xref: /external/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp

//===-- AArch64ISelLowering.cpp - AArch64 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 implements the AArch64TargetLowering class.
//
//===----------------------------------------------------------------------===//

#include "AArch64ISelLowering.h"
#include "AArch64CallingConvention.h"
#include "AArch64MachineFunctionInfo.h"
#include "AArch64PerfectShuffle.h"
#include "AArch64Subtarget.h"
#include "AArch64TargetMachine.h"
#include "AArch64TargetObjectFile.h"
#include "MCTargetDesc/AArch64AddressingModes.h"
#include "llvm/ADT/Statistic.h"
#include "llvm/CodeGen/CallingConvLower.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineInstrBuilder.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
#include "llvm/IR/Function.h"
#include "llvm/IR/GetElementPtrTypeIterator.h"
#include "llvm/IR/Intrinsics.h"
#include "llvm/IR/Type.h"
#include "llvm/Support/CommandLine.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/raw_ostream.h"
#include "llvm/Target/TargetOptions.h"
using namespace llvm;

#define DEBUG_TYPE "aarch64-lower"

STATISTIC(NumTailCalls, "Number of tail calls");
STATISTIC(NumShiftInserts, "Number of vector shift inserts");

// Place holder until extr generation is tested fully.
static cl::opt<bool>
EnableAArch64ExtrGeneration("aarch64-extr-generation", cl::Hidden,
                          cl::desc("Allow AArch64 (or (shift)(shift))->extract"),
                          cl::init(true));

static cl::opt<bool>
EnableAArch64SlrGeneration("aarch64-shift-insert-generation", cl::Hidden,
                           cl::desc("Allow AArch64 SLI/SRI formation"),
                           cl::init(false));

// FIXME: The necessary dtprel relocations don't seem to be supported
// well in the GNU bfd and gold linkers at the moment. Therefore, by
// default, for now, fall back to GeneralDynamic code generation.
cl::opt<bool> EnableAArch64ELFLocalDynamicTLSGeneration(
    "aarch64-elf-ldtls-generation", cl::Hidden,
    cl::desc("Allow AArch64 Local Dynamic TLS code generation"),
    cl::init(false));

/// Value type used for condition codes.
static const MVT MVT_CC = MVT::i32;

AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
                                             const AArch64Subtarget &STI)
    : TargetLowering(TM), Subtarget(&STI) {

  // AArch64 doesn't have comparisons which set GPRs or setcc instructions, so
  // we have to make something up. Arbitrarily, choose ZeroOrOne.
  setBooleanContents(ZeroOrOneBooleanContent);
  // When comparing vectors the result sets the different elements in the
  // vector to all-one or all-zero.
  setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);

  // Set up the register classes.
  addRegisterClass(MVT::i32, &AArch64::GPR32allRegClass);
  addRegisterClass(MVT::i64, &AArch64::GPR64allRegClass);

  if (Subtarget->hasFPARMv8()) {
    addRegisterClass(MVT::f16, &AArch64::FPR16RegClass);
    addRegisterClass(MVT::f32, &AArch64::FPR32RegClass);
    addRegisterClass(MVT::f64, &AArch64::FPR64RegClass);
    addRegisterClass(MVT::f128, &AArch64::FPR128RegClass);
  }

  if (Subtarget->hasNEON()) {
    addRegisterClass(MVT::v16i8, &AArch64::FPR8RegClass);
    addRegisterClass(MVT::v8i16, &AArch64::FPR16RegClass);
    // Someone set us up the NEON.
    addDRTypeForNEON(MVT::v2f32);
    addDRTypeForNEON(MVT::v8i8);
    addDRTypeForNEON(MVT::v4i16);
    addDRTypeForNEON(MVT::v2i32);
    addDRTypeForNEON(MVT::v1i64);
    addDRTypeForNEON(MVT::v1f64);
    addDRTypeForNEON(MVT::v4f16);

    addQRTypeForNEON(MVT::v4f32);
    addQRTypeForNEON(MVT::v2f64);
    addQRTypeForNEON(MVT::v16i8);
    addQRTypeForNEON(MVT::v8i16);
    addQRTypeForNEON(MVT::v4i32);
    addQRTypeForNEON(MVT::v2i64);
    addQRTypeForNEON(MVT::v8f16);
  }

  // Compute derived properties from the register classes
  computeRegisterProperties(Subtarget->getRegisterInfo());

  // Provide all sorts of operation actions
  setOperationAction(ISD::GlobalAddress, MVT::i64, Custom);
  setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
  setOperationAction(ISD::SETCC, MVT::i32, Custom);
  setOperationAction(ISD::SETCC, MVT::i64, Custom);
  setOperationAction(ISD::SETCC, MVT::f32, Custom);
  setOperationAction(ISD::SETCC, MVT::f64, Custom);
  setOperationAction(ISD::BRCOND, MVT::Other, Expand);
  setOperationAction(ISD::BR_CC, MVT::i32, Custom);
  setOperationAction(ISD::BR_CC, MVT::i64, Custom);
  setOperationAction(ISD::BR_CC, MVT::f32, Custom);
  setOperationAction(ISD::BR_CC, MVT::f64, Custom);
  setOperationAction(ISD::SELECT, MVT::i32, Custom);
  setOperationAction(ISD::SELECT, MVT::i64, 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::i64, Custom);
  setOperationAction(ISD::SELECT_CC, MVT::f32, Custom);
  setOperationAction(ISD::SELECT_CC, MVT::f64, Custom);
  setOperationAction(ISD::BR_JT, MVT::Other, Expand);
  setOperationAction(ISD::JumpTable, MVT::i64, Custom);

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

  setOperationAction(ISD::FREM, MVT::f32, Expand);
  setOperationAction(ISD::FREM, MVT::f64, Expand);
  setOperationAction(ISD::FREM, MVT::f80, Expand);

  // Custom lowering hooks are needed for XOR
  // to fold it into CSINC/CSINV.
  setOperationAction(ISD::XOR, MVT::i32, Custom);
  setOperationAction(ISD::XOR, MVT::i64, Custom);

  // Virtually no operation on f128 is legal, but LLVM can't expand them when
  // there's a valid register class, so we need custom operations in most cases.
  setOperationAction(ISD::FABS, MVT::f128, Expand);
  setOperationAction(ISD::FADD, MVT::f128, Custom);
  setOperationAction(ISD::FCOPYSIGN, MVT::f128, Expand);
  setOperationAction(ISD::FCOS, MVT::f128, Expand);
  setOperationAction(ISD::FDIV, MVT::f128, Custom);
  setOperationAction(ISD::FMA, MVT::f128, Expand);
  setOperationAction(ISD::FMUL, MVT::f128, Custom);
  setOperationAction(ISD::FNEG, MVT::f128, Expand);
  setOperationAction(ISD::FPOW, MVT::f128, Expand);
  setOperationAction(ISD::FREM, MVT::f128, Expand);
  setOperationAction(ISD::FRINT, MVT::f128, Expand);
  setOperationAction(ISD::FSIN, MVT::f128, Expand);
  setOperationAction(ISD::FSINCOS, MVT::f128, Expand);
  setOperationAction(ISD::FSQRT, MVT::f128, Expand);
  setOperationAction(ISD::FSUB, MVT::f128, Custom);
  setOperationAction(ISD::FTRUNC, MVT::f128, Expand);
  setOperationAction(ISD::SETCC, MVT::f128, Custom);
  setOperationAction(ISD::BR_CC, MVT::f128, Custom);
  setOperationAction(ISD::SELECT, MVT::f128, Custom);
  setOperationAction(ISD::SELECT_CC, MVT::f128, Custom);
  setOperationAction(ISD::FP_EXTEND, MVT::f128, Custom);

  // Lowering for many of the conversions is actually specified by the non-f128
  // type. The LowerXXX function will be trivial when f128 isn't involved.
  setOperationAction(ISD::FP_TO_SINT, MVT::i32, Custom);
  setOperationAction(ISD::FP_TO_SINT, MVT::i64, Custom);
  setOperationAction(ISD::FP_TO_SINT, MVT::i128, Custom);
  setOperationAction(ISD::FP_TO_UINT, MVT::i32, Custom);
  setOperationAction(ISD::FP_TO_UINT, MVT::i64, Custom);
  setOperationAction(ISD::FP_TO_UINT, MVT::i128, Custom);
  setOperationAction(ISD::SINT_TO_FP, MVT::i32, Custom);
  setOperationAction(ISD::SINT_TO_FP, MVT::i64, Custom);
  setOperationAction(ISD::SINT_TO_FP, MVT::i128, Custom);
  setOperationAction(ISD::UINT_TO_FP, MVT::i32, Custom);
  setOperationAction(ISD::UINT_TO_FP, MVT::i64, Custom);
  setOperationAction(ISD::UINT_TO_FP, MVT::i128, Custom);
  setOperationAction(ISD::FP_ROUND, MVT::f32, Custom);
  setOperationAction(ISD::FP_ROUND, MVT::f64, Custom);

  // Variable arguments.
  setOperationAction(ISD::VASTART, MVT::Other, Custom);
  setOperationAction(ISD::VAARG, MVT::Other, Custom);
  setOperationAction(ISD::VACOPY, MVT::Other, Custom);
  setOperationAction(ISD::VAEND, MVT::Other, Expand);

  // Variable-sized objects.
  setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
  setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
  setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);

  // Constant pool entries
  setOperationAction(ISD::ConstantPool, MVT::i64, Custom);

  // BlockAddress
  setOperationAction(ISD::BlockAddress, MVT::i64, Custom);

  // Add/Sub overflow ops with MVT::Glues are lowered to NZCV dependences.
  setOperationAction(ISD::ADDC, MVT::i32, Custom);
  setOperationAction(ISD::ADDE, MVT::i32, Custom);
  setOperationAction(ISD::SUBC, MVT::i32, Custom);
  setOperationAction(ISD::SUBE, MVT::i32, Custom);
  setOperationAction(ISD::ADDC, MVT::i64, Custom);
  setOperationAction(ISD::ADDE, MVT::i64, Custom);
  setOperationAction(ISD::SUBC, MVT::i64, Custom);
  setOperationAction(ISD::SUBE, MVT::i64, Custom);

  // AArch64 lacks both left-rotate and popcount instructions.
  setOperationAction(ISD::ROTL, MVT::i32, Expand);
  setOperationAction(ISD::ROTL, MVT::i64, Expand);
  for (MVT VT : MVT::vector_valuetypes()) {
    setOperationAction(ISD::ROTL, VT, Expand);
    setOperationAction(ISD::ROTR, VT, Expand);
  }

  // AArch64 doesn't have {U|S}MUL_LOHI.
  setOperationAction(ISD::UMUL_LOHI, MVT::i64, Expand);
  setOperationAction(ISD::SMUL_LOHI, MVT::i64, Expand);


  // Expand the undefined-at-zero variants to cttz/ctlz to their defined-at-zero
  // counterparts, which AArch64 supports directly.
  setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32, Expand);
  setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32, Expand);
  setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
  setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);

  setOperationAction(ISD::CTPOP, MVT::i32, Custom);
  setOperationAction(ISD::CTPOP, MVT::i64, Custom);

  setOperationAction(ISD::SDIVREM, MVT::i32, Expand);
  setOperationAction(ISD::SDIVREM, MVT::i64, Expand);
  for (MVT VT : MVT::vector_valuetypes()) {
    setOperationAction(ISD::SDIVREM, VT, Expand);
    setOperationAction(ISD::UDIVREM, VT, Expand);
  }
  setOperationAction(ISD::SREM, MVT::i32, Expand);
  setOperationAction(ISD::SREM, MVT::i64, Expand);
  setOperationAction(ISD::UDIVREM, MVT::i32, Expand);
  setOperationAction(ISD::UDIVREM, MVT::i64, Expand);
  setOperationAction(ISD::UREM, MVT::i32, Expand);
  setOperationAction(ISD::UREM, MVT::i64, Expand);

  // Custom lower Add/Sub/Mul with overflow.
  setOperationAction(ISD::SADDO, MVT::i32, Custom);
  setOperationAction(ISD::SADDO, MVT::i64, Custom);
  setOperationAction(ISD::UADDO, MVT::i32, Custom);
  setOperationAction(ISD::UADDO, MVT::i64, Custom);
  setOperationAction(ISD::SSUBO, MVT::i32, Custom);
  setOperationAction(ISD::SSUBO, MVT::i64, Custom);
  setOperationAction(ISD::USUBO, MVT::i32, Custom);
  setOperationAction(ISD::USUBO, MVT::i64, Custom);
  setOperationAction(ISD::SMULO, MVT::i32, Custom);
  setOperationAction(ISD::SMULO, MVT::i64, Custom);
  setOperationAction(ISD::UMULO, MVT::i32, Custom);
  setOperationAction(ISD::UMULO, MVT::i64, Custom);

  setOperationAction(ISD::FSIN, MVT::f32, Expand);
  setOperationAction(ISD::FSIN, MVT::f64, Expand);
  setOperationAction(ISD::FCOS, MVT::f32, Expand);
  setOperationAction(ISD::FCOS, MVT::f64, Expand);
  setOperationAction(ISD::FPOW, MVT::f32, Expand);
  setOperationAction(ISD::FPOW, MVT::f64, Expand);
  setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
  setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);

  // f16 is a storage-only type, always promote it to f32.
  setOperationAction(ISD::SETCC,       MVT::f16,  Promote);
  setOperationAction(ISD::BR_CC,       MVT::f16,  Promote);
  setOperationAction(ISD::SELECT_CC,   MVT::f16,  Promote);
  setOperationAction(ISD::SELECT,      MVT::f16,  Promote);
  setOperationAction(ISD::FADD,        MVT::f16,  Promote);
  setOperationAction(ISD::FSUB,        MVT::f16,  Promote);
  setOperationAction(ISD::FMUL,        MVT::f16,  Promote);
  setOperationAction(ISD::FDIV,        MVT::f16,  Promote);
  setOperationAction(ISD::FREM,        MVT::f16,  Promote);
  setOperationAction(ISD::FMA,         MVT::f16,  Promote);
  setOperationAction(ISD::FNEG,        MVT::f16,  Promote);
  setOperationAction(ISD::FABS,        MVT::f16,  Promote);
  setOperationAction(ISD::FCEIL,       MVT::f16,  Promote);
  setOperationAction(ISD::FCOPYSIGN,   MVT::f16,  Promote);
  setOperationAction(ISD::FCOS,        MVT::f16,  Promote);
  setOperationAction(ISD::FFLOOR,      MVT::f16,  Promote);
  setOperationAction(ISD::FNEARBYINT,  MVT::f16,  Promote);
  setOperationAction(ISD::FPOW,        MVT::f16,  Promote);
  setOperationAction(ISD::FPOWI,       MVT::f16,  Promote);
  setOperationAction(ISD::FRINT,       MVT::f16,  Promote);
  setOperationAction(ISD::FSIN,        MVT::f16,  Promote);
  setOperationAction(ISD::FSINCOS,     MVT::f16,  Promote);
  setOperationAction(ISD::FSQRT,       MVT::f16,  Promote);
  setOperationAction(ISD::FEXP,        MVT::f16,  Promote);
  setOperationAction(ISD::FEXP2,       MVT::f16,  Promote);
  setOperationAction(ISD::FLOG,        MVT::f16,  Promote);
  setOperationAction(ISD::FLOG2,       MVT::f16,  Promote);
  setOperationAction(ISD::FLOG10,      MVT::f16,  Promote);
  setOperationAction(ISD::FROUND,      MVT::f16,  Promote);
  setOperationAction(ISD::FTRUNC,      MVT::f16,  Promote);
  setOperationAction(ISD::FMINNUM,     MVT::f16,  Promote);
  setOperationAction(ISD::FMAXNUM,     MVT::f16,  Promote);
  setOperationAction(ISD::FMINNAN,     MVT::f16,  Promote);
  setOperationAction(ISD::FMAXNAN,     MVT::f16,  Promote);

  // v4f16 is also a storage-only type, so promote it to v4f32 when that is
  // known to be safe.
  setOperationAction(ISD::FADD, MVT::v4f16, Promote);
  setOperationAction(ISD::FSUB, MVT::v4f16, Promote);
  setOperationAction(ISD::FMUL, MVT::v4f16, Promote);
  setOperationAction(ISD::FDIV, MVT::v4f16, Promote);
  setOperationAction(ISD::FP_EXTEND, MVT::v4f16, Promote);
  setOperationAction(ISD::FP_ROUND, MVT::v4f16, Promote);
  AddPromotedToType(ISD::FADD, MVT::v4f16, MVT::v4f32);
  AddPromotedToType(ISD::FSUB, MVT::v4f16, MVT::v4f32);
  AddPromotedToType(ISD::FMUL, MVT::v4f16, MVT::v4f32);
  AddPromotedToType(ISD::FDIV, MVT::v4f16, MVT::v4f32);
  AddPromotedToType(ISD::FP_EXTEND, MVT::v4f16, MVT::v4f32);
  AddPromotedToType(ISD::FP_ROUND, MVT::v4f16, MVT::v4f32);

  // Expand all other v4f16 operations.
  // FIXME: We could generate better code by promoting some operations to
  // a pair of v4f32s
  setOperationAction(ISD::FABS, MVT::v4f16, Expand);
  setOperationAction(ISD::FCEIL, MVT::v4f16, Expand);
  setOperationAction(ISD::FCOPYSIGN, MVT::v4f16, Expand);
  setOperationAction(ISD::FCOS, MVT::v4f16, Expand);
  setOperationAction(ISD::FFLOOR, MVT::v4f16, Expand);
  setOperationAction(ISD::FMA, MVT::v4f16, Expand);
  setOperationAction(ISD::FNEARBYINT, MVT::v4f16, Expand);
  setOperationAction(ISD::FNEG, MVT::v4f16, Expand);
  setOperationAction(ISD::FPOW, MVT::v4f16, Expand);
  setOperationAction(ISD::FPOWI, MVT::v4f16, Expand);
  setOperationAction(ISD::FREM, MVT::v4f16, Expand);
  setOperationAction(ISD::FROUND, MVT::v4f16, Expand);
  setOperationAction(ISD::FRINT, MVT::v4f16, Expand);
  setOperationAction(ISD::FSIN, MVT::v4f16, Expand);
  setOperationAction(ISD::FSINCOS, MVT::v4f16, Expand);
  setOperationAction(ISD::FSQRT, MVT::v4f16, Expand);
  setOperationAction(ISD::FTRUNC, MVT::v4f16, Expand);
  setOperationAction(ISD::SETCC, MVT::v4f16, Expand);
  setOperationAction(ISD::BR_CC, MVT::v4f16, Expand);
  setOperationAction(ISD::SELECT, MVT::v4f16, Expand);
  setOperationAction(ISD::SELECT_CC, MVT::v4f16, Expand);
  setOperationAction(ISD::FEXP, MVT::v4f16, Expand);
  setOperationAction(ISD::FEXP2, MVT::v4f16, Expand);
  setOperationAction(ISD::FLOG, MVT::v4f16, Expand);
  setOperationAction(ISD::FLOG2, MVT::v4f16, Expand);
  setOperationAction(ISD::FLOG10, MVT::v4f16, Expand);


  // v8f16 is also a storage-only type, so expand it.
  setOperationAction(ISD::FABS, MVT::v8f16, Expand);
  setOperationAction(ISD::FADD, MVT::v8f16, Expand);
  setOperationAction(ISD::FCEIL, MVT::v8f16, Expand);
  setOperationAction(ISD::FCOPYSIGN, MVT::v8f16, Expand);
  setOperationAction(ISD::FCOS, MVT::v8f16, Expand);
  setOperationAction(ISD::FDIV, MVT::v8f16, Expand);
  setOperationAction(ISD::FFLOOR, MVT::v8f16, Expand);
  setOperationAction(ISD::FMA, MVT::v8f16, Expand);
  setOperationAction(ISD::FMUL, MVT::v8f16, Expand);
  setOperationAction(ISD::FNEARBYINT, MVT::v8f16, Expand);
  setOperationAction(ISD::FNEG, MVT::v8f16, Expand);
  setOperationAction(ISD::FPOW, MVT::v8f16, Expand);
  setOperationAction(ISD::FPOWI, MVT::v8f16, Expand);
  setOperationAction(ISD::FREM, MVT::v8f16, Expand);
  setOperationAction(ISD::FROUND, MVT::v8f16, Expand);
  setOperationAction(ISD::FRINT, MVT::v8f16, Expand);
  setOperationAction(ISD::FSIN, MVT::v8f16, Expand);
  setOperationAction(ISD::FSINCOS, MVT::v8f16, Expand);
  setOperationAction(ISD::FSQRT, MVT::v8f16, Expand);
  setOperationAction(ISD::FSUB, MVT::v8f16, Expand);
  setOperationAction(ISD::FTRUNC, MVT::v8f16, Expand);
  setOperationAction(ISD::SETCC, MVT::v8f16, Expand);
  setOperationAction(ISD::BR_CC, MVT::v8f16, Expand);
  setOperationAction(ISD::SELECT, MVT::v8f16, Expand);
  setOperationAction(ISD::SELECT_CC, MVT::v8f16, Expand);
  setOperationAction(ISD::FP_EXTEND, MVT::v8f16, Expand);
  setOperationAction(ISD::FEXP, MVT::v8f16, Expand);
  setOperationAction(ISD::FEXP2, MVT::v8f16, Expand);
  setOperationAction(ISD::FLOG, MVT::v8f16, Expand);
  setOperationAction(ISD::FLOG2, MVT::v8f16, Expand);
  setOperationAction(ISD::FLOG10, MVT::v8f16, Expand);

  // AArch64 has implementations of a lot of rounding-like FP operations.
  for (MVT Ty : {MVT::f32, MVT::f64}) {
    setOperationAction(ISD::FFLOOR, Ty, Legal);
    setOperationAction(ISD::FNEARBYINT, Ty, Legal);
    setOperationAction(ISD::FCEIL, Ty, Legal);
    setOperationAction(ISD::FRINT, Ty, Legal);
    setOperationAction(ISD::FTRUNC, Ty, Legal);
    setOperationAction(ISD::FROUND, Ty, Legal);
    setOperationAction(ISD::FMINNUM, Ty, Legal);
    setOperationAction(ISD::FMAXNUM, Ty, Legal);
    setOperationAction(ISD::FMINNAN, Ty, Legal);
    setOperationAction(ISD::FMAXNAN, Ty, Legal);
  }

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

  // Lower READCYCLECOUNTER using an mrs from PMCCNTR_EL0.
  // This requires the Performance Monitors extension.
  if (Subtarget->hasPerfMon())
    setOperationAction(ISD::READCYCLECOUNTER, MVT::i64, Legal);

  if (Subtarget->isTargetMachO()) {
    // For iOS, we don't want to the normal expansion of a libcall to
    // sincos. We want to issue a libcall to __sincos_stret to avoid memory
    // traffic.
    setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
    setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
  } else {
    setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
    setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
  }

  // Make floating-point constants legal for the large code model, so they don't
  // become loads from the constant pool.
  if (Subtarget->isTargetMachO() && TM.getCodeModel() == CodeModel::Large) {
    setOperationAction(ISD::ConstantFP, MVT::f32, Legal);
    setOperationAction(ISD::ConstantFP, MVT::f64, Legal);
  }

  // AArch64 does not have floating-point extending loads, i1 sign-extending
  // load, floating-point truncating stores, or v2i32->v2i16 truncating store.
  for (MVT VT : MVT::fp_valuetypes()) {
    setLoadExtAction(ISD::EXTLOAD, VT, MVT::f16, Expand);
    setLoadExtAction(ISD::EXTLOAD, VT, MVT::f32, Expand);
    setLoadExtAction(ISD::EXTLOAD, VT, MVT::f64, Expand);
    setLoadExtAction(ISD::EXTLOAD, VT, MVT::f80, Expand);
  }
  for (MVT VT : MVT::integer_valuetypes())
    setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Expand);

  setTruncStoreAction(MVT::f32, MVT::f16, Expand);
  setTruncStoreAction(MVT::f64, MVT::f32, Expand);
  setTruncStoreAction(MVT::f64, MVT::f16, Expand);
  setTruncStoreAction(MVT::f128, MVT::f80, Expand);
  setTruncStoreAction(MVT::f128, MVT::f64, Expand);
  setTruncStoreAction(MVT::f128, MVT::f32, Expand);
  setTruncStoreAction(MVT::f128, MVT::f16, Expand);

  setOperationAction(ISD::BITCAST, MVT::i16, Custom);
  setOperationAction(ISD::BITCAST, MVT::f16, Custom);

  // Indexed loads and stores are supported.
  for (unsigned im = (unsigned)ISD::PRE_INC;
       im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
    setIndexedLoadAction(im, MVT::i8, Legal);
    setIndexedLoadAction(im, MVT::i16, Legal);
    setIndexedLoadAction(im, MVT::i32, Legal);
    setIndexedLoadAction(im, MVT::i64, Legal);
    setIndexedLoadAction(im, MVT::f64, Legal);
    setIndexedLoadAction(im, MVT::f32, Legal);
    setIndexedLoadAction(im, MVT::f16, Legal);
    setIndexedStoreAction(im, MVT::i8, Legal);
    setIndexedStoreAction(im, MVT::i16, Legal);
    setIndexedStoreAction(im, MVT::i32, Legal);
    setIndexedStoreAction(im, MVT::i64, Legal);
    setIndexedStoreAction(im, MVT::f64, Legal);
    setIndexedStoreAction(im, MVT::f32, Legal);
    setIndexedStoreAction(im, MVT::f16, Legal);
  }

  // Trap.
  setOperationAction(ISD::TRAP, MVT::Other, Legal);

  // We combine OR nodes for bitfield operations.
  setTargetDAGCombine(ISD::OR);

  // Vector add and sub nodes may conceal a high-half opportunity.
  // Also, try to fold ADD into CSINC/CSINV..
  setTargetDAGCombine(ISD::ADD);
  setTargetDAGCombine(ISD::SUB);

  setTargetDAGCombine(ISD::XOR);
  setTargetDAGCombine(ISD::SINT_TO_FP);
  setTargetDAGCombine(ISD::UINT_TO_FP);

  setTargetDAGCombine(ISD::FP_TO_SINT);
  setTargetDAGCombine(ISD::FP_TO_UINT);
  setTargetDAGCombine(ISD::FDIV);

  setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);

  setTargetDAGCombine(ISD::ANY_EXTEND);
  setTargetDAGCombine(ISD::ZERO_EXTEND);
  setTargetDAGCombine(ISD::SIGN_EXTEND);
  setTargetDAGCombine(ISD::BITCAST);
  setTargetDAGCombine(ISD::CONCAT_VECTORS);
  setTargetDAGCombine(ISD::STORE);
  if (Subtarget->supportsAddressTopByteIgnored())
    setTargetDAGCombine(ISD::LOAD);

  setTargetDAGCombine(ISD::MUL);

  setTargetDAGCombine(ISD::SELECT);
  setTargetDAGCombine(ISD::VSELECT);

  setTargetDAGCombine(ISD::INTRINSIC_VOID);
  setTargetDAGCombine(ISD::INTRINSIC_W_CHAIN);
  setTargetDAGCombine(ISD::INSERT_VECTOR_ELT);
  setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);

  MaxStoresPerMemset = MaxStoresPerMemsetOptSize = 8;
  MaxStoresPerMemcpy = MaxStoresPerMemcpyOptSize = 4;
  MaxStoresPerMemmove = MaxStoresPerMemmoveOptSize = 4;

  setStackPointerRegisterToSaveRestore(AArch64::SP);

  setSchedulingPreference(Sched::Hybrid);

  // Enable TBZ/TBNZ
  MaskAndBranchFoldingIsLegal = true;
  EnableExtLdPromotion = true;

  setMinFunctionAlignment(2);

  setHasExtractBitsInsn(true);

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

  if (Subtarget->hasNEON()) {
    // FIXME: v1f64 shouldn't be legal if we can avoid it, because it leads to
    // silliness like this:
    setOperationAction(ISD::FABS, MVT::v1f64, Expand);
    setOperationAction(ISD::FADD, MVT::v1f64, Expand);
    setOperationAction(ISD::FCEIL, MVT::v1f64, Expand);
    setOperationAction(ISD::FCOPYSIGN, MVT::v1f64, Expand);
    setOperationAction(ISD::FCOS, MVT::v1f64, Expand);
    setOperationAction(ISD::FDIV, MVT::v1f64, Expand);
    setOperationAction(ISD::FFLOOR, MVT::v1f64, Expand);
    setOperationAction(ISD::FMA, MVT::v1f64, Expand);
    setOperationAction(ISD::FMUL, MVT::v1f64, Expand);
    setOperationAction(ISD::FNEARBYINT, MVT::v1f64, Expand);
    setOperationAction(ISD::FNEG, MVT::v1f64, Expand);
    setOperationAction(ISD::FPOW, MVT::v1f64, Expand);
    setOperationAction(ISD::FREM, MVT::v1f64, Expand);
    setOperationAction(ISD::FROUND, MVT::v1f64, Expand);
    setOperationAction(ISD::FRINT, MVT::v1f64, Expand);
    setOperationAction(ISD::FSIN, MVT::v1f64, Expand);
    setOperationAction(ISD::FSINCOS, MVT::v1f64, Expand);
    setOperationAction(ISD::FSQRT, MVT::v1f64, Expand);
    setOperationAction(ISD::FSUB, MVT::v1f64, Expand);
    setOperationAction(ISD::FTRUNC, MVT::v1f64, Expand);
    setOperationAction(ISD::SETCC, MVT::v1f64, Expand);
    setOperationAction(ISD::BR_CC, MVT::v1f64, Expand);
    setOperationAction(ISD::SELECT, MVT::v1f64, Expand);
    setOperationAction(ISD::SELECT_CC, MVT::v1f64, Expand);
    setOperationAction(ISD::FP_EXTEND, MVT::v1f64, Expand);

    setOperationAction(ISD::FP_TO_SINT, MVT::v1i64, Expand);
    setOperationAction(ISD::FP_TO_UINT, MVT::v1i64, Expand);
    setOperationAction(ISD::SINT_TO_FP, MVT::v1i64, Expand);
    setOperationAction(ISD::UINT_TO_FP, MVT::v1i64, Expand);
    setOperationAction(ISD::FP_ROUND, MVT::v1f64, Expand);

    setOperationAction(ISD::MUL, MVT::v1i64, Expand);

    // AArch64 doesn't have a direct vector ->f32 conversion instructions for
    // elements smaller than i32, so promote the input to i32 first.
    setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Promote);
    setOperationAction(ISD::SINT_TO_FP, MVT::v4i8, Promote);
    setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Promote);
    setOperationAction(ISD::SINT_TO_FP, MVT::v4i16, Promote);
    // i8 and i16 vector elements also need promotion to i32 for v8i8 or v8i16
    // -> v8f16 conversions.
    setOperationAction(ISD::SINT_TO_FP, MVT::v8i8, Promote);
    setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Promote);
    setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
    setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Promote);
    // Similarly, there is no direct i32 -> f64 vector conversion instruction.
    setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
    setOperationAction(ISD::UINT_TO_FP, MVT::v2i32, Custom);
    setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Custom);
    setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Custom);
    // Or, direct i32 -> f16 vector conversion.  Set it so custom, so the
    // conversion happens in two steps: v4i32 -> v4f32 -> v4f16
    setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Custom);
    setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);

    // AArch64 doesn't have MUL.2d:
    setOperationAction(ISD::MUL, MVT::v2i64, Expand);
    // Custom handling for some quad-vector types to detect MULL.
    setOperationAction(ISD::MUL, MVT::v8i16, Custom);
    setOperationAction(ISD::MUL, MVT::v4i32, Custom);
    setOperationAction(ISD::MUL, MVT::v2i64, Custom);

    setOperationAction(ISD::ANY_EXTEND, MVT::v4i32, Legal);
    setTruncStoreAction(MVT::v2i32, MVT::v2i16, Expand);
    // Likewise, narrowing and extending vector loads/stores aren't handled
    // directly.
    for (MVT VT : MVT::vector_valuetypes()) {
      setOperationAction(ISD::SIGN_EXTEND_INREG, VT, Expand);

      setOperationAction(ISD::MULHS, VT, Expand);
      setOperationAction(ISD::SMUL_LOHI, VT, Expand);
      setOperationAction(ISD::MULHU, VT, Expand);
      setOperationAction(ISD::UMUL_LOHI, VT, Expand);

      setOperationAction(ISD::BSWAP, VT, Expand);

      for (MVT InnerVT : MVT::vector_valuetypes()) {
        setTruncStoreAction(VT, InnerVT, Expand);
        setLoadExtAction(ISD::SEXTLOAD, VT, InnerVT, Expand);
        setLoadExtAction(ISD::ZEXTLOAD, VT, InnerVT, Expand);
        setLoadExtAction(ISD::EXTLOAD, VT, InnerVT, Expand);
      }
    }

    // AArch64 has implementations of a lot of rounding-like FP operations.
    for (MVT Ty : {MVT::v2f32, MVT::v4f32, MVT::v2f64}) {
      setOperationAction(ISD::FFLOOR, Ty, Legal);
      setOperationAction(ISD::FNEARBYINT, Ty, Legal);
      setOperationAction(ISD::FCEIL, Ty, Legal);
      setOperationAction(ISD::FRINT, Ty, Legal);
      setOperationAction(ISD::FTRUNC, Ty, Legal);
      setOperationAction(ISD::FROUND, Ty, Legal);
    }
  }

  // Prefer likely predicted branches to selects on out-of-order cores.
  if (Subtarget->isCortexA57())
    PredictableSelectIsExpensive = true;
}

void AArch64TargetLowering::addTypeForNEON(EVT VT, EVT PromotedBitwiseVT) {
  if (VT == MVT::v2f32 || VT == MVT::v4f16) {
    setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
    AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i32);

    setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
    AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i32);
  } else if (VT == MVT::v2f64 || VT == MVT::v4f32 || VT == MVT::v8f16) {
    setOperationAction(ISD::LOAD, VT.getSimpleVT(), Promote);
    AddPromotedToType(ISD::LOAD, VT.getSimpleVT(), MVT::v2i64);

    setOperationAction(ISD::STORE, VT.getSimpleVT(), Promote);
    AddPromotedToType(ISD::STORE, VT.getSimpleVT(), MVT::v2i64);
  }

  // Mark vector float intrinsics as expand.
  if (VT == MVT::v2f32 || VT == MVT::v4f32 || VT == MVT::v2f64) {
    setOperationAction(ISD::FSIN, VT.getSimpleVT(), Expand);
    setOperationAction(ISD::FCOS, VT.getSimpleVT(), Expand);
    setOperationAction(ISD::FPOWI, VT.getSimpleVT(), Expand);
    setOperationAction(ISD::FPOW, VT.getSimpleVT(), Expand);
    setOperationAction(ISD::FLOG, VT.getSimpleVT(), Expand);
    setOperationAction(ISD::FLOG2, VT.getSimpleVT(), Expand);
    setOperationAction(ISD::FLOG10, VT.getSimpleVT(), Expand);
    setOperationAction(ISD::FEXP, VT.getSimpleVT(), Expand);
    setOperationAction(ISD::FEXP2, VT.getSimpleVT(), Expand);

    // But we do support custom-lowering for FCOPYSIGN.
    setOperationAction(ISD::FCOPYSIGN, VT.getSimpleVT(), Custom);
  }

  setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT.getSimpleVT(), Custom);
  setOperationAction(ISD::INSERT_VECTOR_ELT, VT.getSimpleVT(), Custom);
  setOperationAction(ISD::BUILD_VECTOR, VT.getSimpleVT(), Custom);
  setOperationAction(ISD::VECTOR_SHUFFLE, VT.getSimpleVT(), Custom);
  setOperationAction(ISD::EXTRACT_SUBVECTOR, VT.getSimpleVT(), Custom);
  setOperationAction(ISD::SRA, VT.getSimpleVT(), Custom);
  setOperationAction(ISD::SRL, VT.getSimpleVT(), Custom);
  setOperationAction(ISD::SHL, VT.getSimpleVT(), Custom);
  setOperationAction(ISD::AND, VT.getSimpleVT(), Custom);
  setOperationAction(ISD::OR, VT.getSimpleVT(), Custom);
  setOperationAction(ISD::SETCC, VT.getSimpleVT(), Custom);
  setOperationAction(ISD::CONCAT_VECTORS, VT.getSimpleVT(), Legal);

  setOperationAction(ISD::SELECT, VT.getSimpleVT(), Expand);
  setOperationAction(ISD::SELECT_CC, VT.getSimpleVT(), Expand);
  setOperationAction(ISD::VSELECT, VT.getSimpleVT(), Expand);
  for (MVT InnerVT : MVT::all_valuetypes())
    setLoadExtAction(ISD::EXTLOAD, InnerVT, VT.getSimpleVT(), Expand);

  // CNT supports only B element sizes.
  if (VT != MVT::v8i8 && VT != MVT::v16i8)
    setOperationAction(ISD::CTPOP, VT.getSimpleVT(), Expand);

  setOperationAction(ISD::UDIV, VT.getSimpleVT(), Expand);
  setOperationAction(ISD::SDIV, VT.getSimpleVT(), Expand);
  setOperationAction(ISD::UREM, VT.getSimpleVT(), Expand);
  setOperationAction(ISD::SREM, VT.getSimpleVT(), Expand);
  setOperationAction(ISD::FREM, VT.getSimpleVT(), Expand);

  setOperationAction(ISD::FP_TO_SINT, VT.getSimpleVT(), Custom);
  setOperationAction(ISD::FP_TO_UINT, VT.getSimpleVT(), Custom);

  // [SU][MIN|MAX] are available for all NEON types apart from i64.
  if (!VT.isFloatingPoint() &&
      VT.getSimpleVT() != MVT::v2i64 && VT.getSimpleVT() != MVT::v1i64)
    for (unsigned Opcode : {ISD::SMIN, ISD::SMAX, ISD::UMIN, ISD::UMAX})
      setOperationAction(Opcode, VT.getSimpleVT(), Legal);

  // F[MIN|MAX][NUM|NAN] are available for all FP NEON types (not f16 though!).
  if (VT.isFloatingPoint() && VT.getVectorElementType() != MVT::f16)
    for (unsigned Opcode : {ISD::FMINNAN, ISD::FMAXNAN,
                            ISD::FMINNUM, ISD::FMAXNUM})
      setOperationAction(Opcode, VT.getSimpleVT(), Legal);

  if (Subtarget->isLittleEndian()) {
    for (unsigned im = (unsigned)ISD::PRE_INC;
         im != (unsigned)ISD::LAST_INDEXED_MODE; ++im) {
      setIndexedLoadAction(im, VT.getSimpleVT(), Legal);
      setIndexedStoreAction(im, VT.getSimpleVT(), Legal);
    }
  }
}

void AArch64TargetLowering::addDRTypeForNEON(MVT VT) {
  addRegisterClass(VT, &AArch64::FPR64RegClass);
  addTypeForNEON(VT, MVT::v2i32);
}

void AArch64TargetLowering::addQRTypeForNEON(MVT VT) {
  addRegisterClass(VT, &AArch64::FPR128RegClass);
  addTypeForNEON(VT, MVT::v4i32);
}

EVT AArch64TargetLowering::getSetCCResultType(const DataLayout &, LLVMContext &,
                                              EVT VT) const {
  if (!VT.isVector())
    return MVT::i32;
  return VT.changeVectorElementTypeToInteger();
}

/// computeKnownBitsForTargetNode - Determine which of the bits specified in
/// Mask are known to be either zero or one and return them in the
/// KnownZero/KnownOne bitsets.
void AArch64TargetLowering::computeKnownBitsForTargetNode(
    const SDValue Op, APInt &KnownZero, APInt &KnownOne,
    const SelectionDAG &DAG, unsigned Depth) const {
  switch (Op.getOpcode()) {
  default:
    break;
  case AArch64ISD::CSEL: {
    APInt KnownZero2, KnownOne2;
    DAG.computeKnownBits(Op->getOperand(0), KnownZero, KnownOne, Depth + 1);
    DAG.computeKnownBits(Op->getOperand(1), KnownZero2, KnownOne2, Depth + 1);
    KnownZero &= KnownZero2;
    KnownOne &= KnownOne2;
    break;
  }
  case ISD::INTRINSIC_W_CHAIN: {
    ConstantSDNode *CN = cast<ConstantSDNode>(Op->getOperand(1));
    Intrinsic::ID IntID = static_cast<Intrinsic::ID>(CN->getZExtValue());
    switch (IntID) {
    default: return;
    case Intrinsic::aarch64_ldaxr:
    case Intrinsic::aarch64_ldxr: {
      unsigned BitWidth = KnownOne.getBitWidth();
      EVT VT = cast<MemIntrinsicSDNode>(Op)->getMemoryVT();
      unsigned MemBits = VT.getScalarType().getSizeInBits();
      KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - MemBits);
      return;
    }
    }
    break;
  }
  case ISD::INTRINSIC_WO_CHAIN:
  case ISD::INTRINSIC_VOID: {
    unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
    switch (IntNo) {
    default:
      break;
    case Intrinsic::aarch64_neon_umaxv:
    case Intrinsic::aarch64_neon_uminv: {
      // Figure out the datatype of the vector operand. The UMINV instruction
      // will zero extend the result, so we can mark as known zero all the
      // bits larger than the element datatype. 32-bit or larget doesn't need
      // this as those are legal types and will be handled by isel directly.
      MVT VT = Op.getOperand(1).getValueType().getSimpleVT();
      unsigned BitWidth = KnownZero.getBitWidth();
      if (VT == MVT::v8i8 || VT == MVT::v16i8) {
        assert(BitWidth >= 8 && "Unexpected width!");
        APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 8);
        KnownZero |= Mask;
      } else if (VT == MVT::v4i16 || VT == MVT::v8i16) {
        assert(BitWidth >= 16 && "Unexpected width!");
        APInt Mask = APInt::getHighBitsSet(BitWidth, BitWidth - 16);
        KnownZero |= Mask;
      }
      break;
    } break;
    }
  }
  }
}

MVT AArch64TargetLowering::getScalarShiftAmountTy(const DataLayout &DL,
                                                  EVT) const {
  return MVT::i64;
}

bool AArch64TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
                                                           unsigned AddrSpace,
                                                           unsigned Align,
                                                           bool *Fast) const {
  if (Subtarget->requiresStrictAlign())
    return false;

  // FIXME: This is mostly true for Cyclone, but not necessarily others.
  if (Fast) {
    // FIXME: Define an attribute for slow unaligned accesses instead of
    // relying on the CPU type as a proxy.
    // On Cyclone, unaligned 128-bit stores are slow.
    *Fast = !Subtarget->isCyclone() || VT.getStoreSize() != 16 ||
            // See comments in performSTORECombine() for more details about
            // these conditions.

            // Code that uses clang vector extensions can mark that it
            // wants unaligned accesses to be treated as fast by
            // underspecifying alignment to be 1 or 2.
            Align <= 2 ||

            // Disregard v2i64. Memcpy lowering produces those and splitting
            // them regresses performance on micro-benchmarks and olden/bh.
            VT == MVT::v2i64;
  }
  return true;
}

FastISel *
AArch64TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
                                      const TargetLibraryInfo *libInfo) const {
  return AArch64::createFastISel(funcInfo, libInfo);
}

const char *AArch64TargetLowering::getTargetNodeName(unsigned Opcode) const {
  switch ((AArch64ISD::NodeType)Opcode) {
  case AArch64ISD::FIRST_NUMBER:      break;
  case AArch64ISD::CALL:              return "AArch64ISD::CALL";
  case AArch64ISD::ADRP:              return "AArch64ISD::ADRP";
  case AArch64ISD::ADDlow:            return "AArch64ISD::ADDlow";
  case AArch64ISD::LOADgot:           return "AArch64ISD::LOADgot";
  case AArch64ISD::RET_FLAG:          return "AArch64ISD::RET_FLAG";
  case AArch64ISD::BRCOND:            return "AArch64ISD::BRCOND";
  case AArch64ISD::CSEL:              return "AArch64ISD::CSEL";
  case AArch64ISD::FCSEL:             return "AArch64ISD::FCSEL";
  case AArch64ISD::CSINV:             return "AArch64ISD::CSINV";
  case AArch64ISD::CSNEG:             return "AArch64ISD::CSNEG";
  case AArch64ISD::CSINC:             return "AArch64ISD::CSINC";
  case AArch64ISD::THREAD_POINTER:    return "AArch64ISD::THREAD_POINTER";
  case AArch64ISD::TLSDESC_CALLSEQ:   return "AArch64ISD::TLSDESC_CALLSEQ";
  case AArch64ISD::ADC:               return "AArch64ISD::ADC";
  case AArch64ISD::SBC:               return "AArch64ISD::SBC";
  case AArch64ISD::ADDS:              return "AArch64ISD::ADDS";
  case AArch64ISD::SUBS:              return "AArch64ISD::SUBS";
  case AArch64ISD::ADCS:              return "AArch64ISD::ADCS";
  case AArch64ISD::SBCS:              return "AArch64ISD::SBCS";
  case AArch64ISD::ANDS:              return "AArch64ISD::ANDS";
  case AArch64ISD::CCMP:              return "AArch64ISD::CCMP";
  case AArch64ISD::CCMN:              return "AArch64ISD::CCMN";
  case AArch64ISD::FCCMP:             return "AArch64ISD::FCCMP";
  case AArch64ISD::FCMP:              return "AArch64ISD::FCMP";
  case AArch64ISD::DUP:               return "AArch64ISD::DUP";
  case AArch64ISD::DUPLANE8:          return "AArch64ISD::DUPLANE8";
  case AArch64ISD::DUPLANE16:         return "AArch64ISD::DUPLANE16";
  case AArch64ISD::DUPLANE32:         return "AArch64ISD::DUPLANE32";
  case AArch64ISD::DUPLANE64:         return "AArch64ISD::DUPLANE64";
  case AArch64ISD::MOVI:              return "AArch64ISD::MOVI";
  case AArch64ISD::MOVIshift:         return "AArch64ISD::MOVIshift";
  case AArch64ISD::MOVIedit:          return "AArch64ISD::MOVIedit";
  case AArch64ISD::MOVImsl:           return "AArch64ISD::MOVImsl";
  case AArch64ISD::FMOV:              return "AArch64ISD::FMOV";
  case AArch64ISD::MVNIshift:         return "AArch64ISD::MVNIshift";
  case AArch64ISD::MVNImsl:           return "AArch64ISD::MVNImsl";
  case AArch64ISD::BICi:              return "AArch64ISD::BICi";
  case AArch64ISD::ORRi:              return "AArch64ISD::ORRi";
  case AArch64ISD::BSL:               return "AArch64ISD::BSL";
  case AArch64ISD::NEG:               return "AArch64ISD::NEG";
  case AArch64ISD::EXTR:              return "AArch64ISD::EXTR";
  case AArch64ISD::ZIP1:              return "AArch64ISD::ZIP1";
  case AArch64ISD::ZIP2:              return "AArch64ISD::ZIP2";
  case AArch64ISD::UZP1:              return "AArch64ISD::UZP1";
  case AArch64ISD::UZP2:              return "AArch64ISD::UZP2";
  case AArch64ISD::TRN1:              return "AArch64ISD::TRN1";
  case AArch64ISD::TRN2:              return "AArch64ISD::TRN2";
  case AArch64ISD::REV16:             return "AArch64ISD::REV16";
  case AArch64ISD::REV32:             return "AArch64ISD::REV32";
  case AArch64ISD::REV64:             return "AArch64ISD::REV64";
  case AArch64ISD::EXT:               return "AArch64ISD::EXT";
  case AArch64ISD::VSHL:              return "AArch64ISD::VSHL";
  case AArch64ISD::VLSHR:             return "AArch64ISD::VLSHR";
  case AArch64ISD::VASHR:             return "AArch64ISD::VASHR";
  case AArch64ISD::CMEQ:              return "AArch64ISD::CMEQ";
  case AArch64ISD::CMGE:              return "AArch64ISD::CMGE";
  case AArch64ISD::CMGT:              return "AArch64ISD::CMGT";
  case AArch64ISD::CMHI:              return "AArch64ISD::CMHI";
  case AArch64ISD::CMHS:              return "AArch64ISD::CMHS";
  case AArch64ISD::FCMEQ:             return "AArch64ISD::FCMEQ";
  case AArch64ISD::FCMGE:             return "AArch64ISD::FCMGE";
  case AArch64ISD::FCMGT:             return "AArch64ISD::FCMGT";
  case AArch64ISD::CMEQz:             return "AArch64ISD::CMEQz";
  case AArch64ISD::CMGEz:             return "AArch64ISD::CMGEz";
  case AArch64ISD::CMGTz:             return "AArch64ISD::CMGTz";
  case AArch64ISD::CMLEz:             return "AArch64ISD::CMLEz";
  case AArch64ISD::CMLTz:             return "AArch64ISD::CMLTz";
  case AArch64ISD::FCMEQz:            return "AArch64ISD::FCMEQz";
  case AArch64ISD::FCMGEz:            return "AArch64ISD::FCMGEz";
  case AArch64ISD::FCMGTz:            return "AArch64ISD::FCMGTz";
  case AArch64ISD::FCMLEz:            return "AArch64ISD::FCMLEz";
  case AArch64ISD::FCMLTz:            return "AArch64ISD::FCMLTz";
  case AArch64ISD::SADDV:             return "AArch64ISD::SADDV";
  case AArch64ISD::UADDV:             return "AArch64ISD::UADDV";
  case AArch64ISD::SMINV:             return "AArch64ISD::SMINV";
  case AArch64ISD::UMINV:             return "AArch64ISD::UMINV";
  case AArch64ISD::SMAXV:             return "AArch64ISD::SMAXV";
  case AArch64ISD::UMAXV:             return "AArch64ISD::UMAXV";
  case AArch64ISD::NOT:               return "AArch64ISD::NOT";
  case AArch64ISD::BIT:               return "AArch64ISD::BIT";
  case AArch64ISD::CBZ:               return "AArch64ISD::CBZ";
  case AArch64ISD::CBNZ:              return "AArch64ISD::CBNZ";
  case AArch64ISD::TBZ:               return "AArch64ISD::TBZ";
  case AArch64ISD::TBNZ:              return "AArch64ISD::TBNZ";
  case AArch64ISD::TC_RETURN:         return "AArch64ISD::TC_RETURN";
  case AArch64ISD::PREFETCH:          return "AArch64ISD::PREFETCH";
  case AArch64ISD::SITOF:             return "AArch64ISD::SITOF";
  case AArch64ISD::UITOF:             return "AArch64ISD::UITOF";
  case AArch64ISD::NVCAST:            return "AArch64ISD::NVCAST";
  case AArch64ISD::SQSHL_I:           return "AArch64ISD::SQSHL_I";
  case AArch64ISD::UQSHL_I:           return "AArch64ISD::UQSHL_I";
  case AArch64ISD::SRSHR_I:           return "AArch64ISD::SRSHR_I";
  case AArch64ISD::URSHR_I:           return "AArch64ISD::URSHR_I";
  case AArch64ISD::SQSHLU_I:          return "AArch64ISD::SQSHLU_I";
  case AArch64ISD::WrapperLarge:      return "AArch64ISD::WrapperLarge";
  case AArch64ISD::LD2post:           return "AArch64ISD::LD2post";
  case AArch64ISD::LD3post:           return "AArch64ISD::LD3post";
  case AArch64ISD::LD4post:           return "AArch64ISD::LD4post";
  case AArch64ISD::ST2post:           return "AArch64ISD::ST2post";
  case AArch64ISD::ST3post:           return "AArch64ISD::ST3post";
  case AArch64ISD::ST4post:           return "AArch64ISD::ST4post";
  case AArch64ISD::LD1x2post:         return "AArch64ISD::LD1x2post";
  case AArch64ISD::LD1x3post:         return "AArch64ISD::LD1x3post";
  case AArch64ISD::LD1x4post:         return "AArch64ISD::LD1x4post";
  case AArch64ISD::ST1x2post:         return "AArch64ISD::ST1x2post";
  case AArch64ISD::ST1x3post:         return "AArch64ISD::ST1x3post";
  case AArch64ISD::ST1x4post:         return "AArch64ISD::ST1x4post";
  case AArch64ISD::LD1DUPpost:        return "AArch64ISD::LD1DUPpost";
  case AArch64ISD::LD2DUPpost:        return "AArch64ISD::LD2DUPpost";
  case AArch64ISD::LD3DUPpost:        return "AArch64ISD::LD3DUPpost";
  case AArch64ISD::LD4DUPpost:        return "AArch64ISD::LD4DUPpost";
  case AArch64ISD::LD1LANEpost:       return "AArch64ISD::LD1LANEpost";
  case AArch64ISD::LD2LANEpost:       return "AArch64ISD::LD2LANEpost";
  case AArch64ISD::LD3LANEpost:       return "AArch64ISD::LD3LANEpost";
  case AArch64ISD::LD4LANEpost:       return "AArch64ISD::LD4LANEpost";
  case AArch64ISD::ST2LANEpost:       return "AArch64ISD::ST2LANEpost";
  case AArch64ISD::ST3LANEpost:       return "AArch64ISD::ST3LANEpost";
  case AArch64ISD::ST4LANEpost:       return "AArch64ISD::ST4LANEpost";
  case AArch64ISD::SMULL:             return "AArch64ISD::SMULL";
  case AArch64ISD::UMULL:             return "AArch64ISD::UMULL";
  }
  return nullptr;
}

MachineBasicBlock *
AArch64TargetLowering::EmitF128CSEL(MachineInstr *MI,
                                    MachineBasicBlock *MBB) const {
  // We materialise the F128CSEL pseudo-instruction as some control flow and a
  // phi node:

  // OrigBB:
  //     [... previous instrs leading to comparison ...]
  //     b.ne TrueBB
  //     b EndBB
  // TrueBB:
  //     ; Fallthrough
  // EndBB:
  //     Dest = PHI [IfTrue, TrueBB], [IfFalse, OrigBB]

  MachineFunction *MF = MBB->getParent();
  const TargetInstrInfo *TII = Subtarget->getInstrInfo();
  const BasicBlock *LLVM_BB = MBB->getBasicBlock();
  DebugLoc DL = MI->getDebugLoc();
  MachineFunction::iterator It = ++MBB->getIterator();

  unsigned DestReg = MI->getOperand(0).getReg();
  unsigned IfTrueReg = MI->getOperand(1).getReg();
  unsigned IfFalseReg = MI->getOperand(2).getReg();
  unsigned CondCode = MI->getOperand(3).getImm();
  bool NZCVKilled = MI->getOperand(4).isKill();

  MachineBasicBlock *TrueBB = MF->CreateMachineBasicBlock(LLVM_BB);
  MachineBasicBlock *EndBB = MF->CreateMachineBasicBlock(LLVM_BB);
  MF->insert(It, TrueBB);
  MF->insert(It, EndBB);

  // Transfer rest of current basic-block to EndBB
  EndBB->splice(EndBB->begin(), MBB, std::next(MachineBasicBlock::iterator(MI)),
                MBB->end());
  EndBB->transferSuccessorsAndUpdatePHIs(MBB);

  BuildMI(MBB, DL, TII->get(AArch64::Bcc)).addImm(CondCode).addMBB(TrueBB);
  BuildMI(MBB, DL, TII->get(AArch64::B)).addMBB(EndBB);
  MBB->addSuccessor(TrueBB);
  MBB->addSuccessor(EndBB);

  // TrueBB falls through to the end.
  TrueBB->addSuccessor(EndBB);

  if (!NZCVKilled) {
    TrueBB->addLiveIn(AArch64::NZCV);
    EndBB->addLiveIn(AArch64::NZCV);
  }

  BuildMI(*EndBB, EndBB->begin(), DL, TII->get(AArch64::PHI), DestReg)
      .addReg(IfTrueReg)
      .addMBB(TrueBB)
      .addReg(IfFalseReg)
      .addMBB(MBB);

  MI->eraseFromParent();
  return EndBB;
}

MachineBasicBlock *
AArch64TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
                                                 MachineBasicBlock *BB) const {
  switch (MI->getOpcode()) {
  default:
#ifndef NDEBUG
    MI->dump();
#endif
    llvm_unreachable("Unexpected instruction for custom inserter!");

  case AArch64::F128CSEL:
    return EmitF128CSEL(MI, BB);

  case TargetOpcode::STACKMAP:
  case TargetOpcode::PATCHPOINT:
    return emitPatchPoint(MI, BB);
  }
}

//===----------------------------------------------------------------------===//
// AArch64 Lowering private implementation.
//===----------------------------------------------------------------------===//

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

/// changeIntCCToAArch64CC - Convert a DAG integer condition code to an AArch64
/// CC
static AArch64CC::CondCode changeIntCCToAArch64CC(ISD::CondCode CC) {
  switch (CC) {
  default:
    llvm_unreachable("Unknown condition code!");
  case ISD::SETNE:
    return AArch64CC::NE;
  case ISD::SETEQ:
    return AArch64CC::EQ;
  case ISD::SETGT:
    return AArch64CC::GT;
  case ISD::SETGE:
    return AArch64CC::GE;
  case ISD::SETLT:
    return AArch64CC::LT;
  case ISD::SETLE:
    return AArch64CC::LE;
  case ISD::SETUGT:
    return AArch64CC::HI;
  case ISD::SETUGE:
    return AArch64CC::HS;
  case ISD::SETULT:
    return AArch64CC::LO;
  case ISD::SETULE:
    return AArch64CC::LS;
  }
}

/// changeFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64 CC.
static void changeFPCCToAArch64CC(ISD::CondCode CC,
                                  AArch64CC::CondCode &CondCode,
                                  AArch64CC::CondCode &CondCode2) {
  CondCode2 = AArch64CC::AL;
  switch (CC) {
  default:
    llvm_unreachable("Unknown FP condition!");
  case ISD::SETEQ:
  case ISD::SETOEQ:
    CondCode = AArch64CC::EQ;
    break;
  case ISD::SETGT:
  case ISD::SETOGT:
    CondCode = AArch64CC::GT;
    break;
  case ISD::SETGE:
  case ISD::SETOGE:
    CondCode = AArch64CC::GE;
    break;
  case ISD::SETOLT:
    CondCode = AArch64CC::MI;
    break;
  case ISD::SETOLE:
    CondCode = AArch64CC::LS;
    break;
  case ISD::SETONE:
    CondCode = AArch64CC::MI;
    CondCode2 = AArch64CC::GT;
    break;
  case ISD::SETO:
    CondCode = AArch64CC::VC;
    break;
  case ISD::SETUO:
    CondCode = AArch64CC::VS;
    break;
  case ISD::SETUEQ:
    CondCode = AArch64CC::EQ;
    CondCode2 = AArch64CC::VS;
    break;
  case ISD::SETUGT:
    CondCode = AArch64CC::HI;
    break;
  case ISD::SETUGE:
    CondCode = AArch64CC::PL;
    break;
  case ISD::SETLT:
  case ISD::SETULT:
    CondCode = AArch64CC::LT;
    break;
  case ISD::SETLE:
  case ISD::SETULE:
    CondCode = AArch64CC::LE;
    break;
  case ISD::SETNE:
  case ISD::SETUNE:
    CondCode = AArch64CC::NE;
    break;
  }
}

/// changeVectorFPCCToAArch64CC - Convert a DAG fp condition code to an AArch64
/// CC usable with the vector instructions. Fewer operations are available
/// without a real NZCV register, so we have to use less efficient combinations
/// to get the same effect.
static void changeVectorFPCCToAArch64CC(ISD::CondCode CC,
                                        AArch64CC::CondCode &CondCode,
                                        AArch64CC::CondCode &CondCode2,
                                        bool &Invert) {
  Invert = false;
  switch (CC) {
  default:
    // Mostly the scalar mappings work fine.
    changeFPCCToAArch64CC(CC, CondCode, CondCode2);
    break;
  case ISD::SETUO:
    Invert = true; // Fallthrough
  case ISD::SETO:
    CondCode = AArch64CC::MI;
    CondCode2 = AArch64CC::GE;
    break;
  case ISD::SETUEQ:
  case ISD::SETULT:
  case ISD::SETULE:
  case ISD::SETUGT:
  case ISD::SETUGE:
    // All of the compare-mask comparisons are ordered, but we can switch
    // between the two by a double inversion. E.g. ULE == !OGT.
    Invert = true;
    changeFPCCToAArch64CC(getSetCCInverse(CC, false), CondCode, CondCode2);
    break;
  }
}

static bool isLegalArithImmed(uint64_t C) {
  // Matches AArch64DAGToDAGISel::SelectArithImmed().
  return (C >> 12 == 0) || ((C & 0xFFFULL) == 0 && C >> 24 == 0);
}

static SDValue emitComparison(SDValue LHS, SDValue RHS, ISD::CondCode CC,
                              SDLoc dl, SelectionDAG &DAG) {
  EVT VT = LHS.getValueType();

  if (VT.isFloatingPoint())
    return DAG.getNode(AArch64ISD::FCMP, dl, VT, LHS, RHS);

  // The CMP instruction is just an alias for SUBS, and representing it as
  // SUBS means that it's possible to get CSE with subtract operations.
  // A later phase can perform the optimization of setting the destination
  // register to WZR/XZR if it ends up being unused.
  unsigned Opcode = AArch64ISD::SUBS;

  if (RHS.getOpcode() == ISD::SUB && isNullConstant(RHS.getOperand(0)) &&
      (CC == ISD::SETEQ || CC == ISD::SETNE)) {
    // We'd like to combine a (CMP op1, (sub 0, op2) into a CMN instruction on
    // the grounds that "op1 - (-op2) == op1 + op2". However, the C and V flags
    // can be set differently by this operation. It comes down to whether
    // "SInt(~op2)+1 == SInt(~op2+1)" (and the same for UInt). If they are then
    // everything is fine. If not then the optimization is wrong. Thus general
    // comparisons are only valid if op2 != 0.

    // So, finally, the only LLVM-native comparisons that don't mention C and V
    // are SETEQ and SETNE. They're the only ones we can safely use CMN for in
    // the absence of information about op2.
    Opcode = AArch64ISD::ADDS;
    RHS = RHS.getOperand(1);
  } else if (LHS.getOpcode() == ISD::AND && isNullConstant(RHS) &&
             !isUnsignedIntSetCC(CC)) {
    // Similarly, (CMP (and X, Y), 0) can be implemented with a TST
    // (a.k.a. ANDS) except that the flags are only guaranteed to work for one
    // of the signed comparisons.
    Opcode = AArch64ISD::ANDS;
    RHS = LHS.getOperand(1);
    LHS = LHS.getOperand(0);
  }

  return DAG.getNode(Opcode, dl, DAG.getVTList(VT, MVT_CC), LHS, RHS)
      .getValue(1);
}

/// \defgroup AArch64CCMP CMP;CCMP matching
///
/// These functions deal with the formation of CMP;CCMP;... sequences.
/// The CCMP/CCMN/FCCMP/FCCMPE instructions allow the conditional execution of
/// a comparison. They set the NZCV flags to a predefined value if their
/// predicate is false. This allows to express arbitrary conjunctions, for
/// example "cmp 0 (and (setCA (cmp A)) (setCB (cmp B))))"
/// expressed as:
///   cmp A
///   ccmp B, inv(CB), CA
///   check for CB flags
///
/// In general we can create code for arbitrary "... (and (and A B) C)"
/// sequences. We can also implement some "or" expressions, because "(or A B)"
/// is equivalent to "not (and (not A) (not B))" and we can implement some
/// negation operations:
/// We can negate the results of a single comparison by inverting the flags
/// used when the predicate fails and inverting the flags tested in the next
/// instruction; We can also negate the results of the whole previous
/// conditional compare sequence by inverting the flags tested in the next
/// instruction. However there is no way to negate the result of a partial
/// sequence.
///
/// Therefore on encountering an "or" expression we can negate the subtree on
/// one side and have to be able to push the negate to the leafs of the subtree
/// on the other side (see also the comments in code). As complete example:
/// "or (or (setCA (cmp A)) (setCB (cmp B)))
///     (and (setCC (cmp C)) (setCD (cmp D)))"
/// is transformed to
/// "not (and (not (and (setCC (cmp C)) (setCC (cmp D))))
///           (and (not (setCA (cmp A)) (not (setCB (cmp B))))))"
/// and implemented as:
///   cmp C
///   ccmp D, inv(CD), CC
///   ccmp A, CA, inv(CD)
///   ccmp B, CB, inv(CA)
///   check for CB flags
/// A counterexample is "or (and A B) (and C D)" which cannot be implemented
/// by conditional compare sequences.
/// @{

/// Create a conditional comparison; Use CCMP, CCMN or FCCMP as appropriate.
static SDValue emitConditionalComparison(SDValue LHS, SDValue RHS,
                                         ISD::CondCode CC, SDValue CCOp,
                                         SDValue Condition, unsigned NZCV,
                                         SDLoc DL, SelectionDAG &DAG) {
  unsigned Opcode = 0;
  if (LHS.getValueType().isFloatingPoint())
    Opcode = AArch64ISD::FCCMP;
  else if (RHS.getOpcode() == ISD::SUB) {
    SDValue SubOp0 = RHS.getOperand(0);
    if (isNullConstant(SubOp0) && (CC == ISD::SETEQ || CC == ISD::SETNE)) {
        // See emitComparison() on why we can only do this for SETEQ and SETNE.
        Opcode = AArch64ISD::CCMN;
        RHS = RHS.getOperand(1);
      }
  }
  if (Opcode == 0)
    Opcode = AArch64ISD::CCMP;

  SDValue NZCVOp = DAG.getConstant(NZCV, DL, MVT::i32);
  return DAG.getNode(Opcode, DL, MVT_CC, LHS, RHS, NZCVOp, Condition, CCOp);
}

/// Returns true if @p Val is a tree of AND/OR/SETCC operations.
/// CanPushNegate is set to true if we can push a negate operation through
/// the tree in a was that we are left with AND operations and negate operations
/// at the leafs only. i.e. "not (or (or x y) z)" can be changed to
/// "and (and (not x) (not y)) (not z)"; "not (or (and x y) z)" cannot be
/// brought into such a form.
static bool isConjunctionDisjunctionTree(const SDValue Val, bool &CanPushNegate,
                                         unsigned Depth = 0) {
  if (!Val.hasOneUse())
    return false;
  unsigned Opcode = Val->getOpcode();
  if (Opcode == ISD::SETCC) {
    CanPushNegate = true;
    return true;
  }
  // Protect against stack overflow.
  if (Depth > 15)
    return false;
  if (Opcode == ISD::AND || Opcode == ISD::OR) {
    SDValue O0 = Val->getOperand(0);
    SDValue O1 = Val->getOperand(1);
    bool CanPushNegateL;
    if (!isConjunctionDisjunctionTree(O0, CanPushNegateL, Depth+1))
      return false;
    bool CanPushNegateR;
    if (!isConjunctionDisjunctionTree(O1, CanPushNegateR, Depth+1))
      return false;
    // We cannot push a negate through an AND operation (it would become an OR),
    // we can however change a (not (or x y)) to (and (not x) (not y)) if we can
    // push the negate through the x/y subtrees.
    CanPushNegate = (Opcode == ISD::OR) && CanPushNegateL && CanPushNegateR;
    return true;
  }
  return false;
}

/// Emit conjunction or disjunction tree with the CMP/FCMP followed by a chain
/// of CCMP/CFCMP ops. See @ref AArch64CCMP.
/// Tries to transform the given i1 producing node @p Val to a series compare
/// and conditional compare operations. @returns an NZCV flags producing node
/// and sets @p OutCC to the flags that should be tested or returns SDValue() if
/// transformation was not possible.
/// On recursive invocations @p PushNegate may be set to true to have negation
/// effects pushed to the tree leafs; @p Predicate is an NZCV flag predicate
/// for the comparisons in the current subtree; @p Depth limits the search
/// depth to avoid stack overflow.
static SDValue emitConjunctionDisjunctionTree(SelectionDAG &DAG, SDValue Val,
    AArch64CC::CondCode &OutCC, bool PushNegate = false,
    SDValue CCOp = SDValue(), AArch64CC::CondCode Predicate = AArch64CC::AL,
    unsigned Depth = 0) {
  // We're at a tree leaf, produce a conditional comparison operation.
  unsigned Opcode = Val->getOpcode();
  if (Opcode == ISD::SETCC) {
    SDValue LHS = Val->getOperand(0);
    SDValue RHS = Val->getOperand(1);
    ISD::CondCode CC = cast<CondCodeSDNode>(Val->getOperand(2))->get();
    bool isInteger = LHS.getValueType().isInteger();
    if (PushNegate)
      CC = getSetCCInverse(CC, isInteger);
    SDLoc DL(Val);
    // Determine OutCC and handle FP special case.
    if (isInteger) {
      OutCC = changeIntCCToAArch64CC(CC);
    } else {
      assert(LHS.getValueType().isFloatingPoint());
      AArch64CC::CondCode ExtraCC;
      changeFPCCToAArch64CC(CC, OutCC, ExtraCC);
      // Surpisingly some floating point conditions can't be tested with a
      // single condition code. Construct an additional comparison in this case.
      // See comment below on how we deal with OR conditions.
      if (ExtraCC != AArch64CC::AL) {
        SDValue ExtraCmp;
        if (!CCOp.getNode())
          ExtraCmp = emitComparison(LHS, RHS, CC, DL, DAG);
        else {
          SDValue ConditionOp = DAG.getConstant(Predicate, DL, MVT_CC);
          // Note that we want the inverse of ExtraCC, so NZCV is not inversed.
          unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(ExtraCC);
          ExtraCmp = emitConditionalComparison(LHS, RHS, CC, CCOp, ConditionOp,
                                               NZCV, DL, DAG);
        }
        CCOp = ExtraCmp;
        Predicate = AArch64CC::getInvertedCondCode(ExtraCC);
        OutCC = AArch64CC::getInvertedCondCode(OutCC);
      }
    }

    // Produce a normal comparison if we are first in the chain
    if (!CCOp.getNode())
      return emitComparison(LHS, RHS, CC, DL, DAG);
    // Otherwise produce a ccmp.
    SDValue ConditionOp = DAG.getConstant(Predicate, DL, MVT_CC);
    AArch64CC::CondCode InvOutCC = AArch64CC::getInvertedCondCode(OutCC);
    unsigned NZCV = AArch64CC::getNZCVToSatisfyCondCode(InvOutCC);
    return emitConditionalComparison(LHS, RHS, CC, CCOp, ConditionOp, NZCV, DL,
                                     DAG);
  } else if ((Opcode != ISD::AND && Opcode != ISD::OR) || !Val->hasOneUse())
    return SDValue();

  assert((Opcode == ISD::OR || !PushNegate)
         && "Can only push negate through OR operation");

  // Check if both sides can be transformed.
  SDValue LHS = Val->getOperand(0);
  SDValue RHS = Val->getOperand(1);
  bool CanPushNegateL;
  if (!isConjunctionDisjunctionTree(LHS, CanPushNegateL, Depth+1))
    return SDValue();
  bool CanPushNegateR;
  if (!isConjunctionDisjunctionTree(RHS, CanPushNegateR, Depth+1))
    return SDValue();

  // Do we need to negate our operands?
  bool NegateOperands = Opcode == ISD::OR;
  // We can negate the results of all previous operations by inverting the
  // predicate flags giving us a free negation for one side. For the other side
  // we need to be able to push the negation to the leafs of the tree.
  if (NegateOperands) {
    if (!CanPushNegateL && !CanPushNegateR)
      return SDValue();
    // Order the side where we can push the negate through to LHS.
    if (!CanPushNegateL && CanPushNegateR)
      std::swap(LHS, RHS);
  } else {
    bool NeedsNegOutL = LHS->getOpcode() == ISD::OR;
    bool NeedsNegOutR = RHS->getOpcode() == ISD::OR;
    if (NeedsNegOutL && NeedsNegOutR)
      return SDValue();
    // Order the side where we need to negate the output flags to RHS so it
    // gets emitted first.
    if (NeedsNegOutL)
      std::swap(LHS, RHS);
  }

  // Emit RHS. If we want to negate the tree we only need to push a negate
  // through if we are already in a PushNegate case, otherwise we can negate
  // the "flags to test" afterwards.
  AArch64CC::CondCode RHSCC;
  SDValue CmpR = emitConjunctionDisjunctionTree(DAG, RHS, RHSCC, PushNegate,
                                                CCOp, Predicate, Depth+1);
  if (NegateOperands && !PushNegate)
    RHSCC = AArch64CC::getInvertedCondCode(RHSCC);
  // Emit LHS. We must push the negate through if we need to negate it.
  SDValue CmpL = emitConjunctionDisjunctionTree(DAG, LHS, OutCC, NegateOperands,
                                                CmpR, RHSCC, Depth+1);
  // If we transformed an OR to and AND then we have to negate the result
  // (or absorb a PushNegate resulting in a double negation).
  if (Opcode == ISD::OR && !PushNegate)
    OutCC = AArch64CC::getInvertedCondCode(OutCC);
  return CmpL;
}

/// @}

static SDValue getAArch64Cmp(SDValue LHS, SDValue RHS, ISD::CondCode CC,
                             SDValue &AArch64cc, SelectionDAG &DAG, SDLoc dl) {
  if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS.getNode())) {
    EVT VT = RHS.getValueType();
    uint64_t C = RHSC->getZExtValue();
    if (!isLegalArithImmed(C)) {
      // Constant does not fit, try adjusting it by one?
      switch (CC) {
      default:
        break;
      case ISD::SETLT:
      case ISD::SETGE:
        if ((VT == MVT::i32 && C != 0x80000000 &&
             isLegalArithImmed((uint32_t)(C - 1))) ||
            (VT == MVT::i64 && C != 0x80000000ULL &&
             isLegalArithImmed(C - 1ULL))) {
          CC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGT;
          C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
          RHS = DAG.getConstant(C, dl, VT);
        }
        break;
      case ISD::SETULT:
      case ISD::SETUGE:
        if ((VT == MVT::i32 && C != 0 &&
             isLegalArithImmed((uint32_t)(C - 1))) ||
            (VT == MVT::i64 && C != 0ULL && isLegalArithImmed(C - 1ULL))) {
          CC = (CC == ISD::SETULT) ? ISD::SETULE : ISD::SETUGT;
          C = (VT == MVT::i32) ? (uint32_t)(C - 1) : C - 1;
          RHS = DAG.getConstant(C, dl, VT);
        }
        break;
      case ISD::SETLE:
      case ISD::SETGT:
        if ((VT == MVT::i32 && C != INT32_MAX &&
             isLegalArithImmed((uint32_t)(C + 1))) ||
            (VT == MVT::i64 && C != INT64_MAX &&
             isLegalArithImmed(C + 1ULL))) {
          CC = (CC == ISD::SETLE) ? ISD::SETLT : ISD::SETGE;
          C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
          RHS = DAG.getConstant(C, dl, VT);
        }
        break;
      case ISD::SETULE:
      case ISD::SETUGT:
        if ((VT == MVT::i32 && C != UINT32_MAX &&
             isLegalArithImmed((uint32_t)(C + 1))) ||
            (VT == MVT::i64 && C != UINT64_MAX &&
             isLegalArithImmed(C + 1ULL))) {
          CC = (CC == ISD::SETULE) ? ISD::SETULT : ISD::SETUGE;
          C = (VT == MVT::i32) ? (uint32_t)(C + 1) : C + 1;
          RHS = DAG.getConstant(C, dl, VT);
        }
        break;
      }
    }
  }
  SDValue Cmp;
  AArch64CC::CondCode AArch64CC;
  if ((CC == ISD::SETEQ || CC == ISD::SETNE) && isa<ConstantSDNode>(RHS)) {
    const ConstantSDNode *RHSC = cast<ConstantSDNode>(RHS);

    // The imm operand of ADDS is an unsigned immediate, in the range 0 to 4095.
    // For the i8 operand, the largest immediate is 255, so this can be easily
    // encoded in the compare instruction. For the i16 operand, however, the
    // largest immediate cannot be encoded in the compare.
    // Therefore, use a sign extending load and cmn to avoid materializing the
    // -1 constant. For example,
    // movz w1, #65535
    // ldrh w0, [x0, #0]
    // cmp w0, w1
    // >
    // ldrsh w0, [x0, #0]
    // cmn w0, #1
    // Fundamental, we're relying on the property that (zext LHS) == (zext RHS)
    // if and only if (sext LHS) == (sext RHS). The checks are in place to
    // ensure both the LHS and RHS are truly zero extended and to make sure the
    // transformation is profitable.
    if ((RHSC->getZExtValue() >> 16 == 0) && isa<LoadSDNode>(LHS) &&
        cast<LoadSDNode>(LHS)->getExtensionType() == ISD::ZEXTLOAD &&
        cast<LoadSDNode>(LHS)->getMemoryVT() == MVT::i16 &&
        LHS.getNode()->hasNUsesOfValue(1, 0)) {
      int16_t ValueofRHS = cast<ConstantSDNode>(RHS)->getZExtValue();
      if (ValueofRHS < 0 && isLegalArithImmed(-ValueofRHS)) {
        SDValue SExt =
            DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, LHS.getValueType(), LHS,
                        DAG.getValueType(MVT::i16));
        Cmp = emitComparison(SExt, DAG.getConstant(ValueofRHS, dl,
                                                   RHS.getValueType()),
                             CC, dl, DAG);
        AArch64CC = changeIntCCToAArch64CC(CC);
      }
    }

    if (!Cmp && (RHSC->isNullValue() || RHSC->isOne())) {
      if ((Cmp = emitConjunctionDisjunctionTree(DAG, LHS, AArch64CC))) {
        if ((CC == ISD::SETNE) ^ RHSC->isNullValue())
          AArch64CC = AArch64CC::getInvertedCondCode(AArch64CC);
      }
    }
  }

  if (!Cmp) {
    Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
    AArch64CC = changeIntCCToAArch64CC(CC);
  }
  AArch64cc = DAG.getConstant(AArch64CC, dl, MVT_CC);
  return Cmp;
}

static std::pair<SDValue, SDValue>
getAArch64XALUOOp(AArch64CC::CondCode &CC, SDValue Op, SelectionDAG &DAG) {
  assert((Op.getValueType() == MVT::i32 || Op.getValueType() == MVT::i64) &&
         "Unsupported value type");
  SDValue Value, Overflow;
  SDLoc DL(Op);
  SDValue LHS = Op.getOperand(0);
  SDValue RHS = Op.getOperand(1);
  unsigned Opc = 0;
  switch (Op.getOpcode()) {
  default:
    llvm_unreachable("Unknown overflow instruction!");
  case ISD::SADDO:
    Opc = AArch64ISD::ADDS;
    CC = AArch64CC::VS;
    break;
  case ISD::UADDO:
    Opc = AArch64ISD::ADDS;
    CC = AArch64CC::HS;
    break;
  case ISD::SSUBO:
    Opc = AArch64ISD::SUBS;
    CC = AArch64CC::VS;
    break;
  case ISD::USUBO:
    Opc = AArch64ISD::SUBS;
    CC = AArch64CC::LO;
    break;
  // Multiply needs a little bit extra work.
  case ISD::SMULO:
  case ISD::UMULO: {
    CC = AArch64CC::NE;
    bool IsSigned = Op.getOpcode() == ISD::SMULO;
    if (Op.getValueType() == MVT::i32) {
      unsigned ExtendOpc = IsSigned ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
      // For a 32 bit multiply with overflow check we want the instruction
      // selector to generate a widening multiply (SMADDL/UMADDL). For that we
      // need to generate the following pattern:
      // (i64 add 0, (i64 mul (i64 sext|zext i32 %a), (i64 sext|zext i32 %b))
      LHS = DAG.getNode(ExtendOpc, DL, MVT::i64, LHS);
      RHS = DAG.getNode(ExtendOpc, DL, MVT::i64, RHS);
      SDValue Mul = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
      SDValue Add = DAG.getNode(ISD::ADD, DL, MVT::i64, Mul,
                                DAG.getConstant(0, DL, MVT::i64));
      // On AArch64 the upper 32 bits are always zero extended for a 32 bit
      // operation. We need to clear out the upper 32 bits, because we used a
      // widening multiply that wrote all 64 bits. In the end this should be a
      // noop.
      Value = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, Add);
      if (IsSigned) {
        // The signed overflow check requires more than just a simple check for
        // any bit set in the upper 32 bits of the result. These bits could be
        // just the sign bits of a negative number. To perform the overflow
        // check we have to arithmetic shift right the 32nd bit of the result by
        // 31 bits. Then we compare the result to the upper 32 bits.
        SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Add,
                                        DAG.getConstant(32, DL, MVT::i64));
        UpperBits = DAG.getNode(ISD::TRUNCATE, DL, MVT::i32, UpperBits);
        SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i32, Value,
                                        DAG.getConstant(31, DL, MVT::i64));
        // It is important that LowerBits is last, otherwise the arithmetic
        // shift will not be folded into the compare (SUBS).
        SDVTList VTs = DAG.getVTList(MVT::i32, MVT::i32);
        Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
                       .getValue(1);
      } else {
        // The overflow check for unsigned multiply is easy. We only need to
        // check if any of the upper 32 bits are set. This can be done with a
        // CMP (shifted register). For that we need to generate the following
        // pattern:
        // (i64 AArch64ISD::SUBS i64 0, (i64 srl i64 %Mul, i64 32)
        SDValue UpperBits = DAG.getNode(ISD::SRL, DL, MVT::i64, Mul,
                                        DAG.getConstant(32, DL, MVT::i64));
        SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
        Overflow =
            DAG.getNode(AArch64ISD::SUBS, DL, VTs,
                        DAG.getConstant(0, DL, MVT::i64),
                        UpperBits).getValue(1);
      }
      break;
    }
    assert(Op.getValueType() == MVT::i64 && "Expected an i64 value type");
    // For the 64 bit multiply
    Value = DAG.getNode(ISD::MUL, DL, MVT::i64, LHS, RHS);
    if (IsSigned) {
      SDValue UpperBits = DAG.getNode(ISD::MULHS, DL, MVT::i64, LHS, RHS);
      SDValue LowerBits = DAG.getNode(ISD::SRA, DL, MVT::i64, Value,
                                      DAG.getConstant(63, DL, MVT::i64));
      // It is important that LowerBits is last, otherwise the arithmetic
      // shift will not be folded into the compare (SUBS).
      SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
      Overflow = DAG.getNode(AArch64ISD::SUBS, DL, VTs, UpperBits, LowerBits)
                     .getValue(1);
    } else {
      SDValue UpperBits = DAG.getNode(ISD::MULHU, DL, MVT::i64, LHS, RHS);
      SDVTList VTs = DAG.getVTList(MVT::i64, MVT::i32);
      Overflow =
          DAG.getNode(AArch64ISD::SUBS, DL, VTs,
                      DAG.getConstant(0, DL, MVT::i64),
                      UpperBits).getValue(1);
    }
    break;
  }
  } // switch (...)

  if (Opc) {
    SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::i32);

    // Emit the AArch64 operation with overflow check.
    Value = DAG.getNode(Opc, DL, VTs, LHS, RHS);
    Overflow = Value.getValue(1);
  }
  return std::make_pair(Value, Overflow);
}

SDValue AArch64TargetLowering::LowerF128Call(SDValue Op, SelectionDAG &DAG,
                                             RTLIB::Libcall Call) const {
  SmallVector<SDValue, 2> Ops(Op->op_begin(), Op->op_end());
  return makeLibCall(DAG, Call, MVT::f128, Ops, false, SDLoc(Op)).first;
}

static SDValue LowerXOR(SDValue Op, SelectionDAG &DAG) {
  SDValue Sel = Op.getOperand(0);
  SDValue Other = Op.getOperand(1);

  // If neither operand is a SELECT_CC, give up.
  if (Sel.getOpcode() != ISD::SELECT_CC)
    std::swap(Sel, Other);
  if (Sel.getOpcode() != ISD::SELECT_CC)
    return Op;

  // The folding we want to perform is:
  // (xor x, (select_cc a, b, cc, 0, -1) )
  //   -->
  // (csel x, (xor x, -1), cc ...)
  //
  // The latter will get matched to a CSINV instruction.

  ISD::CondCode CC = cast<CondCodeSDNode>(Sel.getOperand(4))->get();
  SDValue LHS = Sel.getOperand(0);
  SDValue RHS = Sel.getOperand(1);
  SDValue TVal = Sel.getOperand(2);
  SDValue FVal = Sel.getOperand(3);
  SDLoc dl(Sel);

  // FIXME: This could be generalized to non-integer comparisons.
  if (LHS.getValueType() != MVT::i32 && LHS.getValueType() != MVT::i64)
    return Op;

  ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
  ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);

  // The values aren't constants, this isn't the pattern we're looking for.
  if (!CFVal || !CTVal)
    return Op;

  // We can commute the SELECT_CC by inverting the condition.  This
  // might be needed to make this fit into a CSINV pattern.
  if (CTVal->isAllOnesValue() && CFVal->isNullValue()) {
    std::swap(TVal, FVal);
    std::swap(CTVal, CFVal);
    CC = ISD::getSetCCInverse(CC, true);
  }

  // If the constants line up, perform the transform!
  if (CTVal->isNullValue() && CFVal->isAllOnesValue()) {
    SDValue CCVal;
    SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);

    FVal = Other;
    TVal = DAG.getNode(ISD::XOR, dl, Other.getValueType(), Other,
                       DAG.getConstant(-1ULL, dl, Other.getValueType()));

    return DAG.getNode(AArch64ISD::CSEL, dl, Sel.getValueType(), FVal, TVal,
                       CCVal, Cmp);
  }

  return Op;
}

static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
  EVT VT = Op.getValueType();

  // Let legalize expand this if it isn't a legal type yet.
  if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
    return SDValue();

  SDVTList VTs = DAG.getVTList(VT, MVT::i32);

  unsigned Opc;
  bool ExtraOp = false;
  switch (Op.getOpcode()) {
  default:
    llvm_unreachable("Invalid code");
  case ISD::ADDC:
    Opc = AArch64ISD::ADDS;
    break;
  case ISD::SUBC:
    Opc = AArch64ISD::SUBS;
    break;
  case ISD::ADDE:
    Opc = AArch64ISD::ADCS;
    ExtraOp = true;
    break;
  case ISD::SUBE:
    Opc = AArch64ISD::SBCS;
    ExtraOp = true;
    break;
  }

  if (!ExtraOp)
    return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1));
  return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0), Op.getOperand(1),
                     Op.getOperand(2));
}

static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
  // Let legalize expand this if it isn't a legal type yet.
  if (!DAG.getTargetLoweringInfo().isTypeLegal(Op.getValueType()))
    return SDValue();

  SDLoc dl(Op);
  AArch64CC::CondCode CC;
  // The actual operation that sets the overflow or carry flag.
  SDValue Value, Overflow;
  std::tie(Value, Overflow) = getAArch64XALUOOp(CC, Op, DAG);

  // We use 0 and 1 as false and true values.
  SDValue TVal = DAG.getConstant(1, dl, MVT::i32);
  SDValue FVal = DAG.getConstant(0, dl, MVT::i32);

  // We use an inverted condition, because the conditional select is inverted
  // too. This will allow it to be selected to a single instruction:
  // CSINC Wd, WZR, WZR, invert(cond).
  SDValue CCVal = DAG.getConstant(getInvertedCondCode(CC), dl, MVT::i32);
  Overflow = DAG.getNode(AArch64ISD::CSEL, dl, MVT::i32, FVal, TVal,
                         CCVal, Overflow);

  SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
  return DAG.getNode(ISD::MERGE_VALUES, dl, VTs, Value, Overflow);
}

// Prefetch operands are:
// 1: Address to prefetch
// 2: bool isWrite
// 3: int locality (0 = no locality ... 3 = extreme locality)
// 4: bool isDataCache
static SDValue LowerPREFETCH(SDValue Op, SelectionDAG &DAG) {
  SDLoc DL(Op);
  unsigned IsWrite = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
  unsigned Locality = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
  unsigned IsData = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();

  bool IsStream = !Locality;
  // When the locality number is set
  if (Locality) {
    // The front-end should have filtered out the out-of-range values
    assert(Locality <= 3 && "Prefetch locality out-of-range");
    // The locality degree is the opposite of the cache speed.
    // Put the number the other way around.
    // The encoding starts at 0 for level 1
    Locality = 3 - Locality;
  }

  // built the mask value encoding the expected behavior.
  unsigned PrfOp = (IsWrite << 4) |     // Load/Store bit
                   (!IsData << 3) |     // IsDataCache bit
                   (Locality << 1) |    // Cache level bits
                   (unsigned)IsStream;  // Stream bit
  return DAG.getNode(AArch64ISD::PREFETCH, DL, MVT::Other, Op.getOperand(0),
                     DAG.getConstant(PrfOp, DL, MVT::i32), Op.getOperand(1));
}

SDValue AArch64TargetLowering::LowerFP_EXTEND(SDValue Op,
                                              SelectionDAG &DAG) const {
  assert(Op.getValueType() == MVT::f128 && "Unexpected lowering");

  RTLIB::Libcall LC;
  LC = RTLIB::getFPEXT(Op.getOperand(0).getValueType(), Op.getValueType());

  return LowerF128Call(Op, DAG, LC);
}

SDValue AArch64TargetLowering::LowerFP_ROUND(SDValue Op,
                                             SelectionDAG &DAG) const {
  if (Op.getOperand(0).getValueType() != MVT::f128) {
    // It's legal except when f128 is involved
    return Op;
  }

  RTLIB::Libcall LC;
  LC = RTLIB::getFPROUND(Op.getOperand(0).getValueType(), Op.getValueType());

  // FP_ROUND node has a second operand indicating whether it is known to be
  // precise. That doesn't take part in the LibCall so we can't directly use
  // LowerF128Call.
  SDValue SrcVal = Op.getOperand(0);
  return makeLibCall(DAG, LC, Op.getValueType(), SrcVal, /*isSigned*/ false,
                     SDLoc(Op)).first;
}

static SDValue LowerVectorFP_TO_INT(SDValue Op, SelectionDAG &DAG) {
  // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
  // Any additional optimization in this function should be recorded
  // in the cost tables.
  EVT InVT = Op.getOperand(0).getValueType();
  EVT VT = Op.getValueType();
  unsigned NumElts = InVT.getVectorNumElements();

  // f16 vectors are promoted to f32 before a conversion.
  if (InVT.getVectorElementType() == MVT::f16) {
    MVT NewVT = MVT::getVectorVT(MVT::f32, NumElts);
    SDLoc dl(Op);
    return DAG.getNode(
        Op.getOpcode(), dl, Op.getValueType(),
        DAG.getNode(ISD::FP_EXTEND, dl, NewVT, Op.getOperand(0)));
  }

  if (VT.getSizeInBits() < InVT.getSizeInBits()) {
    SDLoc dl(Op);
    SDValue Cv =
        DAG.getNode(Op.getOpcode(), dl, InVT.changeVectorElementTypeToInteger(),
                    Op.getOperand(0));
    return DAG.getNode(ISD::TRUNCATE, dl, VT, Cv);
  }

  if (VT.getSizeInBits() > InVT.getSizeInBits()) {
    SDLoc dl(Op);
    MVT ExtVT =
        MVT::getVectorVT(MVT::getFloatingPointVT(VT.getScalarSizeInBits()),
                         VT.getVectorNumElements());
    SDValue Ext = DAG.getNode(ISD::FP_EXTEND, dl, ExtVT, Op.getOperand(0));
    return DAG.getNode(Op.getOpcode(), dl, VT, Ext);
  }

  // Type changing conversions are illegal.
  return Op;
}

SDValue AArch64TargetLowering::LowerFP_TO_INT(SDValue Op,
                                              SelectionDAG &DAG) const {
  if (Op.getOperand(0).getValueType().isVector())
    return LowerVectorFP_TO_INT(Op, DAG);

  // f16 conversions are promoted to f32.
  if (Op.getOperand(0).getValueType() == MVT::f16) {
    SDLoc dl(Op);
    return DAG.getNode(
        Op.getOpcode(), dl, Op.getValueType(),
        DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, Op.getOperand(0)));
  }

  if (Op.getOperand(0).getValueType() != MVT::f128) {
    // It's legal except when f128 is involved
    return Op;
  }

  RTLIB::Libcall LC;
  if (Op.getOpcode() == ISD::FP_TO_SINT)
    LC = RTLIB::getFPTOSINT(Op.getOperand(0).getValueType(), Op.getValueType());
  else
    LC = RTLIB::getFPTOUINT(Op.getOperand(0).getValueType(), Op.getValueType());

  SmallVector<SDValue, 2> Ops(Op->op_begin(), Op->op_end());
  return makeLibCall(DAG, LC, Op.getValueType(), Ops, false, SDLoc(Op)).first;
}

static SDValue LowerVectorINT_TO_FP(SDValue Op, SelectionDAG &DAG) {
  // Warning: We maintain cost tables in AArch64TargetTransformInfo.cpp.
  // Any additional optimization in this function should be recorded
  // in the cost tables.
  EVT VT = Op.getValueType();
  SDLoc dl(Op);
  SDValue In = Op.getOperand(0);
  EVT InVT = In.getValueType();

  if (VT.getSizeInBits() < InVT.getSizeInBits()) {
    MVT CastVT =
        MVT::getVectorVT(MVT::getFloatingPointVT(InVT.getScalarSizeInBits()),
                         InVT.getVectorNumElements());
    In = DAG.getNode(Op.getOpcode(), dl, CastVT, In);
    return DAG.getNode(ISD::FP_ROUND, dl, VT, In, DAG.getIntPtrConstant(0, dl));
  }

  if (VT.getSizeInBits() > InVT.getSizeInBits()) {
    unsigned CastOpc =
        Op.getOpcode() == ISD::SINT_TO_FP ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
    EVT CastVT = VT.changeVectorElementTypeToInteger();
    In = DAG.getNode(CastOpc, dl, CastVT, In);
    return DAG.getNode(Op.getOpcode(), dl, VT, In);
  }

  return Op;
}

SDValue AArch64TargetLowering::LowerINT_TO_FP(SDValue Op,
                                            SelectionDAG &DAG) const {
  if (Op.getValueType().isVector())
    return LowerVectorINT_TO_FP(Op, DAG);

  // f16 conversions are promoted to f32.
  if (Op.getValueType() == MVT::f16) {
    SDLoc dl(Op);
    return DAG.getNode(
        ISD::FP_ROUND, dl, MVT::f16,
        DAG.getNode(Op.getOpcode(), dl, MVT::f32, Op.getOperand(0)),
        DAG.getIntPtrConstant(0, dl));
  }

  // i128 conversions are libcalls.
  if (Op.getOperand(0).getValueType() == MVT::i128)
    return SDValue();

  // Other conversions are legal, unless it's to the completely software-based
  // fp128.
  if (Op.getValueType() != MVT::f128)
    return Op;

  RTLIB::Libcall LC;
  if (Op.getOpcode() == ISD::SINT_TO_FP)
    LC = RTLIB::getSINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());
  else
    LC = RTLIB::getUINTTOFP(Op.getOperand(0).getValueType(), Op.getValueType());

  return LowerF128Call(Op, DAG, LC);
}

SDValue AArch64TargetLowering::LowerFSINCOS(SDValue Op,
                                            SelectionDAG &DAG) const {
  // For iOS, we want to call an alternative entry point: __sincos_stret,
  // which returns the values in two S / D registers.
  SDLoc dl(Op);
  SDValue Arg = Op.getOperand(0);
  EVT ArgVT = Arg.getValueType();
  Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());

  ArgListTy Args;
  ArgListEntry Entry;

  Entry.Node = Arg;
  Entry.Ty = ArgTy;
  Entry.isSExt = false;
  Entry.isZExt = false;
  Args.push_back(Entry);

  const char *LibcallName =
      (ArgVT == MVT::f64) ? "__sincos_stret" : "__sincosf_stret";
  SDValue Callee =
      DAG.getExternalSymbol(LibcallName, getPointerTy(DAG.getDataLayout()));

  StructType *RetTy = StructType::get(ArgTy, ArgTy, nullptr);
  TargetLowering::CallLoweringInfo CLI(DAG);
  CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
    .setCallee(CallingConv::Fast, RetTy, Callee, std::move(Args), 0);

  std::pair<SDValue, SDValue> CallResult = LowerCallTo(CLI);
  return CallResult.first;
}

static SDValue LowerBITCAST(SDValue Op, SelectionDAG &DAG) {
  if (Op.getValueType() != MVT::f16)
    return SDValue();

  assert(Op.getOperand(0).getValueType() == MVT::i16);
  SDLoc DL(Op);

  Op = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, Op.getOperand(0));
  Op = DAG.getNode(ISD::BITCAST, DL, MVT::f32, Op);
  return SDValue(
      DAG.getMachineNode(TargetOpcode::EXTRACT_SUBREG, DL, MVT::f16, Op,
                         DAG.getTargetConstant(AArch64::hsub, DL, MVT::i32)),
      0);
}

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;
  }
}

static SDValue addRequiredExtensionForVectorMULL(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, SDLoc(N), NewVT, N);
}

static bool isExtendedBUILD_VECTOR(SDNode *N, SelectionDAG &DAG,
                                   bool isSigned) {
  EVT VT = N->getValueType(0);

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

  for (const SDValue &Elt : N->op_values()) {
    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;
}

static SDValue skipExtensionForVectorMULL(SDNode *N, SelectionDAG &DAG) {
  if (N->getOpcode() == ISD::SIGN_EXTEND || N->getOpcode() == ISD::ZERO_EXTEND)
    return addRequiredExtensionForVectorMULL(N->getOperand(0), DAG,
                                             N->getOperand(0)->getValueType(0),
                                             N->getValueType(0),
                                             N->getOpcode());

  assert(N->getOpcode() == ISD::BUILD_VECTOR && "expected BUILD_VECTOR");
  EVT VT = N->getValueType(0);
  SDLoc dl(N);
  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), dl, MVT::i32));
  }
  return DAG.getNode(ISD::BUILD_VECTOR, dl,
                     MVT::getVectorVT(TruncVT, NumElts), Ops);
}

static bool isSignExtended(SDNode *N, SelectionDAG &DAG) {
  if (N->getOpcode() == ISD::SIGN_EXTEND)
    return true;
  if (isExtendedBUILD_VECTOR(N, DAG, true))
    return true;
  return false;
}

static bool isZeroExtended(SDNode *N, SelectionDAG &DAG) {
  if (N->getOpcode() == ISD::ZERO_EXTEND)
    return true;
  if (isExtendedBUILD_VECTOR(N, DAG, false))
    return true;
  return false;
}

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 = AArch64ISD::SMULL;
  else {
    bool isN0ZExt = isZeroExtended(N0, DAG);
    bool isN1ZExt = isZeroExtended(N1, DAG);
    if (isN0ZExt && isN1ZExt)
      NewOpc = AArch64ISD::UMULL;
    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 = AArch64ISD::SMULL;
        isMLA = true;
      } else if (isN1ZExt && isAddSubZExt(N0, DAG)) {
        NewOpc =  AArch64ISD::UMULL;
        isMLA = true;
      } else if (isN0ZExt && isAddSubZExt(N1, DAG)) {
        std::swap(N0, N1);
        NewOpc =  AArch64ISD::UMULL;
        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 S/UMULL instruction
  SDLoc DL(Op);
  SDValue Op0;
  SDValue Op1 = skipExtensionForVectorMULL(N1, DAG);
  if (!isMLA) {
    Op0 = skipExtensionForVectorMULL(N0, DAG);
    assert(Op0.getValueType().is64BitVector() &&
           Op1.getValueType().is64BitVector() &&
           "unexpected types for extended operands to VMULL");
    return DAG.getNode(NewOpc, DL, VT, Op0, Op1);
  }
  // Optimizing (zext A + zext B) * C, to (S/UMULL A, C) + (S/UMULL B, C) during
  // isel lowering to take advantage of no-stall back to back s/umul + s/umla.
  // This is true for CPUs with accumulate forwarding such as Cortex-A53/A57
  SDValue N00 = skipExtensionForVectorMULL(N0->getOperand(0).getNode(), DAG);
  SDValue N01 = skipExtensionForVectorMULL(N0->getOperand(1).getNode(), DAG);
  EVT Op1VT = Op1.getValueType();
  return DAG.getNode(N0->getOpcode(), DL, VT,
                     DAG.getNode(NewOpc, DL, VT,
                               DAG.getNode(ISD::BITCAST, DL, Op1VT, N00), Op1),
                     DAG.getNode(NewOpc, DL, VT,
                               DAG.getNode(ISD::BITCAST, DL, Op1VT, N01), Op1));
}

SDValue AArch64TargetLowering::LowerINTRINSIC_WO_CHAIN(SDValue Op,
                                                     SelectionDAG &DAG) const {
  unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
  SDLoc dl(Op);
  switch (IntNo) {
  default: return SDValue();    // Don't custom lower most intrinsics.
  case Intrinsic::aarch64_thread_pointer: {
    EVT PtrVT = getPointerTy(DAG.getDataLayout());
    return DAG.getNode(AArch64ISD::THREAD_POINTER, dl, PtrVT);
  }
  case Intrinsic::aarch64_neon_smax:
    return DAG.getNode(ISD::SMAX, dl, Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::aarch64_neon_umax:
    return DAG.getNode(ISD::UMAX, dl, Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::aarch64_neon_smin:
    return DAG.getNode(ISD::SMIN, dl, Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  case Intrinsic::aarch64_neon_umin:
    return DAG.getNode(ISD::UMIN, dl, Op.getValueType(),
                       Op.getOperand(1), Op.getOperand(2));
  }
}

SDValue AArch64TargetLowering::LowerOperation(SDValue Op,
                                              SelectionDAG &DAG) const {
  switch (Op.getOpcode()) {
  default:
    llvm_unreachable("unimplemented operand");
    return SDValue();
  case ISD::BITCAST:
    return LowerBITCAST(Op, DAG);
  case ISD::GlobalAddress:
    return LowerGlobalAddress(Op, DAG);
  case ISD::GlobalTLSAddress:
    return LowerGlobalTLSAddress(Op, DAG);
  case ISD::SETCC:
    return LowerSETCC(Op, DAG);
  case ISD::BR_CC:
    return LowerBR_CC(Op, DAG);
  case ISD::SELECT:
    return LowerSELECT(Op, DAG);
  case ISD::SELECT_CC:
    return LowerSELECT_CC(Op, DAG);
  case ISD::JumpTable:
    return LowerJumpTable(Op, DAG);
  case ISD::ConstantPool:
    return LowerConstantPool(Op, DAG);
  case ISD::BlockAddress:
    return LowerBlockAddress(Op, DAG);
  case ISD::VASTART:
    return LowerVASTART(Op, DAG);
  case ISD::VACOPY:
    return LowerVACOPY(Op, DAG);
  case ISD::VAARG:
    return LowerVAARG(Op, DAG);
  case ISD::ADDC:
  case ISD::ADDE:
  case ISD::SUBC:
  case ISD::SUBE:
    return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
  case ISD::SADDO:
  case ISD::UADDO:
  case ISD::SSUBO:
  case ISD::USUBO:
  case ISD::SMULO:
  case ISD::UMULO:
    return LowerXALUO(Op, DAG);
  case ISD::FADD:
    return LowerF128Call(Op, DAG, RTLIB::ADD_F128);
  case ISD::FSUB:
    return LowerF128Call(Op, DAG, RTLIB::SUB_F128);
  case ISD::FMUL:
    return LowerF128Call(Op, DAG, RTLIB::MUL_F128);
  case ISD::FDIV:
    return LowerF128Call(Op, DAG, RTLIB::DIV_F128);
  case ISD::FP_ROUND:
    return LowerFP_ROUND(Op, DAG);
  case ISD::FP_EXTEND:
    return LowerFP_EXTEND(Op, DAG);
  case ISD::FRAMEADDR:
    return LowerFRAMEADDR(Op, DAG);
  case ISD::RETURNADDR:
    return LowerRETURNADDR(Op, DAG);
  case ISD::INSERT_VECTOR_ELT:
    return LowerINSERT_VECTOR_ELT(Op, DAG);
  case ISD::EXTRACT_VECTOR_ELT:
    return LowerEXTRACT_VECTOR_ELT(Op, DAG);
  case ISD::BUILD_VECTOR:
    return LowerBUILD_VECTOR(Op, DAG);
  case ISD::VECTOR_SHUFFLE:
    return LowerVECTOR_SHUFFLE(Op, DAG);
  case ISD::EXTRACT_SUBVECTOR:
    return LowerEXTRACT_SUBVECTOR(Op, DAG);
  case ISD::SRA:
  case ISD::SRL:
  case ISD::SHL:
    return LowerVectorSRA_SRL_SHL(Op, DAG);
  case ISD::SHL_PARTS:
    return LowerShiftLeftParts(Op, DAG);
  case ISD::SRL_PARTS:
  case ISD::SRA_PARTS:
    return LowerShiftRightParts(Op, DAG);
  case ISD::CTPOP:
    return LowerCTPOP(Op, DAG);
  case ISD::FCOPYSIGN:
    return LowerFCOPYSIGN(Op, DAG);
  case ISD::AND:
    return LowerVectorAND(Op, DAG);
  case ISD::OR:
    return LowerVectorOR(Op, DAG);
  case ISD::XOR:
    return LowerXOR(Op, DAG);
  case ISD::PREFETCH:
    return LowerPREFETCH(Op, DAG);
  case ISD::SINT_TO_FP:
  case ISD::UINT_TO_FP:
    return LowerINT_TO_FP(Op, DAG);
  case ISD::FP_TO_SINT:
  case ISD::FP_TO_UINT:
    return LowerFP_TO_INT(Op, DAG);
  case ISD::FSINCOS:
    return LowerFSINCOS(Op, DAG);
  case ISD::MUL:
    return LowerMUL(Op, DAG);
  case ISD::INTRINSIC_WO_CHAIN:
    return LowerINTRINSIC_WO_CHAIN(Op, DAG);
  }
}

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

#include "AArch64GenCallingConv.inc"

/// Selects the correct CCAssignFn for a given CallingConvention value.
CCAssignFn *AArch64TargetLowering::CCAssignFnForCall(CallingConv::ID CC,
                                                     bool IsVarArg) const {
  switch (CC) {
  default:
    llvm_unreachable("Unsupported calling convention.");
  case CallingConv::WebKit_JS:
    return CC_AArch64_WebKit_JS;
  case CallingConv::GHC:
    return CC_AArch64_GHC;
  case CallingConv::C:
  case CallingConv::Fast:
    if (!Subtarget->isTargetDarwin())
      return CC_AArch64_AAPCS;
    return IsVarArg ? CC_AArch64_DarwinPCS_VarArg : CC_AArch64_DarwinPCS;
  }
}

SDValue AArch64TargetLowering::LowerFormalArguments(
    SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
    const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
    SmallVectorImpl<SDValue> &InVals) const {
  MachineFunction &MF = DAG.getMachineFunction();
  MachineFrameInfo *MFI = MF.getFrameInfo();

  // Assign locations to all of the incoming arguments.
  SmallVector<CCValAssign, 16> ArgLocs;
  CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), ArgLocs,
                 *DAG.getContext());

  // At this point, Ins[].VT may already be promoted to i32. To correctly
  // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
  // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
  // Since AnalyzeFormalArguments uses Ins[].VT for both ValVT and LocVT, here
  // we use a special version of AnalyzeFormalArguments to pass in ValVT and
  // LocVT.
  unsigned NumArgs = Ins.size();
  Function::const_arg_iterator CurOrigArg = MF.getFunction()->arg_begin();
  unsigned CurArgIdx = 0;
  for (unsigned i = 0; i != NumArgs; ++i) {
    MVT ValVT = Ins[i].VT;
    if (Ins[i].isOrigArg()) {
      std::advance(CurOrigArg, Ins[i].getOrigArgIndex() - CurArgIdx);
      CurArgIdx = Ins[i].getOrigArgIndex();

      // Get type of the original argument.
      EVT ActualVT = getValueType(DAG.getDataLayout(), CurOrigArg->getType(),
                                  /*AllowUnknown*/ true);
      MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : MVT::Other;
      // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
      if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
        ValVT = MVT::i8;
      else if (ActualMVT == MVT::i16)
        ValVT = MVT::i16;
    }
    CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
    bool Res =
        AssignFn(i, ValVT, ValVT, CCValAssign::Full, Ins[i].Flags, CCInfo);
    assert(!Res && "Call operand has unhandled type");
    (void)Res;
  }
  assert(ArgLocs.size() == Ins.size());
  SmallVector<SDValue, 16> ArgValues;
  for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
    CCValAssign &VA = ArgLocs[i];

    if (Ins[i].Flags.isByVal()) {
      // Byval is used for HFAs in the PCS, but the system should work in a
      // non-compliant manner for larger structs.
      EVT PtrVT = getPointerTy(DAG.getDataLayout());
      int Size = Ins[i].Flags.getByValSize();
      unsigned NumRegs = (Size + 7) / 8;

      // FIXME: This works on big-endian for composite byvals, which are the common
      // case. It should also work for fundamental types too.
      unsigned FrameIdx =
        MFI->CreateFixedObject(8 * NumRegs, VA.getLocMemOffset(), false);
      SDValue FrameIdxN = DAG.getFrameIndex(FrameIdx, PtrVT);
      InVals.push_back(FrameIdxN);

      continue;
    }

    if (VA.isRegLoc()) {
      // Arguments stored in registers.
      EVT RegVT = VA.getLocVT();

      SDValue ArgValue;
      const TargetRegisterClass *RC;

      if (RegVT == MVT::i32)
        RC = &AArch64::GPR32RegClass;
      else if (RegVT == MVT::i64)
        RC = &AArch64::GPR64RegClass;
      else if (RegVT == MVT::f16)
        RC = &AArch64::FPR16RegClass;
      else if (RegVT == MVT::f32)
        RC = &AArch64::FPR32RegClass;
      else if (RegVT == MVT::f64 || RegVT.is64BitVector())
        RC = &AArch64::FPR64RegClass;
      else if (RegVT == MVT::f128 || RegVT.is128BitVector())
        RC = &AArch64::FPR128RegClass;
      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, 16 or 32-bit value, it is really passed promoted
      // to 64 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::AExt:
      case CCValAssign::SExt:
      case CCValAssign::ZExt:
        // SelectionDAGBuilder will insert appropriate AssertZExt & AssertSExt
        // nodes after our lowering.
        assert(RegVT == Ins[i].VT && "incorrect register location selected");
        break;
      }

      InVals.push_back(ArgValue);

    } else { // VA.isRegLoc()
      assert(VA.isMemLoc() && "CCValAssign is neither reg nor mem");
      unsigned ArgOffset = VA.getLocMemOffset();
      unsigned ArgSize = VA.getValVT().getSizeInBits() / 8;

      uint32_t BEAlign = 0;
      if (!Subtarget->isLittleEndian() && ArgSize < 8 &&
          !Ins[i].Flags.isInConsecutiveRegs())
        BEAlign = 8 - ArgSize;

      int FI = MFI->CreateFixedObject(ArgSize, ArgOffset + BEAlign, true);

      // Create load nodes to retrieve arguments from the stack.
      SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
      SDValue ArgValue;

      // For NON_EXTLOAD, generic code in getLoad assert(ValVT == MemVT)
      ISD::LoadExtType ExtType = ISD::NON_EXTLOAD;
      MVT MemVT = VA.getValVT();

      switch (VA.getLocInfo()) {
      default:
        break;
      case CCValAssign::BCvt:
        MemVT = VA.getLocVT();
        break;
      case CCValAssign::SExt:
        ExtType = ISD::SEXTLOAD;
        break;
      case CCValAssign::ZExt:
        ExtType = ISD::ZEXTLOAD;
        break;
      case CCValAssign::AExt:
        ExtType = ISD::EXTLOAD;
        break;
      }

      ArgValue = DAG.getExtLoad(
          ExtType, DL, VA.getLocVT(), Chain, FIN,
          MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
          MemVT, false, false, false, 0);

      InVals.push_back(ArgValue);
    }
  }

  // varargs
  if (isVarArg) {
    if (!Subtarget->isTargetDarwin()) {
      // The AAPCS variadic function ABI is identical to the non-variadic
      // one. As a result there may be more arguments in registers and we should
      // save them for future reference.
      saveVarArgRegisters(CCInfo, DAG, DL, Chain);
    }

    AArch64FunctionInfo *AFI = MF.getInfo<AArch64FunctionInfo>();
    // This will point to the next argument passed via stack.
    unsigned StackOffset = CCInfo.getNextStackOffset();
    // We currently pass all varargs at 8-byte alignment.
    StackOffset = ((StackOffset + 7) & ~7);
    AFI->setVarArgsStackIndex(MFI->CreateFixedObject(4, StackOffset, true));
  }

  AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
  unsigned StackArgSize = CCInfo.getNextStackOffset();
  bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
  if (DoesCalleeRestoreStack(CallConv, TailCallOpt)) {
    // This is a non-standard ABI so by fiat I say we're allowed to make full
    // use of the stack area to be popped, which must be aligned to 16 bytes in
    // any case:
    StackArgSize = RoundUpToAlignment(StackArgSize, 16);

    // If we're expected to restore the stack (e.g. fastcc) then we'll be adding
    // a multiple of 16.
    FuncInfo->setArgumentStackToRestore(StackArgSize);

    // This realignment carries over to the available bytes below. Our own
    // callers will guarantee the space is free by giving an aligned value to
    // CALLSEQ_START.
  }
  // Even if we're not expected to free up the space, it's useful to know how
  // much is there while considering tail calls (because we can reuse it).
  FuncInfo->setBytesInStackArgArea(StackArgSize);

  return Chain;
}

void AArch64TargetLowering::saveVarArgRegisters(CCState &CCInfo,
                                                SelectionDAG &DAG, SDLoc DL,
                                                SDValue &Chain) const {
  MachineFunction &MF = DAG.getMachineFunction();
  MachineFrameInfo *MFI = MF.getFrameInfo();
  AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
  auto PtrVT = getPointerTy(DAG.getDataLayout());

  SmallVector<SDValue, 8> MemOps;

  static const MCPhysReg GPRArgRegs[] = { AArch64::X0, AArch64::X1, AArch64::X2,
                                          AArch64::X3, AArch64::X4, AArch64::X5,
                                          AArch64::X6, AArch64::X7 };
  static const unsigned NumGPRArgRegs = array_lengthof(GPRArgRegs);
  unsigned FirstVariadicGPR = CCInfo.getFirstUnallocated(GPRArgRegs);

  unsigned GPRSaveSize = 8 * (NumGPRArgRegs - FirstVariadicGPR);
  int GPRIdx = 0;
  if (GPRSaveSize != 0) {
    GPRIdx = MFI->CreateStackObject(GPRSaveSize, 8, false);

    SDValue FIN = DAG.getFrameIndex(GPRIdx, PtrVT);

    for (unsigned i = FirstVariadicGPR; i < NumGPRArgRegs; ++i) {
      unsigned VReg = MF.addLiveIn(GPRArgRegs[i], &AArch64::GPR64RegClass);
      SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::i64);
      SDValue Store = DAG.getStore(
          Val.getValue(1), DL, Val, FIN,
          MachinePointerInfo::getStack(DAG.getMachineFunction(), i * 8), false,
          false, 0);
      MemOps.push_back(Store);
      FIN =
          DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getConstant(8, DL, PtrVT));
    }
  }
  FuncInfo->setVarArgsGPRIndex(GPRIdx);
  FuncInfo->setVarArgsGPRSize(GPRSaveSize);

  if (Subtarget->hasFPARMv8()) {
    static const MCPhysReg FPRArgRegs[] = {
        AArch64::Q0, AArch64::Q1, AArch64::Q2, AArch64::Q3,
        AArch64::Q4, AArch64::Q5, AArch64::Q6, AArch64::Q7};
    static const unsigned NumFPRArgRegs = array_lengthof(FPRArgRegs);
    unsigned FirstVariadicFPR = CCInfo.getFirstUnallocated(FPRArgRegs);

    unsigned FPRSaveSize = 16 * (NumFPRArgRegs - FirstVariadicFPR);
    int FPRIdx = 0;
    if (FPRSaveSize != 0) {
      FPRIdx = MFI->CreateStackObject(FPRSaveSize, 16, false);

      SDValue FIN = DAG.getFrameIndex(FPRIdx, PtrVT);

      for (unsigned i = FirstVariadicFPR; i < NumFPRArgRegs; ++i) {
        unsigned VReg = MF.addLiveIn(FPRArgRegs[i], &AArch64::FPR128RegClass);
        SDValue Val = DAG.getCopyFromReg(Chain, DL, VReg, MVT::f128);

        SDValue Store = DAG.getStore(
            Val.getValue(1), DL, Val, FIN,
            MachinePointerInfo::getStack(DAG.getMachineFunction(), i * 16),
            false, false, 0);
        MemOps.push_back(Store);
        FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN,
                          DAG.getConstant(16, DL, PtrVT));
      }
    }
    FuncInfo->setVarArgsFPRIndex(FPRIdx);
    FuncInfo->setVarArgsFPRSize(FPRSaveSize);
  }

  if (!MemOps.empty()) {
    Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
  }
}

/// LowerCallResult - Lower the result values of a call into the
/// appropriate copies out of appropriate physical registers.
SDValue AArch64TargetLowering::LowerCallResult(
    SDValue Chain, SDValue InFlag, CallingConv::ID CallConv, bool isVarArg,
    const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc DL, SelectionDAG &DAG,
    SmallVectorImpl<SDValue> &InVals, bool isThisReturn,
    SDValue ThisVal) const {
  CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
                          ? RetCC_AArch64_WebKit_JS
                          : RetCC_AArch64_AAPCS;
  // Assign locations to each value returned by this call.
  SmallVector<CCValAssign, 16> RVLocs;
  CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
                 *DAG.getContext());
  CCInfo.AnalyzeCallResult(Ins, RetCC);

  // 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::i64 &&
             "unexpected return calling convention register assignment");
      InVals.push_back(ThisVal);
      continue;
    }

    SDValue 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;
}

bool AArch64TargetLowering::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 {
  // For CallingConv::C this function knows whether the ABI needs
  // changing. That's not true for other conventions so they will have to opt in
  // manually.
  if (!IsTailCallConvention(CalleeCC) && CalleeCC != CallingConv::C)
    return false;

  const MachineFunction &MF = DAG.getMachineFunction();
  const Function *CallerF = MF.getFunction();
  CallingConv::ID CallerCC = CallerF->getCallingConv();
  bool CCMatch = CallerCC == CalleeCC;

  // Byval parameters hand the function a pointer directly into the stack area
  // we want to reuse during a tail call. Working around this *is* possible (see
  // X86) but less efficient and uglier in LowerCall.
  for (Function::const_arg_iterator i = CallerF->arg_begin(),
                                    e = CallerF->arg_end();
       i != e; ++i)
    if (i->hasByValAttr())
      return false;

  if (getTargetMachine().Options.GuaranteedTailCallOpt) {
    if (IsTailCallConvention(CalleeCC) && CCMatch)
      return true;
    return false;
  }

  // Externally-defined functions with weak linkage should not be
  // tail-called on AArch64 when the OS does not support dynamic
  // pre-emption of symbols, as the AAELF spec requires normal calls
  // to undefined weak functions to be replaced with a NOP or jump to the
  // next instruction. The behaviour of branch instructions in this
  // situation (as used for tail calls) is implementation-defined, so we
  // cannot rely on the linker replacing the tail call with a return.
  if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
    const GlobalValue *GV = G->getGlobal();
    const Triple &TT = getTargetMachine().getTargetTriple();
    if (GV->hasExternalWeakLinkage() &&
        (!TT.isOSWindows() || TT.isOSBinFormatELF() || TT.isOSBinFormatMachO()))
      return false;
  }

  // Now we search for cases where we can use a tail call without changing the
  // ABI. Sibcall is used in some places (particularly gcc) to refer to this
  // concept.

  // I want anyone implementing a new calling convention to think long and hard
  // about this assert.
  assert((!isVarArg || CalleeCC == CallingConv::C) &&
         "Unexpected variadic calling convention");

  if (isVarArg && !Outs.empty()) {
    // At least two cases here: if caller is fastcc then we can't have any
    // memory arguments (we'd be expected to clean up the stack afterwards). If
    // caller is C then we could potentially use its argument area.

    // FIXME: for now we take the most conservative of these in both cases:
    // disallow all variadic memory operands.
    SmallVector<CCValAssign, 16> ArgLocs;
    CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
                   *DAG.getContext());

    CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, true));
    for (const CCValAssign &ArgLoc : ArgLocs)
      if (!ArgLoc.isRegLoc())
        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;
    CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
                    *DAG.getContext());
    CCInfo1.AnalyzeCallResult(Ins, CCAssignFnForCall(CalleeCC, isVarArg));

    SmallVector<CCValAssign, 16> RVLocs2;
    CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
                    *DAG.getContext());
    CCInfo2.AnalyzeCallResult(Ins, CCAssignFnForCall(CallerCC, 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;
      }
    }
  }

  // Nothing more to check if the callee is taking no arguments
  if (Outs.empty())
    return true;

  SmallVector<CCValAssign, 16> ArgLocs;
  CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
                 *DAG.getContext());

  CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForCall(CalleeCC, isVarArg));

  const AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();

  // If the stack arguments for this call would fit into our own save area then
  // the call can be made tail.
  return CCInfo.getNextStackOffset() <= FuncInfo->getBytesInStackArgArea();
}

SDValue AArch64TargetLowering::addTokenForArgument(SDValue Chain,
                                                   SelectionDAG &DAG,
                                                   MachineFrameInfo *MFI,
                                                   int ClobberedFI) const {
  SmallVector<SDValue, 8> ArgChains;
  int64_t FirstByte = MFI->getObjectOffset(ClobberedFI);
  int64_t LastByte = FirstByte + MFI->getObjectSize(ClobberedFI) - 1;

  // Include the original chain at the beginning of the list. When this is
  // used by target LowerCall hooks, this helps legalize find the
  // CALLSEQ_BEGIN node.
  ArgChains.push_back(Chain);

  // Add a chain value for each stack argument corresponding
  for (SDNode::use_iterator U = DAG.getEntryNode().getNode()->use_begin(),
                            UE = DAG.getEntryNode().getNode()->use_end();
       U != UE; ++U)
    if (LoadSDNode *L = dyn_cast<LoadSDNode>(*U))
      if (FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(L->getBasePtr()))
        if (FI->getIndex() < 0) {
          int64_t InFirstByte = MFI->getObjectOffset(FI->getIndex());
          int64_t InLastByte = InFirstByte;
          InLastByte += MFI->getObjectSize(FI->getIndex()) - 1;

          if ((InFirstByte <= FirstByte && FirstByte <= InLastByte) ||
              (FirstByte <= InFirstByte && InFirstByte <= LastByte))
            ArgChains.push_back(SDValue(L, 1));
        }

  // Build a tokenfactor for all the chains.
  return DAG.getNode(ISD::TokenFactor, SDLoc(Chain), MVT::Other, ArgChains);
}

bool AArch64TargetLowering::DoesCalleeRestoreStack(CallingConv::ID CallCC,
                                                   bool TailCallOpt) const {
  return CallCC == CallingConv::Fast && TailCallOpt;
}

bool AArch64TargetLowering::IsTailCallConvention(CallingConv::ID CallCC) const {
  return CallCC == CallingConv::Fast;
}

/// LowerCall - Lower a call to a callseq_start + CALL + callseq_end chain,
/// and add input and output parameter nodes.
SDValue
AArch64TargetLowering::LowerCall(CallLoweringInfo &CLI,
                                 SmallVectorImpl<SDValue> &InVals) const {
  SelectionDAG &DAG = CLI.DAG;
  SDLoc &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 IsVarArg = CLI.IsVarArg;

  MachineFunction &MF = DAG.getMachineFunction();
  bool IsStructRet = (Outs.empty()) ? false : Outs[0].Flags.isSRet();
  bool IsThisReturn = false;

  AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
  bool TailCallOpt = MF.getTarget().Options.GuaranteedTailCallOpt;
  bool IsSibCall = 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);
    if (!IsTailCall && CLI.CS && CLI.CS->isMustTailCall())
      report_fatal_error("failed to perform tail call elimination on a call "
                         "site marked musttail");

    // A sibling call is one where we're under the usual C ABI and not planning
    // to change that but can still do a tail call:
    if (!TailCallOpt && IsTailCall)
      IsSibCall = true;

    if (IsTailCall)
      ++NumTailCalls;
  }

  // Analyze operands of the call, assigning locations to each operand.
  SmallVector<CCValAssign, 16> ArgLocs;
  CCState CCInfo(CallConv, IsVarArg, DAG.getMachineFunction(), ArgLocs,
                 *DAG.getContext());

  if (IsVarArg) {
    // Handle fixed and variable vector arguments differently.
    // Variable vector arguments always go into memory.
    unsigned NumArgs = Outs.size();

    for (unsigned i = 0; i != NumArgs; ++i) {
      MVT ArgVT = Outs[i].VT;
      ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
      CCAssignFn *AssignFn = CCAssignFnForCall(CallConv,
                                               /*IsVarArg=*/ !Outs[i].IsFixed);
      bool Res = AssignFn(i, ArgVT, ArgVT, CCValAssign::Full, ArgFlags, CCInfo);
      assert(!Res && "Call operand has unhandled type");
      (void)Res;
    }
  } else {
    // At this point, Outs[].VT may already be promoted to i32. To correctly
    // handle passing i8 as i8 instead of i32 on stack, we pass in both i32 and
    // i8 to CC_AArch64_AAPCS with i32 being ValVT and i8 being LocVT.
    // Since AnalyzeCallOperands uses Ins[].VT for both ValVT and LocVT, here
    // we use a special version of AnalyzeCallOperands to pass in ValVT and
    // LocVT.
    unsigned NumArgs = Outs.size();
    for (unsigned i = 0; i != NumArgs; ++i) {
      MVT ValVT = Outs[i].VT;
      // Get type of the original argument.
      EVT ActualVT = getValueType(DAG.getDataLayout(),
                                  CLI.getArgs()[Outs[i].OrigArgIndex].Ty,
                                  /*AllowUnknown*/ true);
      MVT ActualMVT = ActualVT.isSimple() ? ActualVT.getSimpleVT() : ValVT;
      ISD::ArgFlagsTy ArgFlags = Outs[i].Flags;
      // If ActualMVT is i1/i8/i16, we should set LocVT to i8/i8/i16.
      if (ActualMVT == MVT::i1 || ActualMVT == MVT::i8)
        ValVT = MVT::i8;
      else if (ActualMVT == MVT::i16)
        ValVT = MVT::i16;

      CCAssignFn *AssignFn = CCAssignFnForCall(CallConv, /*IsVarArg=*/false);
      bool Res = AssignFn(i, ValVT, ValVT, CCValAssign::Full, ArgFlags, CCInfo);
      assert(!Res && "Call operand has unhandled type");
      (void)Res;
    }
  }

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

  if (IsSibCall) {
    // Since we're not changing the ABI to make this a tail call, the memory
    // operands are already available in the caller's incoming argument space.
    NumBytes = 0;
  }

  // FPDiff is the byte offset of the call's argument area from the callee's.
  // Stores to callee stack arguments will be placed in FixedStackSlots offset
  // by this amount for a tail call. In a sibling call it must be 0 because the
  // caller will deallocate the entire stack and the callee still expects its
  // arguments to begin at SP+0. Completely unused for non-tail calls.
  int FPDiff = 0;

  if (IsTailCall && !IsSibCall) {
    unsigned NumReusableBytes = FuncInfo->getBytesInStackArgArea();

    // Since callee will pop argument stack as a tail call, we must keep the
    // popped size 16-byte aligned.
    NumBytes = RoundUpToAlignment(NumBytes, 16);

    // FPDiff will be negative if this tail call requires more space than we
    // would automatically have in our incoming argument space. Positive if we
    // can actually shrink the stack.
    FPDiff = NumReusableBytes - NumBytes;

    // The stack pointer must be 16-byte aligned at all times it's used for a
    // memory operation, which in practice means at *all* times and in
    // particular across call boundaries. Therefore our own arguments started at
    // a 16-byte aligned SP and the delta applied for the tail call should
    // satisfy the same constraint.
    assert(FPDiff % 16 == 0 && "unaligned stack on tail call");
  }

  // 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, DL,
                                                              true),
                                 DL);

  SDValue StackPtr = DAG.getCopyFromReg(Chain, DL, AArch64::SP,
                                        getPointerTy(DAG.getDataLayout()));

  SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
  SmallVector<SDValue, 8> MemOpChains;
  auto PtrVT = getPointerTy(DAG.getDataLayout());

  // Walk the register/memloc assignments, inserting copies/loads.
  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;

    // 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:
      if (Outs[realArgIdx].ArgVT == MVT::i1) {
        // AAPCS requires i1 to be zero-extended to 8-bits by the caller.
        Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
        Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i8, Arg);
      }
      Arg = DAG.getNode(ISD::ANY_EXTEND, DL, VA.getLocVT(), Arg);
      break;
    case CCValAssign::BCvt:
      Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
      break;
    case CCValAssign::FPExt:
      Arg = DAG.getNode(ISD::FP_EXTEND, DL, VA.getLocVT(), Arg);
      break;
    }

    if (VA.isRegLoc()) {
      if (realArgIdx == 0 && Flags.isReturned() && Outs[0].VT == MVT::i64) {
        assert(VA.getLocVT() == MVT::i64 &&
               "unexpected calling convention register assignment");
        assert(!Ins.empty() && Ins[0].VT == MVT::i64 &&
               "unexpected use of 'returned'");
        IsThisReturn = true;
      }
      RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
    } else {
      assert(VA.isMemLoc());

      SDValue DstAddr;
      MachinePointerInfo DstInfo;

      // FIXME: This works on big-endian for composite byvals, which are the
      // common case. It should also work for fundamental types too.
      uint32_t BEAlign = 0;
      unsigned OpSize = Flags.isByVal() ? Flags.getByValSize() * 8
                                        : VA.getValVT().getSizeInBits();
      OpSize = (OpSize + 7) / 8;
      if (!Subtarget->isLittleEndian() && !Flags.isByVal() &&
          !Flags.isInConsecutiveRegs()) {
        if (OpSize < 8)
          BEAlign = 8 - OpSize;
      }
      unsigned LocMemOffset = VA.getLocMemOffset();
      int32_t Offset = LocMemOffset + BEAlign;
      SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL);
      PtrOff = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff);

      if (IsTailCall) {
        Offset = Offset + FPDiff;
        int FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);

        DstAddr = DAG.getFrameIndex(FI, PtrVT);
        DstInfo =
            MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI);

        // Make sure any stack arguments overlapping with where we're storing
        // are loaded before this eventual operation. Otherwise they'll be
        // clobbered.
        Chain = addTokenForArgument(Chain, DAG, MF.getFrameInfo(), FI);
      } else {
        SDValue PtrOff = DAG.getIntPtrConstant(Offset, DL);

        DstAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackPtr, PtrOff);
        DstInfo = MachinePointerInfo::getStack(DAG.getMachineFunction(),
                                               LocMemOffset);
      }

      if (Outs[i].Flags.isByVal()) {
        SDValue SizeNode =
            DAG.getConstant(Outs[i].Flags.getByValSize(), DL, MVT::i64);
        SDValue Cpy = DAG.getMemcpy(
            Chain, DL, DstAddr, Arg, SizeNode, Outs[i].Flags.getByValAlign(),
            /*isVol = */ false, /*AlwaysInline = */ false,
            /*isTailCall = */ false,
            DstInfo, MachinePointerInfo());

        MemOpChains.push_back(Cpy);
      } else {
        // Since we pass i1/i8/i16 as i1/i8/i16 on stack and Arg is already
        // promoted to a legal register type i32, we should truncate Arg back to
        // i1/i8/i16.
        if (VA.getValVT() == MVT::i1 || VA.getValVT() == MVT::i8 ||
            VA.getValVT() == MVT::i16)
          Arg = DAG.getNode(ISD::TRUNCATE, DL, VA.getValVT(), Arg);

        SDValue Store =
            DAG.getStore(Chain, DL, Arg, DstAddr, DstInfo, false, false, 0);
        MemOpChains.push_back(Store);
      }
    }
  }

  if (!MemOpChains.empty())
    Chain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOpChains);

  // 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;
  for (auto &RegToPass : RegsToPass) {
    Chain = DAG.getCopyToReg(Chain, DL, RegToPass.first,
                             RegToPass.second, InFlag);
    InFlag = Chain.getValue(1);
  }

  // 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.
  if (getTargetMachine().getCodeModel() == CodeModel::Large &&
      Subtarget->isTargetMachO()) {
    if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
      const GlobalValue *GV = G->getGlobal();
      bool InternalLinkage = GV->hasInternalLinkage();
      if (InternalLinkage)
        Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 0);
      else {
        Callee =
            DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_GOT);
        Callee = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, Callee);
      }
    } else if (ExternalSymbolSDNode *S =
                   dyn_cast<ExternalSymbolSDNode>(Callee)) {
      const char *Sym = S->getSymbol();
      Callee = DAG.getTargetExternalSymbol(Sym, PtrVT, AArch64II::MO_GOT);
      Callee = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, Callee);
    }
  } else if (GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee)) {
    const GlobalValue *GV = G->getGlobal();
    Callee = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, 0);
  } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
    const char *Sym = S->getSymbol();
    Callee = DAG.getTargetExternalSymbol(Sym, PtrVT, 0);
  }

  // We don't usually want to end the call-sequence here because we would tidy
  // the frame up *after* the call, however in the ABI-changing tail-call case
  // we've carefully laid out the parameters so that when sp is reset they'll be
  // in the correct location.
  if (IsTailCall && !IsSibCall) {
    Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, DL, true),
                               DAG.getIntPtrConstant(0, DL, true), InFlag, DL);
    InFlag = Chain.getValue(1);
  }

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

  if (IsTailCall) {
    // Each tail call may have to adjust the stack by a different amount, so
    // this information must travel along with the operation for eventual
    // consumption by emitEpilogue.
    Ops.push_back(DAG.getTargetConstant(FPDiff, DL, MVT::i32));
  }

  // Add argument registers to the end of the list so that they are known live
  // into the call.
  for (auto &RegToPass : RegsToPass)
    Ops.push_back(DAG.getRegister(RegToPass.first,
                                  RegToPass.second.getValueType()));

  // Add a register mask operand representing the call-preserved registers.
  const uint32_t *Mask;
  const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
  if (IsThisReturn) {
    // For 'this' returns, use the X0-preserving mask if applicable
    Mask = TRI->getThisReturnPreservedMask(MF, CallConv);
    if (!Mask) {
      IsThisReturn = false;
      Mask = TRI->getCallPreservedMask(MF, CallConv);
    }
  } else
    Mask = TRI->getCallPreservedMask(MF, 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 we're doing a tall call, use a TC_RETURN here rather than an
  // actual call instruction.
  if (IsTailCall) {
    MF.getFrameInfo()->setHasTailCall();
    return DAG.getNode(AArch64ISD::TC_RETURN, DL, NodeTys, Ops);
  }

  // Returns a chain and a flag for retval copy to use.
  Chain = DAG.getNode(AArch64ISD::CALL, DL, NodeTys, Ops);
  InFlag = Chain.getValue(1);

  uint64_t CalleePopBytes = DoesCalleeRestoreStack(CallConv, TailCallOpt)
                                ? RoundUpToAlignment(NumBytes, 16)
                                : 0;

  Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, DL, true),
                             DAG.getIntPtrConstant(CalleePopBytes, DL, true),
                             InFlag, DL);
  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());
}

bool AArch64TargetLowering::CanLowerReturn(
    CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
    const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
  CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
                          ? RetCC_AArch64_WebKit_JS
                          : RetCC_AArch64_AAPCS;
  SmallVector<CCValAssign, 16> RVLocs;
  CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
  return CCInfo.CheckReturn(Outs, RetCC);
}

SDValue
AArch64TargetLowering::LowerReturn(SDValue Chain, CallingConv::ID CallConv,
                                   bool isVarArg,
                                   const SmallVectorImpl<ISD::OutputArg> &Outs,
                                   const SmallVectorImpl<SDValue> &OutVals,
                                   SDLoc DL, SelectionDAG &DAG) const {
  CCAssignFn *RetCC = CallConv == CallingConv::WebKit_JS
                          ? RetCC_AArch64_WebKit_JS
                          : RetCC_AArch64_AAPCS;
  SmallVector<CCValAssign, 16> RVLocs;
  CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
                 *DAG.getContext());
  CCInfo.AnalyzeReturn(Outs, RetCC);

  // Copy the result values into the output registers.
  SDValue Flag;
  SmallVector<SDValue, 4> RetOps(1, Chain);
  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:
      if (Outs[i].ArgVT == MVT::i1) {
        // AAPCS requires i1 to be zero-extended to i8 by the producer of the
        // value. This is strictly redundant on Darwin (which uses "zeroext
        // i1"), but will be optimised out before ISel.
        Arg = DAG.getNode(ISD::TRUNCATE, DL, MVT::i1, Arg);
        Arg = DAG.getNode(ISD::ZERO_EXTEND, DL, VA.getLocVT(), Arg);
      }
      break;
    case CCValAssign::BCvt:
      Arg = DAG.getNode(ISD::BITCAST, DL, VA.getLocVT(), Arg);
      break;
    }

    Chain = DAG.getCopyToReg(Chain, DL, VA.getLocReg(), Arg, Flag);
    Flag = Chain.getValue(1);
    RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
  }
  const AArch64RegisterInfo *TRI = Subtarget->getRegisterInfo();
  const MCPhysReg *I =
      TRI->getCalleeSavedRegsViaCopy(&DAG.getMachineFunction());
  if (I) {
    for (; *I; ++I) {
      if (AArch64::GPR64RegClass.contains(*I))
        RetOps.push_back(DAG.getRegister(*I, MVT::i64));
      else if (AArch64::FPR64RegClass.contains(*I))
        RetOps.push_back(DAG.getRegister(*I, MVT::getFloatingPointVT(64)));
      else
        llvm_unreachable("Unexpected register class in CSRsViaCopy!");
    }
  }

  RetOps[0] = Chain; // Update chain.

  // Add the flag if we have it.
  if (Flag.getNode())
    RetOps.push_back(Flag);

  return DAG.getNode(AArch64ISD::RET_FLAG, DL, MVT::Other, RetOps);
}

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

SDValue AArch64TargetLowering::LowerGlobalAddress(SDValue Op,
                                                  SelectionDAG &DAG) const {
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
  SDLoc DL(Op);
  const GlobalAddressSDNode *GN = cast<GlobalAddressSDNode>(Op);
  const GlobalValue *GV = GN->getGlobal();
  unsigned char OpFlags =
      Subtarget->ClassifyGlobalReference(GV, getTargetMachine());

  assert(cast<GlobalAddressSDNode>(Op)->getOffset() == 0 &&
         "unexpected offset in global node");

  // This also catched the large code model case for Darwin.
  if ((OpFlags & AArch64II::MO_GOT) != 0) {
    SDValue GotAddr = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, OpFlags);
    // FIXME: Once remat is capable of dealing with instructions with register
    // operands, expand this into two nodes instead of using a wrapper node.
    return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
  }

  if ((OpFlags & AArch64II::MO_CONSTPOOL) != 0) {
    assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
           "use of MO_CONSTPOOL only supported on small model");
    SDValue Hi = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, AArch64II::MO_PAGE);
    SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
    unsigned char LoFlags = AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
    SDValue Lo = DAG.getTargetConstantPool(GV, PtrVT, 0, 0, LoFlags);
    SDValue PoolAddr = DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
    SDValue GlobalAddr = DAG.getLoad(
        PtrVT, DL, DAG.getEntryNode(), PoolAddr,
        MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
        /*isVolatile=*/false,
        /*isNonTemporal=*/true,
        /*isInvariant=*/true, 8);
    if (GN->getOffset() != 0)
      return DAG.getNode(ISD::ADD, DL, PtrVT, GlobalAddr,
                         DAG.getConstant(GN->getOffset(), DL, PtrVT));
    return GlobalAddr;
  }

  if (getTargetMachine().getCodeModel() == CodeModel::Large) {
    const unsigned char MO_NC = AArch64II::MO_NC;
    return DAG.getNode(
        AArch64ISD::WrapperLarge, DL, PtrVT,
        DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G3),
        DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
        DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
        DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
  } else {
    // Use ADRP/ADD or ADRP/LDR for everything else: the small model on ELF and
    // the only correct model on Darwin.
    SDValue Hi = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0,
                                            OpFlags | AArch64II::MO_PAGE);
    unsigned char LoFlags = OpFlags | AArch64II::MO_PAGEOFF | AArch64II::MO_NC;
    SDValue Lo = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, LoFlags);

    SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
    return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
  }
}

/// \brief Convert a TLS address reference into the correct sequence of loads
/// and calls to compute the variable's address (for Darwin, currently) and
/// return an SDValue containing the final node.

/// Darwin only has one TLS scheme which must be capable of dealing with the
/// fully general situation, in the worst case. This means:
///     + "extern __thread" declaration.
///     + Defined in a possibly unknown dynamic library.
///
/// The general system is that each __thread variable has a [3 x i64] descriptor
/// which contains information used by the runtime to calculate the address. The
/// only part of this the compiler needs to know about is the first xword, which
/// contains a function pointer that must be called with the address of the
/// entire descriptor in "x0".
///
/// Since this descriptor may be in a different unit, in general even the
/// descriptor must be accessed via an indirect load. The "ideal" code sequence
/// is:
///     adrp x0, _var@TLVPPAGE
///     ldr x0, [x0, _var@TLVPPAGEOFF]   ; x0 now contains address of descriptor
///     ldr x1, [x0]                     ; x1 contains 1st entry of descriptor,
///                                      ; the function pointer
///     blr x1                           ; Uses descriptor address in x0
///     ; Address of _var is now in x0.
///
/// If the address of _var's descriptor *is* known to the linker, then it can
/// change the first "ldr" instruction to an appropriate "add x0, x0, #imm" for
/// a slight efficiency gain.
SDValue
AArch64TargetLowering::LowerDarwinGlobalTLSAddress(SDValue Op,
                                                   SelectionDAG &DAG) const {
  assert(Subtarget->isTargetDarwin() && "TLS only supported on Darwin");

  SDLoc DL(Op);
  MVT PtrVT = getPointerTy(DAG.getDataLayout());
  const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();

  SDValue TLVPAddr =
      DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
  SDValue DescAddr = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TLVPAddr);

  // The first entry in the descriptor is a function pointer that we must call
  // to obtain the address of the variable.
  SDValue Chain = DAG.getEntryNode();
  SDValue FuncTLVGet =
      DAG.getLoad(MVT::i64, DL, Chain, DescAddr,
                  MachinePointerInfo::getGOT(DAG.getMachineFunction()), false,
                  true, true, 8);
  Chain = FuncTLVGet.getValue(1);

  MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
  MFI->setAdjustsStack(true);

  // TLS calls preserve all registers except those that absolutely must be
  // trashed: X0 (it takes an argument), LR (it's a call) and NZCV (let's not be
  // silly).
  const uint32_t *Mask =
      Subtarget->getRegisterInfo()->getTLSCallPreservedMask();

  // Finally, we can make the call. This is just a degenerate version of a
  // normal AArch64 call node: x0 takes the address of the descriptor, and
  // returns the address of the variable in this thread.
  Chain = DAG.getCopyToReg(Chain, DL, AArch64::X0, DescAddr, SDValue());
  Chain =
      DAG.getNode(AArch64ISD::CALL, DL, DAG.getVTList(MVT::Other, MVT::Glue),
                  Chain, FuncTLVGet, DAG.getRegister(AArch64::X0, MVT::i64),
                  DAG.getRegisterMask(Mask), Chain.getValue(1));
  return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Chain.getValue(1));
}

/// When accessing thread-local variables under either the general-dynamic or
/// local-dynamic system, we make a "TLS-descriptor" call. The variable will
/// have a descriptor, accessible via a PC-relative ADRP, and whose first entry
/// is a function pointer to carry out the resolution.
///
/// The sequence is:
///    adrp  x0, :tlsdesc:var
///    ldr   x1, [x0, #:tlsdesc_lo12:var]
///    add   x0, x0, #:tlsdesc_lo12:var
///    .tlsdesccall var
///    blr   x1
///    (TPIDR_EL0 offset now in x0)
///
///  The above sequence must be produced unscheduled, to enable the linker to
///  optimize/relax this sequence.
///  Therefore, a pseudo-instruction (TLSDESC_CALLSEQ) is used to represent the
///  above sequence, and expanded really late in the compilation flow, to ensure
///  the sequence is produced as per above.
SDValue AArch64TargetLowering::LowerELFTLSDescCallSeq(SDValue SymAddr, SDLoc DL,
                                                      SelectionDAG &DAG) const {
  EVT PtrVT = getPointerTy(DAG.getDataLayout());

  SDValue Chain = DAG.getEntryNode();
  SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);

  SmallVector<SDValue, 2> Ops;
  Ops.push_back(Chain);
  Ops.push_back(SymAddr);

  Chain = DAG.getNode(AArch64ISD::TLSDESC_CALLSEQ, DL, NodeTys, Ops);
  SDValue Glue = Chain.getValue(1);

  return DAG.getCopyFromReg(Chain, DL, AArch64::X0, PtrVT, Glue);
}

SDValue
AArch64TargetLowering::LowerELFGlobalTLSAddress(SDValue Op,
                                                SelectionDAG &DAG) const {
  assert(Subtarget->isTargetELF() && "This function expects an ELF target");
  assert(getTargetMachine().getCodeModel() == CodeModel::Small &&
         "ELF TLS only supported in small memory model");
  // Different choices can be made for the maximum size of the TLS area for a
  // module. For the small address model, the default TLS size is 16MiB and the
  // maximum TLS size is 4GiB.
  // FIXME: add -mtls-size command line option and make it control the 16MiB
  // vs. 4GiB code sequence generation.
  const GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);

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

  if (DAG.getTarget().Options.EmulatedTLS)
    return LowerToTLSEmulatedModel(GA, DAG);

  if (!EnableAArch64ELFLocalDynamicTLSGeneration) {
    if (Model == TLSModel::LocalDynamic)
      Model = TLSModel::GeneralDynamic;
  }

  SDValue TPOff;
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
  SDLoc DL(Op);
  const GlobalValue *GV = GA->getGlobal();

  SDValue ThreadBase = DAG.getNode(AArch64ISD::THREAD_POINTER, DL, PtrVT);

  if (Model == TLSModel::LocalExec) {
    SDValue HiVar = DAG.getTargetGlobalAddress(
        GV, DL, PtrVT, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
    SDValue LoVar = DAG.getTargetGlobalAddress(
        GV, DL, PtrVT, 0,
        AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);

    SDValue TPWithOff_lo =
        SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, ThreadBase,
                                   HiVar,
                                   DAG.getTargetConstant(0, DL, MVT::i32)),
                0);
    SDValue TPWithOff =
        SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPWithOff_lo,
                                   LoVar,
                                   DAG.getTargetConstant(0, DL, MVT::i32)),
                0);
    return TPWithOff;
  } else if (Model == TLSModel::InitialExec) {
    TPOff = DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);
    TPOff = DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, TPOff);
  } else if (Model == TLSModel::LocalDynamic) {
    // Local-dynamic accesses proceed in two phases. A general-dynamic TLS
    // descriptor call against the special symbol _TLS_MODULE_BASE_ to calculate
    // the beginning of the module's TLS region, followed by a DTPREL offset
    // calculation.

    // These accesses will need deduplicating if there's more than one.
    AArch64FunctionInfo *MFI =
        DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();
    MFI->incNumLocalDynamicTLSAccesses();

    // The call needs a relocation too for linker relaxation. It doesn't make
    // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
    // the address.
    SDValue SymAddr = DAG.getTargetExternalSymbol("_TLS_MODULE_BASE_", PtrVT,
                                                  AArch64II::MO_TLS);

    // Now we can calculate the offset from TPIDR_EL0 to this module's
    // thread-local area.
    TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG);

    // Now use :dtprel_whatever: operations to calculate this variable's offset
    // in its thread-storage area.
    SDValue HiVar = DAG.getTargetGlobalAddress(
        GV, DL, MVT::i64, 0, AArch64II::MO_TLS | AArch64II::MO_HI12);
    SDValue LoVar = DAG.getTargetGlobalAddress(
        GV, DL, MVT::i64, 0,
        AArch64II::MO_TLS | AArch64II::MO_PAGEOFF | AArch64II::MO_NC);

    TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, HiVar,
                                       DAG.getTargetConstant(0, DL, MVT::i32)),
                    0);
    TPOff = SDValue(DAG.getMachineNode(AArch64::ADDXri, DL, PtrVT, TPOff, LoVar,
                                       DAG.getTargetConstant(0, DL, MVT::i32)),
                    0);
  } else if (Model == TLSModel::GeneralDynamic) {
    // The call needs a relocation too for linker relaxation. It doesn't make
    // sense to call it MO_PAGE or MO_PAGEOFF though so we need another copy of
    // the address.
    SDValue SymAddr =
        DAG.getTargetGlobalAddress(GV, DL, PtrVT, 0, AArch64II::MO_TLS);

    // Finally we can make a call to calculate the offset from tpidr_el0.
    TPOff = LowerELFTLSDescCallSeq(SymAddr, DL, DAG);
  } else
    llvm_unreachable("Unsupported ELF TLS access model");

  return DAG.getNode(ISD::ADD, DL, PtrVT, ThreadBase, TPOff);
}

SDValue AArch64TargetLowering::LowerGlobalTLSAddress(SDValue Op,
                                                     SelectionDAG &DAG) const {
  if (Subtarget->isTargetDarwin())
    return LowerDarwinGlobalTLSAddress(Op, DAG);
  else if (Subtarget->isTargetELF())
    return LowerELFGlobalTLSAddress(Op, DAG);

  llvm_unreachable("Unexpected platform trying to use TLS");
}
SDValue AArch64TargetLowering::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);
  SDLoc dl(Op);

  // Handle f128 first, since lowering it will result in comparing the return
  // value of a libcall against zero, which is just what the rest of LowerBR_CC
  // is expecting to deal with.
  if (LHS.getValueType() == MVT::f128) {
    softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);

    // If softenSetCCOperands returned a scalar, we need to compare the result
    // against zero to select between true and false values.
    if (!RHS.getNode()) {
      RHS = DAG.getConstant(0, dl, LHS.getValueType());
      CC = ISD::SETNE;
    }
  }

  // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a branch
  // instruction.
  unsigned Opc = LHS.getOpcode();
  if (LHS.getResNo() == 1 && isOneConstant(RHS) &&
      (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
       Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
    assert((CC == ISD::SETEQ || CC == ISD::SETNE) &&
           "Unexpected condition code.");
    // Only lower legal XALUO ops.
    if (!DAG.getTargetLoweringInfo().isTypeLegal(LHS->getValueType(0)))
      return SDValue();

    // The actual operation with overflow check.
    AArch64CC::CondCode OFCC;
    SDValue Value, Overflow;
    std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, LHS.getValue(0), DAG);

    if (CC == ISD::SETNE)
      OFCC = getInvertedCondCode(OFCC);
    SDValue CCVal = DAG.getConstant(OFCC, dl, MVT::i32);

    return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
                       Overflow);
  }

  if (LHS.getValueType().isInteger()) {
    assert((LHS.getValueType() == RHS.getValueType()) &&
           (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));

    // If the RHS of the comparison is zero, we can potentially fold this
    // to a specialized branch.
    const ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS);
    if (RHSC && RHSC->getZExtValue() == 0) {
      if (CC == ISD::SETEQ) {
        // See if we can use a TBZ to fold in an AND as well.
        // TBZ has a smaller branch displacement than CBZ.  If the offset is
        // out of bounds, a late MI-layer pass rewrites branches.
        // 403.gcc is an example that hits this case.
        if (LHS.getOpcode() == ISD::AND &&
            isa<ConstantSDNode>(LHS.getOperand(1)) &&
            isPowerOf2_64(LHS.getConstantOperandVal(1))) {
          SDValue Test = LHS.getOperand(0);
          uint64_t Mask = LHS.getConstantOperandVal(1);
          return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, Test,
                             DAG.getConstant(Log2_64(Mask), dl, MVT::i64),
                             Dest);
        }

        return DAG.getNode(AArch64ISD::CBZ, dl, MVT::Other, Chain, LHS, Dest);
      } else if (CC == ISD::SETNE) {
        // See if we can use a TBZ to fold in an AND as well.
        // TBZ has a smaller branch displacement than CBZ.  If the offset is
        // out of bounds, a late MI-layer pass rewrites branches.
        // 403.gcc is an example that hits this case.
        if (LHS.getOpcode() == ISD::AND &&
            isa<ConstantSDNode>(LHS.getOperand(1)) &&
            isPowerOf2_64(LHS.getConstantOperandVal(1))) {
          SDValue Test = LHS.getOperand(0);
          uint64_t Mask = LHS.getConstantOperandVal(1);
          return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, Test,
                             DAG.getConstant(Log2_64(Mask), dl, MVT::i64),
                             Dest);
        }

        return DAG.getNode(AArch64ISD::CBNZ, dl, MVT::Other, Chain, LHS, Dest);
      } else if (CC == ISD::SETLT && LHS.getOpcode() != ISD::AND) {
        // Don't combine AND since emitComparison converts the AND to an ANDS
        // (a.k.a. TST) and the test in the test bit and branch instruction
        // becomes redundant.  This would also increase register pressure.
        uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
        return DAG.getNode(AArch64ISD::TBNZ, dl, MVT::Other, Chain, LHS,
                           DAG.getConstant(Mask, dl, MVT::i64), Dest);
      }
    }
    if (RHSC && RHSC->getSExtValue() == -1 && CC == ISD::SETGT &&
        LHS.getOpcode() != ISD::AND) {
      // Don't combine AND since emitComparison converts the AND to an ANDS
      // (a.k.a. TST) and the test in the test bit and branch instruction
      // becomes redundant.  This would also increase register pressure.
      uint64_t Mask = LHS.getValueType().getSizeInBits() - 1;
      return DAG.getNode(AArch64ISD::TBZ, dl, MVT::Other, Chain, LHS,
                         DAG.getConstant(Mask, dl, MVT::i64), Dest);
    }

    SDValue CCVal;
    SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);
    return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CCVal,
                       Cmp);
  }

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

  // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
  // clean.  Some of them require two branches to implement.
  SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);
  AArch64CC::CondCode CC1, CC2;
  changeFPCCToAArch64CC(CC, CC1, CC2);
  SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
  SDValue BR1 =
      DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, Chain, Dest, CC1Val, Cmp);
  if (CC2 != AArch64CC::AL) {
    SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
    return DAG.getNode(AArch64ISD::BRCOND, dl, MVT::Other, BR1, Dest, CC2Val,
                       Cmp);
  }

  return BR1;
}

SDValue AArch64TargetLowering::LowerFCOPYSIGN(SDValue Op,
                                              SelectionDAG &DAG) const {
  EVT VT = Op.getValueType();
  SDLoc DL(Op);

  SDValue In1 = Op.getOperand(0);
  SDValue In2 = Op.getOperand(1);
  EVT SrcVT = In2.getValueType();

  if (SrcVT.bitsLT(VT))
    In2 = DAG.getNode(ISD::FP_EXTEND, DL, VT, In2);
  else if (SrcVT.bitsGT(VT))
    In2 = DAG.getNode(ISD::FP_ROUND, DL, VT, In2, DAG.getIntPtrConstant(0, DL));

  EVT VecVT;
  EVT EltVT;
  uint64_t EltMask;
  SDValue VecVal1, VecVal2;
  if (VT == MVT::f32 || VT == MVT::v2f32 || VT == MVT::v4f32) {
    EltVT = MVT::i32;
    VecVT = (VT == MVT::v2f32 ? MVT::v2i32 : MVT::v4i32);
    EltMask = 0x80000000ULL;

    if (!VT.isVector()) {
      VecVal1 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
                                          DAG.getUNDEF(VecVT), In1);
      VecVal2 = DAG.getTargetInsertSubreg(AArch64::ssub, DL, VecVT,
                                          DAG.getUNDEF(VecVT), In2);
    } else {
      VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
      VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
    }
  } else if (VT == MVT::f64 || VT == MVT::v2f64) {
    EltVT = MVT::i64;
    VecVT = MVT::v2i64;

    // We want to materialize a mask with the high bit set, but the AdvSIMD
    // immediate moves cannot materialize that in a single instruction for
    // 64-bit elements. Instead, materialize zero and then negate it.
    EltMask = 0;

    if (!VT.isVector()) {
      VecVal1 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
                                          DAG.getUNDEF(VecVT), In1);
      VecVal2 = DAG.getTargetInsertSubreg(AArch64::dsub, DL, VecVT,
                                          DAG.getUNDEF(VecVT), In2);
    } else {
      VecVal1 = DAG.getNode(ISD::BITCAST, DL, VecVT, In1);
      VecVal2 = DAG.getNode(ISD::BITCAST, DL, VecVT, In2);
    }
  } else {
    llvm_unreachable("Invalid type for copysign!");
  }

  SDValue BuildVec = DAG.getConstant(EltMask, DL, VecVT);

  // If we couldn't materialize the mask above, then the mask vector will be
  // the zero vector, and we need to negate it here.
  if (VT == MVT::f64 || VT == MVT::v2f64) {
    BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2f64, BuildVec);
    BuildVec = DAG.getNode(ISD::FNEG, DL, MVT::v2f64, BuildVec);
    BuildVec = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, BuildVec);
  }

  SDValue Sel =
      DAG.getNode(AArch64ISD::BIT, DL, VecVT, VecVal1, VecVal2, BuildVec);

  if (VT == MVT::f32)
    return DAG.getTargetExtractSubreg(AArch64::ssub, DL, VT, Sel);
  else if (VT == MVT::f64)
    return DAG.getTargetExtractSubreg(AArch64::dsub, DL, VT, Sel);
  else
    return DAG.getNode(ISD::BITCAST, DL, VT, Sel);
}

SDValue AArch64TargetLowering::LowerCTPOP(SDValue Op, SelectionDAG &DAG) const {
  if (DAG.getMachineFunction().getFunction()->hasFnAttribute(
          Attribute::NoImplicitFloat))
    return SDValue();

  if (!Subtarget->hasNEON())
    return SDValue();

  // While there is no integer popcount instruction, it can
  // be more efficiently lowered to the following sequence that uses
  // AdvSIMD registers/instructions as long as the copies to/from
  // the AdvSIMD registers are cheap.
  //  FMOV    D0, X0        // copy 64-bit int to vector, high bits zero'd
  //  CNT     V0.8B, V0.8B  // 8xbyte pop-counts
  //  ADDV    B0, V0.8B     // sum 8xbyte pop-counts
  //  UMOV    X0, V0.B[0]   // copy byte result back to integer reg
  SDValue Val = Op.getOperand(0);
  SDLoc DL(Op);
  EVT VT = Op.getValueType();

  if (VT == MVT::i32)
    Val = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Val);
  Val = DAG.getNode(ISD::BITCAST, DL, MVT::v8i8, Val);

  SDValue CtPop = DAG.getNode(ISD::CTPOP, DL, MVT::v8i8, Val);
  SDValue UaddLV = DAG.getNode(
      ISD::INTRINSIC_WO_CHAIN, DL, MVT::i32,
      DAG.getConstant(Intrinsic::aarch64_neon_uaddlv, DL, MVT::i32), CtPop);

  if (VT == MVT::i64)
    UaddLV = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, UaddLV);
  return UaddLV;
}

SDValue AArch64TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {

  if (Op.getValueType().isVector())
    return LowerVSETCC(Op, DAG);

  SDValue LHS = Op.getOperand(0);
  SDValue RHS = Op.getOperand(1);
  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
  SDLoc dl(Op);

  // We chose ZeroOrOneBooleanContents, so use zero and one.
  EVT VT = Op.getValueType();
  SDValue TVal = DAG.getConstant(1, dl, VT);
  SDValue FVal = DAG.getConstant(0, dl, VT);

  // Handle f128 first, since one possible outcome is a normal integer
  // comparison which gets picked up by the next if statement.
  if (LHS.getValueType() == MVT::f128) {
    softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);

    // If softenSetCCOperands returned a scalar, use it.
    if (!RHS.getNode()) {
      assert(LHS.getValueType() == Op.getValueType() &&
             "Unexpected setcc expansion!");
      return LHS;
    }
  }

  if (LHS.getValueType().isInteger()) {
    SDValue CCVal;
    SDValue Cmp =
        getAArch64Cmp(LHS, RHS, ISD::getSetCCInverse(CC, true), CCVal, DAG, dl);

    // Note that we inverted the condition above, so we reverse the order of
    // the true and false operands here.  This will allow the setcc to be
    // matched to a single CSINC instruction.
    return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CCVal, Cmp);
  }

  // Now we know we're dealing with FP values.
  assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);

  // If that fails, we'll need to perform an FCMP + CSEL sequence.  Go ahead
  // and do the comparison.
  SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);

  AArch64CC::CondCode CC1, CC2;
  changeFPCCToAArch64CC(CC, CC1, CC2);
  if (CC2 == AArch64CC::AL) {
    changeFPCCToAArch64CC(ISD::getSetCCInverse(CC, false), CC1, CC2);
    SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);

    // Note that we inverted the condition above, so we reverse the order of
    // the true and false operands here.  This will allow the setcc to be
    // matched to a single CSINC instruction.
    return DAG.getNode(AArch64ISD::CSEL, dl, VT, FVal, TVal, CC1Val, Cmp);
  } else {
    // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't
    // totally clean.  Some of them require two CSELs to implement.  As is in
    // this case, we emit the first CSEL and then emit a second using the output
    // of the first as the RHS.  We're effectively OR'ing the two CC's together.

    // FIXME: It would be nice if we could match the two CSELs to two CSINCs.
    SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
    SDValue CS1 =
        DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);

    SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
    return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
  }
}

SDValue AArch64TargetLowering::LowerSELECT_CC(ISD::CondCode CC, SDValue LHS,
                                              SDValue RHS, SDValue TVal,
                                              SDValue FVal, SDLoc dl,
                                              SelectionDAG &DAG) const {
  // Handle f128 first, because it will result in a comparison of some RTLIB
  // call result against zero.
  if (LHS.getValueType() == MVT::f128) {
    softenSetCCOperands(DAG, MVT::f128, LHS, RHS, CC, dl);

    // If softenSetCCOperands returned a scalar, we need to compare the result
    // against zero to select between true and false values.
    if (!RHS.getNode()) {
      RHS = DAG.getConstant(0, dl, LHS.getValueType());
      CC = ISD::SETNE;
    }
  }

  // Also handle f16, for which we need to do a f32 comparison.
  if (LHS.getValueType() == MVT::f16) {
    LHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, LHS);
    RHS = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f32, RHS);
  }

  // Next, handle integers.
  if (LHS.getValueType().isInteger()) {
    assert((LHS.getValueType() == RHS.getValueType()) &&
           (LHS.getValueType() == MVT::i32 || LHS.getValueType() == MVT::i64));

    unsigned Opcode = AArch64ISD::CSEL;

    // If both the TVal and the FVal are constants, see if we can swap them in
    // order to for a CSINV or CSINC out of them.
    ConstantSDNode *CFVal = dyn_cast<ConstantSDNode>(FVal);
    ConstantSDNode *CTVal = dyn_cast<ConstantSDNode>(TVal);

    if (CTVal && CFVal && CTVal->isAllOnesValue() && CFVal->isNullValue()) {
      std::swap(TVal, FVal);
      std::swap(CTVal, CFVal);
      CC = ISD::getSetCCInverse(CC, true);
    } else if (CTVal && CFVal && CTVal->isOne() && CFVal->isNullValue()) {
      std::swap(TVal, FVal);
      std::swap(CTVal, CFVal);
      CC = ISD::getSetCCInverse(CC, true);
    } else if (TVal.getOpcode() == ISD::XOR) {
      // If TVal is a NOT we want to swap TVal and FVal so that we can match
      // with a CSINV rather than a CSEL.
      if (isAllOnesConstant(TVal.getOperand(1))) {
        std::swap(TVal, FVal);
        std::swap(CTVal, CFVal);
        CC = ISD::getSetCCInverse(CC, true);
      }
    } else if (TVal.getOpcode() == ISD::SUB) {
      // If TVal is a negation (SUB from 0) we want to swap TVal and FVal so
      // that we can match with a CSNEG rather than a CSEL.
      if (isNullConstant(TVal.getOperand(0))) {
        std::swap(TVal, FVal);
        std::swap(CTVal, CFVal);
        CC = ISD::getSetCCInverse(CC, true);
      }
    } else if (CTVal && CFVal) {
      const int64_t TrueVal = CTVal->getSExtValue();
      const int64_t FalseVal = CFVal->getSExtValue();
      bool Swap = false;

      // If both TVal and FVal are constants, see if FVal is the
      // inverse/negation/increment of TVal and generate a CSINV/CSNEG/CSINC
      // instead of a CSEL in that case.
      if (TrueVal == ~FalseVal) {
        Opcode = AArch64ISD::CSINV;
      } else if (TrueVal == -FalseVal) {
        Opcode = AArch64ISD::CSNEG;
      } else if (TVal.getValueType() == MVT::i32) {
        // If our operands are only 32-bit wide, make sure we use 32-bit
        // arithmetic for the check whether we can use CSINC. This ensures that
        // the addition in the check will wrap around properly in case there is
        // an overflow (which would not be the case if we do the check with
        // 64-bit arithmetic).
        const uint32_t TrueVal32 = CTVal->getZExtValue();
        const uint32_t FalseVal32 = CFVal->getZExtValue();

        if ((TrueVal32 == FalseVal32 + 1) || (TrueVal32 + 1 == FalseVal32)) {
          Opcode = AArch64ISD::CSINC;

          if (TrueVal32 > FalseVal32) {
            Swap = true;
          }
        }
        // 64-bit check whether we can use CSINC.
      } else if ((TrueVal == FalseVal + 1) || (TrueVal + 1 == FalseVal)) {
        Opcode = AArch64ISD::CSINC;

        if (TrueVal > FalseVal) {
          Swap = true;
        }
      }

      // Swap TVal and FVal if necessary.
      if (Swap) {
        std::swap(TVal, FVal);
        std::swap(CTVal, CFVal);
        CC = ISD::getSetCCInverse(CC, true);
      }

      if (Opcode != AArch64ISD::CSEL) {
        // Drop FVal since we can get its value by simply inverting/negating
        // TVal.
        FVal = TVal;
      }
    }

    SDValue CCVal;
    SDValue Cmp = getAArch64Cmp(LHS, RHS, CC, CCVal, DAG, dl);

    EVT VT = TVal.getValueType();
    return DAG.getNode(Opcode, dl, VT, TVal, FVal, CCVal, Cmp);
  }

  // Now we know we're dealing with FP values.
  assert(LHS.getValueType() == MVT::f32 || LHS.getValueType() == MVT::f64);
  assert(LHS.getValueType() == RHS.getValueType());
  EVT VT = TVal.getValueType();
  SDValue Cmp = emitComparison(LHS, RHS, CC, dl, DAG);

  // Unfortunately, the mapping of LLVM FP CC's onto AArch64 CC's isn't totally
  // clean.  Some of them require two CSELs to implement.
  AArch64CC::CondCode CC1, CC2;
  changeFPCCToAArch64CC(CC, CC1, CC2);
  SDValue CC1Val = DAG.getConstant(CC1, dl, MVT::i32);
  SDValue CS1 = DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, FVal, CC1Val, Cmp);

  // If we need a second CSEL, emit it, using the output of the first as the
  // RHS.  We're effectively OR'ing the two CC's together.
  if (CC2 != AArch64CC::AL) {
    SDValue CC2Val = DAG.getConstant(CC2, dl, MVT::i32);
    return DAG.getNode(AArch64ISD::CSEL, dl, VT, TVal, CS1, CC2Val, Cmp);
  }

  // Otherwise, return the output of the first CSEL.
  return CS1;
}

SDValue AArch64TargetLowering::LowerSELECT_CC(SDValue Op,
                                              SelectionDAG &DAG) const {
  ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(4))->get();
  SDValue LHS = Op.getOperand(0);
  SDValue RHS = Op.getOperand(1);
  SDValue TVal = Op.getOperand(2);
  SDValue FVal = Op.getOperand(3);
  SDLoc DL(Op);
  return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, DL, DAG);
}

SDValue AArch64TargetLowering::LowerSELECT(SDValue Op,
                                           SelectionDAG &DAG) const {
  SDValue CCVal = Op->getOperand(0);
  SDValue TVal = Op->getOperand(1);
  SDValue FVal = Op->getOperand(2);
  SDLoc DL(Op);

  unsigned Opc = CCVal.getOpcode();
  // Optimize {s|u}{add|sub|mul}.with.overflow feeding into a select
  // instruction.
  if (CCVal.getResNo() == 1 &&
      (Opc == ISD::SADDO || Opc == ISD::UADDO || Opc == ISD::SSUBO ||
       Opc == ISD::USUBO || Opc == ISD::SMULO || Opc == ISD::UMULO)) {
    // Only lower legal XALUO ops.
    if (!DAG.getTargetLoweringInfo().isTypeLegal(CCVal->getValueType(0)))
      return SDValue();

    AArch64CC::CondCode OFCC;
    SDValue Value, Overflow;
    std::tie(Value, Overflow) = getAArch64XALUOOp(OFCC, CCVal.getValue(0), DAG);
    SDValue CCVal = DAG.getConstant(OFCC, DL, MVT::i32);

    return DAG.getNode(AArch64ISD::CSEL, DL, Op.getValueType(), TVal, FVal,
                       CCVal, Overflow);
  }

  // Lower it the same way as we would lower a SELECT_CC node.
  ISD::CondCode CC;
  SDValue LHS, RHS;
  if (CCVal.getOpcode() == ISD::SETCC) {
    LHS = CCVal.getOperand(0);
    RHS = CCVal.getOperand(1);
    CC = cast<CondCodeSDNode>(CCVal->getOperand(2))->get();
  } else {
    LHS = CCVal;
    RHS = DAG.getConstant(0, DL, CCVal.getValueType());
    CC = ISD::SETNE;
  }
  return LowerSELECT_CC(CC, LHS, RHS, TVal, FVal, DL, DAG);
}

SDValue AArch64TargetLowering::LowerJumpTable(SDValue Op,
                                              SelectionDAG &DAG) const {
  // Jump table entries as PC relative offsets. No additional tweaking
  // is necessary here. Just get the address of the jump table.
  JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
  SDLoc DL(Op);

  if (getTargetMachine().getCodeModel() == CodeModel::Large &&
      !Subtarget->isTargetMachO()) {
    const unsigned char MO_NC = AArch64II::MO_NC;
    return DAG.getNode(
        AArch64ISD::WrapperLarge, DL, PtrVT,
        DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G3),
        DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G2 | MO_NC),
        DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_G1 | MO_NC),
        DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
                               AArch64II::MO_G0 | MO_NC));
  }

  SDValue Hi =
      DAG.getTargetJumpTable(JT->getIndex(), PtrVT, AArch64II::MO_PAGE);
  SDValue Lo = DAG.getTargetJumpTable(JT->getIndex(), PtrVT,
                                      AArch64II::MO_PAGEOFF | AArch64II::MO_NC);
  SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
  return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
}

SDValue AArch64TargetLowering::LowerConstantPool(SDValue Op,
                                                 SelectionDAG &DAG) const {
  ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
  SDLoc DL(Op);

  if (getTargetMachine().getCodeModel() == CodeModel::Large) {
    // Use the GOT for the large code model on iOS.
    if (Subtarget->isTargetMachO()) {
      SDValue GotAddr = DAG.getTargetConstantPool(
          CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
          AArch64II::MO_GOT);
      return DAG.getNode(AArch64ISD::LOADgot, DL, PtrVT, GotAddr);
    }

    const unsigned char MO_NC = AArch64II::MO_NC;
    return DAG.getNode(
        AArch64ISD::WrapperLarge, DL, PtrVT,
        DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
                                  CP->getOffset(), AArch64II::MO_G3),
        DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
                                  CP->getOffset(), AArch64II::MO_G2 | MO_NC),
        DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
                                  CP->getOffset(), AArch64II::MO_G1 | MO_NC),
        DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
                                  CP->getOffset(), AArch64II::MO_G0 | MO_NC));
  } else {
    // Use ADRP/ADD or ADRP/LDR for everything else: the small memory model on
    // ELF, the only valid one on Darwin.
    SDValue Hi =
        DAG.getTargetConstantPool(CP->getConstVal(), PtrVT, CP->getAlignment(),
                                  CP->getOffset(), AArch64II::MO_PAGE);
    SDValue Lo = DAG.getTargetConstantPool(
        CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(),
        AArch64II::MO_PAGEOFF | AArch64II::MO_NC);

    SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
    return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
  }
}

SDValue AArch64TargetLowering::LowerBlockAddress(SDValue Op,
                                               SelectionDAG &DAG) const {
  const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
  EVT PtrVT = getPointerTy(DAG.getDataLayout());
  SDLoc DL(Op);
  if (getTargetMachine().getCodeModel() == CodeModel::Large &&
      !Subtarget->isTargetMachO()) {
    const unsigned char MO_NC = AArch64II::MO_NC;
    return DAG.getNode(
        AArch64ISD::WrapperLarge, DL, PtrVT,
        DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G3),
        DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G2 | MO_NC),
        DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G1 | MO_NC),
        DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_G0 | MO_NC));
  } else {
    SDValue Hi = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGE);
    SDValue Lo = DAG.getTargetBlockAddress(BA, PtrVT, 0, AArch64II::MO_PAGEOFF |
                                                             AArch64II::MO_NC);
    SDValue ADRP = DAG.getNode(AArch64ISD::ADRP, DL, PtrVT, Hi);
    return DAG.getNode(AArch64ISD::ADDlow, DL, PtrVT, ADRP, Lo);
  }
}

SDValue AArch64TargetLowering::LowerDarwin_VASTART(SDValue Op,
                                                 SelectionDAG &DAG) const {
  AArch64FunctionInfo *FuncInfo =
      DAG.getMachineFunction().getInfo<AArch64FunctionInfo>();

  SDLoc DL(Op);
  SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(),
                                 getPointerTy(DAG.getDataLayout()));
  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 AArch64TargetLowering::LowerAAPCS_VASTART(SDValue Op,
                                                SelectionDAG &DAG) const {
  // The layout of the va_list struct is specified in the AArch64 Procedure Call
  // Standard, section B.3.
  MachineFunction &MF = DAG.getMachineFunction();
  AArch64FunctionInfo *FuncInfo = MF.getInfo<AArch64FunctionInfo>();
  auto PtrVT = getPointerTy(DAG.getDataLayout());
  SDLoc DL(Op);

  SDValue Chain = Op.getOperand(0);
  SDValue VAList = Op.getOperand(1);
  const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
  SmallVector<SDValue, 4> MemOps;

  // void *__stack at offset 0
  SDValue Stack = DAG.getFrameIndex(FuncInfo->getVarArgsStackIndex(), PtrVT);
  MemOps.push_back(DAG.getStore(Chain, DL, Stack, VAList,
                                MachinePointerInfo(SV), false, false, 8));

  // void *__gr_top at offset 8
  int GPRSize = FuncInfo->getVarArgsGPRSize();
  if (GPRSize > 0) {
    SDValue GRTop, GRTopAddr;

    GRTopAddr =
        DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(8, DL, PtrVT));

    GRTop = DAG.getFrameIndex(FuncInfo->getVarArgsGPRIndex(), PtrVT);
    GRTop = DAG.getNode(ISD::ADD, DL, PtrVT, GRTop,
                        DAG.getConstant(GPRSize, DL, PtrVT));

    MemOps.push_back(DAG.getStore(Chain, DL, GRTop, GRTopAddr,
                                  MachinePointerInfo(SV, 8), false, false, 8));
  }

  // void *__vr_top at offset 16
  int FPRSize = FuncInfo->getVarArgsFPRSize();
  if (FPRSize > 0) {
    SDValue VRTop, VRTopAddr;
    VRTopAddr = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
                            DAG.getConstant(16, DL, PtrVT));

    VRTop = DAG.getFrameIndex(FuncInfo->getVarArgsFPRIndex(), PtrVT);
    VRTop = DAG.getNode(ISD::ADD, DL, PtrVT, VRTop,
                        DAG.getConstant(FPRSize, DL, PtrVT));

    MemOps.push_back(DAG.getStore(Chain, DL, VRTop, VRTopAddr,
                                  MachinePointerInfo(SV, 16), false, false, 8));
  }

  // int __gr_offs at offset 24
  SDValue GROffsAddr =
      DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(24, DL, PtrVT));
  MemOps.push_back(DAG.getStore(Chain, DL,
                                DAG.getConstant(-GPRSize, DL, MVT::i32),
                                GROffsAddr, MachinePointerInfo(SV, 24), false,
                                false, 4));

  // int __vr_offs at offset 28
  SDValue VROffsAddr =
      DAG.getNode(ISD::ADD, DL, PtrVT, VAList, DAG.getConstant(28, DL, PtrVT));
  MemOps.push_back(DAG.getStore(Chain, DL,
                                DAG.getConstant(-FPRSize, DL, MVT::i32),
                                VROffsAddr, MachinePointerInfo(SV, 28), false,
                                false, 4));

  return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
}

SDValue AArch64TargetLowering::LowerVASTART(SDValue Op,
                                            SelectionDAG &DAG) const {
  return Subtarget->isTargetDarwin() ? LowerDarwin_VASTART(Op, DAG)
                                     : LowerAAPCS_VASTART(Op, DAG);
}

SDValue AArch64TargetLowering::LowerVACOPY(SDValue Op,
                                           SelectionDAG &DAG) const {
  // AAPCS has three pointers and two ints (= 32 bytes), Darwin has single
  // pointer.
  SDLoc DL(Op);
  unsigned VaListSize = Subtarget->isTargetDarwin() ? 8 : 32;
  const Value *DestSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
  const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();

  return DAG.getMemcpy(Op.getOperand(0), DL, Op.getOperand(1),
                       Op.getOperand(2),
                       DAG.getConstant(VaListSize, DL, MVT::i32),
                       8, false, false, false, MachinePointerInfo(DestSV),
                       MachinePointerInfo(SrcSV));
}

SDValue AArch64TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
  assert(Subtarget->isTargetDarwin() &&
         "automatic va_arg instruction only works on Darwin");

  const Value *V = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  SDValue Chain = Op.getOperand(0);
  SDValue Addr = Op.getOperand(1);
  unsigned Align = Op.getConstantOperandVal(3);
  auto PtrVT = getPointerTy(DAG.getDataLayout());

  SDValue VAList = DAG.getLoad(PtrVT, DL, Chain, Addr, MachinePointerInfo(V),
                               false, false, false, 0);
  Chain = VAList.getValue(1);

  if (Align > 8) {
    assert(((Align & (Align - 1)) == 0) && "Expected Align to be a power of 2");
    VAList = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
                         DAG.getConstant(Align - 1, DL, PtrVT));
    VAList = DAG.getNode(ISD::AND, DL, PtrVT, VAList,
                         DAG.getConstant(-(int64_t)Align, DL, PtrVT));
  }

  Type *ArgTy = VT.getTypeForEVT(*DAG.getContext());
  uint64_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);

  // Scalar integer and FP values smaller than 64 bits are implicitly extended
  // up to 64 bits.  At the very least, we have to increase the striding of the
  // vaargs list to match this, and for FP values we need to introduce
  // FP_ROUND nodes as well.
  if (VT.isInteger() && !VT.isVector())
    ArgSize = 8;
  bool NeedFPTrunc = false;
  if (VT.isFloatingPoint() && !VT.isVector() && VT != MVT::f64) {
    ArgSize = 8;
    NeedFPTrunc = true;
  }

  // Increment the pointer, VAList, to the next vaarg
  SDValue VANext = DAG.getNode(ISD::ADD, DL, PtrVT, VAList,
                               DAG.getConstant(ArgSize, DL, PtrVT));
  // Store the incremented VAList to the legalized pointer
  SDValue APStore = DAG.getStore(Chain, DL, VANext, Addr, MachinePointerInfo(V),
                                 false, false, 0);

  // Load the actual argument out of the pointer VAList
  if (NeedFPTrunc) {
    // Load the value as an f64.
    SDValue WideFP = DAG.getLoad(MVT::f64, DL, APStore, VAList,
                                 MachinePointerInfo(), false, false, false, 0);
    // Round the value down to an f32.
    SDValue NarrowFP = DAG.getNode(ISD::FP_ROUND, DL, VT, WideFP.getValue(0),
                                   DAG.getIntPtrConstant(1, DL));
    SDValue Ops[] = { NarrowFP, WideFP.getValue(1) };
    // Merge the rounded value with the chain output of the load.
    return DAG.getMergeValues(Ops, DL);
  }

  return DAG.getLoad(VT, DL, APStore, VAList, MachinePointerInfo(), false,
                     false, false, 0);
}

SDValue AArch64TargetLowering::LowerFRAMEADDR(SDValue Op,
                                              SelectionDAG &DAG) const {
  MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
  MFI->setFrameAddressIsTaken(true);

  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
  SDValue FrameAddr =
      DAG.getCopyFromReg(DAG.getEntryNode(), DL, AArch64::FP, VT);
  while (Depth--)
    FrameAddr = DAG.getLoad(VT, DL, DAG.getEntryNode(), FrameAddr,
                            MachinePointerInfo(), false, false, false, 0);
  return FrameAddr;
}

// FIXME? Maybe this could be a TableGen attribute on some registers and
// this table could be generated automatically from RegInfo.
unsigned AArch64TargetLowering::getRegisterByName(const char* RegName, EVT VT,
                                                  SelectionDAG &DAG) const {
  unsigned Reg = StringSwitch<unsigned>(RegName)
                       .Case("sp", AArch64::SP)
                       .Default(0);
  if (Reg)
    return Reg;
  report_fatal_error(Twine("Invalid register name \""
                              + StringRef(RegName)  + "\"."));
}

SDValue AArch64TargetLowering::LowerRETURNADDR(SDValue Op,
                                               SelectionDAG &DAG) const {
  MachineFunction &MF = DAG.getMachineFunction();
  MachineFrameInfo *MFI = MF.getFrameInfo();
  MFI->setReturnAddressIsTaken(true);

  EVT VT = Op.getValueType();
  SDLoc DL(Op);
  unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
  if (Depth) {
    SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
    SDValue Offset = DAG.getConstant(8, DL, getPointerTy(DAG.getDataLayout()));
    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(AArch64::LR, &AArch64::GPR64RegClass);
  return DAG.getCopyFromReg(DAG.getEntryNode(), DL, Reg, VT);
}

/// LowerShiftRightParts - Lower SRA_PARTS, which returns two
/// i64 values and take a 2 x i64 value to shift plus a shift amount.
SDValue AArch64TargetLowering::LowerShiftRightParts(SDValue Op,
                                                    SelectionDAG &DAG) const {
  assert(Op.getNumOperands() == 3 && "Not a double-shift!");
  EVT VT = Op.getValueType();
  unsigned VTBits = VT.getSizeInBits();
  SDLoc dl(Op);
  SDValue ShOpLo = Op.getOperand(0);
  SDValue ShOpHi = Op.getOperand(1);
  SDValue ShAmt = Op.getOperand(2);
  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::i64,
                                 DAG.getConstant(VTBits, dl, MVT::i64), ShAmt);
  SDValue HiBitsForLo = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, RevShAmt);

  // Unfortunately, if ShAmt == 0, we just calculated "(SHL ShOpHi, 64)" which
  // is "undef". We wanted 0, so CSEL it directly.
  SDValue Cmp = emitComparison(ShAmt, DAG.getConstant(0, dl, MVT::i64),
                               ISD::SETEQ, dl, DAG);
  SDValue CCVal = DAG.getConstant(AArch64CC::EQ, dl, MVT::i32);
  HiBitsForLo =
      DAG.getNode(AArch64ISD::CSEL, dl, VT, DAG.getConstant(0, dl, MVT::i64),
                  HiBitsForLo, CCVal, Cmp);

  SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
                                   DAG.getConstant(VTBits, dl, MVT::i64));

  SDValue LoBitsForLo = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, ShAmt);
  SDValue LoForNormalShift =
      DAG.getNode(ISD::OR, dl, VT, LoBitsForLo, HiBitsForLo);

  Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, dl, MVT::i64), ISD::SETGE,
                       dl, DAG);
  CCVal = DAG.getConstant(AArch64CC::GE, dl, MVT::i32);
  SDValue LoForBigShift = DAG.getNode(Opc, dl, VT, ShOpHi, ExtraShAmt);
  SDValue Lo = DAG.getNode(AArch64ISD::CSEL, dl, VT, LoForBigShift,
                           LoForNormalShift, CCVal, Cmp);

  // AArch64 shifts larger than the register width are wrapped rather than
  // clamped, so we can't just emit "hi >> x".
  SDValue HiForNormalShift = DAG.getNode(Opc, dl, VT, ShOpHi, ShAmt);
  SDValue HiForBigShift =
      Opc == ISD::SRA
          ? DAG.getNode(Opc, dl, VT, ShOpHi,
                        DAG.getConstant(VTBits - 1, dl, MVT::i64))
          : DAG.getConstant(0, dl, VT);
  SDValue Hi = DAG.getNode(AArch64ISD::CSEL, dl, VT, HiForBigShift,
                           HiForNormalShift, CCVal, Cmp);

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


/// LowerShiftLeftParts - Lower SHL_PARTS, which returns two
/// i64 values and take a 2 x i64 value to shift plus a shift amount.
SDValue AArch64TargetLowering::LowerShiftLeftParts(SDValue Op,
                                                   SelectionDAG &DAG) const {
  assert(Op.getNumOperands() == 3 && "Not a double-shift!");
  EVT VT = Op.getValueType();
  unsigned VTBits = VT.getSizeInBits();
  SDLoc dl(Op);
  SDValue ShOpLo = Op.getOperand(0);
  SDValue ShOpHi = Op.getOperand(1);
  SDValue ShAmt = Op.getOperand(2);

  assert(Op.getOpcode() == ISD::SHL_PARTS);
  SDValue RevShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64,
                                 DAG.getConstant(VTBits, dl, MVT::i64), ShAmt);
  SDValue LoBitsForHi = DAG.getNode(ISD::SRL, dl, VT, ShOpLo, RevShAmt);

  // Unfortunately, if ShAmt == 0, we just calculated "(SRL ShOpLo, 64)" which
  // is "undef". We wanted 0, so CSEL it directly.
  SDValue Cmp = emitComparison(ShAmt, DAG.getConstant(0, dl, MVT::i64),
                               ISD::SETEQ, dl, DAG);
  SDValue CCVal = DAG.getConstant(AArch64CC::EQ, dl, MVT::i32);
  LoBitsForHi =
      DAG.getNode(AArch64ISD::CSEL, dl, VT, DAG.getConstant(0, dl, MVT::i64),
                  LoBitsForHi, CCVal, Cmp);

  SDValue ExtraShAmt = DAG.getNode(ISD::SUB, dl, MVT::i64, ShAmt,
                                   DAG.getConstant(VTBits, dl, MVT::i64));
  SDValue HiBitsForHi = DAG.getNode(ISD::SHL, dl, VT, ShOpHi, ShAmt);
  SDValue HiForNormalShift =
      DAG.getNode(ISD::OR, dl, VT, LoBitsForHi, HiBitsForHi);

  SDValue HiForBigShift = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ExtraShAmt);

  Cmp = emitComparison(ExtraShAmt, DAG.getConstant(0, dl, MVT::i64), ISD::SETGE,
                       dl, DAG);
  CCVal = DAG.getConstant(AArch64CC::GE, dl, MVT::i32);
  SDValue Hi = DAG.getNode(AArch64ISD::CSEL, dl, VT, HiForBigShift,
                           HiForNormalShift, CCVal, Cmp);

  // AArch64 shifts of larger than register sizes are wrapped rather than
  // clamped, so we can't just emit "lo << a" if a is too big.
  SDValue LoForBigShift = DAG.getConstant(0, dl, VT);
  SDValue LoForNormalShift = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, ShAmt);
  SDValue Lo = DAG.getNode(AArch64ISD::CSEL, dl, VT, LoForBigShift,
                           LoForNormalShift, CCVal, Cmp);

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

bool AArch64TargetLowering::isOffsetFoldingLegal(
    const GlobalAddressSDNode *GA) const {
  // The AArch64 target doesn't support folding offsets into global addresses.
  return false;
}

bool AArch64TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
  // We can materialize #0.0 as fmov $Rd, XZR for 64-bit and 32-bit cases.
  // FIXME: We should be able to handle f128 as well with a clever lowering.
  if (Imm.isPosZero() && (VT == MVT::f64 || VT == MVT::f32))
    return true;

  if (VT == MVT::f64)
    return AArch64_AM::getFP64Imm(Imm) != -1;
  else if (VT == MVT::f32)
    return AArch64_AM::getFP32Imm(Imm) != -1;
  return false;
}

//===----------------------------------------------------------------------===//
//                          AArch64 Optimization Hooks
//===----------------------------------------------------------------------===//

//===----------------------------------------------------------------------===//
//                          AArch64 Inline Assembly Support
//===----------------------------------------------------------------------===//

// Table of Constraints
// TODO: This is the current set of constraints supported by ARM for the
// compiler, not all of them may make sense, e.g. S may be difficult to support.
//
// r - A general register
// w - An FP/SIMD register of some size in the range v0-v31
// x - An FP/SIMD register of some size in the range v0-v15
// I - Constant that can be used with an ADD instruction
// J - Constant that can be used with a SUB instruction
// K - Constant that can be used with a 32-bit logical instruction
// L - Constant that can be used with a 64-bit logical instruction
// M - Constant that can be used as a 32-bit MOV immediate
// N - Constant that can be used as a 64-bit MOV immediate
// Q - A memory reference with base register and no offset
// S - A symbolic address
// Y - Floating point constant zero
// Z - Integer constant zero
//
//   Note that general register operands will be output using their 64-bit x
// register name, whatever the size of the variable, unless the asm operand
// is prefixed by the %w modifier. Floating-point and SIMD register operands
// will be output with the v prefix unless prefixed by the %b, %h, %s, %d or
// %q modifier.

/// getConstraintType - Given a constraint letter, return the type of
/// constraint it is for this target.
AArch64TargetLowering::ConstraintType
AArch64TargetLowering::getConstraintType(StringRef Constraint) const {
  if (Constraint.size() == 1) {
    switch (Constraint[0]) {
    default:
      break;
    case 'z':
      return C_Other;
    case 'x':
    case 'w':
      return C_RegisterClass;
    // An address with a single base register. Due to the way we
    // currently handle addresses it is the same as 'r'.
    case 'Q':
      return C_Memory;
    }
  }
  return TargetLowering::getConstraintType(Constraint);
}

/// Examine constraint type and operand type and determine a weight value.
/// This object must already have been set up with the operand type
/// and the current alternative constraint selected.
TargetLowering::ConstraintWeight
AArch64TargetLowering::getSingleConstraintMatchWeight(
    AsmOperandInfo &info, const char *constraint) const {
  ConstraintWeight weight = CW_Invalid;
  Value *CallOperandVal = info.CallOperandVal;
  // If we don't have a value, we can't do a match,
  // but allow it at the lowest weight.
  if (!CallOperandVal)
    return CW_Default;
  Type *type = CallOperandVal->getType();
  // Look at the constraint type.
  switch (*constraint) {
  default:
    weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
    break;
  case 'x':
  case 'w':
    if (type->isFloatingPointTy() || type->isVectorTy())
      weight = CW_Register;
    break;
  case 'z':
    weight = CW_Constant;
    break;
  }
  return weight;
}

std::pair<unsigned, const TargetRegisterClass *>
AArch64TargetLowering::getRegForInlineAsmConstraint(
    const TargetRegisterInfo *TRI, StringRef Constraint, MVT VT) const {
  if (Constraint.size() == 1) {
    switch (Constraint[0]) {
    case 'r':
      if (VT.getSizeInBits() == 64)
        return std::make_pair(0U, &AArch64::GPR64commonRegClass);
      return std::make_pair(0U, &AArch64::GPR32commonRegClass);
    case 'w':
      if (VT == MVT::f32)
        return std::make_pair(0U, &AArch64::FPR32RegClass);
      if (VT.getSizeInBits() == 64)
        return std::make_pair(0U, &AArch64::FPR64RegClass);
      if (VT.getSizeInBits() == 128)
        return std::make_pair(0U, &AArch64::FPR128RegClass);
      break;
    // The instructions that this constraint is designed for can
    // only take 128-bit registers so just use that regclass.
    case 'x':
      if (VT.getSizeInBits() == 128)
        return std::make_pair(0U, &AArch64::FPR128_loRegClass);
      break;
    }
  }
  if (StringRef("{cc}").equals_lower(Constraint))
    return std::make_pair(unsigned(AArch64::NZCV), &AArch64::CCRRegClass);

  // Use the default implementation in TargetLowering to convert the register
  // constraint into a member of a register class.
  std::pair<unsigned, const TargetRegisterClass *> Res;
  Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);

  // Not found as a standard register?
  if (!Res.second) {
    unsigned Size = Constraint.size();
    if ((Size == 4 || Size == 5) && Constraint[0] == '{' &&
        tolower(Constraint[1]) == 'v' && Constraint[Size - 1] == '}') {
      int RegNo;
      bool Failed = Constraint.slice(2, Size - 1).getAsInteger(10, RegNo);
      if (!Failed && RegNo >= 0 && RegNo <= 31) {
        // v0 - v31 are aliases of q0 - q31.
        // By default we'll emit v0-v31 for this unless there's a modifier where
        // we'll emit the correct register as well.
        Res.first = AArch64::FPR128RegClass.getRegister(RegNo);
        Res.second = &AArch64::FPR128RegClass;
      }
    }
  }

  return Res;
}

/// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
/// vector.  If it is invalid, don't add anything to Ops.
void AArch64TargetLowering::LowerAsmOperandForConstraint(
    SDValue Op, std::string &Constraint, std::vector<SDValue> &Ops,
    SelectionDAG &DAG) const {
  SDValue Result;

  // Currently only support length 1 constraints.
  if (Constraint.length() != 1)
    return;

  char ConstraintLetter = Constraint[0];
  switch (ConstraintLetter) {
  default:
    break;

  // This set of constraints deal with valid constants for various instructions.
  // Validate and return a target constant for them if we can.
  case 'z': {
    // 'z' maps to xzr or wzr so it needs an input of 0.
    if (!isNullConstant(Op))
      return;

    if (Op.getValueType() == MVT::i64)
      Result = DAG.getRegister(AArch64::XZR, MVT::i64);
    else
      Result = DAG.getRegister(AArch64::WZR, MVT::i32);
    break;
  }

  case 'I':
  case 'J':
  case 'K':
  case 'L':
  case 'M':
  case 'N':
    ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op);
    if (!C)
      return;

    // Grab the value and do some validation.
    uint64_t CVal = C->getZExtValue();
    switch (ConstraintLetter) {
    // The I constraint applies only to simple ADD or SUB immediate operands:
    // i.e. 0 to 4095 with optional shift by 12
    // The J constraint applies only to ADD or SUB immediates that would be
    // valid when negated, i.e. if [an add pattern] were to be output as a SUB
    // instruction [or vice versa], in other words -1 to -4095 with optional
    // left shift by 12.
    case 'I':
      if (isUInt<12>(CVal) || isShiftedUInt<12, 12>(CVal))
        break;
      return;
    case 'J': {
      uint64_t NVal = -C->getSExtValue();
      if (isUInt<12>(NVal) || isShiftedUInt<12, 12>(NVal)) {
        CVal = C->getSExtValue();
        break;
      }
      return;
    }
    // The K and L constraints apply *only* to logical immediates, including
    // what used to be the MOVI alias for ORR (though the MOVI alias has now
    // been removed and MOV should be used). So these constraints have to
    // distinguish between bit patterns that are valid 32-bit or 64-bit
    // "bitmask immediates": for example 0xaaaaaaaa is a valid bimm32 (K), but
    // not a valid bimm64 (L) where 0xaaaaaaaaaaaaaaaa would be valid, and vice
    // versa.
    case 'K':
      if (AArch64_AM::isLogicalImmediate(CVal, 32))
        break;
      return;
    case 'L':
      if (AArch64_AM::isLogicalImmediate(CVal, 64))
        break;
      return;
    // The M and N constraints are a superset of K and L respectively, for use
    // with the MOV (immediate) alias. As well as the logical immediates they
    // also match 32 or 64-bit immediates that can be loaded either using a
    // *single* MOVZ or MOVN , such as 32-bit 0x12340000, 0x00001234, 0xffffedca
    // (M) or 64-bit 0x1234000000000000 (N) etc.
    // As a note some of this code is liberally stolen from the asm parser.
    case 'M': {
      if (!isUInt<32>(CVal))
        return;
      if (AArch64_AM::isLogicalImmediate(CVal, 32))
        break;
      if ((CVal & 0xFFFF) == CVal)
        break;
      if ((CVal & 0xFFFF0000ULL) == CVal)
        break;
      uint64_t NCVal = ~(uint32_t)CVal;
      if ((NCVal & 0xFFFFULL) == NCVal)
        break;
      if ((NCVal & 0xFFFF0000ULL) == NCVal)
        break;
      return;
    }
    case 'N': {
      if (AArch64_AM::isLogicalImmediate(CVal, 64))
        break;
      if ((CVal & 0xFFFFULL) == CVal)
        break;
      if ((CVal & 0xFFFF0000ULL) == CVal)
        break;
      if ((CVal & 0xFFFF00000000ULL) == CVal)
        break;
      if ((CVal & 0xFFFF000000000000ULL) == CVal)
        break;
      uint64_t NCVal = ~CVal;
      if ((NCVal & 0xFFFFULL) == NCVal)
        break;
      if ((NCVal & 0xFFFF0000ULL) == NCVal)
        break;
      if ((NCVal & 0xFFFF00000000ULL) == NCVal)
        break;
      if ((NCVal & 0xFFFF000000000000ULL) == NCVal)
        break;
      return;
    }
    default:
      return;
    }

    // All assembler immediates are 64-bit integers.
    Result = DAG.getTargetConstant(CVal, SDLoc(Op), MVT::i64);
    break;
  }

  if (Result.getNode()) {
    Ops.push_back(Result);
    return;
  }

  return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
}

//===----------------------------------------------------------------------===//
//                     AArch64 Advanced SIMD Support
//===----------------------------------------------------------------------===//

/// WidenVector - Given a value in the V64 register class, produce the
/// equivalent value in the V128 register class.
static SDValue WidenVector(SDValue V64Reg, SelectionDAG &DAG) {
  EVT VT = V64Reg.getValueType();
  unsigned NarrowSize = VT.getVectorNumElements();
  MVT EltTy = VT.getVectorElementType().getSimpleVT();
  MVT WideTy = MVT::getVectorVT(EltTy, 2 * NarrowSize);
  SDLoc DL(V64Reg);

  return DAG.getNode(ISD::INSERT_SUBVECTOR, DL, WideTy, DAG.getUNDEF(WideTy),
                     V64Reg, DAG.getConstant(0, DL, MVT::i32));
}

/// getExtFactor - Determine the adjustment factor for the position when
/// generating an "extract from vector registers" instruction.
static unsigned getExtFactor(SDValue &V) {
  EVT EltType = V.getValueType().getVectorElementType();
  return EltType.getSizeInBits() / 8;
}

/// NarrowVector - Given a value in the V128 register class, produce the
/// equivalent value in the V64 register class.
static SDValue NarrowVector(SDValue V128Reg, SelectionDAG &DAG) {
  EVT VT = V128Reg.getValueType();
  unsigned WideSize = VT.getVectorNumElements();
  MVT EltTy = VT.getVectorElementType().getSimpleVT();
  MVT NarrowTy = MVT::getVectorVT(EltTy, WideSize / 2);
  SDLoc DL(V128Reg);

  return DAG.getTargetExtractSubreg(AArch64::dsub, DL, NarrowTy, V128Reg);
}

// Gather data to see if the operation can be modelled as a
// shuffle in combination with VEXTs.
SDValue AArch64TargetLowering::ReconstructShuffle(SDValue Op,
                                                  SelectionDAG &DAG) const {
  assert(Op.getOpcode() == ISD::BUILD_VECTOR && "Unknown opcode!");
  SDLoc dl(Op);
  EVT VT = Op.getValueType();
  unsigned NumElts = VT.getVectorNumElements();

  struct ShuffleSourceInfo {
    SDValue Vec;
    unsigned MinElt;
    unsigned MaxElt;

    // We may insert some combination of BITCASTs and VEXT nodes to force Vec to
    // be compatible with the shuffle we intend to construct. As a result
    // ShuffleVec will be some sliding window into the original Vec.
    SDValue ShuffleVec;

    // Code should guarantee that element i in Vec starts at element "WindowBase
    // + i * WindowScale in ShuffleVec".
    int WindowBase;
    int WindowScale;

    bool operator ==(SDValue OtherVec) { return Vec == OtherVec; }
    ShuffleSourceInfo(SDValue Vec)
        : Vec(Vec), MinElt(UINT_MAX), MaxElt(0), ShuffleVec(Vec), WindowBase(0),
          WindowScale(1) {}
  };

  // First gather all vectors used as an immediate source for this BUILD_VECTOR
  // node.
  SmallVector<ShuffleSourceInfo, 2> Sources;
  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();
    }

    // Add this element source to the list if it's not already there.
    SDValue SourceVec = V.getOperand(0);
    auto Source = std::find(Sources.begin(), Sources.end(), SourceVec);
    if (Source == Sources.end())
      Source = Sources.insert(Sources.end(), ShuffleSourceInfo(SourceVec));

    // Update the minimum and maximum lane number seen.
    unsigned EltNo = cast<ConstantSDNode>(V.getOperand(1))->getZExtValue();
    Source->MinElt = std::min(Source->MinElt, EltNo);
    Source->MaxElt = std::max(Source->MaxElt, EltNo);
  }

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

  // Find out the smallest element size among result and two sources, and use
  // it as element size to build the shuffle_vector.
  EVT SmallestEltTy = VT.getVectorElementType();
  for (auto &Source : Sources) {
    EVT SrcEltTy = Source.Vec.getValueType().getVectorElementType();
    if (SrcEltTy.bitsLT(SmallestEltTy)) {
      SmallestEltTy = SrcEltTy;
    }
  }
  unsigned ResMultiplier =
      VT.getVectorElementType().getSizeInBits() / SmallestEltTy.getSizeInBits();
  NumElts = VT.getSizeInBits() / SmallestEltTy.getSizeInBits();
  EVT ShuffleVT = EVT::getVectorVT(*DAG.getContext(), SmallestEltTy, NumElts);

  // If the source vector is too wide or too narrow, we may nevertheless be able
  // to construct a compatible shuffle either by concatenating it with UNDEF or
  // extracting a suitable range of elements.
  for (auto &Src : Sources) {
    EVT SrcVT = Src.ShuffleVec.getValueType();

    if (SrcVT.getSizeInBits() == VT.getSizeInBits())
      continue;

    // This stage of the search produces a source with the same element type as
    // the original, but with a total width matching the BUILD_VECTOR output.
    EVT EltVT = SrcVT.getVectorElementType();
    unsigned NumSrcElts = VT.getSizeInBits() / EltVT.getSizeInBits();
    EVT DestVT = EVT::getVectorVT(*DAG.getContext(), EltVT, NumSrcElts);

    if (SrcVT.getSizeInBits() < VT.getSizeInBits()) {
      assert(2 * SrcVT.getSizeInBits() == VT.getSizeInBits());
      // We can pad out the smaller vector for free, so if it's part of a
      // shuffle...
      Src.ShuffleVec =
          DAG.getNode(ISD::CONCAT_VECTORS, dl, DestVT, Src.ShuffleVec,
                      DAG.getUNDEF(Src.ShuffleVec.getValueType()));
      continue;
    }

    assert(SrcVT.getSizeInBits() == 2 * VT.getSizeInBits());

    if (Src.MaxElt - Src.MinElt >= NumSrcElts) {
      // Span too large for a VEXT to cope
      return SDValue();
    }

    if (Src.MinElt >= NumSrcElts) {
      // The extraction can just take the second half
      Src.ShuffleVec =
          DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
                      DAG.getConstant(NumSrcElts, dl, MVT::i64));
      Src.WindowBase = -NumSrcElts;
    } else if (Src.MaxElt < NumSrcElts) {
      // The extraction can just take the first half
      Src.ShuffleVec =
          DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
                      DAG.getConstant(0, dl, MVT::i64));
    } else {
      // An actual VEXT is needed
      SDValue VEXTSrc1 =
          DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
                      DAG.getConstant(0, dl, MVT::i64));
      SDValue VEXTSrc2 =
          DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, DestVT, Src.ShuffleVec,
                      DAG.getConstant(NumSrcElts, dl, MVT::i64));
      unsigned Imm = Src.MinElt * getExtFactor(VEXTSrc1);

      Src.ShuffleVec = DAG.getNode(AArch64ISD::EXT, dl, DestVT, VEXTSrc1,
                                   VEXTSrc2,
                                   DAG.getConstant(Imm, dl, MVT::i32));
      Src.WindowBase = -Src.MinElt;
    }
  }

  // Another possible incompatibility occurs from the vector element types. We
  // can fix this by bitcasting the source vectors to the same type we intend
  // for the shuffle.
  for (auto &Src : Sources) {
    EVT SrcEltTy = Src.ShuffleVec.getValueType().getVectorElementType();
    if (SrcEltTy == SmallestEltTy)
      continue;
    assert(ShuffleVT.getVectorElementType() == SmallestEltTy);
    Src.ShuffleVec = DAG.getNode(ISD::BITCAST, dl, ShuffleVT, Src.ShuffleVec);
    Src.WindowScale = SrcEltTy.getSizeInBits() / SmallestEltTy.getSizeInBits();
    Src.WindowBase *= Src.WindowScale;
  }

  // Final sanity check before we try to actually produce a shuffle.
  DEBUG(
    for (auto Src : Sources)
      assert(Src.ShuffleVec.getValueType() == ShuffleVT);
  );

  // The stars all align, our next step is to produce the mask for the shuffle.
  SmallVector<int, 8> Mask(ShuffleVT.getVectorNumElements(), -1);
  int BitsPerShuffleLane = ShuffleVT.getVectorElementType().getSizeInBits();
  for (unsigned i = 0; i < VT.getVectorNumElements(); ++i) {
    SDValue Entry = Op.getOperand(i);
    if (Entry.getOpcode() == ISD::UNDEF)
      continue;

    auto Src = std::find(Sources.begin(), Sources.end(), Entry.getOperand(0));
    int EltNo = cast<ConstantSDNode>(Entry.getOperand(1))->getSExtValue();

    // EXTRACT_VECTOR_ELT performs an implicit any_ext; BUILD_VECTOR an implicit
    // trunc. So only std::min(SrcBits, DestBits) actually get defined in this
    // segment.
    EVT OrigEltTy = Entry.getOperand(0).getValueType().getVectorElementType();
    int BitsDefined = std::min(OrigEltTy.getSizeInBits(),
                               VT.getVectorElementType().getSizeInBits());
    int LanesDefined = BitsDefined / BitsPerShuffleLane;

    // This source is expected to fill ResMultiplier lanes of the final shuffle,
    // starting at the appropriate offset.
    int *LaneMask = &Mask[i * ResMultiplier];

    int ExtractBase = EltNo * Src->WindowScale + Src->WindowBase;
    ExtractBase += NumElts * (Src - Sources.begin());
    for (int j = 0; j < LanesDefined; ++j)
      LaneMask[j] = ExtractBase + j;
  }

  // Final check before we try to produce nonsense...
  if (!isShuffleMaskLegal(Mask, ShuffleVT))
    return SDValue();

  SDValue ShuffleOps[] = { DAG.getUNDEF(ShuffleVT), DAG.getUNDEF(ShuffleVT) };
  for (unsigned i = 0; i < Sources.size(); ++i)
    ShuffleOps[i] = Sources[i].ShuffleVec;

  SDValue Shuffle = DAG.getVectorShuffle(ShuffleVT, dl, ShuffleOps[0],
                                         ShuffleOps[1], &Mask[0]);
  return DAG.getNode(ISD::BITCAST, dl, VT, Shuffle);
}

// check if an EXT instruction can handle the shuffle mask when the
// vector sources of the shuffle are the same.
static bool isSingletonEXTMask(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;
}

// check if an EXT instruction can handle the shuffle mask when the
// vector sources of the shuffle are different.
static bool isEXTMask(ArrayRef<int> M, EVT VT, bool &ReverseEXT,
                      unsigned &Imm) {
  // Look for the first non-undef element.
  const int *FirstRealElt = std::find_if(M.begin(), M.end(),
      [](int Elt) {return Elt >= 0;});

  // Benefit form APInt to handle overflow when calculating expected element.
  unsigned NumElts = VT.getVectorNumElements();
  unsigned MaskBits = APInt(32, NumElts * 2).logBase2();
  APInt ExpectedElt = APInt(MaskBits, *FirstRealElt + 1);
  // The following shuffle indices must be the successive elements after the
  // first real element.
  const int *FirstWrongElt = std::find_if(FirstRealElt + 1, M.end(),
      [&](int Elt) {return Elt != ExpectedElt++ && Elt != -1;});
  if (FirstWrongElt != M.end())
    return false;

  // The index of an EXT is the first element if it is not UNDEF.
  // Watch out for the beginning UNDEFs. The EXT index should be the expected
  // value of the first element.  E.g.
  // <-1, -1, 3, ...> is treated as <1, 2, 3, ...>.
  // <-1, -1, 0, 1, ...> is treated as <2*NumElts-2, 2*NumElts-1, 0, 1, ...>.
  // ExpectedElt is the last mask index plus 1.
  Imm = ExpectedElt.getZExtValue();

  // There are two difference cases requiring to reverse input vectors.
  // For example, for vector <4 x i32> we have the following cases,
  // Case 1: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, -1, 0>)
  // Case 2: shufflevector(<4 x i32>,<4 x i32>,<-1, -1, 7, 0>)
  // For both cases, we finally use mask <5, 6, 7, 0>, which requires
  // to reverse two input vectors.
  if (Imm < NumElts)
    ReverseEXT = true;
  else
    Imm -= NumElts;

  return true;
}

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

  return true;
}

static bool isUZPMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
  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;
  }

  return true;
}

static bool isTRNMask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
  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;
}

/// isZIP_v_undef_Mask - Special case of isZIPMask 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 isZIP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
  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;
  }

  return true;
}

/// isUZP_v_undef_Mask - Special case of isUZPMask 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 isUZP_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
  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;
    }
  }

  return true;
}

/// isTRN_v_undef_Mask - Special case of isTRNMask 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 isTRN_v_undef_Mask(ArrayRef<int> M, EVT VT, unsigned &WhichResult) {
  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 isINSMask(ArrayRef<int> M, int NumInputElements,
                      bool &DstIsLeft, int &Anomaly) {
  if (M.size() != static_cast<size_t>(NumInputElements))
    return false;

  int NumLHSMatch = 0, NumRHSMatch = 0;
  int LastLHSMismatch = -1, LastRHSMismatch = -1;

  for (int i = 0; i < NumInputElements; ++i) {
    if (M[i] == -1) {
      ++NumLHSMatch;
      ++NumRHSMatch;
      continue;
    }

    if (M[i] == i)
      ++NumLHSMatch;
    else
      LastLHSMismatch = i;

    if (M[i] == i + NumInputElements)
      ++NumRHSMatch;
    else
      LastRHSMismatch = i;
  }

  if (NumLHSMatch == NumInputElements - 1) {
    DstIsLeft = true;
    Anomaly = LastLHSMismatch;
    return true;
  } else if (NumRHSMatch == NumInputElements - 1) {
    DstIsLeft = false;
    Anomaly = LastRHSMismatch;
    return true;
  }

  return false;
}

static bool isConcatMask(ArrayRef<int> Mask, EVT VT, bool SplitLHS) {
  if (VT.getSizeInBits() != 128)
    return false;

  unsigned NumElts = VT.getVectorNumElements();

  for (int I = 0, E = NumElts / 2; I != E; I++) {
    if (Mask[I] != I)
      return false;
  }

  int Offset = NumElts / 2;
  for (int I = NumElts / 2, E = NumElts; I != E; I++) {
    if (Mask[I] != I + SplitLHS * Offset)
      return false;
  }

  return true;
}

static SDValue tryFormConcatFromShuffle(SDValue Op, SelectionDAG &DAG) {
  SDLoc DL(Op);
  EVT VT = Op.getValueType();
  SDValue V0 = Op.getOperand(0);
  SDValue V1 = Op.getOperand(1);
  ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(Op)->getMask();

  if (VT.getVectorElementType() != V0.getValueType().getVectorElementType() ||
      VT.getVectorElementType() != V1.getValueType().getVectorElementType())
    return SDValue();

  bool SplitV0 = V0.getValueType().getSizeInBits() == 128;

  if (!isConcatMask(Mask, VT, SplitV0))
    return SDValue();

  EVT CastVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
                                VT.getVectorNumElements() / 2);
  if (SplitV0) {
    V0 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V0,
                     DAG.getConstant(0, DL, MVT::i64));
  }
  if (V1.getValueType().getSizeInBits() == 128) {
    V1 = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, CastVT, V1,
                     DAG.getConstant(0, DL, MVT::i64));
  }
  return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, V0, V1);
}

/// 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,
                                      SDLoc 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(AArch64ISD::REV64, dl, VT, OpLHS);
    // vrev <4 x i16> -> REV32
    if (VT.getVectorElementType() == MVT::i16 ||
        VT.getVectorElementType() == MVT::f16)
      return DAG.getNode(AArch64ISD::REV32, dl, VT, OpLHS);
    // vrev <4 x i8> -> REV16
    assert(VT.getVectorElementType() == MVT::i8);
    return DAG.getNode(AArch64ISD::REV16, dl, VT, OpLHS);
  case OP_VDUP0:
  case OP_VDUP1:
  case OP_VDUP2:
  case OP_VDUP3: {
    EVT EltTy = VT.getVectorElementType();
    unsigned Opcode;
    if (EltTy == MVT::i8)
      Opcode = AArch64ISD::DUPLANE8;
    else if (EltTy == MVT::i16 || EltTy == MVT::f16)
      Opcode = AArch64ISD::DUPLANE16;
    else if (EltTy == MVT::i32 || EltTy == MVT::f32)
      Opcode = AArch64ISD::DUPLANE32;
    else if (EltTy == MVT::i64 || EltTy == MVT::f64)
      Opcode = AArch64ISD::DUPLANE64;
    else
      llvm_unreachable("Invalid vector element type?");

    if (VT.getSizeInBits() == 64)
      OpLHS = WidenVector(OpLHS, DAG);
    SDValue Lane = DAG.getConstant(OpNum - OP_VDUP0, dl, MVT::i64);
    return DAG.getNode(Opcode, dl, VT, OpLHS, Lane);
  }
  case OP_VEXT1:
  case OP_VEXT2:
  case OP_VEXT3: {
    unsigned Imm = (OpNum - OP_VEXT1 + 1) * getExtFactor(OpLHS);
    return DAG.getNode(AArch64ISD::EXT, dl, VT, OpLHS, OpRHS,
                       DAG.getConstant(Imm, dl, MVT::i32));
  }
  case OP_VUZPL:
    return DAG.getNode(AArch64ISD::UZP1, dl, DAG.getVTList(VT, VT), OpLHS,
                       OpRHS);
  case OP_VUZPR:
    return DAG.getNode(AArch64ISD::UZP2, dl, DAG.getVTList(VT, VT), OpLHS,
                       OpRHS);
  case OP_VZIPL:
    return DAG.getNode(AArch64ISD::ZIP1, dl, DAG.getVTList(VT, VT), OpLHS,
                       OpRHS);
  case OP_VZIPR:
    return DAG.getNode(AArch64ISD::ZIP2, dl, DAG.getVTList(VT, VT), OpLHS,
                       OpRHS);
  case OP_VTRNL:
    return DAG.getNode(AArch64ISD::TRN1, dl, DAG.getVTList(VT, VT), OpLHS,
                       OpRHS);
  case OP_VTRNR:
    return DAG.getNode(AArch64ISD::TRN2, dl, DAG.getVTList(VT, VT), OpLHS,
                       OpRHS);
  }
}

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

  EVT EltVT = Op.getValueType().getVectorElementType();
  unsigned BytesPerElt = EltVT.getSizeInBits() / 8;

  SmallVector<SDValue, 8> TBLMask;
  for (int Val : ShuffleMask) {
    for (unsigned Byte = 0; Byte < BytesPerElt; ++Byte) {
      unsigned Offset = Byte + Val * BytesPerElt;
      TBLMask.push_back(DAG.getConstant(Offset, DL, MVT::i32));
    }
  }

  MVT IndexVT = MVT::v8i8;
  unsigned IndexLen = 8;
  if (Op.getValueType().getSizeInBits() == 128) {
    IndexVT = MVT::v16i8;
    IndexLen = 16;
  }

  SDValue V1Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V1);
  SDValue V2Cst = DAG.getNode(ISD::BITCAST, DL, IndexVT, V2);

  SDValue Shuffle;
  if (V2.getNode()->getOpcode() == ISD::UNDEF) {
    if (IndexLen == 8)
      V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V1Cst);
    Shuffle = DAG.getNode(
        ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
        DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), V1Cst,
        DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
                    makeArrayRef(TBLMask.data(), IndexLen)));
  } else {
    if (IndexLen == 8) {
      V1Cst = DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v16i8, V1Cst, V2Cst);
      Shuffle = DAG.getNode(
          ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
          DAG.getConstant(Intrinsic::aarch64_neon_tbl1, DL, MVT::i32), V1Cst,
          DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
                      makeArrayRef(TBLMask.data(), IndexLen)));
    } else {
      // FIXME: We cannot, for the moment, emit a TBL2 instruction because we
      // cannot currently represent the register constraints on the input
      // table registers.
      //  Shuffle = DAG.getNode(AArch64ISD::TBL2, DL, IndexVT, V1Cst, V2Cst,
      //                   DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
      //                               &TBLMask[0], IndexLen));
      Shuffle = DAG.getNode(
          ISD::INTRINSIC_WO_CHAIN, DL, IndexVT,
          DAG.getConstant(Intrinsic::aarch64_neon_tbl2, DL, MVT::i32),
          V1Cst, V2Cst,
          DAG.getNode(ISD::BUILD_VECTOR, DL, IndexVT,
                      makeArrayRef(TBLMask.data(), IndexLen)));
    }
  }
  return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Shuffle);
}

static unsigned getDUPLANEOp(EVT EltType) {
  if (EltType == MVT::i8)
    return AArch64ISD::DUPLANE8;
  if (EltType == MVT::i16 || EltType == MVT::f16)
    return AArch64ISD::DUPLANE16;
  if (EltType == MVT::i32 || EltType == MVT::f32)
    return AArch64ISD::DUPLANE32;
  if (EltType == MVT::i64 || EltType == MVT::f64)
    return AArch64ISD::DUPLANE64;

  llvm_unreachable("Invalid vector element type?");
}

SDValue AArch64TargetLowering::LowerVECTOR_SHUFFLE(SDValue Op,
                                                   SelectionDAG &DAG) const {
  SDLoc dl(Op);
  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.
  ArrayRef<int> ShuffleMask = SVN->getMask();

  SDValue V1 = Op.getOperand(0);
  SDValue V2 = Op.getOperand(1);

  if (ShuffleVectorSDNode::isSplatMask(&ShuffleMask[0],
                                       V1.getValueType().getSimpleVT())) {
    int Lane = SVN->getSplatIndex();
    // If this is undef splat, generate it via "just" vdup, if possible.
    if (Lane == -1)
      Lane = 0;

    if (Lane == 0 && V1.getOpcode() == ISD::SCALAR_TO_VECTOR)
      return DAG.getNode(AArch64ISD::DUP, dl, V1.getValueType(),
                         V1.getOperand(0));
    // Test if V1 is a BUILD_VECTOR and the lane being referenced is a non-
    // constant. If so, we can just reference the lane's definition directly.
    if (V1.getOpcode() == ISD::BUILD_VECTOR &&
        !isa<ConstantSDNode>(V1.getOperand(Lane)))
      return DAG.getNode(AArch64ISD::DUP, dl, VT, V1.getOperand(Lane));

    // Otherwise, duplicate from the lane of the input vector.
    unsigned Opcode = getDUPLANEOp(V1.getValueType().getVectorElementType());

    // SelectionDAGBuilder may have "helpfully" already extracted or conatenated
    // to make a vector of the same size as this SHUFFLE. We can ignore the
    // extract entirely, and canonicalise the concat using WidenVector.
    if (V1.getOpcode() == ISD::EXTRACT_SUBVECTOR) {
      Lane += cast<ConstantSDNode>(V1.getOperand(1))->getZExtValue();
      V1 = V1.getOperand(0);
    } else if (V1.getOpcode() == ISD::CONCAT_VECTORS) {
      unsigned Idx = Lane >= (int)VT.getVectorNumElements() / 2;
      Lane -= Idx * VT.getVectorNumElements() / 2;
      V1 = WidenVector(V1.getOperand(Idx), DAG);
    } else if (VT.getSizeInBits() == 64)
      V1 = WidenVector(V1, DAG);

    return DAG.getNode(Opcode, dl, VT, V1, DAG.getConstant(Lane, dl, MVT::i64));
  }

  if (isREVMask(ShuffleMask, VT, 64))
    return DAG.getNode(AArch64ISD::REV64, dl, V1.getValueType(), V1, V2);
  if (isREVMask(ShuffleMask, VT, 32))
    return DAG.getNode(AArch64ISD::REV32, dl, V1.getValueType(), V1, V2);
  if (isREVMask(ShuffleMask, VT, 16))
    return DAG.getNode(AArch64ISD::REV16, dl, V1.getValueType(), V1, V2);

  bool ReverseEXT = false;
  unsigned Imm;
  if (isEXTMask(ShuffleMask, VT, ReverseEXT, Imm)) {
    if (ReverseEXT)
      std::swap(V1, V2);
    Imm *= getExtFactor(V1);
    return DAG.getNode(AArch64ISD::EXT, dl, V1.getValueType(), V1