setCmpLibcallCC(RTLIB::O_F32,   ISD::SETEQ);
    setLibcallCallingConv(RTLIB::OEQ_F32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::UNE_F32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::OLT_F32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::OLE_F32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::OGE_F32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::OGT_F32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::UO_F32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::O_F32, CallingConv::ARM_AAPCS);

    // Floating-point to integer conversions.
    // RTABI chapter 4.1.2, Table 6
    setLibcallName(RTLIB::FPTOSINT_F64_I32, "__aeabi_d2iz");
    setLibcallName(RTLIB::FPTOUINT_F64_I32, "__aeabi_d2uiz");
    setLibcallName(RTLIB::FPTOSINT_F64_I64, "__aeabi_d2lz");
    setLibcallName(RTLIB::FPTOUINT_F64_I64, "__aeabi_d2ulz");
    setLibcallName(RTLIB::FPTOSINT_F32_I32, "__aeabi_f2iz");
    setLibcallName(RTLIB::FPTOUINT_F32_I32, "__aeabi_f2uiz");
    setLibcallName(RTLIB::FPTOSINT_F32_I64, "__aeabi_f2lz");
    setLibcallName(RTLIB::FPTOUINT_F32_I64, "__aeabi_f2ulz");
    setLibcallCallingConv(RTLIB::FPTOSINT_F64_I32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::FPTOUINT_F64_I32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::FPTOSINT_F64_I64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::FPTOSINT_F32_I32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::FPTOUINT_F32_I32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::FPTOSINT_F32_I64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::ARM_AAPCS);

    // Conversions between floating types.
    // RTABI chapter 4.1.2, Table 7
    setLibcallName(RTLIB::FPROUND_F64_F32, "__aeabi_d2f");
    setLibcallName(RTLIB::FPEXT_F32_F64,   "__aeabi_f2d");
    setLibcallCallingConv(RTLIB::FPROUND_F64_F32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::FPEXT_F32_F64, CallingConv::ARM_AAPCS);

    // Integer to floating-point conversions.
    // RTABI chapter 4.1.2, Table 8
    setLibcallName(RTLIB::SINTTOFP_I32_F64, "__aeabi_i2d");
    setLibcallName(RTLIB::UINTTOFP_I32_F64, "__aeabi_ui2d");
    setLibcallName(RTLIB::SINTTOFP_I64_F64, "__aeabi_l2d");
    setLibcallName(RTLIB::UINTTOFP_I64_F64, "__aeabi_ul2d");
    setLibcallName(RTLIB::SINTTOFP_I32_F32, "__aeabi_i2f");
    setLibcallName(RTLIB::UINTTOFP_I32_F32, "__aeabi_ui2f");
    setLibcallName(RTLIB::SINTTOFP_I64_F32, "__aeabi_l2f");
    setLibcallName(RTLIB::UINTTOFP_I64_F32, "__aeabi_ul2f");
    setLibcallCallingConv(RTLIB::SINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::UINTTOFP_I32_F64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::SINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::UINTTOFP_I64_F64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::SINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::UINTTOFP_I32_F32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::SINTTOFP_I64_F32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::UINTTOFP_I64_F32, CallingConv::ARM_AAPCS);

    // Long long helper functions
    // RTABI chapter 4.2, Table 9
    setLibcallName(RTLIB::MUL_I64,  "__aeabi_lmul");
    setLibcallName(RTLIB::SHL_I64, "__aeabi_llsl");
    setLibcallName(RTLIB::SRL_I64, "__aeabi_llsr");
    setLibcallName(RTLIB::SRA_I64, "__aeabi_lasr");
    setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::SHL_I64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::SRL_I64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::SRA_I64, CallingConv::ARM_AAPCS);

    // Integer division functions
    // RTABI chapter 4.3.1
    setLibcallName(RTLIB::SDIV_I8,  "__aeabi_idiv");
    setLibcallName(RTLIB::SDIV_I16, "__aeabi_idiv");
    setLibcallName(RTLIB::SDIV_I32, "__aeabi_idiv");
    setLibcallName(RTLIB::SDIV_I64, "__aeabi_ldivmod");
    setLibcallName(RTLIB::UDIV_I8,  "__aeabi_uidiv");
    setLibcallName(RTLIB::UDIV_I16, "__aeabi_uidiv");
    setLibcallName(RTLIB::UDIV_I32, "__aeabi_uidiv");
    setLibcallName(RTLIB::UDIV_I64, "__aeabi_uldivmod");
    setLibcallCallingConv(RTLIB::SDIV_I8, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::SDIV_I16, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::SDIV_I32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::UDIV_I8, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::UDIV_I16, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::UDIV_I32, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::ARM_AAPCS);

    // Memory operations
    // RTABI chapter 4.3.4
    setLibcallName(RTLIB::MEMCPY,  "__aeabi_memcpy");
    setLibcallName(RTLIB::MEMMOVE, "__aeabi_memmove");
    setLibcallName(RTLIB::MEMSET,  "__aeabi_memset");
    setLibcallCallingConv(RTLIB::MEMCPY, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::MEMMOVE, CallingConv::ARM_AAPCS);
    setLibcallCallingConv(RTLIB::MEMSET, CallingConv::ARM_AAPCS);
  }

  // Use divmod compiler-rt calls for iOS 5.0 and later.
  if (Subtarget->getTargetTriple().getOS() == Triple::IOS &&
      !Subtarget->getTargetTriple().isOSVersionLT(5, 0)) {
    setLibcallName(RTLIB::SDIVREM_I32, "__divmodsi4");
    setLibcallName(RTLIB::UDIVREM_I32, "__udivmodsi4");
  }

  if (Subtarget->isThumb1Only())
    addRegisterClass(MVT::i32, &ARM::tGPRRegClass);
  else
    addRegisterClass(MVT::i32, &ARM::GPRRegClass);
  if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
      !Subtarget->isThumb1Only()) {
    addRegisterClass(MVT::f32, &ARM::SPRRegClass);
    if (!Subtarget->isFPOnlySP())
      addRegisterClass(MVT::f64, &ARM::DPRRegClass);

    setTruncStoreAction(MVT::f64, MVT::f32, Expand);
  }

  for (unsigned VT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
       VT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++VT) {
    for (unsigned InnerVT = (unsigned)MVT::FIRST_VECTOR_VALUETYPE;
         InnerVT <= (unsigned)MVT::LAST_VECTOR_VALUETYPE; ++InnerVT)
      setTruncStoreAction((MVT::SimpleValueType)VT,
                          (MVT::SimpleValueType)InnerVT, Expand);
    setLoadExtAction(ISD::SEXTLOAD, (MVT::SimpleValueType)VT, Expand);
    setLoadExtAction(ISD::ZEXTLOAD, (MVT::SimpleValueType)VT, Expand);
    setLoadExtAction(ISD::EXTLOAD, (MVT::SimpleValueType)VT, Expand);
  }

  setOperationAction(ISD::ConstantFP, MVT::f32, Custom);

  if (Subtarget->hasNEON()) {
    addDRTypeForNEON(MVT::v2f32);
    addDRTypeForNEON(MVT::v8i8);
    addDRTypeForNEON(MVT::v4i16);
    addDRTypeForNEON(MVT::v2i32);
    addDRTypeForNEON(MVT::v1i64);

    addQRTypeForNEON(MVT::v4f32);
    addQRTypeForNEON(MVT::v2f64);
    addQRTypeForNEON(MVT::v16i8);
    addQRTypeForNEON(MVT::v8i16);
    addQRTypeForNEON(MVT::v4i32);
    addQRTypeForNEON(MVT::v2i64);

    // v2f64 is legal so that QR subregs can be extracted as f64 elements, but
    // neither Neon nor VFP support any arithmetic operations on it.
    // The same with v4f32. But keep in mind that vadd, vsub, vmul are natively
    // supported for v4f32.
    setOperationAction(ISD::FADD, MVT::v2f64, Expand);
    setOperationAction(ISD::FSUB, MVT::v2f64, Expand);
    setOperationAction(ISD::FMUL, MVT::v2f64, Expand);
    // FIXME: Code duplication: FDIV and FREM are expanded always, see
    // ARMTargetLowering::addTypeForNEON method for details.
    setOperationAction(ISD::FDIV, MVT::v2f64, Expand);
    setOperationAction(ISD::FREM, MVT::v2f64, Expand);
    // FIXME: Create unittest.
    // In another words, find a way when "copysign" appears in DAG with vector
    // operands.
    setOperationAction(ISD::FCOPYSIGN, MVT::v2f64, Expand);
    // FIXME: Code duplication: SETCC has custom operation action, see
    // ARMTargetLowering::addTypeForNEON method for details.
    setOperationAction(ISD::SETCC, MVT::v2f64, Expand);
    // FIXME: Create unittest for FNEG and for FABS.
    setOperationAction(ISD::FNEG, MVT::v2f64, Expand);
    setOperationAction(ISD::FABS, MVT::v2f64, Expand);
    setOperationAction(ISD::FSQRT, MVT::v2f64, Expand);
    setOperationAction(ISD::FSIN, MVT::v2f64, Expand);
    setOperationAction(ISD::FCOS, MVT::v2f64, Expand);
    setOperationAction(ISD::FPOWI, MVT::v2f64, Expand);
    setOperationAction(ISD::FPOW, MVT::v2f64, Expand);
    setOperationAction(ISD::FLOG, MVT::v2f64, Expand);
    setOperationAction(ISD::FLOG2, MVT::v2f64, Expand);
    setOperationAction(ISD::FLOG10, MVT::v2f64, Expand);
    setOperationAction(ISD::FEXP, MVT::v2f64, Expand);
    setOperationAction(ISD::FEXP2, MVT::v2f64, Expand);
    // FIXME: Create unittest for FCEIL, FTRUNC, FRINT, FNEARBYINT, FFLOOR.
    setOperationAction(ISD::FCEIL, MVT::v2f64, Expand);
    setOperationAction(ISD::FTRUNC, MVT::v2f64, Expand);
    setOperationAction(ISD::FRINT, MVT::v2f64, Expand);
    setOperationAction(ISD::FNEARBYINT, MVT::v2f64, Expand);
    setOperationAction(ISD::FFLOOR, MVT::v2f64, Expand);
    setOperationAction(ISD::FMA, MVT::v2f64, Expand);

    setOperationAction(ISD::FSQRT, MVT::v4f32, Expand);
    setOperationAction(ISD::FSIN, MVT::v4f32, Expand);
    setOperationAction(ISD::FCOS, MVT::v4f32, Expand);
    setOperationAction(ISD::FPOWI, MVT::v4f32, Expand);
    setOperationAction(ISD::FPOW, MVT::v4f32, Expand);
    setOperationAction(ISD::FLOG, MVT::v4f32, Expand);
    setOperationAction(ISD::FLOG2, MVT::v4f32, Expand);
    setOperationAction(ISD::FLOG10, MVT::v4f32, Expand);
    setOperationAction(ISD::FEXP, MVT::v4f32, Expand);
    setOperationAction(ISD::FEXP2, MVT::v4f32, Expand);
    setOperationAction(ISD::FCEIL, MVT::v4f32, Expand);
    setOperationAction(ISD::FTRUNC, MVT::v4f32, Expand);
    setOperationAction(ISD::FRINT, MVT::v4f32, Expand);
    setOperationAction(ISD::FNEARBYINT, MVT::v4f32, Expand);
    setOperationAction(ISD::FFLOOR, MVT::v4f32, Expand);

    // Mark v2f32 intrinsics.
    setOperationAction(ISD::FSQRT, MVT::v2f32, Expand);
    setOperationAction(ISD::FSIN, MVT::v2f32, Expand);
    setOperationAction(ISD::FCOS, MVT::v2f32, Expand);
    setOperationAction(ISD::FPOWI, MVT::v2f32, Expand);
    setOperationAction(ISD::FPOW, MVT::v2f32, Expand);
    setOperationAction(ISD::FLOG, MVT::v2f32, Expand);
    setOperationAction(ISD::FLOG2, MVT::v2f32, Expand);
    setOperationAction(ISD::FLOG10, MVT::v2f32, Expand);
    setOperationAction(ISD::FEXP, MVT::v2f32, Expand);
    setOperationAction(ISD::FEXP2, MVT::v2f32, Expand);
    setOperationAction(ISD::FCEIL, MVT::v2f32, Expand);
    setOperationAction(ISD::FTRUNC, MVT::v2f32, Expand);
    setOperationAction(ISD::FRINT, MVT::v2f32, Expand);
    setOperationAction(ISD::FNEARBYINT, MVT::v2f32, Expand);
    setOperationAction(ISD::FFLOOR, MVT::v2f32, Expand);

    // Neon does not support some operations on v1i64 and v2i64 types.
    setOperationAction(ISD::MUL, MVT::v1i64, Expand);
    // Custom handling for some quad-vector types to detect VMULL.
    setOperationAction(ISD::MUL, MVT::v8i16, Custom);
    setOperationAction(ISD::MUL, MVT::v4i32, Custom);
    setOperationAction(ISD::MUL, MVT::v2i64, Custom);
    // Custom handling for some vector types to avoid expensive expansions
    setOperationAction(ISD::SDIV, MVT::v4i16, Custom);
    setOperationAction(ISD::SDIV, MVT::v8i8, Custom);
    setOperationAction(ISD::UDIV, MVT::v4i16, Custom);
    setOperationAction(ISD::UDIV, MVT::v8i8, Custom);
    setOperationAction(ISD::SETCC, MVT::v1i64, Expand);
    setOperationAction(ISD::SETCC, MVT::v2i64, Expand);
    // Neon does not have single instruction SINT_TO_FP and UINT_TO_FP with
    // a destination type that is wider than the source, and nor does
    // it have a FP_TO_[SU]INT instruction with a narrower destination than
    // source.
    setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Custom);
    setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
    setOperationAction(ISD::FP_TO_UINT, MVT::v4i16, Custom);
    setOperationAction(ISD::FP_TO_SINT, MVT::v4i16, Custom);

    setOperationAction(ISD::FP_ROUND,   MVT::v2f32, Expand);
    setOperationAction(ISD::FP_EXTEND,  MVT::v2f64, Expand);

    // Custom expand long extensions to vectors.
    setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32,  Custom);
    setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32,  Custom);
    setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64,  Custom);
    setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64,  Custom);
    setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
    setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
    setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64,  Custom);
    setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64,  Custom);

    // NEON does not have single instruction CTPOP for vectors with element
    // types wider than 8-bits.  However, custom lowering can leverage the
    // v8i8/v16i8 vcnt instruction.
    setOperationAction(ISD::CTPOP,      MVT::v2i32, Custom);
    setOperationAction(ISD::CTPOP,      MVT::v4i32, Custom);
    setOperationAction(ISD::CTPOP,      MVT::v4i16, Custom);
    setOperationAction(ISD::CTPOP,      MVT::v8i16, Custom);

    // NEON only has FMA instructions as of VFP4.
    if (!Subtarget->hasVFP4()) {
      setOperationAction(ISD::FMA, MVT::v2f32, Expand);
      setOperationAction(ISD::FMA, MVT::v4f32, Expand);
    }

    setTargetDAGCombine(ISD::INTRINSIC_VOID);
    setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
    setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
    setTargetDAGCombine(ISD::SHL);
    setTargetDAGCombine(ISD::SRL);
    setTargetDAGCombine(ISD::SRA);
    setTargetDAGCombine(ISD::SIGN_EXTEND);
    setTargetDAGCombine(ISD::ZERO_EXTEND);
    setTargetDAGCombine(ISD::ANY_EXTEND);
    setTargetDAGCombine(ISD::SELECT_CC);
    setTargetDAGCombine(ISD::BUILD_VECTOR);
    setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
    setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
    setTargetDAGCombine(ISD::STORE);
    setTargetDAGCombine(ISD::FP_TO_SINT);
    setTargetDAGCombine(ISD::FP_TO_UINT);
    setTargetDAGCombine(ISD::FDIV);

    // It is legal to extload from v4i8 to v4i16 or v4i32.
    MVT Tys[6] = {MVT::v8i8, MVT::v4i8, MVT::v2i8,
                  MVT::v4i16, MVT::v2i16,
                  MVT::v2i32};
    for (unsigned i = 0; i < 6; ++i) {
      setLoadExtAction(ISD::EXTLOAD, Tys[i], Legal);
      setLoadExtAction(ISD::ZEXTLOAD, Tys[i], Legal);
      setLoadExtAction(ISD::SEXTLOAD, Tys[i], Legal);
    }
  }

  // ARM and Thumb2 support UMLAL/SMLAL.
  if (!Subtarget->isThumb1Only())
    setTargetDAGCombine(ISD::ADDC);


  computeRegisterProperties();

  // ARM does not have f32 extending load.
  setLoadExtAction(ISD::EXTLOAD, MVT::f32, Expand);

  // ARM does not have i1 sign extending load.
  setLoadExtAction(ISD::SEXTLOAD, MVT::i1, Promote);

  // ARM supports all 4 flavors of integer indexed load / store.
  if (!Subtarget->isThumb1Only()) {
    for (unsigned im = (unsigned)ISD::PRE_INC;
         im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
      setIndexedLoadAction(im,  MVT::i1,  Legal);
      setIndexedLoadAction(im,  MVT::i8,  Legal);
      setIndexedLoadAction(im,  MVT::i16, Legal);
      setIndexedLoadAction(im,  MVT::i32, Legal);
      setIndexedStoreAction(im, MVT::i1,  Legal);
      setIndexedStoreAction(im, MVT::i8,  Legal);
      setIndexedStoreAction(im, MVT::i16, Legal);
      setIndexedStoreAction(im, MVT::i32, Legal);
    }
  }

  // i64 operation support.
  setOperationAction(ISD::MUL,     MVT::i64, Expand);
  setOperationAction(ISD::MULHU,   MVT::i32, Expand);
  if (Subtarget->isThumb1Only()) {
    setOperationAction(ISD::UMUL_LOHI, MVT::i32, Expand);
    setOperationAction(ISD::SMUL_LOHI, MVT::i32, Expand);
  }
  if (Subtarget->isThumb1Only() || !Subtarget->hasV6Ops()
      || (Subtarget->isThumb2() && !Subtarget->hasThumb2DSP()))
    setOperationAction(ISD::MULHS, MVT::i32, Expand);

  setOperationAction(ISD::SHL_PARTS, MVT::i32, Custom);
  setOperationAction(ISD::SRA_PARTS, MVT::i32, Custom);
  setOperationAction(ISD::SRL_PARTS, MVT::i32, Custom);
  setOperationAction(ISD::SRL,       MVT::i64, Custom);
  setOperationAction(ISD::SRA,       MVT::i64, Custom);

  if (!Subtarget->isThumb1Only()) {
    // FIXME: We should do this for Thumb1 as well.
    setOperationAction(ISD::ADDC,    MVT::i32, Custom);
    setOperationAction(ISD::ADDE,    MVT::i32, Custom);
    setOperationAction(ISD::SUBC,    MVT::i32, Custom);
    setOperationAction(ISD::SUBE,    MVT::i32, Custom);
  }

  // ARM does not have ROTL.
  setOperationAction(ISD::ROTL,  MVT::i32, Expand);
  setOperationAction(ISD::CTTZ,  MVT::i32, Custom);
  setOperationAction(ISD::CTPOP, MVT::i32, Expand);
  if (!Subtarget->hasV5TOps() || Subtarget->isThumb1Only())
    setOperationAction(ISD::CTLZ, MVT::i32, Expand);

  // These just redirect to CTTZ and CTLZ on ARM.
  setOperationAction(ISD::CTTZ_ZERO_UNDEF  , MVT::i32  , Expand);
  setOperationAction(ISD::CTLZ_ZERO_UNDEF  , MVT::i32  , Expand);

  setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Custom);

  // Only ARMv6 has BSWAP.
  if (!Subtarget->hasV6Ops())
    setOperationAction(ISD::BSWAP, MVT::i32, Expand);

  if (!(Subtarget->hasDivide() && Subtarget->isThumb2()) &&
      !(Subtarget->hasDivideInARMMode() && !Subtarget->isThumb())) {
    // These are expanded into libcalls if the cpu doesn't have HW divider.
    setOperationAction(ISD::SDIV,  MVT::i32, Expand);
    setOperationAction(ISD::UDIV,  MVT::i32, Expand);
  }
  setOperationAction(ISD::SREM,  MVT::i32, Expand);
  setOperationAction(ISD::UREM,  MVT::i32, Expand);
  setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
  setOperationAction(ISD::UDIVREM, MVT::i32, Expand);

  setOperationAction(ISD::GlobalAddress, MVT::i32,   Custom);
  setOperationAction(ISD::ConstantPool,  MVT::i32,   Custom);
  setOperationAction(ISD::GLOBAL_OFFSET_TABLE, MVT::i32, Custom);
  setOperationAction(ISD::GlobalTLSAddress, MVT::i32, Custom);
  setOperationAction(ISD::BlockAddress, MVT::i32, Custom);

  setOperationAction(ISD::TRAP, MVT::Other, Legal);

  // Use the default implementation.
  setOperationAction(ISD::VASTART,            MVT::Other, Custom);
  setOperationAction(ISD::VAARG,              MVT::Other, Expand);
  setOperationAction(ISD::VACOPY,             MVT::Other, Expand);
  setOperationAction(ISD::VAEND,              MVT::Other, Expand);
  setOperationAction(ISD::STACKSAVE,          MVT::Other, Expand);
  setOperationAction(ISD::STACKRESTORE,       MVT::Other, Expand);

  if (!Subtarget->isTargetDarwin()) {
    // Non-Darwin platforms may return values in these registers via the
    // personality function.
    setOperationAction(ISD::EHSELECTION,      MVT::i32,   Expand);
    setOperationAction(ISD::EXCEPTIONADDR,    MVT::i32,   Expand);
    setExceptionPointerRegister(ARM::R0);
    setExceptionSelectorRegister(ARM::R1);
  }

  setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
  // ARMv6 Thumb1 (except for CPUs that support dmb / dsb) and earlier use
  // the default expansion.
  // FIXME: This should be checking for v6k, not just v6.
  if (Subtarget->hasDataBarrier() ||
      (Subtarget->hasV6Ops() && !Subtarget->isThumb())) {
    // membarrier needs custom lowering; the rest are legal and handled
    // normally.
    setOperationAction(ISD::ATOMIC_FENCE, MVT::Other, Custom);
    // Custom lowering for 64-bit ops
    setOperationAction(ISD::ATOMIC_LOAD_ADD,  MVT::i64, Custom);
    setOperationAction(ISD::ATOMIC_LOAD_SUB,  MVT::i64, Custom);
    setOperationAction(ISD::ATOMIC_LOAD_AND,  MVT::i64, Custom);
    setOperationAction(ISD::ATOMIC_LOAD_OR,   MVT::i64, Custom);
    setOperationAction(ISD::ATOMIC_LOAD_XOR,  MVT::i64, Custom);
    setOperationAction(ISD::ATOMIC_SWAP,      MVT::i64, Custom);
    setOperationAction(ISD::ATOMIC_LOAD_MIN,  MVT::i64, Custom);
    setOperationAction(ISD::ATOMIC_LOAD_MAX,  MVT::i64, Custom);
    setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i64, Custom);
    setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i64, Custom);
    setOperationAction(ISD::ATOMIC_CMP_SWAP,  MVT::i64, Custom);
    // Automatically insert fences (dmb ist) around ATOMIC_SWAP etc.
    setInsertFencesForAtomic(true);
  } else {
    // Set them all for expansion, which will force libcalls.
    setOperationAction(ISD::ATOMIC_FENCE,   MVT::Other, Expand);
    setOperationAction(ISD::ATOMIC_CMP_SWAP,  MVT::i32, Expand);
    setOperationAction(ISD::ATOMIC_SWAP,      MVT::i32, Expand);
    setOperationAction(ISD::ATOMIC_LOAD_ADD,  MVT::i32, Expand);
    setOperationAction(ISD::ATOMIC_LOAD_SUB,  MVT::i32, Expand);
    setOperationAction(ISD::ATOMIC_LOAD_AND,  MVT::i32, Expand);
    setOperationAction(ISD::ATOMIC_LOAD_OR,   MVT::i32, Expand);
    setOperationAction(ISD::ATOMIC_LOAD_XOR,  MVT::i32, Expand);
    setOperationAction(ISD::ATOMIC_LOAD_NAND, MVT::i32, Expand);
    setOperationAction(ISD::ATOMIC_LOAD_MIN, MVT::i32, Expand);
    setOperationAction(ISD::ATOMIC_LOAD_MAX, MVT::i32, Expand);
    setOperationAction(ISD::ATOMIC_LOAD_UMIN, MVT::i32, Expand);
    setOperationAction(ISD::ATOMIC_LOAD_UMAX, MVT::i32, Expand);
    // Mark ATOMIC_LOAD and ATOMIC_STORE custom so we can handle the
    // Unordered/Monotonic case.
    setOperationAction(ISD::ATOMIC_LOAD, MVT::i32, Custom);
    setOperationAction(ISD::ATOMIC_STORE, MVT::i32, Custom);
  }

  setOperationAction(ISD::PREFETCH,         MVT::Other, Custom);

  // Requires SXTB/SXTH, available on v6 and up in both ARM and Thumb modes.
  if (!Subtarget->hasV6Ops()) {
    setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16, Expand);
    setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8,  Expand);
  }
  setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1, Expand);

  if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
      !Subtarget->isThumb1Only()) {
    // Turn f64->i64 into VMOVRRD, i64 -> f64 to VMOVDRR
    // iff target supports vfp2.
    setOperationAction(ISD::BITCAST, MVT::i64, Custom);
    setOperationAction(ISD::FLT_ROUNDS_, MVT::i32, Custom);
  }

  // We want to custom lower some of our intrinsics.
  setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
  if (Subtarget->isTargetDarwin()) {
    setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
    setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
    setLibcallName(RTLIB::UNWIND_RESUME, "_Unwind_SjLj_Resume");
  }

  setOperationAction(ISD::SETCC,     MVT::i32, Expand);
  setOperationAction(ISD::SETCC,     MVT::f32, Expand);
  setOperationAction(ISD::SETCC,     MVT::f64, Expand);
  setOperationAction(ISD::SELECT,    MVT::i32, Custom);
  setOperationAction(ISD::SELECT,    MVT::f32, Custom);
  setOperationAction(ISD::SELECT,    MVT::f64, Custom);
  setOperationAction(ISD::SELECT_CC, MVT::i32, Custom);
  setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
  setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);

  setOperationAction(ISD::BRCOND,    MVT::Other, Expand);
  setOperationAction(ISD::BR_CC,     MVT::i32,   Custom);
  setOperationAction(ISD::BR_CC,     MVT::f32,   Custom);
  setOperationAction(ISD::BR_CC,     MVT::f64,   Custom);
  setOperationAction(ISD::BR_JT,     MVT::Other, Custom);

  // We don't support sin/cos/fmod/copysign/pow
  setOperationAction(ISD::FSIN,      MVT::f64, Expand);
  setOperationAction(ISD::FSIN,      MVT::f32, Expand);
  setOperationAction(ISD::FCOS,      MVT::f32, Expand);
  setOperationAction(ISD::FCOS,      MVT::f64, Expand);
  setOperationAction(ISD::FSINCOS,   MVT::f64, Expand);
  setOperationAction(ISD::FSINCOS,   MVT::f32, Expand);
  setOperationAction(ISD::FREM,      MVT::f64, Expand);
  setOperationAction(ISD::FREM,      MVT::f32, Expand);
  if (!TM.Options.UseSoftFloat && Subtarget->hasVFP2() &&
      !Subtarget->isThumb1Only()) {
    setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
    setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
  }
  setOperationAction(ISD::FPOW,      MVT::f64, Expand);
  setOperationAction(ISD::FPOW,      MVT::f32, Expand);

  if (!Subtarget->hasVFP4()) {
    setOperationAction(ISD::FMA, MVT::f64, Expand);
    setOperationAction(ISD::FMA, MVT::f32, Expand);
  }

  // Various VFP goodness
  if (!TM.Options.UseSoftFloat && !Subtarget->isThumb1Only()) {
    // int <-> fp are custom expanded into bit_convert + ARMISD ops.
    if (Subtarget->hasVFP2()) {
      setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
      setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
      setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
      setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
    }
    // Special handling for half-precision FP.
    if (!Subtarget->hasFP16()) {
      setOperationAction(ISD::FP16_TO_FP32, MVT::f32, Expand);
      setOperationAction(ISD::FP32_TO_FP16, MVT::i32, Expand);
    }
  }

  // We have target-specific dag combine patterns for the following nodes:
  // ARMISD::VMOVRRD  - No need to call setTargetDAGCombine
  setTargetDAGCombine(ISD::ADD);
  setTargetDAGCombine(ISD::SUB);
  setTargetDAGCombine(ISD::MUL);
  setTargetDAGCombine(ISD::AND);
  setTargetDAGCombine(ISD::OR);
  setTargetDAGCombine(ISD::XOR);

  if (Subtarget->hasV6Ops())
    setTargetDAGCombine(ISD::SRL);

  setStackPointerRegisterToSaveRestore(ARM::SP);

  if (TM.Options.UseSoftFloat || Subtarget->isThumb1Only() ||
      !Subtarget->hasVFP2())
    setSchedulingPreference(Sched::RegPressure);
  else
    setSchedulingPreference(Sched::Hybrid);

  //// temporary - rewrite interface to use type
  MaxStoresPerMemset = 8;
  MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
  MaxStoresPerMemcpy = 4; // For @llvm.memcpy -> sequence of stores
  MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 4 : 2;
  MaxStoresPerMemmove = 4; // For @llvm.memmove -> sequence of stores
  MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 4 : 2;

  // On ARM arguments smaller than 4 bytes are extended, so all arguments
  // are at least 4 bytes aligned.
  setMinStackArgumentAlignment(4);

  // Prefer likely predicted branches to selects on out-of-order cores.
  PredictableSelectIsExpensive = Subtarget->isLikeA9();

  setMinFunctionAlignment(Subtarget->isThumb() ? 1 : 2);
}

// FIXME: It might make sense to define the representative register class as the
// nearest super-register that has a non-null superset. For example, DPR_VFP2 is
// a super-register of SPR, and DPR is a superset if DPR_VFP2. Consequently,
// SPR's representative would be DPR_VFP2. This should work well if register
// pressure tracking were modified such that a register use would increment the
// pressure of the register class's representative and all of it's super
// classes' representatives transitively. We have not implemented this because
// of the difficulty prior to coalescing of modeling operand register classes
// due to the common occurrence of cross class copies and subregister insertions
// and extractions.
std::pair<const TargetRegisterClass*, uint8_t>
ARMTargetLowering::findRepresentativeClass(MVT VT) const{
  const TargetRegisterClass *RRC = 0;
  uint8_t Cost = 1;
  switch (VT.SimpleTy) {
  default:
    return TargetLowering::findRepresentativeClass(VT);
  // Use DPR as representative register class for all floating point
  // and vector types. Since there are 32 SPR registers and 32 DPR registers so
  // the cost is 1 for both f32 and f64.
  case MVT::f32: case MVT::f64: case MVT::v8i8: case MVT::v4i16:
  case MVT::v2i32: case MVT::v1i64: case MVT::v2f32:
    RRC = &ARM::DPRRegClass;
    // When NEON is used for SP, only half of the register file is available
    // because operations that define both SP and DP results will be constrained
    // to the VFP2 class (D0-D15). We currently model this constraint prior to
    // coalescing by double-counting the SP regs. See the FIXME above.
    if (Subtarget->useNEONForSinglePrecisionFP())
      Cost = 2;
    break;
  case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
  case MVT::v4f32: case MVT::v2f64:
    RRC = &ARM::DPRRegClass;
    Cost = 2;
    break;
  case MVT::v4i64:
    RRC = &ARM::DPRRegClass;
    Cost = 4;
    break;
  case MVT::v8i64:
    RRC = &ARM::DPRRegClass;
    Cost = 8;
    break;
  }
  return std::make_pair(RRC, Cost);
}

const char *ARMTargetLowering::getTargetNodeName(unsigned Opcode) const {
  switch (Opcode) {
  default: return 0;
  case ARMISD::Wrapper:       return "ARMISD::Wrapper";
  case ARMISD::WrapperDYN:    return "ARMISD::WrapperDYN";
  case ARMISD::WrapperPIC:    return "ARMISD::WrapperPIC";
  case ARMISD::WrapperJT:     return "ARMISD::WrapperJT";
  case ARMISD::CALL:          return "ARMISD::CALL";
  case ARMISD::CALL_PRED:     return "ARMISD::CALL_PRED";
  case ARMISD::CALL_NOLINK:   return "ARMISD::CALL_NOLINK";
  case ARMISD::tCALL:         return "ARMISD::tCALL";
  case ARMISD::BRCOND:        return "ARMISD::BRCOND";
  case ARMISD::BR_JT:         return "ARMISD::BR_JT";
  case ARMISD::BR2_JT:        return "ARMISD::BR2_JT";
  case ARMISD::RET_FLAG:      return "ARMISD::RET_FLAG";
  case ARMISD::PIC_ADD:       return "ARMISD::PIC_ADD";
  case ARMISD::CMP:           return "ARMISD::CMP";
  case ARMISD::CMN:           return "ARMISD::CMN";
  case ARMISD::CMPZ:          return "ARMISD::CMPZ";
  case ARMISD::CMPFP:         return "ARMISD::CMPFP";
  case ARMISD::CMPFPw0:       return "ARMISD::CMPFPw0";
  case ARMISD::BCC_i64:       return "ARMISD::BCC_i64";
  case ARMISD::FMSTAT:        return "ARMISD::FMSTAT";

  case ARMISD::CMOV:          return "ARMISD::CMOV";

  case ARMISD::RBIT:          return "ARMISD::RBIT";

  case ARMISD::FTOSI:         return "ARMISD::FTOSI";
  case ARMISD::FTOUI:         return "ARMISD::FTOUI";
  case ARMISD::SITOF:         return "ARMISD::SITOF";
  case ARMISD::UITOF:         return "ARMISD::UITOF";

  case ARMISD::SRL_FLAG:      return "ARMISD::SRL_FLAG";
  case ARMISD::SRA_FLAG:      return "ARMISD::SRA_FLAG";
  case ARMISD::RRX:           return "ARMISD::RRX";

  case ARMISD::ADDC:          return "ARMISD::ADDC";
  case ARMISD::ADDE:          return "ARMISD::ADDE";
  case ARMISD::SUBC:          return "ARMISD::SUBC";
  case ARMISD::SUBE:          return "ARMISD::SUBE";

  case ARMISD::VMOVRRD:       return "ARMISD::VMOVRRD";
  case ARMISD::VMOVDRR:       return "ARMISD::VMOVDRR";

  case ARMISD::EH_SJLJ_SETJMP: return "ARMISD::EH_SJLJ_SETJMP";
  case ARMISD::EH_SJLJ_LONGJMP:return "ARMISD::EH_SJLJ_LONGJMP";

  case ARMISD::TC_RETURN:     return "ARMISD::TC_RETURN";

  case ARMISD::THREAD_POINTER:return "ARMISD::THREAD_POINTER";

  case ARMISD::DYN_ALLOC:     return "ARMISD::DYN_ALLOC";

  case ARMISD::MEMBARRIER:    return "ARMISD::MEMBARRIER";
  case ARMISD::MEMBARRIER_MCR: return "ARMISD::MEMBARRIER_MCR";

  case ARMISD::PRELOAD:       return "ARMISD::PRELOAD";

  case ARMISD::VCEQ:          return "ARMISD::VCEQ";
  case ARMISD::VCEQZ:         return "ARMISD::VCEQZ";
  case ARMISD::VCGE:          return "ARMISD::VCGE";
  case ARMISD::VCGEZ:         return "ARMISD::VCGEZ";
  case ARMISD::VCLEZ:         return "ARMISD::VCLEZ";
  case ARMISD::VCGEU:         return "ARMISD::VCGEU";
  case ARMISD::VCGT:          return "ARMISD::VCGT";
  case ARMISD::VCGTZ:         return "ARMISD::VCGTZ";
  case ARMISD::VCLTZ:         return "ARMISD::VCLTZ";
  case ARMISD::VCGTU:         return "ARMISD::VCGTU";
  case ARMISD::VTST:          return "ARMISD::VTST";

  case ARMISD::VSHL:          return "ARMISD::VSHL";
  case ARMISD::VSHRs:         return "ARMISD::VSHRs";
  case ARMISD::VSHRu:         return "ARMISD::VSHRu";
  case ARMISD::VSHLLs:        return "ARMISD::VSHLLs";
  case ARMISD::VSHLLu:        return "ARMISD::VSHLLu";
  case ARMISD::VSHLLi:        return "ARMISD::VSHLLi";
  case ARMISD::VSHRN:         return "ARMISD::VSHRN";
  case ARMISD::VRSHRs:        return "ARMISD::VRSHRs";
  case ARMISD::VRSHRu:        return "ARMISD::VRSHRu";
  case ARMISD::VRSHRN:        return "ARMISD::VRSHRN";
  case ARMISD::VQSHLs:        return "ARMISD::VQSHLs";
  case ARMISD::VQSHLu:        return "ARMISD::VQSHLu";
  case ARMISD::VQSHLsu:       return "ARMISD::VQSHLsu";
  case ARMISD::VQSHRNs:       return "ARMISD::VQSHRNs";
  case ARMISD::VQSHRNu:       return "ARMISD::VQSHRNu";
  case ARMISD::VQSHRNsu:      return "ARMISD::VQSHRNsu";
  case ARMISD::VQRSHRNs:      return "ARMISD::VQRSHRNs";
  case ARMISD::VQRSHRNu:      return "ARMISD::VQRSHRNu";
  case ARMISD::VQRSHRNsu:     return "ARMISD::VQRSHRNsu";
  case ARMISD::VGETLANEu:     return "ARMISD::VGETLANEu";
  case ARMISD::VGETLANEs:     return "ARMISD::VGETLANEs";
  case ARMISD::VMOVIMM:       return "ARMISD::VMOVIMM";
  case ARMISD::VMVNIMM:       return "ARMISD::VMVNIMM";
  case ARMISD::VMOVFPIMM:     return "ARMISD::VMOVFPIMM";
  case ARMISD::VDUP:          return "ARMISD::VDUP";
  case ARMISD::VDUPLANE:      return "ARMISD::VDUPLANE";
  case ARMISD::VEXT:          return "ARMISD::VEXT";
  case ARMISD::VREV64:        return "ARMISD::VREV64";
  case ARMISD::VREV32:        return "ARMISD::VREV32";
  case ARMISD::VREV16:        return "ARMISD::VREV16";
  case ARMISD::VZIP:          return "ARMISD::VZIP";
  case ARMISD::VUZP:          return "ARMISD::VUZP";
  case ARMISD::VTRN:          return "ARMISD::VTRN";
  case ARMISD::VTBL1:         return "ARMISD::VTBL1";
  case ARMISD::VTBL2:         return "ARMISD::VTBL2";
  case ARMISD::VMULLs:        return "ARMISD::VMULLs";
  case ARMISD::VMULLu:        return "ARMISD::VMULLu";
  case ARMISD::UMLAL:         return "ARMISD::UMLAL";
  case ARMISD::SMLAL:         return "ARMISD::SMLAL";
  case ARMISD::BUILD_VECTOR:  return "ARMISD::BUILD_VECTOR";
  case ARMISD::FMAX:          return "ARMISD::FMAX";
  case ARMISD::FMIN:          return "ARMISD::FMIN";
  case ARMISD::BFI:           return "ARMISD::BFI";
  case ARMISD::VORRIMM:       return "ARMISD::VORRIMM";
  case ARMISD::VBICIMM:       return "ARMISD::VBICIMM";
  case ARMISD::VBSL:          return "ARMISD::VBSL";
  case ARMISD::VLD2DUP:       return "ARMISD::VLD2DUP";
  case ARMISD::VLD3DUP:       return "ARMISD::VLD3DUP";
  case ARMISD::VLD4DUP:       return "ARMISD::VLD4DUP";
  case ARMISD::VLD1_UPD:      return "ARMISD::VLD1_UPD";
  case ARMISD::VLD2_UPD:      return "ARMISD::VLD2_UPD";
  case ARMISD::VLD3_UPD:      return "ARMISD::VLD3_UPD";
  case ARMISD::VLD4_UPD:      return "ARMISD::VLD4_UPD";
  case ARMISD::VLD2LN_UPD:    return "ARMISD::VLD2LN_UPD";
  case ARMISD::VLD3LN_UPD:    return "ARMISD::VLD3LN_UPD";
  case ARMISD::VLD4LN_UPD:    return "ARMISD::VLD4LN_UPD";
  case ARMISD::VLD2DUP_UPD:   return "ARMISD::VLD2DUP_UPD";
  case ARMISD::VLD3DUP_UPD:   return "ARMISD::VLD3DUP_UPD";
  case ARMISD::VLD4DUP_UPD:   return "ARMISD::VLD4DUP_UPD";
  case ARMISD::VST1_UPD:      return "ARMISD::VST1_UPD";
  case ARMISD::VST2_UPD:      return "ARMISD::VST2_UPD";
  case ARMISD::VST3_UPD:      return "ARMISD::VST3_UPD";
  case ARMISD::VST4_UPD:      return "ARMISD::VST4_UPD";
  case ARMISD::VST2LN_UPD:    return "ARMISD::VST2LN_UPD";
  case ARMISD::VST3LN_UPD:    return "ARMISD::VST3LN_UPD";
  case ARMISD::VST4LN_UPD:    return "ARMISD::VST4LN_UPD";
  }
}

EVT ARMTargetLowering::getSetCCResultType(LLVMContext &, EVT VT) const {
  if (!VT.isVector()) return getPointerTy();
  return VT.changeVectorElementTypeToInteger();
}

/// getRegClassFor - Return the register class that should be used for the
/// specified value type.
const TargetRegisterClass *ARMTargetLowering::getRegClassFor(MVT VT) const {
  // Map v4i64 to QQ registers but do not make the type legal. Similarly map
  // v8i64 to QQQQ registers. v4i64 and v8i64 are only used for REG_SEQUENCE to
  // load / store 4 to 8 consecutive D registers.
  if (Subtarget->hasNEON()) {
    if (VT == MVT::v4i64)
      return &ARM::QQPRRegClass;
    if (VT == MVT::v8i64)
      return &ARM::QQQQPRRegClass;
  }
  return TargetLowering::getRegClassFor(VT);
}

// Create a fast isel object.
FastISel *
ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
                                  const TargetLibraryInfo *libInfo) const {
  return ARM::createFastISel(funcInfo, libInfo);
}

/// getMaximalGlobalOffset - Returns the maximal possible offset which can
/// be used for loads / stores from the global.
unsigned ARMTargetLowering::getMaximalGlobalOffset() const {
  return (Subtarget->isThumb1Only() ? 127 : 4095);
}

Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {
  unsigned NumVals = N->getNumValues();
  if (!NumVals)
    return Sched::RegPressure;

  for (unsigned i = 0; i != NumVals; ++i) {
    EVT VT = N->getValueType(i);
    if (VT == MVT::Glue || VT == MVT::Other)
      continue;
    if (VT.isFloatingPoint() || VT.isVector())
      return Sched::ILP;
  }

  if (!N->isMachineOpcode())
    return Sched::RegPressure;

  // Load are scheduled for latency even if there instruction itinerary
  // is not available.
  const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
  const MCInstrDesc &MCID = TII->get(N->getMachineOpcode());

  if (MCID.getNumDefs() == 0)
    return Sched::RegPressure;
  if (!Itins->isEmpty() &&
      Itins->getOperandCycle(MCID.getSchedClass(), 0) > 2)
    return Sched::ILP;

  return Sched::RegPressure;
}

//===----------------------------------------------------------------------===//
// Lowering Code
//===----------------------------------------------------------------------===//

/// IntCCToARMCC - Convert a DAG integer condition code to an ARM CC
static ARMCC::CondCodes IntCCToARMCC(ISD::CondCode CC) {
  switch (CC) {
  default: llvm_unreachable("Unknown condition code!");
  case ISD::SETNE:  return ARMCC::NE;
  case ISD::SETEQ:  return ARMCC::EQ;
  case ISD::SETGT:  return ARMCC::GT;
  case ISD::SETGE:  return ARMCC::GE;
  case ISD::SETLT:  return ARMCC::LT;
  case ISD::SETLE:  return ARMCC::LE;
  case ISD::SETUGT: return ARMCC::HI;
  case ISD::SETUGE: return ARMCC::HS;
  case ISD::SETULT: return ARMCC::LO;
  case ISD::SETULE: return ARMCC::LS;
  }
}

/// FPCCToARMCC - Convert a DAG fp condition code to an ARM CC.
static void FPCCToARMCC(ISD::CondCode CC, ARMCC::CondCodes &CondCode,
                        ARMCC::CondCodes &CondCode2) {
  CondCode2 = ARMCC::AL;
  switch (CC) {
  default: llvm_unreachable("Unknown FP condition!");
  case ISD::SETEQ:
  case ISD::SETOEQ: CondCode = ARMCC::EQ; break;
  case ISD::SETGT:
  case ISD::SETOGT: CondCode = ARMCC::GT; break;
  case ISD::SETGE:
  case ISD::SETOGE: CondCode = ARMCC::GE; break;
  case ISD::SETOLT: CondCode = ARMCC::MI; break;
  case ISD::SETOLE: CondCode = ARMCC::LS; break;
  case ISD::SETONE: CondCode = ARMCC::MI; CondCode2 = ARMCC::GT; break;
  case ISD::SETO:   CondCode = ARMCC::VC; break;
  case ISD::SETUO:  CondCode = ARMCC::VS; break;
  case ISD::SETUEQ: CondCode = ARMCC::EQ; CondCode2 = ARMCC::VS; break;
  case ISD::SETUGT: CondCode = ARMCC::HI; break;
  case ISD::SETUGE: CondCode = ARMCC::PL; break;
  case ISD::SETLT:
  case ISD::SETULT: CondCode = ARMCC::LT; break;
  case ISD::SETLE:
  case ISD::SETULE: CondCode = ARMCC::LE; break;
  case ISD::SETNE:
  case ISD::SETUNE: CondCode = ARMCC::NE; break;
  }
}

//===----------------------------------------------------------------------===//
//                      Calling Convention Implementation
//===----------------------------------------------------------------------===//

#include "ARMGenCallingConv.inc"

/// CCAssignFnForNode - Selects the correct CCAssignFn for a the
/// given CallingConvention value.
CCAssignFn *ARMTargetLowering::CCAssignFnForNode(CallingConv::ID CC,
                                                 bool Return,
                                                 bool isVarArg) const {
  switch (CC) {
  default:
    llvm_unreachable("Unsupported calling convention");
  case CallingConv::Fast:
    if (Subtarget->hasVFP2() && !isVarArg) {
      if (!Subtarget->isAAPCS_ABI())
        return (Return ? RetFastCC_ARM_APCS : FastCC_ARM_APCS);
      // For AAPCS ABI targets, just use VFP variant of the calling convention.
      return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
    }
    // Fallthrough
  case CallingConv::C: {
    // Use target triple & subtarget features to do actual dispatch.
    if (!Subtarget->isAAPCS_ABI())
      return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
    else if (Subtarget->hasVFP2() &&
             getTargetMachine().Options.FloatABIType == FloatABI::Hard &&
             !isVarArg)
      return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
    return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
  }
  case CallingConv::ARM_AAPCS_VFP:
    if (!isVarArg)
      return (Return ? RetCC_ARM_AAPCS_VFP : CC_ARM_AAPCS_VFP);
    // Fallthrough
  case CallingConv::ARM_AAPCS:
    return (Return ? RetCC_ARM_AAPCS : CC_ARM_AAPCS);
  case CallingConv::ARM_APCS:
    return (Return ? RetCC_ARM_APCS : CC_ARM_APCS);
  case CallingConv::GHC:
    return (Return ? RetCC_ARM_APCS : CC_ARM_APCS_GHC);
  }
}

/// LowerCallResult - Lower the result values of a call into the
/// appropriate copies out of appropriate physical registers.
SDValue
ARMTargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
                                   CallingConv::ID CallConv, bool isVarArg,
                                   const SmallVectorImpl<ISD::InputArg> &Ins,
                                   DebugLoc dl, SelectionDAG &DAG,
                                   SmallVectorImpl<SDValue> &InVals,
                                   bool isThisReturn, SDValue ThisVal) const {

  // Assign locations to each value returned by this call.
  SmallVector<CCValAssign, 16> RVLocs;
  ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
                    getTargetMachine(), RVLocs, *DAG.getContext(), Call);
  CCInfo.AnalyzeCallResult(Ins,
                           CCAssignFnForNode(CallConv, /* Return*/ true,
                                             isVarArg));

  // Copy all of the result registers out of their specified physreg.
  for (unsigned i = 0; i != RVLocs.size(); ++i) {
    CCValAssign VA = RVLocs[i];

    // Pass 'this' value directly from the argument to return value, to avoid
    // reg unit interference
    if (i == 0 && isThisReturn) {
      assert(!VA.needsCustom() && VA.getLocVT() == MVT::i32 &&
             "unexpected return calling convention register assignment");
      InVals.push_back(ThisVal);
      continue;
    }

    SDValue Val;
    if (VA.needsCustom()) {
      // Handle f64 or half of a v2f64.
      SDValue Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
                                      InFlag);
      Chain = Lo.getValue(1);
      InFlag = Lo.getValue(2);
      VA = RVLocs[++i]; // skip ahead to next loc
      SDValue Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32,
                                      InFlag);
      Chain = Hi.getValue(1);
      InFlag = Hi.getValue(2);
      Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);

      if (VA.getLocVT() == MVT::v2f64) {
        SDValue Vec = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
        Vec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
                          DAG.getConstant(0, MVT::i32));

        VA = RVLocs[++i]; // skip ahead to next loc
        Lo = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
        Chain = Lo.getValue(1);
        InFlag = Lo.getValue(2);
        VA = RVLocs[++i]; // skip ahead to next loc
        Hi = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), MVT::i32, InFlag);
        Chain = Hi.getValue(1);
        InFlag = Hi.getValue(2);
        Val = DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
        Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Vec, Val,
                          DAG.getConstant(1, MVT::i32));
      }
    } else {
      Val = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(), VA.getLocVT(),
                               InFlag);
      Chain = Val.getValue(1);
      InFlag = Val.getValue(2);
    }

    switch (VA.getLocInfo()) {
    default: llvm_unreachable("Unknown loc info!");
    case CCValAssign::Full: break;
    case CCValAssign::BCvt:
      Val = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), Val);
      break;
    }

    InVals.push_back(Val);
  }

  return Chain;
}

/// LowerMemOpCallTo - Store the argument to the stack.
SDValue
ARMTargetLowering::LowerMemOpCallTo(SDValue Chain,
                                    SDValue StackPtr, SDValue Arg,
                                    DebugLoc dl, SelectionDAG &DAG,
                                    const CCValAssign &VA,
                                    ISD::ArgFlagsTy Flags) const {
  unsigned LocMemOffset = VA.getLocMemOffset();
  SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset);
  PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr, PtrOff);
  return DAG.getStore(Chain, dl, Arg, PtrOff,
                      MachinePointerInfo::getStack(LocMemOffset),
                      false, false, 0);
}

void ARMTargetLowering::PassF64ArgInRegs(DebugLoc dl, SelectionDAG &DAG,
                                         SDValue Chain, SDValue &Arg,
                                         RegsToPassVector &RegsToPass,
                                         CCValAssign &VA, CCValAssign &NextVA,
                                         SDValue &StackPtr,
                                         SmallVector<SDValue, 8> &MemOpChains,
                                         ISD::ArgFlagsTy Flags) const {

  SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
                              DAG.getVTList(MVT::i32, MVT::i32), Arg);
  RegsToPass.push_back(std::make_pair(VA.getLocReg(), fmrrd));

  if (NextVA.isRegLoc())
    RegsToPass.push_back(std::make_pair(NextVA.getLocReg(), fmrrd.getValue(1)));
  else {
    assert(NextVA.isMemLoc());
    if (StackPtr.getNode() == 0)
      StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());

    MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, fmrrd.getValue(1),
                                           dl, DAG, NextVA,
                                           Flags));
  }
}

/// LowerCall - Lowering a call into a callseq_start <-
/// ARMISD:CALL <- callseq_end chain. Also add input and output parameter
/// nodes.
SDValue
ARMTargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
                             SmallVectorImpl<SDValue> &InVals) const {
  SelectionDAG &DAG                     = CLI.DAG;
  DebugLoc &dl                          = CLI.DL;
  SmallVector<ISD::OutputArg, 32> &Outs = CLI.Outs;
  SmallVector<SDValue, 32> &OutVals     = CLI.OutVals;
  SmallVector<ISD::InputArg, 32> &Ins   = CLI.Ins;
  SDValue Chain                         = CLI.Chain;
  SDValue Callee                        = CLI.Callee;
  bool &isTailCall                      = CLI.IsTailCall;
  CallingConv::ID CallConv              = CLI.CallConv;
  bool doesNotRet                       = CLI.DoesNotReturn;
  bool isVarArg                         = CLI.IsVarArg;

  MachineFunction &MF = DAG.getMachineFunction();
  bool isStructRet    = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
  bool isThisReturn   = false;
  bool isSibCall      = false;
  // Disable tail calls if they're not supported.
  if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
    isTailCall = false;
  if (isTailCall) {
    // Check if it's really possible to do a tail call.
    isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
                    isVarArg, isStructRet, MF.getFunction()->hasStructRetAttr(),
                                                   Outs, OutVals, Ins, DAG);
    // We don't support GuaranteedTailCallOpt for ARM, only automatically
    // detected sibcalls.
    if (isTailCall) {
      ++NumTailCalls;
      isSibCall = true;
    }
  }

  // Analyze operands of the call, assigning locations to each operand.
  SmallVector<CCValAssign, 16> ArgLocs;
  ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
                 getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
  CCInfo.AnalyzeCallOperands(Outs,
                             CCAssignFnForNode(CallConv, /* Return*/ false,
                                               isVarArg));

  // Get a count of how many bytes are to be pushed on the stack.
  unsigned NumBytes = CCInfo.getNextStackOffset();

  // For tail calls, memory operands are available in our caller's stack.
  if (isSibCall)
    NumBytes = 0;

  // Adjust the stack pointer for the new arguments...
  // These operations are automatically eliminated by the prolog/epilog pass
  if (!isSibCall)
    Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true));

  SDValue StackPtr = DAG.getCopyFromReg(Chain, dl, ARM::SP, getPointerTy());

  RegsToPassVector RegsToPass;
  SmallVector<SDValue, 8> MemOpChains;

  // Walk the register/memloc assignments, inserting copies/loads.  In the case
  // of tail call optimization, arguments are handled later.
  for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
       i != e;
       ++i, ++realArgIdx) {
    CCValAssign &VA = ArgLocs[i];
    SDValue Arg = OutVals[realArgIdx];
    ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
    bool isByVal = Flags.isByVal();

    // Promote the value if needed.
    switch (VA.getLocInfo()) {
    default: llvm_unreachable("Unknown loc info!");
    case CCValAssign::Full: break;
    case CCValAssign::SExt:
      Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), Arg);
      break;
    case CCValAssign::ZExt:
      Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), Arg);
      break;
    case CCValAssign::AExt:
      Arg = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), Arg);
      break;
    case CCValAssign::BCvt:
      Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
      break;
    }

    // f64 and v2f64 might be passed in i32 pairs and must be split into pieces
    if (VA.needsCustom()) {
      if (VA.getLocVT() == MVT::v2f64) {
        SDValue Op0 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
                                  DAG.getConstant(0, MVT::i32));
        SDValue Op1 = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
                                  DAG.getConstant(1, MVT::i32));

        PassF64ArgInRegs(dl, DAG, Chain, Op0, RegsToPass,
                         VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);

        VA = ArgLocs[++i]; // skip ahead to next loc
        if (VA.isRegLoc()) {
          PassF64ArgInRegs(dl, DAG, Chain, Op1, RegsToPass,
                           VA, ArgLocs[++i], StackPtr, MemOpChains, Flags);
        } else {
          assert(VA.isMemLoc());

          MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Op1,
                                                 dl, DAG, VA, Flags));
        }
      } else {
        PassF64ArgInRegs(dl, DAG, Chain, Arg, RegsToPass, VA, ArgLocs[++i],
                         StackPtr, MemOpChains, Flags);
      }
    } else if (VA.isRegLoc()) {
      if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i32) {
        assert(VA.getLocVT() == MVT::i32 &&
               "unexpected calling convention register assignment");
        assert(!Ins.empty() && Ins[0].VT == MVT::i32 &&
               "unexpected use of 'returned'");
        isThisReturn = true;
      }
      RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
    } else if (isByVal) {
      assert(VA.isMemLoc());
      unsigned offset = 0;

      // True if this byval aggregate will be split between registers
      // and memory.
      unsigned ByValArgsCount = CCInfo.getInRegsParamsCount();
      unsigned CurByValIdx = CCInfo.getInRegsParamsProceed();

      if (CurByValIdx < ByValArgsCount) {

        unsigned RegBegin, RegEnd;
        CCInfo.getInRegsParamInfo(CurByValIdx, RegBegin, RegEnd);

        EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
        unsigned int i, j;
        for (i = 0, j = RegBegin; j < RegEnd; i++, j++) {
          SDValue Const = DAG.getConstant(4*i, MVT::i32);
          SDValue AddArg = DAG.getNode(ISD::ADD, dl, PtrVT, Arg, Const);
          SDValue Load = DAG.getLoad(PtrVT, dl, Chain, AddArg,
                                     MachinePointerInfo(),
                                     false, false, false, 0);
          MemOpChains.push_back(Load.getValue(1));
          RegsToPass.push_back(std::make_pair(j, Load));
        }

        // If parameter size outsides register area, "offset" value
        // helps us to calculate stack slot for remained part properly.
        offset = RegEnd - RegBegin;

        CCInfo.nextInRegsParam();
      }

      if (Flags.getByValSize() > 4*offset) {
        unsigned LocMemOffset = VA.getLocMemOffset();
        SDValue StkPtrOff = DAG.getIntPtrConstant(LocMemOffset);
        SDValue Dst = DAG.getNode(ISD::ADD, dl, getPointerTy(), StackPtr,
                                  StkPtrOff);
        SDValue SrcOffset = DAG.getIntPtrConstant(4*offset);
        SDValue Src = DAG.getNode(ISD::ADD, dl, getPointerTy(), Arg, SrcOffset);
        SDValue SizeNode = DAG.getConstant(Flags.getByValSize() - 4*offset,
                                           MVT::i32);
        SDValue AlignNode = DAG.getConstant(Flags.getByValAlign(), MVT::i32);

        SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
        SDValue Ops[] = { Chain, Dst, Src, SizeNode, AlignNode};
        MemOpChains.push_back(DAG.getNode(ARMISD::COPY_STRUCT_BYVAL, dl, VTs,
                                          Ops, array_lengthof(Ops)));
      }
    } else if (!isSibCall) {
      assert(VA.isMemLoc());

      MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
                                             dl, DAG, VA, Flags));
    }
  }

  if (!MemOpChains.empty())
    Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
                        &MemOpChains[0], MemOpChains.size());

  // Build a sequence of copy-to-reg nodes chained together with token chain
  // and flag operands which copy the outgoing args into the appropriate regs.
  SDValue InFlag;
  // Tail call byval lowering might overwrite argument registers so in case of
  // tail call optimization the copies to registers are lowered later.
  if (!isTailCall)
    for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
      Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
                               RegsToPass[i].second, InFlag);
      InFlag = Chain.getValue(1);
    }

  // For tail calls lower the arguments to the 'real' stack slot.
  if (isTailCall) {
    // Force all the incoming stack arguments to be loaded from the stack
    // before any new outgoing arguments are stored to the stack, because the
    // outgoing stack slots may alias the incoming argument stack slots, and
    // the alias isn't otherwise explicit. This is slightly more conservative
    // than necessary, because it means that each store effectively depends
    // on every argument instead of just those arguments it would clobber.

    // Do not flag preceding copytoreg stuff together with the following stuff.
    InFlag = SDValue();
    for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
      Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
                               RegsToPass[i].second, InFlag);
      InFlag = Chain.getValue(1);
    }
    InFlag = SDValue();
  }

  // If the callee is a GlobalAddress/ExternalSymbol node (quite common, every
  // direct call is) turn it into a TargetGlobalAddress/TargetExternalSymbol
  // node so that legalize doesn't hack it.
  bool isDirect = false;
  bool isARMFunc = false;
  bool isLocalARMFunc = false;
  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();

  if (EnableARMLongCalls) {
    assert (getTargetMachine().getRelocationModel() == Reloc::Static
            && "long-calls with non-static relocation model!");
    // Handle a global address or an external symbol. If it's not one of
    // those, the target's already in a register, so we don't need to do
    // anything extra.
    if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
      const GlobalValue *GV = G->getGlobal();
      // Create a constant pool entry for the callee address
      unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
      ARMConstantPoolValue *CPV =
        ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 0);

      // Get the address of the callee into a register
      SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
      CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
      Callee = DAG.getLoad(getPointerTy(), dl,
                           DAG.getEntryNode(), CPAddr,
                           MachinePointerInfo::getConstantPool(),
                           false, false, false, 0);
    } else if (ExternalSymbolSDNode *S=dyn_cast<ExternalSymbolSDNode>(Callee)) {
      const char *Sym = S->getSymbol();

      // Create a constant pool entry for the callee address
      unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
      ARMConstantPoolValue *CPV =
        ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
                                      ARMPCLabelIndex, 0);
      // Get the address of the callee into a register
      SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
      CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
      Callee = DAG.getLoad(getPointerTy(), dl,
                           DAG.getEntryNode(), CPAddr,
                           MachinePointerInfo::getConstantPool(),
                           false, false, false, 0);
    }
  } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
    const GlobalValue *GV = G->getGlobal();
    isDirect = true;
    bool isExt = GV->isDeclaration() || GV->isWeakForLinker();
    bool isStub = (isExt && Subtarget->isTargetDarwin()) &&
                   getTargetMachine().getRelocationModel() != Reloc::Static;
    isARMFunc = !Subtarget->isThumb() || isStub;
    // ARM call to a local ARM function is predicable.
    isLocalARMFunc = !Subtarget->isThumb() && (!isExt || !ARMInterworking);
    // tBX takes a register source operand.
    if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
      unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
      ARMConstantPoolValue *CPV =
        ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue, 4);
      SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
      CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
      Callee = DAG.getLoad(getPointerTy(), dl,
                           DAG.getEntryNode(), CPAddr,
                           MachinePointerInfo::getConstantPool(),
                           false, false, false, 0);
      SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
      Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
                           getPointerTy(), Callee, PICLabel);
    } else {
      // On ELF targets for PIC code, direct calls should go through the PLT
      unsigned OpFlags = 0;
      if (Subtarget->isTargetELF() &&
          getTargetMachine().getRelocationModel() == Reloc::PIC_)
        OpFlags = ARMII::MO_PLT;
      Callee = DAG.getTargetGlobalAddress(GV, dl, getPointerTy(), 0, OpFlags);
    }
  } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
    isDirect = true;
    bool isStub = Subtarget->isTargetDarwin() &&
                  getTargetMachine().getRelocationModel() != Reloc::Static;
    isARMFunc = !Subtarget->isThumb() || isStub;
    // tBX takes a register source operand.
    const char *Sym = S->getSymbol();
    if (isARMFunc && Subtarget->isThumb1Only() && !Subtarget->hasV5TOps()) {
      unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
      ARMConstantPoolValue *CPV =
        ARMConstantPoolSymbol::Create(*DAG.getContext(), Sym,
                                      ARMPCLabelIndex, 4);
      SDValue CPAddr = DAG.getTargetConstantPool(CPV, getPointerTy(), 4);
      CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
      Callee = DAG.getLoad(getPointerTy(), dl,
                           DAG.getEntryNode(), CPAddr,
                           MachinePointerInfo::getConstantPool(),
                           false, false, false, 0);
      SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
      Callee = DAG.getNode(ARMISD::PIC_ADD, dl,
                           getPointerTy(), Callee, PICLabel);
    } else {
      unsigned OpFlags = 0;
      // On ELF targets for PIC code, direct calls should go through the PLT
      if (Subtarget->isTargetELF() &&
                  getTargetMachine().getRelocationModel() == Reloc::PIC_)
        OpFlags = ARMII::MO_PLT;
      Callee = DAG.getTargetExternalSymbol(Sym, getPointerTy(), OpFlags);
    }
  }

  // FIXME: handle tail calls differently.
  unsigned CallOpc;
  bool HasMinSizeAttr = MF.getFunction()->getAttributes().
    hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
  if (Subtarget->isThumb()) {
    if ((!isDirect || isARMFunc) && !Subtarget->hasV5TOps())
      CallOpc = ARMISD::CALL_NOLINK;
    else
      CallOpc = isARMFunc ? ARMISD::CALL : ARMISD::tCALL;
  } else {
    if (!isDirect && !Subtarget->hasV5TOps())
      CallOpc = ARMISD::CALL_NOLINK;
    else if (doesNotRet && isDirect && Subtarget->hasRAS() &&
               // Emit regular call when code size is the priority
               !HasMinSizeAttr)
      // "mov lr, pc; b _foo" to avoid confusing the RSP
      CallOpc = ARMISD::CALL_NOLINK;
    else
      CallOpc = isLocalARMFunc ? ARMISD::CALL_PRED : ARMISD::CALL;
  }

  std::vector<SDValue> Ops;
  Ops.push_back(Chain);
  Ops.push_back(Callee);

  // Add argument registers to the end of the list so that they are known live
  // into the call.
  for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
    Ops.push_back(DAG.getRegister(RegsToPass[i].first,
                                  RegsToPass[i].second.getValueType()));

  // Add a register mask operand representing the call-preserved registers.
  const uint32_t *Mask;
  const TargetRegisterInfo *TRI = getTargetMachine().getRegisterInfo();
  const ARMBaseRegisterInfo *ARI = static_cast<const ARMBaseRegisterInfo*>(TRI);
  if (isThisReturn)
    // For 'this' returns, use the R0-preserving mask
    Mask = ARI->getThisReturnPreservedMask(CallConv);
  else
    Mask = ARI->getCallPreservedMask(CallConv);

  assert(Mask && "Missing call preserved mask for calling convention");
  Ops.push_back(DAG.getRegisterMask(Mask));

  if (InFlag.getNode())
    Ops.push_back(InFlag);

  SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
  if (isTailCall)
    return DAG.getNode(ARMISD::TC_RETURN, dl, NodeTys, &Ops[0], Ops.size());

  // Returns a chain and a flag for retval copy to use.
  Chain = DAG.getNode(CallOpc, dl, NodeTys, &Ops[0], Ops.size());
  InFlag = Chain.getValue(1);

  Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
                             DAG.getIntPtrConstant(0, true), InFlag);
  if (!Ins.empty())
    InFlag = Chain.getValue(1);

  // Handle result values, copying them out of physregs into vregs that we
  // return.
  return LowerCallResult(Chain, InFlag, CallConv, isVarArg, Ins, dl, DAG,
                         InVals, isThisReturn,
                         isThisReturn ? OutVals[0] : SDValue());
}

/// HandleByVal - Every parameter *after* a byval parameter is passed
/// on the stack.  Remember the next parameter register to allocate,
/// and then confiscate the rest of the parameter registers to insure
/// this.
void
ARMTargetLowering::HandleByVal(
    CCState *State, unsigned &size, unsigned Align) const {
  unsigned reg = State->AllocateReg(GPRArgRegs, 4);
  assert((State->getCallOrPrologue() == Prologue ||
          State->getCallOrPrologue() == Call) &&
         "unhandled ParmContext");

  // For in-prologue parameters handling, we also introduce stack offset
  // for byval registers: see CallingConvLower.cpp, CCState::HandleByVal.
  // This behaviour outsides AAPCS rules (5.5 Parameters Passing) of how
  // NSAA should be evaluted (NSAA means "next stacked argument address").
  // So: NextStackOffset = NSAAOffset + SizeOfByValParamsStoredInRegs.
  // Then: NSAAOffset = NextStackOffset - SizeOfByValParamsStoredInRegs.
  unsigned NSAAOffset = State->getNextStackOffset();
  if (State->getCallOrPrologue() != Call) {
    for (unsigned i = 0, e = State->getInRegsParamsCount(); i != e; ++i) {
      unsigned RB, RE;
      State->getInRegsParamInfo(i, RB, RE);
      assert(NSAAOffset >= (RE-RB)*4 &&
             "Stack offset for byval regs doesn't introduced anymore?");
      NSAAOffset -= (RE-RB)*4;
    }
  }
  if ((ARM::R0 <= reg) && (reg <= ARM::R3)) {
    if (Subtarget->isAAPCS_ABI() && Align > 4) {
      unsigned AlignInRegs = Align / 4;
      unsigned Waste = (ARM::R4 - reg) % AlignInRegs;
      for (unsigned i = 0; i < Waste; ++i)
        reg = State->AllocateReg(GPRArgRegs, 4);
    }
    if (reg != 0) {
      unsigned excess = 4 * (ARM::R4 - reg);

      // Special case when NSAA != SP and parameter size greater than size of
      // all remained GPR regs. In that case we can't split parameter, we must
      // send it to stack. We also must set NCRN to R4, so waste all
      // remained registers.
      if (Subtarget->isAAPCS_ABI() && NSAAOffset != 0 && size > excess) {
        while (State->AllocateReg(GPRArgRegs, 4))
          ;
        return;
      }

      // First register for byval parameter is the first register that wasn't
      // allocated before this method call, so it would be "reg".
      // If parameter is small enough to be saved in range [reg, r4), then
      // the end (first after last) register would be reg + param-size-in-regs,
      // else parameter would be splitted between registers and stack,
      // end register would be r4 in this case.
      unsigned ByValRegBegin = reg;
      unsigned ByValRegEnd = (size < excess) ? reg + size/4 : (unsigned)ARM::R4;
      State->addInRegsParamInfo(ByValRegBegin, ByValRegEnd);
      // Note, first register is allocated in the beginning of function already,
      // allocate remained amount of registers we need.
      for (unsigned i = reg+1; i != ByValRegEnd; ++i)
        State->AllocateReg(GPRArgRegs, 4);
      // At a call site, a byval parameter that is split between
      // registers and memory needs its size truncated here.  In a
      // function prologue, such byval parameters are reassembled in
      // memory, and are not truncated.
      if (State->getCallOrPrologue() == Call) {
        // Make remained size equal to 0 in case, when
        // the whole structure may be stored into registers.
        if (size < excess)
          size = 0;
        else
          size -= excess;
      }
    }
  }
}

/// MatchingStackOffset - Return true if the given stack call argument is
/// already available in the same position (relatively) of the caller's
/// incoming argument stack.
static
bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
                         MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
                         const TargetInstrInfo *TII) {
  unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
  int FI = INT_MAX;
  if (Arg.getOpcode() == ISD::CopyFromReg) {
    unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
    if (!TargetRegisterInfo::isVirtualRegister(VR))
      return false;
    MachineInstr *Def = MRI->getVRegDef(VR);
    if (!Def)
      return false;
    if (!Flags.isByVal()) {
      if (!TII->isLoadFromStackSlot(Def, FI))
        return false;
    } else {
      return false;
    }
  } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
    if (Flags.isByVal())
      // ByVal argument is passed in as a pointer but it's now being
      // dereferenced. e.g.
      // define @foo(%struct.X* %A) {
      //   tail call @bar(%struct.X* byval %A)
      // }
      return false;
    SDValue Ptr = Ld->getBasePtr();
    FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
    if (!FINode)
      return false;
    FI = FINode->getIndex();
  } else
    return false;

  assert(FI != INT_MAX);
  if (!MFI->isFixedObjectIndex(FI))
    return false;
  return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
}

/// IsEligibleForTailCallOptimization - Check whether the call is eligible
/// for tail call optimization. Targets which want to do tail call
/// optimization should implement this function.
bool
ARMTargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
                                                     CallingConv::ID CalleeCC,
                                                     bool isVarArg,
                                                     bool isCalleeStructRet,
                                                     bool isCallerStructRet,
                                    const SmallVectorImpl<ISD::OutputArg> &Outs,
                                    const SmallVectorImpl<SDValue> &OutVals,
                                    const SmallVectorImpl<ISD::InputArg> &Ins,
                                                     SelectionDAG& DAG) const {
  const Function *CallerF = DAG.getMachineFunction().getFunction();
  CallingConv::ID CallerCC = CallerF->getCallingConv();
  bool CCMatch = CallerCC == CalleeCC;

  // Look for obvious safe cases to perform tail call optimization that do not
  // require ABI changes. This is what gcc calls sibcall.

  // Do not sibcall optimize vararg calls unless the call site is not passing
  // any arguments.
  if (isVarArg && !Outs.empty())
    return false;

  // Also avoid sibcall optimization if either caller or callee uses struct
  // return semantics.
  if (isCalleeStructRet || isCallerStructRet)
    return false;

  // FIXME: Completely disable sibcall for Thumb1 since Thumb1RegisterInfo::
  // emitEpilogue is not ready for them. Thumb tail calls also use t2B, as
  // the Thumb1 16-bit unconditional branch doesn't have sufficient relocation
  // support in the assembler and linker to be used. This would need to be
  // fixed to fully support tail calls in Thumb1.
  //
  // Doing this is tricky, since the LDM/POP instruction on Thumb doesn't take
  // LR.  This means if we need to reload LR, it takes an extra instructions,
  // which outweighs the value of the tail call; but here we don't know yet
  // whether LR is going to be used.  Probably the right approach is to
  // generate the tail call here and turn it back into CALL/RET in
  // emitEpilogue if LR is used.

  // Thumb1 PIC calls to external symbols use BX, so they can be tail calls,
  // but we need to make sure there are enough registers; the only valid
  // registers are the 4 used for parameters.  We don't currently do this
  // case.
  if (Subtarget->isThumb1Only())
    return false;

  // If the calling conventions do not match, then we'd better make sure the
  // results are returned in the same way as what the caller expects.
  if (!CCMatch) {
    SmallVector<CCValAssign, 16> RVLocs1;
    ARMCCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(),
                       getTargetMachine(), RVLocs1, *DAG.getContext(), Call);
    CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForNode(CalleeCC, true, isVarArg));

    SmallVector<CCValAssign, 16> RVLocs2;
    ARMCCState CCInfo2(CallerCC, false, DAG.getMachineFunction(),
                       getTargetMachine(), RVLocs2, *DAG.getContext(), Call);
    CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForNode(CallerCC, true, isVarArg));

    if (RVLocs1.size() != RVLocs2.size())
      return false;
    for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
      if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
        return false;
      if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
        return false;
      if (RVLocs1[i].isRegLoc()) {
        if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
          return false;
      } else {
        if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
          return false;
      }
    }
  }

  // If Caller's vararg or byval argument has been split between registers and
  // stack, do not perform tail call, since part of the argument is in caller's
  // local frame.
  const ARMFunctionInfo *AFI_Caller = DAG.getMachineFunction().
                                      getInfo<ARMFunctionInfo>();
  if (AFI_Caller->getArgRegsSaveSize())
    return false;

  // If the callee takes no arguments then go on to check the results of the
  // call.
  if (!Outs.empty()) {
    // Check if stack adjustment is needed. For now, do not do this if any
    // argument is passed on the stack.
    SmallVector<CCValAssign, 16> ArgLocs;
    ARMCCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(),
                      getTargetMachine(), ArgLocs, *DAG.getContext(), Call);
    CCInfo.AnalyzeCallOperands(Outs,
                               CCAssignFnForNode(CalleeCC, false, isVarArg));
    if (CCInfo.getNextStackOffset()) {
      MachineFunction &MF = DAG.getMachineFunction();

      // Check if the arguments are already laid out in the right way as
      // the caller's fixed stack objects.
      MachineFrameInfo *MFI = MF.getFrameInfo();
      const MachineRegisterInfo *MRI = &MF.getRegInfo();
      const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
      for (unsigned i = 0, realArgIdx = 0, e = ArgLocs.size();
           i != e;
           ++i, ++realArgIdx) {
        CCValAssign &VA = ArgLocs[i];
        EVT RegVT = VA.getLocVT();
        SDValue Arg = OutVals[realArgIdx];
        ISD::ArgFlagsTy Flags = Outs[realArgIdx].Flags;
        if (VA.getLocInfo() == CCValAssign::Indirect)
          return false;
        if (VA.needsCustom()) {
          // f64 and vector types are split into multiple registers or
          // register/stack-slot combinations.  The types will not match
          // the registers; give up on memory f64 refs until we figure
          // out what to do about this.
          if (!VA.isRegLoc())
            return false;
          if (!ArgLocs[++i].isRegLoc())
            return false;
          if (RegVT == MVT::v2f64) {
            if (!ArgLocs[++i].isRegLoc())
              return false;
            if (!ArgLocs[++i].isRegLoc())
              return false;
          }
        } else if (!VA.isRegLoc()) {
          if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
                                   MFI, MRI, TII))
            return false;
        }
      }
    }
  }

  return true;
}

bool
ARMTargetLowering::CanLowerReturn(CallingConv::ID CallConv,
                                  MachineFunction &MF, bool isVarArg,
                                  const SmallVectorImpl<ISD::OutputArg> &Outs,
                                  LLVMContext &Context) const {
  SmallVector<CCValAssign, 16> RVLocs;
  CCState CCInfo(CallConv, isVarArg, MF, getTargetMachine(), RVLocs, Context);
  return CCInfo.CheckReturn(Outs, CCAssignFnForNode(CallConv, /*Return=*/true,
                                                    isVarArg));
}

SDValue
ARMTargetLowering::LowerReturn(SDValue Chain,
                               CallingConv::ID CallConv, bool isVarArg,
                               const SmallVectorImpl<ISD::OutputArg> &Outs,
                               const SmallVectorImpl<SDValue> &OutVals,
                               DebugLoc dl, SelectionDAG &DAG) const {

  // CCValAssign - represent the assignment of the return value to a location.
  SmallVector<CCValAssign, 16> RVLocs;

  // CCState - Info about the registers and stack slots.
  ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
                    getTargetMachine(), RVLocs, *DAG.getContext(), Call);

  // Analyze outgoing return values.
  CCInfo.AnalyzeReturn(Outs, CCAssignFnForNode(CallConv, /* Return */ true,
                                               isVarArg));

  SDValue Flag;
  SmallVector<SDValue, 4> RetOps;
  RetOps.push_back(Chain); // Operand #0 = Chain (updated below)

  // Copy the result values into the output registers.
  for (unsigned i = 0, realRVLocIdx = 0;
       i != RVLocs.size();
       ++i, ++realRVLocIdx) {
    CCValAssign &VA = RVLocs[i];
    assert(VA.isRegLoc() && "Can only return in registers!");

    SDValue Arg = OutVals[realRVLocIdx];

    switch (VA.getLocInfo()) {
    default: llvm_unreachable("Unknown loc info!");
    case CCValAssign::Full: break;
    case CCValAssign::BCvt:
      Arg = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), Arg);
      break;
    }

    if (VA.needsCustom()) {
      if (VA.getLocVT() == MVT::v2f64) {
        // Extract the first half and return it in two registers.
        SDValue Half = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
                                   DAG.getConstant(0, MVT::i32));
        SDValue HalfGPRs = DAG.getNode(ARMISD::VMOVRRD, dl,
                                       DAG.getVTList(MVT::i32, MVT::i32), Half);

        Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), HalfGPRs, Flag);
        Flag = Chain.getValue(1);
        RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
        VA = RVLocs[++i]; // skip ahead to next loc
        Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(),
                                 HalfGPRs.getValue(1), Flag);
        Flag = Chain.getValue(1);
        RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
        VA = RVLocs[++i]; // skip ahead to next loc

        // Extract the 2nd half and fall through to handle it as an f64 value.
        Arg = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Arg,
                          DAG.getConstant(1, MVT::i32));
      }
      // Legalize ret f64 -> ret 2 x i32.  We always have fmrrd if f64 is
      // available.
      SDValue fmrrd = DAG.getNode(ARMISD::VMOVRRD, dl,
                                  DAG.getVTList(MVT::i32, MVT::i32), &Arg, 1);
      Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd, Flag);
      Flag = Chain.getValue(1);
      RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
      VA = RVLocs[++i]; // skip ahead to next loc
      Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), fmrrd.getValue(1),
                               Flag);
    } else
      Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), Arg, Flag);

    // Guarantee that all emitted copies are
    // stuck together, avoiding something bad.
    Flag = Chain.getValue(1);
    RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
  }

  // Update chain and glue.
  RetOps[0] = Chain;
  if (Flag.getNode())
    RetOps.push_back(Flag);

  return DAG.getNode(ARMISD::RET_FLAG, dl, MVT::Other,
                     RetOps.data(), RetOps.size());
}

bool ARMTargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
  if (N->getNumValues() != 1)
    return false;
  if (!N->hasNUsesOfValue(1, 0))
    return false;

  SDValue TCChain = Chain;
  SDNode *Copy = *N->use_begin();
  if (Copy->getOpcode() == ISD::CopyToReg) {
    // If the copy has a glue operand, we conservatively assume it isn't safe to
    // perform a tail call.
    if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
      return false;
    TCChain = Copy->getOperand(0);
  } else if (Copy->getOpcode() == ARMISD::VMOVRRD) {
    SDNode *VMov = Copy;
    // f64 returned in a pair of GPRs.
    SmallPtrSet<SDNode*, 2> Copies;
    for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
         UI != UE; ++UI) {
      if (UI->getOpcode() != ISD::CopyToReg)
        return false;
      Copies.insert(*UI);
    }
    if (Copies.size() > 2)
      return false;

    for (SDNode::use_iterator UI = VMov->use_begin(), UE = VMov->use_end();
         UI != UE; ++UI) {
      SDValue UseChain = UI->getOperand(0);
      if (Copies.count(UseChain.getNode()))
        // Second CopyToReg
        Copy = *UI;
      else
        // First CopyToReg
        TCChain = UseChain;
    }
  } else if (Copy->getOpcode() == ISD::BITCAST) {
    // f32 returned in a single GPR.
    if (!Copy->hasOneUse())
      return false;
    Copy = *Copy->use_begin();
    if (Copy->getOpcode() != ISD::CopyToReg || !Copy->hasNUsesOfValue(1, 0))
      return false;
    TCChain = Copy->getOperand(0);
  } else {
    return false;
  }

  bool HasRet = false;
  for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
       UI != UE; ++UI) {
    if (UI->getOpcode() != ARMISD::RET_FLAG)
      return false;
    HasRet = true;
  }

  if (!HasRet)
    return false;

  Chain = TCChain;
  return true;
}

bool ARMTargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
  if (!EnableARMTailCalls && !Subtarget->supportsTailCall())
    return false;

  if (!CI->isTailCall())
    return false;

  return !Subtarget->isThumb1Only();
}

// ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
// their target counterpart wrapped in the ARMISD::Wrapper node. Suppose N is
// one of the above mentioned nodes. It has to be wrapped because otherwise
// Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
// be used to form addressing mode. These wrapped nodes will be selected
// into MOVi.
static SDValue LowerConstantPool(SDValue Op, SelectionDAG &DAG) {
  EVT PtrVT = Op.getValueType();
  // FIXME there is no actual debug info here
  DebugLoc dl = Op.getDebugLoc();
  ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
  SDValue Res;
  if (CP->isMachineConstantPoolEntry())
    Res = DAG.getTargetConstantPool(CP->getMachineCPVal(), PtrVT,
                                    CP->getAlignment());
  else
    Res = DAG.getTargetConstantPool(CP->getConstVal(), PtrVT,
                                    CP->getAlignment());
  return DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Res);
}

unsigned ARMTargetLowering::getJumpTableEncoding() const {
  return MachineJumpTableInfo::EK_Inline;
}

SDValue ARMTargetLowering::LowerBlockAddress(SDValue Op,
                                             SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();
  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
  unsigned ARMPCLabelIndex = 0;
  DebugLoc DL = Op.getDebugLoc();
  EVT PtrVT = getPointerTy();
  const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
  Reloc::Model RelocM = getTargetMachine().getRelocationModel();
  SDValue CPAddr;
  if (RelocM == Reloc::Static) {
    CPAddr = DAG.getTargetConstantPool(BA, PtrVT, 4);
  } else {
    unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
    ARMPCLabelIndex = AFI->createPICLabelUId();
    ARMConstantPoolValue *CPV =
      ARMConstantPoolConstant::Create(BA, ARMPCLabelIndex,
                                      ARMCP::CPBlockAddress, PCAdj);
    CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
  }
  CPAddr = DAG.getNode(ARMISD::Wrapper, DL, PtrVT, CPAddr);
  SDValue Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), CPAddr,
                               MachinePointerInfo::getConstantPool(),
                               false, false, false, 0);
  if (RelocM == Reloc::Static)
    return Result;
  SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
  return DAG.getNode(ARMISD::PIC_ADD, DL, PtrVT, Result, PICLabel);
}

// Lower ISD::GlobalTLSAddress using the "general dynamic" model
SDValue
ARMTargetLowering::LowerToTLSGeneralDynamicModel(GlobalAddressSDNode *GA,
                                                 SelectionDAG &DAG) const {
  DebugLoc dl = GA->getDebugLoc();
  EVT PtrVT = getPointerTy();
  unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
  MachineFunction &MF = DAG.getMachineFunction();
  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
  unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
  ARMConstantPoolValue *CPV =
    ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
                                    ARMCP::CPValue, PCAdj, ARMCP::TLSGD, true);
  SDValue Argument = DAG.getTargetConstantPool(CPV, PtrVT, 4);
  Argument = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Argument);
  Argument = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Argument,
                         MachinePointerInfo::getConstantPool(),
                         false, false, false, 0);
  SDValue Chain = Argument.getValue(1);

  SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
  Argument = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Argument, PICLabel);

  // call __tls_get_addr.
  ArgListTy Args;
  ArgListEntry Entry;
  Entry.Node = Argument;
  Entry.Ty = (Type *) Type::getInt32Ty(*DAG.getContext());
  Args.push_back(Entry);
  // FIXME: is there useful debug info available here?
  TargetLowering::CallLoweringInfo CLI(Chain,
                (Type *) Type::getInt32Ty(*DAG.getContext()),
                false, false, false, false,
                0, CallingConv::C, /*isTailCall=*/false,
                /*doesNotRet=*/false, /*isReturnValueUsed=*/true,
                DAG.getExternalSymbol("__tls_get_addr", PtrVT), Args, DAG, dl);
  std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
  return CallResult.first;
}

// Lower ISD::GlobalTLSAddress using the "initial exec" or
// "local exec" model.
SDValue
ARMTargetLowering::LowerToTLSExecModels(GlobalAddressSDNode *GA,
                                        SelectionDAG &DAG,
                                        TLSModel::Model model) const {
  const GlobalValue *GV = GA->getGlobal();
  DebugLoc dl = GA->getDebugLoc();
  SDValue Offset;
  SDValue Chain = DAG.getEntryNode();
  EVT PtrVT = getPointerTy();
  // Get the Thread Pointer
  SDValue ThreadPointer = DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);

  if (model == TLSModel::InitialExec) {
    MachineFunction &MF = DAG.getMachineFunction();
    ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
    unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
    // Initial exec model.
    unsigned char PCAdj = Subtarget->isThumb() ? 4 : 8;
    ARMConstantPoolValue *CPV =
      ARMConstantPoolConstant::Create(GA->getGlobal(), ARMPCLabelIndex,
                                      ARMCP::CPValue, PCAdj, ARMCP::GOTTPOFF,
                                      true);
    Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
    Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
    Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
                         MachinePointerInfo::getConstantPool(),
                         false, false, false, 0);
    Chain = Offset.getValue(1);

    SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
    Offset = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Offset, PICLabel);

    Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
                         MachinePointerInfo::getConstantPool(),
                         false, false, false, 0);
  } else {
    // local exec model
    assert(model == TLSModel::LocalExec);
    ARMConstantPoolValue *CPV =
      ARMConstantPoolConstant::Create(GV, ARMCP::TPOFF);
    Offset = DAG.getTargetConstantPool(CPV, PtrVT, 4);
    Offset = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, Offset);
    Offset = DAG.getLoad(PtrVT, dl, Chain, Offset,
                         MachinePointerInfo::getConstantPool(),
                         false, false, false, 0);
  }

  // The address of the thread local variable is the add of the thread
  // pointer with the offset of the variable.
  return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
}

SDValue
ARMTargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
  // TODO: implement the "local dynamic" model
  assert(Subtarget->isTargetELF() &&
         "TLS not implemented for non-ELF targets");
  GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);

  TLSModel::Model model = getTargetMachine().getTLSModel(GA->getGlobal());

  switch (model) {
    case TLSModel::GeneralDynamic:
    case TLSModel::LocalDynamic:
      return LowerToTLSGeneralDynamicModel(GA, DAG);
    case TLSModel::InitialExec:
    case TLSModel::LocalExec:
      return LowerToTLSExecModels(GA, DAG, model);
  }
  llvm_unreachable("bogus TLS model");
}

SDValue ARMTargetLowering::LowerGlobalAddressELF(SDValue Op,
                                                 SelectionDAG &DAG) const {
  EVT PtrVT = getPointerTy();
  DebugLoc dl = Op.getDebugLoc();
  const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
  if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
    bool UseGOTOFF = GV->hasLocalLinkage() || GV->hasHiddenVisibility();
    ARMConstantPoolValue *CPV =
      ARMConstantPoolConstant::Create(GV,
                                      UseGOTOFF ? ARMCP::GOTOFF : ARMCP::GOT);
    SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
    CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
    SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
                                 CPAddr,
                                 MachinePointerInfo::getConstantPool(),
                                 false, false, false, 0);
    SDValue Chain = Result.getValue(1);
    SDValue GOT = DAG.getGLOBAL_OFFSET_TABLE(PtrVT);
    Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result, GOT);
    if (!UseGOTOFF)
      Result = DAG.getLoad(PtrVT, dl, Chain, Result,
                           MachinePointerInfo::getGOT(),
                           false, false, false, 0);
    return Result;
  }

  // If we have T2 ops, we can materialize the address directly via movt/movw
  // pair. This is always cheaper.
  if (Subtarget->useMovt()) {
    ++NumMovwMovt;
    // FIXME: Once remat is capable of dealing with instructions with register
    // operands, expand this into two nodes.
    return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
                       DAG.getTargetGlobalAddress(GV, dl, PtrVT));
  } else {
    SDValue CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
    CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
    return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
                       MachinePointerInfo::getConstantPool(),
                       false, false, false, 0);
  }
}

SDValue ARMTargetLowering::LowerGlobalAddressDarwin(SDValue Op,
                                                    SelectionDAG &DAG) const {
  EVT PtrVT = getPointerTy();
  DebugLoc dl = Op.getDebugLoc();
  const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
  Reloc::Model RelocM = getTargetMachine().getRelocationModel();

  // FIXME: Enable this for static codegen when tool issues are fixed.  Also
  // update ARMFastISel::ARMMaterializeGV.
  if (Subtarget->useMovt() && RelocM != Reloc::Static) {
    ++NumMovwMovt;
    // FIXME: Once remat is capable of dealing with instructions with register
    // operands, expand this into two nodes.
    if (RelocM == Reloc::Static)
      return DAG.getNode(ARMISD::Wrapper, dl, PtrVT,
                                 DAG.getTargetGlobalAddress(GV, dl, PtrVT));

    unsigned Wrapper = (RelocM == Reloc::PIC_)
      ? ARMISD::WrapperPIC : ARMISD::WrapperDYN;
    SDValue Result = DAG.getNode(Wrapper, dl, PtrVT,
                                 DAG.getTargetGlobalAddress(GV, dl, PtrVT));
    if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
      Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
                           MachinePointerInfo::getGOT(),
                           false, false, false, 0);
    return Result;
  }

  unsigned ARMPCLabelIndex = 0;
  SDValue CPAddr;
  if (RelocM == Reloc::Static) {
    CPAddr = DAG.getTargetConstantPool(GV, PtrVT, 4);
  } else {
    ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
    ARMPCLabelIndex = AFI->createPICLabelUId();
    unsigned PCAdj = (RelocM != Reloc::PIC_) ? 0 : (Subtarget->isThumb()?4:8);
    ARMConstantPoolValue *CPV =
      ARMConstantPoolConstant::Create(GV, ARMPCLabelIndex, ARMCP::CPValue,
                                      PCAdj);
    CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
  }
  CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);

  SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
                               MachinePointerInfo::getConstantPool(),
                               false, false, false, 0);
  SDValue Chain = Result.getValue(1);

  if (RelocM == Reloc::PIC_) {
    SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
    Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
  }

  if (Subtarget->GVIsIndirectSymbol(GV, RelocM))
    Result = DAG.getLoad(PtrVT, dl, Chain, Result, MachinePointerInfo::getGOT(),
                         false, false, false, 0);

  return Result;
}

SDValue ARMTargetLowering::LowerGLOBAL_OFFSET_TABLE(SDValue Op,
                                                    SelectionDAG &DAG) const {
  assert(Subtarget->isTargetELF() &&
         "GLOBAL OFFSET TABLE not implemented for non-ELF targets");
  MachineFunction &MF = DAG.getMachineFunction();
  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
  unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
  EVT PtrVT = getPointerTy();
  DebugLoc dl = Op.getDebugLoc();
  unsigned PCAdj = Subtarget->isThumb() ? 4 : 8;
  ARMConstantPoolValue *CPV =
    ARMConstantPoolSymbol::Create(*DAG.getContext(), "_GLOBAL_OFFSET_TABLE_",
                                  ARMPCLabelIndex, PCAdj);
  SDValue CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
  CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
  SDValue Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
                               MachinePointerInfo::getConstantPool(),
                               false, false, false, 0);
  SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
  return DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
}

SDValue
ARMTargetLowering::LowerEH_SJLJ_SETJMP(SDValue Op, SelectionDAG &DAG) const {
  DebugLoc dl = Op.getDebugLoc();
  SDValue Val = DAG.getConstant(0, MVT::i32);
  return DAG.getNode(ARMISD::EH_SJLJ_SETJMP, dl,
                     DAG.getVTList(MVT::i32, MVT::Other), Op.getOperand(0),
                     Op.getOperand(1), Val);
}

SDValue
ARMTargetLowering::LowerEH_SJLJ_LONGJMP(SDValue Op, SelectionDAG &DAG) const {
  DebugLoc dl = Op.getDebugLoc();
  return DAG.getNode(ARMISD::EH_SJLJ_LONGJMP, dl, MVT::Other, Op.getOperand(0),
                     Op.getOperand(1), DAG.getConstant(0, MVT::i32));
}

SDValue
ARMTargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op, SelectionDAG &DAG,
                                          const ARMSubtarget *Subtarget) const {
  unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
  DebugLoc dl = Op.getDebugLoc();
  switch (IntNo) {
  default: return SDValue();    // Don't custom lower most intrinsics.
  case Intrinsic::arm_thread_pointer: {
    EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
    return DAG.getNode(ARMISD::THREAD_POINTER, dl, PtrVT);
  }
  case Intrinsic::eh_sjlj_lsda: {
    MachineFunction &MF = DAG.getMachineFunction();
    ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
    unsigned ARMPCLabelIndex = AFI->createPICLabelUId();
    EVT PtrVT = getPointerTy();
    Reloc::Model RelocM = getTargetMachine().getRelocationModel();
    SDValue CPAddr;
    unsigned PCAdj = (RelocM != Reloc::PIC_)
      ? 0 : (Subtarget->isThumb() ? 4 : 8);
    ARMConstantPoolValue *CPV =
      ARMConstantPoolConstant::Create(MF.getFunction(), ARMPCLabelIndex,
                                      ARMCP::CPLSDA, PCAdj);
    CPAddr = DAG.getTargetConstantPool(CPV, PtrVT, 4);
    CPAddr = DAG.getNode(ARMISD::Wrapper, dl, MVT::i32, CPAddr);
    SDValue Result =
      DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), CPAddr,
                  MachinePointerInfo::getConstantPool(),
                  false, false, false, 0);

    if (RelocM == Reloc::PIC_) {
      SDValue PICLabel = DAG.getConstant(ARMPCLabelIndex, MVT::i32);
      Result = DAG.getNode(ARMISD::PIC_ADD, dl, PtrVT, Result, PICLabel);
    }
    return Result;
  }
  case Intrinsic::arm_neon_vmulls:
  case Intrinsic::arm_neon_vmullu: {
    unsigned NewOpc = (IntNo == Intrinsic::arm_neon_vmulls)
      ? ARMISD::VMULLs : ARMISD::VMULLu;
    return DAG.getNode(NewOpc, Op.getDebugLoc(), Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  }
  }
}

static SDValue LowerATOMIC_FENCE(SDValue Op, SelectionDAG &DAG,
                                 const ARMSubtarget *Subtarget) {
  // FIXME: handle "fence singlethread" more efficiently.
  DebugLoc dl = Op.getDebugLoc();
  if (!Subtarget->hasDataBarrier()) {
    // Some ARMv6 cpus can support data barriers with an mcr instruction.
    // Thumb1 and pre-v6 ARM mode use a libcall instead and should never get
    // here.
    assert(Subtarget->hasV6Ops() && !Subtarget->isThumb() &&
           "Unexpected ISD::MEMBARRIER encountered. Should be libcall!");
    return DAG.getNode(ARMISD::MEMBARRIER_MCR, dl, MVT::Other, Op.getOperand(0),
                       DAG.getConstant(0, MVT::i32));
  }

  return DAG.getNode(ARMISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0),
                     DAG.getConstant(ARM_MB::ISH, MVT::i32));
}

static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG,
                             const ARMSubtarget *Subtarget) {
  // ARM pre v5TE and Thumb1 does not have preload instructions.
  if (!(Subtarget->isThumb2() ||
        (!Subtarget->isThumb1Only() && Subtarget->hasV5TEOps())))
    // Just preserve the chain.
    return Op.getOperand(0);

  DebugLoc dl = Op.getDebugLoc();
  unsigned isRead = ~cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue() & 1;
  if (!isRead &&
      (!Subtarget->hasV7Ops() || !Subtarget->hasMPExtension()))
    // ARMv7 with MP extension has PLDW.
    return Op.getOperand(0);

  unsigned isData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
  if (Subtarget->isThumb()) {
    // Invert the bits.
    isRead = ~isRead & 1;
    isData = ~isData & 1;
  }

  return DAG.getNode(ARMISD::PRELOAD, dl, MVT::Other, Op.getOperand(0),
                     Op.getOperand(1), DAG.getConstant(isRead, MVT::i32),
                     DAG.getConstant(isData, MVT::i32));
}

static SDValue LowerVASTART(SDValue Op, SelectionDAG &DAG) {
  MachineFunction &MF = DAG.getMachineFunction();
  ARMFunctionInfo *FuncInfo = MF.getInfo<ARMFunctionInfo>();

  // vastart just stores the address of the VarArgsFrameIndex slot into the
  // memory location argument.
  DebugLoc dl = Op.getDebugLoc();
  EVT PtrVT = DAG.getTargetLoweringInfo().getPointerTy();
  SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
  const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
  return DAG.getStore(Op.getOperand(0), dl, FR, Op.getOperand(1),
                      MachinePointerInfo(SV), false, false, 0);
}

SDValue
ARMTargetLowering::GetF64FormalArgument(CCValAssign &VA, CCValAssign &NextVA,
                                        SDValue &Root, SelectionDAG &DAG,
                                        DebugLoc dl) const {
  MachineFunction &MF = DAG.getMachineFunction();
  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();

  const TargetRegisterClass *RC;
  if (AFI->isThumb1OnlyFunction())
    RC = &ARM::tGPRRegClass;
  else
    RC = &ARM::GPRRegClass;

  // Transform the arguments stored in physical registers into virtual ones.
  unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
  SDValue ArgValue = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);

  SDValue ArgValue2;
  if (NextVA.isMemLoc()) {
    MachineFrameInfo *MFI = MF.getFrameInfo();
    int FI = MFI->CreateFixedObject(4, NextVA.getLocMemOffset(), true);

    // Create load node to retrieve arguments from the stack.
    SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
    ArgValue2 = DAG.getLoad(MVT::i32, dl, Root, FIN,
                            MachinePointerInfo::getFixedStack(FI),
                            false, false, false, 0);
  } else {
    Reg = MF.addLiveIn(NextVA.getLocReg(), RC);
    ArgValue2 = DAG.getCopyFromReg(Root, dl, Reg, MVT::i32);
  }

  return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, ArgValue, ArgValue2);
}

void
ARMTargetLowering::computeRegArea(CCState &CCInfo, MachineFunction &MF,
                                  unsigned InRegsParamRecordIdx,
                                  unsigned ArgSize,
                                  unsigned &ArgRegsSize,
                                  unsigned &ArgRegsSaveSize)
  const {
  unsigned NumGPRs;
  if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
    unsigned RBegin, REnd;
    CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
    NumGPRs = REnd - RBegin;
  } else {
    unsigned int firstUnalloced;
    firstUnalloced = CCInfo.getFirstUnallocated(GPRArgRegs,
                                                sizeof(GPRArgRegs) /
                                                sizeof(GPRArgRegs[0]));
    NumGPRs = (firstUnalloced <= 3) ? (4 - firstUnalloced) : 0;
  }

  unsigned Align = MF.getTarget().getFrameLowering()->getStackAlignment();
  ArgRegsSize = NumGPRs * 4;

  // If parameter is split between stack and GPRs...
  if (NumGPRs && Align == 8 &&
      (ArgRegsSize < ArgSize ||
        InRegsParamRecordIdx >= CCInfo.getInRegsParamsCount())) {
    // Add padding for part of param recovered from GPRs, so
    // its last byte must be at address K*8 - 1.
    // We need to do it, since remained (stack) part of parameter has
    // stack alignment, and we need to "attach" "GPRs head" without gaps
    // to it:
    // Stack:
    // |---- 8 bytes block ----| |---- 8 bytes block ----| |---- 8 bytes...
    // [ [padding] [GPRs head] ] [        Tail passed via stack       ....
    //
    ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
    unsigned Padding =
        ((ArgRegsSize + AFI->getArgRegsSaveSize() + Align - 1) & ~(Align-1)) -
        (ArgRegsSize + AFI->getArgRegsSaveSize());
    ArgRegsSaveSize = ArgRegsSize + Padding;
  } else
    // We don't need to extend regs save size for byval parameters if they
    // are passed via GPRs only.
    ArgRegsSaveSize = ArgRegsSize;
}

// The remaining GPRs hold either the beginning of variable-argument
// data, or the beginning of an aggregate passed by value (usually
// byval).  Either way, we allocate stack slots adjacent to the data
// provided by our caller, and store the unallocated registers there.
// If this is a variadic function, the va_list pointer will begin with
// these values; otherwise, this reassembles a (byval) structure that
// was split between registers and memory.
// Return: The frame index registers were stored into.
int
ARMTargetLowering::StoreByValRegs(CCState &CCInfo, SelectionDAG &DAG,
                                  DebugLoc dl, SDValue &Chain,
                                  const Value *OrigArg,
                                  unsigned InRegsParamRecordIdx,
                                  unsigned OffsetFromOrigArg,
                                  unsigned ArgOffset,
                                  unsigned ArgSize,
                                  bool ForceMutable) const {

  // Currently, two use-cases possible:
  // Case #1. Non var-args function, and we meet first byval parameter.
  //          Setup first unallocated register as first byval register;
  //          eat all remained registers
  //          (these two actions are performed by HandleByVal method).
  //          Then, here, we initialize stack frame with
  //          "store-reg" instructions.
  // Case #2. Var-args function, that doesn't contain byval parameters.
  //          The same: eat all remained unallocated registers,
  //          initialize stack frame.

  MachineFunction &MF = DAG.getMachineFunction();
  MachineFrameInfo *MFI = MF.getFrameInfo();
  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();
  unsigned firstRegToSaveIndex, lastRegToSaveIndex;
  unsigned RBegin, REnd;
  if (InRegsParamRecordIdx < CCInfo.getInRegsParamsCount()) {
    CCInfo.getInRegsParamInfo(InRegsParamRecordIdx, RBegin, REnd);
    firstRegToSaveIndex = RBegin - ARM::R0;
    lastRegToSaveIndex = REnd - ARM::R0;
  } else {
    firstRegToSaveIndex = CCInfo.getFirstUnallocated
      (GPRArgRegs, sizeof(GPRArgRegs) / sizeof(GPRArgRegs[0]));
    lastRegToSaveIndex = 4;
  }

  unsigned ArgRegsSize, ArgRegsSaveSize;
  computeRegArea(CCInfo, MF, InRegsParamRecordIdx, ArgSize,
                 ArgRegsSize, ArgRegsSaveSize);

  // Store any by-val regs to their spots on the stack so that they may be
  // loaded by deferencing the result of formal parameter pointer or va_next.
  // Note: once stack area for byval/varargs registers
  // was initialized, it can't be initialized again.
  if (ArgRegsSaveSize) {

    unsigned Padding = ArgRegsSaveSize - ArgRegsSize;

    if (Padding) {
      assert(AFI->getStoredByValParamsPadding() == 0 &&
             "The only parameter may be padded.");
      AFI->setStoredByValParamsPadding(Padding);
    }

    int FrameIndex = MFI->CreateFixedObject(
                      ArgRegsSaveSize,
                      Padding + ArgOffset,
                      false);
    SDValue FIN = DAG.getFrameIndex(FrameIndex, getPointerTy());

    SmallVector<SDValue, 4> MemOps;
    for (unsigned i = 0; firstRegToSaveIndex < lastRegToSaveIndex;
         ++firstRegToSaveIndex, ++i) {
      const TargetRegisterClass *RC;
      if (AFI->isThumb1OnlyFunction())
        RC = &ARM::tGPRRegClass;
      else
        RC = &ARM::GPRRegClass;

      unsigned VReg = MF.addLiveIn(GPRArgRegs[firstRegToSaveIndex], RC);
      SDValue Val = DAG.getCopyFromReg(Chain, dl, VReg, MVT::i32);
      SDValue Store =
        DAG.getStore(Val.getValue(1), dl, Val, FIN,
                     MachinePointerInfo(OrigArg, OffsetFromOrigArg + 4*i),
                     false, false, 0);
      MemOps.push_back(Store);
      FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(), FIN,
                        DAG.getConstant(4, getPointerTy()));
    }

    AFI->setArgRegsSaveSize(ArgRegsSaveSize + AFI->getArgRegsSaveSize());

    if (!MemOps.empty())
      Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
                          &MemOps[0], MemOps.size());
    return FrameIndex;
  } else
    // This will point to the next argument passed via stack.
    return MFI->CreateFixedObject(
        4, AFI->getStoredByValParamsPadding() + ArgOffset, !ForceMutable);
}

// Setup stack frame, the va_list pointer will start from.
void
ARMTargetLowering::VarArgStyleRegisters(CCState &CCInfo, SelectionDAG &DAG,
                                        DebugLoc dl, SDValue &Chain,
                                        unsigned ArgOffset,
                                        bool ForceMutable) const {
  MachineFunction &MF = DAG.getMachineFunction();
  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();

  // Try to store any remaining integer argument regs
  // to their spots on the stack so that they may be loaded by deferencing
  // the result of va_next.
  // If there is no regs to be stored, just point address after last
  // argument passed via stack.
  int FrameIndex =
    StoreByValRegs(CCInfo, DAG, dl, Chain, 0, CCInfo.getInRegsParamsCount(),
                   0, ArgOffset, 0, ForceMutable);

  AFI->setVarArgsFrameIndex(FrameIndex);
}

SDValue
ARMTargetLowering::LowerFormalArguments(SDValue Chain,
                                        CallingConv::ID CallConv, bool isVarArg,
                                        const SmallVectorImpl<ISD::InputArg>
                                          &Ins,
                                        DebugLoc dl, SelectionDAG &DAG,
                                        SmallVectorImpl<SDValue> &InVals)
                                          const {
  MachineFunction &MF = DAG.getMachineFunction();
  MachineFrameInfo *MFI = MF.getFrameInfo();

  ARMFunctionInfo *AFI = MF.getInfo<ARMFunctionInfo>();

  // Assign locations to all of the incoming arguments.
  SmallVector<CCValAssign, 16> ArgLocs;
  ARMCCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(),
                    getTargetMachine(), ArgLocs, *DAG.getContext(), Prologue);
  CCInfo.AnalyzeFormalArguments(Ins,
                                CCAssignFnForNode(CallConv, /* Return*/ false,
                                                  isVarArg));

  SmallVector<SDValue, 16> ArgValues;
  int lastInsIndex = -1;
  SDValue ArgValue;
  Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
  unsigned CurArgIdx = 0;

  // Initially ArgRegsSaveSize is zero.
  // Then we increase this value each time we meet byval parameter.
  // We also increase this value in case of varargs function.
  AFI->setArgRegsSaveSize(0);

  for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
    CCValAssign &VA = ArgLocs[i];
    std::advance(CurOrigArg, Ins[VA.getValNo()].OrigArgIndex - CurArgIdx);
    CurArgIdx = Ins[VA.getValNo()].OrigArgIndex;
    // Arguments stored in registers.
    if (VA.isRegLoc()) {
      EVT RegVT = VA.getLocVT();

      if (VA.needsCustom()) {
        // f64 and vector types are split up into multiple registers or
        // combinations of registers and stack slots.
        if (VA.getLocVT() == MVT::v2f64) {
          SDValue ArgValue1 = GetF64FormalArgument(VA, ArgLocs[++i],
                                                   Chain, DAG, dl);
          VA = ArgLocs[++i]; // skip ahead to next loc
          SDValue ArgValue2;
          if (VA.isMemLoc()) {
            int FI = MFI->CreateFixedObject(8, VA.getLocMemOffset(), true);
            SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
            ArgValue2 = DAG.getLoad(MVT::f64, dl, Chain, FIN,
                                    MachinePointerInfo::getFixedStack(FI),
                                    false, false, false, 0);
          } else {
            ArgValue2 = GetF64FormalArgument(VA, ArgLocs[++i],
                                             Chain, DAG, dl);
          }
          ArgValue = DAG.getNode(ISD::UNDEF, dl, MVT::v2f64);
          ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
                                 ArgValue, ArgValue1, DAG.getIntPtrConstant(0));
          ArgValue = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64,
                                 ArgValue, ArgValue2, DAG.getIntPtrConstant(1));
        } else
          ArgValue = GetF64FormalArgument(VA, ArgLocs[++i], Chain, DAG, dl);

      } else {
        const TargetRegisterClass *RC;

        if (RegVT == MVT::f32)
          RC = &ARM::SPRRegClass;
        else if (RegVT == MVT::f64)
          RC = &ARM::DPRRegClass;
        else if (RegVT == MVT::v2f64)
          RC = &ARM::QPRRegClass;
        else if (RegVT == MVT::i32)
          RC = AFI->isThumb1OnlyFunction() ?
            (const TargetRegisterClass*)&ARM::tGPRRegClass :
            (const TargetRegisterClass*)&ARM::GPRRegClass;
        else
          llvm_unreachable("RegVT not supported by FORMAL_ARGUMENTS Lowering");

        // Transform the arguments in physical registers into virtual ones.
        unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
        ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
      }

      // If this is an 8 or 16-bit value, it is really passed promoted
      // to 32 bits.  Insert an assert[sz]ext to capture this, then
      // truncate to the right size.
      switch (VA.getLocInfo()) {
      default: llvm_unreachable("Unknown loc info!");
      case CCValAssign::Full: break;
      case CCValAssign::BCvt:
        ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
        break;
      case CCValAssign::SExt:
        ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
                               DAG.getValueType(VA.getValVT()));
        ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
        break;
      case CCValAssign::ZExt:
        ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
                               DAG.getValueType(VA.getValVT()));
        ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
        break;
      }

      InVals.push_back(ArgValue);

    } else { // VA.isRegLoc()

      // sanity check
      assert(VA.isMemLoc());
      assert(VA.getValVT() != MVT::i64 && "i64 should already be lowered");

      int index = ArgLocs[i].getValNo();

      // Some Ins[] entries become multiple ArgLoc[] entries.
      // Process them only once.
      if (index != lastInsIndex)
        {
          ISD::ArgFlagsTy Flags = Ins[index].Flags;
          // FIXME: For now, all byval parameter objects are marked mutable.
          // This can be changed with more analysis.
          // In case of tail call optimization mark all arguments mutable.
          // Since they could be overwritten by lowering of arguments in case of
          // a tail call.
          if (Flags.isByVal()) {
            unsigned CurByValIndex = CCInfo.getInRegsParamsProceed();
            int FrameIndex = StoreByValRegs(
                CCInfo, DAG, dl, Chain, CurOrigArg,
                CurByValIndex,
                Ins[VA.getValNo()].PartOffset,
                VA.getLocMemOffset(),
                Flags.getByValSize(),
                true /*force mutable frames*/);
            InVals.push_back(DAG.getFrameIndex(FrameIndex, getPointerTy()));
            CCInfo.nextInRegsParam();
          } else {
            unsigned FIOffset = VA.getLocMemOffset() +
                                AFI->getStoredByValParamsPadding();
            int FI = MFI->CreateFixedObject(VA.getLocVT().getSizeInBits()/8,
                                            FIOffset, true);

            // Create load nodes to retrieve arguments from the stack.
            SDValue FIN = DAG.getFrameIndex(FI, getPointerTy());
            InVals.push_back(DAG.getLoad(VA.getValVT(), dl, Chain, FIN,
                                         MachinePointerInfo::getFixedStack(FI),
                                         false, false, false, 0));
          }
          lastInsIndex = index;
        }
    }
  }

  // varargs
  if (isVarArg)
    VarArgStyleRegisters(CCInfo, DAG, dl, Chain,
                         CCInfo.getNextStackOffset());

  return Chain;
}

/// isFloatingPointZero - Return true if this is +0.0.
static bool isFloatingPointZero(SDValue Op) {
  if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Op))
    return CFP->getValueAPF().isPosZero();
  else if (ISD::isEXTLoad(Op.getNode()) || ISD::isNON_EXTLoad(Op.getNode())) {
    // Maybe this has already been legalized into the constant pool?
    if (Op.getOperand(1).getOpcode() == ARMISD::Wrapper) {
      SDValue WrapperOp = Op.getOperand(1).getOperand(0);
      if (ConstantPoolSDNode *CP = dyn_cast<ConstantPoolSDNode>(WrapperOp))
        if (const ConstantFP *CFP = dyn_cast<ConstantFP>(CP->getConstVal()))
          return CFP->getValueAPF().isPosZero();
    }
  }
  return false;
}

/// Returns appropriate ARM CMP (cmp) and corresponding condition code for
/// the given operands.
SDValue
ARMTargetLowering::getARMCmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
                             SDValue &ARMcc, SelectionDAG &DAG,
                             DebugLoc dl) const {
  if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
    unsigned C = RHSC->getZExtValue();
    if (!isLegalICmpImmediate(C)) {
      // Constant does not fit, try adjusting it by one?
      switch (CC) {
      default: break;
      case ISD::SETLT:
      case ISD::SETGE:
        if (C != 0x80000000 && isLegalICmpImmediate(C-1)) {
          CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
          RHS = DAG.getConstant(C-1, MVT::i32);
        }
        break;
      case ISD::SETULT:
      case ISD::SETUGE:
        if (C != 0 && isLegalICmpImmediate(C-1)) {
          CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
          RHS = DAG.getConstant(C-1, MVT::i32);
        }
        break;
      case ISD::SETLE:
      case ISD::SETGT:
        if (C != 0x7fffffff && isLegalICmpImmediate(C+1)) {
          CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
          RHS = DAG.getConstant(C+1, MVT::i32);
        }
        break;
      case ISD::SETULE:
      case ISD::SETUGT:
        if (C != 0xffffffff && isLegalICmpImmediate(C+1)) {
          CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
          RHS = DAG.getConstant(C+1, MVT::i32);
        }
        break;
      }
    }
  }

  ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
  ARMISD::NodeType CompareType;
  switch (CondCode) {
  default:
    CompareType = ARMISD::CMP;
    break;
  case ARMCC::EQ:
  case ARMCC::NE:
    // Uses only Z Flag
    CompareType = ARMISD::CMPZ;
    break;
  }
  ARMcc = DAG.getConstant(CondCode, MVT::i32);
  return DAG.getNode(CompareType, dl, MVT::Glue, LHS, RHS);
}

/// Returns a appropriate VFP CMP (fcmp{s|d}+fmstat) for the given operands.
SDValue
ARMTargetLowering::getVFPCmp(SDValue LHS, SDValue RHS, SelectionDAG &DAG,
                             DebugLoc dl) const {
  SDValue Cmp;
  if (!isFloatingPointZero(RHS))
    Cmp = DAG.getNode(ARMISD::CMPFP, dl, MVT::Glue, LHS, RHS);
  else
    Cmp = DAG.getNode(ARMISD::CMPFPw0, dl, MVT::Glue, LHS);
  return DAG.getNode(ARMISD::FMSTAT, dl, MVT::Glue, Cmp);
}

/// duplicateCmp - Glue values can have only one use, so this function
/// duplicates a comparison node.
SDValue
ARMTargetLowering::duplicateCmp(SDValue Cmp, SelectionDAG &DAG) const {
  unsigned Opc = Cmp.getOpcode();
  DebugLoc DL = Cmp.getDebugLoc();
  if (Opc == ARMISD::CMP || Opc == ARMISD::CMPZ)
    return DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));

  assert(Opc == ARMISD::FMSTAT && "unexpected comparison operation");
  Cmp = Cmp.getOperand(0);
  Opc = Cmp.getOpcode();
  if (Opc == ARMISD::CMPFP)
    Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0),Cmp.getOperand(1));
  else {
    assert(Opc == ARMISD::CMPFPw0 && "unexpected operand of FMSTAT");
    Cmp = DAG.getNode(Opc, DL, MVT::Glue, Cmp.getOperand(0));
  }
  return DAG.getNode(ARMISD::FMSTAT, DL, MVT::Glue, Cmp);
}

SDValue ARMTargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
  SDValue Cond = Op.getOperand(0);
  SDValue SelectTrue = Op.getOperand(1);
  SDValue SelectFalse = Op.getOperand(2);
  DebugLoc dl = Op.getDebugLoc();

  // Convert:
  //
  //   (select (cmov 1, 0, cond), t, f) -> (cmov t, f, cond)
  //   (select (cmov 0, 1, cond), t, f) -> (cmov f, t, cond)
  //
  if (Cond.getOpcode() == ARMISD::CMOV && Cond.hasOneUse()) {
    const ConstantSDNode *CMOVTrue =
      dyn_cast<ConstantSDNode>(Cond.getOperand(0));
    const ConstantSDNode *CMOVFalse =
      dyn_cast<ConstantSDNode>(Cond.getOperand(1));

    if (CMOVTrue && CMOVFalse) {
      unsigned CMOVTrueVal = CMOVTrue->getZExtValue();
      unsigned CMOVFalseVal = CMOVFalse->getZExtValue();

      SDValue True;
      SDValue False;
      if (CMOVTrueVal == 1 && CMOVFalseVal == 0) {
        True = SelectTrue;
        False = SelectFalse;
      } else if (CMOVTrueVal == 0 && CMOVFalseVal == 1) {
        True = SelectFalse;
        False = SelectTrue;
      }

      if (True.getNode() && False.getNode()) {
        EVT VT = Op.getValueType();
        SDValue ARMcc = Cond.getOperand(2);
        SDValue CCR = Cond.getOperand(3);
        SDValue Cmp = duplicateCmp(Cond.getOperand(4), DAG);
        assert(True.getValueType() == VT);
        return DAG.getNode(ARMISD::CMOV, dl, VT, True, False, ARMcc, CCR, Cmp);
      }
    }
  }

  // ARM's BooleanContents value is UndefinedBooleanContent. Mask out the
  // undefined bits before doing a full-word comparison with zero.
  Cond = DAG.getNode(ISD::AND, dl, Cond.getValueType(), Cond,
                     DAG.getConstant(1, Cond.getValueType()));

  return DAG.getSelectCC(dl, Cond,
                         DAG.getConstant(0, Cond.getValueType()),
                         SelectTrue, SelectFalse, ISD::SETNE);
}

SDValue ARMTargetLowering::LowerSELECT_CC(SDValue Op, SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  SDValue LHS = Op.getOperand(0);
  SDValue RHS = Op.getOperand(1);
  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
  SDValue TrueVal = Op.getOperand(2);
  SDValue FalseVal = Op.getOperand(3);
  DebugLoc dl = Op.getDebugLoc();

  if (LHS.getValueType() == MVT::i32) {
    SDValue ARMcc;
    SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
    SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
    return DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc, CCR,Cmp);
  }

  ARMCC::CondCodes CondCode, CondCode2;
  FPCCToARMCC(CC, CondCode, CondCode2);

  SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
  SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
  SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
  SDValue Result = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal,
                               ARMcc, CCR, Cmp);
  if (CondCode2 != ARMCC::AL) {
    SDValue ARMcc2 = DAG.getConstant(CondCode2, MVT::i32);
    // FIXME: Needs another CMP because flag can have but one use.
    SDValue Cmp2 = getVFPCmp(LHS, RHS, DAG, dl);
    Result = DAG.getNode(ARMISD::CMOV, dl, VT,
                         Result, TrueVal, ARMcc2, CCR, Cmp2);
  }
  return Result;
}

/// canChangeToInt - Given the fp compare operand, return true if it is suitable
/// to morph to an integer compare sequence.
static bool canChangeToInt(SDValue Op, bool &SeenZero,
                           const ARMSubtarget *Subtarget) {
  SDNode *N = Op.getNode();
  if (!N->hasOneUse())
    // Otherwise it requires moving the value from fp to integer registers.
    return false;
  if (!N->getNumValues())
    return false;
  EVT VT = Op.getValueType();
  if (VT != MVT::f32 && !Subtarget->isFPBrccSlow())
    // f32 case is generally profitable. f64 case only makes sense when vcmpe +
    // vmrs are very slow, e.g. cortex-a8.
    return false;

  if (isFloatingPointZero(Op)) {
    SeenZero = true;
    return true;
  }
  return ISD::isNormalLoad(N);
}

static SDValue bitcastf32Toi32(SDValue Op, SelectionDAG &DAG) {
  if (isFloatingPointZero(Op))
    return DAG.getConstant(0, MVT::i32);

  if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op))
    return DAG.getLoad(MVT::i32, Op.getDebugLoc(),
                       Ld->getChain(), Ld->getBasePtr(), Ld->getPointerInfo(),
                       Ld->isVolatile(), Ld->isNonTemporal(),
                       Ld->isInvariant(), Ld->getAlignment());

  llvm_unreachable("Unknown VFP cmp argument!");
}

static void expandf64Toi32(SDValue Op, SelectionDAG &DAG,
                           SDValue &RetVal1, SDValue &RetVal2) {
  if (isFloatingPointZero(Op)) {
    RetVal1 = DAG.getConstant(0, MVT::i32);
    RetVal2 = DAG.getConstant(0, MVT::i32);
    return;
  }

  if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Op)) {
    SDValue Ptr = Ld->getBasePtr();
    RetVal1 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
                          Ld->getChain(), Ptr,
                          Ld->getPointerInfo(),
                          Ld->isVolatile(), Ld->isNonTemporal(),
                          Ld->isInvariant(), Ld->getAlignment());

    EVT PtrType = Ptr.getValueType();
    unsigned NewAlign = MinAlign(Ld->getAlignment(), 4);
    SDValue NewPtr = DAG.getNode(ISD::ADD, Op.getDebugLoc(),
                                 PtrType, Ptr, DAG.getConstant(4, PtrType));
    RetVal2 = DAG.getLoad(MVT::i32, Op.getDebugLoc(),
                          Ld->getChain(), NewPtr,
                          Ld->getPointerInfo().getWithOffset(4),
                          Ld->isVolatile(), Ld->isNonTemporal(),
                          Ld->isInvariant(), NewAlign);
    return;
  }

  llvm_unreachable("Unknown VFP cmp argument!");
}

/// OptimizeVFPBrcond - With -enable-unsafe-fp-math, it's legal to optimize some
/// f32 and even f64 comparisons to integer ones.
SDValue
ARMTargetLowering::OptimizeVFPBrcond(SDValue Op, SelectionDAG &DAG) const {
  SDValue Chain = Op.getOperand(0);
  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
  SDValue LHS = Op.getOperand(2);
  SDValue RHS = Op.getOperand(3);
  SDValue Dest = Op.getOperand(4);
  DebugLoc dl = Op.getDebugLoc();

  bool LHSSeenZero = false;
  bool LHSOk = canChangeToInt(LHS, LHSSeenZero, Subtarget);
  bool RHSSeenZero = false;
  bool RHSOk = canChangeToInt(RHS, RHSSeenZero, Subtarget);
  if (LHSOk && RHSOk && (LHSSeenZero || RHSSeenZero)) {
    // If unsafe fp math optimization is enabled and there are no other uses of
    // the CMP operands, and the condition code is EQ or NE, we can optimize it
    // to an integer comparison.
    if (CC == ISD::SETOEQ)
      CC = ISD::SETEQ;
    else if (CC == ISD::SETUNE)
      CC = ISD::SETNE;

    SDValue Mask = DAG.getConstant(0x7fffffff, MVT::i32);
    SDValue ARMcc;
    if (LHS.getValueType() == MVT::f32) {
      LHS = DAG.getNode(ISD::AND, dl, MVT::i32,
                        bitcastf32Toi32(LHS, DAG), Mask);
      RHS = DAG.getNode(ISD::AND, dl, MVT::i32,
                        bitcastf32Toi32(RHS, DAG), Mask);
      SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
      SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
      return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
                         Chain, Dest, ARMcc, CCR, Cmp);
    }

    SDValue LHS1, LHS2;
    SDValue RHS1, RHS2;
    expandf64Toi32(LHS, DAG, LHS1, LHS2);
    expandf64Toi32(RHS, DAG, RHS1, RHS2);
    LHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, LHS2, Mask);
    RHS2 = DAG.getNode(ISD::AND, dl, MVT::i32, RHS2, Mask);
    ARMCC::CondCodes CondCode = IntCCToARMCC(CC);
    ARMcc = DAG.getConstant(CondCode, MVT::i32);
    SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
    SDValue Ops[] = { Chain, ARMcc, LHS1, LHS2, RHS1, RHS2, Dest };
    return DAG.getNode(ARMISD::BCC_i64, dl, VTList, Ops, 7);
  }

  return SDValue();
}

SDValue ARMTargetLowering::LowerBR_CC(SDValue Op, SelectionDAG &DAG) const {
  SDValue Chain = Op.getOperand(0);
  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(1))->get();
  SDValue LHS = Op.getOperand(2);
  SDValue RHS = Op.getOperand(3);
  SDValue Dest = Op.getOperand(4);
  DebugLoc dl = Op.getDebugLoc();

  if (LHS.getValueType() == MVT::i32) {
    SDValue ARMcc;
    SDValue Cmp = getARMCmp(LHS, RHS, CC, ARMcc, DAG, dl);
    SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
    return DAG.getNode(ARMISD::BRCOND, dl, MVT::Other,
                       Chain, Dest, ARMcc, CCR, Cmp);
  }

  assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);

  if (getTargetMachine().Options.UnsafeFPMath &&
      (CC == ISD::SETEQ || CC == ISD::SETOEQ ||
       CC == ISD::SETNE || CC == ISD::SETUNE)) {
    SDValue Result = OptimizeVFPBrcond(Op, DAG);
    if (Result.getNode())
      return Result;
  }

  ARMCC::CondCodes CondCode, CondCode2;
  FPCCToARMCC(CC, CondCode, CondCode2);

  SDValue ARMcc = DAG.getConstant(CondCode, MVT::i32);
  SDValue Cmp = getVFPCmp(LHS, RHS, DAG, dl);
  SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
  SDVTList VTList = DAG.getVTList(MVT::Other, MVT::Glue);
  SDValue Ops[] = { Chain, Dest, ARMcc, CCR, Cmp };
  SDValue Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
  if (CondCode2 != ARMCC::AL) {
    ARMcc = DAG.getConstant(CondCode2, MVT::i32);
    SDValue Ops[] = { Res, Dest, ARMcc, CCR, Res.getValue(1) };
    Res = DAG.getNode(ARMISD::BRCOND, dl, VTList, Ops, 5);
  }
  return Res;
}

SDValue ARMTargetLowering::LowerBR_JT(SDValue Op, SelectionDAG &DAG) const {
  SDValue Chain = Op.getOperand(0);
  SDValue Table = Op.getOperand(1);
  SDValue Index = Op.getOperand(2);
  DebugLoc dl = Op.getDebugLoc();

  EVT PTy = getPointerTy();
  JumpTableSDNode *JT = cast<JumpTableSDNode>(Table);
  ARMFunctionInfo *AFI = DAG.getMachineFunction().getInfo<ARMFunctionInfo>();
  SDValue UId = DAG.getConstant(AFI->createJumpTableUId(), PTy);
  SDValue JTI = DAG.getTargetJumpTable(JT->getIndex(), PTy);
  Table = DAG.getNode(ARMISD::WrapperJT, dl, MVT::i32, JTI, UId);
  Index = DAG.getNode(ISD::MUL, dl, PTy, Index, DAG.getConstant(4, PTy));
  SDValue Addr = DAG.getNode(ISD::ADD, dl, PTy, Index, Table);
  if (Subtarget->isThumb2()) {
    // Thumb2 uses a two-level jump. That is, it jumps into the jump table
    // which does another jump to the destination. This also makes it easier
    // to translate it to TBB / TBH later.
    // FIXME: This might not work if the function is extremely large.
    return DAG.getNode(ARMISD::BR2_JT, dl, MVT::Other, Chain,
                       Addr, Op.getOperand(2), JTI, UId);
  }
  if (getTargetMachine().getRelocationModel() == Reloc::PIC_) {
    Addr = DAG.getLoad((EVT)MVT::i32, dl, Chain, Addr,
                       MachinePointerInfo::getJumpTable(),
                       false, false, false, 0);
    Chain = Addr.getValue(1);
    Addr = DAG.getNode(ISD::ADD, dl, PTy, Addr, Table);
    return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
  } else {
    Addr = DAG.getLoad(PTy, dl, Chain, Addr,
                       MachinePointerInfo::getJumpTable(),
                       false, false, false, 0);
    Chain = Addr.getValue(1);
    return DAG.getNode(ARMISD::BR_JT, dl, MVT::Other, Chain, Addr, JTI, UId);
  }
}

static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
  EVT VT = Op.getValueType();
  DebugLoc dl = Op.getDebugLoc();

  if (Op.getValueType().getVectorElementType() == MVT::i32) {
    if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::f32)
      return Op;
    return DAG.UnrollVectorOp(Op.getNode());
  }

  assert(Op.getOperand(0).getValueType() == MVT::v4f32 &&
         "Invalid type for custom lowering!");
  if (VT != MVT::v4i16)
    return DAG.UnrollVectorOp(Op.getNode());

  Op = DAG.getNode(Op.getOpcode(), dl, MVT::v4i32, Op.getOperand(0));
  return DAG.getNode(ISD::TRUNCATE, dl, VT, Op);
}

static SDValue LowerFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
  EVT VT = Op.getValueType();
  if (VT.isVector())
    return LowerVectorFP_TO_INT(Op, DAG);

  DebugLoc dl = Op.getDebugLoc();
  unsigned Opc;

  switch (Op.getOpcode()) {
  default: llvm_unreachable("Invalid opcode!");
  case ISD::FP_TO_SINT:
    Opc = ARMISD::FTOSI;
    break;
  case ISD::FP_TO_UINT:
    Opc = ARMISD::FTOUI;
    break;
  }
  Op = DAG.getNode(Opc, dl, MVT::f32, Op.getOperand(0));
  return DAG.getNode(ISD::BITCAST, dl, MVT::i32, Op);
}

static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
  EVT VT = Op.getValueType();
  DebugLoc dl = Op.getDebugLoc();

  if (Op.getOperand(0).getValueType().getVectorElementType() == MVT::i32) {
    if (VT.getVectorElementType() == MVT::f32)
      return Op;
    return DAG.UnrollVectorOp(Op.getNode());
  }

  assert(Op.getOperand(0).getValueType() == MVT::v4i16 &&
         "Invalid type for custom lowering!");
  if (VT != MVT::v4f32)
    return DAG.UnrollVectorOp(Op.getNode());

  unsigned CastOpc;
  unsigned Opc;
  switch (Op.getOpcode()) {
  default: llvm_unreachable("Invalid opcode!");
  case ISD::SINT_TO_FP:
    CastOpc = ISD::SIGN_EXTEND;
    Opc = ISD::SINT_TO_FP;
    break;
  case ISD::UINT_TO_FP:
    CastOpc = ISD::ZERO_EXTEND;
    Opc = ISD::UINT_TO_FP;
    break;
  }

  Op = DAG.getNode(CastOpc, dl, MVT::v4i32, Op.getOperand(0));
  return DAG.getNode(Opc, dl, VT, Op);
}

static SDValue LowerINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
  EVT VT = Op.getValueType();
  if (VT.isVector())
    return LowerVectorINT_TO_FP(Op, DAG);

  DebugLoc dl = Op.getDebugLoc();
  unsigned Opc;

  switch (Op.getOpcode()) {
  default: llvm_unreachable("Invalid opcode!");
  case ISD::SINT_TO_FP:
    Opc = ARMISD::SITOF;
    break;
  case ISD::UINT_TO_FP:
    Opc = ARMISD::UITOF;
    break;
  }

  Op = DAG.getNode(ISD::BITCAST, dl, MVT::f32, Op.getOperand(0));
  return DAG.getNode(Opc, dl, VT, Op);
}

SDValue ARMTargetLowering::LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) const {
  // Implement fcopysign with a fabs and a conditional fneg.
  SDValue Tmp0 = Op.getOperand(0);
  SDValue Tmp1 = Op.getOperand(1);
  DebugLoc dl = Op.getDebugLoc();
  EVT VT = Op.getValueType();
  EVT SrcVT = Tmp1.getValueType();
  bool InGPR = Tmp0.getOpcode() == ISD::BITCAST ||
    Tmp0.getOpcode() == ARMISD::VMOVDRR;
  bool UseNEON = !InGPR && Subtarget->hasNEON();

  if (UseNEON) {
    // Use VBSL to copy the sign bit.
    unsigned EncodedVal = ARM_AM::createNEONModImm(0x6, 0x80);
    SDValue Mask = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v2i32,
                               DAG.getTargetConstant(EncodedVal, MVT::i32));
    EVT OpVT = (VT == MVT::f32) ? MVT::v2i32 : MVT::v1i64;
    if (VT == MVT::f64)
      Mask = DAG.getNode(ARMISD::VSHL, dl, OpVT,
                         DAG.getNode(ISD::BITCAST, dl, OpVT, Mask),
                         DAG.getConstant(32, MVT::i32));
    else /*if (VT == MVT::f32)*/
      Tmp0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp0);
    if (SrcVT == MVT::f32) {
      Tmp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f32, Tmp1);
      if (VT == MVT::f64)
        Tmp1 = DAG.getNode(ARMISD::VSHL, dl, OpVT,
                           DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1),
                           DAG.getConstant(32, MVT::i32));
    } else if (VT == MVT::f32)
      Tmp1 = DAG.getNode(ARMISD::VSHRu, dl, MVT::v1i64,
                         DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, Tmp1),
                         DAG.getConstant(32, MVT::i32));
    Tmp0 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp0);
    Tmp1 = DAG.getNode(ISD::BITCAST, dl, OpVT, Tmp1);

    SDValue AllOnes = DAG.getTargetConstant(ARM_AM::createNEONModImm(0xe, 0xff),
                                            MVT::i32);
    AllOnes = DAG.getNode(ARMISD::VMOVIMM, dl, MVT::v8i8, AllOnes);
    SDValue MaskNot = DAG.getNode(ISD::XOR, dl, OpVT, Mask,
                                  DAG.getNode(ISD::BITCAST, dl, OpVT, AllOnes));

    SDValue Res = DAG.getNode(ISD::OR, dl, OpVT,
                              DAG.getNode(ISD::AND, dl, OpVT, Tmp1, Mask),
                              DAG.getNode(ISD::AND, dl, OpVT, Tmp0, MaskNot));
    if (VT == MVT::f32) {
      Res = DAG.getNode(ISD::BITCAST, dl, MVT::v2f32, Res);
      Res = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, Res,
                        DAG.getConstant(0, MVT::i32));
    } else {
      Res = DAG.getNode(ISD::BITCAST, dl, MVT::f64, Res);
    }

    return Res;
  }

  // Bitcast operand 1 to i32.
  if (SrcVT == MVT::f64)
    Tmp1 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
                       &Tmp1, 1).getValue(1);
  Tmp1 = DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp1);

  // Or in the signbit with integer operations.
  SDValue Mask1 = DAG.getConstant(0x80000000, MVT::i32);
  SDValue Mask2 = DAG.getConstant(0x7fffffff, MVT::i32);
  Tmp1 = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp1, Mask1);
  if (VT == MVT::f32) {
    Tmp0 = DAG.getNode(ISD::AND, dl, MVT::i32,
                       DAG.getNode(ISD::BITCAST, dl, MVT::i32, Tmp0), Mask2);
    return DAG.getNode(ISD::BITCAST, dl, MVT::f32,
                       DAG.getNode(ISD::OR, dl, MVT::i32, Tmp0, Tmp1));
  }

  // f64: Or the high part with signbit and then combine two parts.
  Tmp0 = DAG.getNode(ARMISD::VMOVRRD, dl, DAG.getVTList(MVT::i32, MVT::i32),
                     &Tmp0, 1);
  SDValue Lo = Tmp0.getValue(0);
  SDValue Hi = DAG.getNode(ISD::AND, dl, MVT::i32, Tmp0.getValue(1), Mask2);
  Hi = DAG.getNode(ISD::OR, dl, MVT::i32, Hi, Tmp1);
  return DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi);
}

SDValue ARMTargetLowering::LowerRETURNADDR(SDValue Op, SelectionDAG &DAG) const{
  MachineFunction &MF = DAG.getMachineFunction();
  MachineFrameInfo *MFI = MF.getFrameInfo();
  MFI->setReturnAddressIsTaken(true);

  EVT VT = Op.getValueType();
  DebugLoc dl = Op.getDebugLoc();
  unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
  if (Depth) {
    SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
    SDValue Offset = DAG.getConstant(4, MVT::i32);
    return DAG.getLoad(VT, dl, DAG.getEntryNode(),
                       DAG.getNode(ISD::ADD, dl, VT, FrameAddr, Offset),
                       MachinePointerInfo(), false, false, false, 0);
  }

  // Return LR, which contains the return address. Mark it an implicit live-in.
  unsigned Reg = MF.addLiveIn(ARM::LR, getRegClassFor(MVT::i32));
  return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
}

SDValue ARMTargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
  MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
  MFI->setFrameAddressIsTaken(true);

  EVT VT = Op.getValueType();
  DebugLoc dl = Op.getDebugLoc();  // FIXME probably not meaningful
  unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
  unsigned FrameReg = (Subtarget->isThumb() || Subtarget->isTargetDarwin())
    ? ARM::R7 : ARM::R11;
  SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
  while (Depth--)
    FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
                            MachinePointerInfo(),
                            false, false, false, 0);
  return FrameAddr;
}

/// Custom Expand long vector extensions, where size(DestVec) > 2*size(SrcVec),
/// and size(DestVec) > 128-bits.
/// This is achieved by doing the one extension from the SrcVec, splitting the
/// result, extending these parts, and then concatenating these into the
/// destination.
static SDValue ExpandVectorExtension(SDNode *N, SelectionDAG &DAG) {
  SDValue Op = N->getOperand(0);
  EVT SrcVT = Op.getValueType();
  EVT DestVT = N->getValueType(0);

  assert(DestVT.getSizeInBits() > 128 &&
         "Custom sext/zext expansion needs >128-bit vector.");
  // If this is a normal length extension, use the default expansion.
  if (SrcVT.getSizeInBits()*4 != DestVT.getSizeInBits() &&
      SrcVT.getSizeInBits()*8 != DestVT.getSizeInBits())
    return SDValue();

  DebugLoc dl = N->getDebugLoc();
  unsigned SrcEltSize = SrcVT.getVectorElementType().getSizeInBits();
  unsigned DestEltSize = DestVT.getVectorElementType().getSizeInBits();
  unsigned NumElts = SrcVT.getVectorNumElements();
  LLVMContext &Ctx = *DAG.getContext();
  SDValue Mid, SplitLo, SplitHi, ExtLo, ExtHi;

  EVT MidVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, SrcEltSize*2),
                               NumElts);
  EVT SplitVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, SrcEltSize*2),
                                 NumElts/2);
  EVT ExtVT = EVT::getVectorVT(Ctx, EVT::getIntegerVT(Ctx, DestEltSize),
                               NumElts/2);

  Mid = DAG.getNode(N->getOpcode(), dl, MidVT, Op);
  SplitLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SplitVT, Mid,
                        DAG.getIntPtrConstant(0));
  SplitHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SplitVT, Mid,
                        DAG.getIntPtrConstant(NumElts/2));
  ExtLo = DAG.getNode(N->getOpcode(), dl, ExtVT, SplitLo);
  ExtHi = DAG.getNode(N->getOpcode(), dl, ExtVT, SplitHi);
  return DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, ExtLo, ExtHi);
}

/// ExpandBITCAST - If the target supports VFP, this function is called to
/// expand a bit convert where either the source or destination type is i64 to
/// use a VMOVDRR or VMOVRRD node.  This should not be done when the non-i64
/// operand type is illegal (e.g., v2f32 for a target that doesn't support
/// vectors), since the legalizer won't know what to do with that.
static SDValue ExpandBITCAST(SDNode *N, SelectionDAG &DAG) {
  const TargetLowering &TLI = DAG.getTargetLoweringInfo();
  DebugLoc dl = N->getDebugLoc();
  SDValue Op = N->getOperand(0);

  // This function is only supposed to be called for i64 types, either as the
  // source or destination of the bit convert.
  EVT SrcVT = Op.getValueType();
  EVT DstVT = N->getValueType(0);
  assert((SrcVT == MVT::i64 || DstVT == MVT::i64) &&
         "ExpandBITCAST called for non-i64 type");

  // Turn i64->f64 into VMOVDRR.
  if (SrcVT == MVT::i64 && TLI.isTypeLegal(DstVT)) {
    SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
                             DAG.getConstant(0, MVT::i32));
    SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, Op,
                             DAG.getConstant(1, MVT::i32));
    return DAG.getNode(ISD::BITCAST, dl, DstVT,
                       DAG.getNode(ARMISD::VMOVDRR, dl, MVT::f64, Lo, Hi));
  }

  // Turn f64->i64 into VMOVRRD.
  if (DstVT == MVT::i64 && TLI.isTypeLegal(SrcVT)) {
    SDValue Cvt = DAG.getNode(ARMISD::VMOVRRD, dl,
                              DAG.getVTList(MVT::i32, MVT::i32), &Op, 1);
    // Merge the pieces into a single i64 value.
    return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Cvt, Cvt.getValue(1));
  }

  return SDValue();
}

/// getZeroVector - Returns a vector of specified type with all zero elements.
/// Zero vectors are used to represent vector negation and in those cases
/// will be implemented with the NEON VNEG instruction.  However, VNEG does
/// not support i64 elements, so sometimes the zero vectors will need to be
/// explicitly constructed.  Regardless, use a canonical VMOV to create the
/// zero vector.
static SDValue getZeroVector(EVT VT, SelectionDAG &DAG, DebugLoc dl) {
  assert(VT.isVector() && "Expected a vector type");
  // The canonical modified immediate encoding of a zero vector is....0!
  SDValue EncodedVal = DAG.getTargetConstant(0, MVT::i32);
  EVT VmovVT = VT.is128BitVector() ? MVT::v4i32 : MVT::v2i32;
  SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, EncodedVal);
  return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
}

/// LowerShiftRightParts - Lower SRA_PARTS, which returns two
/// i32 values and take a 2 x i32 value to shift plus a shift amount.
SDValue ARMTargetLowering::LowerShiftRightParts(SDValue Op,
                                                SelectionDAG &DAG) const {
  assert(Op.getNumOperands() == 3 && "Not a double-shift!");
  EVT VT = Op.getValueType();
  unsigned VTBits = VT.getSizeInBits();
  DebugLoc dl = Op.getDebugLoc();
  SDValue ShOpLo = Op.getOperand(0);
  SDValue ShOpHi = Op.getOperand(1);
  SDValue ShAmt  = Op.getOperand(2);
  SDValue ARMcc;
  unsigned Opc = (Op.getOpcode() == ISD::SRA_PARTS) ? ISD::SRA : ISD::SRL;

  assert(Op.getOpcode() == ISD::SRA_PARTS || Op.getOpcode() == ISD::SRL_PARTS);

  SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
                                 DAG.getConstant(VTBits, MVT::i32), ShAmt);
  SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
  SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
                                   DAG.getConstant(VTBits, MVT::i32));
  SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);
  SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
  SDValue TrueVal = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);

  SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
  SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
                          ARMcc, DAG, dl);
  SDValue Hi = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
  SDValue Lo = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, TrueVal, ARMcc,
                           CCR, Cmp);

  SDValue Ops[2] = { Lo, Hi };
  return DAG.getMergeValues(Ops, 2, dl);
}

/// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
/// i32 values and take a 2 x i32 value to shift plus a shift amount.
SDValue ARMTargetLowering::LowerShiftLeftParts(SDValue Op,
                                               SelectionDAG &DAG) const {
  assert(Op.getNumOperands() == 3 && "Not a double-shift!");
  EVT VT = Op.getValueType();
  unsigned VTBits = VT.getSizeInBits();
  DebugLoc dl = Op.getDebugLoc();
  SDValue ShOpLo = Op.getOperand(0);
  SDValue ShOpHi = Op.getOperand(1);
  SDValue ShAmt  = Op.getOperand(2);
  SDValue ARMcc;

  assert(Op.getOpcode() == ISD::SHL_PARTS);
  SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32,
                                 DAG.getConstant(VTBits, MVT::i32), ShAmt);
  SDValue Tmp1 = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);
  SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i32, ShAmt,
                                   DAG.getConstant(VTBits, MVT::i32));
  SDValue Tmp2 = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
  SDValue Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);

  SDValue FalseVal = DAG.getNode(ISD::OR, dl, VT, Tmp1, Tmp2);
  SDValue CCR = DAG.getRegister(ARM::CPSR, MVT::i32);
  SDValue Cmp = getARMCmp(ExtraShAmt, DAG.getConstant(0, MVT::i32), ISD::SETGE,
                          ARMcc, DAG, dl);
  SDValue Lo = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
  SDValue Hi = DAG.getNode(ARMISD::CMOV, dl, VT, FalseVal, Tmp3, ARMcc,
                           CCR, Cmp);

  SDValue Ops[2] = { Lo, Hi };
  return DAG.getMergeValues(Ops, 2, dl);
}

SDValue ARMTargetLowering::LowerFLT_ROUNDS_(SDValue Op,
                                            SelectionDAG &DAG) const {
  // The rounding mode is in bits 23:22 of the FPSCR.
  // The ARM rounding mode value to FLT_ROUNDS mapping is 0->1, 1->2, 2->3, 3->0
  // The formula we use to implement this is (((FPSCR + 1 << 22) >> 22) & 3)
  // so that the shift + and get folded into a bitfield extract.
  DebugLoc dl = Op.getDebugLoc();
  SDValue FPSCR = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, MVT::i32,
                              DAG.getConstant(Intrinsic::arm_get_fpscr,
                                              MVT::i32));
  SDValue FltRounds = DAG.getNode(ISD::ADD, dl, MVT::i32, FPSCR,
                                  DAG.getConstant(1U << 22, MVT::i32));
  SDValue RMODE = DAG.getNode(ISD::SRL, dl, MVT::i32, FltRounds,
                              DAG.getConstant(22, MVT::i32));
  return DAG.getNode(ISD::AND, dl, MVT::i32, RMODE,
                     DAG.getConstant(3, MVT::i32));
}

static SDValue LowerCTTZ(SDNode *N, SelectionDAG &DAG,
                         const ARMSubtarget *ST) {
  EVT VT = N->getValueType(0);
  DebugLoc dl = N->getDebugLoc();

  if (!ST->hasV6T2Ops())
    return SDValue();

  SDValue rbit = DAG.getNode(ARMISD::RBIT, dl, VT, N->getOperand(0));
  return DAG.getNode(ISD::CTLZ, dl, VT, rbit);
}

/// getCTPOP16BitCounts - Returns a v8i8/v16i8 vector containing the bit-count
/// for each 16-bit element from operand, repeated.  The basic idea is to
/// leverage vcnt to get the 8-bit counts, gather and add the results.
///
/// Trace for v4i16:
/// input    = [v0    v1    v2    v3   ] (vi 16-bit element)
/// cast: N0 = [w0 w1 w2 w3 w4 w5 w6 w7] (v0 = [w0 w1], wi 8-bit element)
/// vcnt: N1 = [b0 b1 b2 b3 b4 b5 b6 b7] (bi = bit-count of 8-bit element wi)
/// vrev: N2 = [b1 b0 b3 b2 b5 b4 b7 b6]
///            [b0 b1 b2 b3 b4 b5 b6 b7]
///           +[b1 b0 b3 b2 b5 b4 b7 b6]
/// N3=N1+N2 = [k0 k0 k1 k1 k2 k2 k3 k3] (k0 = b0+b1 = bit-count of 16-bit v0,
/// vuzp:    = [k0 k1 k2 k3 k0 k1 k2 k3]  each ki is 8-bits)
static SDValue getCTPOP16BitCounts(SDNode *N, SelectionDAG &DAG) {
  EVT VT = N->getValueType(0);
  DebugLoc DL = N->getDebugLoc();

  EVT VT8Bit = VT.is64BitVector() ? MVT::v8i8 : MVT::v16i8;
  SDValue N0 = DAG.getNode(ISD::BITCAST, DL, VT8Bit, N->getOperand(0));
  SDValue N1 = DAG.getNode(ISD::CTPOP, DL, VT8Bit, N0);
  SDValue N2 = DAG.getNode(ARMISD::VREV16, DL, VT8Bit, N1);
  SDValue N3 = DAG.getNode(ISD::ADD, DL, VT8Bit, N1, N2);
  return DAG.getNode(ARMISD::VUZP, DL, VT8Bit, N3, N3);
}

/// lowerCTPOP16BitElements - Returns a v4i16/v8i16 vector containing the
/// bit-count for each 16-bit element from the operand.  We need slightly
/// different sequencing for v4i16 and v8i16 to stay within NEON's available
/// 64/128-bit registers.
///
/// Trace for v4i16:
/// input           = [v0    v1    v2    v3    ] (vi 16-bit element)
/// v8i8: BitCounts = [k0 k1 k2 k3 k0 k1 k2 k3 ] (ki is the bit-count of vi)
/// v8i16:Extended  = [k0    k1    k2    k3    k0    k1    k2    k3    ]
/// v4i16:Extracted = [k0    k1    k2    k3    ]
static SDValue lowerCTPOP16BitElements(SDNode *N, SelectionDAG &DAG) {
  EVT VT = N->getValueType(0);
  DebugLoc DL = N->getDebugLoc();

  SDValue BitCounts = getCTPOP16BitCounts(N, DAG);
  if (VT.is64BitVector()) {
    SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, BitCounts);
    return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, Extended,
                       DAG.getIntPtrConstant(0));
  } else {
    SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v8i8,
                                    BitCounts, DAG.getIntPtrConstant(0));
    return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v8i16, Extracted);
  }
}

/// lowerCTPOP32BitElements - Returns a v2i32/v4i32 vector containing the
/// bit-count for each 32-bit element from the operand.  The idea here is
/// to split the vector into 16-bit elements, leverage the 16-bit count
/// routine, and then combine the results.
///
/// Trace for v2i32 (v4i32 similar with Extracted/Extended exchanged):
/// input    = [v0    v1    ] (vi: 32-bit elements)
/// Bitcast  = [w0 w1 w2 w3 ] (wi: 16-bit elements, v0 = [w0 w1])
/// Counts16 = [k0 k1 k2 k3 ] (ki: 16-bit elements, bit-count of wi)
/// vrev: N0 = [k1 k0 k3 k2 ]
///            [k0 k1 k2 k3 ]
///       N1 =+[k1 k0 k3 k2 ]
///            [k0 k2 k1 k3 ]
///       N2 =+[k1 k3 k0 k2 ]
///            [k0    k2    k1    k3    ]
/// Extended =+[k1    k3    k0    k2    ]
///            [k0    k2    ]
/// Extracted=+[k1    k3    ]
///
static SDValue lowerCTPOP32BitElements(SDNode *N, SelectionDAG &DAG) {
  EVT VT = N->getValueType(0);
  DebugLoc DL = N->getDebugLoc();

  EVT VT16Bit = VT.is64BitVector() ? MVT::v4i16 : MVT::v8i16;

  SDValue Bitcast = DAG.getNode(ISD::BITCAST, DL, VT16Bit, N->getOperand(0));
  SDValue Counts16 = lowerCTPOP16BitElements(Bitcast.getNode(), DAG);
  SDValue N0 = DAG.getNode(ARMISD::VREV32, DL, VT16Bit, Counts16);
  SDValue N1 = DAG.getNode(ISD::ADD, DL, VT16Bit, Counts16, N0);
  SDValue N2 = DAG.getNode(ARMISD::VUZP, DL, VT16Bit, N1, N1);

  if (VT.is64BitVector()) {
    SDValue Extended = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, N2);
    return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i32, Extended,
                       DAG.getIntPtrConstant(0));
  } else {
    SDValue Extracted = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i16, N2,
                                    DAG.getIntPtrConstant(0));
    return DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::v4i32, Extracted);
  }
}

static SDValue LowerCTPOP(SDNode *N, SelectionDAG &DAG,
                          const ARMSubtarget *ST) {
  EVT VT = N->getValueType(0);

  assert(ST->hasNEON() && "Custom ctpop lowering requires NEON.");
  assert((VT == MVT::v2i32 || VT == MVT::v4i32 ||
          VT == MVT::v4i16 || VT == MVT::v8i16) &&
         "Unexpected type for custom ctpop lowering");

  if (VT.getVectorElementType() == MVT::i32)
    return lowerCTPOP32BitElements(N, DAG);
  else
    return lowerCTPOP16BitElements(N, DAG);
}

static SDValue LowerShift(SDNode *N, SelectionDAG &DAG,
                          const ARMSubtarget *ST) {
  EVT VT = N->getValueType(0);
  DebugLoc dl = N->getDebugLoc();

  if (!VT.isVector())
    return SDValue();

  // Lower vector shifts on NEON to use VSHL.
  assert(ST->hasNEON() && "unexpected vector shift");

  // Left shifts translate directly to the vshiftu intrinsic.
  if (N->getOpcode() == ISD::SHL)
    return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
                       DAG.getConstant(Intrinsic::arm_neon_vshiftu, MVT::i32),
                       N->getOperand(0), N->getOperand(1));

  assert((N->getOpcode() == ISD::SRA ||
          N->getOpcode() == ISD::SRL) && "unexpected vector shift opcode");

  // NEON uses the same intrinsics for both left and right shifts.  For
  // right shifts, the shift amounts are negative, so negate the vector of
  // shift amounts.
  EVT ShiftVT = N->getOperand(1).getValueType();
  SDValue NegatedCount = DAG.getNode(ISD::SUB, dl, ShiftVT,
                                     getZeroVector(ShiftVT, DAG, dl),
                                     N->getOperand(1));
  Intrinsic::ID vshiftInt = (N->getOpcode() == ISD::SRA ?
                             Intrinsic::arm_neon_vshifts :
                             Intrinsic::arm_neon_vshiftu);
  return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
                     DAG.getConstant(vshiftInt, MVT::i32),
                     N->getOperand(0), NegatedCount);
}

static SDValue Expand64BitShift(SDNode *N, SelectionDAG &DAG,
                                const ARMSubtarget *ST) {
  EVT VT = N->getValueType(0);
  DebugLoc dl = N->getDebugLoc();

  // We can get here for a node like i32 = ISD::SHL i32, i64
  if (VT != MVT::i64)
    return SDValue();

  assert((N->getOpcode() == ISD::SRL || N->getOpcode() == ISD::SRA) &&
         "Unknown shift to lower!");

  // We only lower SRA, SRL of 1 here, all others use generic lowering.
  if (!isa<ConstantSDNode>(N->getOperand(1)) ||
      cast<ConstantSDNode>(N->getOperand(1))->getZExtValue() != 1)
    return SDValue();

  // If we are in thumb mode, we don't have RRX.
  if (ST->isThumb1Only()) return SDValue();

  // Okay, we have a 64-bit SRA or SRL of 1.  Lower this to an RRX expr.
  SDValue Lo = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
                           DAG.getConstant(0, MVT::i32));
  SDValue Hi = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, MVT::i32, N->getOperand(0),
                           DAG.getConstant(1, MVT::i32));

  // First, build a SRA_FLAG/SRL_FLAG op, which shifts the top part by one and
  // captures the result into a carry flag.
  unsigned Opc = N->getOpcode() == ISD::SRL ? ARMISD::SRL_FLAG:ARMISD::SRA_FLAG;
  Hi = DAG.getNode(Opc, dl, DAG.getVTList(MVT::i32, MVT::Glue), &Hi, 1);

  // The low part is an ARMISD::RRX operand, which shifts the carry in.
  Lo = DAG.getNode(ARMISD::RRX, dl, MVT::i32, Lo, Hi.getValue(1));

  // Merge the pieces into a single i64 value.
 return DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Lo, Hi);
}

static SDValue LowerVSETCC(SDValue Op, SelectionDAG &DAG) {
  SDValue TmpOp0, TmpOp1;
  bool Invert = false;
  bool Swap = false;
  unsigned Opc = 0;

  SDValue Op0 = Op.getOperand(0);
  SDValue Op1 = Op.getOperand(1);
  SDValue CC = Op.getOperand(2);
  EVT VT = Op.getValueType();
  ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
  DebugLoc dl = Op.getDebugLoc();

  if (Op.getOperand(1).getValueType().isFloatingPoint()) {
    switch (SetCCOpcode) {
    default: llvm_unreachable("Illegal FP comparison");
    case ISD::SETUNE:
    case ISD::SETNE:  Invert = true; // Fallthrough
    case ISD::SETOEQ:
    case ISD::SETEQ:  Opc = ARMISD::VCEQ; break;
    case ISD::SETOLT:
    case ISD::SETLT: Swap = true; // Fallthrough
    case ISD::SETOGT:
    case ISD::SETGT:  Opc = ARMISD::VCGT; break;
    case ISD::SETOLE:
    case ISD::SETLE:  Swap = true; // Fallthrough
    case ISD::SETOGE:
    case ISD::SETGE: Opc = ARMISD::VCGE; break;
    case ISD::SETUGE: Swap = true; // Fallthrough
    case ISD::SETULE: Invert = true; Opc = ARMISD::VCGT; break;
    case ISD::SETUGT: Swap = true; // Fallthrough
    case ISD::SETULT: Invert = true; Opc = ARMISD::VCGE; break;
    case ISD::SETUEQ: Invert = true; // Fallthrough
    case ISD::SETONE:
      // Expand this to (OLT | OGT).
      TmpOp0 = Op0;
      TmpOp1 = Op1;
      Opc = ISD::OR;
      Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
      Op1 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp0, TmpOp1);
      break;
    case ISD::SETUO: Invert = true; // Fallthrough
    case ISD::SETO:
      // Expand this to (OLT | OGE).
      TmpOp0 = Op0;
      TmpOp1 = Op1;
      Opc = ISD::OR;
      Op0 = DAG.getNode(ARMISD::VCGT, dl, VT, TmpOp1, TmpOp0);
      Op1 = DAG.getNode(ARMISD::VCGE, dl, VT, TmpOp0, TmpOp1);
      break;
    }
  } else {
    // Integer comparisons.
    switch (SetCCOpcode) {
    default: llvm_unreachable("Illegal integer comparison");
    case ISD::SETNE:  Invert = true;
    case ISD::SETEQ:  Opc = ARMISD::VCEQ; break;
    case ISD::SETLT:  Swap = true;
    case ISD::SETGT:  Opc = ARMISD::VCGT; break;
    case ISD::SETLE:  Swap = true;
    case ISD::SETGE:  Opc = ARMISD::VCGE; break;
    case ISD::SETULT: Swap = true;
    case ISD::SETUGT: Opc = ARMISD::VCGTU; break;
    case ISD::SETULE: Swap = true;
    case ISD::SETUGE: Opc = ARMISD::VCGEU; break;
    }

    // Detect VTST (Vector Test Bits) = icmp ne (and (op0, op1), zero).
    if (Opc == ARMISD::VCEQ) {

      SDValue AndOp;
      if (ISD::isBuildVectorAllZeros(Op1.getNode()))
        AndOp = Op0;
      else if (ISD::isBuildVectorAllZeros(Op0.getNode()))
        AndOp = Op1;

      // Ignore bitconvert.
      if (AndOp.getNode() && AndOp.getOpcode() == ISD::BITCAST)
        AndOp = AndOp.getOperand(0);

      if (AndOp.getNode() && AndOp.getOpcode() == ISD::AND) {
        Opc = ARMISD::VTST;
        Op0 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(0));
        Op1 = DAG.getNode(ISD::BITCAST, dl, VT, AndOp.getOperand(1));
        Invert = !Invert;
      }
    }
  }

  if (Swap)
    std::swap(Op0, Op1);

  // If one of the operands is a constant vector zero, attempt to fold the
  // comparison to a specialized compare-against-zero form.
  SDValue SingleOp;
  if (ISD::isBuildVectorAllZeros(Op1.getNode()))
    SingleOp = Op0;
  else if (ISD::isBuildVectorAllZeros(Op0.getNode())) {
    if (Opc == ARMISD::VCGE)
      Opc = ARMISD::VCLEZ;
    else if (Opc == ARMISD::VCGT)
      Opc = ARMISD::VCLTZ;
    SingleOp = Op1;
  }

  SDValue Result;
  if (SingleOp.getNode()) {
    switch (Opc) {
    case ARMISD::VCEQ:
      Result = DAG.getNode(ARMISD::VCEQZ, dl, VT, SingleOp); break;
    case ARMISD::VCGE:
      Result = DAG.getNode(ARMISD::VCGEZ, dl, VT, SingleOp); break;
    case ARMISD::VCLEZ:
      Result = DAG.getNode(ARMISD::VCLEZ, dl, VT, SingleOp); break;
    case ARMISD::VCGT:
      Result = DAG.getNode(ARMISD::VCGTZ, dl, VT, SingleOp); break;
    case ARMISD::VCLTZ:
      Result = DAG.getNode(ARMISD::VCLTZ, dl, VT, SingleOp); break;
    default:
      Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
    }
  } else {
     Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
  }

  if (Invert)
    Result = DAG.getNOT(dl, Result, VT);

  return Result;
}

/// isNEONModifiedImm - Check if the specified splat value corresponds to a
/// valid vector constant for a NEON instruction with a "modified immediate"
/// operand (e.g., VMOV).  If so, return the encoded value.
static SDValue isNEONModifiedImm(uint64_t SplatBits, uint64_t SplatUndef,
                                 unsigned SplatBitSize, SelectionDAG &DAG,
                                 EVT &VT, bool is128Bits, NEONModImmType type) {
  unsigned OpCmode, Imm;

  // SplatBitSize is set to the smallest size that splats the vector, so a
  // zero vector will always have SplatBitSize == 8.  However, NEON modified
  // immediate instructions others than VMOV do not support the 8-bit encoding
  // of a zero vector, and the default encoding of zero is supposed to be the
  // 32-bit version.
  if (SplatBits == 0)
    SplatBitSize = 32;

  switch (SplatBitSize) {
  case 8:
    if (type != VMOVModImm)
      return SDValue();
    // Any 1-byte value is OK.  Op=0, Cmode=1110.
    assert((SplatBits & ~0xff) == 0 && "one byte splat value is too big");
    OpCmode = 0xe;
    Imm = SplatBits;
    VT = is128Bits ? MVT::v16i8 : MVT::v8i8;
    break;

  case 16:
    // NEON's 16-bit VMOV supports splat values where only one byte is nonzero.
    VT = is128Bits ? MVT::v8i16 : MVT::v4i16;
    if ((SplatBits & ~0xff) == 0) {
      // Value = 0x00nn: Op=x, Cmode=100x.
      OpCmode = 0x8;
      Imm = SplatBits;
      break;
    }
    if ((SplatBits & ~0xff00) == 0) {
      // Value = 0xnn00: Op=x, Cmode=101x.
      OpCmode = 0xa;
      Imm = SplatBits >> 8;
      break;
    }
    return SDValue();

  case 32:
    // NEON's 32-bit VMOV supports splat values where:
    // * only one byte is nonzero, or
    // * the least significant byte is 0xff and the second byte is nonzero, or
    // * the least significant 2 bytes are 0xff and the third is nonzero.
    VT = is128Bits ? MVT::v4i32 : MVT::v2i32;
    if ((SplatBits & ~0xff) == 0) {
      // Value = 0x000000nn: Op=x, Cmode=000x.
      OpCmode = 0;
      Imm = SplatBits;
      break;
    }
    if ((SplatBits & ~0xff00) == 0) {
      // Value = 0x0000nn00: Op=x, Cmode=001x.
      OpCmode = 0x2;
      Imm = SplatBits >> 8;
      break;
    }
    if ((SplatBits & ~0xff0000) == 0) {
      // Value = 0x00nn0000: Op=x, Cmode=010x.
      OpCmode = 0x4;
      Imm = SplatBits >> 16;
      break;
    }
    if ((SplatBits & ~0xff000000) == 0) {
      // Value = 0xnn000000: Op=x, Cmode=011x.
      OpCmode = 0x6;
      Imm = SplatBits >> 24;
      break;
    }

    // cmode == 0b1100 and cmode == 0b1101 are not supported for VORR or VBIC
    if (type == OtherModImm) return SDValue();

    if ((SplatBits & ~0xffff) == 0 &&
        ((SplatBits | SplatUndef) & 0xff) == 0xff) {
      // Value = 0x0000nnff: Op=x, Cmode=1100.
      OpCmode = 0xc;
      Imm = SplatBits >> 8;
      SplatBits |= 0xff;
      break;
    }

    if ((SplatBits & ~0xffffff) == 0 &&
        ((SplatBits | SplatUndef) & 0xffff) == 0xffff) {
      // Value = 0x00nnffff: Op=x, Cmode=1101.
      OpCmode = 0xd;
      Imm = SplatBits >> 16;
      SplatBits |= 0xffff;
      break;
    }

    // Note: there are a few 32-bit splat values (specifically: 00ffff00,
    // ff000000, ff0000ff, and ffff00ff) that are valid for VMOV.I64 but not
    // VMOV.I32.  A (very) minor optimization would be to replicate the value
    // and fall through here to test for a valid 64-bit splat.  But, then the
    // caller would also need to check and handle the change in size.
    return SDValue();

  case 64: {
    if (type != VMOVModImm)
      return SDValue();
    // NEON has a 64-bit VMOV splat where each byte is either 0 or 0xff.
    uint64_t BitMask = 0xff;
    uint64_t Val = 0;
    unsigned ImmMask = 1;
    Imm = 0;
    for (int ByteNum = 0; ByteNum < 8; ++ByteNum) {
      if (((SplatBits | SplatUndef) & BitMask) == BitMask) {
        Val |= BitMask;
        Imm |= ImmMask;
      } else if ((SplatBits & BitMask) != 0) {
        return SDValue();
      }
      BitMask <<= 8;
      ImmMask <<= 1;
    }
    // Op=1, Cmode=1110.
    OpCmode = 0x1e;
    SplatBits = Val;
    VT = is128Bits ? MVT::v2i64 : MVT::v1i64;
    break;
  }

  default:
    llvm_unreachable("unexpected size for isNEONModifiedImm");
  }

  unsigned EncodedVal = ARM_AM::createNEONModImm(OpCmode, Imm);
  return DAG.getTargetConstant(EncodedVal, MVT::i32);
}

SDValue ARMTargetLowering::LowerConstantFP(SDValue Op, SelectionDAG &DAG,
                                           const ARMSubtarget *ST) const {
  if (!ST->useNEONForSinglePrecisionFP() || !ST->hasVFP3() || ST->hasD16())
    return SDValue();

  ConstantFPSDNode *CFP = cast<ConstantFPSDNode>(Op);
  assert(Op.getValueType() == MVT::f32 &&
         "ConstantFP custom lowering should only occur for f32.");

  // Try splatting with a VMOV.f32...
  APFloat FPVal = CFP->getValueAPF();
  int ImmVal = ARM_AM::getFP32Imm(FPVal);
  if (ImmVal != -1) {
    DebugLoc DL = Op.getDebugLoc();
    SDValue NewVal = DAG.getTargetConstant(ImmVal, MVT::i32);
    SDValue VecConstant = DAG.getNode(ARMISD::VMOVFPIMM, DL, MVT::v2f32,
                                      NewVal);
    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecConstant,
                       DAG.getConstant(0, MVT::i32));
  }

  // If that fails, try a VMOV.i32
  EVT VMovVT;
  unsigned iVal = FPVal.bitcastToAPInt().getZExtValue();
  SDValue NewVal = isNEONModifiedImm(iVal, 0, 32, DAG, VMovVT, false,
                                     VMOVModImm);
  if (NewVal != SDValue()) {
    DebugLoc DL = Op.getDebugLoc();
    SDValue VecConstant = DAG.getNode(ARMISD::VMOVIMM, DL, VMovVT,
                                      NewVal);
    SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
                                       VecConstant);
    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
                       DAG.getConstant(0, MVT::i32));
  }

  // Finally, try a VMVN.i32
  NewVal = isNEONModifiedImm(~iVal & 0xffffffff, 0, 32, DAG, VMovVT, false,
                             VMVNModImm);
  if (NewVal != SDValue()) {
    DebugLoc DL = Op.getDebugLoc();
    SDValue VecConstant = DAG.getNode(ARMISD::VMVNIMM, DL, VMovVT, NewVal);
    SDValue VecFConstant = DAG.getNode(ISD::BITCAST, DL, MVT::v2f32,
                                       VecConstant);
    return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32, VecFConstant,
                       DAG.getConstant(0, MVT::i32));
  }

  return SDValue();
}

// check if an VEXT instruction can handle the shuffle mask when the
// vector sources of the shuffle are the same.
static bool isSingletonVEXTMask(ArrayRef<int> M, EVT VT, unsigned &Imm) {
  unsigned NumElts = VT.getVectorNumElements();

  // Assume that the first shuffle index is not UNDEF.  Fail if it is.
  if (M[0] < 0)
    return false;

  Imm = M[0];

  // If this is a VEXT shuffle, the immediate value is the index of the first
  // element.  The other shuffle indices must be the successive elements after
  // the first one.
  unsigned ExpectedElt = Imm;
  for (unsigned i = 1; i < NumElts; ++i) {
    // Increment the expected index.  If it wraps around, just follow it
    // back to index zero and keep going.
    ++ExpectedElt;
    if (ExpectedElt == NumElts)
      ExpectedElt = 0;

    if (M[i] < 0) continue; // ignore UNDEF indices
    if (ExpectedElt != static_cast<unsigned>(M[i]))
      return false;
  }

  return true;
}


static bool isVEXTMask(ArrayRef<int> M, EVT VT,
                       bool &ReverseVEXT, unsigned &Imm) {
  unsigned NumElts = VT.getVectorNumElements();
  ReverseVEXT = false;

  // Assume that the first shuffle index is not UNDEF.  Fail if it is.
  if (M[0] < 0)
    return false;

  Imm = M[0];

  // If this is a VEXT shuffle, the immediate value is the index of the first
  // element.  The other shuffle indices must be the successive elements after
  // the first one.
  unsigned ExpectedElt = Imm;
  for (unsigned i = 1; i < NumElts; ++i) {
    // Increment the expected index.  If it wraps around, it may still be
    // a VEXT but the source vectors must be swapped.
    ExpectedElt += 1;
    if (ExpectedElt == NumElts * 2) {
      ExpectedElt = 0;
      ReverseVEXT = true;
    }

    if (M[i] < 0) continue; // ignore UNDEF indices
    if (ExpectedElt != static_cast<unsigned>(M[i]))
      return false;
  }

  // Adjust the index value if the source operands will be swapped.
  if (ReverseVEXT)
    Imm -= NumElts;

  return true;
}

/// isVREVMask - Check if a vector shuffle corresponds to a VREV
/// instruction with the specified blocksize.  (The order of the elements
/// within each block of the vector is reversed.)
static bool isVREVMask(ArrayRef<int> M, EVT VT, unsigned BlockSize) {
  assert((BlockSize==16 || BlockSize==32 || BlockSize==64) &&
         "Only possible block sizes for VREV are: 16, 32, 64");

  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
  if (EltSz == 64)
    return false;

  unsigned NumElts = VT.getVectorNumElements();
  unsigned BlockElts = M[0] + 1;
  // If the first shuffle index is UNDEF, be optimistic.
  if (M[0] < 0)
    BlockElts = BlockSize / EltSz;

  if (BlockSize <= EltSz || BlockSize != BlockElts * EltSz)
    return false;

  for (unsigned i = 0; i < NumElts; ++i) {
    if (M[i] < 0) continue; // ignore UNDEF indices
    if ((unsigned) M[i] != (i - i%BlockElts) + (BlockElts - 1 - i%BlockElts))
      return false;
  }

  return true;
}

static bool isVTBLMask(ArrayRef<int> M, EVT VT) {
  // We can handle <8 x i8> vector shuffles. If the index in the mask is out of
  // range, then 0 is placed into the resulting vector. So pretty much any mask
  // of 8 elements can work here.
  return VT == MVT::v8i8 && M.size() == 8;
}

static bool isVTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
  if (EltSz == 64)
    return false;

  unsigned NumElts = VT.getVectorNumElements();
  WhichResult = (M[0] == 0 ? 0 : 1);
  for (unsigned i = 0; i < NumElts; i += 2) {
    if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
        (M[i+1] >= 0 && (unsigned) M[i+1] != i + NumElts + WhichResult))
      return false;
  }
  return true;
}

/// isVTRN_v_undef_Mask - Special case of isVTRNMask for canonical form of
/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
/// Mask is e.g., <0, 0, 2, 2> instead of <0, 4, 2, 6>.
static bool isVTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
  if (EltSz == 64)
    return false;

  unsigned NumElts = VT.getVectorNumElements();
  WhichResult = (M[0] == 0 ? 0 : 1);
  for (unsigned i = 0; i < NumElts; i += 2) {
    if ((M[i] >= 0 && (unsigned) M[i] != i + WhichResult) ||
        (M[i+1] >= 0 && (unsigned) M[i+1] != i + WhichResult))
      return false;
  }
  return true;
}

static bool isVUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
  if (EltSz == 64)
    return false;

  unsigned NumElts = VT.getVectorNumElements();
  WhichResult = (M[0] == 0 ? 0 : 1);
  for (unsigned i = 0; i != NumElts; ++i) {
    if (M[i] < 0) continue; // ignore UNDEF indices
    if ((unsigned) M[i] != 2 * i + WhichResult)
      return false;
  }

  // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
  if (VT.is64BitVector() && EltSz == 32)
    return false;

  return true;
}

/// isVUZP_v_undef_Mask - Special case of isVUZPMask for canonical form of
/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
/// Mask is e.g., <0, 2, 0, 2> instead of <0, 2, 4, 6>,
static bool isVUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
  if (EltSz == 64)
    return false;

  unsigned Half = VT.getVectorNumElements() / 2;
  WhichResult = (M[0] == 0 ? 0 : 1);
  for (unsigned j = 0; j != 2; ++j) {
    unsigned Idx = WhichResult;
    for (unsigned i = 0; i != Half; ++i) {
      int MIdx = M[i + j * Half];
      if (MIdx >= 0 && (unsigned) MIdx != Idx)
        return false;
      Idx += 2;
    }
  }

  // VUZP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
  if (VT.is64BitVector() && EltSz == 32)
    return false;

  return true;
}

static bool isVZIPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
  if (EltSz == 64)
    return false;

  unsigned NumElts = VT.getVectorNumElements();
  WhichResult = (M[0] == 0 ? 0 : 1);
  unsigned Idx = WhichResult * NumElts / 2;
  for (unsigned i = 0; i != NumElts; i += 2) {
    if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
        (M[i+1] >= 0 && (unsigned) M[i+1] != Idx + NumElts))
      return false;
    Idx += 1;
  }

  // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
  if (VT.is64BitVector() && EltSz == 32)
    return false;

  return true;
}

/// isVZIP_v_undef_Mask - Special case of isVZIPMask for canonical form of
/// "vector_shuffle v, v", i.e., "vector_shuffle v, undef".
/// Mask is e.g., <0, 0, 1, 1> instead of <0, 4, 1, 5>.
static bool isVZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult){
  unsigned EltSz = VT.getVectorElementType().getSizeInBits();
  if (EltSz == 64)
    return false;

  unsigned NumElts = VT.getVectorNumElements();
  WhichResult = (M[0] == 0 ? 0 : 1);
  unsigned Idx = WhichResult * NumElts / 2;
  for (unsigned i = 0; i != NumElts; i += 2) {
    if ((M[i] >= 0 && (unsigned) M[i] != Idx) ||
        (M[i+1] >= 0 && (unsigned) M[i+1] != Idx))
      return false;
    Idx += 1;
  }

  // VZIP.32 for 64-bit vectors is a pseudo-instruction alias for VTRN.32.
  if (VT.is64BitVector() && EltSz == 32)
    return false;

  return true;
}

/// \return true if this is a reverse operation on an vector.
static bool isReverseMask(ArrayRef<int> M, EVT VT) {
  unsigned NumElts = VT.getVectorNumElements();
  // Make sure the mask has the right size.
  if (NumElts != M.size())
      return false;

  // Look for <15, ..., 3, -1, 1, 0>.
  for (unsigned i = 0; i != NumElts; ++i)
    if (M[i] >= 0 && M[i] != (int) (NumElts - 1 - i))
      return false;

  return true;
}

// If N is an integer constant that can be moved into a register in one
// instruction, return an SDValue of such a constant (will become a MOV
// instruction).  Otherwise return null.
static SDValue IsSingleInstrConstant(SDValue N, SelectionDAG &DAG,
                                     const ARMSubtarget *ST, DebugLoc dl) {
  uint64_t Val;
  if (!isa<ConstantSDNode>(N))
    return SDValue();
  Val = cast<ConstantSDNode>(N)->getZExtValue();

  if (ST->isThumb1Only()) {
    if (Val <= 255 || ~Val <= 255)
      return DAG.getConstant(Val, MVT::i32);
  } else {
    if (ARM_AM::getSOImmVal(Val) != -1 || ARM_AM::getSOImmVal(~Val) != -1)
      return DAG.getConstant(Val, MVT::i32);
  }
  return SDValue();
}

// If this is a case we can't handle, return null and let the default
// expansion code take care of it.
SDValue ARMTargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG,
                                             const ARMSubtarget *ST) const {
  BuildVectorSDNode *BVN = cast<BuildVectorSDNode>(Op.getNode());
  DebugLoc dl = Op.getDebugLoc();
  EVT VT = Op.getValueType();

  APInt SplatBits, SplatUndef;
  unsigned SplatBitSize;
  bool HasAnyUndefs;
  if (BVN->isConstantSplat(SplatBits, SplatUndef, SplatBitSize, HasAnyUndefs)) {
    if (SplatBitSize <= 64) {
      // Check if an immediate VMOV works.
      EVT VmovVT;
      SDValue Val = isNEONModifiedImm(SplatBits.getZExtValue(),
                                      SplatUndef.getZExtValue(), SplatBitSize,
                                      DAG, VmovVT, VT.is128BitVector(),
                                      VMOVModImm);
      if (Val.getNode()) {
        SDValue Vmov = DAG.getNode(ARMISD::VMOVIMM, dl, VmovVT, Val);
        return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
      }

      // Try an immediate VMVN.
      uint64_t NegatedImm = (~SplatBits).getZExtValue();
      Val = isNEONModifiedImm(NegatedImm,
                                      SplatUndef.getZExtValue(), SplatBitSize,
                                      DAG, VmovVT, VT.is128BitVector(),
                                      VMVNModImm);
      if (Val.getNode()) {
        SDValue Vmov = DAG.getNode(ARMISD::VMVNIMM, dl, VmovVT, Val);
        return DAG.getNode(ISD::BITCAST, dl, VT, Vmov);
      }

      // Use vmov.f32 to materialize other v2f32 and v4f32 splats.
      if ((VT == MVT::v2f32 || VT == MVT::v4f32) && SplatBitSize == 32) {
        int ImmVal = ARM_AM::getFP32Imm(SplatBits);
        if (ImmVal != -1) {
          SDValue Val = DAG.getTargetConstant(ImmVal, MVT::i32);
          return DAG.getNode(ARMISD::VMOVFPIMM, dl, VT, Val);
        }
      }
    }
  }

  // Scan through the operands to see if only one value is used.
  //
  // As an optimisation, even if more than one value is used it may be more
  // profitable to splat with one value then change some lanes.
  //
  // Heuristically we decide to do this if the vector has a "dominant" value,
  // defined as splatted to more than half of the lanes.
  unsigned NumElts = VT.getVectorNumElements();
  bool isOnlyLowElement = true;
  bool usesOnlyOneValue = true;
  bool hasDominantValue = false;
  bool isConstant = true;

  // Map of the number of times a particular SDValue appears in the
  // element list.
  DenseMap<SDValue, unsigned> ValueCounts;
  SDValue Value;
  for (unsigned i = 0; i < NumElts; ++i) {
    SDValue V = Op.getOperand(i);
    if (V.getOpcode() == ISD::UNDEF)
      continue;
    if (i > 0)
      isOnlyLowElement = false;
    if (!isa<ConstantFPSDNode>(V) && !isa<ConstantSDNode>(V))
      isConstant = false;

    ValueCounts.insert(std::make_pair(V, 0));
    unsigned &Count = ValueCounts[V];

    // Is this value dominant? (takes up more than half of the lanes)
    if (++Count > (NumElts / 2)) {
      hasDominantValue = true;
      Value = V;
    }
  }
  if (ValueCounts.size() != 1)
    usesOnlyOneValue = false;
  if (!Value.getNode() && ValueCounts.size() > 0)
    Value = ValueCounts.begin()->first;

  if (ValueCounts.size() == 0)
    return DAG.getUNDEF(VT);

  if (isOnlyLowElement)
    return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Value);

  unsigned EltSize = VT.getVectorElementType().getSizeInBits();

  // Use VDUP for non-constant splats.  For f32 constant splats, reduce to
  // i32 and try again.
  if (hasDominantValue && EltSize <= 32) {
    if (!isConstant) {
      SDValue N;

      // If we are VDUPing a value that comes directly from a vector, that will
      // cause an unnecessary move to and from a GPR, where instead we could
      // just use VDUPLANE. We can only do this if the lane being extracted
      // is at a constant index, as the VDUP from lane instructions only have
      // constant-index forms.
      if (Value->getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
          isa<ConstantSDNode>(Value->getOperand(1))) {
        // We need to create a new undef vector to use for the VDUPLANE if the
        // size of the vector from which we get the value is different than the
        // size of the vector that we need to create. We will insert the element
        // such that the register coalescer will remove unnecessary copies.
        if (VT != Value->getOperand(0).getValueType()) {
          ConstantSDNode *constIndex;
          constIndex = dyn_cast<ConstantSDNode>(Value->getOperand(1));
          assert(constIndex && "The index is not a constant!");
          unsigned index = constIndex->getAPIntValue().getLimitedValue() %
                             VT.getVectorNumElements();
          N =  DAG.getNode(ARMISD::VDUPLANE, dl, VT,
                 DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DAG.getUNDEF(VT),
                        Value, DAG.getConstant(index, MVT::i32)),
                           DAG.getConstant(index, MVT::i32));
        } else
          N = DAG.getNode(ARMISD::VDUPLANE, dl, VT,
                        Value->getOperand(0), Value->getOperand(1));
      } else
        N = DAG.getNode(ARMISD::VDUP, dl, VT, Value);

      if (!usesOnlyOneValue) {
        // The dominant value was splatted as 'N', but we now have to insert
        // all differing elements.
        for (unsigned I = 0; I < NumElts; ++I) {
          if (Op.getOperand(I) == Value)
            continue;
          SmallVector<SDValue, 3> Ops;
          Ops.push_back(N);
          Ops.push_back(Op.getOperand(I));
          Ops.push_back(DAG.getConstant(I, MVT::i32));
          N = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, &Ops[0], 3);
        }
      }
      return N;
    }
    if (VT.getVectorElementType().isFloatingPoint()) {
      SmallVector<SDValue, 8> Ops;
      for (unsigned i = 0; i < NumElts; ++i)
        Ops.push_back(DAG.getNode(ISD::BITCAST, dl, MVT::i32,
                                  Op.getOperand(i)));
      EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32, NumElts);
      SDValue Val = DAG.getNode(ISD::BUILD_VECTOR, dl, VecVT, &Ops[0], NumElts);
      Val = LowerBUILD_VECTOR(Val, DAG, ST);
      if (Val.getNode())
        return DAG.getNode(ISD::BITCAST, dl, VT, Val);
    }
    if (usesOnlyOneValue) {
      SDValue Val = IsSingleInstrConstant(Value, DAG, ST, dl);
      if (isConstant && Val.getNode())
        return DAG.getNode(ARMISD::VDUP, dl, VT, Val);
    }
  }

  // If all elements are constants and the case above didn't get hit, fall back
  // to the default expansion, which will generate a load from the constant
  // pool.
  if (isConstant)
    return SDValue();

  // Empirical tests suggest this is rarely worth it for vectors of length <= 2.
  if (NumElts >= 4) {
    SDValue shuffle = ReconstructShuffle(Op, DAG);
    if (shuffle != SDValue())
      return shuffle;
  }

  // Vectors with 32- or 64-bit elements can be built by directly assigning
  // the subregisters.  Lower it to an ARMISD::BUILD_VECTOR so the operands
  // will be legalized.
  if (EltSize >= 32) {
    // Do the expansion with floating-point types, since that is what the VFP
    // registers are defined to use, and since i64 is not legal.
    EVT EltVT = EVT::getFloatingPointVT(EltSize);
    EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
    SmallVector<SDValue, 8> Ops;
    for (unsigned i = 0; i < NumElts; ++i)
      Ops.push_back(DAG.getNode(ISD::BITCAST, dl, EltVT, Op.getOperand(i)));
    SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
    return DAG.getNode(ISD::BITCAST, dl, VT, Val);
  }

  return SDValue();
}

// Gather data to see if the operation can be modelled as a
// shuffle in combination with VEXTs.
SDValue ARMTargetLowering::ReconstructShuffle(SDValue Op,
                                              SelectionDAG &DAG) const {
  DebugLoc dl = Op.getDebugLoc();
  EVT VT = Op.getValueType();
  unsigned NumElts = VT.getVectorNumElements();

  SmallVector<SDValue, 2> SourceVecs;
  SmallVector<unsigned, 2> MinElts;
  SmallVector<unsigned, 2> MaxElts;

  for (unsigned i = 0; i < NumElts; ++i) {
    SDValue V = Op.getOperand(i);
    if (V.getOpcode() == ISD::UNDEF)
      continue;
    else if (V.getOpcode() != ISD::EXTRACT_VECTOR_ELT) {
      // A shuffle can only come from building a vector from various
      // elements of other vectors.
      return SDValue();
    } else if (V.getOperand(0).getValueType().getVectorElementType() !=
               VT.getVectorElementType()) {
      // This code doesn't know how to handle shuffles where the vector
      // element types do not match (this happens because type legalization
      // promotes the return type of EXTRACT_VECTOR_ELT).
      // FIXME: It might be appropriate to extend this code to handle
      // mismatched types.
      return SDValue();
    }

    // Record this extraction against the appropriate vector if possible...
    SDValue SourceVec = V.getOperand(0);
    // If the element number isn't a constant, we can't effectively
    // analyze what's going on.
    if (!isa<ConstantSDNode>(V.getOperand(1)))
      return SDValue();
    unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
    bool FoundSource = false;
    for (unsigned j = 0; j < SourceVecs.size(); ++j) {
      if (SourceVecs[j] == SourceVec) {
        if (MinElts[j] > EltNo)
          MinElts[j] = EltNo;
        if (MaxElts[j] < EltNo)
          MaxElts[j] = EltNo;
        FoundSource = true;
        break;
      }
    }

    // Or record a new source if not...
    if (!FoundSource) {
      SourceVecs.push_back(SourceVec);
      MinElts.push_back(EltNo);
      MaxElts.push_back(EltNo);
    }
  }

  // Currently only do something sane when at most two source vectors
  // involved.
  if (SourceVecs.size() > 2)
    return SDValue();

  SDValue ShuffleSrcs[2] = {DAG.getUNDEF(VT), DAG.getUNDEF(VT) };
  int VEXTOffsets[2] = {0, 0};

  // This loop extracts the usage patterns of the source vectors
  // and prepares appropriate SDValues for a shuffle if possible.
  for (unsigned i = 0; i < SourceVecs.size(); ++i) {
    if (SourceVecs[i].getValueType() == VT) {
      // No VEXT necessary
      ShuffleSrcs[i] = SourceVecs[i];
      VEXTOffsets[i] = 0;
      continue;
    } else if (SourceVecs[i].getValueType().getVectorNumElements() < NumElts) {
      // It probably isn't worth padding out a smaller vector just to
      // break it down again in a shuffle.
      return SDValue();
    }

    // Since only 64-bit and 128-bit vectors are legal on ARM and
    // we've eliminated the other cases...
    assert(SourceVecs[i].getValueType().getVectorNumElements() == 2*NumElts &&
           "unexpected vector sizes in ReconstructShuffle");

    if (MaxElts[i] - MinElts[i] >= NumElts) {
      // Span too large for a VEXT to cope
      return SDValue();
    }

    if (MinElts[i] >= NumElts) {
      // The extraction can just take the second half
      VEXTOffsets[i] = NumElts;
      ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
                                   SourceVecs[i],
                                   DAG.getIntPtrConstant(NumElts));
    } else if (MaxElts[i] < NumElts) {
      // The extraction can just take the first half
      VEXTOffsets[i] = 0;
      ShuffleSrcs[i] = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
                                   SourceVecs[i],
                                   DAG.getIntPtrConstant(0));
    } else {
      // An actual VEXT is needed
      VEXTOffsets[i] = MinElts[i];
      SDValue VEXTSrc1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
                                     SourceVecs[i],
                                     DAG.getIntPtrConstant(0));
      SDValue VEXTSrc2 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT,
                                     SourceVecs[i],
                                     DAG.getIntPtrConstant(NumElts));
      ShuffleSrcs[i] = DAG.getNode(ARMISD::VEXT, dl, VT, VEXTSrc1, VEXTSrc2,
                                   DAG.getConstant(VEXTOffsets[i], MVT::i32));
    }
  }

  SmallVector<int, 8> Mask;

  for (unsigned i = 0; i < NumElts; ++i) {
    SDValue Entry = Op.getOperand(i);
    if (Entry.getOpcode() == ISD::UNDEF) {
      Mask.push_back(-1);
      continue;
    }

    SDValue ExtractVec = Entry.getOperand(0);
    int ExtractElt = cast<ConstantSDNode>(Op.getOperand(i)
                                          .getOperand(1))->getSExtValue();
    if (ExtractVec == SourceVecs[0]) {
      Mask.push_back(ExtractElt - VEXTOffsets[0]);
    } else {
      Mask.push_back(ExtractElt + NumElts - VEXTOffsets[1]);
    }
  }

  // Final check before we try to produce nonsense...
  if (isShuffleMaskLegal(Mask, VT))
    return DAG.getVectorShuffle(VT, dl, ShuffleSrcs[0], ShuffleSrcs[1],
                                &Mask[0]);

  return SDValue();
}

/// isShuffleMaskLegal - Targets can use this to indicate that they only
/// support *some* VECTOR_SHUFFLE operations, those with specific masks.
/// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
/// are assumed to be legal.
bool
ARMTargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
                                      EVT VT) const {
  if (VT.getVectorNumElements() == 4 &&
      (VT.is128BitVector() || VT.is64BitVector())) {
    unsigned PFIndexes[4];
    for (unsigned i = 0; i != 4; ++i) {
      if (M[i] < 0)
        PFIndexes[i] = 8;
      else
        PFIndexes[i] = M[i];
    }

    // Compute the index in the perfect shuffle table.
    unsigned PFTableIndex =
      PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
    unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
    unsigned Cost = (PFEntry >> 30);

    if (Cost <= 4)
      return true;
  }

  bool ReverseVEXT;
  unsigned Imm, WhichResult;

  unsigned EltSize = VT.getVectorElementType().getSizeInBits();
  return (EltSize >= 32 ||
          ShuffleVectorSDNode::isSplatMask(&M[0], VT) ||
          isVREVMask(M, VT, 64) ||
          isVREVMask(M, VT, 32) ||
          isVREVMask(M, VT, 16) ||
          isVEXTMask(M, VT, ReverseVEXT, Imm) ||
          isVTBLMask(M, VT) ||
          isVTRNMask(M, VT, WhichResult) ||
          isVUZPMask(M, VT, WhichResult) ||
          isVZIPMask(M, VT, WhichResult) ||
          isVTRN_v_undef_Mask(M, VT, WhichResult) ||
          isVUZP_v_undef_Mask(M, VT, WhichResult) ||
          isVZIP_v_undef_Mask(M, VT, WhichResult) ||
          ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(M, VT)));
}

/// GeneratePerfectShuffle - Given an entry in the perfect-shuffle table, emit
/// the specified operations to build the shuffle.
static SDValue GeneratePerfectShuffle(unsigned PFEntry, SDValue LHS,
                                      SDValue RHS, SelectionDAG &DAG,
                                      DebugLoc dl) {
  unsigned OpNum = (PFEntry >> 26) & 0x0F;
  unsigned LHSID = (PFEntry >> 13) & ((1 << 13)-1);
  unsigned RHSID = (PFEntry >>  0) & ((1 << 13)-1);

  enum {
    OP_COPY = 0, // Copy, used for things like <u,u,u,3> to say it is <0,1,2,3>
    OP_VREV,
    OP_VDUP0,
    OP_VDUP1,
    OP_VDUP2,
    OP_VDUP3,
    OP_VEXT1,
    OP_VEXT2,
    OP_VEXT3,
    OP_VUZPL, // VUZP, left result
    OP_VUZPR, // VUZP, right result
    OP_VZIPL, // VZIP, left result
    OP_VZIPR, // VZIP, right result
    OP_VTRNL, // VTRN, left result
    OP_VTRNR  // VTRN, right result
  };

  if (OpNum == OP_COPY) {
    if (LHSID == (1*9+2)*9+3) return LHS;
    assert(LHSID == ((4*9+5)*9+6)*9+7 && "Illegal OP_COPY!");
    return RHS;
  }

  SDValue OpLHS, OpRHS;
  OpLHS = GeneratePerfectShuffle(PerfectShuffleTable[LHSID], LHS, RHS, DAG, dl);
  OpRHS = GeneratePerfectShuffle(PerfectShuffleTable[RHSID], LHS, RHS, DAG, dl);
  EVT VT = OpLHS.getValueType();

  switch (OpNum) {
  default: llvm_unreachable("Unknown shuffle opcode!");
  case OP_VREV:
    // VREV divides the vector in half and swaps within the half.
    if (VT.getVectorElementType() == MVT::i32 ||
        VT.getVectorElementType() == MVT::f32)
      return DAG.getNode(ARMISD::VREV64, dl, VT, OpLHS);
    // vrev <4 x i16> -> VREV32
    if (VT.getVectorElementType() == MVT::i16)
      return DAG.getNode(ARMISD::VREV32, dl, VT, OpLHS);
    // vrev <4 x i8> -> VREV16
    assert(VT.getVectorElementType() == MVT::i8);
    return DAG.getNode(ARMISD::VREV16, dl, VT, OpLHS);
  case OP_VDUP0:
  case OP_VDUP1:
  case OP_VDUP2:
  case OP_VDUP3:
    return DAG.getNode(ARMISD::VDUPLANE, dl, VT,
                       OpLHS, DAG.getConstant(OpNum-OP_VDUP0, MVT::i32));
  case OP_VEXT1:
  case OP_VEXT2:
  case OP_VEXT3:
    return DAG.getNode(ARMISD::VEXT, dl, VT,
                       OpLHS, OpRHS,
                       DAG.getConstant(OpNum-OP_VEXT1+1, MVT::i32));
  case OP_VUZPL:
  case OP_VUZPR:
    return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
                       OpLHS, OpRHS).getValue(OpNum-OP_VUZPL);
  case OP_VZIPL:
  case OP_VZIPR:
    return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
                       OpLHS, OpRHS).getValue(OpNum-OP_VZIPL);
  case OP_VTRNL:
  case OP_VTRNR:
    return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
                       OpLHS, OpRHS).getValue(OpNum-OP_VTRNL);
  }
}

static SDValue LowerVECTOR_SHUFFLEv8i8(SDValue Op,
                                       ArrayRef<int> ShuffleMask,
                                       SelectionDAG &DAG) {
  // Check to see if we can use the VTBL instruction.
  SDValue V1 = Op.getOperand(0);
  SDValue V2 = Op.getOperand(1);
  DebugLoc DL = Op.getDebugLoc();

  SmallVector<SDValue, 8> VTBLMask;
  for (ArrayRef<int>::iterator
         I = ShuffleMask.begin(), E = ShuffleMask.end(); I != E; ++I)
    VTBLMask.push_back(DAG.getConstant(*I, MVT::i32));

  if (V2.getNode()->getOpcode() == ISD::UNDEF)
    return DAG.getNode(ARMISD::VTBL1, DL, MVT::v8i8, V1,
                       DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
                                   &VTBLMask[0], 8));

  return DAG.getNode(ARMISD::VTBL2, DL, MVT::v8i8, V1, V2,
                     DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i8,
                                 &VTBLMask[0], 8));
}

static SDValue LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(SDValue Op,
                                                      SelectionDAG &DAG) {
  DebugLoc DL = Op.getDebugLoc();
  SDValue OpLHS = Op.getOperand(0);
  EVT VT = OpLHS.getValueType();

  assert((VT == MVT::v8i16 || VT == MVT::v16i8) &&
         "Expect an v8i16/v16i8 type");
  OpLHS = DAG.getNode(ARMISD::VREV64, DL, VT, OpLHS);
  // For a v16i8 type: After the VREV, we have got <8, ...15, 8, ..., 0>. Now,
  // extract the first 8 bytes into the top double word and the last 8 bytes
  // into the bottom double word. The v8i16 case is similar.
  unsigned ExtractNum = (VT == MVT::v16i8) ? 8 : 4;
  return DAG.getNode(ARMISD::VEXT, DL, VT, OpLHS, OpLHS,
                     DAG.getConstant(ExtractNum, MVT::i32));
}

static SDValue LowerVECTOR_SHUFFLE(SDValue Op, SelectionDAG &DAG) {
  SDValue V1 = Op.getOperand(0);
  SDValue V2 = Op.getOperand(1);
  DebugLoc dl = Op.getDebugLoc();
  EVT VT = Op.getValueType();
  ShuffleVectorSDNode *SVN = cast<ShuffleVectorSDNode>(Op.getNode());

  // Convert shuffles that are directly supported on NEON to target-specific
  // DAG nodes, instead of keeping them as shuffles and matching them again
  // during code selection.  This is more efficient and avoids the possibility
  // of inconsistencies between legalization and selection.
  // FIXME: floating-point vectors should be canonicalized to integer vectors
  // of the same time so that they get CSEd properly.
  ArrayRef<int> ShuffleMask = SVN->getMask();

  unsigned EltSize = VT.getVectorElementType().getSizeInBits();
  if (EltSize <= 32) {
    if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0], VT)) {
      int Lane = SVN->getSplatIndex();
      // If this is undef splat, generate it via "just" vdup, if possible.
      if (Lane == -1) Lane = 0;

      // Test if V1 is a SCALAR_TO_VECTOR.
      if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR) {
        return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
      }
      // Test if V1 is a BUILD_VECTOR which is equivalent to a SCALAR_TO_VECTOR
      // (and probably will turn into a SCALAR_TO_VECTOR once legalization
      // reaches it).
      if (Lane == 0 && V1.getOpcode() == ISD::BUILD_VECTOR &&
          !isa<ConstantSDNode>(V1.getOperand(0))) {
        bool IsScalarToVector = true;
        for (unsigned i = 1, e = V1.getNumOperands(); i != e; ++i)
          if (V1.getOperand(i).getOpcode() != ISD::UNDEF) {
            IsScalarToVector = false;
            break;
          }
        if (IsScalarToVector)
          return DAG.getNode(ARMISD::VDUP, dl, VT, V1.getOperand(0));
      }
      return DAG.getNode(ARMISD::VDUPLANE, dl, VT, V1,
                         DAG.getConstant(Lane, MVT::i32));
    }

    bool ReverseVEXT;
    unsigned Imm;
    if (isVEXTMask(ShuffleMask, VT, ReverseVEXT, Imm)) {
      if (ReverseVEXT)
        std::swap(V1, V2);
      return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V2,
                         DAG.getConstant(Imm, MVT::i32));
    }

    if (isVREVMask(ShuffleMask, VT, 64))
      return DAG.getNode(ARMISD::VREV64, dl, VT, V1);
    if (isVREVMask(ShuffleMask, VT, 32))
      return DAG.getNode(ARMISD::VREV32, dl, VT, V1);
    if (isVREVMask(ShuffleMask, VT, 16))
      return DAG.getNode(ARMISD::VREV16, dl, VT, V1);

    if (V2->getOpcode() == ISD::UNDEF &&
        isSingletonVEXTMask(ShuffleMask, VT, Imm)) {
      return DAG.getNode(ARMISD::VEXT, dl, VT, V1, V1,
                         DAG.getConstant(Imm, MVT::i32));
    }

    // Check for Neon shuffles that modify both input vectors in place.
    // If both results are used, i.e., if there are two shuffles with the same
    // source operands and with masks corresponding to both results of one of
    // these operations, DAG memoization will ensure that a single node is
    // used for both shuffles.
    unsigned WhichResult;
    if (isVTRNMask(ShuffleMask, VT, WhichResult))
      return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
                         V1, V2).getValue(WhichResult);
    if (isVUZPMask(ShuffleMask, VT, WhichResult))
      return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
                         V1, V2).getValue(WhichResult);
    if (isVZIPMask(ShuffleMask, VT, WhichResult))
      return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
                         V1, V2).getValue(WhichResult);

    if (isVTRN_v_undef_Mask(ShuffleMask, VT, WhichResult))
      return DAG.getNode(ARMISD::VTRN, dl, DAG.getVTList(VT, VT),
                         V1, V1).getValue(WhichResult);
    if (isVUZP_v_undef_Mask(ShuffleMask, VT, WhichResult))
      return DAG.getNode(ARMISD::VUZP, dl, DAG.getVTList(VT, VT),
                         V1, V1).getValue(WhichResult);
    if (isVZIP_v_undef_Mask(ShuffleMask, VT, WhichResult))
      return DAG.getNode(ARMISD::VZIP, dl, DAG.getVTList(VT, VT),
                         V1, V1).getValue(WhichResult);
  }

  // If the shuffle is not directly supported and it has 4 elements, use
  // the PerfectShuffle-generated table to synthesize it from other shuffles.
  unsigned NumElts = VT.getVectorNumElements();
  if (NumElts == 4) {
    unsigned PFIndexes[4];
    for (unsigned i = 0; i != 4; ++i) {
      if (ShuffleMask[i] < 0)
        PFIndexes[i] = 8;
      else
        PFIndexes[i] = ShuffleMask[i];
    }

    // Compute the index in the perfect shuffle table.
    unsigned PFTableIndex =
      PFIndexes[0]*9*9*9+PFIndexes[1]*9*9+PFIndexes[2]*9+PFIndexes[3];
    unsigned PFEntry = PerfectShuffleTable[PFTableIndex];
    unsigned Cost = (PFEntry >> 30);

    if (Cost <= 4)
      return GeneratePerfectShuffle(PFEntry, V1, V2, DAG, dl);
  }

  // Implement shuffles with 32- or 64-bit elements as ARMISD::BUILD_VECTORs.
  if (EltSize >= 32) {
    // Do the expansion with floating-point types, since that is what the VFP
    // registers are defined to use, and since i64 is not legal.
    EVT EltVT = EVT::getFloatingPointVT(EltSize);
    EVT VecVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumElts);
    V1 = DAG.getNode(ISD::BITCAST, dl, VecVT, V1);
    V2 = DAG.getNode(ISD::BITCAST, dl, VecVT, V2);
    SmallVector<SDValue, 8> Ops;
    for (unsigned i = 0; i < NumElts; ++i) {
      if (ShuffleMask[i] < 0)
        Ops.push_back(DAG.getUNDEF(EltVT));
      else
        Ops.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT,
                                  ShuffleMask[i] < (int)NumElts ? V1 : V2,
                                  DAG.getConstant(ShuffleMask[i] & (NumElts-1),
                                                  MVT::i32)));
    }
    SDValue Val = DAG.getNode(ARMISD::BUILD_VECTOR, dl, VecVT, &Ops[0],NumElts);
    return DAG.getNode(ISD::BITCAST, dl, VT, Val);
  }

  if ((VT == MVT::v8i16 || VT == MVT::v16i8) && isReverseMask(ShuffleMask, VT))
    return LowerReverse_VECTOR_SHUFFLEv16i8_v8i16(Op, DAG);

  if (VT == MVT::v8i8) {
    SDValue NewOp = LowerVECTOR_SHUFFLEv8i8(Op, ShuffleMask, DAG);
    if (NewOp.getNode())
      return NewOp;
  }

  return SDValue();
}

static SDValue LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
  // INSERT_VECTOR_ELT is legal only for immediate indexes.
  SDValue Lane = Op.getOperand(2);
  if (!isa<ConstantSDNode>(Lane))
    return SDValue();

  return Op;
}

static SDValue LowerEXTRACT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) {
  // EXTRACT_VECTOR_ELT is legal only for immediate indexes.
  SDValue Lane = Op.getOperand(1);
  if (!isa<ConstantSDNode>(Lane))
    return SDValue();

  SDValue Vec = Op.getOperand(0);
  if (Op.getValueType() == MVT::i32 &&
      Vec.getValueType().getVectorElementType().getSizeInBits() < 32) {
    DebugLoc dl = Op.getDebugLoc();
    return DAG.getNode(ARMISD::VGETLANEu, dl, MVT::i32, Vec, Lane);
  }

  return Op;
}

static SDValue LowerCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
  // The only time a CONCAT_VECTORS operation can have legal types is when
  // two 64-bit vectors are concatenated to a 128-bit vector.
  assert(Op.getValueType().is128BitVector() && Op.getNumOperands() == 2 &&
         "unexpected CONCAT_VECTORS");
  DebugLoc dl = Op.getDebugLoc();
  SDValue Val = DAG.getUNDEF(MVT::v2f64);
  SDValue Op0 = Op.getOperand(0);
  SDValue Op1 = Op.getOperand(1);
  if (Op0.getOpcode() != ISD::UNDEF)
    Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
                      DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op0),
                      DAG.getIntPtrConstant(0));
  if (Op1.getOpcode() != ISD::UNDEF)
    Val = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v2f64, Val,
                      DAG.getNode(ISD::BITCAST, dl, MVT::f64, Op1),
                      DAG.getIntPtrConstant(1));
  return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Val);
}

/// isExtendedBUILD_VECTOR - Check if N is a constant BUILD_VECTOR where each
/// element has been zero/sign-extended, depending on the isSigned parameter,
/// from an integer type half its size.
static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
                                   bool isSigned) {
  // A v2i64 BUILD_VECTOR will have been legalized to a BITCAST from v4i32.
  EVT VT = N->getValueType(0);
  if (VT == MVT::v2i64 && N->getOpcode() == ISD::BITCAST) {
    SDNode *BVN = N->getOperand(0).getNode();
    if (BVN->getValueType(0) != MVT::v4i32 ||
        BVN->getOpcode() != ISD::BUILD_VECTOR)
      return false;
    unsigned LoElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
    unsigned HiElt = 1 - LoElt;
    ConstantSDNode *Lo0 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt));
    ConstantSDNode *Hi0 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt));
    ConstantSDNode *Lo1 = dyn_cast<ConstantSDNode>(BVN->getOperand(LoElt+2));
    ConstantSDNode *Hi1 = dyn_cast<ConstantSDNode>(BVN->getOperand(HiElt+2));
    if (!Lo0 || !Hi0 || !Lo1 || !Hi1)
      return false;
    if (isSigned) {
      if (Hi0->getSExtValue() == Lo0->getSExtValue() >> 32 &&
          Hi1->getSExtValue() == Lo1->getSExtValue() >> 32)
        return true;
    } else {
      if (Hi0->isNullValue() && Hi1->isNullValue())
        return true;
    }
    return false;
  }

  if (N->getOpcode() != ISD::BUILD_VECTOR)
    return false;

  for (unsigned i = 0, e = N->getNumOperands(); i != e; ++i) {
    SDNode *Elt = N->getOperand(i).getNode();
    if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Elt)) {
      unsigned EltSize = VT.getVectorElementType().getSizeInBits();
      unsigned HalfSize = EltSize / 2;
      if (isSigned) {
        if (!isIntN(HalfSize, C->getSExtValue()))
          return false;
      } else {
        if (!isUIntN(HalfSize, C->getZExtValue()))
          return false;
      }
      continue;
    }
    return false;
  }

  return true;
}

/// isSignExtended - Check if a node is a vector value that is sign-extended
/// or a constant BUILD_VECTOR with sign-extended elements.
static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
  if (N->getOpcode() == ISD::SIGN_EXTEND || ISD::isSEXTLoad(N))
    return true;
  if (isExtendedBUILD_VECTOR(N, DAG, true))
    return true;
  return false;
}

/// isZeroExtended - Check if a node is a vector value that is zero-extended
/// or a constant BUILD_VECTOR with zero-extended elements.
static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
  if (N->getOpcode() == ISD::ZERO_EXTEND || ISD::isZEXTLoad(N))
    return true;
  if (isExtendedBUILD_VECTOR(N, DAG, false))
    return true;
  return false;
}

static EVT getExtensionTo64Bits(const EVT &OrigVT) {
  if (OrigVT.getSizeInBits() >= 64)
    return OrigVT;

  assert(OrigVT.isSimple() && "Expecting a simple value type");

  MVT::SimpleValueType OrigSimpleTy = OrigVT.getSimpleVT().SimpleTy;
  switch (OrigSimpleTy) {
  default: llvm_unreachable("Unexpected Vector Type");
  case MVT::v2i8:
  case MVT::v2i16:
     return MVT::v2i32;
  case MVT::v4i8:
    return  MVT::v4i16;
  }
}

/// AddRequiredExtensionForVMULL - Add a sign/zero extension to extend the total
/// value size to 64 bits. We need a 64-bit D register as an operand to VMULL.
/// We insert the required extension here to get the vector to fill a D register.
static SDValue AddRequiredExtensionForVMULL(SDValue N, SelectionDAG &DAG,
                                            const EVT &OrigTy,
                                            const EVT &ExtTy,
                                            unsigned ExtOpcode) {
  // The vector originally had a size of OrigTy. It was then extended to ExtTy.
  // We expect the ExtTy to be 128-bits total. If the OrigTy is less than
  // 64-bits we need to insert a new extension so that it will be 64-bits.
  assert(ExtTy.is128BitVector() && "Unexpected extension size");
  if (OrigTy.getSizeInBits() >= 64)
    return N;

  // Must extend size to at least 64 bits to be used as an operand for VMULL.
  EVT NewVT = getExtensionTo64Bits(OrigTy);

  return DAG.getNode(ExtOpcode, N->getDebugLoc(), NewVT, N);
}

/// SkipLoadExtensionForVMULL - return a load of the original vector size that
/// does not do any sign/zero extension. If the original vector is less
/// than 64 bits, an appropriate extension will be added after the load to
/// reach a total size of 64 bits. We have to add the extension separately
/// because ARM does not have a sign/zero extending load for vectors.
static SDValue SkipLoadExtensionForVMULL(LoadSDNode *LD, SelectionDAG& DAG) {
  EVT ExtendedTy = getExtensionTo64Bits(LD->getMemoryVT());

  // The load already has the right type.
  if (ExtendedTy == LD->getMemoryVT())
    return DAG.getLoad(LD->getMemoryVT(), LD->getDebugLoc(), LD->getChain(),
                LD->getBasePtr(), LD->getPointerInfo(), LD->isVolatile(),
                LD->isNonTemporal(), LD->isInvariant(),
                LD->getAlignment());

  // We need to create a zextload/sextload. We cannot just create a load
  // followed by a zext/zext node because LowerMUL is also run during normal
  // operation legalization where we can't create illegal types.
  return DAG.getExtLoad(LD->getExtensionType(), LD->getDebugLoc(), ExtendedTy,
                        LD->getChain(), LD->getBasePtr(), LD->getPointerInfo(),
                        LD->getMemoryVT(), LD->isVolatile(),
                        LD->isNonTemporal(), LD->getAlignment());
}

/// SkipExtensionForVMULL - For a node that is a SIGN_EXTEND, ZERO_EXTEND,
/// extending load, or BUILD_VECTOR with extended elements, return the
/// unextended value. The unextended vector should be 64 bits so that it can
/// be used as an operand to a VMULL instruction. If the original vector size
/// before extension is less than 64 bits we add a an extension to resize
/// the vector to 64 bits.
static SDValue SkipExtensionForVMULL(SDNode *N, SelectionDAG &DAG) {
  if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
    return AddRequiredExtensionForVMULL(N->getOperand(0), DAG,
                                        N->getOperand(0)->getValueType(0),
                                        N->getValueType(0),
                                        N->getOpcode());

  if (LoadSDNode *LD = dyn_cast<LoadSDNode>(N))
    return SkipLoadExtensionForVMULL(LD, DAG);

  // Otherwise, the value must be a BUILD_VECTOR.  For v2i64, it will
  // have been legalized as a BITCAST from v4i32.
  if (N->getOpcode() == ISD::BITCAST) {
    SDNode *BVN = N->getOperand(0).getNode();
    assert(BVN->getOpcode() == ISD::BUILD_VECTOR &&
           BVN->getValueType(0) == MVT::v4i32 && "expected v4i32 BUILD_VECTOR");
    unsigned LowElt = DAG.getTargetLoweringInfo().isBigEndian() ? 1 : 0;
    return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(), MVT::v2i32,
                       BVN->getOperand(LowElt), BVN->getOperand(LowElt+2));
  }
  // Construct a new BUILD_VECTOR with elements truncated to half the size.
  assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
  EVT VT = N->getValueType(0);
  unsigned EltSize = VT.getVectorElementType().getSizeInBits() / 2;
  unsigned NumElts = VT.getVectorNumElements();
  MVT TruncVT = MVT::getIntegerVT(EltSize);
  SmallVector<SDValue, 8> Ops;
  for (unsigned i = 0; i != NumElts; ++i) {
    ConstantSDNode *C = cast<ConstantSDNode>(N->getOperand(i));
    const APInt &CInt = C->getAPIntValue();
    // Element types smaller than 32 bits are not legal, so use i32 elements.
    // The values are implicitly truncated so sext vs. zext doesn't matter.
    Ops.push_back(DAG.getConstant(CInt.zextOrTrunc(32), MVT::i32));
  }
  return DAG.getNode(ISD::BUILD_VECTOR, N->getDebugLoc(),
                     MVT::getVectorVT(TruncVT, NumElts), Ops.data(), NumElts);
}

static bool isAddSubSExt(SDNode *N, SelectionDAG &DAG) {
  unsigned Opcode = N->getOpcode();
  if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
    SDNode *N0 = N->getOperand(0).getNode();
    SDNode *N1 = N->getOperand(1).getNode();
    return N0->hasOneUse() && N1->hasOneUse() &&
      isSignExtended(N0, DAG) && isSignExtended(N1, DAG);
  }
  return false;
}

static bool isAddSubZExt(SDNode *N, SelectionDAG &DAG) {
  unsigned Opcode = N->getOpcode();
  if (Opcode == ISD::ADD || Opcode == ISD::SUB) {
    SDNode *N0 = N->getOperand(0).getNode();
    SDNode *N1 = N->getOperand(1).getNode();
    return N0->hasOneUse() && N1->hasOneUse() &&
      isZeroExtended(N0, DAG) && isZeroExtended(N1, DAG);
  }
  return false;
}

static SDValue LowerMUL(SDValue Op, SelectionDAG &DAG) {
  // Multiplications are only custom-lowered for 128-bit vectors so that
  // VMULL can be detected.  Otherwise v2i64 multiplications are not legal.
  EVT VT = Op.getValueType();
  assert(VT.is128BitVector() && VT.isInteger() &&
         "unexpected type for custom-lowering ISD::MUL");
  SDNode *N0 = Op.getOperand(0).getNode();
  SDNode *N1 = Op.getOperand(1).getNode();
  unsigned NewOpc = 0;
  bool isMLA = false;
  bool isN0SExt = isSignExtended(N0, DAG);
  bool isN1SExt = isSignExtended(N1, DAG);
  if (isN0SExt && isN1SExt)
    NewOpc = ARMISD::VMULLs;
  else {
    bool isN0ZExt = isZeroExtended(N0, DAG);
    bool isN1ZExt = isZeroExtended(N1, DAG);
    if (isN0ZExt && isN1ZExt)
      NewOpc = ARMISD::VMULLu;
    else if (isN1SExt || isN1ZExt) {
      // Look for (s/zext A + s/zext B) * (s/zext C). We want to turn these
      // into (s/zext A * s/zext C) + (s/zext B * s/zext C)
      if (isN1SExt && isAddSubSExt(N0, DAG)) {
        NewOpc = ARMISD::VMULLs;
        isMLA = true;
      } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
        NewOpc = ARMISD::VMULLu;
        isMLA = true;
      } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
        std::swap(N0, N1);
        NewOpc = ARMISD::VMULLu;
        isMLA = true;
      }
    }

    if (!NewOpc) {
      if (VT == MVT::v2i64)
        // Fall through to expand this.  It is not legal.
        return SDValue();
      else
        // Other vector multiplications are legal.
        return Op;
    }
  }

  // Legalize to a VMULL instruction.
  DebugLoc DL = Op.