js/src/jit/arm/Assembler-arm.h
author Nicolas B. Pierron <nicolas.b.pierron@nbp.name>
Tue, 16 Apr 2019 13:56:58 +0000
changeset 469672 e55ace0633daf9eb7bd5f260300fdbd770834bb3
parent 467003 7de6e431f1ae32cacc2b5d4aea1c47bf7900c4c4
child 470367 44928a7f57454b09e22399f2bb45f715015c267b
permissions -rw-r--r--
Bug 1534840 part 3 - Prevent ARM from generating nops within jump tables. r=sstangl Depends on D26522 Differential Revision: https://phabricator.services.mozilla.com/D26523

/* -*- Mode: C++; tab-width: 8; indent-tabs-mode: nil; c-basic-offset: 2 -*-
 * vim: set ts=8 sts=2 et sw=2 tw=80:
 * This Source Code Form is subject to the terms of the Mozilla Public
 * License, v. 2.0. If a copy of the MPL was not distributed with this
 * file, You can obtain one at http://mozilla.org/MPL/2.0/. */

#ifndef jit_arm_Assembler_arm_h
#define jit_arm_Assembler_arm_h

#include "mozilla/ArrayUtils.h"
#include "mozilla/Attributes.h"
#include "mozilla/MathAlgorithms.h"

#include "jit/arm/Architecture-arm.h"
#include "jit/arm/disasm/Disasm-arm.h"
#include "jit/CompactBuffer.h"
#include "jit/IonCode.h"
#include "jit/JitRealm.h"
#include "jit/shared/Assembler-shared.h"
#include "jit/shared/Disassembler-shared.h"
#include "jit/shared/IonAssemblerBufferWithConstantPools.h"

union PoolHintPun;

namespace js {
namespace jit {

using LiteralDoc = DisassemblerSpew::LiteralDoc;
using LabelDoc = DisassemblerSpew::LabelDoc;

// NOTE: there are duplicates in this list! Sometimes we want to specifically
// refer to the link register as a link register (bl lr is much clearer than bl
// r14). HOWEVER, this register can easily be a gpr when it is not busy holding
// the return address.
static constexpr Register r0{Registers::r0};
static constexpr Register r1{Registers::r1};
static constexpr Register r2{Registers::r2};
static constexpr Register r3{Registers::r3};
static constexpr Register r4{Registers::r4};
static constexpr Register r5{Registers::r5};
static constexpr Register r6{Registers::r6};
static constexpr Register r7{Registers::r7};
static constexpr Register r8{Registers::r8};
static constexpr Register r9{Registers::r9};
static constexpr Register r10{Registers::r10};
static constexpr Register r11{Registers::r11};
static constexpr Register r12{Registers::ip};
static constexpr Register ip{Registers::ip};
static constexpr Register sp{Registers::sp};
static constexpr Register r14{Registers::lr};
static constexpr Register lr{Registers::lr};
static constexpr Register pc{Registers::pc};

static constexpr Register ScratchRegister{Registers::ip};

// Helper class for ScratchRegister usage. Asserts that only one piece
// of code thinks it has exclusive ownership of the scratch register.
struct ScratchRegisterScope : public AutoRegisterScope {
  explicit ScratchRegisterScope(MacroAssembler& masm)
      : AutoRegisterScope(masm, ScratchRegister) {}
};

struct SecondScratchRegisterScope : public AutoRegisterScope {
  explicit SecondScratchRegisterScope(MacroAssembler& masm);
};

static constexpr Register OsrFrameReg = r3;
static constexpr Register CallTempReg0 = r5;
static constexpr Register CallTempReg1 = r6;
static constexpr Register CallTempReg2 = r7;
static constexpr Register CallTempReg3 = r8;
static constexpr Register CallTempReg4 = r0;
static constexpr Register CallTempReg5 = r1;

static constexpr Register IntArgReg0 = r0;
static constexpr Register IntArgReg1 = r1;
static constexpr Register IntArgReg2 = r2;
static constexpr Register IntArgReg3 = r3;
static constexpr Register HeapReg = r10;
static constexpr Register CallTempNonArgRegs[] = {r5, r6, r7, r8};
static const uint32_t NumCallTempNonArgRegs =
    mozilla::ArrayLength(CallTempNonArgRegs);

// These register assignments for the 64-bit atomic ops are frequently too
// constraining, but we have no way of expressing looser constraints to the
// register allocator.

// CompareExchange: Any two odd/even pairs would do for `new` and `out`, and any
// pair would do for `old`, so long as none of them overlap.

static constexpr Register CmpXchgOldLo = r4;
static constexpr Register CmpXchgOldHi = r5;
static constexpr Register64 CmpXchgOld64 =
    Register64(CmpXchgOldHi, CmpXchgOldLo);
static constexpr Register CmpXchgNewLo = IntArgReg2;
static constexpr Register CmpXchgNewHi = IntArgReg3;
static constexpr Register64 CmpXchgNew64 =
    Register64(CmpXchgNewHi, CmpXchgNewLo);
static constexpr Register CmpXchgOutLo = IntArgReg0;
static constexpr Register CmpXchgOutHi = IntArgReg1;

// Exchange: Any two non-equal odd/even pairs would do for `new` and `out`.

static constexpr Register XchgNewLo = IntArgReg2;
static constexpr Register XchgNewHi = IntArgReg3;
static constexpr Register64 XchgNew64 = Register64(XchgNewHi, XchgNewLo);
static constexpr Register XchgOutLo = IntArgReg0;
static constexpr Register XchgOutHi = IntArgReg1;

// Atomic rmw operations: Any two odd/even pairs would do for `tmp` and `out`,
// and any pair would do for `val`, so long as none of them overlap.

static constexpr Register FetchOpValLo = r4;
static constexpr Register FetchOpValHi = r5;
static constexpr Register64 FetchOpVal64 =
    Register64(FetchOpValHi, FetchOpValLo);
static constexpr Register FetchOpTmpLo = IntArgReg2;
static constexpr Register FetchOpTmpHi = IntArgReg3;
static constexpr Register FetchOpOutLo = IntArgReg0;
static constexpr Register FetchOpOutHi = IntArgReg1;

class ABIArgGenerator {
  unsigned intRegIndex_;
  unsigned floatRegIndex_;
  uint32_t stackOffset_;
  ABIArg current_;

  // ARM can either use HardFp (use float registers for float arguments), or
  // SoftFp (use general registers for float arguments) ABI.  We keep this
  // switch as a runtime switch because wasm always use the HardFp back-end
  // while the calls to native functions have to use the one provided by the
  // system.
  bool useHardFp_;

  ABIArg softNext(MIRType argType);
  ABIArg hardNext(MIRType argType);

 public:
  ABIArgGenerator();

  void setUseHardFp(bool useHardFp) {
    MOZ_ASSERT(intRegIndex_ == 0 && floatRegIndex_ == 0);
    useHardFp_ = useHardFp;
  }
  ABIArg next(MIRType argType);
  ABIArg& current() { return current_; }
  uint32_t stackBytesConsumedSoFar() const { return stackOffset_; }
};

bool IsUnaligned(const wasm::MemoryAccessDesc& access);

// These registers may be volatile or nonvolatile.
static constexpr Register ABINonArgReg0 = r4;
static constexpr Register ABINonArgReg1 = r5;
static constexpr Register ABINonArgReg2 = r6;
static constexpr Register ABINonArgReg3 = r7;

// This register may be volatile or nonvolatile. Avoid d15 which is the
// ScratchDoubleReg.
static constexpr FloatRegister ABINonArgDoubleReg{FloatRegisters::d8,
                                                  VFPRegister::Double};

// These registers may be volatile or nonvolatile.
// Note: these three registers are all guaranteed to be different
static constexpr Register ABINonArgReturnReg0 = r4;
static constexpr Register ABINonArgReturnReg1 = r5;
static constexpr Register ABINonVolatileReg = r6;

// This register is guaranteed to be clobberable during the prologue and
// epilogue of an ABI call which must preserve both ABI argument, return
// and non-volatile registers.
static constexpr Register ABINonArgReturnVolatileReg = lr;

// TLS pointer argument register for WebAssembly functions. This must not alias
// any other register used for passing function arguments or return values.
// Preserved by WebAssembly functions.
static constexpr Register WasmTlsReg = r9;

// Registers used for wasm table calls. These registers must be disjoint
// from the ABI argument registers, WasmTlsReg and each other.
static constexpr Register WasmTableCallScratchReg0 = ABINonArgReg0;
static constexpr Register WasmTableCallScratchReg1 = ABINonArgReg1;
static constexpr Register WasmTableCallSigReg = ABINonArgReg2;
static constexpr Register WasmTableCallIndexReg = ABINonArgReg3;

static constexpr Register PreBarrierReg = r1;

static constexpr Register InvalidReg{Registers::invalid_reg};
static constexpr FloatRegister InvalidFloatReg;

static constexpr Register JSReturnReg_Type = r3;
static constexpr Register JSReturnReg_Data = r2;
static constexpr Register StackPointer = sp;
static constexpr Register FramePointer = r11;
static constexpr Register ReturnReg = r0;
static constexpr Register64 ReturnReg64(r1, r0);

// The attribute '__value_in_regs' alters the calling convention of a function
// so that a structure of up to four elements can be returned via the argument
// registers rather than being written to memory.
static constexpr Register ReturnRegVal0 = IntArgReg0;
static constexpr Register ReturnRegVal1 = IntArgReg1;
static constexpr Register ReturnRegVal2 = IntArgReg2;
static constexpr Register ReturnRegVal3 = IntArgReg3;

static constexpr FloatRegister ReturnFloat32Reg = {FloatRegisters::d0,
                                                   VFPRegister::Single};
static constexpr FloatRegister ReturnDoubleReg = {FloatRegisters::d0,
                                                  VFPRegister::Double};
static constexpr FloatRegister ReturnSimd128Reg = InvalidFloatReg;
static constexpr FloatRegister ScratchFloat32Reg = {FloatRegisters::s30,
                                                    VFPRegister::Single};
static constexpr FloatRegister ScratchDoubleReg = {FloatRegisters::d15,
                                                   VFPRegister::Double};
static constexpr FloatRegister ScratchSimd128Reg = InvalidFloatReg;
static constexpr FloatRegister ScratchUIntReg = {FloatRegisters::d15,
                                                 VFPRegister::UInt};
static constexpr FloatRegister ScratchIntReg = {FloatRegisters::d15,
                                                VFPRegister::Int};

struct ScratchFloat32Scope : public AutoFloatRegisterScope {
  explicit ScratchFloat32Scope(MacroAssembler& masm)
      : AutoFloatRegisterScope(masm, ScratchFloat32Reg) {}
};
struct ScratchDoubleScope : public AutoFloatRegisterScope {
  explicit ScratchDoubleScope(MacroAssembler& masm)
      : AutoFloatRegisterScope(masm, ScratchDoubleReg) {}
};

// Registerd used in RegExpMatcher instruction (do not use JSReturnOperand).
static constexpr Register RegExpMatcherRegExpReg = CallTempReg0;
static constexpr Register RegExpMatcherStringReg = CallTempReg1;
static constexpr Register RegExpMatcherLastIndexReg = CallTempReg2;

// Registerd used in RegExpTester instruction (do not use ReturnReg).
static constexpr Register RegExpTesterRegExpReg = CallTempReg0;
static constexpr Register RegExpTesterStringReg = CallTempReg1;
static constexpr Register RegExpTesterLastIndexReg = CallTempReg2;

static constexpr FloatRegister d0 = {FloatRegisters::d0, VFPRegister::Double};
static constexpr FloatRegister d1 = {FloatRegisters::d1, VFPRegister::Double};
static constexpr FloatRegister d2 = {FloatRegisters::d2, VFPRegister::Double};
static constexpr FloatRegister d3 = {FloatRegisters::d3, VFPRegister::Double};
static constexpr FloatRegister d4 = {FloatRegisters::d4, VFPRegister::Double};
static constexpr FloatRegister d5 = {FloatRegisters::d5, VFPRegister::Double};
static constexpr FloatRegister d6 = {FloatRegisters::d6, VFPRegister::Double};
static constexpr FloatRegister d7 = {FloatRegisters::d7, VFPRegister::Double};
static constexpr FloatRegister d8 = {FloatRegisters::d8, VFPRegister::Double};
static constexpr FloatRegister d9 = {FloatRegisters::d9, VFPRegister::Double};
static constexpr FloatRegister d10 = {FloatRegisters::d10, VFPRegister::Double};
static constexpr FloatRegister d11 = {FloatRegisters::d11, VFPRegister::Double};
static constexpr FloatRegister d12 = {FloatRegisters::d12, VFPRegister::Double};
static constexpr FloatRegister d13 = {FloatRegisters::d13, VFPRegister::Double};
static constexpr FloatRegister d14 = {FloatRegisters::d14, VFPRegister::Double};
static constexpr FloatRegister d15 = {FloatRegisters::d15, VFPRegister::Double};

// For maximal awesomeness, 8 should be sufficent. ldrd/strd (dual-register
// load/store) operate in a single cycle when the address they are dealing with
// is 8 byte aligned. Also, the ARM abi wants the stack to be 8 byte aligned at
// function boundaries. I'm trying to make sure this is always true.
static constexpr uint32_t ABIStackAlignment = 8;
static constexpr uint32_t CodeAlignment = 8;
static constexpr uint32_t JitStackAlignment = 8;

static constexpr uint32_t JitStackValueAlignment =
    JitStackAlignment / sizeof(Value);
static_assert(JitStackAlignment % sizeof(Value) == 0 &&
                  JitStackValueAlignment >= 1,
              "Stack alignment should be a non-zero multiple of sizeof(Value)");

// This boolean indicates whether we support SIMD instructions flavoured for
// this architecture or not. Rather than a method in the LIRGenerator, it is
// here such that it is accessible from the entire codebase. Once full support
// for SIMD is reached on all tier-1 platforms, this constant can be deleted.
static constexpr bool SupportsSimd = false;
static constexpr uint32_t SimdMemoryAlignment = 8;

static_assert(CodeAlignment % SimdMemoryAlignment == 0,
              "Code alignment should be larger than any of the alignments "
              "which are used for "
              "the constant sections of the code buffer.  Thus it should be "
              "larger than the "
              "alignment for SIMD constants.");

static_assert(JitStackAlignment % SimdMemoryAlignment == 0,
              "Stack alignment should be larger than any of the alignments "
              "which are used for "
              "spilled values.  Thus it should be larger than the alignment "
              "for SIMD accesses.");

static const uint32_t WasmStackAlignment = SimdMemoryAlignment;
static const uint32_t WasmTrapInstructionLength = 4;

// Does this architecture support SIMD conversions between Uint32x4 and
// Float32x4?
static constexpr bool SupportsUint32x4FloatConversions = false;

// Does this architecture support comparisons of unsigned integer vectors?
static constexpr bool SupportsUint8x16Compares = false;
static constexpr bool SupportsUint16x8Compares = false;
static constexpr bool SupportsUint32x4Compares = false;

static const Scale ScalePointer = TimesFour;

class Instruction;
class InstBranchImm;
uint32_t RM(Register r);
uint32_t RS(Register r);
uint32_t RD(Register r);
uint32_t RT(Register r);
uint32_t RN(Register r);

uint32_t maybeRD(Register r);
uint32_t maybeRT(Register r);
uint32_t maybeRN(Register r);

Register toRN(Instruction i);
Register toRM(Instruction i);
Register toRD(Instruction i);
Register toR(Instruction i);

class VFPRegister;
uint32_t VD(VFPRegister vr);
uint32_t VN(VFPRegister vr);
uint32_t VM(VFPRegister vr);

// For being passed into the generic vfp instruction generator when there is an
// instruction that only takes two registers.
static constexpr VFPRegister NoVFPRegister(VFPRegister::Double, 0, false, true);

struct ImmTag : public Imm32 {
  explicit ImmTag(JSValueTag mask) : Imm32(int32_t(mask)) {}
};

struct ImmType : public ImmTag {
  explicit ImmType(JSValueType type) : ImmTag(JSVAL_TYPE_TO_TAG(type)) {}
};

enum Index {
  Offset = 0 << 21 | 1 << 24,
  PreIndex = 1 << 21 | 1 << 24,
  PostIndex = 0 << 21 | 0 << 24
  // The docs were rather unclear on this. It sounds like
  // 1 << 21 | 0 << 24 encodes dtrt.
};

enum IsImmOp2_ { IsImmOp2 = 1 << 25, IsNotImmOp2 = 0 << 25 };
enum IsImmDTR_ { IsImmDTR = 0 << 25, IsNotImmDTR = 1 << 25 };
// For the extra memory operations, ldrd, ldrsb, ldrh.
enum IsImmEDTR_ { IsImmEDTR = 1 << 22, IsNotImmEDTR = 0 << 22 };

enum ShiftType {
  LSL = 0,   // << 5
  LSR = 1,   // << 5
  ASR = 2,   // << 5
  ROR = 3,   // << 5
  RRX = ROR  // RRX is encoded as ROR with a 0 offset.
};

// Modes for STM/LDM. Names are the suffixes applied to the instruction.
enum DTMMode {
  A = 0 << 24,  // empty / after
  B = 1 << 24,  // full / before
  D = 0 << 23,  // decrement
  I = 1 << 23,  // increment
  DA = D | A,
  DB = D | B,
  IA = I | A,
  IB = I | B
};

enum DTMWriteBack { WriteBack = 1 << 21, NoWriteBack = 0 << 21 };

// Condition code updating mode.
enum SBit {
  SetCC = 1 << 20,   // Set condition code.
  LeaveCC = 0 << 20  // Leave condition code unchanged.
};

enum LoadStore { IsLoad = 1 << 20, IsStore = 0 << 20 };

// You almost never want to use this directly. Instead, you wantto pass in a
// signed constant, and let this bit be implicitly set for you. This is however,
// necessary if we want a negative index.
enum IsUp_ { IsUp = 1 << 23, IsDown = 0 << 23 };
enum ALUOp {
  OpMov = 0xd << 21,
  OpMvn = 0xf << 21,
  OpAnd = 0x0 << 21,
  OpBic = 0xe << 21,
  OpEor = 0x1 << 21,
  OpOrr = 0xc << 21,
  OpAdc = 0x5 << 21,
  OpAdd = 0x4 << 21,
  OpSbc = 0x6 << 21,
  OpSub = 0x2 << 21,
  OpRsb = 0x3 << 21,
  OpRsc = 0x7 << 21,
  OpCmn = 0xb << 21,
  OpCmp = 0xa << 21,
  OpTeq = 0x9 << 21,
  OpTst = 0x8 << 21,
  OpInvalid = -1
};

enum MULOp {
  OpmMul = 0 << 21,
  OpmMla = 1 << 21,
  OpmUmaal = 2 << 21,
  OpmMls = 3 << 21,
  OpmUmull = 4 << 21,
  OpmUmlal = 5 << 21,
  OpmSmull = 6 << 21,
  OpmSmlal = 7 << 21
};
enum BranchTag {
  OpB = 0x0a000000,
  OpBMask = 0x0f000000,
  OpBDestMask = 0x00ffffff,
  OpBl = 0x0b000000,
  OpBlx = 0x012fff30,
  OpBx = 0x012fff10
};

// Just like ALUOp, but for the vfp instruction set.
enum VFPOp {
  OpvMul = 0x2 << 20,
  OpvAdd = 0x3 << 20,
  OpvSub = 0x3 << 20 | 0x1 << 6,
  OpvDiv = 0x8 << 20,
  OpvMov = 0xB << 20 | 0x1 << 6,
  OpvAbs = 0xB << 20 | 0x3 << 6,
  OpvNeg = 0xB << 20 | 0x1 << 6 | 0x1 << 16,
  OpvSqrt = 0xB << 20 | 0x3 << 6 | 0x1 << 16,
  OpvCmp = 0xB << 20 | 0x1 << 6 | 0x4 << 16,
  OpvCmpz = 0xB << 20 | 0x1 << 6 | 0x5 << 16
};

// Negate the operation, AND negate the immediate that we were passed in.
ALUOp ALUNeg(ALUOp op, Register dest, Register scratch, Imm32* imm,
             Register* negDest);
bool can_dbl(ALUOp op);
bool condsAreSafe(ALUOp op);

// If there is a variant of op that has a dest (think cmp/sub) return that
// variant of it.
ALUOp getDestVariant(ALUOp op);

static constexpr ValueOperand JSReturnOperand{JSReturnReg_Type,
                                              JSReturnReg_Data};
static const ValueOperand softfpReturnOperand = ValueOperand(r1, r0);

// All of these classes exist solely to shuffle data into the various operands.
// For example Operand2 can be an imm8, a register-shifted-by-a-constant or a
// register-shifted-by-a-register. We represent this in C++ by having a base
// class Operand2, which just stores the 32 bits of data as they will be encoded
// in the instruction. You cannot directly create an Operand2 since it is
// tricky, and not entirely sane to do so. Instead, you create one of its child
// classes, e.g. Imm8. Imm8's constructor takes a single integer argument. Imm8
// will verify that its argument can be encoded as an ARM 12 bit imm8, encode it
// using an Imm8data, and finally call its parent's (Operand2) constructor with
// the Imm8data. The Operand2 constructor will then call the Imm8data's encode()
// function to extract the raw bits from it.
//
// In the future, we should be able to extract data from the Operand2 by asking
// it for its component Imm8data structures. The reason this is so horribly
// round-about is we wanted to have Imm8 and RegisterShiftedRegister inherit
// directly from Operand2 but have all of them take up only a single word of
// storage. We also wanted to avoid passing around raw integers at all since
// they are error prone.
class Op2Reg;
class O2RegImmShift;
class O2RegRegShift;

namespace datastore {

class Reg {
  // The "second register".
  uint32_t rm_ : 4;
  // Do we get another register for shifting.
  uint32_t rrs_ : 1;
  uint32_t type_ : 2;
  // We'd like this to be a more sensible encoding, but that would need to be
  // a struct and that would not pack :(
  uint32_t shiftAmount_ : 5;

 protected:
  // Mark as a protected field to avoid unused private field warnings.
  uint32_t pad_ : 20;

 public:
  Reg(uint32_t rm, ShiftType type, uint32_t rsr, uint32_t shiftAmount)
      : rm_(rm), rrs_(rsr), type_(type), shiftAmount_(shiftAmount), pad_(0) {}
  explicit Reg(const Op2Reg& op) { memcpy(this, &op, sizeof(*this)); }

  uint32_t shiftAmount() const { return shiftAmount_; }

  uint32_t encode() const {
    return rm_ | (rrs_ << 4) | (type_ << 5) | (shiftAmount_ << 7);
  }
};

// Op2 has a mode labelled "<imm8m>", which is arm's magical immediate encoding.
// Some instructions actually get 8 bits of data, which is called Imm8Data
// below. These should have edit distance > 1, but this is how it is for now.
class Imm8mData {
  uint32_t data_ : 8;
  uint32_t rot_ : 4;

 protected:
  // Mark as a protected field to avoid unused private field warnings.
  uint32_t buff_ : 19;

 private:
  // Throw in an extra bit that will be 1 if we can't encode this properly.
  // if we can encode it properly, a simple "|" will still suffice to meld it
  // into the instruction.
  uint32_t invalid_ : 1;

 public:
  // Default constructor makes an invalid immediate.
  Imm8mData() : data_(0xff), rot_(0xf), buff_(0), invalid_(true) {}

  Imm8mData(uint32_t data, uint32_t rot)
      : data_(data), rot_(rot), buff_(0), invalid_(false) {
    MOZ_ASSERT(data == data_);
    MOZ_ASSERT(rot == rot_);
  }

  bool invalid() const { return invalid_; }

  uint32_t encode() const {
    MOZ_ASSERT(!invalid_);
    return data_ | (rot_ << 8);
  };
};

class Imm8Data {
  uint32_t imm4L_ : 4;

 protected:
  // Mark as a protected field to avoid unused private field warnings.
  uint32_t pad_ : 4;

 private:
  uint32_t imm4H_ : 4;

 public:
  explicit Imm8Data(uint32_t imm) : imm4L_(imm & 0xf), imm4H_(imm >> 4) {
    MOZ_ASSERT(imm <= 0xff);
  }

  uint32_t encode() const { return imm4L_ | (imm4H_ << 8); };
};

// VLDR/VSTR take an 8 bit offset, which is implicitly left shifted by 2.
class Imm8VFPOffData {
  uint32_t data_;

 public:
  explicit Imm8VFPOffData(uint32_t imm) : data_(imm) {
    MOZ_ASSERT((imm & ~(0xff)) == 0);
  }
  uint32_t encode() const { return data_; };
};

// ARM can magically encode 256 very special immediates to be moved into a
// register.
struct Imm8VFPImmData {
  // This structure's members are public and it has no constructor to
  // initialize them, for a very special reason. Were this structure to
  // have a constructor, the initialization for DoubleEncoder's internal
  // table (see below) would require a rather large static constructor on
  // some of our supported compilers. The known solution to this is to mark
  // the constructor constexpr, but, again, some of our supported
  // compilers don't support constexpr! So we are reduced to public
  // members and eschewing a constructor in hopes that the initialization
  // of DoubleEncoder's table is correct.
  uint32_t imm4L : 4;
  uint32_t imm4H : 4;
  int32_t isInvalid : 24;

  uint32_t encode() const {
    // This assert is an attempting at ensuring that we don't create random
    // instances of this structure and then asking to encode() it.
    MOZ_ASSERT(isInvalid == 0);
    return imm4L | (imm4H << 16);
  };
};

class Imm12Data {
  uint32_t data_ : 12;

 public:
  explicit Imm12Data(uint32_t imm) : data_(imm) { MOZ_ASSERT(data_ == imm); }

  uint32_t encode() const { return data_; }
};

class RIS {
  uint32_t shiftAmount_ : 5;

 public:
  explicit RIS(uint32_t imm) : shiftAmount_(imm) {
    MOZ_ASSERT(shiftAmount_ == imm);
  }

  explicit RIS(Reg r) : shiftAmount_(r.shiftAmount()) {}

  uint32_t encode() const { return shiftAmount_; }
};

class RRS {
 protected:
  // Mark as a protected field to avoid unused private field warnings.
  uint32_t mustZero_ : 1;

 private:
  // The register that holds the shift amount.
  uint32_t rs_ : 4;

 public:
  explicit RRS(uint32_t rs) : rs_(rs) { MOZ_ASSERT(rs_ == rs); }

  uint32_t encode() const { return rs_ << 1; }
};

}  // namespace datastore

class MacroAssemblerARM;
class Operand;

class Operand2 {
  friend class Operand;
  friend class MacroAssemblerARM;
  friend class InstALU;

  uint32_t oper_ : 31;
  uint32_t invalid_ : 1;

 protected:
  explicit Operand2(datastore::Imm8mData base)
      : oper_(base.invalid() ? -1 : (base.encode() | uint32_t(IsImmOp2))),
        invalid_(base.invalid()) {}

  explicit Operand2(datastore::Reg base)
      : oper_(base.encode() | uint32_t(IsNotImmOp2)), invalid_(false) {}

 private:
  explicit Operand2(uint32_t blob) : oper_(blob), invalid_(false) {}

 public:
  bool isO2Reg() const { return !(oper_ & IsImmOp2); }

  Op2Reg toOp2Reg() const;

  bool isImm8() const { return oper_ & IsImmOp2; }

  bool invalid() const { return invalid_; }

  uint32_t encode() const { return oper_; }
};

class Imm8 : public Operand2 {
 public:
  explicit Imm8(uint32_t imm) : Operand2(EncodeImm(imm)) {}

  static datastore::Imm8mData EncodeImm(uint32_t imm) {
    // RotateLeft below may not be called with a shift of zero.
    if (imm <= 0xFF) {
      return datastore::Imm8mData(imm, 0);
    }

    // An encodable integer has a maximum of 8 contiguous set bits,
    // with an optional wrapped left rotation to even bit positions.
    for (int rot = 1; rot < 16; rot++) {
      uint32_t rotimm = mozilla::RotateLeft(imm, rot * 2);
      if (rotimm <= 0xFF) {
        return datastore::Imm8mData(rotimm, rot);
      }
    }
    return datastore::Imm8mData();
  }

  // Pair template?
  struct TwoImm8mData {
    datastore::Imm8mData fst_, snd_;

    TwoImm8mData() = default;

    TwoImm8mData(datastore::Imm8mData fst, datastore::Imm8mData snd)
        : fst_(fst), snd_(snd) {}

    datastore::Imm8mData fst() const { return fst_; }
    datastore::Imm8mData snd() const { return snd_; }
  };

  static TwoImm8mData EncodeTwoImms(uint32_t);
};

class Op2Reg : public Operand2 {
 public:
  explicit Op2Reg(Register rm, ShiftType type, datastore::RIS shiftImm)
      : Operand2(datastore::Reg(rm.code(), type, 0, shiftImm.encode())) {}

  explicit Op2Reg(Register rm, ShiftType type, datastore::RRS shiftReg)
      : Operand2(datastore::Reg(rm.code(), type, 1, shiftReg.encode())) {}
};

static_assert(sizeof(Op2Reg) == sizeof(datastore::Reg),
              "datastore::Reg(const Op2Reg&) constructor relies on Reg/Op2Reg "
              "having same size");

class O2RegImmShift : public Op2Reg {
 public:
  explicit O2RegImmShift(Register rn, ShiftType type, uint32_t shift)
      : Op2Reg(rn, type, datastore::RIS(shift)) {}
};

class O2RegRegShift : public Op2Reg {
 public:
  explicit O2RegRegShift(Register rn, ShiftType type, Register rs)
      : Op2Reg(rn, type, datastore::RRS(rs.code())) {}
};

O2RegImmShift O2Reg(Register r);
O2RegImmShift lsl(Register r, int amt);
O2RegImmShift lsr(Register r, int amt);
O2RegImmShift asr(Register r, int amt);
O2RegImmShift rol(Register r, int amt);
O2RegImmShift ror(Register r, int amt);

O2RegRegShift lsl(Register r, Register amt);
O2RegRegShift lsr(Register r, Register amt);
O2RegRegShift asr(Register r, Register amt);
O2RegRegShift ror(Register r, Register amt);

// An offset from a register to be used for ldr/str. This should include the
// sign bit, since ARM has "signed-magnitude" offsets. That is it encodes an
// unsigned offset, then the instruction specifies if the offset is positive or
// negative. The +/- bit is necessary if the instruction set wants to be able to
// have a negative register offset e.g. ldr pc, [r1,-r2];
class DtrOff {
  uint32_t data_;

 protected:
  explicit DtrOff(datastore::Imm12Data immdata, IsUp_ iu)
      : data_(immdata.encode() | uint32_t(IsImmDTR) | uint32_t(iu)) {}

  explicit DtrOff(datastore::Reg reg, IsUp_ iu = IsUp)
      : data_(reg.encode() | uint32_t(IsNotImmDTR) | iu) {}

 public:
  uint32_t encode() const { return data_; }
};

class DtrOffImm : public DtrOff {
 public:
  explicit DtrOffImm(int32_t imm)
      : DtrOff(datastore::Imm12Data(mozilla::Abs(imm)),
               imm >= 0 ? IsUp : IsDown) {
    MOZ_ASSERT(mozilla::Abs(imm) < 4096);
  }
};

class DtrOffReg : public DtrOff {
  // These are designed to be called by a constructor of a subclass.
  // Constructing the necessary RIS/RRS structures is annoying.

 protected:
  explicit DtrOffReg(Register rn, ShiftType type, datastore::RIS shiftImm,
                     IsUp_ iu = IsUp)
      : DtrOff(datastore::Reg(rn.code(), type, 0, shiftImm.encode()), iu) {}

  explicit DtrOffReg(Register rn, ShiftType type, datastore::RRS shiftReg,
                     IsUp_ iu = IsUp)
      : DtrOff(datastore::Reg(rn.code(), type, 1, shiftReg.encode()), iu) {}
};

class DtrRegImmShift : public DtrOffReg {
 public:
  explicit DtrRegImmShift(Register rn, ShiftType type, uint32_t shift,
                          IsUp_ iu = IsUp)
      : DtrOffReg(rn, type, datastore::RIS(shift), iu) {}
};

class DtrRegRegShift : public DtrOffReg {
 public:
  explicit DtrRegRegShift(Register rn, ShiftType type, Register rs,
                          IsUp_ iu = IsUp)
      : DtrOffReg(rn, type, datastore::RRS(rs.code()), iu) {}
};

// We will frequently want to bundle a register with its offset so that we have
// an "operand" to a load instruction.
class DTRAddr {
  friend class Operand;

  uint32_t data_;

 public:
  explicit DTRAddr(Register reg, DtrOff dtr)
      : data_(dtr.encode() | (reg.code() << 16)) {}

  uint32_t encode() const { return data_; }

  Register getBase() const { return Register::FromCode((data_ >> 16) & 0xf); }
};

// Offsets for the extended data transfer instructions:
// ldrsh, ldrd, ldrsb, etc.
class EDtrOff {
  uint32_t data_;

 protected:
  explicit EDtrOff(datastore::Imm8Data imm8, IsUp_ iu = IsUp)
      : data_(imm8.encode() | IsImmEDTR | uint32_t(iu)) {}

  explicit EDtrOff(Register rm, IsUp_ iu = IsUp)
      : data_(rm.code() | IsNotImmEDTR | iu) {}

 public:
  uint32_t encode() const { return data_; }
};

class EDtrOffImm : public EDtrOff {
 public:
  explicit EDtrOffImm(int32_t imm)
      : EDtrOff(datastore::Imm8Data(mozilla::Abs(imm)),
                (imm >= 0) ? IsUp : IsDown) {
    MOZ_ASSERT(mozilla::Abs(imm) < 256);
  }
};

// This is the most-derived class, since the extended data transfer instructions
// don't support any sort of modifying the "index" operand.
class EDtrOffReg : public EDtrOff {
 public:
  explicit EDtrOffReg(Register rm) : EDtrOff(rm) {}
};

class EDtrAddr {
  uint32_t data_;

 public:
  explicit EDtrAddr(Register r, EDtrOff off) : data_(RN(r) | off.encode()) {}

  uint32_t encode() const { return data_; }
#ifdef DEBUG
  Register maybeOffsetRegister() const {
    if (data_ & IsImmEDTR) {
      return InvalidReg;
    }
    return Register::FromCode(data_ & 0xf);
  }
#endif
};

class VFPOff {
  uint32_t data_;

 protected:
  explicit VFPOff(datastore::Imm8VFPOffData imm, IsUp_ isup)
      : data_(imm.encode() | uint32_t(isup)) {}

 public:
  uint32_t encode() const { return data_; }
};

class VFPOffImm : public VFPOff {
 public:
  explicit VFPOffImm(int32_t imm)
      : VFPOff(datastore::Imm8VFPOffData(mozilla::Abs(imm) / 4),
               imm < 0 ? IsDown : IsUp) {
    MOZ_ASSERT(mozilla::Abs(imm) <= 255 * 4);
  }
};

class VFPAddr {
  friend class Operand;

  uint32_t data_;

 public:
  explicit VFPAddr(Register base, VFPOff off)
      : data_(RN(base) | off.encode()) {}

  uint32_t encode() const { return data_; }
};

class VFPImm {
  uint32_t data_;

 public:
  explicit VFPImm(uint32_t topWordOfDouble);

  static const VFPImm One;

  uint32_t encode() const { return data_; }
  bool isValid() const { return data_ != (~0U); }
};

// A BOffImm is an immediate that is used for branches. Namely, it is the offset
// that will be encoded in the branch instruction. This is the only sane way of
// constructing a branch.
class BOffImm {
  friend class InstBranchImm;

  uint32_t data_;

 public:
  explicit BOffImm(int offset) : data_((offset - 8) >> 2 & 0x00ffffff) {
    MOZ_ASSERT((offset & 0x3) == 0);
    if (!IsInRange(offset)) {
      MOZ_CRASH("BOffImm offset out of range");
    }
  }

  explicit BOffImm() : data_(INVALID) {}

 private:
  explicit BOffImm(const Instruction& inst);

 public:
  static const uint32_t INVALID = 0x00800000;

  uint32_t encode() const { return data_; }
  int32_t decode() const { return ((int32_t(data_) << 8) >> 6) + 8; }

  static bool IsInRange(int offset) {
    if ((offset - 8) < -33554432) {
      return false;
    }
    if ((offset - 8) > 33554428) {
      return false;
    }
    return true;
  }

  bool isInvalid() const { return data_ == INVALID; }
  Instruction* getDest(Instruction* src) const;
};

class Imm16 {
  uint32_t lower_ : 12;

 protected:
  // Mark as a protected field to avoid unused private field warnings.
  uint32_t pad_ : 4;

 private:
  uint32_t upper_ : 4;
  uint32_t invalid_ : 12;

 public:
  explicit Imm16();
  explicit Imm16(uint32_t imm);
  explicit Imm16(Instruction& inst);

  uint32_t encode() const { return lower_ | (upper_ << 16); }
  uint32_t decode() const { return lower_ | (upper_ << 12); }

  bool isInvalid() const { return invalid_; }
};

// I would preffer that these do not exist, since there are essentially no
// instructions that would ever take more than one of these, however, the MIR
// wants to only have one type of arguments to functions, so bugger.
class Operand {
  // The encoding of registers is the same for OP2, DTR and EDTR yet the type
  // system doesn't let us express this, so choices must be made.
 public:
  enum class Tag : uint8_t { OP2, MEM, FOP };

 private:
  uint32_t tag_ : 8;
  uint32_t reg_ : 5;
  int32_t offset_;

 protected:
  Operand(Tag tag, uint32_t regCode, int32_t offset)
      : tag_(static_cast<uint32_t>(tag)), reg_(regCode), offset_(offset) {}

 public:
  explicit Operand(Register reg) : Operand(Tag::OP2, reg.code(), 0) {}

  explicit Operand(FloatRegister freg) : Operand(Tag::FOP, freg.code(), 0) {}

  explicit Operand(Register base, Imm32 off)
      : Operand(Tag::MEM, base.code(), off.value) {}

  explicit Operand(Register base, int32_t off)
      : Operand(Tag::MEM, base.code(), off) {}

  explicit Operand(const Address& addr)
      : Operand(Tag::MEM, addr.base.code(), addr.offset) {}

 public:
  Tag tag() const { return static_cast<Tag>(tag_); }

  Operand2 toOp2() const {
    MOZ_ASSERT(tag() == Tag::OP2);
    return O2Reg(Register::FromCode(reg_));
  }

  Register toReg() const {
    MOZ_ASSERT(tag() == Tag::OP2);
    return Register::FromCode(reg_);
  }

  Address toAddress() const {
    MOZ_ASSERT(tag() == Tag::MEM);
    return Address(Register::FromCode(reg_), offset_);
  }
  int32_t disp() const {
    MOZ_ASSERT(tag() == Tag::MEM);
    return offset_;
  }

  int32_t base() const {
    MOZ_ASSERT(tag() == Tag::MEM);
    return reg_;
  }
  Register baseReg() const {
    MOZ_ASSERT(tag() == Tag::MEM);
    return Register::FromCode(reg_);
  }
  DTRAddr toDTRAddr() const {
    MOZ_ASSERT(tag() == Tag::MEM);
    return DTRAddr(baseReg(), DtrOffImm(offset_));
  }
  VFPAddr toVFPAddr() const {
    MOZ_ASSERT(tag() == Tag::MEM);
    return VFPAddr(baseReg(), VFPOffImm(offset_));
  }
};

inline Imm32 Imm64::firstHalf() const { return low(); }

inline Imm32 Imm64::secondHalf() const { return hi(); }

void PatchJump(CodeLocationJump& jump_, CodeLocationLabel label);

class InstructionIterator {
 private:
  Instruction* inst_;

 public:
  explicit InstructionIterator(Instruction* inst) : inst_(inst) {
    maybeSkipAutomaticInstructions();
  }

  // Advances to the next intentionally-inserted instruction.
  Instruction* next();

  // Advances past any automatically-inserted instructions.
  Instruction* maybeSkipAutomaticInstructions();

  Instruction* cur() const { return inst_; }

 protected:
  // Advances past the given number of instruction-length bytes.
  inline void advanceRaw(ptrdiff_t instructions = 1);
};

class Assembler;
typedef js::jit::AssemblerBufferWithConstantPools<1024, 4, Instruction,
                                                  Assembler>
    ARMBuffer;

class Assembler : public AssemblerShared {
 public:
  // ARM conditional constants:
  enum ARMCondition : uint32_t {
    EQ = 0x00000000,  // Zero
    NE = 0x10000000,  // Non-zero
    CS = 0x20000000,
    CC = 0x30000000,
    MI = 0x40000000,
    PL = 0x50000000,
    VS = 0x60000000,
    VC = 0x70000000,
    HI = 0x80000000,
    LS = 0x90000000,
    GE = 0xa0000000,
    LT = 0xb0000000,
    GT = 0xc0000000,
    LE = 0xd0000000,
    AL = 0xe0000000
  };

  enum Condition : uint32_t {
    Equal = EQ,
    NotEqual = NE,
    Above = HI,
    AboveOrEqual = CS,
    Below = CC,
    BelowOrEqual = LS,
    GreaterThan = GT,
    GreaterThanOrEqual = GE,
    LessThan = LT,
    LessThanOrEqual = LE,
    Overflow = VS,
    CarrySet = CS,
    CarryClear = CC,
    Signed = MI,
    NotSigned = PL,
    Zero = EQ,
    NonZero = NE,
    Always = AL,

    VFP_NotEqualOrUnordered = NE,
    VFP_Equal = EQ,
    VFP_Unordered = VS,
    VFP_NotUnordered = VC,
    VFP_GreaterThanOrEqualOrUnordered = CS,
    VFP_GreaterThanOrEqual = GE,
    VFP_GreaterThanOrUnordered = HI,
    VFP_GreaterThan = GT,
    VFP_LessThanOrEqualOrUnordered = LE,
    VFP_LessThanOrEqual = LS,
    VFP_LessThanOrUnordered = LT,
    VFP_LessThan = CC  // MI is valid too.
  };

  // Bit set when a DoubleCondition does not map to a single ARM condition.
  // The macro assembler has to special-case these conditions, or else
  // ConditionFromDoubleCondition will complain.
  static const int DoubleConditionBitSpecial = 0x1;

  enum DoubleCondition : uint32_t {
    // These conditions will only evaluate to true if the comparison is
    // ordered - i.e. neither operand is NaN.
    DoubleOrdered = VFP_NotUnordered,
    DoubleEqual = VFP_Equal,
    DoubleNotEqual = VFP_NotEqualOrUnordered | DoubleConditionBitSpecial,
    DoubleGreaterThan = VFP_GreaterThan,
    DoubleGreaterThanOrEqual = VFP_GreaterThanOrEqual,
    DoubleLessThan = VFP_LessThan,
    DoubleLessThanOrEqual = VFP_LessThanOrEqual,
    // If either operand is NaN, these conditions always evaluate to true.
    DoubleUnordered = VFP_Unordered,
    DoubleEqualOrUnordered = VFP_Equal | DoubleConditionBitSpecial,
    DoubleNotEqualOrUnordered = VFP_NotEqualOrUnordered,
    DoubleGreaterThanOrUnordered = VFP_GreaterThanOrUnordered,
    DoubleGreaterThanOrEqualOrUnordered = VFP_GreaterThanOrEqualOrUnordered,
    DoubleLessThanOrUnordered = VFP_LessThanOrUnordered,
    DoubleLessThanOrEqualOrUnordered = VFP_LessThanOrEqualOrUnordered
  };

  Condition getCondition(uint32_t inst) {
    return (Condition)(0xf0000000 & inst);
  }
  static inline Condition ConditionFromDoubleCondition(DoubleCondition cond) {
    MOZ_ASSERT(!(cond & DoubleConditionBitSpecial));
    return static_cast<Condition>(cond);
  }

  enum BarrierOption {
    BarrierSY = 15,  // Full system barrier
    BarrierST = 14   // StoreStore barrier
  };

  // This should be protected, but since CodeGenerator wants to use it, it
  // needs to go out here :(

  BufferOffset nextOffset() { return m_buffer.nextOffset(); }

 protected:
  // Shim around AssemblerBufferWithConstantPools::allocEntry.
  BufferOffset allocLiteralLoadEntry(size_t numInst, unsigned numPoolEntries,
                                     PoolHintPun& php, uint8_t* data,
                                     const LiteralDoc& doc = LiteralDoc(),
                                     ARMBuffer::PoolEntry* pe = nullptr,
                                     bool loadToPC = false);

  Instruction* editSrc(BufferOffset bo) { return m_buffer.getInst(bo); }

#ifdef JS_DISASM_ARM
  typedef disasm::EmbeddedVector<char, disasm::ReasonableBufferSize>
      DisasmBuffer;

  static void disassembleInstruction(const Instruction* i,
                                     DisasmBuffer& buffer);

  void initDisassembler();
  void finishDisassembler();
  void spew(Instruction* i);
  void spewBranch(Instruction* i, const LabelDoc& target);
  void spewLiteralLoad(PoolHintPun& php, bool loadToPC, const Instruction* offs,
                       const LiteralDoc& doc);
#endif

 public:
  void resetCounter();
  uint32_t actualIndex(uint32_t) const;
  static uint8_t* PatchableJumpAddress(JitCode* code, uint32_t index);
  static uint32_t NopFill;
  static uint32_t GetNopFill();
  static uint32_t AsmPoolMaxOffset;
  static uint32_t GetPoolMaxOffset();

 protected:
  // Structure for fixing up pc-relative loads/jumps when a the machine code
  // gets moved (executable copy, gc, etc.).
  class RelativePatch {
    void* target_;
    RelocationKind kind_;

   public:
    RelativePatch(void* target, RelocationKind kind)
        : target_(target), kind_(kind) {}
    void* target() const { return target_; }
    RelocationKind kind() const { return kind_; }
  };

  // TODO: this should actually be a pool-like object. It is currently a big
  // hack, and probably shouldn't exist.
  js::Vector<RelativePatch, 8, SystemAllocPolicy> jumps_;

  CompactBufferWriter jumpRelocations_;
  CompactBufferWriter dataRelocations_;

  ARMBuffer m_buffer;

#ifdef JS_DISASM_ARM
  DisassemblerSpew spew_;
#endif

 public:
  // For the alignment fill use NOP: 0x0320f000 or (Always | InstNOP::NopInst).
  // For the nopFill use a branch to the next instruction: 0xeaffffff.
  Assembler()
      : m_buffer(1, 1, 8, GetPoolMaxOffset(), 8, 0xe320f000, 0xeaffffff,
                 GetNopFill()),
        isFinished(false),
        dtmActive(false),
        dtmCond(Always) {
#ifdef JS_DISASM_ARM
    initDisassembler();
#endif
  }

  ~Assembler() {
#ifdef JS_DISASM_ARM
    finishDisassembler();
#endif
  }

  // We need to wait until an AutoJitContextAlloc is created by the
  // MacroAssembler, before allocating any space.
  void initWithAllocator() { m_buffer.initWithAllocator(); }

  void setUnlimitedBuffer() { m_buffer.setUnlimited(); }

  static Condition InvertCondition(Condition cond);
  static Condition UnsignedCondition(Condition cond);
  static Condition ConditionWithoutEqual(Condition cond);

  static DoubleCondition InvertCondition(DoubleCondition cond);

  void writeRelocation(BufferOffset src) {
    jumpRelocations_.writeUnsigned(src.getOffset());
  }

  void writeDataRelocation(BufferOffset offset, ImmGCPtr ptr) {
    if (ptr.value) {
      if (gc::IsInsideNursery(ptr.value)) {
        embedsNurseryPointers_ = true;
      }
      dataRelocations_.writeUnsigned(offset.getOffset());
    }
  }

  enum RelocBranchStyle { B_MOVWT, B_LDR_BX, B_LDR, B_MOVW_ADD };

  enum RelocStyle { L_MOVWT, L_LDR };

 public:
  // Given the start of a Control Flow sequence, grab the value that is
  // finally branched to given the start of a function that loads an address
  // into a register get the address that ends up in the register.
  template <class Iter>
  static const uint32_t* GetCF32Target(Iter* iter);

  static uintptr_t GetPointer(uint8_t*);
  static const uint32_t* GetPtr32Target(InstructionIterator iter,
                                        Register* dest = nullptr,
                                        RelocStyle* rs = nullptr);

  bool oom() const;

  void setPrinter(Sprinter* sp) {
#ifdef JS_DISASM_ARM
    spew_.setPrinter(sp);
#endif
  }

  static const Register getStackPointer() { return StackPointer; }

 private:
  bool isFinished;

 protected:
  LabelDoc refLabel(const Label* label) {
#ifdef JS_DISASM_ARM
    return spew_.refLabel(label);
#else
    return LabelDoc();
#endif
  }

 public:
  void finish();
  bool appendRawCode(const uint8_t* code, size_t numBytes);
  bool reserve(size_t size);
  bool swapBuffer(wasm::Bytes& bytes);
  void copyJumpRelocationTable(uint8_t* dest);
  void copyDataRelocationTable(uint8_t* dest);

  // Size of the instruction stream, in bytes, after pools are flushed.
  size_t size() const;
  // Size of the jump relocation table, in bytes.
  size_t jumpRelocationTableBytes() const;
  size_t dataRelocationTableBytes() const;

  // Size of the data table, in bytes.
  size_t bytesNeeded() const;

  // Write a single instruction into the instruction stream.  Very hot,
  // inlined for performance
  MOZ_ALWAYS_INLINE BufferOffset writeInst(uint32_t x) {
    BufferOffset offs = m_buffer.putInt(x);
#ifdef JS_DISASM_ARM
    spew(m_buffer.getInstOrNull(offs));
#endif
    return offs;
  }

  // As above, but also mark the instruction as a branch.  Very hot, inlined
  // for performance
  MOZ_ALWAYS_INLINE BufferOffset
  writeBranchInst(uint32_t x, const LabelDoc& documentation) {
    BufferOffset offs = m_buffer.putInt(x);
#ifdef JS_DISASM_ARM
    spewBranch(m_buffer.getInstOrNull(offs), documentation);
#endif
    return offs;
  }

  // Write a placeholder NOP for a branch into the instruction stream
  // (in order to adjust assembler addresses and mark it as a branch), it will
  // be overwritten subsequently.
  BufferOffset allocBranchInst();

  // A static variant for the cases where we don't want to have an assembler
  // object.
  static void WriteInstStatic(uint32_t x, uint32_t* dest);

 public:
  void writeCodePointer(CodeLabel* label);

  void haltingAlign(int alignment);
  void nopAlign(int alignment);
  BufferOffset as_nop();
  BufferOffset as_alu(Register dest, Register src1, Operand2 op2, ALUOp op,
                      SBit s = LeaveCC, Condition c = Always);
  BufferOffset as_mov(Register dest, Operand2 op2, SBit s = LeaveCC,
                      Condition c = Always);
  BufferOffset as_mvn(Register dest, Operand2 op2, SBit s = LeaveCC,
                      Condition c = Always);

  static void as_alu_patch(Register dest, Register src1, Operand2 op2, ALUOp op,
                           SBit s, Condition c, uint32_t* pos);
  static void as_mov_patch(Register dest, Operand2 op2, SBit s, Condition c,
                           uint32_t* pos);

  // Logical operations:
  BufferOffset as_and(Register dest, Register src1, Operand2 op2,
                      SBit s = LeaveCC, Condition c = Always);
  BufferOffset as_bic(Register dest, Register src1, Operand2 op2,
                      SBit s = LeaveCC, Condition c = Always);
  BufferOffset as_eor(Register dest, Register src1, Operand2 op2,
                      SBit s = LeaveCC, Condition c = Always);
  BufferOffset as_orr(Register dest, Register src1, Operand2 op2,
                      SBit s = LeaveCC, Condition c = Always);
  // Mathematical operations:
  BufferOffset as_adc(Register dest, Register src1, Operand2 op2,
                      SBit s = LeaveCC, Condition c = Always);
  BufferOffset as_add(Register dest, Register src1, Operand2 op2,
                      SBit s = LeaveCC, Condition c = Always);
  BufferOffset as_sbc(Register dest, Register src1, Operand2 op2,
                      SBit s = LeaveCC, Condition c = Always);
  BufferOffset as_sub(Register dest, Register src1, Operand2 op2,
                      SBit s = LeaveCC, Condition c = Always);
  BufferOffset as_rsb(Register dest, Register src1, Operand2 op2,
                      SBit s = LeaveCC, Condition c = Always);
  BufferOffset as_rsc(Register dest, Register src1, Operand2 op2,
                      SBit s = LeaveCC, Condition c = Always);
  // Test operations:
  BufferOffset as_cmn(Register src1, Operand2 op2, Condition c = Always);
  BufferOffset as_cmp(Register src1, Operand2 op2, Condition c = Always);
  BufferOffset as_teq(Register src1, Operand2 op2, Condition c = Always);
  BufferOffset as_tst(Register src1, Operand2 op2, Condition c = Always);

  // Sign extension operations:
  BufferOffset as_sxtb(Register dest, Register src, int rotate,
                       Condition c = Always);
  BufferOffset as_sxth(Register dest, Register src, int rotate,
                       Condition c = Always);
  BufferOffset as_uxtb(Register dest, Register src, int rotate,
                       Condition c = Always);
  BufferOffset as_uxth(Register dest, Register src, int rotate,
                       Condition c = Always);

  // Not quite ALU worthy, but useful none the less: These also have the issue
  // of these being formatted completly differently from the standard ALU
  // operations.
  BufferOffset as_movw(Register dest, Imm16 imm, Condition c = Always);
  BufferOffset as_movt(Register dest, Imm16 imm, Condition c = Always);

  static void as_movw_patch(Register dest, Imm16 imm, Condition c,
                            Instruction* pos);
  static void as_movt_patch(Register dest, Imm16 imm, Condition c,
                            Instruction* pos);

  BufferOffset as_genmul(Register d1, Register d2, Register rm, Register rn,
                         MULOp op, SBit s, Condition c = Always);
  BufferOffset as_mul(Register dest, Register src1, Register src2,
                      SBit s = LeaveCC, Condition c = Always);
  BufferOffset as_mla(Register dest, Register acc, Register src1, Register src2,
                      SBit s = LeaveCC, Condition c = Always);
  BufferOffset as_umaal(Register dest1, Register dest2, Register src1,
                        Register src2, Condition c = Always);
  BufferOffset as_mls(Register dest, Register acc, Register src1, Register src2,
                      Condition c = Always);
  BufferOffset as_umull(Register dest1, Register dest2, Register src1,
                        Register src2, SBit s = LeaveCC, Condition c = Always);
  BufferOffset as_umlal(Register dest1, Register dest2, Register src1,
                        Register src2, SBit s = LeaveCC, Condition c = Always);
  BufferOffset as_smull(Register dest1, Register dest2, Register src1,
                        Register src2, SBit s = LeaveCC, Condition c = Always);
  BufferOffset as_smlal(Register dest1, Register dest2, Register src1,
                        Register src2, SBit s = LeaveCC, Condition c = Always);

  BufferOffset as_sdiv(Register dest, Register num, Register div,
                       Condition c = Always);
  BufferOffset as_udiv(Register dest, Register num, Register div,
                       Condition c = Always);
  BufferOffset as_clz(Register dest, Register src, Condition c = Always);

  // Data transfer instructions: ldr, str, ldrb, strb.
  // Using an int to differentiate between 8 bits and 32 bits is overkill.
  BufferOffset as_dtr(LoadStore ls, int size, Index mode, Register rt,
                      DTRAddr addr, Condition c = Always);

  static void as_dtr_patch(LoadStore ls, int size, Index mode, Register rt,
                           DTRAddr addr, Condition c, uint32_t* dest);

  // Handles all of the other integral data transferring functions:
  // ldrsb, ldrsh, ldrd, etc. The size is given in bits.
  BufferOffset as_extdtr(LoadStore ls, int size, bool IsSigned, Index mode,
                         Register rt, EDtrAddr addr, Condition c = Always);

  BufferOffset as_dtm(LoadStore ls, Register rn, uint32_t mask, DTMMode mode,
                      DTMWriteBack wb, Condition c = Always);

  // Overwrite a pool entry with new data.
  static void WritePoolEntry(Instruction* addr, Condition c, uint32_t data);

  // Load a 32 bit immediate from a pool into a register.
  BufferOffset as_Imm32Pool(Register dest, uint32_t value,
                            Condition c = Always);
  // Make a patchable jump that can target the entire 32 bit address space.
  BufferOffset as_BranchPool(uint32_t value, RepatchLabel* label,
                             const LabelDoc& documentation,
                             ARMBuffer::PoolEntry* pe = nullptr,
                             Condition c = Always);

  // Load a 64 bit floating point immediate from a pool into a register.
  BufferOffset as_FImm64Pool(VFPRegister dest, double value,
                             Condition c = Always);
  // Load a 32 bit floating point immediate from a pool into a register.
  BufferOffset as_FImm32Pool(VFPRegister dest, float value,
                             Condition c = Always);

  // Atomic instructions: ldrexd, ldrex, ldrexh, ldrexb, strexd, strex, strexh,
  // strexb.
  //
  // The doubleword, halfword, and byte versions are available from ARMv6K
  // forward.
  //
  // The word versions are available from ARMv6 forward and can be used to
  // implement the halfword and byte versions on older systems.

  // LDREXD rt, rt2, [rn].  Constraint: rt even register, rt2=rt+1.
  BufferOffset as_ldrexd(Register rt, Register rt2, Register rn,
                         Condition c = Always);

  // LDREX rt, [rn]
  BufferOffset as_ldrex(Register rt, Register rn, Condition c = Always);
  BufferOffset as_ldrexh(Register rt, Register rn, Condition c = Always);
  BufferOffset as_ldrexb(Register rt, Register rn, Condition c = Always);

  // STREXD rd, rt, rt2, [rn].  Constraint: rt even register, rt2=rt+1.
  BufferOffset as_strexd(Register rd, Register rt, Register rt2, Register rn,
                         Condition c = Always);

  // STREX rd, rt, [rn].  Constraint: rd != rn, rd != rt.
  BufferOffset as_strex(Register rd, Register rt, Register rn,
                        Condition c = Always);
  BufferOffset as_strexh(Register rd, Register rt, Register rn,
                         Condition c = Always);
  BufferOffset as_strexb(Register rd, Register rt, Register rn,
                         Condition c = Always);

  // CLREX
  BufferOffset as_clrex();

  // Memory synchronization.
  // These are available from ARMv7 forward.
  BufferOffset as_dmb(BarrierOption option = BarrierSY);
  BufferOffset as_dsb(BarrierOption option = BarrierSY);
  BufferOffset as_isb();

  // Memory synchronization for architectures before ARMv7.
  BufferOffset as_dsb_trap();
  BufferOffset as_dmb_trap();
  BufferOffset as_isb_trap();

  // Speculation barrier
  BufferOffset as_csdb();

  // Control flow stuff:

  // bx can *only* branch to a register never to an immediate.
  BufferOffset as_bx(Register r, Condition c = Always);

  // Branch can branch to an immediate *or* to a register. Branches to
  // immediates are pc relative, branches to registers are absolute.
  BufferOffset as_b(BOffImm off, Condition c, Label* documentation = nullptr);

  BufferOffset as_b(Label* l, Condition c = Always);
  BufferOffset as_b(BOffImm off, Condition c, BufferOffset inst);

  // blx can go to either an immediate or a register. When blx'ing to a
  // register, we change processor mode depending on the low bit of the
  // register when blx'ing to an immediate, we *always* change processor
  // state.
  BufferOffset as_blx(Label* l);

  BufferOffset as_blx(Register r, Condition c = Always);
  BufferOffset as_bl(BOffImm off, Condition c, Label* documentation = nullptr);
  // bl can only branch+link to an immediate, never to a register it never
  // changes processor state.
  BufferOffset as_bl();
  // bl #imm can have a condition code, blx #imm cannot.
  // blx reg can be conditional.
  BufferOffset as_bl(Label* l, Condition c);
  BufferOffset as_bl(BOffImm off, Condition c, BufferOffset inst);

  BufferOffset as_mrs(Register r, Condition c = Always);
  BufferOffset as_msr(Register r, Condition c = Always);

  // VFP instructions!
 private:
  enum vfp_size { IsDouble = 1 << 8, IsSingle = 0 << 8 };

  BufferOffset writeVFPInst(vfp_size sz, uint32_t blob);

  static void WriteVFPInstStatic(vfp_size sz, uint32_t blob, uint32_t* dest);

  // Unityped variants: all registers hold the same (ieee754 single/double)
  // notably not included are vcvt; vmov vd, #imm; vmov rt, vn.
  BufferOffset as_vfp_float(VFPRegister vd, VFPRegister vn, VFPRegister vm,
                            VFPOp op, Condition c = Always);

 public:
  BufferOffset as_vadd(VFPRegister vd, VFPRegister vn, VFPRegister vm,
                       Condition c = Always);
  BufferOffset as_vdiv(VFPRegister vd, VFPRegister vn, VFPRegister vm,
                       Condition c = Always);
  BufferOffset as_vmul(VFPRegister vd, VFPRegister vn, VFPRegister vm,
                       Condition c = Always);
  BufferOffset as_vnmul(VFPRegister vd, VFPRegister vn, VFPRegister vm,
                        Condition c = Always);
  BufferOffset as_vnmla(VFPRegister vd, VFPRegister vn, VFPRegister vm,
                        Condition c = Always);
  BufferOffset as_vnmls(VFPRegister vd, VFPRegister vn, VFPRegister vm,
                        Condition c = Always);
  BufferOffset as_vneg(VFPRegister vd, VFPRegister vm, Condition c = Always);
  BufferOffset as_vsqrt(VFPRegister vd, VFPRegister vm, Condition c = Always);
  BufferOffset as_vabs(VFPRegister vd, VFPRegister vm, Condition c = Always);
  BufferOffset as_vsub(VFPRegister vd, VFPRegister vn, VFPRegister vm,
                       Condition c = Always);
  BufferOffset as_vcmp(VFPRegister vd, VFPRegister vm, Condition c = Always);
  BufferOffset as_vcmpz(VFPRegister vd, Condition c = Always);

  // Specifically, a move between two same sized-registers.
  BufferOffset as_vmov(VFPRegister vd, VFPRegister vsrc, Condition c = Always);

  // Transfer between Core and VFP.
  enum FloatToCore_ { FloatToCore = 1 << 20, CoreToFloat = 0 << 20 };

 private:
  enum VFPXferSize { WordTransfer = 0x02000010, DoubleTransfer = 0x00400010 };

 public:
  // Unlike the next function, moving between the core registers and vfp
  // registers can't be *that* properly typed. Namely, since I don't want to
  // munge the type VFPRegister to also include core registers. Thus, the core
  // and vfp registers are passed in based on their type, and src/dest is
  // determined by the float2core.

  BufferOffset as_vxfer(Register vt1, Register vt2, VFPRegister vm,
                        FloatToCore_ f2c, Condition c = Always, int idx = 0);

  // Our encoding actually allows just the src and the dest (and their types)
  // to uniquely specify the encoding that we are going to use.
  BufferOffset as_vcvt(VFPRegister vd, VFPRegister vm, bool useFPSCR = false,
                       Condition c = Always);

  // Hard coded to a 32 bit fixed width result for now.
  BufferOffset as_vcvtFixed(VFPRegister vd, bool isSigned, uint32_t fixedPoint,
                            bool toFixed, Condition c = Always);

  // Transfer between VFP and memory.
  BufferOffset as_vdtr(LoadStore ls, VFPRegister vd, VFPAddr addr,
                       Condition c = Always /* vfp doesn't have a wb option*/);

  static void as_vdtr_patch(LoadStore ls, VFPRegister vd, VFPAddr addr,
                            Condition c /* vfp doesn't have a wb option */,
                            uint32_t* dest);

  // VFP's ldm/stm work differently from the standard arm ones. You can only
  // transfer a range.

  BufferOffset as_vdtm(LoadStore st, Register rn, VFPRegister vd, int length,
                       /* also has update conditions */ Condition c = Always);

  BufferOffset as_vimm(VFPRegister vd, VFPImm imm, Condition c = Always);

  BufferOffset as_vmrs(Register r, Condition c = Always);
  BufferOffset as_vmsr(Register r, Condition c = Always);

  // Label operations.
  bool nextLink(BufferOffset b, BufferOffset* next);
  void bind(Label* label, BufferOffset boff = BufferOffset());
  void bind(RepatchLabel* label);
  uint32_t currentOffset() { return nextOffset().getOffset(); }
  void retarget(Label* label, Label* target);
  // I'm going to pretend this doesn't exist for now.
  void retarget(Label* label, void* target, RelocationKind reloc);

  static void Bind(uint8_t* rawCode, const CodeLabel& label);

  void as_bkpt();
  BufferOffset as_illegal_trap();

 public:
  static void TraceJumpRelocations(JSTracer* trc, JitCode* code,
                                   CompactBufferReader& reader);
  static void TraceDataRelocations(JSTracer* trc, JitCode* code,
                                   CompactBufferReader& reader);

  void assertNoGCThings() const {
#ifdef DEBUG
    MOZ_ASSERT(dataRelocations_.length() == 0);
    for (auto& j : jumps_) {
      MOZ_ASSERT(j.kind() == RelocationKind::HARDCODED);
    }
#endif
  }

  static bool SupportsFloatingPoint() { return HasVFP(); }
  static bool SupportsUnalignedAccesses() { return HasARMv7(); }
  static bool SupportsSimd() { return js::jit::SupportsSimd; }

  static bool HasRoundInstruction(RoundingMode mode) { return false; }

 protected:
  void addPendingJump(BufferOffset src, ImmPtr target, RelocationKind kind) {
    enoughMemory_ &= jumps_.append(RelativePatch(target.value, kind));
    if (kind == RelocationKind::JITCODE) {
      writeRelocation(src);
    }
  }

 public:
  // The buffer is about to be linked, make sure any constant pools or excess
  // bookkeeping has been flushed to the instruction stream.
  void flush() {
    MOZ_ASSERT(!isFinished);
    m_buffer.flushPool();
    return;
  }

  void comment(const char* msg) {
#ifdef JS_DISASM_ARM
    spew_.spew("; %s", msg);
#endif
  }

  // Copy the assembly code to the given buffer, and perform any pending
  // relocations relying on the target address.
  void executableCopy(uint8_t* buffer, bool flushICache = true);

  // Actual assembly emitting functions.

  // Since I can't think of a reasonable default for the mode, I'm going to
  // leave it as a required argument.
  void startDataTransferM(LoadStore ls, Register rm, DTMMode mode,
                          DTMWriteBack update = NoWriteBack,
                          Condition c = Always) {
    MOZ_ASSERT(!dtmActive);
    dtmUpdate = update;
    dtmBase = rm;
    dtmLoadStore = ls;
    dtmLastReg = -1;
    dtmRegBitField = 0;
    dtmActive = 1;
    dtmCond = c;
    dtmMode = mode;
  }

  void transferReg(Register rn) {
    MOZ_ASSERT(dtmActive);
    MOZ_ASSERT(rn.code() > dtmLastReg);
    dtmRegBitField |= 1 << rn.code();
    if (dtmLoadStore == IsLoad && rn.code() == 13 && dtmBase.code() == 13) {
      MOZ_CRASH("ARM Spec says this is invalid");
    }
  }
  void finishDataTransfer() {
    dtmActive = false;
    as_dtm(dtmLoadStore, dtmBase, dtmRegBitField, dtmMode, dtmUpdate, dtmCond);
  }

  void startFloatTransferM(LoadStore ls, Register rm, DTMMode mode,
                           DTMWriteBack update = NoWriteBack,
                           Condition c = Always) {
    MOZ_ASSERT(!dtmActive);
    dtmActive = true;
    dtmUpdate = update;
    dtmLoadStore = ls;
    dtmBase = rm;
    dtmCond = c;
    dtmLastReg = -1;
    dtmMode = mode;
    dtmDelta = 0;
  }
  void transferFloatReg(VFPRegister rn) {
    if (dtmLastReg == -1) {
      vdtmFirstReg = rn.code();
    } else {
      if (dtmDelta == 0) {
        dtmDelta = rn.code() - dtmLastReg;
        MOZ_ASSERT(dtmDelta == 1 || dtmDelta == -1);
      }
      MOZ_ASSERT(dtmLastReg >= 0);
      MOZ_ASSERT(rn.code() == unsigned(dtmLastReg) + dtmDelta);
    }

    dtmLastReg = rn.code();
  }
  void finishFloatTransfer() {
    MOZ_ASSERT(dtmActive);
    dtmActive = false;
    MOZ_ASSERT(dtmLastReg != -1);
    dtmDelta = dtmDelta ? dtmDelta : 1;
    // The operand for the vstr/vldr instruction is the lowest register in the
    // range.
    int low = Min(dtmLastReg, vdtmFirstReg);
    int high = Max(dtmLastReg, vdtmFirstReg);
    // Fencepost problem.
    int len = high - low + 1;
    // vdtm can only transfer 16 registers at once.  If we need to transfer
    // more, then either hoops are necessary, or we need to be updating the
    // register.
    MOZ_ASSERT_IF(len > 16, dtmUpdate == WriteBack);

    int adjustLow = dtmLoadStore == IsStore ? 0 : 1;
    int adjustHigh = dtmLoadStore == IsStore ? -1 : 0;
    while (len > 0) {
      // Limit the instruction to 16 registers.
      int curLen = Min(len, 16);
      // If it is a store, we want to start at the high end and move down
      // (e.g. vpush d16-d31; vpush d0-d15).
      int curStart = (dtmLoadStore == IsStore) ? high - curLen + 1 : low;
      as_vdtm(dtmLoadStore, dtmBase,
              VFPRegister(FloatRegister::FromCode(curStart)), curLen, dtmCond);
      // Update the bounds.
      low += adjustLow * curLen;
      high += adjustHigh * curLen;
      // Update the length parameter.
      len -= curLen;
    }
  }

 private:
  int dtmRegBitField;
  int vdtmFirstReg;
  int dtmLastReg;
  int dtmDelta;
  Register dtmBase;
  DTMWriteBack dtmUpdate;
  DTMMode dtmMode;
  LoadStore dtmLoadStore;
  bool dtmActive;
  Condition dtmCond;

 public:
  enum {
    PadForAlign8 = (int)0x00,
    PadForAlign16 = (int)0x0000,
    PadForAlign32 = (int)0xe12fff7f  // 'bkpt 0xffff'
  };

  // API for speaking with the IonAssemblerBufferWithConstantPools generate an
  // initial placeholder instruction that we want to later fix up.
  static void InsertIndexIntoTag(uint8_t* load, uint32_t index);

  // Take the stub value that was written in before, and write in an actual
  // load using the index we'd computed previously as well as the address of
  // the pool start.
  static void PatchConstantPoolLoad(void* loadAddr, void* constPoolAddr);

  // We're not tracking short-range branches for ARM for now.
  static void PatchShortRangeBranchToVeneer(ARMBuffer*, unsigned rangeIdx,
                                            BufferOffset deadline,
                                            BufferOffset veneer) {
    MOZ_CRASH();
  }
  // END API

  // Move our entire pool into the instruction stream. This is to force an
  // opportunistic dump of the pool, prefferably when it is more convenient to
  // do a dump.
  void flushBuffer();
  void enterNoPool(size_t maxInst);
  void leaveNoPool();
  void enterNoNops();
  void leaveNoNops();
  // This should return a BOffImm, but we didn't want to require everyplace
  // that used the AssemblerBuffer to make that class.
  static ptrdiff_t GetBranchOffset(const Instruction* i);
  static void RetargetNearBranch(Instruction* i, int offset, Condition cond,
                                 bool final = true);
  static void RetargetNearBranch(Instruction* i, int offset, bool final = true);
  static void RetargetFarBranch(Instruction* i, uint8_t** slot, uint8_t* dest,
                                Condition cond);

  static void WritePoolHeader(uint8_t* start, Pool* p, bool isNatural);
  static void WritePoolGuard(BufferOffset branch, Instruction* inst,
                             BufferOffset dest);

  static uint32_t PatchWrite_NearCallSize();
  static uint32_t NopSize() { return 4; }
  static void PatchWrite_NearCall(CodeLocationLabel start,
                                  CodeLocationLabel toCall);
  static void PatchDataWithValueCheck(CodeLocationLabel label,
                                      PatchedImmPtr newValue,
                                      PatchedImmPtr expectedValue);
  static void PatchDataWithValueCheck(CodeLocationLabel label, ImmPtr newValue,
                                      ImmPtr expectedValue);
  static void PatchWrite_Imm32(CodeLocationLabel label, Imm32 imm);

  static uint32_t AlignDoubleArg(uint32_t offset) { return (offset + 1) & ~1; }
  static uint8_t* NextInstruction(uint8_t* instruction,
                                  uint32_t* count = nullptr);

  // Toggle a jmp or cmp emitted by toggledJump().
  static void ToggleToJmp(CodeLocationLabel inst_);
  static void ToggleToCmp(CodeLocationLabel inst_);

  static uint8_t* BailoutTableStart(uint8_t* code);

  static size_t ToggledCallSize(uint8_t* code);
  static void ToggleCall(CodeLocationLabel inst_, bool enabled);

  void processCodeLabels(uint8_t* rawCode);

  void verifyHeapAccessDisassembly(uint32_t begin, uint32_t end,
                                   const Disassembler::HeapAccess& heapAccess) {
    // Implement this if we implement a disassembler.
  }
};  // Assembler

// An Instruction is a structure for both encoding and decoding any and all ARM
// instructions. Many classes have not been implemented thus far.
class Instruction {
  uint32_t data;

 protected:
  // This is not for defaulting to always, this is for instructions that
  // cannot be made conditional, and have the usually invalid 4b1111 cond
  // field.
  explicit Instruction(uint32_t data_, bool fake = false)
      : data(data_ | 0xf0000000) {
    MOZ_ASSERT(fake || ((data_ & 0xf0000000) == 0));
  }
  // Standard constructor.
  Instruction(uint32_t data_, Assembler::Condition c)
      : data(data_ | (uint32_t)c) {
    MOZ_ASSERT((data_ & 0xf0000000) == 0);
  }
  // You should never create an instruction directly. You should create a more
  // specific instruction which will eventually call one of these constructors
  // for you.
 public:
  uint32_t encode() const { return data; }
  // Check if this instruction is really a particular case.
  template <class C>
  bool is() const {
    return C::IsTHIS(*this);
  }

  // Safely get a more specific variant of this pointer.
  template <class C>
  C* as() const {
    return C::AsTHIS(*this);
  }

  const Instruction& operator=(Instruction src) {
    data = src.data;
    return *this;
  }
  // Since almost all instructions have condition codes, the condition code
  // extractor resides in the base class.
  Assembler::Condition extractCond() const {
    MOZ_ASSERT(data >> 28 != 0xf,
               "The instruction does not have condition code");
    return (Assembler::Condition)(data & 0xf0000000);
  }

  // Sometimes, an api wants a uint32_t (or a pointer to it) rather than an
  // instruction. raw() just coerces this into a pointer to a uint32_t.
  const uint32_t* raw() const { return &data; }
  uint32_t size() const { return 4; }
};  // Instruction

// Make sure that it is the right size.
JS_STATIC_ASSERT(sizeof(Instruction) == 4);

inline void InstructionIterator::advanceRaw(ptrdiff_t instructions) {
  inst_ = inst_ + instructions;
}

// Data Transfer Instructions.
class InstDTR : public Instruction {
 public:
  enum IsByte_ { IsByte = 0x00400000, IsWord = 0x00000000 };
  static const int IsDTR = 0x04000000;
  static const int IsDTRMask = 0x0c000000;

  // TODO: Replace the initialization with something that is safer.
  InstDTR(LoadStore ls, IsByte_ ib, Index mode, Register rt, DTRAddr addr,
          Assembler::Condition c)
      : Instruction(ls | ib | mode | RT(rt) | addr.encode() | IsDTR, c) {}

  static bool IsTHIS(const Instruction& i);
  static InstDTR* AsTHIS(const Instruction& i);
};
JS_STATIC_ASSERT(sizeof(InstDTR) == sizeof(Instruction));

class InstLDR : public InstDTR {
 public:
  InstLDR(Index mode, Register rt, DTRAddr addr, Assembler::Condition c)
      : InstDTR(IsLoad, IsWord, mode, rt, addr, c) {}

  static bool IsTHIS(const Instruction& i);
  static InstLDR* AsTHIS(const Instruction& i);

  int32_t signedOffset() const {
    int32_t offset = encode() & 0xfff;
    if (IsUp_(encode() & IsUp) != IsUp) {
      return -offset;
    }
    return offset;
  }
  uint32_t* dest() const {
    int32_t offset = signedOffset();
    // When patching the load in PatchConstantPoolLoad, we ensure that the
    // offset is a multiple of 4, offset by 8 bytes from the actual
    // location.  Indeed, when the base register is PC, ARM's 3 stages
    // pipeline design makes it that PC is off by 8 bytes (= 2 *
    // sizeof(uint32*)) when we actually executed it.
    MOZ_ASSERT(offset % 4 == 0);
    offset >>= 2;
    return (uint32_t*)raw() + offset + 2;
  }
};
JS_STATIC_ASSERT(sizeof(InstDTR) == sizeof(InstLDR));

class InstNOP : public Instruction {
 public:
  static const uint32_t NopInst = 0x0320f000;

  InstNOP() : Instruction(NopInst, Assembler::Always) {}

  static bool IsTHIS(const Instruction& i);
  static InstNOP* AsTHIS(Instruction& i);
};

// Branching to a register, or calling a register
class InstBranchReg : public Instruction {
 protected:
  // Don't use BranchTag yourself, use a derived instruction.
  enum BranchTag { IsBX = 0x012fff10, IsBLX = 0x012fff30 };

  static const uint32_t IsBRegMask = 0x0ffffff0;

  InstBranchReg(BranchTag tag, Register rm, Assembler::Condition c)
      : Instruction(tag | rm.code(), c) {}

 public:
  static bool IsTHIS(const Instruction& i);
  static InstBranchReg* AsTHIS(const Instruction& i);

  // Get the register that is being branched to
  void extractDest(Register* dest);
  // Make sure we are branching to a pre-known register
  bool checkDest(Register dest);
};
JS_STATIC_ASSERT(sizeof(InstBranchReg) == sizeof(Instruction));

// Branching to an immediate offset, or calling an immediate offset
class InstBranchImm : public Instruction {
 protected:
  enum BranchTag { IsB = 0x0a000000, IsBL = 0x0b000000 };

  static const uint32_t IsBImmMask = 0x0f000000;

  InstBranchImm(BranchTag tag, BOffImm off, Assembler::Condition c)
      : Instruction(tag | off.encode(), c) {}

 public:
  static bool IsTHIS(const Instruction& i);
  static InstBranchImm* AsTHIS(const Instruction& i);

  void extractImm(BOffImm* dest);
};
JS_STATIC_ASSERT(sizeof(InstBranchImm) == sizeof(Instruction));

// Very specific branching instructions.
class InstBXReg : public InstBranchReg {
 public:
  static bool IsTHIS(const Instruction& i);
  static InstBXReg* AsTHIS(const Instruction& i);
};

class InstBLXReg : public InstBranchReg {
 public:
  InstBLXReg(Register reg, Assembler::Condition c)
      : InstBranchReg(IsBLX, reg, c) {}

  static bool IsTHIS(const Instruction& i);
  static InstBLXReg* AsTHIS(const Instruction& i);
};

class InstBImm : public InstBranchImm {
 public:
  InstBImm(BOffImm off, Assembler::Condition c) : InstBranchImm(IsB, off, c) {}

  static bool IsTHIS(const Instruction& i);
  static InstBImm* AsTHIS(const Instruction& i);
};

class InstBLImm : public InstBranchImm {
 public:
  InstBLImm(BOffImm off, Assembler::Condition c)
      : InstBranchImm(IsBL, off, c) {}

  static bool IsTHIS(const Instruction& i);
  static InstBLImm* AsTHIS(const Instruction& i);
};

// Both movw and movt. The layout of both the immediate and the destination
// register is the same so the code is being shared.
class InstMovWT : public Instruction {
 protected:
  enum WT { IsW = 0x03000000, IsT = 0x03400000 };
  static const uint32_t IsWTMask = 0x0ff00000;

  InstMovWT(Register rd, Imm16 imm, WT wt, Assembler::Condition c)
      : Instruction(RD(rd) | imm.encode() | wt, c) {}

 public:
  void extractImm(Imm16* dest);
  void extractDest(Register* dest);
  bool checkImm(Imm16 dest);
  bool checkDest(Register dest);

  static bool IsTHIS(Instruction& i);
  static InstMovWT* AsTHIS(Instruction& i);
};
JS_STATIC_ASSERT(sizeof(InstMovWT) == sizeof(Instruction));

class InstMovW : public InstMovWT {
 public:
  InstMovW(Register rd, Imm16 imm, Assembler::Condition c)
      : InstMovWT(rd, imm, IsW, c) {}

  static bool IsTHIS(const Instruction& i);
  static InstMovW* AsTHIS(const Instruction& i);
};

class InstMovT : public InstMovWT {
 public:
  InstMovT(Register rd, Imm16 imm, Assembler::Condition c)
      : InstMovWT(rd, imm, IsT, c) {}

  static bool IsTHIS(const Instruction& i);
  static InstMovT* AsTHIS(const Instruction& i);
};

class InstALU : public Instruction {
  static const int32_t ALUMask = 0xc << 24;

 public:
  InstALU(Register rd, Register rn, Operand2 op2, ALUOp op, SBit s,
          Assembler::Condition c)
      : Instruction(maybeRD(rd) | maybeRN(rn) | op2.encode() | op | s, c) {}

  static bool IsTHIS(const Instruction& i);
  static InstALU* AsTHIS(const Instruction& i);

  void extractOp(ALUOp* ret);
  bool checkOp(ALUOp op);
  void extractDest(Register* ret);
  bool checkDest(Register rd);
  void extractOp1(Register* ret);
  bool checkOp1(Register rn);
  Operand2 extractOp2();
};

class InstCMP : public InstALU {
 public:
  static bool IsTHIS(const Instruction& i);
  static InstCMP* AsTHIS(const Instruction& i);
};

class InstMOV : public InstALU {
 public:
  static bool IsTHIS(const Instruction& i);
  static InstMOV* AsTHIS(const Instruction& i);
};

// Compile-time iterator over instructions, with a safe interface that
// references not-necessarily-linear Instructions by linear BufferOffset.
class BufferInstructionIterator
    : public ARMBuffer::AssemblerBufferInstIterator {
 public:
  BufferInstructionIterator(BufferOffset bo, ARMBuffer* buffer)
      : ARMBuffer::AssemblerBufferInstIterator(bo, buffer) {}

  // Advances the buffer to the next intentionally-inserted instruction.
  Instruction* next() {
    advance(cur()->size());
    maybeSkipAutomaticInstructions();
    return cur();
  }

  // Advances the BufferOffset past any automatically-inserted instructions.
  Instruction* maybeSkipAutomaticInstructions();
};

static const uint32_t NumIntArgRegs = 4;

// There are 16 *float* registers available for arguments
// If doubles are used, only half the number of registers are available.
static const uint32_t NumFloatArgRegs = 16;

static inline bool GetIntArgReg(uint32_t usedIntArgs, uint32_t usedFloatArgs,
                                Register* out) {
  if (usedIntArgs >= NumIntArgRegs) {
    return false;
  }

  *out = Register::FromCode(usedIntArgs);
  return true;
}

// Get a register in which we plan to put a quantity that will be used as an
// integer argument. This differs from GetIntArgReg in that if we have no more
// actual argument registers to use we will fall back on using whatever
// CallTempReg* don't overlap the argument registers, and only fail once those
// run out too.
static inline bool GetTempRegForIntArg(uint32_t usedIntArgs,
                                       uint32_t usedFloatArgs, Register* out) {
  if (GetIntArgReg(usedIntArgs, usedFloatArgs, out)) {
    return true;
  }

  // Unfortunately, we have to assume things about the point at which
  // GetIntArgReg returns false, because we need to know how many registers it
  // can allocate.
  usedIntArgs -= NumIntArgRegs;
  if (usedIntArgs >= NumCallTempNonArgRegs) {
    return false;
  }

  *out = CallTempNonArgRegs[usedIntArgs];
  return true;
}

#if !defined(JS_CODEGEN_ARM_HARDFP) || defined(JS_SIMULATOR_ARM)

static inline uint32_t GetArgStackDisp(uint32_t arg) {
  MOZ_ASSERT(!UseHardFpABI());
  MOZ_ASSERT(arg >= NumIntArgRegs);
  return (arg - NumIntArgRegs) * sizeof(intptr_t);
}

#endif

#if defined(JS_CODEGEN_ARM_HARDFP) || defined(JS_SIMULATOR_ARM)

static inline bool GetFloat32ArgReg(uint32_t usedIntArgs,
                                    uint32_t usedFloatArgs,
                                    FloatRegister* out) {
  MOZ_ASSERT(UseHardFpABI());
  if (usedFloatArgs >= NumFloatArgRegs) {
    return false;
  }
  *out = VFPRegister(usedFloatArgs, VFPRegister::Single);
  return true;
}
static inline bool GetDoubleArgReg(uint32_t usedIntArgs, uint32_t usedFloatArgs,
                                   FloatRegister* out) {
  MOZ_ASSERT(UseHardFpABI());
  MOZ_ASSERT((usedFloatArgs % 2) == 0);
  if (usedFloatArgs >= NumFloatArgRegs) {
    return false;
  }
  *out = VFPRegister(usedFloatArgs >> 1, VFPRegister::Double);
  return true;
}

static inline uint32_t GetIntArgStackDisp(uint32_t usedIntArgs,
                                          uint32_t usedFloatArgs,
                                          uint32_t* padding) {
  MOZ_ASSERT(UseHardFpABI());
  MOZ_ASSERT(usedIntArgs >= NumIntArgRegs);
  uint32_t doubleSlots =
      Max(0, (int32_t)usedFloatArgs - (int32_t)NumFloatArgRegs);
  doubleSlots *= 2;
  int intSlots = usedIntArgs - NumIntArgRegs;
  return (intSlots + doubleSlots + *padding) * sizeof(intptr_t);
}

static inline uint32_t GetFloat32ArgStackDisp(uint32_t usedIntArgs,
                                              uint32_t usedFloatArgs,
                                              uint32_t* padding) {
  MOZ_ASSERT(UseHardFpABI());
  MOZ_ASSERT(usedFloatArgs >= NumFloatArgRegs);
  uint32_t intSlots = 0;
  if (usedIntArgs > NumIntArgRegs) {
    intSlots = usedIntArgs - NumIntArgRegs;
  }
  uint32_t float32Slots = usedFloatArgs - NumFloatArgRegs;
  return (intSlots + float32Slots + *padding) * sizeof(intptr_t);
}

static inline uint32_t GetDoubleArgStackDisp(uint32_t usedIntArgs,
                                             uint32_t usedFloatArgs,
                                             uint32_t* padding) {
  MOZ_ASSERT(UseHardFpABI());
  MOZ_ASSERT(usedFloatArgs >= NumFloatArgRegs);
  uint32_t intSlots = 0;
  if (usedIntArgs > NumIntArgRegs) {
    intSlots = usedIntArgs - NumIntArgRegs;
    // Update the amount of padding required.
    *padding += (*padding + usedIntArgs) % 2;
  }
  uint32_t doubleSlots = usedFloatArgs - NumFloatArgRegs;
  doubleSlots *= 2;
  return (intSlots + doubleSlots + *padding) * sizeof(intptr_t);
}

#endif

class DoubleEncoder {
  struct DoubleEntry {
    uint32_t dblTop;
    datastore::Imm8VFPImmData data;
  };

  static const DoubleEntry table[256];

 public:
  bool lookup(uint32_t top, datastore::Imm8VFPImmData* ret) const {
    for (int i = 0; i < 256; i++) {
      if (table[i].dblTop == top) {
        *ret = table[i].data;
        return true;
      }
    }
    return false;
  }
};

// Forbids nop filling for testing purposes. Not nestable.
class AutoForbidNops {
 protected:
  Assembler* masm_;

 public:
  explicit AutoForbidNops(Assembler* masm) : masm_(masm) {
    masm_->enterNoNops();
  }
  ~AutoForbidNops() { masm_->leaveNoNops(); }
};

class AutoForbidPoolsAndNops : public AutoForbidNops {
 public:
  // The maxInst argument is the maximum number of word sized instructions
  // that will be allocated within this context. It is used to determine if
  // the pool needs to be dumped before entering this content. The debug code
  // checks that no more than maxInst instructions are actually allocated.
  //
  // Allocation of pool entries is not supported within this content so the
  // code can not use large integers or float constants etc.
  AutoForbidPoolsAndNops(Assembler* masm, size_t maxInst) : AutoForbidNops(masm) {
    masm_->enterNoPool(maxInst);
  }

  ~AutoForbidPoolsAndNops() { masm_->leaveNoPool(); }
};

}  // namespace jit
}  // namespace js

#endif /* jit_arm_Assembler_arm_h */