/* -*- 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/. */

#include "jit/arm64/CodeGenerator-arm64.h"

#include "mozilla/DebugOnly.h"
#include "mozilla/MathAlgorithms.h"

#include <bit>

#include "builtin/Number.h"
#include "jit/CodeGenerator.h"
#include "jit/InlineScriptTree.h"
#include "jit/JitRuntime.h"
#include "jit/MIR-wasm.h"
#include "jit/MIR.h"
#include "jit/MIRGraph.h"
#include "jit/ReciprocalMulConstants.h"
#include "vm/JSContext.h"
#include "vm/Realm.h"
#include "vm/Shape.h"

#include "jit/shared/CodeGenerator-shared-inl.h"
#include "vm/JSScript-inl.h"

using namespace js;
using namespace js::jit;

using JS::GenericNaN;
using mozilla::FloorLog2;
using mozilla::Maybe;
using mozilla::NegativeInfinity;
using mozilla::Nothing;
using mozilla::Some;

// shared
CodeGeneratorARM64::CodeGeneratorARM64(MIRGenerator* gen, LIRGraph* graph,
                                       MacroAssembler* masm,
                                       const wasm::CodeMetadata* wasmCodeMeta)
    : CodeGeneratorShared(gen, graph, masm, wasmCodeMeta) {}

bool CodeGeneratorARM64::generateOutOfLineCode() {
  AutoCreatedBy acb(masm, "CodeGeneratorARM64::generateOutOfLineCode");

  if (!CodeGeneratorShared::generateOutOfLineCode()) {
    return false;
  }

  if (deoptLabel_.used()) {
    // All non-table-based bailouts will go here.
    masm.bind(&deoptLabel_);

    // Store the frame size, so the handler can recover the IonScript.
    masm.push(Imm32(frameSize()));

    TrampolinePtr handler = gen->jitRuntime()->getGenericBailoutHandler();
    masm.jump(handler);
  }

  return !masm.oom();
}

void CodeGeneratorARM64::emitBranch(Assembler::Condition cond,
                                    MBasicBlock* mirTrue,
                                    MBasicBlock* mirFalse) {
  if (isNextBlock(mirFalse->lir())) {
    jumpToBlock(mirTrue, cond);
  } else {
    jumpToBlock(mirFalse, Assembler::InvertCondition(cond));
    jumpToBlock(mirTrue);
  }
}

void CodeGeneratorARM64::emitBailoutOOL(LSnapshot* snapshot) {
  masm.push(Imm32(snapshot->snapshotOffset()));
  masm.B(&deoptLabel_);
}

void CodeGeneratorARM64::bailoutIf(Assembler::Condition condition,
                                   LSnapshot* snapshot) {
  encode(snapshot);

  InlineScriptTree* tree = snapshot->mir()->block()->trackedTree();
  auto* ool = new (alloc()) LambdaOutOfLineCode(
      [=, this](OutOfLineCode& ool) { emitBailoutOOL(snapshot); });
  addOutOfLineCode(ool,
                   new (alloc()) BytecodeSite(tree, tree->script()->code()));

  masm.B(ool->entry(), condition);
}

void CodeGeneratorARM64::bailoutIfZero(Assembler::Condition condition,
                                       ARMRegister rt, LSnapshot* snapshot) {
  MOZ_ASSERT(condition == Assembler::Zero || condition == Assembler::NonZero);

  encode(snapshot);

  InlineScriptTree* tree = snapshot->mir()->block()->trackedTree();
  auto* ool = new (alloc()) LambdaOutOfLineCode(
      [=, this](OutOfLineCode& ool) { emitBailoutOOL(snapshot); });
  addOutOfLineCode(ool,
                   new (alloc()) BytecodeSite(tree, tree->script()->code()));

  if (condition == Assembler::Zero) {
    masm.Cbz(rt, ool->entry());
  } else {
    masm.Cbnz(rt, ool->entry());
  }
}

void CodeGeneratorARM64::bailoutFrom(Label* label, LSnapshot* snapshot) {
  MOZ_ASSERT_IF(!masm.oom(), label->used());
  MOZ_ASSERT_IF(!masm.oom(), !label->bound());

  encode(snapshot);

  InlineScriptTree* tree = snapshot->mir()->block()->trackedTree();
  auto* ool = new (alloc()) LambdaOutOfLineCode(
      [=, this](OutOfLineCode& ool) { emitBailoutOOL(snapshot); });
  addOutOfLineCode(ool,
                   new (alloc()) BytecodeSite(tree, tree->script()->code()));

  masm.retarget(label, ool->entry());
}

void CodeGeneratorARM64::bailout(LSnapshot* snapshot) {
  Label label;
  masm.b(&label);
  bailoutFrom(&label, snapshot);
}

void CodeGenerator::visitMinMaxD(LMinMaxD* ins) {
  ARMFPRegister lhs(ToFloatRegister(ins->first()), 64);
  ARMFPRegister rhs(ToFloatRegister(ins->second()), 64);
  ARMFPRegister output(ToFloatRegister(ins->output()), 64);
  if (ins->mir()->isMax()) {
    masm.Fmax(output, lhs, rhs);
  } else {
    masm.Fmin(output, lhs, rhs);
  }
}

void CodeGenerator::visitMinMaxF(LMinMaxF* ins) {
  ARMFPRegister lhs(ToFloatRegister(ins->first()), 32);
  ARMFPRegister rhs(ToFloatRegister(ins->second()), 32);
  ARMFPRegister output(ToFloatRegister(ins->output()), 32);
  if (ins->mir()->isMax()) {
    masm.Fmax(output, lhs, rhs);
  } else {
    masm.Fmin(output, lhs, rhs);
  }
}

template <typename T>
static ARMRegister toWRegister(const T* a) {
  return ARMRegister(ToRegister(a), 32);
}

template <typename T>
static ARMRegister toXRegister(const T* a) {
  return ARMRegister(ToRegister(a), 64);
}

static ARMRegister toXRegister(const LInt64Allocation& a) {
  return ARMRegister(ToRegister64(a).reg, 64);
}

static Operand toWOperand(const LAllocation* a) {
  if (a->isConstant()) {
    return Operand(ToInt32(a));
  }
  return Operand(toWRegister(a));
}

static Operand toXOperand(const LAllocation* a) {
  if (a->isConstant()) {
    return Operand(ToIntPtr(a));
  }
  return Operand(toXRegister(a));
}

static Operand toXOperand(const LInt64Allocation& a) {
  if (IsConstant(a)) {
    return Operand(ToInt64(a));
  }
  return Operand(toXRegister(a));
}

void CodeGenerator::visitAddI(LAddI* ins) {
  const LAllocation* lhs = ins->lhs();
  const LAllocation* rhs = ins->rhs();
  const LDefinition* dest = ins->output();

  // Platforms with three-operand arithmetic ops don't need recovery.
  MOZ_ASSERT(!ins->recoversInput());

  if (ins->snapshot()) {
    masm.Adds(toWRegister(dest), toWRegister(lhs), toWOperand(rhs));
    bailoutIf(Assembler::Overflow, ins->snapshot());
  } else {
    masm.Add(toWRegister(dest), toWRegister(lhs), toWOperand(rhs));
  }
}

void CodeGenerator::visitSubI(LSubI* ins) {
  const LAllocation* lhs = ins->lhs();
  const LAllocation* rhs = ins->rhs();
  const LDefinition* dest = ins->output();

  // Platforms with three-operand arithmetic ops don't need recovery.
  MOZ_ASSERT(!ins->recoversInput());

  if (ins->snapshot()) {
    masm.Subs(toWRegister(dest), toWRegister(lhs), toWOperand(rhs));
    bailoutIf(Assembler::Overflow, ins->snapshot());
  } else {
    masm.Sub(toWRegister(dest), toWRegister(lhs), toWOperand(rhs));
  }
}

void CodeGenerator::visitMulI(LMulI* ins) {
  const LAllocation* lhs = ins->lhs();
  const LAllocation* rhs = ins->rhs();
  const LDefinition* dest = ins->output();
  MMul* mul = ins->mir();
  MOZ_ASSERT_IF(mul->mode() == MMul::Integer,
                !mul->canBeNegativeZero() && !mul->canOverflow());

  Register lhsreg = ToRegister(lhs);
  const ARMRegister lhsreg32 = ARMRegister(lhsreg, 32);
  Register destreg = ToRegister(dest);
  const ARMRegister destreg32 = ARMRegister(destreg, 32);

  if (rhs->isConstant()) {
    // Bailout on -0.0.
    int32_t constant = ToInt32(rhs);
    if (mul->canBeNegativeZero() && constant <= 0) {
      Assembler::Condition bailoutCond =
          (constant == 0) ? Assembler::LessThan : Assembler::Equal;
      bailoutCmp32(bailoutCond, lhsreg, Imm32(0), ins->snapshot());
    }

    switch (constant) {
      case -1:
        masm.Negs(destreg32, Operand(lhsreg32));
        break;  // Go to overflow check.
      case 0:
        masm.Mov(destreg32, wzr);
        return;  // Avoid overflow check.
      case 1:
        if (destreg != lhsreg) {
          masm.Mov(destreg32, lhsreg32);
        }
        return;  // Avoid overflow check.
      case 2:
        if (!mul->canOverflow()) {
          masm.Add(destreg32, lhsreg32, Operand(lhsreg32));
          return;  // Avoid overflow check.
        }
        masm.Adds(destreg32, lhsreg32, Operand(lhsreg32));
        break;  // Go to overflow check.
      default:
        // Use shift if cannot overflow and constant is a power of 2
        if (!mul->canOverflow() && constant > 0) {
          int32_t shift = FloorLog2(uint32_t(constant));
          if ((1 << shift) == constant) {
            masm.Lsl(destreg32, lhsreg32, shift);
            return;
          }
        }

        // Otherwise, just multiply. We have to check for overflow.
        // Negative zero was handled above.
        Label bailout;
        Label* onOverflow = mul->canOverflow() ? &bailout : nullptr;

        vixl::UseScratchRegisterScope temps(&masm.asVIXL());
        const Register scratch = temps.AcquireW().asUnsized();

        masm.move32(Imm32(constant), scratch);
        masm.mul32(lhsreg, scratch, destreg, onOverflow);

        if (onOverflow) {
          bailoutFrom(&bailout, ins->snapshot());
        }
        return;
    }

    // Overflow check.
    if (mul->canOverflow()) {
      bailoutIf(Assembler::Overflow, ins->snapshot());
    }
  } else {
    Register rhsreg = ToRegister(rhs);
    const ARMRegister rhsreg32 = ARMRegister(rhsreg, 32);

    Label bailout;
    Label* onOverflow = mul->canOverflow() ? &bailout : nullptr;

    if (mul->canBeNegativeZero()) {
      // The product of two integer operands is negative zero iff one
      // operand is zero, and the other is negative. Therefore, the
      // sum of the two operands will also be negative (specifically,
      // it will be the non-zero operand). If the result of the
      // multiplication is 0, we can check the sign of the sum to
      // determine whether we should bail out.

      // This code can bailout, so lowering guarantees that the input
      // operands are not overwritten.
      MOZ_ASSERT(destreg != lhsreg);
      MOZ_ASSERT(destreg != rhsreg);

      // Do the multiplication.
      masm.mul32(lhsreg, rhsreg, destreg, onOverflow);

      // Set Zero flag if destreg is 0.
      masm.test32(destreg, destreg);

      // ccmn is 'conditional compare negative'.
      // If the Zero flag is set:
      //    perform a compare negative (compute lhs+rhs and set flags)
      // else:
      //    clear flags
      masm.Ccmn(lhsreg32, rhsreg32, vixl::NoFlag, Assembler::Zero);

      // Bails out if (lhs * rhs == 0) && (lhs + rhs < 0):
      bailoutIf(Assembler::LessThan, ins->snapshot());

    } else {
      masm.mul32(lhsreg, rhsreg, destreg, onOverflow);
    }
    if (onOverflow) {
      bailoutFrom(&bailout, ins->snapshot());
    }
  }
}

template <class LIR>
static void TrapIfDivideByZero(MacroAssembler& masm, LIR* lir,
                               ARMRegister rhs) {
  auto* mir = lir->mir();
  MOZ_ASSERT(mir->trapOnError());

  if (mir->canBeDivideByZero()) {
    Label nonZero;
    masm.Cbnz(rhs, &nonZero);
    masm.wasmTrap(wasm::Trap::IntegerDivideByZero, mir->trapSiteDesc());
    masm.bind(&nonZero);
  }
}

void CodeGenerator::visitDivI(LDivI* ins) {
  Register lhs = ToRegister(ins->lhs());
  Register rhs = ToRegister(ins->rhs());

  ARMRegister lhs32 = toWRegister(ins->lhs());
  ARMRegister rhs32 = toWRegister(ins->rhs());
  ARMRegister output32 = toWRegister(ins->output());

  MDiv* mir = ins->mir();

  // Handle division by zero.
  if (mir->canBeDivideByZero()) {
    if (mir->trapOnError()) {
      TrapIfDivideByZero(masm, ins, rhs32);
    } else if (mir->canTruncateInfinities()) {
      // SDIV returns zero for division by zero, exactly what we want for
      // truncated division. Remainder computation expects a non-zero divisor,
      // so we must also be allowed to truncate the remainder.
      MOZ_ASSERT(mir->canTruncateRemainder(),
                 "remainder computation expects a non-zero divisor");
    } else {
      MOZ_ASSERT(mir->fallible());
      bailoutTest32(Assembler::Zero, rhs, rhs, ins->snapshot());
    }
  }

  // Handle an integer overflow from (INT32_MIN / -1).
  // The integer division gives INT32_MIN, but should be -(double)INT32_MIN.
  //
  // SDIV returns INT32_MIN for (INT32_MIN / -1), so no extra code needed when
  // truncation is allowed.
  if (mir->canBeNegativeOverflow() &&
      (mir->trapOnError() || !mir->canTruncateOverflow())) {
    Label notOverflow;

    // Branch to handle the non-overflow cases.
    masm.branch32(Assembler::NotEqual, lhs, Imm32(INT32_MIN), &notOverflow);

    // Handle overflow.
    if (mir->trapOnError()) {
      masm.branch32(Assembler::NotEqual, rhs, Imm32(-1), &notOverflow);
      masm.wasmTrap(wasm::Trap::IntegerOverflow, mir->trapSiteDesc());
    } else {
      MOZ_ASSERT(mir->fallible());
      bailoutCmp32(Assembler::Equal, rhs, Imm32(-1), ins->snapshot());
    }
    masm.bind(&notOverflow);
  }

  // Handle negative zero: lhs == 0 && rhs < 0.
  if (!mir->canTruncateNegativeZero() && mir->canBeNegativeZero()) {
    Label nonZero;
    masm.branch32(Assembler::NotEqual, lhs, Imm32(0), &nonZero);
    bailoutCmp32(Assembler::LessThan, rhs, Imm32(0), ins->snapshot());
    masm.bind(&nonZero);
  }

  // Perform integer division.
  masm.Sdiv(output32, lhs32, rhs32);

  if (!mir->canTruncateRemainder()) {
    vixl::UseScratchRegisterScope temps(&masm.asVIXL());
    ARMRegister remainder32 = temps.AcquireW();
    Register remainder = remainder32.asUnsized();

    // Compute the remainder: remainder = lhs - (output * rhs).
    masm.Msub(remainder32, output32, rhs32, lhs32);

    bailoutTest32(Assembler::NonZero, remainder, remainder, ins->snapshot());
  }
}

void CodeGenerator::visitDivPowTwoI(LDivPowTwoI* ins) {
  const Register numerator = ToRegister(ins->numerator());
  const ARMRegister numerator32 = toWRegister(ins->numerator());
  const ARMRegister output32 = toWRegister(ins->output());

  int32_t shift = ins->shift();
  bool negativeDivisor = ins->negativeDivisor();
  MDiv* mir = ins->mir();

  if (!mir->isTruncated() && negativeDivisor) {
    // 0 divided by a negative number returns a -0 double.
    bailoutTest32(Assembler::Zero, numerator, numerator, ins->snapshot());
  }

  if (shift) {
    if (!mir->isTruncated()) {
      // If the remainder is != 0, bailout since this must be a double.
      bailoutTest32(Assembler::NonZero, numerator,
                    Imm32(UINT32_MAX >> (32 - shift)), ins->snapshot());
    }

    if (mir->isUnsigned()) {
      // shift right
      masm.Lsr(output32, numerator32, shift);
    } else {
      ARMRegister temp32 = numerator32;
      // Adjust the value so that shifting produces a correctly
      // rounded result when the numerator is negative. See 10-1
      // "Signed Division by a Known Power of 2" in Henry
      // S. Warren, Jr.'s Hacker's Delight.
      if (mir->canBeNegativeDividend() && mir->isTruncated()) {
        if (shift > 1) {
          // Copy the sign bit of the numerator. (= (2^32 - 1) or 0)
          masm.Asr(output32, numerator32, 31);
          temp32 = output32;
        }
        // Divide by 2^(32 - shift)
        // i.e. (= (2^32 - 1) / 2^(32 - shift) or 0)
        // i.e. (= (2^shift - 1) or 0)
        masm.Lsr(output32, temp32, 32 - shift);
        // If signed, make any 1 bit below the shifted bits to bubble up, such
        // that once shifted the value would be rounded towards 0.
        masm.Add(output32, output32, numerator32);
        temp32 = output32;
      }
      masm.Asr(output32, temp32, shift);

      if (negativeDivisor) {
        masm.Neg(output32, output32);
      }
    }
    return;
  }

  if (negativeDivisor) {
    // INT32_MIN / -1 overflows.
    if (!mir->isTruncated()) {
      masm.Negs(output32, numerator32);
      bailoutIf(Assembler::Overflow, ins->snapshot());
    } else if (mir->trapOnError()) {
      Label ok;
      masm.Negs(output32, numerator32);
      masm.branch(Assembler::NoOverflow, &ok);
      masm.wasmTrap(wasm::Trap::IntegerOverflow, mir->trapSiteDesc());
      masm.bind(&ok);
    } else {
      // Do not set condition flags.
      masm.Neg(output32, numerator32);
    }
  } else {
    if (mir->isUnsigned() && !mir->isTruncated()) {
      // Copy and set flags.
      masm.Adds(output32, numerator32, 0);
      // Unsigned division by 1 can overflow if output is not truncated, as we
      // do not have an Unsigned type for MIR instructions.
      bailoutIf(Assembler::Signed, ins->snapshot());
    } else {
      // Copy the result.
      masm.Mov(output32, numerator32);
    }
  }
}

void CodeGenerator::visitDivPowTwoI64(LDivPowTwoI64* ins) {
  ARMRegister numerator64 = toXRegister(ins->numerator());
  ARMRegister output64 = toXRegister(ins->output());

  int32_t shift = ins->shift();
  bool negativeDivisor = ins->negativeDivisor();
  MDiv* mir = ins->mir();

  if (shift) {
    if (mir->isUnsigned()) {
      // shift right
      masm.Lsr(output64, numerator64, shift);
    } else {
      ARMRegister temp64 = numerator64;
      // Adjust the value so that shifting produces a correctly
      // rounded result when the numerator is negative. See 10-1
      // "Signed Division by a Known Power of 2" in Henry
      // S. Warren, Jr.'s Hacker's Delight.
      if (mir->canBeNegativeDividend()) {
        if (shift > 1) {
          // Copy the sign bit of the numerator. (= (2^64 - 1) or 0)
          masm.Asr(output64, numerator64, 63);
          temp64 = output64;
        }
        // Divide by 2^(64 - shift)
        // i.e. (= (2^64 - 1) / 2^(64 - shift) or 0)
        // i.e. (= (2^shift - 1) or 0)
        masm.Lsr(output64, temp64, 64 - shift);
        // If signed, make any 1 bit below the shifted bits to bubble up, such
        // that once shifted the value would be rounded towards 0.
        masm.Add(output64, output64, numerator64);
        temp64 = output64;
      }
      masm.Asr(output64, temp64, shift);

      if (negativeDivisor) {
        masm.Neg(output64, output64);
      }
    }
    return;
  }

  if (negativeDivisor) {
    // INT64_MIN / -1 overflows.
    Label ok;
    masm.Negs(output64, numerator64);
    masm.branch(Assembler::NoOverflow, &ok);
    masm.wasmTrap(wasm::Trap::IntegerOverflow, mir->trapSiteDesc());
    masm.bind(&ok);
  } else {
    // Copy the result.
    masm.Mov(output64, numerator64);
  }
}

template <class LDivOrMod>
static void DivideWithConstant(MacroAssembler& masm, LDivOrMod* ins) {
  ARMRegister lhs32 = toWRegister(ins->numerator());
  ARMRegister lhs64 = toXRegister(ins->numerator());
  ARMRegister output32 = toWRegister(ins->output());
  ARMRegister output64 = toXRegister(ins->output());
  int32_t d = ins->denominator();

  vixl::UseScratchRegisterScope temps(&masm.asVIXL());
  ARMRegister const32 = temps.AcquireW();

  // The absolute value of the denominator isn't a power of 2.
  MOZ_ASSERT(!std::has_single_bit(mozilla::Abs(d)));

  auto* mir = ins->mir();

  // We will first divide by Abs(d), and negate the answer if d is negative.
  // If desired, this can be avoided by generalizing computeDivisionConstants.
  auto rmc = ReciprocalMulConstants::computeSignedDivisionConstants(d);

  // We first compute (M * n) >> 32, where M = rmc.multiplier.
  masm.Mov(const32, int32_t(rmc.multiplier));
  if (rmc.multiplier > INT32_MAX) {
    MOZ_ASSERT(rmc.multiplier < (int64_t(1) << 32));

    // We actually compute (int32_t(M) * n) instead, without the upper bit.
    // Thus, (M * n) = (int32_t(M) * n) + n << 32.
    //
    // ((int32_t(M) * n) + n << 32) can't overflow, as both operands have
    // opposite signs because int32_t(M) is negative.
    masm.Lsl(output64, lhs64, 32);

    // Store (M * n) in output64.
    masm.Smaddl(output64, const32, lhs32, output64);
  } else {
    // Store (M * n) in output64.
    masm.Smull(output64, const32, lhs32);
  }

  // (M * n) >> (32 + shift) is the truncated division answer if n is
  // non-negative, as proved in the comments of computeDivisionConstants. We
  // must add 1 later if n is negative to get the right answer in all cases.
  masm.Asr(output64, output64, 32 + rmc.shiftAmount);

  // We'll subtract -1 instead of adding 1, because (n < 0 ? -1 : 0) can be
  // computed with just a sign-extending shift of 31 bits.
  if (mir->canBeNegativeDividend()) {
    masm.Sub(output32, output32, Operand(lhs32, vixl::ASR, 31));
  }

  // After this, output32 contains the correct truncated division result.
  if (d < 0) {
    masm.Neg(output32, output32);
  }
}

void CodeGenerator::visitDivConstantI(LDivConstantI* ins) {
  ARMRegister lhs32 = toWRegister(ins->numerator());
  ARMRegister output32 = toWRegister(ins->output());
  int32_t d = ins->denominator();

  MDiv* mir = ins->mir();

  if (d == 0) {
    if (mir->trapOnError()) {
      masm.wasmTrap(wasm::Trap::IntegerDivideByZero, mir->trapSiteDesc());
    } else if (mir->canTruncateInfinities()) {
      masm.Mov(output32, wzr);
    } else {
      MOZ_ASSERT(mir->fallible());
      bailout(ins->snapshot());
    }
    return;
  }

  // Compute the truncated division result in output32.
  DivideWithConstant(masm, ins);

  if (!mir->isTruncated()) {
    vixl::UseScratchRegisterScope temps(&masm.asVIXL());
    ARMRegister temp32 = temps.AcquireW();
    Register temp = temp32.asUnsized();

    // This is a division op. Multiply the obtained value by d to check if
    // the correct answer is an integer. This cannot overflow, since |d| > 1.
    masm.Mov(temp32, d);
    masm.Msub(temp32, output32, temp32, lhs32);

    if (d > 0) {
      // bailout if (lhs - output * d != 0)
      bailoutTest32(Assembler::NonZero, temp, temp, ins->snapshot());
    } else {
      MOZ_ASSERT(d < 0);

      // bailout if (lhs - output * d != 0)
      masm.Cmp(temp32, wzr);

      // If lhs is zero and the divisor is negative, the answer should have
      // been -0.
      //
      // or bailout if (lhs == 0).
      // ^                  ^
      // |                  '-- masm.Ccmp(lhs32, lhs32, .., ..)
      // '-- masm.Ccmp(.., .., vixl::ZFlag, Assembler::Zero)
      masm.Ccmp(lhs32, wzr, vixl::ZFlag, Assembler::Zero);

      // bailout if (lhs - output * d != 0) or (lhs == 0)
      bailoutIf(Assembler::Zero, ins->snapshot());
    }
  }
}

template <class LUDivOrUMod>
static void UnsignedDivideWithConstant(MacroAssembler& masm, LUDivOrUMod* ins) {
  ARMRegister lhs32 = toWRegister(ins->numerator());
  ARMRegister lhs64 = toXRegister(ins->numerator());
  ARMRegister output64 = toXRegister(ins->output());
  uint32_t d = ins->denominator();

  vixl::UseScratchRegisterScope temps(&masm.asVIXL());
  ARMRegister const32 = temps.AcquireW();

  // The denominator isn't a power of 2 (see LDivPowTwoI).
  MOZ_ASSERT(!std::has_single_bit(d));

  auto rmc = ReciprocalMulConstants::computeUnsignedDivisionConstants(d);

  // We first compute (M * n), where M = rmc.multiplier.
  masm.Mov(const32, int32_t(rmc.multiplier));
  masm.Umull(output64, const32, lhs32);
  if (rmc.multiplier > UINT32_MAX) {
    // M >= 2^32 and shift == 0 is impossible, as d >= 2 implies that
    // ((M * n) >> (32 + shift)) >= n > floor(n/d) whenever n >= d,
    // contradicting the proof of correctness in computeDivisionConstants.
    MOZ_ASSERT(rmc.shiftAmount > 0);
    MOZ_ASSERT(rmc.multiplier < (int64_t(1) << 33));

    // We actually compute (uint32_t(M) * n) instead, without the upper bit.
    // Thus, (M * n) = (uint32_t(M) * n) + n << 32.
    //
    // ((uint32_t(M) * n) + n << 32) can overflow. Hacker's Delight explains a
    // trick to avoid this overflow case, but we can avoid it by computing the
    // addition on 64 bits registers.
    //
    // Compute ((uint32_t(M) * n) >> 32 + n)
    masm.Add(output64, lhs64, Operand(output64, vixl::LSR, 32));

    // (M * n) >> (32 + shift) is the truncated division answer.
    masm.Lsr(output64, output64, rmc.shiftAmount);
  } else {
    // (M * n) >> (32 + shift) is the truncated division answer.
    masm.Lsr(output64, output64, 32 + rmc.shiftAmount);
  }
}

void CodeGenerator::visitUDivConstant(LUDivConstant* ins) {
  ARMRegister lhs32 = toWRegister(ins->numerator());
  ARMRegister output32 = toWRegister(ins->output());
  uint32_t d = ins->denominator();

  MDiv* mir = ins->mir();

  if (d == 0) {
    if (ins->mir()->trapOnError()) {
      masm.wasmTrap(wasm::Trap::IntegerDivideByZero, mir->trapSiteDesc());
    } else if (mir->canTruncateInfinities()) {
      masm.Mov(output32, wzr);
    } else {
      MOZ_ASSERT(mir->fallible());
      bailout(ins->snapshot());
    }
    return;
  }

  // Compute the truncated division result in output32.
  UnsignedDivideWithConstant(masm, ins);

  // We now have the truncated division value. We are checking whether the
  // division resulted in an integer, we multiply the obtained value by d and
  // check the remainder of the division.
  if (!mir->isTruncated()) {
    vixl::UseScratchRegisterScope temps(&masm.asVIXL());
    ARMRegister temp32 = temps.AcquireW();
    Register temp = temp32.asUnsized();

    masm.Mov(temp32, d);
    masm.Msub(temp32, output32, temp32, lhs32);

    // bailout if (lhs - output * d != 0)
    bailoutTest32(Assembler::NonZero, temp, temp, ins->snapshot());
  }
}

template <class LDivOrMod>
static void Divide64WithConstant(MacroAssembler& masm, LDivOrMod* ins) {
  ARMRegister lhs64 = toXRegister(ins->numerator());
  ARMRegister output64 = toXRegister(ins->output());
  int64_t d = ins->denominator();

  vixl::UseScratchRegisterScope temps(&masm.asVIXL());
  ARMRegister const64 = temps.AcquireX();

  // The absolute value of the denominator isn't a power of 2.
  MOZ_ASSERT(!std::has_single_bit(mozilla::Abs(d)));

  auto* mir = ins->mir();

  // We will first divide by Abs(d), and negate the answer if d is negative.
  // If desired, this can be avoided by generalizing computeDivisionConstants.
  auto rmc = ReciprocalMulConstants::computeSignedDivisionConstants(d);

  // We first compute (M * n) >> 64, where M = rmc.multiplier.
  masm.Mov(const64, uint64_t(rmc.multiplier));
  masm.Smulh(output64, lhs64, const64);
  if (rmc.multiplier > Int128(INT64_MAX)) {
    MOZ_ASSERT(rmc.multiplier < (Int128(1) << 64));

    // We actually computed output = ((int64_t(M) * n) >> 64) instead. Since
    // (M * n) >> 64 is the same as (output + n), we can correct for the
    // overflow. (output + n) can't overflow, as n and output have opposite
    // signs because int64_t(M) is negative.
    masm.Add(output64, output64, lhs64);
  }

  // (M * n) >> (64 + shift) is the truncated division answer if n is
  // non-negative, as proved in the comments of computeDivisionConstants. We
  // must add 1 later if n is negative to get the right answer in all cases.
  masm.Asr(output64, output64, rmc.shiftAmount);

  // We'll subtract -1 instead of adding 1, because (n < 0 ? -1 : 0) can be
  // computed with just a sign-extending shift of 63 bits.
  if (mir->canBeNegativeDividend()) {
    masm.Sub(output64, output64, Operand(lhs64, vixl::ASR, 63));
  }

  // After this, output64 contains the correct truncated division result.
  if (d < 0) {
    masm.Neg(output64, output64);
  }
}

void CodeGenerator::visitDivConstantI64(LDivConstantI64* ins) {
  int64_t d = ins->denominator();

  if (d == 0) {
    masm.wasmTrap(wasm::Trap::IntegerDivideByZero, ins->mir()->trapSiteDesc());
    return;
  }

  // Compute the truncated division result.
  Divide64WithConstant(masm, ins);
}

template <class LUDivOrUMod>
static void UnsignedDivide64WithConstant(MacroAssembler& masm,
                                         LUDivOrUMod* ins) {
  ARMRegister lhs64 = toXRegister(ins->numerator());
  ARMRegister output64 = toXRegister(ins->output());
  uint64_t d = ins->denominator();

  vixl::UseScratchRegisterScope temps(&masm.asVIXL());
  ARMRegister const64 = temps.AcquireX();

  // The denominator isn't a power of 2 (see LDivPowTwoI).
  MOZ_ASSERT(!std::has_single_bit(d));

  auto rmc = ReciprocalMulConstants::computeUnsignedDivisionConstants(d);

  // We first compute (M * n) >> 64, where M = rmc.multiplier.
  masm.Mov(const64, uint64_t(rmc.multiplier));
  masm.Umulh(output64, lhs64, const64);
  if (rmc.multiplier > Int128(UINT64_MAX)) {
    // M >= 2^64 and shift == 0 is impossible, as d >= 2 implies that
    // ((M * n) >> (64 + shift)) >= n > floor(n/d) whenever n >= d,
    // contradicting the proof of correctness in computeDivisionConstants.
    MOZ_ASSERT(rmc.shiftAmount > 0);
    MOZ_ASSERT(rmc.multiplier < (Int128(1) << 65));

    // We actually computed output = ((uint64_t(M) * n) >> 64) instead. Since
    // (M * n) >> (64 + shift) is the same as (output + n) >> shift, we can
    // correct for the overflow. This case is a bit trickier than the signed
    // case, though, as the (output + n) addition itself can overflow; however,
    // note that
    // (output + n) >> shift == (((n - output) >> 1) + output) >> (shift - 1),
    // which is overflow-free. See Hacker's Delight, section 10-8 for details.

    masm.Sub(const64, lhs64, output64);
    masm.Add(output64, output64, Operand(const64, vixl::LSR, 1));
    masm.Lsr(output64, output64, rmc.shiftAmount - 1);
  } else {
    masm.Lsr(output64, output64, rmc.shiftAmount);
  }
}

void CodeGenerator::visitUDivConstantI64(LUDivConstantI64* ins) {
  uint64_t d = ins->denominator();

  if (d == 0) {
    masm.wasmTrap(wasm::Trap::IntegerDivideByZero, ins->mir()->trapSiteDesc());
    return;
  }

  // Compute the truncated division result.
  UnsignedDivide64WithConstant(masm, ins);
}

void CodeGenerator::visitModI(LModI* ins) {
  Register lhs = ToRegister(ins->lhs());
  Register rhs = ToRegister(ins->rhs());

  ARMRegister lhs32 = toWRegister(ins->lhs());
  ARMRegister rhs32 = toWRegister(ins->rhs());
  ARMRegister output32 = toWRegister(ins->output());
  Label done;

  MMod* mir = ins->mir();

  // Prevent divide by zero.
  if (mir->canBeDivideByZero()) {
    if (mir->trapOnError()) {
      TrapIfDivideByZero(masm, ins, rhs32);
    } else if (mir->isTruncated()) {
      // Truncated division by zero yields integer zero.
      masm.Mov(output32, wzr);
      masm.Cbz(rhs32, &done);
    } else {
      // Non-truncated division by zero produces a non-integer.
      MOZ_ASSERT(mir->fallible());
      bailoutTest32(Assembler::Zero, rhs, rhs, ins->snapshot());
    }
  }

  // Signed division.
  masm.Sdiv(output32, lhs32, rhs32);

  // Compute the remainder: output = lhs - (output * rhs).
  masm.Msub(output32, output32, rhs32, lhs32);

  if (mir->canBeNegativeDividend() && !mir->isTruncated()) {
    // If output == 0 and lhs < 0, then the result should be double -0.0.
    // Note that this guard handles lhs == INT_MIN and rhs == -1:
    //   output = INT_MIN - (INT_MIN / -1) * -1
    //          = INT_MIN - INT_MIN
    //          = 0
    masm.Cbnz(output32, &done);
    bailoutCmp32(Assembler::LessThan, lhs, Imm32(0), ins->snapshot());
  }

  if (done.used()) {
    masm.bind(&done);
  }
}

void CodeGenerator::visitModPowTwoI(LModPowTwoI* ins) {
  Register lhs = ToRegister(ins->input());
  ARMRegister lhsw = toWRegister(ins->input());
  ARMRegister outw = toWRegister(ins->output());

  int32_t shift = ins->shift();
  bool canBeNegative =
      !ins->mir()->isUnsigned() && ins->mir()->canBeNegativeDividend();

  if (shift == 0) {
    if (canBeNegative && !ins->mir()->isTruncated()) {
      bailoutTest32(Assembler::Signed, lhs, lhs, ins->snapshot());
    }
    masm.Mov(outw, wzr);
    return;
  }

  Label negative;
  if (canBeNegative) {
    // Switch based on sign of the lhs.
    // Positive numbers are just a bitmask.
    masm.branchTest32(Assembler::Signed, lhs, lhs, &negative);
  }

  masm.And(outw, lhsw, Operand((uint32_t(1) << shift) - 1));

  if (canBeNegative) {
    Label done;
    masm.jump(&done);

    // Negative numbers need a negate, bitmask, negate.
    masm.bind(&negative);
    masm.Neg(outw, Operand(lhsw));
    masm.And(outw, outw, Operand((uint32_t(1) << shift) - 1));

    // Since a%b has the same sign as b, and a is negative in this branch,
    // an answer of 0 means the correct result is actually -0. Bail out.
    if (!ins->mir()->isTruncated()) {
      masm.Negs(outw, Operand(outw));
      bailoutIf(Assembler::Zero, ins->snapshot());
    } else {
      masm.Neg(outw, Operand(outw));
    }

    masm.bind(&done);
  }
}

void CodeGenerator::visitModPowTwoI64(LModPowTwoI64* ins) {
  Register lhs = ToRegister(ins->input());
  ARMRegister lhs64 = toXRegister(ins->input());
  ARMRegister out64 = toXRegister(ins->output());

  int32_t shift = ins->shift();
  bool canBeNegative =
      !ins->mir()->isUnsigned() && ins->mir()->canBeNegativeDividend();

  if (shift == 0) {
    masm.Mov(out64, xzr);
    return;
  }

  auto clearHighBits = [&](ARMRegister reg) {
    switch (shift) {
      case 32:
        masm.Mov(out64.W(), reg.W());
        break;
      default:
        masm.And(out64, reg, Operand((uint64_t(1) << shift) - 1));
        break;
    }
  };

  Label negative;
  if (canBeNegative) {
    // Switch based on sign of the lhs.
    // Positive numbers are just a bitmask.
    masm.branchTestPtr(Assembler::Signed, lhs, lhs, &negative);
  }

  clearHighBits(lhs64);

  if (canBeNegative) {
    Label done;
    masm.jump(&done);

    // Negative numbers need a negate, bitmask, negate.
    masm.bind(&negative);
    masm.Neg(out64, Operand(lhs64));
    clearHighBits(out64);
    masm.Neg(out64, Operand(out64));

    masm.bind(&done);
  }
}

void CodeGenerator::visitModConstantI(LModConstantI* ins) {
  Register lhs = ToRegister(ins->numerator());
  ARMRegister lhs32 = toWRegister(ins->numerator());
  ARMRegister output32 = toWRegister(ins->output());

  MMod* mir = ins->mir();

  int32_t d = ins->denominator();
  if (d == 0) {
    if (mir->trapOnError()) {
      masm.wasmTrap(wasm::Trap::IntegerDivideByZero, mir->trapSiteDesc());
    } else if (mir->isTruncated()) {
      masm.Mov(output32, wzr);
    } else {
      MOZ_ASSERT(mir->fallible());
      bailout(ins->snapshot());
    }
    return;
  }

  // Compute the truncated division result in output32.
  DivideWithConstant(masm, ins);

  // Compute the remainder: output = lhs - (output * rhs).
  {
    vixl::UseScratchRegisterScope temps(&masm.asVIXL());
    ARMRegister rhs32 = temps.AcquireW();

    masm.Mov(rhs32, d);
    masm.Msub(output32, output32, rhs32, lhs32);
  }

  if (mir->canBeNegativeDividend() && !mir->isTruncated()) {
    // If output == 0 and lhs < 0, then the result should be double -0.0.
    Label done;
    masm.Cbnz(output32, &done);
    bailoutCmp32(Assembler::LessThan, lhs, Imm32(0), ins->snapshot());
    masm.bind(&done);
  }
}

void CodeGenerator::visitUModConstant(LUModConstant* ins) {
  Register output = ToRegister(ins->output());
  ARMRegister lhs32 = toWRegister(ins->numerator());
  ARMRegister output32 = toWRegister(ins->output());

  MMod* mir = ins->mir();

  uint32_t d = ins->denominator();
  if (d == 0) {
    if (ins->mir()->trapOnError()) {
      masm.wasmTrap(wasm::Trap::IntegerDivideByZero, mir->trapSiteDesc());
    } else if (mir->isTruncated()) {
      masm.Mov(output32, wzr);
    } else {
      MOZ_ASSERT(mir->fallible());
      bailout(ins->snapshot());
    }
    return;
  }

  // Compute the truncated division result in output32.
  UnsignedDivideWithConstant(masm, ins);

  // Compute the remainder: output = lhs - (output * rhs).
  {
    vixl::UseScratchRegisterScope temps(&masm.asVIXL());
    ARMRegister rhs32 = temps.AcquireW();

    masm.Mov(rhs32, d);
    masm.Msub(output32, output32, rhs32, lhs32);
  }

  // Bail if not truncated and the remainder is in the range [2^31, 2^32).
  if (!ins->mir()->isTruncated()) {
    bailoutTest32(Assembler::Signed, output, output, ins->snapshot());
  }
}

void CodeGenerator::visitModConstantI64(LModConstantI64* ins) {
  ARMRegister lhs64 = toXRegister(ins->numerator());
  ARMRegister output64 = toXRegister(ins->output());

  int64_t d = ins->denominator();

  if (d == 0) {
    masm.wasmTrap(wasm::Trap::IntegerDivideByZero, ins->mir()->trapSiteDesc());
    return;
  }

  // Compute the truncated division result in output64.
  Divide64WithConstant(masm, ins);

  // Compute the remainder: output = lhs - (output * rhs).
  {
    vixl::UseScratchRegisterScope temps(&masm.asVIXL());
    ARMRegister rhs64 = temps.AcquireX();

    masm.Mov(rhs64, d);
    masm.Msub(output64, output64, rhs64, lhs64);
  }
}

void CodeGenerator::visitUModConstantI64(LUModConstantI64* ins) {
  ARMRegister lhs64 = toXRegister(ins->numerator());
  ARMRegister output64 = toXRegister(ins->output());

  uint64_t d = ins->denominator();

  if (d == 0) {
    masm.wasmTrap(wasm::Trap::IntegerDivideByZero, ins->mir()->trapSiteDesc());
    return;
  }

  // Compute the truncated division result in output64.
  UnsignedDivide64WithConstant(masm, ins);

  // Compute the remainder: output = lhs - (output * rhs).
  {
    vixl::UseScratchRegisterScope temps(&masm.asVIXL());
    ARMRegister rhs64 = temps.AcquireX();

    masm.Mov(rhs64, d);
    masm.Msub(output64, output64, rhs64, lhs64);
  }
}

void CodeGeneratorARM64::emitBigIntPtrDiv(LBigIntPtrDiv* ins, Register dividend,
                                          Register divisor, Register output) {
  // Callers handle division by zero and integer overflow.

  const ARMRegister dividend64(dividend, 64);
  const ARMRegister divisor64(divisor, 64);
  const ARMRegister output64(output, 64);

  masm.Sdiv(/* result= */ output64, dividend64, divisor64);
}

void CodeGeneratorARM64::emitBigIntPtrMod(LBigIntPtrMod* ins, Register dividend,
                                          Register divisor, Register output) {
  // Callers handle division by zero and integer overflow.

  const ARMRegister dividend64(dividend, 64);
  const ARMRegister divisor64(divisor, 64);
  const ARMRegister output64(output, 64);

  // Signed division.
  masm.Sdiv(output64, dividend64, divisor64);

  // Compute the remainder: output = dividend - (output * divisor).
  masm.Msub(/* result= */ output64, output64, divisor64, dividend64);
}

void CodeGenerator::visitBitNotI(LBitNotI* ins) {
  const LAllocation* input = ins->input();
  const LDefinition* output = ins->output();
  masm.Mvn(toWRegister(output), toWOperand(input));
}

void CodeGenerator::visitBitNotI64(LBitNotI64* ins) {
  Register64 input = ToRegister64(ins->input());
  Register64 output = ToOutRegister64(ins);
  masm.Mvn(vixl::Register(output.reg, 64), vixl::Register(input.reg, 64));
}

void CodeGenerator::visitBitOpI(LBitOpI* ins) {
  const ARMRegister lhs = toWRegister(ins->lhs());
  const Operand rhs = toWOperand(ins->rhs());
  const ARMRegister dest = toWRegister(ins->output());

  switch (ins->bitop()) {
    case JSOp::BitOr:
      masm.Orr(dest, lhs, rhs);
      break;
    case JSOp::BitXor:
      masm.Eor(dest, lhs, rhs);
      break;
    case JSOp::BitAnd:
      masm.And(dest, lhs, rhs);
      break;
    default:
      MOZ_CRASH("unexpected binary opcode");
  }
}

void CodeGenerator::visitShiftI(LShiftI* ins) {
  const ARMRegister lhs = toWRegister(ins->lhs());
  const LAllocation* rhs = ins->rhs();
  const ARMRegister dest = toWRegister(ins->output());

  if (rhs->isConstant()) {
    int32_t shift = ToInt32(rhs) & 0x1F;
    switch (ins->bitop()) {
      case JSOp::Lsh:
        if (shift) {
          masm.Lsl(dest, lhs, shift);
        } else {
          masm.Mov(dest, lhs);
        }
        break;
      case JSOp::Rsh:
        if (shift) {
          masm.Asr(dest, lhs, shift);
        } else {
          masm.Mov(dest, lhs);
        }
        break;
      case JSOp::Ursh:
        if (shift) {
          masm.Lsr(dest, lhs, shift);
        } else if (ins->mir()->toUrsh()->fallible()) {
          // x >>> 0 can overflow.
          masm.Ands(dest, lhs, Operand(0xFFFFFFFF));
          bailoutIf(Assembler::Signed, ins->snapshot());
        } else {
          masm.Mov(dest, lhs);
        }
        break;
      default:
        MOZ_CRASH("Unexpected shift op");
    }
  } else {
    const ARMRegister rhsreg = toWRegister(rhs);
    switch (ins->bitop()) {
      case JSOp::Lsh:
        masm.Lsl(dest, lhs, rhsreg);
        break;
      case JSOp::Rsh:
        masm.Asr(dest, lhs, rhsreg);
        break;
      case JSOp::Ursh:
        masm.Lsr(dest, lhs, rhsreg);
        if (ins->mir()->toUrsh()->fallible()) {
          /// x >>> 0 can overflow.
          masm.Cmp(dest, Operand(0));
          bailoutIf(Assembler::LessThan, ins->snapshot());
        }
        break;
      default:
        MOZ_CRASH("Unexpected shift op");
    }
  }
}

void CodeGenerator::visitShiftIntPtr(LShiftIntPtr* ins) {
  ARMRegister lhs = toXRegister(ins->lhs());
  const LAllocation* rhs = ins->rhs();
  ARMRegister dest = toXRegister(ins->output());

  if (rhs->isConstant()) {
    int32_t shift = ToIntPtr(rhs) & 0x3F;
    if (shift == 0) {
      masm.Mov(dest, lhs);
    } else {
      switch (ins->bitop()) {
        case JSOp::Lsh:
          masm.Lsl(dest, lhs, shift);
          break;
        case JSOp::Rsh:
          masm.Asr(dest, lhs, shift);
          break;
        case JSOp::Ursh:
          masm.Lsr(dest, lhs, shift);
          break;
        default:
          MOZ_CRASH("Unexpected shift op");
      }
    }
  } else {
    ARMRegister rhsreg = toXRegister(rhs);
    switch (ins->bitop()) {
      case JSOp::Lsh:
        masm.Lsl(dest, lhs, rhsreg);
        break;
      case JSOp::Rsh:
        masm.Asr(dest, lhs, rhsreg);
        break;
      case JSOp::Ursh:
        masm.Lsr(dest, lhs, rhsreg);
        break;
      default:
        MOZ_CRASH("Unexpected shift op");
    }
  }
}

void CodeGenerator::visitUrshD(LUrshD* ins) {
  ARMRegister lhs = toWRegister(ins->lhs());
  const LAllocation* rhs = ins->rhs();
  FloatRegister out = ToFloatRegister(ins->output());
  ARMRegister temp = toWRegister(ins->temp0());

  if (rhs->isConstant()) {
    int32_t shift = ToInt32(rhs) & 0x1F;
    if (shift) {
      masm.Lsr(temp, lhs, shift);
      masm.convertUInt32ToDouble(temp.asUnsized(), out);
    } else {
      masm.convertUInt32ToDouble(lhs.asUnsized(), out);
    }
  } else {
    masm.Lsr(temp, lhs, toWRegister(rhs));
    masm.convertUInt32ToDouble(temp.asUnsized(), out);
  }
}

void CodeGenerator::visitPowHalfD(LPowHalfD* ins) {
  FloatRegister input = ToFloatRegister(ins->input());
  FloatRegister output = ToFloatRegister(ins->output());

  ScratchDoubleScope scratch(masm);

  Label done, sqrt;

  if (!ins->mir()->operandIsNeverNegativeInfinity()) {
    // Branch if not -Infinity.
    masm.loadConstantDouble(NegativeInfinity<double>(), scratch);

    Assembler::DoubleCondition cond = Assembler::DoubleNotEqualOrUnordered;
    if (ins->mir()->operandIsNeverNaN()) {
      cond = Assembler::DoubleNotEqual;
    }
    masm.branchDouble(cond, input, scratch, &sqrt);

    // Math.pow(-Infinity, 0.5) == Infinity.
    masm.Fneg(ARMFPRegister(output, 64), ARMFPRegister(scratch, 64));
    masm.jump(&done);

    masm.bind(&sqrt);
  }

  if (!ins->mir()->operandIsNeverNegativeZero()) {
    // Math.pow(-0, 0.5) == 0 == Math.pow(0, 0.5).
    // Adding 0 converts any -0 to 0.
    masm.zeroDouble(scratch);
    masm.addDouble(input, scratch);
    masm.sqrtDouble(scratch, output);
  } else {
    masm.sqrtDouble(input, output);
  }

  masm.bind(&done);
}

MoveOperand CodeGeneratorARM64::toMoveOperand(const LAllocation a) const {
  if (a.isGeneralReg()) {
    return MoveOperand(ToRegister(a));
  }
  if (a.isFloatReg()) {
    return MoveOperand(ToFloatRegister(a));
  }
  MoveOperand::Kind kind = a.isStackArea() ? MoveOperand::Kind::EffectiveAddress
                                           : MoveOperand::Kind::Memory;
  return MoveOperand(ToAddress(a), kind);
}

class js::jit::OutOfLineTableSwitch
    : public OutOfLineCodeBase<CodeGeneratorARM64> {
  MTableSwitch* mir_;
  CodeLabel jumpLabel_;

  void accept(CodeGeneratorARM64* codegen) override {
    codegen->visitOutOfLineTableSwitch(this);
  }

 public:
  explicit OutOfLineTableSwitch(MTableSwitch* mir) : mir_(mir) {}

  MTableSwitch* mir() const { return mir_; }

  CodeLabel* jumpLabel() { return &jumpLabel_; }
};

void CodeGeneratorARM64::visitOutOfLineTableSwitch(OutOfLineTableSwitch* ool) {
  MTableSwitch* mir = ool->mir();

  // Prevent nop and pools sequences to appear in the jump table.
  AutoForbidPoolsAndNops afp(
      &masm, (mir->numCases() + 1) * (sizeof(void*) / vixl::kInstructionSize));
  masm.haltingAlign(sizeof(void*));
  masm.bind(ool->jumpLabel());
  masm.addCodeLabel(*ool->jumpLabel());

  for (size_t i = 0; i < mir->numCases(); i++) {
    LBlock* caseblock = skipTrivialBlocks(mir->getCase(i))->lir();
    Label* caseheader = caseblock->label();
    uint32_t caseoffset = caseheader->offset();

    // The entries of the jump table need to be absolute addresses,
    // and thus must be patched after codegen is finished.
    CodeLabel cl;
    masm.writeCodePointer(&cl);
    cl.target()->bind(caseoffset);
    masm.addCodeLabel(cl);
  }
}

void CodeGeneratorARM64::emitTableSwitchDispatch(MTableSwitch* mir,
                                                 Register index,
                                                 Register base) {
  Label* defaultcase = skipTrivialBlocks(mir->getDefault())->lir()->label();

  // Let the lowest table entry be indexed at 0.
  if (mir->low() != 0) {
    masm.sub32(Imm32(mir->low()), index);
  }

  // Jump to the default case if input is out of range.
  int32_t cases = mir->numCases();
  masm.branch32(Assembler::AboveOrEqual, index, Imm32(cases), defaultcase);

  // Because the target code has not yet been generated, we cannot know the
  // instruction offsets for use as jump targets. Therefore we construct
  // an OutOfLineTableSwitch that winds up holding the jump table.
  //
  // Because the jump table is generated as part of out-of-line code,
  // it is generated after all the regular codegen, so the jump targets
  // are guaranteed to exist when generating the jump table.
  OutOfLineTableSwitch* ool = new (alloc()) OutOfLineTableSwitch(mir);
  addOutOfLineCode(ool, mir);

  // Use the index to get the address of the jump target from the table.
  masm.mov(ool->jumpLabel(), base);
  BaseIndex pointer(base, index, ScalePointer);

  // Load the target from the jump table and branch to it.
  masm.branchToComputedAddress(pointer);
}

void CodeGenerator::visitMathD(LMathD* math) {
  ARMFPRegister lhs(ToFloatRegister(math->lhs()), 64);
  ARMFPRegister rhs(ToFloatRegister(math->rhs()), 64);
  ARMFPRegister output(ToFloatRegister(math->output()), 64);

  switch (math->jsop()) {
    case JSOp::Add:
      masm.Fadd(output, lhs, rhs);
      break;
    case JSOp::Sub:
      masm.Fsub(output, lhs, rhs);
      break;
    case JSOp::Mul:
      masm.Fmul(output, lhs, rhs);
      break;
    case JSOp::Div:
      masm.Fdiv(output, lhs, rhs);
      break;
    default:
      MOZ_CRASH("unexpected opcode");
  }
}

void CodeGenerator::visitMathF(LMathF* math) {
  ARMFPRegister lhs(ToFloatRegister(math->lhs()), 32);
  ARMFPRegister rhs(ToFloatRegister(math->rhs()), 32);
  ARMFPRegister output(ToFloatRegister(math->output()), 32);

  switch (math->jsop()) {
    case JSOp::Add:
      masm.Fadd(output, lhs, rhs);
      break;
    case JSOp::Sub:
      masm.Fsub(output, lhs, rhs);
      break;
    case JSOp::Mul:
      masm.Fmul(output, lhs, rhs);
      break;
    case JSOp::Div:
      masm.Fdiv(output, lhs, rhs);
      break;
    default:
      MOZ_CRASH("unexpected opcode");
  }
}

void CodeGenerator::visitTruncateDToInt32(LTruncateDToInt32* ins) {
  FloatRegister input = ToFloatRegister(ins->input());
  Register output = ToRegister(ins->output());

  // Directly call Fjcvtzs if available to avoid generating unused OOL code in
  // emitTruncateDouble.
  if (masm.hasFjcvtzs()) {
    masm.Fjcvtzs(ARMRegister(output, 32), ARMFPRegister(input, 64));
  } else {
    emitTruncateDouble(input, output, ins->mir());
  }
}

void CodeGenerator::visitWasmBuiltinTruncateDToInt32(
    LWasmBuiltinTruncateDToInt32* lir) {
  FloatRegister input = ToFloatRegister(lir->input());
  Register output = ToRegister(lir->output());

  // Directly call Fjcvtzs if available to avoid generating unused OOL code in
  // emitTruncateDouble.
  if (masm.hasFjcvtzs()) {
    masm.Fjcvtzs(ARMRegister(output, 32), ARMFPRegister(input, 64));
  } else {
    emitTruncateDouble(input, output, lir->mir());
  }
}

void CodeGenerator::visitTruncateFToInt32(LTruncateFToInt32* ins) {
  masm.truncateFloat32ModUint32(ToFloatRegister(ins->input()),
                                ToRegister(ins->output()));
}

void CodeGenerator::visitWasmBuiltinTruncateFToInt32(
    LWasmBuiltinTruncateFToInt32* lir) {
  MOZ_ASSERT(lir->instance()->isBogus(), "instance not used for arm64");
  masm.truncateFloat32ModUint32(ToFloatRegister(lir->input()),
                                ToRegister(lir->output()));
}

void CodeGenerator::visitBox(LBox* box) {
  const LAllocation* in = box->payload();
  ValueOperand result = ToOutValue(box);

  masm.moveValue(TypedOrValueRegister(box->type(), ToAnyRegister(in)), result);
}

void CodeGenerator::visitUnbox(LUnbox* unbox) {
  MUnbox* mir = unbox->mir();

  Register result = ToRegister(unbox->output());

  if (mir->fallible()) {
    ValueOperand value = ToValue(unbox->input());
    Label bail;
    switch (mir->type()) {
      case MIRType::Int32:
        masm.fallibleUnboxInt32(value, result, &bail);
        break;
      case MIRType::Boolean:
        masm.fallibleUnboxBoolean(value, result, &bail);
        break;
      case MIRType::Object:
        masm.fallibleUnboxObject(value, result, &bail);
        break;
      case MIRType::String:
        masm.fallibleUnboxString(value, result, &bail);
        break;
      case MIRType::Symbol:
        masm.fallibleUnboxSymbol(value, result, &bail);
        break;
      case MIRType::BigInt:
        masm.fallibleUnboxBigInt(value, result, &bail);
        break;
      default:
        MOZ_CRASH("Given MIRType cannot be unboxed.");
    }
    bailoutFrom(&bail, unbox->snapshot());
    return;
  }

  // Infallible unbox.

  ValueOperand input = ToValue(unbox->input());

#ifdef DEBUG
  // Assert the types match.
  JSValueTag tag = MIRTypeToTag(mir->type());
  Label ok;
  {
    ScratchTagScope scratch(masm, input);
    masm.splitTagForTest(input, scratch);
    masm.cmpTag(scratch, ImmTag(tag));
  }
  masm.B(&ok, Assembler::Condition::Equal);
  masm.assumeUnreachable("Infallible unbox type mismatch");
  masm.bind(&ok);
#endif

  switch (mir->type()) {
    case MIRType::Int32:
      masm.unboxInt32(input, result);
      break;
    case MIRType::Boolean:
      masm.unboxBoolean(input, result);
      break;
    case MIRType::Object:
      masm.unboxObject(input, result);
      break;
    case MIRType::String:
      masm.unboxString(input, result);
      break;
    case MIRType::Symbol:
      masm.unboxSymbol(input, result);
      break;
    case MIRType::BigInt:
      masm.unboxBigInt(input, result);
      break;
    default:
      MOZ_CRASH("Given MIRType cannot be unboxed.");
  }
}

void CodeGenerator::visitTestDAndBranch(LTestDAndBranch* test) {
  const LAllocation* opd = test->input();
  MBasicBlock* ifTrue = test->ifTrue();
  MBasicBlock* ifFalse = test->ifFalse();

  masm.Fcmp(ARMFPRegister(ToFloatRegister(opd), 64), 0.0);

  // If the compare set the 0 bit, then the result is definitely false.
  jumpToBlock(ifFalse, Assembler::Zero);

  // Overflow means one of the operands was NaN, which is also false.
  jumpToBlock(ifFalse, Assembler::Overflow);
  jumpToBlock(ifTrue);
}

void CodeGenerator::visitTestFAndBranch(LTestFAndBranch* test) {
  const LAllocation* opd = test->input();
  MBasicBlock* ifTrue = test->ifTrue();
  MBasicBlock* ifFalse = test->ifFalse();

  masm.Fcmp(ARMFPRegister(ToFloatRegister(opd), 32), 0.0);

  // If the compare set the 0 bit, then the result is definitely false.
  jumpToBlock(ifFalse, Assembler::Zero);

  // Overflow means one of the operands was NaN, which is also false.
  jumpToBlock(ifFalse, Assembler::Overflow);
  jumpToBlock(ifTrue);
}

void CodeGenerator::visitCompareD(LCompareD* comp) {
  const FloatRegister left = ToFloatRegister(comp->left());
  const FloatRegister right = ToFloatRegister(comp->right());
  ARMRegister output = toWRegister(comp->output());
  Assembler::DoubleCondition cond = JSOpToDoubleCondition(comp->mir()->jsop());

  masm.compareDouble(left, right);
  masm.cset(output, Assembler::ConditionFromDoubleCondition(cond));
}

void CodeGenerator::visitCompareF(LCompareF* comp) {
  const FloatRegister left = ToFloatRegister(comp->left());
  const FloatRegister right = ToFloatRegister(comp->right());
  ARMRegister output = toWRegister(comp->output());
  Assembler::DoubleCondition cond = JSOpToDoubleCondition(comp->mir()->jsop());

  masm.compareFloat(left, right);
  masm.cset(output, Assembler::ConditionFromDoubleCondition(cond));
}

void CodeGenerator::visitCompareDAndBranch(LCompareDAndBranch* comp) {
  const FloatRegister left = ToFloatRegister(comp->left());
  const FloatRegister right = ToFloatRegister(comp->right());
  Assembler::DoubleCondition doubleCond =
      JSOpToDoubleCondition(comp->cmpMir()->jsop());
  Assembler::Condition cond =
      Assembler::ConditionFromDoubleCondition(doubleCond);

  masm.compareDouble(left, right);
  emitBranch(cond, comp->ifTrue(), comp->ifFalse());
}

void CodeGenerator::visitCompareFAndBranch(LCompareFAndBranch* comp) {
  const FloatRegister left = ToFloatRegister(comp->left());
  const FloatRegister right = ToFloatRegister(comp->right());
  Assembler::DoubleCondition doubleCond =
      JSOpToDoubleCondition(comp->cmpMir()->jsop());
  Assembler::Condition cond =
      Assembler::ConditionFromDoubleCondition(doubleCond);

  masm.compareFloat(left, right);
  emitBranch(cond, comp->ifTrue(), comp->ifFalse());
}

void CodeGenerator::visitWasmUint32ToDouble(LWasmUint32ToDouble* lir) {
  masm.convertUInt32ToDouble(ToRegister(lir->input()),
                             ToFloatRegister(lir->output()));
}

void CodeGenerator::visitWasmUint32ToFloat32(LWasmUint32ToFloat32* lir) {
  masm.convertUInt32ToFloat32(ToRegister(lir->input()),
                              ToFloatRegister(lir->output()));
}

//        NZCV
// NAN -> 0011
// ==  -> 0110
// <   -> 1000
// >   -> 0010
void CodeGenerator::visitNotD(LNotD* ins) {
  ARMFPRegister input(ToFloatRegister(ins->input()), 64);
  ARMRegister output = toWRegister(ins->output());

  // Set output to 1 if input compares equal to 0.0, else 0.
  masm.Fcmp(input, 0.0);
  masm.Cset(output, Assembler::Equal);

  // Comparison with NaN sets V in the NZCV register.
  // If the input was NaN, output must now be zero, so it can be incremented.
  // The instruction is read: "output = if NoOverflow then output else 0+1".
  masm.Csinc(output, output, ZeroRegister32, Assembler::NoOverflow);
}

void CodeGenerator::visitNotF(LNotF* ins) {
  ARMFPRegister input(ToFloatRegister(ins->input()), 32);
  ARMRegister output = toWRegister(ins->output());

  // Set output to 1 input compares equal to 0.0, else 0.
  masm.Fcmp(input, 0.0);
  masm.Cset(output, Assembler::Equal);

  // Comparison with NaN sets V in the NZCV register.
  // If the input was NaN, output must now be zero, so it can be incremented.
  // The instruction is read: "output = if NoOverflow then output else 0+1".
  masm.Csinc(output, output, ZeroRegister32, Assembler::NoOverflow);
}

void CodeGeneratorARM64::generateInvalidateEpilogue() {
  // Ensure that there is enough space in the buffer for the OsiPoint patching
  // to occur. Otherwise, we could overwrite the invalidation epilogue.
  for (size_t i = 0; i < sizeof(void*); i += Assembler::NopSize()) {
    masm.nop();
  }

  masm.bind(&invalidate_);

  // Push the return address of the point that we bailout out onto the stack.
  masm.push(lr);

  // Push the Ion script onto the stack (when we determine what that pointer
  // is).
  invalidateEpilogueData_ = masm.pushWithPatch(ImmWord(uintptr_t(-1)));

  // Jump to the invalidator which will replace the current frame.
  TrampolinePtr thunk = gen->jitRuntime()->getInvalidationThunk();
  masm.jump(thunk);
}

template <class U>
Register getBase(U* mir) {
  switch (mir->base()) {
    case U::Heap:
      return HeapReg;
  }
  return InvalidReg;
}

void CodeGenerator::visitAsmJSLoadHeap(LAsmJSLoadHeap* ins) {
  const MAsmJSLoadHeap* mir = ins->mir();
  MOZ_ASSERT(!mir->hasMemoryBase());

  const LAllocation* ptr = ins->ptr();
  const LAllocation* boundsCheckLimit = ins->boundsCheckLimit();

  Register ptrReg = ToRegister(ptr);
  Scalar::Type accessType = mir->accessType();
  bool isFloat = accessType == Scalar::Float32 || accessType == Scalar::Float64;
  Label done;

  if (mir->needsBoundsCheck()) {
    Label boundsCheckPassed;
    Register boundsCheckLimitReg = ToRegister(boundsCheckLimit);
    masm.wasmBoundsCheck32(Assembler::Below, ptrReg, boundsCheckLimitReg,
                           &boundsCheckPassed);
    // Return a default value in case of a bounds-check failure.
    if (isFloat) {
      if (accessType == Scalar::Float32) {
        masm.loadConstantFloat32(GenericNaN(), ToFloatRegister(ins->output()));
      } else {
        masm.loadConstantDouble(GenericNaN(), ToFloatRegister(ins->output()));
      }
    } else {
      masm.Mov(ARMRegister(ToRegister(ins->output()), 64), 0);
    }
    masm.jump(&done);
    masm.bind(&boundsCheckPassed);
  }

  MemOperand addr(ARMRegister(HeapReg, 64), ARMRegister(ptrReg, 64));
  switch (accessType) {
    case Scalar::Int8:
      masm.Ldrb(toWRegister(ins->output()), addr);
      masm.Sxtb(toWRegister(ins->output()), toWRegister(ins->output()));
      break;
    case Scalar::Uint8:
      masm.Ldrb(toWRegister(ins->output()), addr);
      break;
    case Scalar::Int16:
      masm.Ldrh(toWRegister(ins->output()), addr);
      masm.Sxth(toWRegister(ins->output()), toWRegister(ins->output()));
      break;
    case Scalar::Uint16:
      masm.Ldrh(toWRegister(ins->output()), addr);
      break;
    case Scalar::Int32:
    case Scalar::Uint32:
      masm.Ldr(toWRegister(ins->output()), addr);
      break;
    case Scalar::Float64:
      masm.Ldr(ARMFPRegister(ToFloatRegister(ins->output()), 64), addr);
      break;
    case Scalar::Float32:
      masm.Ldr(ARMFPRegister(ToFloatRegister(ins->output()), 32), addr);
      break;
    default:
      MOZ_CRASH("unexpected array type");
  }
  if (done.used()) {
    masm.bind(&done);
  }
}

void CodeGenerator::visitAsmJSStoreHeap(LAsmJSStoreHeap* ins) {
  const MAsmJSStoreHeap* mir = ins->mir();
  MOZ_ASSERT(!mir->hasMemoryBase());

  const LAllocation* ptr = ins->ptr();
  const LAllocation* boundsCheckLimit = ins->boundsCheckLimit();

  Register ptrReg = ToRegister(ptr);

  Label done;
  if (mir->needsBoundsCheck()) {
    Register boundsCheckLimitReg = ToRegister(boundsCheckLimit);
    masm.wasmBoundsCheck32(Assembler::AboveOrEqual, ptrReg, boundsCheckLimitReg,
                           &done);
  }

  MemOperand addr(ARMRegister(HeapReg, 64), ARMRegister(ptrReg, 64));
  switch (mir->accessType()) {
    case Scalar::Int8:
    case Scalar::Uint8:
      masm.Strb(toWRegister(ins->value()), addr);
      break;
    case Scalar::Int16:
    case Scalar::Uint16:
      masm.Strh(toWRegister(ins->value()), addr);
      break;
    case Scalar::Int32:
    case Scalar::Uint32:
      masm.Str(toWRegister(ins->value()), addr);
      break;
    case Scalar::Float64:
      masm.Str(ARMFPRegister(ToFloatRegister(ins->value()), 64), addr);
      break;
    case Scalar::Float32:
      masm.Str(ARMFPRegister(ToFloatRegister(ins->value()), 32), addr);
      break;
    default:
      MOZ_CRASH("unexpected array type");
  }
  if (done.used()) {
    masm.bind(&done);
  }
}

void CodeGenerator::visitWasmCompareExchangeHeap(
    LWasmCompareExchangeHeap* ins) {
  MWasmCompareExchangeHeap* mir = ins->mir();

  Register memoryBase = ToRegister(ins->memoryBase());
  Register ptr = ToRegister(ins->ptr());
  Register oldval = ToRegister(ins->oldValue());
  Register newval = ToRegister(ins->newValue());
  Register out = ToRegister(ins->output());

  BaseIndex srcAddr(memoryBase, ptr, TimesOne, mir->access().offset32());

  if (mir->access().type() == Scalar::Int64) {
    masm.wasmCompareExchange64(mir->access(), srcAddr, Register64(oldval),
                               Register64(newval), Register64(out));
  } else {
    masm.wasmCompareExchange(mir->access(), srcAddr, oldval, newval, out);
  }
}

void CodeGenerator::visitWasmAtomicExchangeHeap(LWasmAtomicExchangeHeap* ins) {
  MWasmAtomicExchangeHeap* mir = ins->mir();

  Register memoryBase = ToRegister(ins->memoryBase());
  Register ptr = ToRegister(ins->ptr());
  Register oldval = ToRegister(ins->value());
  Register out = ToRegister(ins->output());

  BaseIndex srcAddr(memoryBase, ptr, TimesOne, mir->access().offset32());

  if (mir->access().type() == Scalar::Int64) {
    masm.wasmAtomicExchange64(mir->access(), srcAddr, Register64(oldval),
                              Register64(out));
  } else {
    masm.wasmAtomicExchange(mir->access(), srcAddr, oldval, out);
  }
}

void CodeGenerator::visitWasmAtomicBinopHeap(LWasmAtomicBinopHeap* ins) {
  MWasmAtomicBinopHeap* mir = ins->mir();

  MOZ_ASSERT(mir->hasUses());

  Register memoryBase = ToRegister(ins->memoryBase());
  Register ptr = ToRegister(ins->ptr());
  Register value = ToRegister(ins->value());
  Register flagTemp = ToRegister(ins->temp0());
  Register out = ToRegister(ins->output());

  BaseIndex srcAddr(memoryBase, ptr, TimesOne, mir->access().offset32());
  AtomicOp op = mir->operation();

  if (mir->access().type() == Scalar::Int64) {
    masm.wasmAtomicFetchOp64(mir->access(), op, Register64(value), srcAddr,
                             Register64(flagTemp), Register64(out));
  } else {
    masm.wasmAtomicFetchOp(mir->access(), op, value, srcAddr, flagTemp, out);
  }
}

void CodeGenerator::visitWasmAtomicBinopHeapForEffect(
    LWasmAtomicBinopHeapForEffect* ins) {
  MWasmAtomicBinopHeap* mir = ins->mir();

  MOZ_ASSERT(!mir->hasUses());

  Register memoryBase = ToRegister(ins->memoryBase());
  Register ptr = ToRegister(ins->ptr());
  Register value = ToRegister(ins->value());
  Register flagTemp = ToRegister(ins->temp0());

  BaseIndex srcAddr(memoryBase, ptr, TimesOne, mir->access().offset32());
  AtomicOp op = mir->operation();

  if (mir->access().type() == Scalar::Int64) {
    masm.wasmAtomicEffectOp64(mir->access(), op, Register64(value), srcAddr,
                              Register64(flagTemp));
  } else {
    masm.wasmAtomicEffectOp(mir->access(), op, value, srcAddr, flagTemp);
  }
}

void CodeGenerator::visitWasmStackArg(LWasmStackArg* ins) {
  const MWasmStackArg* mir = ins->mir();
  Address dst(masm.getStackPointer(), mir->spOffset());
  if (ins->arg()->isConstant()) {
    masm.storePtr(ImmWord(ToInt32(ins->arg())), dst);
  } else if (ins->arg()->isGeneralReg()) {
    masm.storePtr(ToRegister(ins->arg()), dst);
  } else {
    switch (mir->input()->type()) {
      case MIRType::Double:
        masm.storeDouble(ToFloatRegister(ins->arg()), dst);
        return;
      case MIRType::Float32:
        masm.storeFloat32(ToFloatRegister(ins->arg()), dst);
        return;
#ifdef ENABLE_WASM_SIMD
      case MIRType::Simd128:
        masm.storeUnalignedSimd128(ToFloatRegister(ins->arg()), dst);
        return;
#endif
      default:
        break;
    }
    MOZ_MAKE_COMPILER_ASSUME_IS_UNREACHABLE(
        "unexpected mir type in WasmStackArg");
  }
}

void CodeGenerator::visitUDiv(LUDiv* ins) {
  MDiv* mir = ins->mir();
  Register lhs = ToRegister(ins->lhs());
  Register rhs = ToRegister(ins->rhs());
  Register output = ToRegister(ins->output());
  ARMRegister lhs32 = ARMRegister(lhs, 32);
  ARMRegister rhs32 = ARMRegister(rhs, 32);
  ARMRegister output32 = ARMRegister(output, 32);

  // Prevent divide by zero.
  if (mir->canBeDivideByZero()) {
    if (mir->trapOnError()) {
      TrapIfDivideByZero(masm, ins, rhs32);
    } else if (mir->canTruncateInfinities()) {
      // Udiv returns zero for division by zero, exactly what we want for
      // truncated division. Remainder computation expects a non-zero divisor,
      // so we must also be allowed to truncate the remainder.
      MOZ_ASSERT(mir->canTruncateRemainder(),
                 "remainder computation expects a non-zero divisor");
    } else {
      MOZ_ASSERT(mir->fallible());
      bailoutTest32(Assembler::Zero, rhs, rhs, ins->snapshot());
    }
  }

  // Unsigned division.
  masm.Udiv(output32, lhs32, rhs32);

  // If the remainder is > 0, bailout since this must be a double.
  if (!mir->canTruncateRemainder()) {
    vixl::UseScratchRegisterScope temps(&masm.asVIXL());
    ARMRegister remainder32 = temps.AcquireW();
    Register remainder = remainder32.asUnsized();

    // Compute the remainder: remainder = lhs - (output * rhs).
    masm.Msub(remainder32, output32, rhs32, lhs32);

    bailoutTest32(Assembler::NonZero, remainder, remainder, ins->snapshot());
  }

  // Unsigned div can return a value that's not a signed int32.
  // If our users aren't expecting that, bail.
  if (!mir->isTruncated()) {
    bailoutTest32(Assembler::Signed, output, output, ins->snapshot());
  }
}

void CodeGenerator::visitUMod(LUMod* ins) {
  Register rhs = ToRegister(ins->rhs());
  Register output = ToRegister(ins->output());

  ARMRegister lhs32 = toWRegister(ins->lhs());
  ARMRegister rhs32 = toWRegister(ins->rhs());
  ARMRegister output32 = toWRegister(ins->output());
  Label done;

  MMod* mir = ins->mir();

  if (mir->canBeDivideByZero()) {
    if (mir->trapOnError()) {
      TrapIfDivideByZero(masm, ins, rhs32);
    } else if (mir->isTruncated()) {
      // Truncated division by zero yields integer zero.
      masm.Mov(output32, wzr);
      masm.Cbz(rhs32, &done);
    } else {
      // Non-truncated division by zero produces a non-integer.
      bailoutTest32(Assembler::Zero, rhs, rhs, ins->snapshot());
    }
  }

  // Unsigned division.
  masm.Udiv(output32, lhs32, rhs32);

  // Compute the remainder: output = lhs - (output * rhs).
  masm.Msub(output32, output32, rhs32, lhs32);

  if (!mir->isTruncated()) {
    // Bail if the output would be negative.
    //
    // LUMod inputs may be Uint32, so care is taken to ensure the result
    // is not unexpectedly signed.
    bailoutCmp32(Assembler::LessThan, output, Imm32(0), ins->snapshot());
  }

  if (done.used()) {
    masm.bind(&done);
  }
}

void CodeGenerator::visitEffectiveAddress3(LEffectiveAddress3* ins) {
  const MEffectiveAddress3* mir = ins->mir();
  const ARMRegister base = toWRegister(ins->base());
  const ARMRegister index = toWRegister(ins->index());
  const ARMRegister output = toWRegister(ins->output());

  if (mir->scale() == Scale::TimesOne) {
    masm.Add(output, base, Operand(index));
  } else {
    masm.Add(output, base, Operand(index, vixl::LSL, mir->scale()));
  }
  if (mir->displacement() != 0) {
    masm.Add(output, output, Operand(mir->displacement()));
  }
}

void CodeGenerator::visitEffectiveAddress2(LEffectiveAddress2* ins) {
  const MEffectiveAddress2* mir = ins->mir();
  const ARMRegister index = toWRegister(ins->index());
  const ARMRegister output = toWRegister(ins->output());
  // Ensured because the LIR's `index` input is not an AtStart variant.
  // If this ever fails, we'll need to generate a slower sequence the same
  // as ::visitEffectiveAddress, but with `base` being `wzr`.
  MOZ_RELEASE_ASSERT(output.code() != index.code());

  masm.Mov(output, mir->displacement());
  masm.Add(output, output, Operand(index, vixl::LSL, mir->scale()));
}

void CodeGenerator::visitNegI(LNegI* ins) {
  const ARMRegister input = toWRegister(ins->input());
  const ARMRegister output = toWRegister(ins->output());
  masm.Neg(output, input);
}

void CodeGenerator::visitNegI64(LNegI64* ins) {
  const ARMRegister input = toXRegister(ins->input());
  const ARMRegister output = toXRegister(ins->output());
  masm.Neg(output, input);
}

void CodeGenerator::visitNegD(LNegD* ins) {
  const ARMFPRegister input(ToFloatRegister(ins->input()), 64);
  const ARMFPRegister output(ToFloatRegister(ins->output()), 64);
  masm.Fneg(output, input);
}

void CodeGenerator::visitNegF(LNegF* ins) {
  const ARMFPRegister input(ToFloatRegister(ins->input()), 32);
  const ARMFPRegister output(ToFloatRegister(ins->output()), 32);
  masm.Fneg(output, input);
}

void CodeGenerator::visitCompareExchangeTypedArrayElement(
    LCompareExchangeTypedArrayElement* lir) {
  Register elements = ToRegister(lir->elements());
  AnyRegister output = ToAnyRegister(lir->output());
  Register temp = ToTempRegisterOrInvalid(lir->temp0());

  Register oldval = ToRegister(lir->oldval());
  Register newval = ToRegister(lir->newval());

  Scalar::Type arrayType = lir->mir()->arrayType();

  auto dest = ToAddressOrBaseIndex(elements, lir->index(), arrayType);

  dest.match([&](const auto& dest) {
    masm.compareExchangeJS(arrayType, Synchronization::Full(), dest, oldval,
                           newval, temp, output);
  });
}

void CodeGenerator::visitAtomicExchangeTypedArrayElement(
    LAtomicExchangeTypedArrayElement* lir) {
  Register elements = ToRegister(lir->elements());
  AnyRegister output = ToAnyRegister(lir->output());
  Register temp = ToTempRegisterOrInvalid(lir->temp0());

  Register value = ToRegister(lir->value());

  Scalar::Type arrayType = lir->mir()->arrayType();

  auto dest = ToAddressOrBaseIndex(elements, lir->index(), arrayType);

  dest.match([&](const auto& dest) {
    masm.atomicExchangeJS(arrayType, Synchronization::Full(), dest, value, temp,
                          output);
  });
}

void CodeGenerator::visitAtomicLoad64(LAtomicLoad64* lir) {
  Register elements = ToRegister(lir->elements());
  Register64 out = ToOutRegister64(lir);

  Scalar::Type storageType = lir->mir()->storageType();

  auto source = ToAddressOrBaseIndex(elements, lir->index(), storageType);

  // NOTE: the generated code must match the assembly code in gen_load in
  // GenerateAtomicOperations.py
  auto sync = Synchronization::Load();

  masm.memoryBarrierBefore(sync);
  source.match([&](const auto& source) { masm.load64(source, out); });
  masm.memoryBarrierAfter(sync);
}

void CodeGenerator::visitAtomicStore64(LAtomicStore64* lir) {
  Register elements = ToRegister(lir->elements());
  Register64 value = ToRegister64(lir->value());

  Scalar::Type writeType = lir->mir()->writeType();

  auto dest = ToAddressOrBaseIndex(elements, lir->index(), writeType);

  // NOTE: the generated code must match the assembly code in gen_store in
  // GenerateAtomicOperations.py
  auto sync = Synchronization::Store();

  masm.memoryBarrierBefore(sync);
  dest.match([&](const auto& dest) { masm.store64(value, dest); });
  masm.memoryBarrierAfter(sync);
}

void CodeGenerator::visitCompareExchangeTypedArrayElement64(
    LCompareExchangeTypedArrayElement64* lir) {
  Register elements = ToRegister(lir->elements());
  Register64 oldval = ToRegister64(lir->oldval());
  Register64 newval = ToRegister64(lir->newval());
  Register64 out = ToOutRegister64(lir);

  Scalar::Type arrayType = lir->mir()->arrayType();

  auto dest = ToAddressOrBaseIndex(elements, lir->index(), arrayType);

  dest.match([&](const auto& dest) {
    masm.compareExchange64(Synchronization::Full(), dest, oldval, newval, out);
  });
}

void CodeGenerator::visitAtomicExchangeTypedArrayElement64(
    LAtomicExchangeTypedArrayElement64* lir) {
  Register elements = ToRegister(lir->elements());
  Register64 value = ToRegister64(lir->value());
  Register64 out = ToOutRegister64(lir);

  Scalar::Type arrayType = lir->mir()->arrayType();

  auto dest = ToAddressOrBaseIndex(elements, lir->index(), arrayType);

  dest.match([&](const auto& dest) {
    masm.atomicExchange64(Synchronization::Full(), dest, value, out);
  });
}

void CodeGenerator::visitAtomicTypedArrayElementBinop64(
    LAtomicTypedArrayElementBinop64* lir) {
  MOZ_ASSERT(!lir->mir()->isForEffect());

  Register elements = ToRegister(lir->elements());
  Register64 value = ToRegister64(lir->value());
  Register64 temp = ToRegister64(lir->temp0());
  Register64 out = ToOutRegister64(lir);

  Scalar::Type arrayType = lir->mir()->arrayType();
  AtomicOp atomicOp = lir->mir()->operation();

  auto dest = ToAddressOrBaseIndex(elements, lir->index(), arrayType);

  dest.match([&](const auto& dest) {
    masm.atomicFetchOp64(Synchronization::Full(), atomicOp, value, dest, temp,
                         out);
  });
}

void CodeGenerator::visitAtomicTypedArrayElementBinopForEffect64(
    LAtomicTypedArrayElementBinopForEffect64* lir) {
  MOZ_ASSERT(lir->mir()->isForEffect());

  Register elements = ToRegister(lir->elements());
  Register64 value = ToRegister64(lir->value());
  Register64 temp = ToRegister64(lir->temp0());

  Scalar::Type arrayType = lir->mir()->arrayType();
  AtomicOp atomicOp = lir->mir()->operation();

  auto dest = ToAddressOrBaseIndex(elements, lir->index(), arrayType);

  dest.match([&](const auto& dest) {
    masm.atomicEffectOp64(Synchronization::Full(), atomicOp, value, dest, temp);
  });
}

void CodeGenerator::visitAddI64(LAddI64* lir) {
  ARMRegister dest = toXRegister(lir->output());
  ARMRegister lhs = toXRegister(lir->lhs());
  Operand rhs = toXOperand(lir->rhs());

  masm.Add(dest, lhs, rhs);
}

void CodeGenerator::visitMulI64(LMulI64* lir) {
  LInt64Allocation lhs = lir->lhs();
  LInt64Allocation rhs = lir->rhs();
  Register64 output = ToOutRegister64(lir);

  if (IsConstant(rhs)) {
    int64_t constant = ToInt64(rhs);
    // Ad-hoc strength reduction, cf the x64 code as well as the 32-bit code
    // higher up in this file.  Bug 1712298 will lift this code to the MIR
    // constant folding pass, or to lowering.
    //
    // This is for wasm integers only, so no input guards or overflow checking
    // are needed.
    switch (constant) {
      case -1:
        masm.Neg(ARMRegister(output.reg, 64),
                 ARMRegister(ToRegister64(lhs).reg, 64));
        break;
      case 0:
        masm.Mov(ARMRegister(output.reg, 64), xzr);
        break;
      case 1:
        if (ToRegister64(lhs) != output) {
          masm.move64(ToRegister64(lhs), output);
        }
        break;
      case 2:
        masm.Add(ARMRegister(output.reg, 64),
                 ARMRegister(ToRegister64(lhs).reg, 64),
                 ARMRegister(ToRegister64(lhs).reg, 64));
        break;
      default:
        // Use shift if constant is nonnegative power of 2.
        if (constant > 0) {
          int32_t shift = mozilla::FloorLog2(uint64_t(constant));
          if (int64_t(1) << shift == constant) {
            masm.Lsl(ARMRegister(output.reg, 64),
                     ARMRegister(ToRegister64(lhs).reg, 64), shift);
            break;
          }
        }
        masm.mul64(Imm64(constant), ToRegister64(lhs), output);
        break;
    }
  } else {
    masm.mul64(ToRegister64(lhs), ToRegister64(rhs), output);
  }
}

void CodeGenerator::visitSubI64(LSubI64* lir) {
  ARMRegister dest = toXRegister(lir->output());
  ARMRegister lhs = toXRegister(lir->lhs());
  Operand rhs = toXOperand(lir->rhs());

  masm.Sub(dest, lhs, rhs);
}

void CodeGenerator::visitBitOpI64(LBitOpI64* lir) {
  ARMRegister dest = toXRegister(lir->output());
  ARMRegister lhs = toXRegister(lir->lhs());
  Operand rhs = toXOperand(lir->rhs());

  switch (lir->bitop()) {
    case JSOp::BitOr:
      masm.Orr(dest, lhs, rhs);
      break;
    case JSOp::BitXor:
      masm.Eor(dest, lhs, rhs);
      break;
    case JSOp::BitAnd:
      masm.And(dest, lhs, rhs);
      break;
    default:
      MOZ_CRASH("unexpected binary opcode");
  }
}

void CodeGenerator::visitShiftI64(LShiftI64* lir) {
  ARMRegister lhs(ToRegister64(lir->lhs()).reg, 64);
  const LAllocation* rhsAlloc = lir->rhs();
  ARMRegister dest(ToOutRegister64(lir).reg, 64);

  if (rhsAlloc->isConstant()) {
    int32_t shift = int32_t(rhsAlloc->toConstant()->toInt64() & 0x3F);
    if (shift == 0) {
      masm.Mov(dest, lhs);
    } else {
      switch (lir->bitop()) {
        case JSOp::Lsh:
          masm.Lsl(dest, lhs, shift);
          break;
        case JSOp::Rsh:
          masm.Asr(dest, lhs, shift);
          break;
        case JSOp::Ursh:
          masm.Lsr(dest, lhs, shift);
          break;
        default:
          MOZ_CRASH("Unexpected shift op");
      }
    }
  } else {
    ARMRegister rhs(ToRegister(rhsAlloc), 64);
    switch (lir->bitop()) {
      case JSOp::Lsh:
        masm.Lsl(dest, lhs, rhs);
        break;
      case JSOp::Rsh:
        masm.Asr(dest, lhs, rhs);
        break;
      case JSOp::Ursh:
        masm.Lsr(dest, lhs, rhs);
        break;
      default:
        MOZ_CRASH("Unexpected shift op");
    }
  }
}

void CodeGenerator::visitAddIntPtr(LAddIntPtr* ins) {
  ARMRegister lhs = toXRegister(ins->lhs());
  Operand rhs = toXOperand(ins->rhs());
  ARMRegister dest = toXRegister(ins->output());

  masm.Add(dest, lhs, rhs);
}

void CodeGenerator::visitSubIntPtr(LSubIntPtr* ins) {
  ARMRegister lhs = toXRegister(ins->lhs());
  Operand rhs = toXOperand(ins->rhs());
  ARMRegister dest = toXRegister(ins->output());

  masm.Sub(dest, lhs, rhs);
}

void CodeGenerator::visitMulIntPtr(LMulIntPtr* ins) {
  ARMRegister lhs = toXRegister(ins->lhs());
  const LAllocation* rhs = ins->rhs();
  ARMRegister dest = toXRegister(ins->output());

  if (rhs->isConstant()) {
    intptr_t constant = ToIntPtr(rhs);

    switch (constant) {
      case -1:
        masm.Neg(dest, lhs);
        return;
      case 0:
        masm.Mov(dest, xzr);
        return;
      case 1:
        if (!dest.Is(lhs)) {
          masm.Mov(dest, lhs);
        }
        return;
      case 2:
        masm.Add(dest, lhs, lhs);
        return;
    }

    // Use shift if constant is a power of 2.
    if (constant > 0 && std::has_single_bit(uintptr_t(constant))) {
      uint32_t shift = mozilla::FloorLog2(uintptr_t(constant));
      masm.Lsl(dest, lhs, shift);
      return;
    }

    vixl::UseScratchRegisterScope temps(&masm.asVIXL());
    vixl::Register scratch = temps.AcquireX();

    masm.Mov(scratch, constant);
    masm.Mul(dest, lhs, scratch);
  } else {
    masm.Mul(dest, lhs, toXRegister(rhs));
  }
}

// If we have a constant base ptr, try to add the offset to it, to generate
// better code when the full address is known.  The addition may overflow past
// 32 bits because the front end does nothing special if the base is a large
// constant and base+offset overflows; sidestep this by performing the addition
// anyway, overflowing to 64-bit.

static Maybe<uint64_t> IsAbsoluteAddress(const LAllocation* ptr,
                                         const wasm::MemoryAccessDesc& access) {
  if (ptr->isConstantValue()) {
    const MConstant* c = ptr->toConstant();
    uint64_t base_address = c->type() == MIRType::Int32
                                ? uint64_t(uint32_t(c->toInt32()))
                                : uint64_t(c->toInt64());
    uint64_t offset = access.offset32();
    return Some(base_address + offset);
  }
  return Nothing();
}

void CodeGenerator::visitWasmLoad(LWasmLoad* lir) {
  const MWasmLoad* mir = lir->mir();

  if (Maybe<uint64_t> absAddr = IsAbsoluteAddress(lir->ptr(), mir->access())) {
    masm.wasmLoadAbsolute(mir->access(), ToRegister(lir->memoryBase()),
                          absAddr.value(), ToAnyRegister(lir->output()),
                          Register64::Invalid());
    return;
  }

  // ptr is a GPR and is either a 32-bit value zero-extended to 64-bit, or a
  // true 64-bit value.
  masm.wasmLoad(mir->access(), ToRegister(lir->memoryBase()),
                ToRegister(lir->ptr()), ToAnyRegister(lir->output()));
}

void CodeGenerator::visitWasmStore(LWasmStore* lir) {
  const MWasmStore* mir = lir->mir();

  if (Maybe<uint64_t> absAddr = IsAbsoluteAddress(lir->ptr(), mir->access())) {
    masm.wasmStoreAbsolute(mir->access(), ToAnyRegister(lir->value()),
                           Register64::Invalid(), ToRegister(lir->memoryBase()),
                           absAddr.value());
    return;
  }

  masm.wasmStore(mir->access(), ToAnyRegister(lir->value()),
                 ToRegister(lir->memoryBase()), ToRegister(lir->ptr()));
}

void CodeGenerator::visitWasmSelect(LWasmSelect* lir) {
  MIRType mirType = lir->mir()->type();
  Register condReg = ToRegister(lir->condExpr());

  masm.test32(condReg, condReg);

  switch (mirType) {
    case MIRType::Int32:
    case MIRType::WasmAnyRef:
    case MIRType::WasmArrayData: {
      Register outReg = ToRegister(lir->output());
      Register trueReg = ToRegister(lir->trueExpr());
      Register falseReg = ToRegister(lir->falseExpr());

      if (mirType == MIRType::Int32) {
        masm.Csel(ARMRegister(outReg, 32), ARMRegister(trueReg, 32),
                  ARMRegister(falseReg, 32), Assembler::NonZero);
      } else {
        masm.Csel(ARMRegister(outReg, 64), ARMRegister(trueReg, 64),
                  ARMRegister(falseReg, 64), Assembler::NonZero);
      }
      break;
    }

    case MIRType::Float32:
    case MIRType::Double:
    case MIRType::Simd128: {
      FloatRegister outReg = ToFloatRegister(lir->output());
      FloatRegister trueReg = ToFloatRegister(lir->trueExpr());
      FloatRegister falseReg = ToFloatRegister(lir->falseExpr());

      switch (mirType) {
        case MIRType::Float32:
          masm.Fcsel(ARMFPRegister(outReg, 32), ARMFPRegister(trueReg, 32),
                     ARMFPRegister(falseReg, 32), Assembler::NonZero);
          break;
        case MIRType::Double:
          masm.Fcsel(ARMFPRegister(outReg, 64), ARMFPRegister(trueReg, 64),
                     ARMFPRegister(falseReg, 64), Assembler::NonZero);
          break;
#ifdef ENABLE_WASM_SIMD
        case MIRType::Simd128: {
          MOZ_ASSERT(outReg == trueReg);
          Label done;
          masm.j(Assembler::NonZero, &done);
          masm.moveSimd128(falseReg, outReg);
          masm.bind(&done);
          break;
        }
#endif
        default:
          MOZ_CRASH();
      }
      break;
    }

    default: {
      MOZ_CRASH("unhandled type in visitWasmSelect!");
    }
  }
}

// We expect to handle the cases: compare is {{U,}Int32, {U,}Int64}, Float32,
// Double}, and select is {{U,}Int32, {U,}Int64}, Float32, Double},
// independently.
void CodeGenerator::visitWasmCompareAndSelect(LWasmCompareAndSelect* ins) {
  MCompare::CompareType compTy = ins->compareType();

  // Set flag.
  if (compTy == MCompare::Compare_Int32 || compTy == MCompare::Compare_UInt32) {
    Register lhs = ToRegister(ins->leftExpr());
    if (ins->rightExpr()->isConstant()) {
      masm.cmp32(lhs, Imm32(ins->rightExpr()->toConstant()->toInt32()));
    } else {
      masm.cmp32(lhs, ToRegister(ins->rightExpr()));
    }
  } else if (compTy == MCompare::Compare_Int64 ||
             compTy == MCompare::Compare_UInt64) {
    Register lhs = ToRegister(ins->leftExpr());
    if (ins->rightExpr()->isConstant()) {
      masm.cmpPtr(lhs, Imm64(ins->rightExpr()->toConstant()->toInt64()));
    } else {
      masm.cmpPtr(lhs, ToRegister(ins->rightExpr()));
    }
  } else if (compTy == MCompare::Compare_Float32) {
    masm.compareFloat(ToFloatRegister(ins->leftExpr()),
                      ToFloatRegister(ins->rightExpr()));
  } else if (compTy == MCompare::Compare_Double) {
    masm.compareDouble(ToFloatRegister(ins->leftExpr()),
                       ToFloatRegister(ins->rightExpr()));
  } else {
    // Ref types not supported yet; v128 is not yet observed to be worth
    // optimizing.
    MOZ_CRASH("CodeGenerator::visitWasmCompareAndSelect: unexpected type (1)");
  }

  // Act on flag.
  Assembler::Condition cond;
  if (compTy == MCompare::Compare_Float32 ||
      compTy == MCompare::Compare_Double) {
    cond = Assembler::ConditionFromDoubleCondition(
        JSOpToDoubleCondition(ins->jsop()));
  } else {
    cond = JSOpToCondition(compTy, ins->jsop());
  }
  MIRType insTy = ins->mir()->type();
  if (insTy == MIRType::Int32 || insTy == MIRType::Int64) {
    Register destReg = ToRegister(ins->output());
    Register trueReg = ToRegister(ins->ifTrueExpr());
    Register falseReg = ToRegister(ins->ifFalseExpr());
    size_t size = insTy == MIRType::Int32 ? 32 : 64;
    masm.Csel(ARMRegister(destReg, size), ARMRegister(trueReg, size),
              ARMRegister(falseReg, size), cond);
  } else if (insTy == MIRType::Float32 || insTy == MIRType::Double) {
    FloatRegister destReg = ToFloatRegister(ins->output());
    FloatRegister trueReg = ToFloatRegister(ins->ifTrueExpr());
    FloatRegister falseReg = ToFloatRegister(ins->ifFalseExpr());
    size_t size = MIRTypeToSize(insTy) * 8;
    masm.Fcsel(ARMFPRegister(destReg, size), ARMFPRegister(trueReg, size),
               ARMFPRegister(falseReg, size), cond);
  } else {
    // See above.
    MOZ_CRASH("CodeGenerator::visitWasmCompareAndSelect: unexpected type (2)");
  }
}

void CodeGenerator::visitWasmLoadI64(LWasmLoadI64* lir) {
  const MWasmLoad* mir = lir->mir();

  if (Maybe<uint64_t> absAddr = IsAbsoluteAddress(lir->ptr(), mir->access())) {
    masm.wasmLoadAbsolute(mir->access(), ToRegister(lir->memoryBase()),
                          absAddr.value(), AnyRegister(), ToOutRegister64(lir));
    return;
  }

  masm.wasmLoadI64(mir->access(), ToRegister(lir->memoryBase()),
                   ToRegister(lir->ptr()), ToOutRegister64(lir));
}

void CodeGenerator::visitWasmStoreI64(LWasmStoreI64* lir) {
  const MWasmStore* mir = lir->mir();

  if (Maybe<uint64_t> absAddr = IsAbsoluteAddress(lir->ptr(), mir->access())) {
    masm.wasmStoreAbsolute(mir->access(), AnyRegister(),
                           ToRegister64(lir->value()),
                           ToRegister(lir->memoryBase()), absAddr.value());
    return;
  }

  masm.wasmStoreI64(mir->access(), ToRegister64(lir->value()),
                    ToRegister(lir->memoryBase()), ToRegister(lir->ptr()));
}

void CodeGenerator::visitWasmAddOffset(LWasmAddOffset* lir) {
  MWasmAddOffset* mir = lir->mir();
  Register base = ToRegister(lir->base());
  Register out = ToRegister(lir->output());

  masm.Adds(ARMRegister(out, 32), ARMRegister(base, 32),
            Operand(mir->offset()));
  auto* ool = new (alloc()) LambdaOutOfLineCode([=, this](OutOfLineCode& ool) {
    masm.wasmTrap(wasm::Trap::OutOfBounds, mir->trapSiteDesc());
  });
  addOutOfLineCode(ool, mir);
  masm.j(Assembler::CarrySet, ool->entry());
}

void CodeGenerator::visitWasmAddOffset64(LWasmAddOffset64* lir) {
  MWasmAddOffset* mir = lir->mir();
  Register64 base = ToRegister64(lir->base());
  Register64 out = ToOutRegister64(lir);

  masm.Adds(ARMRegister(out.reg, 64), ARMRegister(base.reg, 64),
            Operand(mir->offset()));
  auto* ool = new (alloc()) LambdaOutOfLineCode([=, this](OutOfLineCode& ool) {
    masm.wasmTrap(wasm::Trap::OutOfBounds, mir->trapSiteDesc());
  });
  addOutOfLineCode(ool, mir);
  masm.j(Assembler::CarrySet, ool->entry());
}

void CodeGenerator::visitWasmSelectI64(LWasmSelectI64* lir) {
  MOZ_ASSERT(lir->mir()->type() == MIRType::Int64);
  Register condReg = ToRegister(lir->condExpr());
  Register64 trueReg = ToRegister64(lir->trueExpr());
  Register64 falseReg = ToRegister64(lir->falseExpr());
  Register64 outReg = ToOutRegister64(lir);

  masm.test32(condReg, condReg);
  masm.Csel(ARMRegister(outReg.reg, 64), ARMRegister(trueReg.reg, 64),
            ARMRegister(falseReg.reg, 64), Assembler::NonZero);
}

void CodeGenerator::visitSignExtendInt64(LSignExtendInt64* ins) {
  Register64 input = ToRegister64(ins->input());
  Register64 output = ToOutRegister64(ins);
  switch (ins->mir()->mode()) {
    case MSignExtendInt64::Byte:
      masm.move8To64SignExtend(input.reg, output);
      break;
    case MSignExtendInt64::Half:
      masm.move16To64SignExtend(input.reg, output);
      break;
    case MSignExtendInt64::Word:
      masm.move32To64SignExtend(input.reg, output);
      break;
  }
}

void CodeGenerator::visitWasmStackArgI64(LWasmStackArgI64* ins) {
  const MWasmStackArg* mir = ins->mir();
  Address dst(masm.getStackPointer(), mir->spOffset());
  if (IsConstant(ins->arg())) {
    masm.store64(Imm64(ToInt64(ins->arg())), dst);
  } else {
    masm.store64(ToRegister64(ins->arg()), dst);
  }
}

void CodeGenerator::visitWrapInt64ToInt32(LWrapInt64ToInt32* lir) {
  LInt64Allocation input = lir->input();
  Register output = ToRegister(lir->output());

  if (lir->mir()->bottomHalf()) {
    if (input.value().isMemory()) {
      masm.load32(ToAddress(input), output);
    } else {
      masm.move64To32(ToRegister64(input), output);
    }
  } else {
    MOZ_CRASH("Not implemented.");
  }
}

void CodeGenerator::visitExtendInt32ToInt64(LExtendInt32ToInt64* lir) {
  Register input = ToRegister(lir->input());
  Register64 output = ToOutRegister64(lir);

  if (lir->mir()->isUnsigned()) {
    masm.move32To64ZeroExtend(input, output);
  } else {
    masm.move32To64SignExtend(input, output);
  }
}

void CodeGenerator::visitWasmExtendU32Index(LWasmExtendU32Index* lir) {
  // Generates no code on this platform because the input is assumed to have
  // canonical form.
  Register output = ToRegister(lir->output());
  MOZ_ASSERT(ToRegister(lir->input()) == output);
  masm.debugAssertCanonicalInt32(output);
}

void CodeGenerator::visitWasmWrapU32Index(LWasmWrapU32Index* lir) {
  // Generates no code on this platform because the input is assumed to have
  // canonical form.
  Register output = ToRegister(lir->output());
  MOZ_ASSERT(ToRegister(lir->input()) == output);
  masm.debugAssertCanonicalInt32(output);
}

void CodeGenerator::visitWasmTruncateToInt32(LWasmTruncateToInt32* lir) {
  auto input = ToFloatRegister(lir->input());
  auto output = ToRegister(lir->output());

  MWasmTruncateToInt32* mir = lir->mir();
  MIRType fromType = mir->input()->type();

  MOZ_ASSERT(fromType == MIRType::Double || fromType == MIRType::Float32);

  auto* ool = new (alloc()) OutOfLineWasmTruncateCheck(mir, input, output);
  addOutOfLineCode(ool, mir);

  Label* oolEntry = ool->entry();
  if (mir->isUnsigned()) {
    if (fromType == MIRType::Double) {
      masm.wasmTruncateDoubleToUInt32(input, output, mir->isSaturating(),
                                      oolEntry);
    } else if (fromType == MIRType::Float32) {
      masm.wasmTruncateFloat32ToUInt32(input, output, mir->isSaturating(),
                                       oolEntry);
    } else {
      MOZ_CRASH("unexpected type");
    }

    masm.bind(ool->rejoin());
    return;
  }

  if (fromType == MIRType::Double) {
    masm.wasmTruncateDoubleToInt32(input, output, mir->isSaturating(),
                                   oolEntry);
  } else if (fromType == MIRType::Float32) {
    masm.wasmTruncateFloat32ToInt32(input, output, mir->isSaturating(),
                                    oolEntry);
  } else {
    MOZ_CRASH("unexpected type");
  }

  masm.bind(ool->rejoin());
}

void CodeGenerator::visitWasmTruncateToInt64(LWasmTruncateToInt64* lir) {
  FloatRegister input = ToFloatRegister(lir->input());
  Register64 output = ToOutRegister64(lir);

  MWasmTruncateToInt64* mir = lir->mir();
  MIRType fromType = mir->input()->type();

  MOZ_ASSERT(fromType == MIRType::Double || fromType == MIRType::Float32);

  auto* ool = new (alloc()) OutOfLineWasmTruncateCheck(mir, input, output);
  addOutOfLineCode(ool, mir);

  Label* oolEntry = ool->entry();
  Label* oolRejoin = ool->rejoin();
  bool isSaturating = mir->isSaturating();

  if (fromType == MIRType::Double) {
    if (mir->isUnsigned()) {
      masm.wasmTruncateDoubleToUInt64(input, output, isSaturating, oolEntry,
                                      oolRejoin, InvalidFloatReg);
    } else {
      masm.wasmTruncateDoubleToInt64(input, output, isSaturating, oolEntry,
                                     oolRejoin, InvalidFloatReg);
    }
  } else {
    if (mir->isUnsigned()) {
      masm.wasmTruncateFloat32ToUInt64(input, output, isSaturating, oolEntry,
                                       oolRejoin, InvalidFloatReg);
    } else {
      masm.wasmTruncateFloat32ToInt64(input, output, isSaturating, oolEntry,
                                      oolRejoin, InvalidFloatReg);
    }
  }
}

void CodeGeneratorARM64::visitOutOfLineWasmTruncateCheck(
    OutOfLineWasmTruncateCheck* ool) {
  FloatRegister input = ool->input();
  Register output = ool->output();
  Register64 output64 = ool->output64();
  MIRType fromType = ool->fromType();
  MIRType toType = ool->toType();
  Label* oolRejoin = ool->rejoin();
  TruncFlags flags = ool->flags();
  const wasm::TrapSiteDesc& trapSiteDesc = ool->trapSiteDesc();

  if (fromType == MIRType::Float32) {
    if (toType == MIRType::Int32) {
      masm.oolWasmTruncateCheckF32ToI32(input, output, flags, trapSiteDesc,
                                        oolRejoin);
    } else if (toType == MIRType::Int64) {
      masm.oolWasmTruncateCheckF32ToI64(input, output64, flags, trapSiteDesc,
                                        oolRejoin);
    } else {
      MOZ_CRASH("unexpected type");
    }
  } else if (fromType == MIRType::Double) {
    if (toType == MIRType::Int32) {
      masm.oolWasmTruncateCheckF64ToI32(input, output, flags, trapSiteDesc,
                                        oolRejoin);
    } else if (toType == MIRType::Int64) {
      masm.oolWasmTruncateCheckF64ToI64(input, output64, flags, trapSiteDesc,
                                        oolRejoin);
    } else {
      MOZ_CRASH("unexpected type");
    }
  } else {
    MOZ_CRASH("unexpected type");
  }
}

void CodeGenerator::visitAtomicTypedArrayElementBinop(
    LAtomicTypedArrayElementBinop* lir) {
  MOZ_ASSERT(!lir->mir()->isForEffect());

  AnyRegister output = ToAnyRegister(lir->output());
  Register elements = ToRegister(lir->elements());
  Register flagTemp = ToRegister(lir->temp0());
  Register outTemp = ToTempRegisterOrInvalid(lir->temp1());
  Register value = ToRegister(lir->value());

  Scalar::Type arrayType = lir->mir()->arrayType();

  auto mem = ToAddressOrBaseIndex(elements, lir->index(), arrayType);

  mem.match([&](const auto& mem) {
    masm.atomicFetchOpJS(arrayType, Synchronization::Full(),
                         lir->mir()->operation(), value, mem, flagTemp, outTemp,
                         output);
  });
}

void CodeGenerator::visitAtomicTypedArrayElementBinopForEffect(
    LAtomicTypedArrayElementBinopForEffect* lir) {
  MOZ_ASSERT(lir->mir()->isForEffect());

  Register elements = ToRegister(lir->elements());
  Register flagTemp = ToRegister(lir->temp0());
  Register value = ToRegister(lir->value());

  Scalar::Type arrayType = lir->mir()->arrayType();

  auto mem = ToAddressOrBaseIndex(elements, lir->index(), arrayType);

  mem.match([&](const auto& mem) {
    masm.atomicEffectOpJS(arrayType, Synchronization::Full(),
                          lir->mir()->operation(), value, mem, flagTemp);
  });
}

void CodeGenerator::visitInt64ToFloatingPoint(LInt64ToFloatingPoint* lir) {
  Register64 input = ToRegister64(lir->input());
  FloatRegister output = ToFloatRegister(lir->output());

  MIRType outputType = lir->mir()->type();
  MOZ_ASSERT(outputType == MIRType::Double || outputType == MIRType::Float32);

  if (outputType == MIRType::Double) {
    if (lir->mir()->isUnsigned()) {
      masm.convertUInt64ToDouble(input, output, Register::Invalid());
    } else {
      masm.convertInt64ToDouble(input, output);
    }
  } else {
    if (lir->mir()->isUnsigned()) {
      masm.convertUInt64ToFloat32(input, output, Register::Invalid());
    } else {
      masm.convertInt64ToFloat32(input, output);
    }
  }
}

void CodeGenerator::visitDivI64(LDivI64* lir) {
  Register lhs = ToRegister(lir->lhs());
  Register rhs = ToRegister(lir->rhs());

  ARMRegister lhs64 = toXRegister(lir->lhs());
  ARMRegister rhs64 = toXRegister(lir->rhs());
  ARMRegister output64 = toXRegister(lir->output());

  MDiv* mir = lir->mir();

  // Handle divide by zero.
  TrapIfDivideByZero(masm, lir, rhs64);

  // Handle an integer overflow exception from INT64_MIN / -1.
  if (mir->canBeNegativeOverflow()) {
    Label noOverflow;
    masm.branchPtr(Assembler::NotEqual, lhs, ImmWord(INT64_MIN), &noOverflow);
    masm.branchPtr(Assembler::NotEqual, rhs, ImmWord(-1), &noOverflow);
    masm.wasmTrap(wasm::Trap::IntegerOverflow, mir->trapSiteDesc());
    masm.bind(&noOverflow);
  }

  masm.Sdiv(output64, lhs64, rhs64);
}

void CodeGenerator::visitModI64(LModI64* lir) {
  ARMRegister lhs64 = toXRegister(lir->lhs());
  ARMRegister rhs64 = toXRegister(lir->rhs());
  ARMRegister output64 = toXRegister(lir->output());

  // Handle divide by zero.
  TrapIfDivideByZero(masm, lir, rhs64);

  masm.Sdiv(output64, lhs64, rhs64);

  // Compute the remainder: output = lhs - (output * rhs).
  masm.Msub(output64, output64, rhs64, lhs64);
}

void CodeGenerator::visitUDivI64(LUDivI64* lir) {
  ARMRegister lhs64 = toXRegister(lir->lhs());
  ARMRegister rhs64 = toXRegister(lir->rhs());
  ARMRegister output64 = toXRegister(lir->output());

  // Handle divide by zero.
  TrapIfDivideByZero(masm, lir, rhs64);

  masm.Udiv(output64, lhs64, rhs64);
}

void CodeGenerator::visitUModI64(LUModI64* lir) {
  ARMRegister lhs64 = toXRegister(lir->lhs());
  ARMRegister rhs64 = toXRegister(lir->rhs());
  ARMRegister output64 = toXRegister(lir->output());

  // Handle divide by zero.
  TrapIfDivideByZero(masm, lir, rhs64);

  masm.Udiv(output64, lhs64, rhs64);

  // Compute the remainder: output = lhs - (output * rhs).
  masm.Msub(output64, output64, rhs64, lhs64);
}

void CodeGenerator::visitSimd128(LSimd128* ins) {
#ifdef ENABLE_WASM_SIMD
  const LDefinition* out = ins->output();
  masm.loadConstantSimd128(ins->simd128(), ToFloatRegister(out));
#else
  MOZ_CRASH("No SIMD");
#endif
}

void CodeGenerator::visitWasmTernarySimd128(LWasmTernarySimd128* ins) {
#ifdef ENABLE_WASM_SIMD
  switch (ins->simdOp()) {
    case wasm::SimdOp::V128Bitselect: {
      FloatRegister lhs = ToFloatRegister(ins->v0());
      FloatRegister rhs = ToFloatRegister(ins->v1());
      FloatRegister controlDest = ToFloatRegister(ins->v2());
      masm.bitwiseSelectSimd128(lhs, rhs, controlDest);
      break;
    }
    case wasm::SimdOp::F32x4RelaxedMadd:
      masm.fmaFloat32x4(ToFloatRegister(ins->v0()), ToFloatRegister(ins->v1()),
                        ToFloatRegister(ins->v2()));
      break;
    case wasm::SimdOp::F32x4RelaxedNmadd:
      masm.fnmaFloat32x4(ToFloatRegister(ins->v0()), ToFloatRegister(ins->v1()),
                         ToFloatRegister(ins->v2()));
      break;
    case wasm::SimdOp::F64x2RelaxedMadd:
      masm.fmaFloat64x2(ToFloatRegister(ins->v0()), ToFloatRegister(ins->v1()),
                        ToFloatRegister(ins->v2()));
      break;
    case wasm::SimdOp::F64x2RelaxedNmadd:
      masm.fnmaFloat64x2(ToFloatRegister(ins->v0()), ToFloatRegister(ins->v1()),
                         ToFloatRegister(ins->v2()));
      break;
    case wasm::SimdOp::I8x16RelaxedLaneSelect:
    case wasm::SimdOp::I16x8RelaxedLaneSelect:
    case wasm::SimdOp::I32x4RelaxedLaneSelect:
    case wasm::SimdOp::I64x2RelaxedLaneSelect: {
      FloatRegister lhs = ToFloatRegister(ins->v0());
      FloatRegister rhs = ToFloatRegister(ins->v1());
      FloatRegister maskDest = ToFloatRegister(ins->v2());
      masm.laneSelectSimd128(maskDest, lhs, rhs, maskDest);
      break;
    }
    case wasm::SimdOp::I32x4RelaxedDotI8x16I7x16AddS:
      masm.dotInt8x16Int7x16ThenAdd(
          ToFloatRegister(ins->v0()), ToFloatRegister(ins->v1()),
          ToFloatRegister(ins->v2()), ToFloatRegister(ins->temp0()));
      break;
    default:
      MOZ_CRASH("NYI");
  }
#else
  MOZ_CRASH("No SIMD");
#endif
}

void CodeGenerator::visitWasmBinarySimd128(LWasmBinarySimd128* ins) {
#ifdef ENABLE_WASM_SIMD
  FloatRegister lhs = ToFloatRegister(ins->lhs());
  FloatRegister rhs = ToFloatRegister(ins->rhs());
  FloatRegister dest = ToFloatRegister(ins->output());

  switch (ins->simdOp()) {
    case wasm::SimdOp::V128And:
      masm.bitwiseAndSimd128(lhs, rhs, dest);
      break;
    case wasm::SimdOp::V128Or:
      masm.bitwiseOrSimd128(lhs, rhs, dest);
      break;
    case wasm::SimdOp::V128Xor:
      masm.bitwiseXorSimd128(lhs, rhs, dest);
      break;
    case wasm::SimdOp::V128AndNot:
      masm.bitwiseAndNotSimd128(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16AvgrU:
      masm.unsignedAverageInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8AvgrU:
      masm.unsignedAverageInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16Add:
      masm.addInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16AddSatS:
      masm.addSatInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16AddSatU:
      masm.unsignedAddSatInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16Sub:
      masm.subInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16SubSatS:
      masm.subSatInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16SubSatU:
      masm.unsignedSubSatInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16MinS:
      masm.minInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16MinU:
      masm.unsignedMinInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16MaxS:
      masm.maxInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16MaxU:
      masm.unsignedMaxInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8Add:
      masm.addInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8AddSatS:
      masm.addSatInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8AddSatU:
      masm.unsignedAddSatInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8Sub:
      masm.subInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8SubSatS:
      masm.subSatInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8SubSatU:
      masm.unsignedSubSatInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8Mul:
      masm.mulInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8MinS:
      masm.minInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8MinU:
      masm.unsignedMinInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8MaxS:
      masm.maxInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8MaxU:
      masm.unsignedMaxInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4Add:
      masm.addInt32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4Sub:
      masm.subInt32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4Mul:
      masm.mulInt32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4MinS:
      masm.minInt32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4MinU:
      masm.unsignedMinInt32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4MaxS:
      masm.maxInt32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4MaxU:
      masm.unsignedMaxInt32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I64x2Add:
      masm.addInt64x2(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I64x2Sub:
      masm.subInt64x2(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I64x2Mul: {
      auto temp1 = ToFloatRegister(ins->temp0());
      auto temp2 = ToFloatRegister(ins->temp1());
      masm.mulInt64x2(lhs, rhs, dest, temp1, temp2);
      break;
    }
    case wasm::SimdOp::F32x4Add:
      masm.addFloat32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::F32x4Sub:
      masm.subFloat32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::F32x4Mul:
      masm.mulFloat32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::F32x4Div:
      masm.divFloat32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::F32x4Min:
      masm.minFloat32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::F32x4Max:
      masm.maxFloat32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::F64x2Add:
      masm.addFloat64x2(lhs, rhs, dest);
      break;
    case wasm::SimdOp::F64x2Sub:
      masm.subFloat64x2(lhs, rhs, dest);
      break;
    case wasm::SimdOp::F64x2Mul:
      masm.mulFloat64x2(lhs, rhs, dest);
      break;
    case wasm::SimdOp::F64x2Div:
      masm.divFloat64x2(lhs, rhs, dest);
      break;
    case wasm::SimdOp::F64x2Min:
      masm.minFloat64x2(lhs, rhs, dest);
      break;
    case wasm::SimdOp::F64x2Max:
      masm.maxFloat64x2(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16Swizzle:
      masm.swizzleInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16RelaxedSwizzle:
      masm.swizzleInt8x16Relaxed(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16NarrowI16x8S:
      masm.narrowInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16NarrowI16x8U:
      masm.unsignedNarrowInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8NarrowI32x4S:
      masm.narrowInt32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8NarrowI32x4U:
      masm.unsignedNarrowInt32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16Eq:
      masm.compareInt8x16(Assembler::Equal, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16Ne:
      masm.compareInt8x16(Assembler::NotEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16LtS:
      masm.compareInt8x16(Assembler::LessThan, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16GtS:
      masm.compareInt8x16(Assembler::GreaterThan, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16LeS:
      masm.compareInt8x16(Assembler::LessThanOrEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16GeS:
      masm.compareInt8x16(Assembler::GreaterThanOrEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16LtU:
      masm.compareInt8x16(Assembler::Below, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16GtU:
      masm.compareInt8x16(Assembler::Above, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16LeU:
      masm.compareInt8x16(Assembler::BelowOrEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16GeU:
      masm.compareInt8x16(Assembler::AboveOrEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8Eq:
      masm.compareInt16x8(Assembler::Equal, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8Ne:
      masm.compareInt16x8(Assembler::NotEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8LtS:
      masm.compareInt16x8(Assembler::LessThan, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8GtS:
      masm.compareInt16x8(Assembler::GreaterThan, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8LeS:
      masm.compareInt16x8(Assembler::LessThanOrEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8GeS:
      masm.compareInt16x8(Assembler::GreaterThanOrEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8LtU:
      masm.compareInt16x8(Assembler::Below, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8GtU:
      masm.compareInt16x8(Assembler::Above, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8LeU:
      masm.compareInt16x8(Assembler::BelowOrEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8GeU:
      masm.compareInt16x8(Assembler::AboveOrEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4Eq:
      masm.compareInt32x4(Assembler::Equal, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4Ne:
      masm.compareInt32x4(Assembler::NotEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4LtS:
      masm.compareInt32x4(Assembler::LessThan, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4GtS:
      masm.compareInt32x4(Assembler::GreaterThan, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4LeS:
      masm.compareInt32x4(Assembler::LessThanOrEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4GeS:
      masm.compareInt32x4(Assembler::GreaterThanOrEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4LtU:
      masm.compareInt32x4(Assembler::Below, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4GtU:
      masm.compareInt32x4(Assembler::Above, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4LeU:
      masm.compareInt32x4(Assembler::BelowOrEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4GeU:
      masm.compareInt32x4(Assembler::AboveOrEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I64x2Eq:
      masm.compareInt64x2(Assembler::Equal, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I64x2LtS:
      masm.compareInt64x2(Assembler::LessThan, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I64x2GtS:
      masm.compareInt64x2(Assembler::GreaterThan, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I64x2LeS:
      masm.compareInt64x2(Assembler::LessThanOrEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I64x2GeS:
      masm.compareInt64x2(Assembler::GreaterThanOrEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::I64x2Ne:
      masm.compareInt64x2(Assembler::NotEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::F32x4Eq:
      masm.compareFloat32x4(Assembler::Equal, lhs, rhs, dest);
      break;
    case wasm::SimdOp::F32x4Ne:
      masm.compareFloat32x4(Assembler::NotEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::F32x4Lt:
      masm.compareFloat32x4(Assembler::LessThan, lhs, rhs, dest);
      break;
    case wasm::SimdOp::F32x4Gt:
      masm.compareFloat32x4(Assembler::GreaterThan, lhs, rhs, dest);
      break;
    case wasm::SimdOp::F32x4Le:
      masm.compareFloat32x4(Assembler::LessThanOrEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::F32x4Ge:
      masm.compareFloat32x4(Assembler::GreaterThanOrEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::F64x2Eq:
      masm.compareFloat64x2(Assembler::Equal, lhs, rhs, dest);
      break;
    case wasm::SimdOp::F64x2Ne:
      masm.compareFloat64x2(Assembler::NotEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::F64x2Lt:
      masm.compareFloat64x2(Assembler::LessThan, lhs, rhs, dest);
      break;
    case wasm::SimdOp::F64x2Gt:
      masm.compareFloat64x2(Assembler::GreaterThan, lhs, rhs, dest);
      break;
    case wasm::SimdOp::F64x2Le:
      masm.compareFloat64x2(Assembler::LessThanOrEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::F64x2Ge:
      masm.compareFloat64x2(Assembler::GreaterThanOrEqual, lhs, rhs, dest);
      break;
    case wasm::SimdOp::F32x4PMax:
      masm.pseudoMaxFloat32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::F32x4PMin:
      masm.pseudoMinFloat32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::F64x2PMax:
      masm.pseudoMaxFloat64x2(lhs, rhs, dest);
      break;
    case wasm::SimdOp::F64x2PMin:
      masm.pseudoMinFloat64x2(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4DotI16x8S:
      masm.widenDotInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8ExtmulLowI8x16S:
      masm.extMulLowInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8ExtmulHighI8x16S:
      masm.extMulHighInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8ExtmulLowI8x16U:
      masm.unsignedExtMulLowInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8ExtmulHighI8x16U:
      masm.unsignedExtMulHighInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4ExtmulLowI16x8S:
      masm.extMulLowInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4ExtmulHighI16x8S:
      masm.extMulHighInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4ExtmulLowI16x8U:
      masm.unsignedExtMulLowInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4ExtmulHighI16x8U:
      masm.unsignedExtMulHighInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I64x2ExtmulLowI32x4S:
      masm.extMulLowInt32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I64x2ExtmulHighI32x4S:
      masm.extMulHighInt32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I64x2ExtmulLowI32x4U:
      masm.unsignedExtMulLowInt32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I64x2ExtmulHighI32x4U:
      masm.unsignedExtMulHighInt32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8Q15MulrSatS:
      masm.q15MulrSatInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::F32x4RelaxedMin:
      masm.minFloat32x4Relaxed(lhs, rhs, dest);
      break;
    case wasm::SimdOp::F32x4RelaxedMax:
      masm.maxFloat32x4Relaxed(lhs, rhs, dest);
      break;
    case wasm::SimdOp::F64x2RelaxedMin:
      masm.minFloat64x2Relaxed(lhs, rhs, dest);
      break;
    case wasm::SimdOp::F64x2RelaxedMax:
      masm.maxFloat64x2Relaxed(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8RelaxedQ15MulrS:
      masm.q15MulrInt16x8Relaxed(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8RelaxedDotI8x16I7x16S:
      masm.dotInt8x16Int7x16(lhs, rhs, dest);
      break;
    default:
      MOZ_CRASH("Binary SimdOp not implemented");
  }
#else
  MOZ_CRASH("No SIMD");
#endif
}

void CodeGenerator::visitWasmBinarySimd128WithConstant(
    LWasmBinarySimd128WithConstant* ins) {
  MOZ_CRASH("No SIMD");
}

void CodeGenerator::visitWasmVariableShiftSimd128(
    LWasmVariableShiftSimd128* ins) {
#ifdef ENABLE_WASM_SIMD
  FloatRegister lhs = ToFloatRegister(ins->lhs());
  Register rhs = ToRegister(ins->rhs());
  FloatRegister dest = ToFloatRegister(ins->output());

  switch (ins->mir()->simdOp()) {
    case wasm::SimdOp::I8x16Shl:
      masm.leftShiftInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16ShrS:
      masm.rightShiftInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I8x16ShrU:
      masm.unsignedRightShiftInt8x16(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8Shl:
      masm.leftShiftInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8ShrS:
      masm.rightShiftInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I16x8ShrU:
      masm.unsignedRightShiftInt16x8(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4Shl:
      masm.leftShiftInt32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4ShrS:
      masm.rightShiftInt32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I32x4ShrU:
      masm.unsignedRightShiftInt32x4(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I64x2Shl:
      masm.leftShiftInt64x2(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I64x2ShrS:
      masm.rightShiftInt64x2(lhs, rhs, dest);
      break;
    case wasm::SimdOp::I64x2ShrU:
      masm.unsignedRightShiftInt64x2(lhs, rhs, dest);
      break;
    default:
      MOZ_CRASH("Shift SimdOp not implemented");
  }
#else
  MOZ_CRASH("No SIMD");
#endif
}

void CodeGenerator::visitWasmConstantShiftSimd128(
    LWasmConstantShiftSimd128* ins) {
#ifdef ENABLE_WASM_SIMD
  FloatRegister src = ToFloatRegister(ins->src());
  FloatRegister dest = ToFloatRegister(ins->output());
  int32_t shift = ins->shift();

  if (shift == 0) {
    if (src != dest) {
      masm.moveSimd128(src, dest);
    }
    return;
  }

  switch (ins->mir()->simdOp()) {
    case wasm::SimdOp::I8x16Shl:
      masm.leftShiftInt8x16(Imm32(shift), src, dest);
      break;
    case wasm::SimdOp::I8x16ShrS:
      masm.rightShiftInt8x16(Imm32(shift), src, dest);
      break;
    case wasm::SimdOp::I8x16ShrU:
      masm.unsignedRightShiftInt8x16(Imm32(shift), src, dest);
      break;
    case wasm::SimdOp::I16x8Shl:
      masm.leftShiftInt16x8(Imm32(shift), src, dest);
      break;
    case wasm::SimdOp::I16x8ShrS:
      masm.rightShiftInt16x8(Imm32(shift), src, dest);
      break;
    case wasm::SimdOp::I16x8ShrU:
      masm.unsignedRightShiftInt16x8(Imm32(shift), src, dest);
      break;
    case wasm::SimdOp::I32x4Shl:
      masm.leftShiftInt32x4(Imm32(shift), src, dest);
      break;
    case wasm::SimdOp::I32x4ShrS:
      masm.rightShiftInt32x4(Imm32(shift), src, dest);
      break;
    case wasm::SimdOp::I32x4ShrU:
      masm.unsignedRightShiftInt32x4(Imm32(shift), src, dest);
      break;
    case wasm::SimdOp::I64x2Shl:
      masm.leftShiftInt64x2(Imm32(shift), src, dest);
      break;
    case wasm::SimdOp::I64x2ShrS:
      masm.rightShiftInt64x2(Imm32(shift), src, dest);
      break;
    case wasm::SimdOp::I64x2ShrU:
      masm.unsignedRightShiftInt64x2(Imm32(shift), src, dest);
      break;
    default:
      MOZ_CRASH("Shift SimdOp not implemented");
  }
#else
  MOZ_CRASH("No SIMD");
#endif
}

void CodeGenerator::visitWasmSignReplicationSimd128(
    LWasmSignReplicationSimd128* ins) {
  MOZ_CRASH("No SIMD");
}

void CodeGenerator::visitWasmShuffleSimd128(LWasmShuffleSimd128* ins) {
#ifdef ENABLE_WASM_SIMD
  FloatRegister lhs = ToFloatRegister(ins->lhs());
  FloatRegister rhs = ToFloatRegister(ins->rhs());
  FloatRegister dest = ToFloatRegister(ins->output());
  SimdConstant control = ins->control();
  switch (ins->op()) {
    case SimdShuffleOp::BLEND_8x16: {
      masm.blendInt8x16(reinterpret_cast<const uint8_t*>(control.asInt8x16()),
                        lhs, rhs, dest);
      break;
    }
    case SimdShuffleOp::BLEND_16x8: {
      masm.blendInt16x8(reinterpret_cast<const uint16_t*>(control.asInt16x8()),
                        lhs, rhs, dest);
      break;
    }
    case SimdShuffleOp::CONCAT_RIGHT_SHIFT_8x16: {
      int8_t count = 16 - control.asInt8x16()[0];
      MOZ_ASSERT(count > 0, "Should have been a MOVE operation");
      masm.concatAndRightShiftSimd128(lhs, rhs, dest, count);
      break;
    }
    case SimdShuffleOp::INTERLEAVE_HIGH_8x16: {
      masm.interleaveHighInt8x16(lhs, rhs, dest);
      break;
    }
    case SimdShuffleOp::INTERLEAVE_HIGH_16x8: {
      masm.interleaveHighInt16x8(lhs, rhs, dest);
      break;
    }
    case SimdShuffleOp::INTERLEAVE_HIGH_32x4: {
      masm.interleaveHighInt32x4(lhs, rhs, dest);
      break;
    }
    case SimdShuffleOp::INTERLEAVE_HIGH_64x2: {
      masm.interleaveHighInt64x2(lhs, rhs, dest);
      break;
    }
    case SimdShuffleOp::INTERLEAVE_LOW_8x16: {
      masm.interleaveLowInt8x16(lhs, rhs, dest);
      break;
    }
    case SimdShuffleOp::INTERLEAVE_LOW_16x8: {
      masm.interleaveLowInt16x8(lhs, rhs, dest);
      break;
    }
    case SimdShuffleOp::INTERLEAVE_LOW_32x4: {
      masm.interleaveLowInt32x4(lhs, rhs, dest);
      break;
    }
    case SimdShuffleOp::INTERLEAVE_LOW_64x2: {
      masm.interleaveLowInt64x2(lhs, rhs, dest);
      break;
    }
    case SimdShuffleOp::SHUFFLE_BLEND_8x16: {
      masm.shuffleInt8x16(reinterpret_cast<const uint8_t*>(control.asInt8x16()),
                          lhs, rhs, dest);
      break;
    }
    default: {
      MOZ_CRASH("Unsupported SIMD shuffle operation");
    }
  }
#else
  MOZ_CRASH("No SIMD");
#endif
}

void CodeGenerator::visitWasmPermuteSimd128(LWasmPermuteSimd128* ins) {
#ifdef ENABLE_WASM_SIMD
  FloatRegister src = ToFloatRegister(ins->src());
  FloatRegister dest = ToFloatRegister(ins->output());
  SimdConstant control = ins->control();
  switch (ins->op()) {
    case SimdPermuteOp::BROADCAST_8x16: {
      const SimdConstant::I8x16& mask = control.asInt8x16();
      int8_t source = mask[0];
      masm.splatX16(source, src, dest);
      break;
    }
    case SimdPermuteOp::BROADCAST_16x8: {
      const SimdConstant::I16x8& mask = control.asInt16x8();
      int16_t source = mask[0];
      masm.splatX8(source, src, dest);
      break;
    }
    case SimdPermuteOp::MOVE: {
      masm.moveSimd128(src, dest);
      break;
    }
    case SimdPermuteOp::PERMUTE_8x16: {
      const SimdConstant::I8x16& mask = control.asInt8x16();
#  ifdef DEBUG
      mozilla::DebugOnly<int> i;
      for (i = 0; i < 16 && mask[i] == i; i++) {
      }
      MOZ_ASSERT(i < 16, "Should have been a MOVE operation");
#  endif
      masm.permuteInt8x16(reinterpret_cast<const uint8_t*>(mask), src, dest);
      break;
    }
    case SimdPermuteOp::PERMUTE_16x8: {
      const SimdConstant::I16x8& mask = control.asInt16x8();
#  ifdef DEBUG
      mozilla::DebugOnly<int> i;
      for (i = 0; i < 8 && mask[i] == i; i++) {
      }
      MOZ_ASSERT(i < 8, "Should have been a MOVE operation");
#  endif
      masm.permuteInt16x8(reinterpret_cast<const uint16_t*>(mask), src, dest);
      break;
    }
    case SimdPermuteOp::PERMUTE_32x4: {
      const SimdConstant::I32x4& mask = control.asInt32x4();
#  ifdef DEBUG
      mozilla::DebugOnly<int> i;
      for (i = 0; i < 4 && mask[i] == i; i++) {
      }
      MOZ_ASSERT(i < 4, "Should have been a MOVE operation");
#  endif
      masm.permuteInt32x4(reinterpret_cast<const uint32_t*>(mask), src, dest);
      break;
    }
    case SimdPermuteOp::ROTATE_RIGHT_8x16: {
      int8_t count = control.asInt8x16()[0];
      MOZ_ASSERT(count > 0, "Should have been a MOVE operation");
      masm.rotateRightSimd128(src, dest, count);
      break;
    }
    case SimdPermuteOp::SHIFT_LEFT_8x16: {
      int8_t count = control.asInt8x16()[0];
      MOZ_ASSERT(count > 0, "Should have been a MOVE operation");
      masm.leftShiftSimd128(Imm32(count), src, dest);
      break;
    }
    case SimdPermuteOp::SHIFT_RIGHT_8x16: {
      int8_t count = control.asInt8x16()[0];
      MOZ_ASSERT(count > 0, "Should have been a MOVE operation");
      masm.rightShiftSimd128(Imm32(count), src, dest);
      break;
    }
    case SimdPermuteOp::ZERO_EXTEND_8x16_TO_16x8:
      masm.zeroExtend8x16To16x8(src, dest);
      break;
    case SimdPermuteOp::ZERO_EXTEND_8x16_TO_32x4:
      masm.zeroExtend8x16To32x4(src, dest);
      break;
    case SimdPermuteOp::ZERO_EXTEND_8x16_TO_64x2:
      masm.zeroExtend8x16To64x2(src, dest);
      break;
    case SimdPermuteOp::ZERO_EXTEND_16x8_TO_32x4:
      masm.zeroExtend16x8To32x4(src, dest);
      break;
    case SimdPermuteOp::ZERO_EXTEND_16x8_TO_64x2:
      masm.zeroExtend16x8To64x2(src, dest);
      break;
    case SimdPermuteOp::ZERO_EXTEND_32x4_TO_64x2:
      masm.zeroExtend32x4To64x2(src, dest);
      break;
    case SimdPermuteOp::REVERSE_16x8:
      masm.reverseInt16x8(src, dest);
      break;
    case SimdPermuteOp::REVERSE_32x4:
      masm.reverseInt32x4(src, dest);
      break;
    case SimdPermuteOp::REVERSE_64x2:
      masm.reverseInt64x2(src, dest);
      break;
    default: {
      MOZ_CRASH("Unsupported SIMD permutation operation");
    }
  }
#else
  MOZ_CRASH("No SIMD");
#endif
}

void CodeGenerator::visitWasmReplaceLaneSimd128(LWasmReplaceLaneSimd128* ins) {
#ifdef ENABLE_WASM_SIMD
  MOZ_ASSERT(ToFloatRegister(ins->lhs()) == ToFloatRegister(ins->output()));
  FloatRegister lhsDest = ToFloatRegister(ins->lhs());
  const LAllocation* rhs = ins->rhs();
  uint32_t laneIndex = ins->mir()->laneIndex();

  switch (ins->mir()->simdOp()) {
    case wasm::SimdOp::I8x16ReplaceLane:
      masm.replaceLaneInt8x16(laneIndex, ToRegister(rhs), lhsDest);
      break;
    case wasm::SimdOp::I16x8ReplaceLane:
      masm.replaceLaneInt16x8(laneIndex, ToRegister(rhs), lhsDest);
      break;
    case wasm::SimdOp::I32x4ReplaceLane:
      masm.replaceLaneInt32x4(laneIndex, ToRegister(rhs), lhsDest);
      break;
    case wasm::SimdOp::F32x4ReplaceLane:
      masm.replaceLaneFloat32x4(laneIndex, ToFloatRegister(rhs), lhsDest);
      break;
    case wasm::SimdOp::F64x2ReplaceLane:
      masm.replaceLaneFloat64x2(laneIndex, ToFloatRegister(rhs), lhsDest);
      break;
    default:
      MOZ_CRASH("ReplaceLane SimdOp not implemented");
  }
#else
  MOZ_CRASH("No SIMD");
#endif
}

void CodeGenerator::visitWasmReplaceInt64LaneSimd128(
    LWasmReplaceInt64LaneSimd128* ins) {
#ifdef ENABLE_WASM_SIMD
  MOZ_RELEASE_ASSERT(ins->mir()->simdOp() == wasm::SimdOp::I64x2ReplaceLane);
  MOZ_ASSERT(ToFloatRegister(ins->lhs()) == ToFloatRegister(ins->output()));
  masm.replaceLaneInt64x2(ins->mir()->laneIndex(), ToRegister64(ins->rhs()),
                          ToFloatRegister(ins->lhs()));
#else
  MOZ_CRASH("No SIMD");
#endif
}

void CodeGenerator::visitWasmScalarToSimd128(LWasmScalarToSimd128* ins) {
#ifdef ENABLE_WASM_SIMD
  FloatRegister dest = ToFloatRegister(ins->output());

  switch (ins->mir()->simdOp()) {
    case wasm::SimdOp::I8x16Splat:
      masm.splatX16(ToRegister(ins->src()), dest);
      break;
    case wasm::SimdOp::I16x8Splat:
      masm.splatX8(ToRegister(ins->src()), dest);
      break;
    case wasm::SimdOp::I32x4Splat:
      masm.splatX4(ToRegister(ins->src()), dest);
      break;
    case wasm::SimdOp::F32x4Splat:
      masm.splatX4(ToFloatRegister(ins->src()), dest);
      break;
    case wasm::SimdOp::F64x2Splat:
      masm.splatX2(ToFloatRegister(ins->src()), dest);
      break;
    default:
      MOZ_CRASH("ScalarToSimd128 SimdOp not implemented");
  }
#else
  MOZ_CRASH("No SIMD");
#endif
}

void CodeGenerator::visitWasmInt64ToSimd128(LWasmInt64ToSimd128* ins) {
#ifdef ENABLE_WASM_SIMD
  Register64 src = ToRegister64(ins->src());
  FloatRegister dest = ToFloatRegister(ins->output());

  switch (ins->mir()->simdOp()) {
    case wasm::SimdOp::I64x2Splat:
      masm.splatX2(src, dest);
      break;
    case wasm::SimdOp::V128Load8x8S:
      masm.moveGPR64ToDouble(src, dest);
      masm.widenLowInt8x16(dest, dest);
      break;
    case wasm::SimdOp::V128Load8x8U:
      masm.moveGPR64ToDouble(src, dest);
      masm.unsignedWidenLowInt8x16(dest, dest);
      break;
    case wasm::SimdOp::V128Load16x4S:
      masm.moveGPR64ToDouble(src, dest);
      masm.widenLowInt16x8(dest, dest);
      break;
    case wasm::SimdOp::V128Load16x4U:
      masm.moveGPR64ToDouble(src, dest);
      masm.unsignedWidenLowInt16x8(dest, dest);
      break;
    case wasm::SimdOp::V128Load32x2S:
      masm.moveGPR64ToDouble(src, dest);
      masm.widenLowInt32x4(dest, dest);
      break;
    case wasm::SimdOp::V128Load32x2U:
      masm.moveGPR64ToDouble(src, dest);
      masm.unsignedWidenLowInt32x4(dest, dest);
      break;
    default:
      MOZ_CRASH("Int64ToSimd128 SimdOp not implemented");
  }
#else
  MOZ_CRASH("No SIMD");
#endif
}

void CodeGenerator::visitWasmUnarySimd128(LWasmUnarySimd128* ins) {
#ifdef ENABLE_WASM_SIMD
  FloatRegister src = ToFloatRegister(ins->src());
  FloatRegister dest = ToFloatRegister(ins->output());

  switch (ins->mir()->simdOp()) {
    case wasm::SimdOp::I8x16Neg:
      masm.negInt8x16(src, dest);
      break;
    case wasm::SimdOp::I16x8Neg:
      masm.negInt16x8(src, dest);
      break;
    case wasm::SimdOp::I16x8ExtendLowI8x16S:
      masm.widenLowInt8x16(src, dest);
      break;
    case wasm::SimdOp::I16x8ExtendHighI8x16S:
      masm.widenHighInt8x16(src, dest);
      break;
    case wasm::SimdOp::I16x8ExtendLowI8x16U:
      masm.unsignedWidenLowInt8x16(src, dest);
      break;
    case wasm::SimdOp::I16x8ExtendHighI8x16U:
      masm.unsignedWidenHighInt8x16(src, dest);
      break;
    case wasm::SimdOp::I32x4Neg:
      masm.negInt32x4(src, dest);
      break;
    case wasm::SimdOp::I32x4ExtendLowI16x8S:
      masm.widenLowInt16x8(src, dest);
      break;
    case wasm::SimdOp::I32x4ExtendHighI16x8S:
      masm.widenHighInt16x8(src, dest);
      break;
    case wasm::SimdOp::I32x4ExtendLowI16x8U:
      masm.unsignedWidenLowInt16x8(src, dest);
      break;
    case wasm::SimdOp::I32x4ExtendHighI16x8U:
      masm.unsignedWidenHighInt16x8(src, dest);
      break;
    case wasm::SimdOp::I32x4TruncSatF32x4S:
      masm.truncSatFloat32x4ToInt32x4(src, dest);
      break;
    case wasm::SimdOp::I32x4TruncSatF32x4U:
      masm.unsignedTruncSatFloat32x4ToInt32x4(src, dest);
      break;
    case wasm::SimdOp::I64x2Neg:
      masm.negInt64x2(src, dest);
      break;
    case wasm::SimdOp::I64x2ExtendLowI32x4S:
      masm.widenLowInt32x4(src, dest);
      break;
    case wasm::SimdOp::I64x2ExtendHighI32x4S:
      masm.widenHighInt32x4(src, dest);
      break;
    case wasm::SimdOp::I64x2ExtendLowI32x4U:
      masm.unsignedWidenLowInt32x4(src, dest);
      break;
    case wasm::SimdOp::I64x2ExtendHighI32x4U:
      masm.unsignedWidenHighInt32x4(src, dest);
      break;
    case wasm::SimdOp::F32x4Abs:
      masm.absFloat32x4(src, dest);
      break;
    case wasm::SimdOp::F32x4Neg:
      masm.negFloat32x4(src, dest);
      break;
    case wasm::SimdOp::F32x4Sqrt:
      masm.sqrtFloat32x4(src, dest);
      break;
    case wasm::SimdOp::F32x4ConvertI32x4S:
      masm.convertInt32x4ToFloat32x4(src, dest);
      break;
    case wasm::SimdOp::F32x4ConvertI32x4U:
      masm.unsignedConvertInt32x4ToFloat32x4(src, dest);
      break;
    case wasm::SimdOp::F64x2Abs:
      masm.absFloat64x2(src, dest);
      break;
    case wasm::SimdOp::F64x2Neg:
      masm.negFloat64x2(src, dest);
      break;
    case wasm::SimdOp::F64x2Sqrt:
      masm.sqrtFloat64x2(src, dest);
      break;
    case wasm::SimdOp::V128Not:
      masm.bitwiseNotSimd128(src, dest);
      break;
    case wasm::SimdOp::I8x16Abs:
      masm.absInt8x16(src, dest);
      break;
    case wasm::SimdOp::I16x8Abs:
      masm.absInt16x8(src, dest);
      break;
    case wasm::SimdOp::I32x4Abs:
      masm.absInt32x4(src, dest);
      break;
    case wasm::SimdOp::I64x2Abs:
      masm.absInt64x2(src, dest);
      break;
    case wasm::SimdOp::F32x4Ceil:
      masm.ceilFloat32x4(src, dest);
      break;
    case wasm::SimdOp::F32x4Floor:
      masm.floorFloat32x4(src, dest);
      break;
    case wasm::SimdOp::F32x4Trunc:
      masm.truncFloat32x4(src, dest);
      break;
    case wasm::SimdOp::F32x4Nearest:
      masm.nearestFloat32x4(src, dest);
      break;
    case wasm::SimdOp::F64x2Ceil:
      masm.ceilFloat64x2(src, dest);
      break;
    case wasm::SimdOp::F64x2Floor:
      masm.floorFloat64x2(src, dest);
      break;
    case wasm::SimdOp::F64x2Trunc:
      masm.truncFloat64x2(src, dest);
      break;
    case wasm::SimdOp::F64x2Nearest:
      masm.nearestFloat64x2(src, dest);
      break;
    case wasm::SimdOp::F32x4DemoteF64x2Zero:
      masm.convertFloat64x2ToFloat32x4(src, dest);
      break;
    case wasm::SimdOp::F64x2PromoteLowF32x4:
      masm.convertFloat32x4ToFloat64x2(src, dest);
      break;
    case wasm::SimdOp::F64x2ConvertLowI32x4S:
      masm.convertInt32x4ToFloat64x2(src, dest);
      break;
    case wasm::SimdOp::F64x2ConvertLowI32x4U:
      masm.unsignedConvertInt32x4ToFloat64x2(src, dest);
      break;
    case wasm::SimdOp::I32x4TruncSatF64x2SZero:
      masm.truncSatFloat64x2ToInt32x4(src, dest, ToFloatRegister(ins->temp0()));
      break;
    case wasm::SimdOp::I32x4TruncSatF64x2UZero:
      masm.unsignedTruncSatFloat64x2ToInt32x4(src, dest,
                                              ToFloatRegister(ins->temp0()));
      break;
    case wasm::SimdOp::I16x8ExtaddPairwiseI8x16S:
      masm.extAddPairwiseInt8x16(src, dest);
      break;
    case wasm::SimdOp::I16x8ExtaddPairwiseI8x16U:
      masm.unsignedExtAddPairwiseInt8x16(src, dest);
      break;
    case wasm::SimdOp::I32x4ExtaddPairwiseI16x8S:
      masm.extAddPairwiseInt16x8(src, dest);
      break;
    case wasm::SimdOp::I32x4ExtaddPairwiseI16x8U:
      masm.unsignedExtAddPairwiseInt16x8(src, dest);
      break;
    case wasm::SimdOp::I8x16Popcnt:
      masm.popcntInt8x16(src, dest);
      break;
    case wasm::SimdOp::I32x4RelaxedTruncF32x4S:
      masm.truncFloat32x4ToInt32x4Relaxed(src, dest);
      break;
    case wasm::SimdOp::I32x4RelaxedTruncF32x4U:
      masm.unsignedTruncFloat32x4ToInt32x4Relaxed(src, dest);
      break;
    case wasm::SimdOp::I32x4RelaxedTruncF64x2SZero:
      masm.truncFloat64x2ToInt32x4Relaxed(src, dest);
      break;
    case wasm::SimdOp::I32x4RelaxedTruncF64x2UZero:
      masm.unsignedTruncFloat64x2ToInt32x4Relaxed(src, dest);
      break;
    default:
      MOZ_CRASH("Unary SimdOp not implemented");
  }
#else
  MOZ_CRASH("No SIMD");
#endif
}

void CodeGenerator::visitWasmReduceSimd128(LWasmReduceSimd128* ins) {
#ifdef ENABLE_WASM_SIMD
  FloatRegister src = ToFloatRegister(ins->src());
  const LDefinition* dest = ins->output();
  uint32_t imm = ins->mir()->imm();
  FloatRegister temp = ToTempFloatRegisterOrInvalid(ins->temp0());

  switch (ins->mir()->simdOp()) {
    case wasm::SimdOp::V128AnyTrue:
      masm.anyTrueSimd128(src, ToRegister(dest));
      break;
    case wasm::SimdOp::I8x16AllTrue:
      masm.allTrueInt8x16(src, ToRegister(dest));
      break;
    case wasm::SimdOp::I16x8AllTrue:
      masm.allTrueInt16x8(src, ToRegister(dest));
      break;
    case wasm::SimdOp::I32x4AllTrue:
      masm.allTrueInt32x4(src, ToRegister(dest));
      break;
    case wasm::SimdOp::I64x2AllTrue:
      masm.allTrueInt64x2(src, ToRegister(dest));
      break;
    case wasm::SimdOp::I8x16Bitmask:
      masm.bitmaskInt8x16(src, ToRegister(dest), temp);
      break;
    case wasm::SimdOp::I16x8Bitmask:
      masm.bitmaskInt16x8(src, ToRegister(dest), temp);
      break;
    case wasm::SimdOp::I32x4Bitmask:
      masm.bitmaskInt32x4(src, ToRegister(dest), temp);
      break;
    case wasm::SimdOp::I64x2Bitmask:
      masm.bitmaskInt64x2(src, ToRegister(dest), temp);
      break;
    case wasm::SimdOp::I8x16ExtractLaneS:
      masm.extractLaneInt8x16(imm, src, ToRegister(dest));
      break;
    case wasm::SimdOp::I8x16ExtractLaneU:
      masm.unsignedExtractLaneInt8x16(imm, src, ToRegister(dest));
      break;
    case wasm::SimdOp::I16x8ExtractLaneS:
      masm.extractLaneInt16x8(imm, src, ToRegister(dest));
      break;
    case wasm::SimdOp::I16x8ExtractLaneU:
      masm.unsignedExtractLaneInt16x8(imm, src, ToRegister(dest));
      break;
    case wasm::SimdOp::I32x4ExtractLane:
      masm.extractLaneInt32x4(imm, src, ToRegister(dest));
      break;
    case wasm::SimdOp::F32x4ExtractLane:
      masm.extractLaneFloat32x4(imm, src, ToFloatRegister(dest));
      break;
    case wasm::SimdOp::F64x2ExtractLane:
      masm.extractLaneFloat64x2(imm, src, ToFloatRegister(dest));
      break;
    default:
      MOZ_CRASH("Reduce SimdOp not implemented");
  }
#else
  MOZ_CRASH("No SIMD");
#endif
}

void CodeGenerator::visitWasmReduceAndBranchSimd128(
    LWasmReduceAndBranchSimd128* ins) {
#ifdef ENABLE_WASM_SIMD
  FloatRegister src = ToFloatRegister(ins->src());

  ScratchSimd128Scope scratch(masm);
  vixl::UseScratchRegisterScope temps(&masm.asVIXL());
  const Register test = temps.AcquireX().asUnsized();

  switch (ins->simdOp()) {
    case wasm::SimdOp::V128AnyTrue:
      masm.Addp(Simd1D(scratch), Simd2D(src));
      masm.Umov(ARMRegister(test, 64), Simd1D(scratch), 0);
      masm.branch64(Assembler::Equal, Register64(test), Imm64(0),
                    getJumpLabelForBranch(ins->ifFalse()));
      jumpToBlock(ins->ifTrue());
      break;
    case wasm::SimdOp::I8x16AllTrue:
    case wasm::SimdOp::I16x8AllTrue:
    case wasm::SimdOp::I32x4AllTrue:
    case wasm::SimdOp::I64x2AllTrue: {
      // Compare all lanes to zero.
      switch (ins->simdOp()) {
        case wasm::SimdOp::I8x16AllTrue:
          masm.Cmeq(Simd16B(scratch), Simd16B(src), 0);
          break;
        case wasm::SimdOp::I16x8AllTrue:
          masm.Cmeq(Simd8H(scratch), Simd8H(src), 0);
          break;
        case wasm::SimdOp::I32x4AllTrue:
          masm.Cmeq(Simd4S(scratch), Simd4S(src), 0);
          break;
        case wasm::SimdOp::I64x2AllTrue:
          masm.Cmeq(Simd2D(scratch), Simd2D(src), 0);
          break;
        default:
          MOZ_CRASH();
      }
      masm.Addp(Simd1D(scratch), Simd2D(scratch));
      masm.Umov(ARMRegister(test, 64), Simd1D(scratch), 0);
      masm.branch64(Assembler::NotEqual, Register64(test), Imm64(0),
                    getJumpLabelForBranch(ins->ifFalse()));
      jumpToBlock(ins->ifTrue());
      break;
    }
    default:
      MOZ_CRASH("Reduce-and-branch SimdOp not implemented");
  }
#else
  MOZ_CRASH("No SIMD");
#endif
}

void CodeGenerator::visitWasmReduceSimd128ToInt64(
    LWasmReduceSimd128ToInt64* ins) {
#ifdef ENABLE_WASM_SIMD
  FloatRegister src = ToFloatRegister(ins->src());
  Register64 dest = ToOutRegister64(ins);
  uint32_t imm = ins->mir()->imm();

  switch (ins->mir()->simdOp()) {
    case wasm::SimdOp::I64x2ExtractLane:
      masm.extractLaneInt64x2(imm, src, dest);
      break;
    default:
      MOZ_CRASH("Reduce SimdOp not implemented");
  }
#else
  MOZ_CRASH("No SIMD");
#endif
}

static inline wasm::MemoryAccessDesc DeriveMemoryAccessDesc(
    const wasm::MemoryAccessDesc& access, Scalar::Type type) {
  return wasm::MemoryAccessDesc(access.memoryIndex(), type, access.align(),
                                access.offset32(), access.trapDesc(),
                                access.isHugeMemory());
}

void CodeGenerator::visitWasmLoadLaneSimd128(LWasmLoadLaneSimd128* ins) {
#ifdef ENABLE_WASM_SIMD
  // Forward loading to wasmLoad, and use replaceLane after that.
  const MWasmLoadLaneSimd128* mir = ins->mir();
  Register memoryBase = ToRegister(ins->memoryBase());
  Register temp = ToRegister(ins->temp0());
  FloatRegister src = ToFloatRegister(ins->src());
  FloatRegister dest = ToFloatRegister(ins->output());
  // replaceLane takes an lhsDest argument.
  masm.moveSimd128(src, dest);
  switch (mir->laneSize()) {
    case 1: {
      masm.wasmLoad(DeriveMemoryAccessDesc(mir->access(), Scalar::Int8),
                    memoryBase, ToRegister(ins->ptr()), AnyRegister(temp));
      masm.replaceLaneInt8x16(mir->laneIndex(), temp, dest);
      break;
    }
    case 2: {
      masm.wasmLoad(DeriveMemoryAccessDesc(mir->access(), Scalar::Int16),
                    memoryBase, ToRegister(ins->ptr()), AnyRegister(temp));
      masm.replaceLaneInt16x8(mir->laneIndex(), temp, dest);
      break;
    }
    case 4: {
      masm.wasmLoad(DeriveMemoryAccessDesc(mir->access(), Scalar::Int32),
                    memoryBase, ToRegister(ins->ptr()), AnyRegister(temp));
      masm.replaceLaneInt32x4(mir->laneIndex(), temp, dest);
      break;
    }
    case 8: {
      masm.wasmLoadI64(DeriveMemoryAccessDesc(mir->access(), Scalar::Int64),
                       memoryBase, ToRegister(ins->ptr()), Register64(temp));
      masm.replaceLaneInt64x2(mir->laneIndex(), Register64(temp), dest);
      break;
    }
    default:
      MOZ_CRASH("Unsupported load lane size");
  }
#else
  MOZ_CRASH("No SIMD");
#endif
}

void CodeGenerator::visitWasmStoreLaneSimd128(LWasmStoreLaneSimd128* ins) {
#ifdef ENABLE_WASM_SIMD
  // Forward storing to wasmStore for the result of extractLane.
  const MWasmStoreLaneSimd128* mir = ins->mir();
  Register memoryBase = ToRegister(ins->memoryBase());
  Register temp = ToRegister(ins->temp0());
  FloatRegister src = ToFloatRegister(ins->src());
  switch (mir->laneSize()) {
    case 1: {
      masm.extractLaneInt8x16(mir->laneIndex(), src, temp);
      masm.wasmStore(DeriveMemoryAccessDesc(mir->access(), Scalar::Int8),
                     AnyRegister(temp), memoryBase, ToRegister(ins->ptr()));
      break;
    }
    case 2: {
      masm.extractLaneInt16x8(mir->laneIndex(), src, temp);
      masm.wasmStore(DeriveMemoryAccessDesc(mir->access(), Scalar::Int16),
                     AnyRegister(temp), memoryBase, ToRegister(ins->ptr()));
      break;
    }
    case 4: {
      masm.extractLaneInt32x4(mir->laneIndex(), src, temp);
      masm.wasmStore(DeriveMemoryAccessDesc(mir->access(), Scalar::Int32),
                     AnyRegister(temp), memoryBase, ToRegister(ins->ptr()));
      break;
    }
    case 8: {
      masm.extractLaneInt64x2(mir->laneIndex(), src, Register64(temp));
      masm.wasmStoreI64(DeriveMemoryAccessDesc(mir->access(), Scalar::Int64),
                        Register64(temp), memoryBase, ToRegister(ins->ptr()));
      break;
    }
    default:
      MOZ_CRASH("Unsupported store lane size");
  }
#else
  MOZ_CRASH("No SIMD");
#endif
}

void CodeGenerator::visitWasmMulI64WideHI64(LWasmMulI64WideHI64* lir) {
  Register lhs = ToRegister(lir->lhs());
  Register rhs = ToRegister(lir->rhs());
  Register output = ToRegister(lir->output());
  // This holds because both operands are non-AtStart variants.
  MOZ_ASSERT(output != lhs && output != rhs);
  masm.wasmMulI64WideHI64(lhs, rhs, output, lir->isSigned());
}