// Copyright (c) 2013, the Dart project authors.  Please see the AUTHORS file
// for details. All rights reserved. Use of this source code is governed by a
// BSD-style license that can be found in the LICENSE file.

#include "vm/globals.h"  // Needed here to get TARGET_ARCH_X64.
#if defined(TARGET_ARCH_X64) && !defined(DART_PRECOMPILED_RUNTIME)

#include "vm/compiler/backend/il.h"

#include "vm/compiler/backend/flow_graph.h"
#include "vm/compiler/backend/flow_graph_compiler.h"
#include "vm/compiler/backend/locations.h"
#include "vm/compiler/backend/locations_helpers.h"
#include "vm/compiler/backend/range_analysis.h"
#include "vm/compiler/jit/compiler.h"
#include "vm/dart_entry.h"
#include "vm/instructions.h"
#include "vm/object_store.h"
#include "vm/parser.h"
#include "vm/stack_frame.h"
#include "vm/stub_code.h"
#include "vm/symbols.h"

#define __ compiler->assembler()->
#define Z (compiler->zone())

namespace dart {

// Generic summary for call instructions that have all arguments pushed
// on the stack and return the result in a fixed register RAX.
LocationSummary* Instruction::MakeCallSummary(Zone* zone) {
  LocationSummary* result =
      new (zone) LocationSummary(zone, 0, 0, LocationSummary::kCall);
  result->set_out(0, Location::RegisterLocation(RAX));
  return result;
}

LocationSummary* PushArgumentInstr::MakeLocationSummary(Zone* zone,
                                                        bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  locs->set_in(0, Location::AnyOrConstant(value()));
  return locs;
}

void PushArgumentInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  // In SSA mode, we need an explicit push. Nothing to do in non-SSA mode
  // where PushArgument is handled by BindInstr::EmitNativeCode.
  if (compiler->is_optimizing()) {
    Location value = locs()->in(0);
    if (value.IsRegister()) {
      __ pushq(value.reg());
    } else if (value.IsConstant()) {
      __ PushObject(value.constant());
    } else {
      ASSERT(value.IsStackSlot());
      __ pushq(value.ToStackSlotAddress());
    }
  }
}

LocationSummary* ReturnInstr::MakeLocationSummary(Zone* zone, bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  locs->set_in(0, Location::RegisterLocation(RAX));
  return locs;
}

// Attempt optimized compilation at return instruction instead of at the entry.
// The entry needs to be patchable, no inlined objects are allowed in the area
// that will be overwritten by the patch instruction: a jump).
void ReturnInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register result = locs()->in(0).reg();
  ASSERT(result == RAX);

  if (compiler->intrinsic_mode()) {
    // Intrinsics don't have a frame.
    __ ret();
    return;
  }

#if defined(DEBUG)
  __ Comment("Stack Check");
  Label done;
  const intptr_t fp_sp_dist =
      (kFirstLocalSlotFromFp + 1 - compiler->StackSize()) * kWordSize;
  ASSERT(fp_sp_dist <= 0);
  __ movq(RDI, RSP);
  __ subq(RDI, RBP);
  __ CompareImmediate(RDI, Immediate(fp_sp_dist));
  __ j(EQUAL, &done, Assembler::kNearJump);
  __ int3();
  __ Bind(&done);
#endif
  ASSERT(__ constant_pool_allowed());
  __ LeaveDartFrame();  // Disallows constant pool use.
  __ ret();
  // This ReturnInstr may be emitted out of order by the optimizer. The next
  // block may be a target expecting a properly set constant pool pointer.
  __ set_constant_pool_allowed(true);
}

static Condition NegateCondition(Condition condition) {
  switch (condition) {
    case EQUAL:
      return NOT_EQUAL;
    case NOT_EQUAL:
      return EQUAL;
    case LESS:
      return GREATER_EQUAL;
    case LESS_EQUAL:
      return GREATER;
    case GREATER:
      return LESS_EQUAL;
    case GREATER_EQUAL:
      return LESS;
    case BELOW:
      return ABOVE_EQUAL;
    case BELOW_EQUAL:
      return ABOVE;
    case ABOVE:
      return BELOW_EQUAL;
    case ABOVE_EQUAL:
      return BELOW;
    case PARITY_EVEN:
      return PARITY_ODD;
    case PARITY_ODD:
      return PARITY_EVEN;
    default:
      UNIMPLEMENTED();
      return EQUAL;
  }
}

// Detect pattern when one value is zero and another is a power of 2.
static bool IsPowerOfTwoKind(intptr_t v1, intptr_t v2) {
  return (Utils::IsPowerOfTwo(v1) && (v2 == 0)) ||
         (Utils::IsPowerOfTwo(v2) && (v1 == 0));
}

LocationSummary* IfThenElseInstr::MakeLocationSummary(Zone* zone,
                                                      bool opt) const {
  comparison()->InitializeLocationSummary(zone, opt);
  // TODO(dartbug.com/30952) support convertion of Register to corresponding
  // least significant byte register (e.g. RAX -> AL, RSI -> SIL, r15 -> r15b).
  comparison()->locs()->set_out(0, Location::RegisterLocation(RDX));
  return comparison()->locs();
}

void IfThenElseInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  ASSERT(locs()->out(0).reg() == RDX);

  // Clear upper part of the out register. We are going to use setcc on it
  // which is a byte move.
  __ xorq(RDX, RDX);

  // Emit comparison code. This must not overwrite the result register.
  // IfThenElseInstr::Supports() should prevent EmitComparisonCode from using
  // the labels or returning an invalid condition.
  BranchLabels labels = {NULL, NULL, NULL};
  Condition true_condition = comparison()->EmitComparisonCode(compiler, labels);
  ASSERT(true_condition != INVALID_CONDITION);

  const bool is_power_of_two_kind = IsPowerOfTwoKind(if_true_, if_false_);

  intptr_t true_value = if_true_;
  intptr_t false_value = if_false_;

  if (is_power_of_two_kind) {
    if (true_value == 0) {
      // We need to have zero in RDX on true_condition.
      true_condition = NegateCondition(true_condition);
    }
  } else {
    if (true_value == 0) {
      // Swap values so that false_value is zero.
      intptr_t temp = true_value;
      true_value = false_value;
      false_value = temp;
    } else {
      true_condition = NegateCondition(true_condition);
    }
  }

  __ setcc(true_condition, DL);

  if (is_power_of_two_kind) {
    const intptr_t shift =
        Utils::ShiftForPowerOfTwo(Utils::Maximum(true_value, false_value));
    __ shlq(RDX, Immediate(shift + kSmiTagSize));
  } else {
    __ decq(RDX);
    __ AndImmediate(
        RDX, Immediate(Smi::RawValue(true_value) - Smi::RawValue(false_value)));
    if (false_value != 0) {
      __ AddImmediate(RDX, Immediate(Smi::RawValue(false_value)));
    }
  }
}

LocationSummary* LoadLocalInstr::MakeLocationSummary(Zone* zone,
                                                     bool opt) const {
  const intptr_t kNumInputs = 0;
  const intptr_t stack_index = (local().index() < 0)
                                   ? kFirstLocalSlotFromFp - local().index()
                                   : kParamEndSlotFromFp - local().index();
  return LocationSummary::Make(zone, kNumInputs,
                               Location::StackSlot(stack_index),
                               LocationSummary::kNoCall);
}

void LoadLocalInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  ASSERT(!compiler->is_optimizing());
  // Nothing to do.
}

LocationSummary* StoreLocalInstr::MakeLocationSummary(Zone* zone,
                                                      bool opt) const {
  const intptr_t kNumInputs = 1;
  return LocationSummary::Make(zone, kNumInputs, Location::SameAsFirstInput(),
                               LocationSummary::kNoCall);
}

void StoreLocalInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register value = locs()->in(0).reg();
  Register result = locs()->out(0).reg();
  ASSERT(result == value);  // Assert that register assignment is correct.
  __ movq(Address(RBP, local().index() * kWordSize), value);
}

LocationSummary* ConstantInstr::MakeLocationSummary(Zone* zone,
                                                    bool opt) const {
  const intptr_t kNumInputs = 0;
  return LocationSummary::Make(zone, kNumInputs,
                               Assembler::IsSafe(value())
                                   ? Location::Constant(this)
                                   : Location::RequiresRegister(),
                               LocationSummary::kNoCall);
}

void ConstantInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  // The register allocator drops constant definitions that have no uses.
  Location out = locs()->out(0);
  ASSERT(out.IsRegister() || out.IsConstant() || out.IsInvalid());
  if (out.IsRegister()) {
    Register result = out.reg();
    __ LoadObject(result, value());
  }
}

LocationSummary* UnboxedConstantInstr::MakeLocationSummary(Zone* zone,
                                                           bool opt) const {
  const intptr_t kNumInputs = 0;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  switch (representation()) {
    case kUnboxedDouble:
      locs->set_out(0, Location::RequiresFpuRegister());
      break;
    case kUnboxedInt32:
      locs->set_out(0, Location::RequiresRegister());
      break;
    default:
      UNREACHABLE();
      break;
  }
  return locs;
}

void UnboxedConstantInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  // The register allocator drops constant definitions that have no uses.
  if (!locs()->out(0).IsInvalid()) {
    switch (representation()) {
      case kUnboxedDouble: {
        XmmRegister result = locs()->out(0).fpu_reg();
        if (Utils::DoublesBitEqual(Double::Cast(value()).value(), 0.0)) {
          __ xorps(result, result);
        } else {
          __ LoadObject(TMP, value());
          __ movsd(result, FieldAddress(TMP, Double::value_offset()));
        }
        break;
      }
      case kUnboxedInt32:
        __ movl(locs()->out(0).reg(),
                Immediate(static_cast<int32_t>(Smi::Cast(value()).Value())));
        break;
      default:
        UNREACHABLE();
    }
  }
}

LocationSummary* AssertAssignableInstr::MakeLocationSummary(Zone* zone,
                                                            bool opt) const {
  const intptr_t kNumInputs = 3;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kCall);
  summary->set_in(0, Location::RegisterLocation(RAX));  // Value.
  summary->set_in(1, Location::RegisterLocation(RDX));  // Instant. type args.
  summary->set_in(2, Location::RegisterLocation(RCX));  // Function type args.
  summary->set_out(0, Location::RegisterLocation(RAX));
  return summary;
}

LocationSummary* AssertSubtypeInstr::MakeLocationSummary(Zone* zone,
                                                         bool opt) const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kCall);
  summary->set_in(0, Location::RegisterLocation(RDX));  // Instant. type args
  summary->set_in(1, Location::RegisterLocation(RCX));  // Function type args
  return summary;
}

LocationSummary* AssertBooleanInstr::MakeLocationSummary(Zone* zone,
                                                         bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kCall);
  locs->set_in(0, Location::RegisterLocation(RAX));
  locs->set_out(0, Location::RegisterLocation(RAX));
  return locs;
}

static void EmitAssertBoolean(Register reg,
                              TokenPosition token_pos,
                              intptr_t deopt_id,
                              LocationSummary* locs,
                              FlowGraphCompiler* compiler) {
  // Check that the type of the value is allowed in conditional context.
  // Call the runtime if the object is not bool::true or bool::false.
  ASSERT(locs->always_calls());
  Label done;

  if (Isolate::Current()->type_checks()) {
    __ CompareObject(reg, Bool::True());
    __ j(EQUAL, &done, Assembler::kNearJump);
    __ CompareObject(reg, Bool::False());
    __ j(EQUAL, &done, Assembler::kNearJump);
  } else {
    ASSERT(Isolate::Current()->asserts());
    __ CompareObject(reg, Object::null_instance());
    __ j(NOT_EQUAL, &done, Assembler::kNearJump);
  }

  __ pushq(reg);  // Push the source object.
  compiler->GenerateRuntimeCall(token_pos, deopt_id,
                                kNonBoolTypeErrorRuntimeEntry, 1, locs);
  // We should never return here.
  __ int3();
  __ Bind(&done);
}

void AssertBooleanInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register obj = locs()->in(0).reg();
  Register result = locs()->out(0).reg();

  EmitAssertBoolean(obj, token_pos(), deopt_id(), locs(), compiler);
  ASSERT(obj == result);
}

static Condition TokenKindToIntCondition(Token::Kind kind) {
  switch (kind) {
    case Token::kEQ:
      return EQUAL;
    case Token::kNE:
      return NOT_EQUAL;
    case Token::kLT:
      return LESS;
    case Token::kGT:
      return GREATER;
    case Token::kLTE:
      return LESS_EQUAL;
    case Token::kGTE:
      return GREATER_EQUAL;
    default:
      UNREACHABLE();
      return OVERFLOW;
  }
}

LocationSummary* EqualityCompareInstr::MakeLocationSummary(Zone* zone,
                                                           bool opt) const {
  const intptr_t kNumInputs = 2;
  if (operation_cid() == kMintCid) {
    const intptr_t kNumTemps = 0;
    LocationSummary* locs = new (zone)
        LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
    locs->set_in(0, Location::RequiresRegister());
    locs->set_in(1, Location::RequiresRegister());
    locs->set_out(0, Location::RequiresRegister());
    return locs;
  }
  if (operation_cid() == kDoubleCid) {
    const intptr_t kNumTemps = 0;
    LocationSummary* locs = new (zone)
        LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
    locs->set_in(0, Location::RequiresFpuRegister());
    locs->set_in(1, Location::RequiresFpuRegister());
    locs->set_out(0, Location::RequiresRegister());
    return locs;
  }
  if (operation_cid() == kSmiCid) {
    const intptr_t kNumTemps = 0;
    LocationSummary* locs = new (zone)
        LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
    locs->set_in(0, Location::RegisterOrConstant(left()));
    // Only one input can be a constant operand. The case of two constant
    // operands should be handled by constant propagation.
    // Only right can be a stack slot.
    locs->set_in(1, locs->in(0).IsConstant()
                        ? Location::RequiresRegister()
                        : Location::RegisterOrConstant(right()));
    locs->set_out(0, Location::RequiresRegister());
    return locs;
  }
  UNREACHABLE();
  return NULL;
}

static void LoadValueCid(FlowGraphCompiler* compiler,
                         Register value_cid_reg,
                         Register value_reg,
                         Label* value_is_smi = NULL) {
  Label done;
  if (value_is_smi == NULL) {
    __ LoadImmediate(value_cid_reg, Immediate(kSmiCid));
  }
  __ testq(value_reg, Immediate(kSmiTagMask));
  if (value_is_smi == NULL) {
    __ j(ZERO, &done, Assembler::kNearJump);
  } else {
    __ j(ZERO, value_is_smi);
  }
  __ LoadClassId(value_cid_reg, value_reg);
  __ Bind(&done);
}

static Condition FlipCondition(Condition condition) {
  switch (condition) {
    case EQUAL:
      return EQUAL;
    case NOT_EQUAL:
      return NOT_EQUAL;
    case LESS:
      return GREATER;
    case LESS_EQUAL:
      return GREATER_EQUAL;
    case GREATER:
      return LESS;
    case GREATER_EQUAL:
      return LESS_EQUAL;
    case BELOW:
      return ABOVE;
    case BELOW_EQUAL:
      return ABOVE_EQUAL;
    case ABOVE:
      return BELOW;
    case ABOVE_EQUAL:
      return BELOW_EQUAL;
    default:
      UNIMPLEMENTED();
      return EQUAL;
  }
}

static void EmitBranchOnCondition(FlowGraphCompiler* compiler,
                                  Condition true_condition,
                                  BranchLabels labels) {
  if (labels.fall_through == labels.false_label) {
    // If the next block is the false successor, fall through to it.
    __ j(true_condition, labels.true_label);
  } else {
    // If the next block is not the false successor, branch to it.
    Condition false_condition = NegateCondition(true_condition);
    __ j(false_condition, labels.false_label);

    // Fall through or jump to the true successor.
    if (labels.fall_through != labels.true_label) {
      __ jmp(labels.true_label);
    }
  }
}

static Condition EmitInt64ComparisonOp(FlowGraphCompiler* compiler,
                                       const LocationSummary& locs,
                                       Token::Kind kind) {
  Location left = locs.in(0);
  Location right = locs.in(1);
  ASSERT(!left.IsConstant() || !right.IsConstant());

  Condition true_condition = TokenKindToIntCondition(kind);

  if (left.IsConstant()) {
    __ CompareObject(right.reg(), left.constant());
    true_condition = FlipCondition(true_condition);
  } else if (right.IsConstant()) {
    __ CompareObject(left.reg(), right.constant());
  } else if (right.IsStackSlot()) {
    __ cmpq(left.reg(), right.ToStackSlotAddress());
  } else {
    __ cmpq(left.reg(), right.reg());
  }
  return true_condition;
}

static Condition TokenKindToDoubleCondition(Token::Kind kind) {
  switch (kind) {
    case Token::kEQ:
      return EQUAL;
    case Token::kNE:
      return NOT_EQUAL;
    case Token::kLT:
      return BELOW;
    case Token::kGT:
      return ABOVE;
    case Token::kLTE:
      return BELOW_EQUAL;
    case Token::kGTE:
      return ABOVE_EQUAL;
    default:
      UNREACHABLE();
      return OVERFLOW;
  }
}

static Condition EmitDoubleComparisonOp(FlowGraphCompiler* compiler,
                                        const LocationSummary& locs,
                                        Token::Kind kind,
                                        BranchLabels labels) {
  XmmRegister left = locs.in(0).fpu_reg();
  XmmRegister right = locs.in(1).fpu_reg();

  __ comisd(left, right);

  Condition true_condition = TokenKindToDoubleCondition(kind);
  Label* nan_result =
      (true_condition == NOT_EQUAL) ? labels.true_label : labels.false_label;
  __ j(PARITY_EVEN, nan_result);
  return true_condition;
}

Condition EqualityCompareInstr::EmitComparisonCode(FlowGraphCompiler* compiler,
                                                   BranchLabels labels) {
  if ((operation_cid() == kSmiCid) || (operation_cid() == kMintCid)) {
    return EmitInt64ComparisonOp(compiler, *locs(), kind());
  } else {
    ASSERT(operation_cid() == kDoubleCid);
    return EmitDoubleComparisonOp(compiler, *locs(), kind(), labels);
  }
}

void ComparisonInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Label is_true, is_false;
  BranchLabels labels = {&is_true, &is_false, &is_false};
  Condition true_condition = EmitComparisonCode(compiler, labels);
  if (true_condition != INVALID_CONDITION) {
    EmitBranchOnCondition(compiler, true_condition, labels);
  }

  Register result = locs()->out(0).reg();
  Label done;
  __ Bind(&is_false);
  __ LoadObject(result, Bool::False());
  __ jmp(&done);
  __ Bind(&is_true);
  __ LoadObject(result, Bool::True());
  __ Bind(&done);
}

void ComparisonInstr::EmitBranchCode(FlowGraphCompiler* compiler,
                                     BranchInstr* branch) {
  BranchLabels labels = compiler->CreateBranchLabels(branch);
  Condition true_condition = EmitComparisonCode(compiler, labels);
  if (true_condition != INVALID_CONDITION) {
    EmitBranchOnCondition(compiler, true_condition, labels);
  }
}

LocationSummary* TestSmiInstr::MakeLocationSummary(Zone* zone, bool opt) const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  locs->set_in(0, Location::RequiresRegister());
  // Only one input can be a constant operand. The case of two constant
  // operands should be handled by constant propagation.
  locs->set_in(1, Location::RegisterOrConstant(right()));
  return locs;
}

Condition TestSmiInstr::EmitComparisonCode(FlowGraphCompiler* compiler,
                                           BranchLabels labels) {
  Register left_reg = locs()->in(0).reg();
  Location right = locs()->in(1);
  if (right.IsConstant()) {
    ASSERT(right.constant().IsSmi());
    const int64_t imm = reinterpret_cast<int64_t>(right.constant().raw());
    __ TestImmediate(left_reg, Immediate(imm));
  } else {
    __ testq(left_reg, right.reg());
  }
  Condition true_condition = (kind() == Token::kNE) ? NOT_ZERO : ZERO;
  return true_condition;
}

LocationSummary* TestCidsInstr::MakeLocationSummary(Zone* zone,
                                                    bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 1;
  LocationSummary* locs = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  locs->set_in(0, Location::RequiresRegister());
  locs->set_temp(0, Location::RequiresRegister());
  locs->set_out(0, Location::RequiresRegister());
  return locs;
}

Condition TestCidsInstr::EmitComparisonCode(FlowGraphCompiler* compiler,
                                            BranchLabels labels) {
  ASSERT((kind() == Token::kIS) || (kind() == Token::kISNOT));
  Register val_reg = locs()->in(0).reg();
  Register cid_reg = locs()->temp(0).reg();

  Label* deopt = CanDeoptimize() ? compiler->AddDeoptStub(
                                       deopt_id(), ICData::kDeoptTestCids,
                                       licm_hoisted_ ? ICData::kHoisted : 0)
                                 : NULL;

  const intptr_t true_result = (kind() == Token::kIS) ? 1 : 0;
  const ZoneGrowableArray<intptr_t>& data = cid_results();
  ASSERT(data[0] == kSmiCid);
  bool result = data[1] == true_result;
  __ testq(val_reg, Immediate(kSmiTagMask));
  __ j(ZERO, result ? labels.true_label : labels.false_label);
  __ LoadClassId(cid_reg, val_reg);
  for (intptr_t i = 2; i < data.length(); i += 2) {
    const intptr_t test_cid = data[i];
    ASSERT(test_cid != kSmiCid);
    result = data[i + 1] == true_result;
    __ cmpq(cid_reg, Immediate(test_cid));
    __ j(EQUAL, result ? labels.true_label : labels.false_label);
  }
  // No match found, deoptimize or default action.
  if (deopt == NULL) {
    // If the cid is not in the list, jump to the opposite label from the cids
    // that are in the list.  These must be all the same (see asserts in the
    // constructor).
    Label* target = result ? labels.false_label : labels.true_label;
    if (target != labels.fall_through) {
      __ jmp(target);
    }
  } else {
    __ jmp(deopt);
  }
  // Dummy result as this method already did the jump, there's no need
  // for the caller to branch on a condition.
  return INVALID_CONDITION;
}

LocationSummary* RelationalOpInstr::MakeLocationSummary(Zone* zone,
                                                        bool opt) const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  if (operation_cid() == kDoubleCid) {
    LocationSummary* summary = new (zone)
        LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
    summary->set_in(0, Location::RequiresFpuRegister());
    summary->set_in(1, Location::RequiresFpuRegister());
    summary->set_out(0, Location::RequiresRegister());
    return summary;
  } else if (operation_cid() == kMintCid) {
    LocationSummary* summary = new (zone)
        LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
    summary->set_in(0, Location::RequiresRegister());
    summary->set_in(1, Location::RequiresRegister());
    summary->set_out(0, Location::RequiresRegister());
    return summary;
  }
  ASSERT(operation_cid() == kSmiCid);
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RegisterOrConstant(left()));
  // Only one input can be a constant operand. The case of two constant
  // operands should be handled by constant propagation.
  summary->set_in(1, summary->in(0).IsConstant()
                         ? Location::RequiresRegister()
                         : Location::RegisterOrConstant(right()));
  summary->set_out(0, Location::RequiresRegister());
  return summary;
}

Condition RelationalOpInstr::EmitComparisonCode(FlowGraphCompiler* compiler,
                                                BranchLabels labels) {
  if ((operation_cid() == kSmiCid) || (operation_cid() == kMintCid)) {
    return EmitInt64ComparisonOp(compiler, *locs(), kind());
  } else {
    ASSERT(operation_cid() == kDoubleCid);
    return EmitDoubleComparisonOp(compiler, *locs(), kind(), labels);
  }
}

LocationSummary* NativeCallInstr::MakeLocationSummary(Zone* zone,
                                                      bool opt) const {
  return MakeCallSummary(zone);
}

void NativeCallInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  SetupNative();
  Register result = locs()->out(0).reg();
  const intptr_t argc_tag = NativeArguments::ComputeArgcTag(function());

  // All arguments are already @RSP due to preceding PushArgument()s.
  ASSERT(ArgumentCount() == function().NumParameters());

  // Push the result place holder initialized to NULL.
  __ PushObject(Object::null_object());

  // Pass a pointer to the first argument in RAX.
  __ leaq(RAX, Address(RSP, ArgumentCount() * kWordSize));

  __ LoadImmediate(R10, Immediate(argc_tag));
  const StubEntry* stub_entry;
  if (link_lazily()) {
    stub_entry = StubCode::CallBootstrapNative_entry();
    ExternalLabel label(NativeEntry::LinkNativeCallEntry());
    __ LoadNativeEntry(RBX, &label, kPatchable);
    compiler->GeneratePatchableCall(token_pos(), *stub_entry,
                                    RawPcDescriptors::kOther, locs());
  } else {
    if (is_bootstrap_native()) {
      stub_entry = StubCode::CallBootstrapNative_entry();
    } else if (is_auto_scope()) {
      stub_entry = StubCode::CallAutoScopeNative_entry();
    } else {
      stub_entry = StubCode::CallNoScopeNative_entry();
    }
    const ExternalLabel label(reinterpret_cast<uword>(native_c_function()));
    __ LoadNativeEntry(RBX, &label, kNotPatchable);
    compiler->GenerateCall(token_pos(), *stub_entry, RawPcDescriptors::kOther,
                           locs());
  }
  __ popq(result);

  __ Drop(ArgumentCount());  // Drop the arguments.
}

static bool CanBeImmediateIndex(Value* index, intptr_t cid) {
  if (!index->definition()->IsConstant()) return false;
  const Object& constant = index->definition()->AsConstant()->value();
  if (!constant.IsSmi()) return false;
  const Smi& smi_const = Smi::Cast(constant);
  const intptr_t scale = Instance::ElementSizeFor(cid);
  const intptr_t data_offset = Instance::DataOffsetFor(cid);
  const int64_t disp = smi_const.AsInt64Value() * scale + data_offset;
  return Utils::IsInt(32, disp);
}

LocationSummary* OneByteStringFromCharCodeInstr::MakeLocationSummary(
    Zone* zone,
    bool opt) const {
  const intptr_t kNumInputs = 1;
  // TODO(fschneider): Allow immediate operands for the char code.
  return LocationSummary::Make(zone, kNumInputs, Location::RequiresRegister(),
                               LocationSummary::kNoCall);
}

void OneByteStringFromCharCodeInstr::EmitNativeCode(
    FlowGraphCompiler* compiler) {
  ASSERT(compiler->is_optimizing());
  Register char_code = locs()->in(0).reg();
  Register result = locs()->out(0).reg();

  __ movq(result, Address(THR, Thread::predefined_symbols_address_offset()));
  __ movq(result, Address(result, char_code,
                          TIMES_HALF_WORD_SIZE,  // Char code is a smi.
                          Symbols::kNullCharCodeSymbolOffset * kWordSize));
}

LocationSummary* StringToCharCodeInstr::MakeLocationSummary(Zone* zone,
                                                            bool opt) const {
  const intptr_t kNumInputs = 1;
  return LocationSummary::Make(zone, kNumInputs, Location::RequiresRegister(),
                               LocationSummary::kNoCall);
}

void StringToCharCodeInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  ASSERT(cid_ == kOneByteStringCid);
  Register str = locs()->in(0).reg();
  Register result = locs()->out(0).reg();
  Label is_one, done;
  __ movq(result, FieldAddress(str, String::length_offset()));
  __ cmpq(result, Immediate(Smi::RawValue(1)));
  __ j(EQUAL, &is_one, Assembler::kNearJump);
  __ movq(result, Immediate(Smi::RawValue(-1)));
  __ jmp(&done);
  __ Bind(&is_one);
  __ movzxb(result, FieldAddress(str, OneByteString::data_offset()));
  __ SmiTag(result);
  __ Bind(&done);
}

LocationSummary* StringInterpolateInstr::MakeLocationSummary(Zone* zone,
                                                             bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kCall);
  summary->set_in(0, Location::RegisterLocation(RAX));
  summary->set_out(0, Location::RegisterLocation(RAX));
  return summary;
}

void StringInterpolateInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register array = locs()->in(0).reg();
  __ pushq(array);
  const int kTypeArgsLen = 0;
  const int kNumberOfArguments = 1;
  const Array& kNoArgumentNames = Object::null_array();
  ArgumentsInfo args_info(kTypeArgsLen, kNumberOfArguments, kNoArgumentNames);
  compiler->GenerateStaticCall(deopt_id(), token_pos(), CallFunction(),
                               args_info, locs(), ICData::Handle(),
                               ICData::kStatic);
  ASSERT(locs()->out(0).reg() == RAX);
}

LocationSummary* LoadUntaggedInstr::MakeLocationSummary(Zone* zone,
                                                        bool opt) const {
  const intptr_t kNumInputs = 1;
  return LocationSummary::Make(zone, kNumInputs, Location::RequiresRegister(),
                               LocationSummary::kNoCall);
}

void LoadUntaggedInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register obj = locs()->in(0).reg();
  Register result = locs()->out(0).reg();
  if (object()->definition()->representation() == kUntagged) {
    __ movq(result, Address(obj, offset()));
  } else {
    ASSERT(object()->definition()->representation() == kTagged);
    __ movq(result, FieldAddress(obj, offset()));
  }
}

LocationSummary* LoadClassIdInstr::MakeLocationSummary(Zone* zone,
                                                       bool opt) const {
  const intptr_t kNumInputs = 1;
  return LocationSummary::Make(zone, kNumInputs, Location::RequiresRegister(),
                               LocationSummary::kNoCall);
}

void LoadClassIdInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  const Register object = locs()->in(0).reg();
  const Register result = locs()->out(0).reg();
  const AbstractType& value_type = *this->object()->Type()->ToAbstractType();
  if (CompileType::Smi().IsAssignableTo(value_type) ||
      value_type.IsTypeParameter()) {
    // We don't use Assembler::LoadTaggedClassIdMayBeSmi() here---which uses
    // a conditional move instead---because it is slower, probably due to
    // branch prediction usually working just fine in this case.
    Label load, done;
    __ testq(object, Immediate(kSmiTagMask));
    __ j(NOT_ZERO, &load, Assembler::kNearJump);
    __ LoadImmediate(result, Immediate(Smi::RawValue(kSmiCid)));
    __ jmp(&done);
    __ Bind(&load);
    __ LoadClassId(result, object);
    __ SmiTag(result);
    __ Bind(&done);
  } else {
    __ LoadClassId(result, object);
    __ SmiTag(result);
  }
}

CompileType LoadIndexedInstr::ComputeType() const {
  switch (class_id_) {
    case kArrayCid:
    case kImmutableArrayCid:
      return CompileType::Dynamic();

    case kTypedDataFloat32ArrayCid:
    case kTypedDataFloat64ArrayCid:
      return CompileType::FromCid(kDoubleCid);
    case kTypedDataFloat32x4ArrayCid:
      return CompileType::FromCid(kFloat32x4Cid);
    case kTypedDataInt32x4ArrayCid:
      return CompileType::FromCid(kInt32x4Cid);
    case kTypedDataFloat64x2ArrayCid:
      return CompileType::FromCid(kFloat64x2Cid);

    case kTypedDataInt8ArrayCid:
    case kTypedDataUint8ArrayCid:
    case kTypedDataUint8ClampedArrayCid:
    case kExternalTypedDataUint8ArrayCid:
    case kExternalTypedDataUint8ClampedArrayCid:
    case kTypedDataInt16ArrayCid:
    case kTypedDataUint16ArrayCid:
    case kOneByteStringCid:
    case kTwoByteStringCid:
    case kExternalOneByteStringCid:
    case kExternalTwoByteStringCid:
    case kTypedDataInt32ArrayCid:
    case kTypedDataUint32ArrayCid:
      return CompileType::FromCid(kSmiCid);

    case kTypedDataInt64ArrayCid:
      return CompileType::Int();

    default:
      UNIMPLEMENTED();
      return CompileType::Dynamic();
  }
}

Representation LoadIndexedInstr::representation() const {
  switch (class_id_) {
    case kArrayCid:
    case kImmutableArrayCid:
    case kTypedDataInt8ArrayCid:
    case kTypedDataUint8ArrayCid:
    case kTypedDataUint8ClampedArrayCid:
    case kExternalTypedDataUint8ArrayCid:
    case kExternalTypedDataUint8ClampedArrayCid:
    case kTypedDataInt16ArrayCid:
    case kTypedDataUint16ArrayCid:
    case kOneByteStringCid:
    case kTwoByteStringCid:
    case kExternalOneByteStringCid:
    case kExternalTwoByteStringCid:
      return kTagged;
    case kTypedDataInt32ArrayCid:
      return kUnboxedInt32;
    case kTypedDataUint32ArrayCid:
      return kUnboxedUint32;
    case kTypedDataInt64ArrayCid:
      return kUnboxedInt64;
    case kTypedDataFloat32ArrayCid:
    case kTypedDataFloat64ArrayCid:
      return kUnboxedDouble;
    case kTypedDataInt32x4ArrayCid:
      return kUnboxedInt32x4;
    case kTypedDataFloat32x4ArrayCid:
      return kUnboxedFloat32x4;
    case kTypedDataFloat64x2ArrayCid:
      return kUnboxedFloat64x2;
    default:
      UNIMPLEMENTED();
      return kTagged;
  }
}

LocationSummary* LoadIndexedInstr::MakeLocationSummary(Zone* zone,
                                                       bool opt) const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  locs->set_in(0, Location::RequiresRegister());
  // The smi index is either untagged (element size == 1), or it is left smi
  // tagged (for all element sizes > 1).
  if (index_scale() == 1) {
    locs->set_in(1,
                 CanBeImmediateIndex(index(), class_id())
                     ? Location::Constant(index()->definition()->AsConstant())
                     : Location::WritableRegister());
  } else {
    locs->set_in(1,
                 CanBeImmediateIndex(index(), class_id())
                     ? Location::Constant(index()->definition()->AsConstant())
                     : Location::RequiresRegister());
  }
  if ((representation() == kUnboxedDouble) ||
      (representation() == kUnboxedFloat32x4) ||
      (representation() == kUnboxedInt32x4) ||
      (representation() == kUnboxedFloat64x2)) {
    locs->set_out(0, Location::RequiresFpuRegister());
  } else {
    locs->set_out(0, Location::RequiresRegister());
  }
  return locs;
}

void LoadIndexedInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  // The array register points to the backing store for external arrays.
  const Register array = locs()->in(0).reg();
  const Location index = locs()->in(1);

  Address element_address =
      index.IsRegister()
          ? Assembler::ElementAddressForRegIndex(
                IsExternal(), class_id(), index_scale(), array, index.reg())
          : Assembler::ElementAddressForIntIndex(
                IsExternal(), class_id(), index_scale(), array,
                Smi::Cast(index.constant()).Value());

  if ((representation() == kUnboxedDouble) ||
      (representation() == kUnboxedFloat32x4) ||
      (representation() == kUnboxedInt32x4) ||
      (representation() == kUnboxedFloat64x2)) {
    if ((index_scale() == 1) && index.IsRegister()) {
      __ SmiUntag(index.reg());
    }

    XmmRegister result = locs()->out(0).fpu_reg();
    if (class_id() == kTypedDataFloat32ArrayCid) {
      // Load single precision float.
      __ movss(result, element_address);
    } else if (class_id() == kTypedDataFloat64ArrayCid) {
      __ movsd(result, element_address);
    } else {
      ASSERT((class_id() == kTypedDataInt32x4ArrayCid) ||
             (class_id() == kTypedDataFloat32x4ArrayCid) ||
             (class_id() == kTypedDataFloat64x2ArrayCid));
      __ movups(result, element_address);
    }
    return;
  }

  if ((representation() == kUnboxedUint32) ||
      (representation() == kUnboxedInt32)) {
    if ((index_scale() == 1) && index.IsRegister()) {
      __ SmiUntag(index.reg());
    }
    Register result = locs()->out(0).reg();
    switch (class_id()) {
      case kTypedDataInt32ArrayCid:
        ASSERT(representation() == kUnboxedInt32);
        __ movsxd(result, element_address);
        break;
      case kTypedDataUint32ArrayCid:
        ASSERT(representation() == kUnboxedUint32);
        __ movl(result, element_address);
        break;
      default:
        UNREACHABLE();
    }
    return;
  }

  if (representation() == kUnboxedInt64) {
    ASSERT(class_id() == kTypedDataInt64ArrayCid);
    if ((index_scale() == 1) && index.IsRegister()) {
      __ SmiUntag(index.reg());
    }
    Register result = locs()->out(0).reg();
    __ movq(result, element_address);
    return;
  }

  ASSERT(representation() == kTagged);

  if ((index_scale() == 1) && index.IsRegister()) {
    __ SmiUntag(index.reg());
  }
  Register result = locs()->out(0).reg();
  switch (class_id()) {
    case kTypedDataInt8ArrayCid:
      __ movsxb(result, element_address);
      __ SmiTag(result);
      break;
    case kTypedDataUint8ArrayCid:
    case kTypedDataUint8ClampedArrayCid:
    case kExternalTypedDataUint8ArrayCid:
    case kExternalTypedDataUint8ClampedArrayCid:
    case kOneByteStringCid:
    case kExternalOneByteStringCid:
      __ movzxb(result, element_address);
      __ SmiTag(result);
      break;
    case kTypedDataInt16ArrayCid:
      __ movsxw(result, element_address);
      __ SmiTag(result);
      break;
    case kTypedDataUint16ArrayCid:
    case kTwoByteStringCid:
    case kExternalTwoByteStringCid:
      __ movzxw(result, element_address);
      __ SmiTag(result);
      break;
    default:
      ASSERT((class_id() == kArrayCid) || (class_id() == kImmutableArrayCid));
      __ movq(result, element_address);
      break;
  }
}

LocationSummary* LoadCodeUnitsInstr::MakeLocationSummary(Zone* zone,
                                                         bool opt) const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  // The smi index is either untagged (element size == 1), or it is left smi
  // tagged (for all element sizes > 1).
  summary->set_in(1, index_scale() == 1 ? Location::WritableRegister()
                                        : Location::RequiresRegister());
  summary->set_out(0, Location::RequiresRegister());
  return summary;
}

void LoadCodeUnitsInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  // The string register points to the backing store for external strings.
  const Register str = locs()->in(0).reg();
  const Location index = locs()->in(1);

  Address element_address = Assembler::ElementAddressForRegIndex(
      IsExternal(), class_id(), index_scale(), str, index.reg());

  if ((index_scale() == 1)) {
    __ SmiUntag(index.reg());
  }
  Register result = locs()->out(0).reg();
  switch (class_id()) {
    case kOneByteStringCid:
    case kExternalOneByteStringCid:
      switch (element_count()) {
        case 1:
          __ movzxb(result, element_address);
          break;
        case 2:
          __ movzxw(result, element_address);
          break;
        case 4:
          __ movl(result, element_address);
          break;
        default:
          UNREACHABLE();
      }
      __ SmiTag(result);
      break;
    case kTwoByteStringCid:
    case kExternalTwoByteStringCid:
      switch (element_count()) {
        case 1:
          __ movzxw(result, element_address);
          break;
        case 2:
          __ movl(result, element_address);
          break;
        default:
          UNREACHABLE();
      }
      __ SmiTag(result);
      break;
    default:
      UNREACHABLE();
      break;
  }
}

Representation StoreIndexedInstr::RequiredInputRepresentation(
    intptr_t idx) const {
  if (idx == 0) return kNoRepresentation;
  if (idx == 1) return kTagged;
  ASSERT(idx == 2);
  switch (class_id_) {
    case kArrayCid:
    case kOneByteStringCid:
    case kTypedDataInt8ArrayCid:
    case kTypedDataUint8ArrayCid:
    case kExternalTypedDataUint8ArrayCid:
    case kTypedDataUint8ClampedArrayCid:
    case kExternalTypedDataUint8ClampedArrayCid:
    case kTypedDataInt16ArrayCid:
    case kTypedDataUint16ArrayCid:
      return kTagged;
    case kTypedDataInt32ArrayCid:
      return kUnboxedInt32;
    case kTypedDataUint32ArrayCid:
      return kUnboxedUint32;
    case kTypedDataInt64ArrayCid:
      return kUnboxedInt64;
    case kTypedDataFloat32ArrayCid:
    case kTypedDataFloat64ArrayCid:
      return kUnboxedDouble;
    case kTypedDataFloat32x4ArrayCid:
      return kUnboxedFloat32x4;
    case kTypedDataInt32x4ArrayCid:
      return kUnboxedInt32x4;
    case kTypedDataFloat64x2ArrayCid:
      return kUnboxedFloat64x2;
    default:
      UNIMPLEMENTED();
      return kTagged;
  }
}

LocationSummary* StoreIndexedInstr::MakeLocationSummary(Zone* zone,
                                                        bool opt) const {
  const intptr_t kNumInputs = 3;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  locs->set_in(0, Location::RequiresRegister());
  // The smi index is either untagged (element size == 1), or it is left smi
  // tagged (for all element sizes > 1).
  if (index_scale() == 1) {
    locs->set_in(1,
                 CanBeImmediateIndex(index(), class_id())
                     ? Location::Constant(index()->definition()->AsConstant())
                     : Location::WritableRegister());
  } else {
    locs->set_in(1,
                 CanBeImmediateIndex(index(), class_id())
                     ? Location::Constant(index()->definition()->AsConstant())
                     : Location::RequiresRegister());
  }
  switch (class_id()) {
    case kArrayCid:
      locs->set_in(2, ShouldEmitStoreBarrier()
                          ? Location::WritableRegister()
                          : Location::RegisterOrConstant(value()));
      break;
    case kExternalTypedDataUint8ArrayCid:
    case kExternalTypedDataUint8ClampedArrayCid:
    case kTypedDataInt8ArrayCid:
    case kTypedDataUint8ArrayCid:
    case kTypedDataUint8ClampedArrayCid:
    case kOneByteStringCid:
      // TODO(fschneider): Add location constraint for byte registers (RAX,
      // RBX, RCX, RDX) instead of using a fixed register.
      locs->set_in(2, Location::FixedRegisterOrSmiConstant(value(), RAX));
      break;
    case kTypedDataInt16ArrayCid:
    case kTypedDataUint16ArrayCid:
      // Writable register because the value must be untagged before storing.
      locs->set_in(2, Location::WritableRegister());
      break;
    case kTypedDataInt32ArrayCid:
    case kTypedDataUint32ArrayCid:
    case kTypedDataInt64ArrayCid:
      locs->set_in(2, Location::RequiresRegister());
      break;
    case kTypedDataFloat32ArrayCid:
    case kTypedDataFloat64ArrayCid:
      // TODO(srdjan): Support Float64 constants.
      locs->set_in(2, Location::RequiresFpuRegister());
      break;
    case kTypedDataInt32x4ArrayCid:
    case kTypedDataFloat64x2ArrayCid:
    case kTypedDataFloat32x4ArrayCid:
      locs->set_in(2, Location::RequiresFpuRegister());
      break;
    default:
      UNREACHABLE();
      return NULL;
  }
  return locs;
}

void StoreIndexedInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  // The array register points to the backing store for external arrays.
  const Register array = locs()->in(0).reg();
  const Location index = locs()->in(1);

  Address element_address =
      index.IsRegister()
          ? Assembler::ElementAddressForRegIndex(
                IsExternal(), class_id(), index_scale(), array, index.reg())
          : Assembler::ElementAddressForIntIndex(
                IsExternal(), class_id(), index_scale(), array,
                Smi::Cast(index.constant()).Value());

  if ((index_scale() == 1) && index.IsRegister()) {
    __ SmiUntag(index.reg());
  }
  switch (class_id()) {
    case kArrayCid:
      if (ShouldEmitStoreBarrier()) {
        Register value = locs()->in(2).reg();
        __ StoreIntoObject(array, element_address, value);
      } else if (locs()->in(2).IsConstant()) {
        const Object& constant = locs()->in(2).constant();
        __ StoreIntoObjectNoBarrier(array, element_address, constant);
      } else {
        Register value = locs()->in(2).reg();
        __ StoreIntoObjectNoBarrier(array, element_address, value);
      }
      break;
    case kTypedDataInt8ArrayCid:
    case kTypedDataUint8ArrayCid:
    case kExternalTypedDataUint8ArrayCid:
    case kOneByteStringCid:
      if (locs()->in(2).IsConstant()) {
        const Smi& constant = Smi::Cast(locs()->in(2).constant());
        __ movb(element_address,
                Immediate(static_cast<int8_t>(constant.Value())));
      } else {
        ASSERT(locs()->in(2).reg() == RAX);
        __ SmiUntag(RAX);
        __ movb(element_address, RAX);
      }
      break;
    case kTypedDataUint8ClampedArrayCid:
    case kExternalTypedDataUint8ClampedArrayCid: {
      if (locs()->in(2).IsConstant()) {
        const Smi& constant = Smi::Cast(locs()->in(2).constant());
        intptr_t value = constant.Value();
        // Clamp to 0x0 or 0xFF respectively.
        if (value > 0xFF) {
          value = 0xFF;
        } else if (value < 0) {
          value = 0;
        }
        __ movb(element_address, Immediate(static_cast<int8_t>(value)));
      } else {
        ASSERT(locs()->in(2).reg() == RAX);
        Label store_value, store_0xff;
        __ SmiUntag(RAX);
        __ CompareImmediate(RAX, Immediate(0xFF));
        __ j(BELOW_EQUAL, &store_value, Assembler::kNearJump);
        // Clamp to 0x0 or 0xFF respectively.
        __ j(GREATER, &store_0xff);
        __ xorq(RAX, RAX);
        __ jmp(&store_value, Assembler::kNearJump);
        __ Bind(&store_0xff);
        __ LoadImmediate(RAX, Immediate(0xFF));
        __ Bind(&store_value);
        __ movb(element_address, RAX);
      }
      break;
    }
    case kTypedDataInt16ArrayCid:
    case kTypedDataUint16ArrayCid: {
      Register value = locs()->in(2).reg();
      __ SmiUntag(value);
      __ movw(element_address, value);
      break;
    }
    case kTypedDataInt32ArrayCid:
    case kTypedDataUint32ArrayCid: {
      Register value = locs()->in(2).reg();
      __ movl(element_address, value);
      break;
    }
    case kTypedDataInt64ArrayCid: {
      Register value = locs()->in(2).reg();
      __ movq(element_address, value);
      break;
    }
    case kTypedDataFloat32ArrayCid:
      __ movss(element_address, locs()->in(2).fpu_reg());
      break;
    case kTypedDataFloat64ArrayCid:
      __ movsd(element_address, locs()->in(2).fpu_reg());
      break;
    case kTypedDataInt32x4ArrayCid:
    case kTypedDataFloat64x2ArrayCid:
    case kTypedDataFloat32x4ArrayCid:
      __ movups(element_address, locs()->in(2).fpu_reg());
      break;
    default:
      UNREACHABLE();
  }
}

LocationSummary* GuardFieldClassInstr::MakeLocationSummary(Zone* zone,
                                                           bool opt) const {
  const intptr_t kNumInputs = 1;

  const intptr_t value_cid = value()->Type()->ToCid();
  const intptr_t field_cid = field().guarded_cid();

  const bool emit_full_guard = !opt || (field_cid == kIllegalCid);
  const bool needs_value_cid_temp_reg =
      (value_cid == kDynamicCid) && (emit_full_guard || (field_cid != kSmiCid));
  const bool needs_field_temp_reg = emit_full_guard;

  intptr_t num_temps = 0;
  if (needs_value_cid_temp_reg) {
    num_temps++;
  }
  if (needs_field_temp_reg) {
    num_temps++;
  }

  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, num_temps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());

  for (intptr_t i = 0; i < num_temps; i++) {
    summary->set_temp(i, Location::RequiresRegister());
  }

  return summary;
}

void GuardFieldClassInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  ASSERT(sizeof(classid_t) == kInt16Size);
  const intptr_t value_cid = value()->Type()->ToCid();
  const intptr_t field_cid = field().guarded_cid();
  const intptr_t nullability = field().is_nullable() ? kNullCid : kIllegalCid;

  if (field_cid == kDynamicCid) {
    if (Compiler::IsBackgroundCompilation()) {
      // Field state changed while compiling.
      Compiler::AbortBackgroundCompilation(
          deopt_id(),
          "GuardFieldClassInstr: field state changed while compiling");
    }
    ASSERT(!compiler->is_optimizing());
    return;  // Nothing to emit.
  }

  const bool emit_full_guard =
      !compiler->is_optimizing() || (field_cid == kIllegalCid);

  const bool needs_value_cid_temp_reg =
      (value_cid == kDynamicCid) && (emit_full_guard || (field_cid != kSmiCid));

  const bool needs_field_temp_reg = emit_full_guard;

  const Register value_reg = locs()->in(0).reg();

  const Register value_cid_reg =
      needs_value_cid_temp_reg ? locs()->temp(0).reg() : kNoRegister;

  const Register field_reg = needs_field_temp_reg
                                 ? locs()->temp(locs()->temp_count() - 1).reg()
                                 : kNoRegister;

  Label ok, fail_label;

  Label* deopt =
      compiler->is_optimizing()
          ? compiler->AddDeoptStub(deopt_id(), ICData::kDeoptGuardField)
          : NULL;

  Label* fail = (deopt != NULL) ? deopt : &fail_label;

  if (emit_full_guard) {
    __ LoadObject(field_reg, Field::ZoneHandle(field().Original()));

    FieldAddress field_cid_operand(field_reg, Field::guarded_cid_offset());
    FieldAddress field_nullability_operand(field_reg,
                                           Field::is_nullable_offset());

    if (value_cid == kDynamicCid) {
      LoadValueCid(compiler, value_cid_reg, value_reg);

      __ cmpw(value_cid_reg, field_cid_operand);
      __ j(EQUAL, &ok);
      __ cmpw(value_cid_reg, field_nullability_operand);
    } else if (value_cid == kNullCid) {
      __ cmpw(field_nullability_operand, Immediate(value_cid));
    } else {
      __ cmpw(field_cid_operand, Immediate(value_cid));
    }
    __ j(EQUAL, &ok);

    // Check if the tracked state of the guarded field can be initialized
    // inline. If the field needs length check we fall through to runtime
    // which is responsible for computing offset of the length field
    // based on the class id.
    if (!field().needs_length_check()) {
      // Uninitialized field can be handled inline. Check if the
      // field is still unitialized.
      __ cmpw(field_cid_operand, Immediate(kIllegalCid));
      __ j(NOT_EQUAL, fail);

      if (value_cid == kDynamicCid) {
        __ movw(field_cid_operand, value_cid_reg);
        __ movw(field_nullability_operand, value_cid_reg);
      } else {
        ASSERT(field_reg != kNoRegister);
        __ movw(field_cid_operand, Immediate(value_cid));
        __ movw(field_nullability_operand, Immediate(value_cid));
      }

      if (deopt == NULL) {
        ASSERT(!compiler->is_optimizing());
        __ jmp(&ok);
      }
    }

    if (deopt == NULL) {
      ASSERT(!compiler->is_optimizing());
      __ Bind(fail);

      __ cmpw(FieldAddress(field_reg, Field::guarded_cid_offset()),
              Immediate(kDynamicCid));
      __ j(EQUAL, &ok);

      __ pushq(field_reg);
      __ pushq(value_reg);
      __ CallRuntime(kUpdateFieldCidRuntimeEntry, 2);
      __ Drop(2);  // Drop the field and the value.
    }
  } else {
    ASSERT(compiler->is_optimizing());
    ASSERT(deopt != NULL);

    // Field guard class has been initialized and is known.
    if (value_cid == kDynamicCid) {
      // Value's class id is not known.
      __ testq(value_reg, Immediate(kSmiTagMask));

      if (field_cid != kSmiCid) {
        __ j(ZERO, fail);
        __ LoadClassId(value_cid_reg, value_reg);
        __ CompareImmediate(value_cid_reg, Immediate(field_cid));
      }

      if (field().is_nullable() && (field_cid != kNullCid)) {
        __ j(EQUAL, &ok);
        __ CompareObject(value_reg, Object::null_object());
      }

      __ j(NOT_EQUAL, fail);
    } else {
      // Both value's and field's class id is known.
      ASSERT((value_cid != field_cid) && (value_cid != nullability));
      __ jmp(fail);
    }
  }
  __ Bind(&ok);
}

LocationSummary* GuardFieldLengthInstr::MakeLocationSummary(Zone* zone,
                                                            bool opt) const {
  const intptr_t kNumInputs = 1;
  if (!opt || (field().guarded_list_length() == Field::kUnknownFixedLength)) {
    const intptr_t kNumTemps = 3;
    LocationSummary* summary = new (zone)
        LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
    summary->set_in(0, Location::RequiresRegister());
    // We need temporaries for field object, length offset and expected length.
    summary->set_temp(0, Location::RequiresRegister());
    summary->set_temp(1, Location::RequiresRegister());
    summary->set_temp(2, Location::RequiresRegister());
    return summary;
  } else {
    LocationSummary* summary = new (zone)
        LocationSummary(zone, kNumInputs, 0, LocationSummary::kNoCall);
    summary->set_in(0, Location::RequiresRegister());
    return summary;
  }
  UNREACHABLE();
}

void GuardFieldLengthInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  if (field().guarded_list_length() == Field::kNoFixedLength) {
    if (Compiler::IsBackgroundCompilation()) {
      // Field state changed while compiling.
      Compiler::AbortBackgroundCompilation(
          deopt_id(),
          "GuardFieldLengthInstr: field state changed while compiling");
    }
    ASSERT(!compiler->is_optimizing());
    return;  // Nothing to emit.
  }

  Label* deopt =
      compiler->is_optimizing()
          ? compiler->AddDeoptStub(deopt_id(), ICData::kDeoptGuardField)
          : NULL;

  const Register value_reg = locs()->in(0).reg();

  if (!compiler->is_optimizing() ||
      (field().guarded_list_length() == Field::kUnknownFixedLength)) {
    const Register field_reg = locs()->temp(0).reg();
    const Register offset_reg = locs()->temp(1).reg();
    const Register length_reg = locs()->temp(2).reg();

    Label ok;

    __ LoadObject(field_reg, Field::ZoneHandle(field().Original()));

    __ movsxb(
        offset_reg,
        FieldAddress(field_reg,
                     Field::guarded_list_length_in_object_offset_offset()));
    __ movq(length_reg,
            FieldAddress(field_reg, Field::guarded_list_length_offset()));

    __ cmpq(offset_reg, Immediate(0));
    __ j(NEGATIVE, &ok);

    // Load the length from the value. GuardFieldClass already verified that
    // value's class matches guarded class id of the field.
    // offset_reg contains offset already corrected by -kHeapObjectTag that is
    // why we use Address instead of FieldAddress.
    __ cmpq(length_reg, Address(value_reg, offset_reg, TIMES_1, 0));

    if (deopt == NULL) {
      __ j(EQUAL, &ok);

      __ pushq(field_reg);
      __ pushq(value_reg);
      __ CallRuntime(kUpdateFieldCidRuntimeEntry, 2);
      __ Drop(2);  // Drop the field and the value.
    } else {
      __ j(NOT_EQUAL, deopt);
    }

    __ Bind(&ok);
  } else {
    ASSERT(compiler->is_optimizing());
    ASSERT(field().guarded_list_length() >= 0);
    ASSERT(field().guarded_list_length_in_object_offset() !=
           Field::kUnknownLengthOffset);

    __ CompareImmediate(
        FieldAddress(value_reg, field().guarded_list_length_in_object_offset()),
        Immediate(Smi::RawValue(field().guarded_list_length())));
    __ j(NOT_EQUAL, deopt);
  }
}

class BoxAllocationSlowPath : public SlowPathCode {
 public:
  BoxAllocationSlowPath(Instruction* instruction,
                        const Class& cls,
                        Register result)
      : instruction_(instruction), cls_(cls), result_(result) {}

  virtual void EmitNativeCode(FlowGraphCompiler* compiler) {
    if (Assembler::EmittingComments()) {
      __ Comment("%s slow path allocation of %s", instruction_->DebugName(),
                 String::Handle(cls_.ScrubbedName()).ToCString());
    }
    __ Bind(entry_label());
    const Code& stub = Code::ZoneHandle(
        compiler->zone(), StubCode::GetAllocationStubForClass(cls_));
    const StubEntry stub_entry(stub);

    LocationSummary* locs = instruction_->locs();

    locs->live_registers()->Remove(Location::RegisterLocation(result_));

    compiler->SaveLiveRegisters(locs);
    compiler->GenerateCall(TokenPosition::kNoSource,  // No token position.
                           stub_entry, RawPcDescriptors::kOther, locs);
    compiler->AddStubCallTarget(stub);
    __ MoveRegister(result_, RAX);
    compiler->RestoreLiveRegisters(locs);
    __ jmp(exit_label());
  }

  static void Allocate(FlowGraphCompiler* compiler,
                       Instruction* instruction,
                       const Class& cls,
                       Register result,
                       Register temp) {
    if (compiler->intrinsic_mode()) {
      __ TryAllocate(cls, compiler->intrinsic_slow_path_label(),
                     Assembler::kFarJump, result, temp);
    } else {
      BoxAllocationSlowPath* slow_path =
          new BoxAllocationSlowPath(instruction, cls, result);
      compiler->AddSlowPathCode(slow_path);

      __ TryAllocate(cls, slow_path->entry_label(), Assembler::kFarJump, result,
                     temp);
      __ Bind(slow_path->exit_label());
    }
  }

 private:
  Instruction* instruction_;
  const Class& cls_;
  const Register result_;
};

LocationSummary* StoreInstanceFieldInstr::MakeLocationSummary(Zone* zone,
                                                              bool opt) const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps =
      (IsUnboxedStore() && opt) ? 2 : ((IsPotentialUnboxedStore()) ? 3 : 0);
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps,
                      ((IsUnboxedStore() && opt && is_initialization()) ||
                       IsPotentialUnboxedStore())
                          ? LocationSummary::kCallOnSlowPath
                          : LocationSummary::kNoCall);

  summary->set_in(0, Location::RequiresRegister());
  if (IsUnboxedStore() && opt) {
    summary->set_in(1, Location::RequiresFpuRegister());
    summary->set_temp(0, Location::RequiresRegister());
    summary->set_temp(1, Location::RequiresRegister());
  } else if (IsPotentialUnboxedStore()) {
    summary->set_in(1, ShouldEmitStoreBarrier() ? Location::WritableRegister()
                                                : Location::RequiresRegister());
    summary->set_temp(0, Location::RequiresRegister());
    summary->set_temp(1, Location::RequiresRegister());
    summary->set_temp(2, opt ? Location::RequiresFpuRegister()
                             : Location::FpuRegisterLocation(XMM1));
  } else {
    summary->set_in(1, ShouldEmitStoreBarrier()
                           ? Location::WritableRegister()
                           : Location::RegisterOrConstant(value()));
  }
  return summary;
}

static void EnsureMutableBox(FlowGraphCompiler* compiler,
                             StoreInstanceFieldInstr* instruction,
                             Register box_reg,
                             const Class& cls,
                             Register instance_reg,
                             intptr_t offset,
                             Register temp) {
  Label done;
  __ movq(box_reg, FieldAddress(instance_reg, offset));
  __ CompareObject(box_reg, Object::null_object());
  __ j(NOT_EQUAL, &done);
  BoxAllocationSlowPath::Allocate(compiler, instruction, cls, box_reg, temp);
  __ movq(temp, box_reg);
  __ StoreIntoObject(instance_reg, FieldAddress(instance_reg, offset), temp);

  __ Bind(&done);
}

void StoreInstanceFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  ASSERT(sizeof(classid_t) == kInt16Size);
  Label skip_store;

  Register instance_reg = locs()->in(0).reg();

  if (IsUnboxedStore() && compiler->is_optimizing()) {
    XmmRegister value = locs()->in(1).fpu_reg();
    Register temp = locs()->temp(0).reg();
    Register temp2 = locs()->temp(1).reg();
    const intptr_t cid = field().UnboxedFieldCid();

    if (is_initialization()) {
      const Class* cls = NULL;
      switch (cid) {
        case kDoubleCid:
          cls = &compiler->double_class();
          break;
        case kFloat32x4Cid:
          cls = &compiler->float32x4_class();
          break;
        case kFloat64x2Cid:
          cls = &compiler->float64x2_class();
          break;
        default:
          UNREACHABLE();
      }

      BoxAllocationSlowPath::Allocate(compiler, this, *cls, temp, temp2);
      __ movq(temp2, temp);
      __ StoreIntoObject(instance_reg,
                         FieldAddress(instance_reg, offset_in_bytes_), temp2);
    } else {
      __ movq(temp, FieldAddress(instance_reg, offset_in_bytes_));
    }
    switch (cid) {
      case kDoubleCid:
        __ Comment("UnboxedDoubleStoreInstanceFieldInstr");
        __ movsd(FieldAddress(temp, Double::value_offset()), value);
        break;
      case kFloat32x4Cid:
        __ Comment("UnboxedFloat32x4StoreInstanceFieldInstr");
        __ movups(FieldAddress(temp, Float32x4::value_offset()), value);
        break;
      case kFloat64x2Cid:
        __ Comment("UnboxedFloat64x2StoreInstanceFieldInstr");
        __ movups(FieldAddress(temp, Float64x2::value_offset()), value);
        break;
      default:
        UNREACHABLE();
    }
    return;
  }

  if (IsPotentialUnboxedStore()) {
    Register value_reg = locs()->in(1).reg();
    Register temp = locs()->temp(0).reg();
    Register temp2 = locs()->temp(1).reg();
    FpuRegister fpu_temp = locs()->temp(2).fpu_reg();

    if (ShouldEmitStoreBarrier()) {
      // Value input is a writable register and should be manually preserved
      // across allocation slow-path.
      locs()->live_registers()->Add(locs()->in(1), kTagged);
    }

    Label store_pointer;
    Label store_double;
    Label store_float32x4;
    Label store_float64x2;

    __ LoadObject(temp, Field::ZoneHandle(Z, field().Original()));

    __ cmpw(FieldAddress(temp, Field::is_nullable_offset()),
            Immediate(kNullCid));
    __ j(EQUAL, &store_pointer);

    __ movzxb(temp2, FieldAddress(temp, Field::kind_bits_offset()));
    __ testq(temp2, Immediate(1 << Field::kUnboxingCandidateBit));
    __ j(ZERO, &store_pointer);

    __ cmpw(FieldAddress(temp, Field::guarded_cid_offset()),
            Immediate(kDoubleCid));
    __ j(EQUAL, &store_double);

    __ cmpw(FieldAddress(temp, Field::guarded_cid_offset()),
            Immediate(kFloat32x4Cid));
    __ j(EQUAL, &store_float32x4);

    __ cmpw(FieldAddress(temp, Field::guarded_cid_offset()),
            Immediate(kFloat64x2Cid));
    __ j(EQUAL, &store_float64x2);

    // Fall through.
    __ jmp(&store_pointer);

    if (!compiler->is_optimizing()) {
      locs()->live_registers()->Add(locs()->in(0));
      locs()->live_registers()->Add(locs()->in(1));
    }

    {
      __ Bind(&store_double);
      EnsureMutableBox(compiler, this, temp, compiler->double_class(),
                       instance_reg, offset_in_bytes_, temp2);
      __ movsd(fpu_temp, FieldAddress(value_reg, Double::value_offset()));
      __ movsd(FieldAddress(temp, Double::value_offset()), fpu_temp);
      __ jmp(&skip_store);
    }

    {
      __ Bind(&store_float32x4);
      EnsureMutableBox(compiler, this, temp, compiler->float32x4_class(),
                       instance_reg, offset_in_bytes_, temp2);
      __ movups(fpu_temp, FieldAddress(value_reg, Float32x4::value_offset()));
      __ movups(FieldAddress(temp, Float32x4::value_offset()), fpu_temp);
      __ jmp(&skip_store);
    }

    {
      __ Bind(&store_float64x2);
      EnsureMutableBox(compiler, this, temp, compiler->float64x2_class(),
                       instance_reg, offset_in_bytes_, temp2);
      __ movups(fpu_temp, FieldAddress(value_reg, Float64x2::value_offset()));
      __ movups(FieldAddress(temp, Float64x2::value_offset()), fpu_temp);
      __ jmp(&skip_store);
    }

    __ Bind(&store_pointer);
  }

  if (ShouldEmitStoreBarrier()) {
    Register value_reg = locs()->in(1).reg();
    __ StoreIntoObject(instance_reg,
                       FieldAddress(instance_reg, offset_in_bytes_), value_reg,
                       CanValueBeSmi());
  } else {
    if (locs()->in(1).IsConstant()) {
      __ StoreIntoObjectNoBarrier(instance_reg,
                                  FieldAddress(instance_reg, offset_in_bytes_),
                                  locs()->in(1).constant());
    } else {
      Register value_reg = locs()->in(1).reg();
      __ StoreIntoObjectNoBarrier(instance_reg,
                                  FieldAddress(instance_reg, offset_in_bytes_),
                                  value_reg);
    }
  }
  __ Bind(&skip_store);
}

LocationSummary* LoadStaticFieldInstr::MakeLocationSummary(Zone* zone,
                                                           bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  summary->set_out(0, Location::RequiresRegister());
  return summary;
}

// When the parser is building an implicit static getter for optimization,
// it can generate a function body where deoptimization ids do not line up
// with the unoptimized code.
//
// This is safe only so long as LoadStaticFieldInstr cannot deoptimize.
void LoadStaticFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register field = locs()->in(0).reg();
  Register result = locs()->out(0).reg();
  __ movq(result, FieldAddress(field, Field::static_value_offset()));
}

LocationSummary* StoreStaticFieldInstr::MakeLocationSummary(Zone* zone,
                                                            bool opt) const {
  LocationSummary* locs =
      new (zone) LocationSummary(zone, 1, 1, LocationSummary::kNoCall);
  locs->set_in(0, value()->NeedsStoreBuffer() ? Location::WritableRegister()
                                              : Location::RequiresRegister());
  locs->set_temp(0, Location::RequiresRegister());
  return locs;
}

void StoreStaticFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register value = locs()->in(0).reg();
  Register temp = locs()->temp(0).reg();

  __ LoadObject(temp, Field::ZoneHandle(Z, field().Original()));
  if (this->value()->NeedsStoreBuffer()) {
    __ StoreIntoObject(temp, FieldAddress(temp, Field::static_value_offset()),
                       value, CanValueBeSmi());
  } else {
    __ StoreIntoObjectNoBarrier(
        temp, FieldAddress(temp, Field::static_value_offset()), value);
  }
}

LocationSummary* InstanceOfInstr::MakeLocationSummary(Zone* zone,
                                                      bool opt) const {
  const intptr_t kNumInputs = 3;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kCall);
  summary->set_in(0, Location::RegisterLocation(RAX));  // Instance.
  summary->set_in(1, Location::RegisterLocation(RDX));  // Instant. type args.
  summary->set_in(2, Location::RegisterLocation(RCX));  // Function type args.
  summary->set_out(0, Location::RegisterLocation(RAX));
  return summary;
}

void InstanceOfInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  ASSERT(locs()->in(0).reg() == RAX);  // Value.
  ASSERT(locs()->in(1).reg() == RDX);  // Instantiator type arguments.
  ASSERT(locs()->in(2).reg() == RCX);  // Function type arguments.

  compiler->GenerateInstanceOf(token_pos(), deopt_id(), type(), locs());
  ASSERT(locs()->out(0).reg() == RAX);
}

// TODO(srdjan): In case of constant inputs make CreateArray kNoCall and
// use slow path stub.
LocationSummary* CreateArrayInstr::MakeLocationSummary(Zone* zone,
                                                       bool opt) const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kCall);
  locs->set_in(0, Location::RegisterLocation(RBX));
  locs->set_in(1, Location::RegisterLocation(R10));
  locs->set_out(0, Location::RegisterLocation(RAX));
  return locs;
}

// Inlines array allocation for known constant values.
static void InlineArrayAllocation(FlowGraphCompiler* compiler,
                                  intptr_t num_elements,
                                  Label* slow_path,
                                  Label* done) {
  const int kInlineArraySize = 12;  // Same as kInlineInstanceSize.
  const Register kLengthReg = R10;
  const Register kElemTypeReg = RBX;
  const intptr_t instance_size = Array::InstanceSize(num_elements);

  __ TryAllocateArray(kArrayCid, instance_size, slow_path, Assembler::kFarJump,
                      RAX,   // instance
                      RCX,   // end address
                      R13);  // temp

  // RAX: new object start as a tagged pointer.
  // Store the type argument field.
  __ StoreIntoObjectNoBarrier(
      RAX, FieldAddress(RAX, Array::type_arguments_offset()), kElemTypeReg);

  // Set the length field.
  __ StoreIntoObjectNoBarrier(RAX, FieldAddress(RAX, Array::length_offset()),
                              kLengthReg);

  // Initialize all array elements to raw_null.
  // RAX: new object start as a tagged pointer.
  // RCX: new object end address.
  // RDI: iterator which initially points to the start of the variable
  // data area to be initialized.
  if (num_elements > 0) {
    const intptr_t array_size = instance_size - sizeof(RawArray);
    __ LoadObject(R12, Object::null_object());
    __ leaq(RDI, FieldAddress(RAX, sizeof(RawArray)));
    if (array_size < (kInlineArraySize * kWordSize)) {
      intptr_t current_offset = 0;
      while (current_offset < array_size) {
        __ StoreIntoObjectNoBarrier(RAX, Address(RDI, current_offset), R12);
        current_offset += kWordSize;
      }
    } else {
      Label init_loop;
      __ Bind(&init_loop);
      __ StoreIntoObjectNoBarrier(RAX, Address(RDI, 0), R12);
      __ addq(RDI, Immediate(kWordSize));
      __ cmpq(RDI, RCX);
      __ j(BELOW, &init_loop, Assembler::kNearJump);
    }
  }
  __ jmp(done, Assembler::kNearJump);
}

void CreateArrayInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  // Allocate the array.  R10 = length, RBX = element type.
  const Register kLengthReg = R10;
  const Register kElemTypeReg = RBX;
  const Register kResultReg = RAX;
  ASSERT(locs()->in(0).reg() == kElemTypeReg);
  ASSERT(locs()->in(1).reg() == kLengthReg);

  Label slow_path, done;
  if (compiler->is_optimizing() && !FLAG_precompiled_mode &&
      num_elements()->BindsToConstant() &&
      num_elements()->BoundConstant().IsSmi()) {
    const intptr_t length = Smi::Cast(num_elements()->BoundConstant()).Value();
    if ((length >= 0) && (length <= Array::kMaxElements)) {
      Label slow_path, done;
      InlineArrayAllocation(compiler, length, &slow_path, &done);
      __ Bind(&slow_path);
      __ PushObject(Object::null_object());  // Make room for the result.
      __ pushq(kLengthReg);
      __ pushq(kElemTypeReg);
      compiler->GenerateRuntimeCall(token_pos(), deopt_id(),
                                    kAllocateArrayRuntimeEntry, 2, locs());
      __ Drop(2);
      __ popq(kResultReg);
      __ Bind(&done);
      return;
    }
  }

  __ Bind(&slow_path);
  const Code& stub = Code::ZoneHandle(compiler->zone(),
                                      StubCode::AllocateArray_entry()->code());
  compiler->AddStubCallTarget(stub);
  compiler->GenerateCallWithDeopt(token_pos(), deopt_id(),
                                  *StubCode::AllocateArray_entry(),
                                  RawPcDescriptors::kOther, locs());
  __ Bind(&done);
  ASSERT(locs()->out(0).reg() == kResultReg);
}

LocationSummary* LoadFieldInstr::MakeLocationSummary(Zone* zone,
                                                     bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps =
      (IsUnboxedLoad() && opt) ? 1 : ((IsPotentialUnboxedLoad()) ? 2 : 0);
  LocationSummary* locs = new (zone) LocationSummary(
      zone, kNumInputs, kNumTemps,
      (opt && !IsPotentialUnboxedLoad()) ? LocationSummary::kNoCall
                                         : LocationSummary::kCallOnSlowPath);

  locs->set_in(0, Location::RequiresRegister());

  if (IsUnboxedLoad() && opt) {
    locs->set_temp(0, Location::RequiresRegister());
  } else if (IsPotentialUnboxedLoad()) {
    locs->set_temp(0, opt ? Location::RequiresFpuRegister()
                          : Location::FpuRegisterLocation(XMM1));
    locs->set_temp(1, Location::RequiresRegister());
  }
  locs->set_out(0, Location::RequiresRegister());
  return locs;
}

void LoadFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  ASSERT(sizeof(classid_t) == kInt16Size);
  Register instance_reg = locs()->in(0).reg();
  if (IsUnboxedLoad() && compiler->is_optimizing()) {
    XmmRegister result = locs()->out(0).fpu_reg();
    Register temp = locs()->temp(0).reg();
    __ movq(temp, FieldAddress(instance_reg, offset_in_bytes()));
    intptr_t cid = field()->UnboxedFieldCid();
    switch (cid) {
      case kDoubleCid:
        __ Comment("UnboxedDoubleLoadFieldInstr");
        __ movsd(result, FieldAddress(temp, Double::value_offset()));
        break;
      case kFloat32x4Cid:
        __ Comment("UnboxedFloat32x4LoadFieldInstr");
        __ movups(result, FieldAddress(temp, Float32x4::value_offset()));
        break;
      case kFloat64x2Cid:
        __ Comment("UnboxedFloat64x2LoadFieldInstr");
        __ movups(result, FieldAddress(temp, Float64x2::value_offset()));
        break;
      default:
        UNREACHABLE();
    }
    return;
  }

  Label done;
  Register result = locs()->out(0).reg();
  if (IsPotentialUnboxedLoad()) {
    Register temp = locs()->temp(1).reg();
    XmmRegister value = locs()->temp(0).fpu_reg();

    Label load_pointer;
    Label load_double;
    Label load_float32x4;
    Label load_float64x2;

    __ LoadObject(result, Field::ZoneHandle(field()->Original()));

    FieldAddress field_cid_operand(result, Field::guarded_cid_offset());
    FieldAddress field_nullability_operand(result, Field::is_nullable_offset());

    __ cmpw(field_nullability_operand, Immediate(kNullCid));
    __ j(EQUAL, &load_pointer);

    __ cmpw(field_cid_operand, Immediate(kDoubleCid));
    __ j(EQUAL, &load_double);

    __ cmpw(field_cid_operand, Immediate(kFloat32x4Cid));
    __ j(EQUAL, &load_float32x4);

    __ cmpw(field_cid_operand, Immediate(kFloat64x2Cid));
    __ j(EQUAL, &load_float64x2);

    // Fall through.
    __ jmp(&load_pointer);

    if (!compiler->is_optimizing()) {
      locs()->live_registers()->Add(locs()->in(0));
    }

    {
      __ Bind(&load_double);
      BoxAllocationSlowPath::Allocate(compiler, this, compiler->double_class(),
                                      result, temp);
      __ movq(temp, FieldAddress(instance_reg, offset_in_bytes()));
      __ movsd(value, FieldAddress(temp, Double::value_offset()));
      __ movsd(FieldAddress(result, Double::value_offset()), value);
      __ jmp(&done);
    }

    {
      __ Bind(&load_float32x4);
      BoxAllocationSlowPath::Allocate(
          compiler, this, compiler->float32x4_class(), result, temp);
      __ movq(temp, FieldAddress(instance_reg, offset_in_bytes()));
      __ movups(value, FieldAddress(temp, Float32x4::value_offset()));
      __ movups(FieldAddress(result, Float32x4::value_offset()), value);
      __ jmp(&done);
    }

    {
      __ Bind(&load_float64x2);
      BoxAllocationSlowPath::Allocate(
          compiler, this, compiler->float64x2_class(), result, temp);
      __ movq(temp, FieldAddress(instance_reg, offset_in_bytes()));
      __ movups(value, FieldAddress(temp, Float64x2::value_offset()));
      __ movups(FieldAddress(result, Float64x2::value_offset()), value);
      __ jmp(&done);
    }

    __ Bind(&load_pointer);
  }
  __ movq(result, FieldAddress(instance_reg, offset_in_bytes()));
  __ Bind(&done);
}

LocationSummary* InstantiateTypeInstr::MakeLocationSummary(Zone* zone,
                                                           bool opt) const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kCall);
  locs->set_in(0, Location::RegisterLocation(RAX));  // Instant. type args.
  locs->set_in(1, Location::RegisterLocation(RDX));  // Function type args.
  locs->set_out(0, Location::RegisterLocation(RAX));
  return locs;
}

void InstantiateTypeInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register instantiator_type_args_reg = locs()->in(0).reg();
  Register function_type_args_reg = locs()->in(1).reg();
  Register result_reg = locs()->out(0).reg();

  // 'instantiator_type_args_reg' is a TypeArguments object (or null).
  // 'function_type_args_reg' is a TypeArguments object (or null).
  // A runtime call to instantiate the type is required.
  __ PushObject(Object::null_object());  // Make room for the result.
  __ PushObject(type());
  __ pushq(instantiator_type_args_reg);  // Push instantiator type arguments.
  __ pushq(function_type_args_reg);      // Push function type arguments.
  compiler->GenerateRuntimeCall(token_pos(), deopt_id(),
                                kInstantiateTypeRuntimeEntry, 3, locs());
  __ Drop(3);           // Drop 2 type argument vectors and uninstantiated type.
  __ popq(result_reg);  // Pop instantiated type.
}

LocationSummary* InstantiateTypeArgumentsInstr::MakeLocationSummary(
    Zone* zone,
    bool opt) const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kCall);
  locs->set_in(0, Location::RegisterLocation(RAX));  // Instant. type args.
  locs->set_in(1, Location::RegisterLocation(RCX));  // Function type args.
  locs->set_out(0, Location::RegisterLocation(RAX));
  return locs;
}

void InstantiateTypeArgumentsInstr::EmitNativeCode(
    FlowGraphCompiler* compiler) {
  Register instantiator_type_args_reg = locs()->in(0).reg();
  Register function_type_args_reg = locs()->in(1).reg();
  Register result_reg = locs()->out(0).reg();
  ASSERT(instantiator_type_args_reg == RAX);
  ASSERT(instantiator_type_args_reg == result_reg);

  // 'instantiator_type_args_reg' is a TypeArguments object (or null).
  // 'function_type_args_reg' is a TypeArguments object (or null).
  ASSERT(!type_arguments().IsUninstantiatedIdentity() &&
         !type_arguments().CanShareInstantiatorTypeArguments(
             instantiator_class()));
  // If both the instantiator and function type arguments are null and if the
  // type argument vector instantiated from null becomes a vector of dynamic,
  // then use null as the type arguments.
  Label type_arguments_instantiated;
  const intptr_t len = type_arguments().Length();
  if (type_arguments().IsRawWhenInstantiatedFromRaw(len)) {
    Label non_null_type_args;
    __ CompareObject(instantiator_type_args_reg, Object::null_object());
    __ j(NOT_EQUAL, &non_null_type_args, Assembler::kNearJump);
    __ CompareObject(function_type_args_reg, Object::null_object());
    __ j(EQUAL, &type_arguments_instantiated, Assembler::kNearJump);
    __ Bind(&non_null_type_args);
  }

  // Lookup cache before calling runtime.
  // TODO(regis): Consider moving this into a shared stub to reduce
  // generated code size.
  __ LoadObject(RDI, type_arguments());
  __ movq(RDI, FieldAddress(RDI, TypeArguments::instantiations_offset()));
  __ leaq(RDI, FieldAddress(RDI, Array::data_offset()));
  // The instantiations cache is initialized with Object::zero_array() and is
  // therefore guaranteed to contain kNoInstantiator. No length check needed.
  Label loop, next, found, slow_case;
  __ Bind(&loop);
  __ movq(RDX, Address(RDI, 0 * kWordSize));  // Cached instantiator type args.
  __ cmpq(RDX, instantiator_type_args_reg);
  __ j(NOT_EQUAL, &next, Assembler::kNearJump);
  __ movq(R10, Address(RDI, 1 * kWordSize));  // Cached function type args.
  __ cmpq(R10, function_type_args_reg);
  __ j(EQUAL, &found, Assembler::kNearJump);
  __ Bind(&next);
  __ addq(RDI, Immediate(StubCode::kInstantiationSizeInWords * kWordSize));
  __ cmpq(RDX, Immediate(Smi::RawValue(StubCode::kNoInstantiator)));
  __ j(NOT_EQUAL, &loop, Assembler::kNearJump);
  __ jmp(&slow_case, Assembler::kNearJump);
  __ Bind(&found);
  __ movq(result_reg, Address(RDI, 2 * kWordSize));  // Cached instantiated ta.
  __ jmp(&type_arguments_instantiated, Assembler::kNearJump);

  __ Bind(&slow_case);
  // Instantiate non-null type arguments.
  // A runtime call to instantiate the type arguments is required.
  __ PushObject(Object::null_object());  // Make room for the result.
  __ PushObject(type_arguments());
  __ pushq(instantiator_type_args_reg);  // Push instantiator type arguments.
  __ pushq(function_type_args_reg);      // Push function type arguments.
  compiler->GenerateRuntimeCall(token_pos(), deopt_id(),
                                kInstantiateTypeArgumentsRuntimeEntry, 3,
                                locs());
  __ Drop(3);           // Drop 2 type argument vectors and uninstantiated args.
  __ popq(result_reg);  // Pop instantiated type arguments.
  __ Bind(&type_arguments_instantiated);
}

LocationSummary* AllocateUninitializedContextInstr::MakeLocationSummary(
    Zone* zone,
    bool opt) const {
  ASSERT(opt);
  const intptr_t kNumInputs = 0;
  const intptr_t kNumTemps = 2;
  LocationSummary* locs = new (zone) LocationSummary(
      zone, kNumInputs, kNumTemps, LocationSummary::kCallOnSlowPath);
  locs->set_temp(0, Location::RegisterLocation(R10));
  locs->set_temp(1, Location::RegisterLocation(R13));
  locs->set_out(0, Location::RegisterLocation(RAX));
  return locs;
}

class AllocateContextSlowPath : public SlowPathCode {
 public:
  explicit AllocateContextSlowPath(
      AllocateUninitializedContextInstr* instruction)
      : instruction_(instruction) {}

  virtual void EmitNativeCode(FlowGraphCompiler* compiler) {
    __ Comment("AllocateContextSlowPath");
    __ Bind(entry_label());

    LocationSummary* locs = instruction_->locs();
    locs->live_registers()->Remove(locs->out(0));

    compiler->SaveLiveRegisters(locs);

    __ LoadImmediate(R10, Immediate(instruction_->num_context_variables()));
    const Code& stub = Code::ZoneHandle(
        compiler->zone(), StubCode::AllocateContext_entry()->code());
    compiler->AddStubCallTarget(stub);
    compiler->GenerateCall(instruction_->token_pos(),
                           *StubCode::AllocateContext_entry(),
                           RawPcDescriptors::kOther, locs);
    ASSERT(instruction_->locs()->out(0).reg() == RAX);
    compiler->RestoreLiveRegisters(instruction_->locs());
    __ jmp(exit_label());
  }

 private:
  AllocateUninitializedContextInstr* instruction_;
};

void AllocateUninitializedContextInstr::EmitNativeCode(
    FlowGraphCompiler* compiler) {
  ASSERT(compiler->is_optimizing());
  Register temp = locs()->temp(0).reg();
  Register result = locs()->out(0).reg();
  // Try allocate the object.
  AllocateContextSlowPath* slow_path = new AllocateContextSlowPath(this);
  compiler->AddSlowPathCode(slow_path);
  intptr_t instance_size = Context::InstanceSize(num_context_variables());

  __ TryAllocateArray(kContextCid, instance_size, slow_path->entry_label(),
                      Assembler::kFarJump,
                      result,  // instance
                      temp,    // end address
                      locs()->temp(1).reg());

  // Setup up number of context variables field.
  __ movq(FieldAddress(result, Context::num_variables_offset()),
          Immediate(num_context_variables()));

  __ Bind(slow_path->exit_label());
}

LocationSummary* AllocateContextInstr::MakeLocationSummary(Zone* zone,
                                                           bool opt) const {
  const intptr_t kNumInputs = 0;
  const intptr_t kNumTemps = 1;
  LocationSummary* locs = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kCall);
  locs->set_temp(0, Location::RegisterLocation(R10));
  locs->set_out(0, Location::RegisterLocation(RAX));
  return locs;
}

void AllocateContextInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  ASSERT(locs()->temp(0).reg() == R10);
  ASSERT(locs()->out(0).reg() == RAX);

  __ LoadImmediate(R10, Immediate(num_context_variables()));
  compiler->GenerateCall(token_pos(), *StubCode::AllocateContext_entry(),
                         RawPcDescriptors::kOther, locs());
}

LocationSummary* InitStaticFieldInstr::MakeLocationSummary(Zone* zone,
                                                           bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 1;
  LocationSummary* locs = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kCall);
  locs->set_in(0, Location::RegisterLocation(RAX));
  locs->set_temp(0, Location::RegisterLocation(RCX));
  return locs;
}

void InitStaticFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register field = locs()->in(0).reg();
  Register temp = locs()->temp(0).reg();

  Label call_runtime, no_call;

  __ movq(temp, FieldAddress(field, Field::static_value_offset()));
  __ CompareObject(temp, Object::sentinel());
  __ j(EQUAL, &call_runtime);

  __ CompareObject(temp, Object::transition_sentinel());
  __ j(NOT_EQUAL, &no_call);

  __ Bind(&call_runtime);
  __ PushObject(Object::null_object());  // Make room for (unused) result.
  __ pushq(field);
  compiler->GenerateRuntimeCall(token_pos(), deopt_id(),
                                kInitStaticFieldRuntimeEntry, 1, locs());
  __ Drop(2);  // Remove argument and unused result.
  __ Bind(&no_call);
}

LocationSummary* CloneContextInstr::MakeLocationSummary(Zone* zone,
                                                        bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kCall);
  locs->set_in(0, Location::RegisterLocation(RAX));
  locs->set_out(0, Location::RegisterLocation(RAX));
  return locs;
}

void CloneContextInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register context_value = locs()->in(0).reg();
  Register result = locs()->out(0).reg();

  __ PushObject(Object::null_object());  // Make room for the result.
  __ pushq(context_value);
  compiler->GenerateRuntimeCall(token_pos(), deopt_id(),
                                kCloneContextRuntimeEntry, 1, locs());
  __ popq(result);  // Remove argument.
  __ popq(result);  // Get result (cloned context).
}

LocationSummary* CatchBlockEntryInstr::MakeLocationSummary(Zone* zone,
                                                           bool opt) const {
  UNREACHABLE();
  return NULL;
}

void CatchBlockEntryInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  __ Bind(compiler->GetJumpLabel(this));
  compiler->AddExceptionHandler(catch_try_index(), try_index(),
                                compiler->assembler()->CodeSize(),
                                handler_token_pos(), is_generated(),
                                catch_handler_types_, needs_stacktrace());
  // On lazy deoptimization we patch the optimized code here to enter the
  // deoptimization stub.
  const intptr_t deopt_id = Thread::ToDeoptAfter(GetDeoptId());
  if (compiler->is_optimizing()) {
    compiler->AddDeoptIndexAtCall(deopt_id);
  } else {
    compiler->AddCurrentDescriptor(RawPcDescriptors::kDeopt, deopt_id,
                                   TokenPosition::kNoSource);
  }
  if (HasParallelMove()) {
    compiler->parallel_move_resolver()->EmitNativeCode(parallel_move());
  }

  // Restore RSP from RBP as we are coming from a throw and the code for
  // popping arguments has not been run.
  const intptr_t fp_sp_dist =
      (kFirstLocalSlotFromFp + 1 - compiler->StackSize()) * kWordSize;
  ASSERT(fp_sp_dist <= 0);
  __ leaq(RSP, Address(RBP, fp_sp_dist));

  // Auxiliary variables introduced by the try catch can be captured if we are
  // inside a function with yield/resume points. In this case we first need
  // to restore the context to match the context at entry into the closure.
  if (should_restore_closure_context()) {
    const ParsedFunction& parsed_function = compiler->parsed_function();
    ASSERT(parsed_function.function().IsClosureFunction());
    LocalScope* scope = parsed_function.node_sequence()->scope();

    LocalVariable* closure_parameter = scope->VariableAt(0);
    ASSERT(!closure_parameter->is_captured());
    __ movq(CTX, Address(RBP, closure_parameter->index() * kWordSize));
    __ movq(CTX, FieldAddress(CTX, Closure::context_offset()));

#ifdef DEBUG
    Label ok;
    __ LoadClassId(RBX, CTX);
    __ cmpq(RBX, Immediate(kContextCid));
    __ j(EQUAL, &ok, Assembler::kNearJump);
    __ Stop("Incorrect context at entry");
    __ Bind(&ok);
#endif

    const intptr_t context_index =
        parsed_function.current_context_var()->index();
    __ movq(Address(RBP, context_index * kWordSize), CTX);
  }

  // Initialize exception and stack trace variables.
  if (exception_var().is_captured()) {
    ASSERT(stacktrace_var().is_captured());
    __ StoreIntoObject(
        CTX,
        FieldAddress(CTX, Context::variable_offset(exception_var().index())),
        kExceptionObjectReg);
    __ StoreIntoObject(
        CTX,
        FieldAddress(CTX, Context::variable_offset(stacktrace_var().index())),
        kStackTraceObjectReg);
  } else {
    __ movq(Address(RBP, exception_var().index() * kWordSize),
            kExceptionObjectReg);
    __ movq(Address(RBP, stacktrace_var().index() * kWordSize),
            kStackTraceObjectReg);
  }
}

LocationSummary* CheckStackOverflowInstr::MakeLocationSummary(Zone* zone,
                                                              bool opt) const {
  const intptr_t kNumInputs = 0;
  const intptr_t kNumTemps = 1;
  LocationSummary* summary = new (zone) LocationSummary(
      zone, kNumInputs, kNumTemps, LocationSummary::kCallOnSlowPath);
  summary->set_temp(0, Location::RequiresRegister());
  return summary;
}

class CheckStackOverflowSlowPath : public SlowPathCode {
 public:
  explicit CheckStackOverflowSlowPath(CheckStackOverflowInstr* instruction)
      : instruction_(instruction) {}

  virtual void EmitNativeCode(FlowGraphCompiler* compiler) {
    if (compiler->isolate()->use_osr() && osr_entry_label()->IsLinked()) {
      __ Comment("CheckStackOverflowSlowPathOsr");
      __ Bind(osr_entry_label());
      __ movq(Address(THR, Thread::stack_overflow_flags_offset()),
              Immediate(Thread::kOsrRequest));
    }
    __ Comment("CheckStackOverflowSlowPath");
    __ Bind(entry_label());
    compiler->SaveLiveRegisters(instruction_->locs());
    // pending_deoptimization_env_ is needed to generate a runtime call that
    // may throw an exception.
    ASSERT(compiler->pending_deoptimization_env_ == NULL);
    Environment* env = compiler->SlowPathEnvironmentFor(instruction_);
    compiler->pending_deoptimization_env_ = env;
    compiler->GenerateRuntimeCall(
        instruction_->token_pos(), instruction_->deopt_id(),
        kStackOverflowRuntimeEntry, 0, instruction_->locs());

    if (compiler->isolate()->use_osr() && !compiler->is_optimizing() &&
        instruction_->in_loop()) {
      // In unoptimized code, record loop stack checks as possible OSR entries.
      compiler->AddCurrentDescriptor(RawPcDescriptors::kOsrEntry,
                                     instruction_->deopt_id(),
                                     TokenPosition::kNoSource);
    }
    compiler->pending_deoptimization_env_ = NULL;
    compiler->RestoreLiveRegisters(instruction_->locs());
    __ jmp(exit_label());
  }

  Label* osr_entry_label() {
    ASSERT(Isolate::Current()->use_osr());
    return &osr_entry_label_;
  }

 private:
  CheckStackOverflowInstr* instruction_;
  Label osr_entry_label_;
};

void CheckStackOverflowInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  CheckStackOverflowSlowPath* slow_path = new CheckStackOverflowSlowPath(this);
  compiler->AddSlowPathCode(slow_path);

  Register temp = locs()->temp(0).reg();
  // Generate stack overflow check.
  __ cmpq(RSP, Address(THR, Thread::stack_limit_offset()));
  __ j(BELOW_EQUAL, slow_path->entry_label());
  if (compiler->CanOSRFunction() && in_loop()) {
    // In unoptimized code check the usage counter to trigger OSR at loop
    // stack checks.  Use progressively higher thresholds for more deeply
    // nested loops to attempt to hit outer loops with OSR when possible.
    __ LoadObject(temp, compiler->parsed_function().function());
    int32_t threshold =
        FLAG_optimization_counter_threshold * (loop_depth() + 1);
    __ cmpl(FieldAddress(temp, Function::usage_counter_offset()),
            Immediate(threshold));
    __ j(GREATER_EQUAL, slow_path->osr_entry_label());
  }
  if (compiler->ForceSlowPathForStackOverflow()) {
    __ jmp(slow_path->entry_label());
  }
  __ Bind(slow_path->exit_label());
}

static void EmitSmiShiftLeft(FlowGraphCompiler* compiler,
                             BinarySmiOpInstr* shift_left) {
  const LocationSummary& locs = *shift_left->locs();
  Register left = locs.in(0).reg();
  Register result = locs.out(0).reg();
  ASSERT(left == result);
  Label* deopt = shift_left->CanDeoptimize()
                     ? compiler->AddDeoptStub(shift_left->deopt_id(),
                                              ICData::kDeoptBinarySmiOp)
                     : NULL;
  if (locs.in(1).IsConstant()) {
    const Object& constant = locs.in(1).constant();
    ASSERT(constant.IsSmi());
    // shlq operation masks the count to 6 bits.
    const intptr_t kCountLimit = 0x3F;
    const intptr_t value = Smi::Cast(constant).Value();
    ASSERT((0 < value) && (value < kCountLimit));
    if (shift_left->can_overflow()) {
      if (value == 1) {
        // Use overflow flag.
        __ shlq(left, Immediate(1));
        __ j(OVERFLOW, deopt);
        return;
      }
      // Check for overflow.
      Register temp = locs.temp(0).reg();
      __ movq(temp, left);
      __ shlq(left, Immediate(value));
      __ sarq(left, Immediate(value));
      __ cmpq(left, temp);
      __ j(NOT_EQUAL, deopt);  // Overflow.
    }
    // Shift for result now we know there is no overflow.
    __ shlq(left, Immediate(value));
    return;
  }

  // Right (locs.in(1)) is not constant.
  Register right = locs.in(1).reg();
  Range* right_range = shift_left->right_range();
  if (shift_left->left()->BindsToConstant() && shift_left->can_overflow()) {
    // TODO(srdjan): Implement code below for is_truncating().
    // If left is constant, we know the maximal allowed size for right.
    const Object& obj = shift_left->left()->BoundConstant();
    if (obj.IsSmi()) {
      const intptr_t left_int = Smi::Cast(obj).Value();
      if (left_int == 0) {
        __ CompareImmediate(right, Immediate(0));
        __ j(NEGATIVE, deopt);
        return;
      }
      const intptr_t max_right = kSmiBits - Utils::HighestBit(left_int);
      const bool right_needs_check =
          !RangeUtils::IsWithin(right_range, 0, max_right - 1);
      if (right_needs_check) {
        __ CompareImmediate(
            right, Immediate(reinterpret_cast<int64_t>(Smi::New(max_right))));
        __ j(ABOVE_EQUAL, deopt);
      }
      __ SmiUntag(right);
      __ shlq(left, right);
    }
    return;
  }

  const bool right_needs_check =
      !RangeUtils::IsWithin(right_range, 0, (Smi::kBits - 1));
  ASSERT(right == RCX);  // Count must be in RCX
  if (!shift_left->can_overflow()) {
    if (right_needs_check) {
      const bool right_may_be_negative =
          (right_range == NULL) || !right_range->IsPositive();
      if (right_may_be_negative) {
        ASSERT(shift_left->CanDeoptimize());
        __ CompareImmediate(right, Immediate(0));
        __ j(NEGATIVE, deopt);
      }
      Label done, is_not_zero;
      __ CompareImmediate(
          right, Immediate(reinterpret_cast<int64_t>(Smi::New(Smi::kBits))));
      __ j(BELOW, &is_not_zero, Assembler::kNearJump);
      __ xorq(left, left);
      __ jmp(&done, Assembler::kNearJump);
      __ Bind(&is_not_zero);
      __ SmiUntag(right);
      __ shlq(left, right);
      __ Bind(&done);
    } else {
      __ SmiUntag(right);
      __ shlq(left, right);
    }
  } else {
    if (right_needs_check) {
      ASSERT(shift_left->CanDeoptimize());
      __ CompareImmediate(
          right, Immediate(reinterpret_cast<int64_t>(Smi::New(Smi::kBits))));
      __ j(ABOVE_EQUAL, deopt);
    }
    // Left is not a constant.
    Register temp = locs.temp(0).reg();
    // Check if count too large for handling it inlined.
    __ movq(temp, left);
    __ SmiUntag(right);
    // Overflow test (preserve temp and right);
    __ shlq(left, right);
    __ sarq(left, right);
    __ cmpq(left, temp);
    __ j(NOT_EQUAL, deopt);  // Overflow.
    // Shift for result now we know there is no overflow.
    __ shlq(left, right);
  }
}

class CheckedSmiSlowPath : public SlowPathCode {
 public:
  CheckedSmiSlowPath(CheckedSmiOpInstr* instruction, intptr_t try_index)
      : instruction_(instruction), try_index_(try_index) {}

  virtual void EmitNativeCode(FlowGraphCompiler* compiler) {
    if (Assembler::EmittingComments()) {
      __ Comment("slow path smi operation");
    }
    __ Bind(entry_label());
    LocationSummary* locs = instruction_->locs();
    Register result = locs->out(0).reg();
    locs->live_registers()->Remove(Location::RegisterLocation(result));

    compiler->SaveLiveRegisters(locs);
    if (instruction_->env() != NULL) {
      Environment* env = compiler->SlowPathEnvironmentFor(instruction_);
      compiler->pending_deoptimization_env_ = env;
    }
    __ pushq(locs->in(0).reg());
    __ pushq(locs->in(1).reg());
    const String& selector =
        String::Handle(instruction_->call()->ic_data()->target_name());
    const Array& arguments_descriptor =
        Array::Handle(instruction_->call()->ic_data()->arguments_descriptor());
    compiler->EmitMegamorphicInstanceCall(
        selector, arguments_descriptor, instruction_->call()->deopt_id(),
        instruction_->call()->token_pos(), locs, try_index_,
        /* slow_path_argument_count = */ 2);
    __ MoveRegister(result, RAX);
    compiler->RestoreLiveRegisters(locs);
    __ jmp(exit_label());
    compiler->pending_deoptimization_env_ = NULL;
  }

 private:
  CheckedSmiOpInstr* instruction_;
  intptr_t try_index_;
};

LocationSummary* CheckedSmiOpInstr::MakeLocationSummary(Zone* zone,
                                                        bool opt) const {
  bool is_shift = (op_kind() == Token::kSHL) || (op_kind() == Token::kSHR);
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = is_shift ? 1 : 0;
  LocationSummary* summary = new (zone) LocationSummary(
      zone, kNumInputs, kNumTemps, LocationSummary::kCallOnSlowPath);
  summary->set_in(0, Location::RequiresRegister());
  summary->set_in(1, Location::RequiresRegister());
  switch (op_kind()) {
    case Token::kADD:
    case Token::kSUB:
    case Token::kMUL:
    case Token::kSHL:
    case Token::kSHR:
      summary->set_out(0, Location::RequiresRegister());
      break;
    case Token::kBIT_OR:
    case Token::kBIT_AND:
    case Token::kBIT_XOR:
      summary->set_out(0, Location::SameAsFirstInput());
      break;
    default:
      UNIMPLEMENTED();
  }
  if (is_shift) {
    summary->set_temp(0, Location::RegisterLocation(RCX));
  }
  return summary;
}

void CheckedSmiOpInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  CheckedSmiSlowPath* slow_path =
      new CheckedSmiSlowPath(this, compiler->CurrentTryIndex());
  compiler->AddSlowPathCode(slow_path);
  // Test operands if necessary.

  intptr_t left_cid = left()->Type()->ToCid();
  intptr_t right_cid = right()->Type()->ToCid();
  Register left = locs()->in(0).reg();
  Register right = locs()->in(1).reg();
  if (this->left()->definition() == this->right()->definition()) {
    __ testq(left, Immediate(kSmiTagMask));
  } else if (left_cid == kSmiCid) {
    __ testq(right, Immediate(kSmiTagMask));
  } else if (right_cid == kSmiCid) {
    __ testq(left, Immediate(kSmiTagMask));
  } else {
    __ movq(TMP, left);
    __ orq(TMP, right);
    __ testq(TMP, Immediate(kSmiTagMask));
  }
  __ j(NOT_ZERO, slow_path->entry_label());
  Register result = locs()->out(0).reg();
  switch (op_kind()) {
    case Token::kADD:
      __ movq(result, left);
      __ addq(result, right);
      __ j(OVERFLOW, slow_path->entry_label());
      break;
    case Token::kSUB:
      __ movq(result, left);
      __ subq(result, right);
      __ j(OVERFLOW, slow_path->entry_label());
      break;
    case Token::kMUL:
      __ movq(result, left);
      __ SmiUntag(result);
      __ imulq(result, right);
      __ j(OVERFLOW, slow_path->entry_label());
      break;
    case Token::kBIT_OR:
      ASSERT(left == result);
      __ orq(result, right);
      break;
    case Token::kBIT_AND:
      ASSERT(left == result);
      __ andq(result, right);
      break;
    case Token::kBIT_XOR:
      ASSERT(left == result);
      __ xorq(result, right);
      break;
    case Token::kSHL:
      ASSERT(result != right);
      ASSERT(locs()->temp(0).reg() == RCX);
      __ cmpq(right, Immediate(Smi::RawValue(Smi::kBits)));
      __ j(ABOVE_EQUAL, slow_path->entry_label());

      __ movq(RCX, right);
      __ SmiUntag(RCX);
      __ movq(result, left);
      __ shlq(result, RCX);
      __ movq(TMP, result);
      __ sarq(TMP, RCX);
      __ cmpq(TMP, left);
      __ j(NOT_EQUAL, slow_path->entry_label());
      break;
    case Token::kSHR: {
      Label shift_count_ok;
      ASSERT(result != right);
      ASSERT(locs()->temp(0).reg() == RCX);
      __ cmpq(right, Immediate(Smi::RawValue(Smi::kBits)));
      __ j(ABOVE_EQUAL, slow_path->entry_label());

      __ movq(RCX, right);
      __ SmiUntag(RCX);
      __ movq(result, left);
      __ SmiUntag(result);
      __ sarq(result, RCX);
      __ SmiTag(result);
      break;
    }
    default:
      UNIMPLEMENTED();
  }
  __ Bind(slow_path->exit_label());
}

class CheckedSmiComparisonSlowPath : public SlowPathCode {
 public:
  CheckedSmiComparisonSlowPath(CheckedSmiComparisonInstr* instruction,
                               intptr_t try_index,
                               BranchLabels labels,
                               bool merged = false)
      : instruction_(instruction),
        try_index_(try_index),
        labels_(labels),
        merged_(merged) {}

  virtual void EmitNativeCode(FlowGraphCompiler* compiler) {
    if (Assembler::EmittingComments()) {
      __ Comment("slow path smi comparison");
    }
    __ Bind(entry_label());
    LocationSummary* locs = instruction_->locs();
    Register result = merged_ ? locs->temp(0).reg() : locs->out(0).reg();
    locs->live_registers()->Remove(Location::RegisterLocation(result));

    compiler->SaveLiveRegisters(locs);
    if (instruction_->env() != NULL) {
      Environment* env = compiler->SlowPathEnvironmentFor(instruction_);
      compiler->pending_deoptimization_env_ = env;
    }
    __ pushq(locs->in(0).reg());
    __ pushq(locs->in(1).reg());
    String& selector =
        String::Handle(instruction_->call()->ic_data()->target_name());
    const Array& arguments_descriptor =
        Array::Handle(instruction_->call()->ic_data()->arguments_descriptor());
    compiler->EmitMegamorphicInstanceCall(
        selector, arguments_descriptor, instruction_->call()->deopt_id(),
        instruction_->call()->token_pos(), locs, try_index_,
        /* slow_path_argument_count = */ 2);
    __ MoveRegister(result, RAX);
    compiler->RestoreLiveRegisters(locs);
    compiler->pending_deoptimization_env_ = NULL;
    if (merged_) {
      __ CompareObject(result, Bool::True());
      __ j(EQUAL, instruction_->is_negated() ? labels_.false_label
                                             : labels_.true_label);
      __ jmp(instruction_->is_negated() ? labels_.true_label
                                        : labels_.false_label);
    } else {
      __ jmp(exit_label());
    }
  }

 private:
  CheckedSmiComparisonInstr* instruction_;
  intptr_t try_index_;
  BranchLabels labels_;
  bool merged_;
};

LocationSummary* CheckedSmiComparisonInstr::MakeLocationSummary(
    Zone* zone,
    bool opt) const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 1;
  LocationSummary* summary = new (zone) LocationSummary(
      zone, kNumInputs, kNumTemps, LocationSummary::kCallOnSlowPath);
  summary->set_in(0, Location::RequiresRegister());
  summary->set_in(1, Location::RequiresRegister());
  summary->set_temp(0, Location::RequiresRegister());
  summary->set_out(0, Location::RequiresRegister());
  return summary;
}

Condition CheckedSmiComparisonInstr::EmitComparisonCode(
    FlowGraphCompiler* compiler,
    BranchLabels labels) {
  return EmitInt64ComparisonOp(compiler, *locs(), kind());
}

#define EMIT_SMI_CHECK                                                         \
  intptr_t left_cid = left()->Type()->ToCid();                                 \
  intptr_t right_cid = right()->Type()->ToCid();                               \
  Register left = locs()->in(0).reg();                                         \
  Register right = locs()->in(1).reg();                                        \
  if (this->left()->definition() == this->right()->definition()) {             \
    __ testq(left, Immediate(kSmiTagMask));                                    \
  } else if (left_cid == kSmiCid) {                                            \
    __ testq(right, Immediate(kSmiTagMask));                                   \
  } else if (right_cid == kSmiCid) {                                           \
    __ testq(left, Immediate(kSmiTagMask));                                    \
  } else {                                                                     \
    __ movq(TMP, left);                                                        \
    __ orq(TMP, right);                                                        \
    __ testq(TMP, Immediate(kSmiTagMask));                                     \
  }                                                                            \
  __ j(NOT_ZERO, slow_path->entry_label())

void CheckedSmiComparisonInstr::EmitBranchCode(FlowGraphCompiler* compiler,
                                               BranchInstr* branch) {
  BranchLabels labels = compiler->CreateBranchLabels(branch);
  CheckedSmiComparisonSlowPath* slow_path = new CheckedSmiComparisonSlowPath(
      this, compiler->CurrentTryIndex(), labels,
      /* merged = */ true);
  compiler->AddSlowPathCode(slow_path);
  EMIT_SMI_CHECK;
  Condition true_condition = EmitComparisonCode(compiler, labels);
  ASSERT(true_condition != INVALID_CONDITION);
  EmitBranchOnCondition(compiler, true_condition, labels);
  __ Bind(slow_path->exit_label());
}

void CheckedSmiComparisonInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Label true_label, false_label, done;
  BranchLabels labels = {&true_label, &false_label, &false_label};
  CheckedSmiComparisonSlowPath* slow_path = new CheckedSmiComparisonSlowPath(
      this, compiler->CurrentTryIndex(), labels,
      /* merged = */ false);
  compiler->AddSlowPathCode(slow_path);
  EMIT_SMI_CHECK;
  Condition true_condition = EmitComparisonCode(compiler, labels);
  ASSERT(true_condition != INVALID_CONDITION);
  EmitBranchOnCondition(compiler, true_condition, labels);
  Register result = locs()->out(0).reg();
  __ Bind(&false_label);
  __ LoadObject(result, Bool::False());
  __ jmp(&done);
  __ Bind(&true_label);
  __ LoadObject(result, Bool::True());
  __ Bind(&done);
  __ Bind(slow_path->exit_label());
}

static bool CanBeImmediate(const Object& constant) {
  return constant.IsSmi() &&
         Immediate(reinterpret_cast<int64_t>(constant.raw())).is_int32();
}

static bool IsSmiValue(const Object& constant, intptr_t value) {
  return constant.IsSmi() && (Smi::Cast(constant).Value() == value);
}

LocationSummary* BinarySmiOpInstr::MakeLocationSummary(Zone* zone,
                                                       bool opt) const {
  const intptr_t kNumInputs = 2;

  ConstantInstr* right_constant = right()->definition()->AsConstant();
  if ((right_constant != NULL) && (op_kind() != Token::kTRUNCDIV) &&
      (op_kind() != Token::kSHL) && (op_kind() != Token::kMUL) &&
      (op_kind() != Token::kMOD) && CanBeImmediate(right_constant->value())) {
    const intptr_t kNumTemps = 0;
    LocationSummary* summary = new (zone)
        LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
    summary->set_in(0, Location::RequiresRegister());
    summary->set_in(1, Location::Constant(right_constant));
    summary->set_out(0, Location::SameAsFirstInput());
    return summary;
  }

  if (op_kind() == Token::kTRUNCDIV) {
    const intptr_t kNumTemps = 1;
    LocationSummary* summary = new (zone)
        LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
    if (RightIsPowerOfTwoConstant()) {
      summary->set_in(0, Location::RequiresRegister());
      ConstantInstr* right_constant = right()->definition()->AsConstant();
      summary->set_in(1, Location::Constant(right_constant));
      summary->set_temp(0, Location::RequiresRegister());
      summary->set_out(0, Location::SameAsFirstInput());
    } else {
      // Both inputs must be writable because they will be untagged.
      summary->set_in(0, Location::RegisterLocation(RAX));
      summary->set_in(1, Location::WritableRegister());
      summary->set_out(0, Location::SameAsFirstInput());
      // Will be used for sign extension and division.
      summary->set_temp(0, Location::RegisterLocation(RDX));
    }
    return summary;
  } else if (op_kind() == Token::kMOD) {
    const intptr_t kNumTemps = 1;
    LocationSummary* summary = new (zone)
        LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
    // Both inputs must be writable because they will be untagged.
    summary->set_in(0, Location::RegisterLocation(RDX));
    summary->set_in(1, Location::WritableRegister());
    summary->set_out(0, Location::SameAsFirstInput());
    // Will be used for sign extension and division.
    summary->set_temp(0, Location::RegisterLocation(RAX));
    return summary;
  } else if (op_kind() == Token::kSHR) {
    const intptr_t kNumTemps = 0;
    LocationSummary* summary = new (zone)
        LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
    summary->set_in(0, Location::RequiresRegister());
    summary->set_in(1, Location::FixedRegisterOrSmiConstant(right(), RCX));
    summary->set_out(0, Location::SameAsFirstInput());
    return summary;
  } else if (op_kind() == Token::kSHL) {
    // Shift-by-1 overflow checking can use flags, otherwise we need a temp.
    const bool shiftBy1 =
        (right_constant != NULL) && IsSmiValue(right_constant->value(), 1);
    const intptr_t kNumTemps = (can_overflow() && !shiftBy1) ? 1 : 0;
    LocationSummary* summary = new (zone)
        LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
    summary->set_in(0, Location::RequiresRegister());
    summary->set_in(1, Location::FixedRegisterOrSmiConstant(right(), RCX));
    if (kNumTemps == 1) {
      summary->set_temp(0, Location::RequiresRegister());
    }
    summary->set_out(0, Location::SameAsFirstInput());
    return summary;
  } else {
    const intptr_t kNumTemps = 0;
    LocationSummary* summary = new (zone)
        LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
    summary->set_in(0, Location::RequiresRegister());
    ConstantInstr* constant = right()->definition()->AsConstant();
    if (constant != NULL) {
      summary->set_in(1, Location::RegisterOrSmiConstant(right()));
    } else {
      summary->set_in(1, Location::PrefersRegister());
    }
    summary->set_out(0, Location::SameAsFirstInput());
    return summary;
  }
}

void BinarySmiOpInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  if (op_kind() == Token::kSHL) {
    EmitSmiShiftLeft(compiler, this);
    return;
  }

  Register left = locs()->in(0).reg();
  Register result = locs()->out(0).reg();
  ASSERT(left == result);
  Label* deopt = NULL;
  if (CanDeoptimize()) {
    deopt = compiler->AddDeoptStub(deopt_id(), ICData::kDeoptBinarySmiOp);
  }

  if (locs()->in(1).IsConstant()) {
    const Object& constant = locs()->in(1).constant();
    ASSERT(constant.IsSmi());
    const int64_t imm = reinterpret_cast<int64_t>(constant.raw());
    switch (op_kind()) {
      case Token::kADD: {
        __ AddImmediate(left, Immediate(imm));
        if (deopt != NULL) __ j(OVERFLOW, deopt);
        break;
      }
      case Token::kSUB: {
        __ SubImmediate(left, Immediate(imm));
        if (deopt != NULL) __ j(OVERFLOW, deopt);
        break;
      }
      case Token::kMUL: {
        // Keep left value tagged and untag right value.
        const intptr_t value = Smi::Cast(constant).Value();
        __ MulImmediate(left, Immediate(value));
        if (deopt != NULL) __ j(OVERFLOW, deopt);
        break;
      }
      case Token::kTRUNCDIV: {
        const intptr_t value = Smi::Cast(constant).Value();
        ASSERT(value != kIntptrMin);
        ASSERT(Utils::IsPowerOfTwo(Utils::Abs(value)));
        const intptr_t shift_count =
            Utils::ShiftForPowerOfTwo(Utils::Abs(value)) + kSmiTagSize;
        ASSERT(kSmiTagSize == 1);
        Register temp = locs()->temp(0).reg();
        __ movq(temp, left);
        __ sarq(temp, Immediate(63));
        ASSERT(shift_count > 1);  // 1, -1 case handled above.
        __ shrq(temp, Immediate(64 - shift_count));
        __ addq(left, temp);
        ASSERT(shift_count > 0);
        __ sarq(left, Immediate(shift_count));
        if (value < 0) {
          __ negq(left);
        }
        __ SmiTag(left);
        break;
      }
      case Token::kBIT_AND: {
        // No overflow check.
        __ AndImmediate(left, Immediate(imm));
        break;
      }
      case Token::kBIT_OR: {
        // No overflow check.
        __ OrImmediate(left, Immediate(imm));
        break;
      }
      case Token::kBIT_XOR: {
        // No overflow check.
        __ XorImmediate(left, Immediate(imm));
        break;
      }

      case Token::kSHR: {
        // sarq operation masks the count to 6 bits.
        const intptr_t kCountLimit = 0x3F;
        const intptr_t value = Smi::Cast(constant).Value();
        __ sarq(left,
                Immediate(Utils::Minimum(value + kSmiTagSize, kCountLimit)));
        __ SmiTag(left);
        break;
      }

      default:
        UNREACHABLE();
        break;
    }
    return;
  }  // locs()->in(1).IsConstant().

  if (locs()->in(1).IsStackSlot()) {
    const Address& right = locs()->in(1).ToStackSlotAddress();
    switch (op_kind()) {
      case Token::kADD: {
        __ addq(left, right);
        if (deopt != NULL) __ j(OVERFLOW, deopt);
        break;
      }
      case Token::kSUB: {
        __ subq(left, right);
        if (deopt != NULL) __ j(OVERFLOW, deopt);
        break;
      }
      case Token::kMUL: {
        __ SmiUntag(left);
        __ imulq(left, right);
        if (deopt != NULL) __ j(OVERFLOW, deopt);
        break;
      }
      case Token::kBIT_AND: {
        // No overflow check.
        __ andq(left, right);
        break;
      }
      case Token::kBIT_OR: {
        // No overflow check.
        __ orq(left, right);
        break;
      }
      case Token::kBIT_XOR: {
        // No overflow check.
        __ xorq(left, right);
        break;
      }
      default:
        UNREACHABLE();
        break;
    }
    return;
  }  // locs()->in(1).IsStackSlot().

  // if locs()->in(1).IsRegister.
  Register right = locs()->in(1).reg();
  switch (op_kind()) {
    case Token::kADD: {
      __ addq(left, right);
      if (deopt != NULL) __ j(OVERFLOW, deopt);
      break;
    }
    case Token::kSUB: {
      __ subq(left, right);
      if (deopt != NULL) __ j(OVERFLOW, deopt);
      break;
    }
    case Token::kMUL: {
      __ SmiUntag(left);
      __ imulq(left, right);
      if (deopt != NULL) __ j(OVERFLOW, deopt);
      break;
    }
    case Token::kBIT_AND: {
      // No overflow check.
      __ andq(left, right);
      break;
    }
    case Token::kBIT_OR: {
      // No overflow check.
      __ orq(left, right);
      break;
    }
    case Token::kBIT_XOR: {
      // No overflow check.
      __ xorq(left, right);
      break;
    }
    case Token::kTRUNCDIV: {
      Label not_32bit, done;

      Register temp = locs()->temp(0).reg();
      ASSERT(left == RAX);
      ASSERT((right != RDX) && (right != RAX));
      ASSERT(temp == RDX);
      ASSERT(result == RAX);
      if (RangeUtils::CanBeZero(right_range())) {
        // Handle divide by zero in runtime.
        __ testq(right, right);
        __ j(ZERO, deopt);
      }
      // Check if both operands fit into 32bits as idiv with 64bit operands
      // requires twice as many cycles and has much higher latency.
      // We are checking this before untagging them to avoid corner case
      // dividing INT_MAX by -1 that raises exception because quotient is
      // too large for 32bit register.
      __ movsxd(temp, left);
      __ cmpq(temp, left);
      __ j(NOT_EQUAL, &not_32bit);
      __ movsxd(temp, right);
      __ cmpq(temp, right);
      __ j(NOT_EQUAL, &not_32bit);

      // Both operands are 31bit smis. Divide using 32bit idiv.
      __ SmiUntag(left);
      __ SmiUntag(right);
      __ cdq();
      __ idivl(right);
      __ movsxd(result, result);
      __ jmp(&done);

      // Divide using 64bit idiv.
      __ Bind(&not_32bit);
      __ SmiUntag(left);
      __ SmiUntag(right);
      __ cqo();         // Sign extend RAX -> RDX:RAX.
      __ idivq(right);  //  RAX: quotient, RDX: remainder.
      // Check the corner case of dividing the 'MIN_SMI' with -1, in which
      // case we cannot tag the result.
      __ CompareImmediate(result, Immediate(0x4000000000000000));
      __ j(EQUAL, deopt);
      __ Bind(&done);
      __ SmiTag(result);
      break;
    }
    case Token::kMOD: {
      Label not_32bit, div_done;

      Register temp = locs()->temp(0).reg();
      ASSERT(left == RDX);
      ASSERT((right != RDX) && (right != RAX));
      ASSERT(temp == RAX);
      ASSERT(result == RDX);
      if (RangeUtils::CanBeZero(right_range())) {
        // Handle divide by zero in runtime.
        __ testq(right, right);
        __ j(ZERO, deopt);
      }
      // Check if both operands fit into 32bits as idiv with 64bit operands
      // requires twice as many cycles and has much higher latency.
      // We are checking this before untagging them to avoid corner case
      // dividing INT_MAX by -1 that raises exception because quotient is
      // too large for 32bit register.
      __ movsxd(temp, left);
      __ cmpq(temp, left);
      __ j(NOT_EQUAL, &not_32bit);
      __ movsxd(temp, right);
      __ cmpq(temp, right);
      __ j(NOT_EQUAL, &not_32bit);
      // Both operands are 31bit smis. Divide using 32bit idiv.
      __ SmiUntag(left);
      __ SmiUntag(right);
      __ movq(RAX, RDX);
      __ cdq();
      __ idivl(right);
      __ movsxd(result, result);
      __ jmp(&div_done);

      // Divide using 64bit idiv.
      __ Bind(&not_32bit);
      __ SmiUntag(left);
      __ SmiUntag(right);
      __ movq(RAX, RDX);
      __ cqo();         // Sign extend RAX -> RDX:RAX.
      __ idivq(right);  //  RAX: quotient, RDX: remainder.
      __ Bind(&div_done);
      //  res = left % right;
      //  if (res < 0) {
      //    if (right < 0) {
      //      res = res - right;
      //    } else {
      //      res = res + right;
      //    }
      //  }
      Label all_done;
      __ cmpq(result, Immediate(0));
      __ j(GREATER_EQUAL, &all_done, Assembler::kNearJump);
      // Result is negative, adjust it.
      if (RangeUtils::Overlaps(right_range(), -1, 1)) {
        Label subtract;
        __ cmpq(right, Immediate(0));
        __ j(LESS, &subtract, Assembler::kNearJump);
        __ addq(result, right);
        __ jmp(&all_done, Assembler::kNearJump);
        __ Bind(&subtract);
        __ subq(result, right);
      } else if (right_range()->IsPositive()) {
        // Right is positive.
        __ addq(result, right);
      } else {
        // Right is negative.
        __ subq(result, right);
      }
      __ Bind(&all_done);
      __ SmiTag(result);
      break;
    }
    case Token::kSHR: {
      if (CanDeoptimize()) {
        __ CompareImmediate(right, Immediate(0));
        __ j(LESS, deopt);
      }
      __ SmiUntag(right);
      // sarq operation masks the count to 6 bits.
      const intptr_t kCountLimit = 0x3F;
      if (!RangeUtils::OnlyLessThanOrEqualTo(right_range(), kCountLimit)) {
        __ CompareImmediate(right, Immediate(kCountLimit));
        Label count_ok;
        __ j(LESS, &count_ok, Assembler::kNearJump);
        __ LoadImmediate(right, Immediate(kCountLimit));
        __ Bind(&count_ok);
      }
      ASSERT(right == RCX);  // Count must be in RCX
      __ SmiUntag(left);
      __ sarq(left, right);
      __ SmiTag(left);
      break;
    }
    case Token::kDIV: {
      // Dispatches to 'Double./'.
      // TODO(srdjan): Implement as conversion to double and double division.
      UNREACHABLE();
      break;
    }
    case Token::kOR:
    case Token::kAND: {
      // Flow graph builder has dissected this operation to guarantee correct
      // behavior (short-circuit evaluation).
      UNREACHABLE();
      break;
    }
    default:
      UNREACHABLE();
      break;
  }
}

LocationSummary* CheckEitherNonSmiInstr::MakeLocationSummary(Zone* zone,
                                                             bool opt) const {
  intptr_t left_cid = left()->Type()->ToCid();
  intptr_t right_cid = right()->Type()->ToCid();
  ASSERT((left_cid != kDoubleCid) && (right_cid != kDoubleCid));
  const intptr_t kNumInputs = 2;
  const bool need_temp = (left()->definition() != right()->definition()) &&
                         (left_cid != kSmiCid) && (right_cid != kSmiCid);
  const intptr_t kNumTemps = need_temp ? 1 : 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  summary->set_in(1, Location::RequiresRegister());
  if (need_temp) summary->set_temp(0, Location::RequiresRegister());
  return summary;
}

void CheckEitherNonSmiInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Label* deopt =
      compiler->AddDeoptStub(deopt_id(), ICData::kDeoptBinaryDoubleOp,
                             licm_hoisted_ ? ICData::kHoisted : 0);
  intptr_t left_cid = left()->Type()->ToCid();
  intptr_t right_cid = right()->Type()->ToCid();
  Register left = locs()->in(0).reg();
  Register right = locs()->in(1).reg();
  if (this->left()->definition() == this->right()->definition()) {
    __ testq(left, Immediate(kSmiTagMask));
  } else if (left_cid == kSmiCid) {
    __ testq(right, Immediate(kSmiTagMask));
  } else if (right_cid == kSmiCid) {
    __ testq(left, Immediate(kSmiTagMask));
  } else {
    Register temp = locs()->temp(0).reg();
    __ movq(temp, left);
    __ orq(temp, right);
    __ testq(temp, Immediate(kSmiTagMask));
  }
  __ j(ZERO, deopt);
}

LocationSummary* BoxInstr::MakeLocationSummary(Zone* zone, bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 1;
  LocationSummary* summary = new (zone) LocationSummary(
      zone, kNumInputs, kNumTemps, LocationSummary::kCallOnSlowPath);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_temp(0, Location::RequiresRegister());
  summary->set_out(0, Location::RequiresRegister());
  return summary;
}

void BoxInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register out_reg = locs()->out(0).reg();
  Register temp = locs()->temp(0).reg();
  XmmRegister value = locs()->in(0).fpu_reg();

  BoxAllocationSlowPath::Allocate(compiler, this,
                                  compiler->BoxClassFor(from_representation()),
                                  out_reg, temp);

  switch (from_representation()) {
    case kUnboxedDouble:
      __ movsd(FieldAddress(out_reg, ValueOffset()), value);
      break;
    case kUnboxedFloat32x4:
    case kUnboxedFloat64x2:
    case kUnboxedInt32x4:
      __ movups(FieldAddress(out_reg, ValueOffset()), value);
      break;
    default:
      UNREACHABLE();
      break;
  }
}

LocationSummary* UnboxInstr::MakeLocationSummary(Zone* zone, bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  const bool needs_writable_input =
      (representation() != kUnboxedInt64) &&
      (value()->Type()->ToNullableCid() != BoxCid());
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, needs_writable_input ? Location::WritableRegister()
                                          : Location::RequiresRegister());
  if (representation() == kUnboxedInt64) {
    summary->set_out(0, Location::SameAsFirstInput());
  } else {
    summary->set_out(0, Location::RequiresFpuRegister());
  }
  return summary;
}

void UnboxInstr::EmitLoadFromBox(FlowGraphCompiler* compiler) {
  const Register box = locs()->in(0).reg();

  switch (representation()) {
    case kUnboxedInt64: {
      const Register result = locs()->out(0).reg();
      __ movq(result, FieldAddress(box, ValueOffset()));
      break;
    }

    case kUnboxedDouble: {
      const FpuRegister result = locs()->out(0).fpu_reg();
      __ movsd(result, FieldAddress(box, ValueOffset()));
      break;
    }

    case kUnboxedFloat32x4:
    case kUnboxedFloat64x2:
    case kUnboxedInt32x4: {
      const FpuRegister result = locs()->out(0).fpu_reg();
      __ movups(result, FieldAddress(box, ValueOffset()));
      break;
    }

    default:
      UNREACHABLE();
      break;
  }
}

void UnboxInstr::EmitSmiConversion(FlowGraphCompiler* compiler) {
  const Register box = locs()->in(0).reg();

  switch (representation()) {
    case kUnboxedInt64: {
      const Register result = locs()->out(0).reg();
      ASSERT(result == box);
      __ SmiUntag(box);
      break;
    }

    case kUnboxedDouble: {
      const FpuRegister result = locs()->out(0).fpu_reg();
      __ SmiUntag(box);
      __ cvtsi2sdq(result, box);
      break;
    }

    default:
      UNREACHABLE();
      break;
  }
}

void UnboxInstr::EmitLoadInt64FromBoxOrSmi(FlowGraphCompiler* compiler) {
  const Register value = locs()->in(0).reg();
  const Register result = locs()->out(0).reg();
  ASSERT(value == result);
  Label done;
  __ SmiUntag(value);
  __ j(NOT_CARRY, &done, Assembler::kNearJump);
  __ movq(value, Address(value, TIMES_2, Mint::value_offset()));
  __ Bind(&done);
}

LocationSummary* UnboxInteger32Instr::MakeLocationSummary(Zone* zone,
                                                          bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = (!is_truncating() && CanDeoptimize()) ? 1 : 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  summary->set_out(0, Location::SameAsFirstInput());
  if (kNumTemps > 0) {
    summary->set_temp(0, Location::RequiresRegister());
  }
  return summary;
}

void UnboxInteger32Instr::EmitNativeCode(FlowGraphCompiler* compiler) {
  const intptr_t value_cid = value()->Type()->ToCid();
  const Register value = locs()->in(0).reg();
  Label* deopt = CanDeoptimize() ? compiler->AddDeoptStub(
                                       GetDeoptId(), ICData::kDeoptUnboxInteger)
                                 : NULL;
  ASSERT(value == locs()->out(0).reg());

  if (value_cid == kSmiCid) {
    __ SmiUntag(value);
  } else if (value_cid == kMintCid) {
    __ movq(value, FieldAddress(value, Mint::value_offset()));
  } else if (!CanDeoptimize()) {
    // Type information is not conclusive, but range analysis found
    // the value to be in int64 range. Therefore it must be a smi
    // or mint value.
    ASSERT(is_truncating());
    Label done;
    __ SmiUntag(value);
    __ j(NOT_CARRY, &done, Assembler::kNearJump);
    __ movq(value, Address(value, TIMES_2, Mint::value_offset()));
    __ Bind(&done);
    return;
  } else {
    Label done;
    // Optimistically untag value.
    __ SmiUntagOrCheckClass(value, kMintCid, &done);
    __ j(NOT_EQUAL, deopt);
    // Undo untagging by multiplying value with 2.
    __ movq(value, Address(value, TIMES_2, Mint::value_offset()));
    __ Bind(&done);
  }

  // TODO(vegorov): as it is implemented right now truncating unboxing would
  // leave "garbage" in the higher word.
  if (!is_truncating() && (deopt != NULL)) {
    ASSERT(representation() == kUnboxedInt32);
    Register temp = locs()->temp(0).reg();
    __ movsxd(temp, value);
    __ cmpq(temp, value);
    __ j(NOT_EQUAL, deopt);
  }
}

LocationSummary* BoxInteger32Instr::MakeLocationSummary(Zone* zone,
                                                        bool opt) const {
  ASSERT((from_representation() == kUnboxedInt32) ||
         (from_representation() == kUnboxedUint32));
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  summary->set_out(0, Location::RequiresRegister());
  return summary;
}

void BoxInteger32Instr::EmitNativeCode(FlowGraphCompiler* compiler) {
  const Register value = locs()->in(0).reg();
  const Register out = locs()->out(0).reg();
  ASSERT(value != out);

  ASSERT(kSmiTagSize == 1);
  if (from_representation() == kUnboxedInt32) {
    __ movsxd(out, value);
  } else {
    ASSERT(from_representation() == kUnboxedUint32);
    __ movl(out, value);
  }
  __ SmiTag(out);
}

LocationSummary* BoxInt64Instr::MakeLocationSummary(Zone* zone,
                                                    bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = ValueFitsSmi() ? 0 : 1;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps,
                      ValueFitsSmi() ? LocationSummary::kNoCall
                                     : LocationSummary::kCallOnSlowPath);
  summary->set_in(0, Location::RequiresRegister());
  if (!ValueFitsSmi()) {
    summary->set_temp(0, Location::RequiresRegister());
  }
  summary->set_out(0, Location::RequiresRegister());
  return summary;
}

void BoxInt64Instr::EmitNativeCode(FlowGraphCompiler* compiler) {
  const Register out = locs()->out(0).reg();
  const Register value = locs()->in(0).reg();
  __ MoveRegister(out, value);
  __ SmiTag(out);
  if (!ValueFitsSmi()) {
    const Register temp = locs()->temp(0).reg();
    Label done;
    __ j(NO_OVERFLOW, &done);
    BoxAllocationSlowPath::Allocate(compiler, this, compiler->mint_class(), out,
                                    temp);
    __ movq(FieldAddress(out, Mint::value_offset()), value);
    __ Bind(&done);
  }
}

LocationSummary* BinaryDoubleOpInstr::MakeLocationSummary(Zone* zone,
                                                          bool opt) const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_in(1, Location::RequiresFpuRegister());
  summary->set_out(0, Location::SameAsFirstInput());
  return summary;
}

void BinaryDoubleOpInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister left = locs()->in(0).fpu_reg();
  XmmRegister right = locs()->in(1).fpu_reg();

  ASSERT(locs()->out(0).fpu_reg() == left);

  switch (op_kind()) {
    case Token::kADD:
      __ addsd(left, right);
      break;
    case Token::kSUB:
      __ subsd(left, right);
      break;
    case Token::kMUL:
      __ mulsd(left, right);
      break;
    case Token::kDIV:
      __ divsd(left, right);
      break;
    default:
      UNREACHABLE();
  }
}

LocationSummary* DoubleTestOpInstr::MakeLocationSummary(Zone* zone,
                                                        bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps =
      (op_kind() == MethodRecognizer::kDouble_getIsInfinite) ? 1 : 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  if (op_kind() == MethodRecognizer::kDouble_getIsInfinite) {
    summary->set_temp(0, Location::RequiresRegister());
  }
  summary->set_out(0, Location::RequiresRegister());
  return summary;
}

Condition DoubleTestOpInstr::EmitComparisonCode(FlowGraphCompiler* compiler,
                                                BranchLabels labels) {
  ASSERT(compiler->is_optimizing());
  const XmmRegister value = locs()->in(0).fpu_reg();
  const bool is_negated = kind() != Token::kEQ;
  if (op_kind() == MethodRecognizer::kDouble_getIsNaN) {
    Label is_nan;
    __ comisd(value, value);
    return is_negated ? PARITY_ODD : PARITY_EVEN;
  } else {
    ASSERT(op_kind() == MethodRecognizer::kDouble_getIsInfinite);
    const Register temp = locs()->temp(0).reg();
    __ AddImmediate(RSP, Immediate(-kDoubleSize));
    __ movsd(Address(RSP, 0), value);
    __ movq(temp, Address(RSP, 0));
    __ AddImmediate(RSP, Immediate(kDoubleSize));
    // Mask off the sign.
    __ AndImmediate(temp, Immediate(0x7FFFFFFFFFFFFFFFLL));
    // Compare with +infinity.
    __ CompareImmediate(temp, Immediate(0x7FF0000000000000LL));
    return is_negated ? NOT_EQUAL : EQUAL;
  }
}

// SIMD

#define DEFINE_EMIT(Name, Args)                                                \
  static void Emit##Name(FlowGraphCompiler* compiler, SimdOpInstr* instr,      \
                         PP_APPLY(PP_UNPACK, Args))

#define SIMD_OP_FLOAT_ARITH(V, Name, op)                                       \
  V(Float32x4##Name, op##ps)                                                   \
  V(Float64x2##Name, op##pd)

#define SIMD_OP_SIMPLE_BINARY(V)                                               \
  SIMD_OP_FLOAT_ARITH(V, Add, add)                                             \
  SIMD_OP_FLOAT_ARITH(V, Sub, sub)                                             \
  SIMD_OP_FLOAT_ARITH(V, Mul, mul)                                             \
  SIMD_OP_FLOAT_ARITH(V, Div, div)                                             \
  SIMD_OP_FLOAT_ARITH(V, Min, min)                                             \
  SIMD_OP_FLOAT_ARITH(V, Max, max)                                             \
  V(Int32x4Add, addpl)                                                         \
  V(Int32x4Sub, subpl)                                                         \
  V(Int32x4BitAnd, andps)                                                      \
  V(Int32x4BitOr, orps)                                                        \
  V(Int32x4BitXor, xorps)                                                      \
  V(Float32x4Equal, cmppseq)                                                   \
  V(Float32x4NotEqual, cmppsneq)                                               \
  V(Float32x4GreaterThan, cmppsnle)                                            \
  V(Float32x4GreaterThanOrEqual, cmppsnlt)                                     \
  V(Float32x4LessThan, cmppslt)                                                \
  V(Float32x4LessThanOrEqual, cmppsle)

DEFINE_EMIT(SimdBinaryOp,
            (SameAsFirstInput, XmmRegister left, XmmRegister right)) {
  switch (instr->kind()) {
#define EMIT(Name, op)                                                         \
  case SimdOpInstr::k##Name:                                                   \
    __ op(left, right);                                                        \
    break;
    SIMD_OP_SIMPLE_BINARY(EMIT)
#undef EMIT
    case SimdOpInstr::kFloat32x4Scale:
      __ cvtsd2ss(left, left);
      __ shufps(left, left, Immediate(0x00));
      __ mulps(left, right);
      break;
    case SimdOpInstr::kFloat32x4ShuffleMix:
    case SimdOpInstr::kInt32x4ShuffleMix:
      __ shufps(left, right, Immediate(instr->mask()));
      break;
    case SimdOpInstr::kFloat64x2Constructor:
      // shufpd mask 0x0 results in:
      // Lower 64-bits of left = Lower 64-bits of left.
      // Upper 64-bits of left = Lower 64-bits of right.
      __ shufpd(left, right, Immediate(0x0));
      break;
    case SimdOpInstr::kFloat64x2Scale:
      __ shufpd(right, right, Immediate(0x00));
      __ mulpd(left, right);
      break;
    case SimdOpInstr::kFloat64x2WithX:
    case SimdOpInstr::kFloat64x2WithY: {
      // TODO(dartbug.com/30949) avoid transfer through memory.
      COMPILE_ASSERT(SimdOpInstr::kFloat64x2WithY ==
                     (SimdOpInstr::kFloat64x2WithX + 1));
      const intptr_t lane_index = instr->kind() - SimdOpInstr::kFloat64x2WithX;
      ASSERT(0 <= lane_index && lane_index < 2);

      __ SubImmediate(RSP, Immediate(kSimd128Size));
      __ movups(Address(RSP, 0), left);
      __ movsd(Address(RSP, lane_index * kDoubleSize), right);
      __ movups(left, Address(RSP, 0));
      __ AddImmediate(RSP, Immediate(kSimd128Size));
      break;
    }
    case SimdOpInstr::kFloat32x4WithX:
    case SimdOpInstr::kFloat32x4WithY:
    case SimdOpInstr::kFloat32x4WithZ:
    case SimdOpInstr::kFloat32x4WithW: {
      // TODO(dartbug.com/30949) avoid transfer through memory. SSE4.1 has
      // insertps. SSE2 these instructions can be implemented via a combination
      // of shufps/movss/movlhps.
      COMPILE_ASSERT(
          SimdOpInstr::kFloat32x4WithY == (SimdOpInstr::kFloat32x4WithX + 1) &&
          SimdOpInstr::kFloat32x4WithZ == (SimdOpInstr::kFloat32x4WithX + 2) &&
          SimdOpInstr::kFloat32x4WithW == (SimdOpInstr::kFloat32x4WithX + 3));
      const intptr_t lane_index = instr->kind() - SimdOpInstr::kFloat32x4WithX;
      ASSERT(0 <= lane_index && lane_index < 4);
      __ cvtsd2ss(left, left);
      __ SubImmediate(RSP, Immediate(kSimd128Size));
      __ movups(Address(RSP, 0), right);
      __ movss(Address(RSP, lane_index * kFloatSize), left);
      __ movups(left, Address(RSP, 0));
      __ AddImmediate(RSP, Immediate(kSimd128Size));
      break;
    }

    default:
      UNREACHABLE();
  }
}

#define SIMD_OP_SIMPLE_UNARY(V)                                                \
  SIMD_OP_FLOAT_ARITH(V, Sqrt, sqrt)                                           \
  SIMD_OP_FLOAT_ARITH(V, Negate, negate)                                       \
  SIMD_OP_FLOAT_ARITH(V, Abs, abs)                                             \
  V(Float32x4Reciprocal, reciprocalps)                                         \
  V(Float32x4ReciprocalSqrt, rsqrtps)

DEFINE_EMIT(SimdUnaryOp, (SameAsFirstInput, XmmRegister value)) {
  // TODO(dartbug.com/30949) select better register constraints to avoid
  // redundant move of input into a different register because all instructions
  // below support two operand forms.
  switch (instr->kind()) {
#define EMIT(Name, op)                                                         \
  case SimdOpInstr::k##Name:                                                   \
    __ op(value);                                                              \
    break;
    SIMD_OP_SIMPLE_UNARY(EMIT)
#undef EMIT
    case SimdOpInstr::kFloat32x4ShuffleX:
      // Shuffle not necessary.
      __ cvtss2sd(value, value);
      break;
    case SimdOpInstr::kFloat32x4ShuffleY:
      __ shufps(value, value, Immediate(0x55));
      __ cvtss2sd(value, value);
      break;
    case SimdOpInstr::kFloat32x4ShuffleZ:
      __ shufps(value, value, Immediate(0xAA));
      __ cvtss2sd(value, value);
      break;
    case SimdOpInstr::kFloat32x4ShuffleW:
      __ shufps(value, value, Immediate(0xFF));
      __ cvtss2sd(value, value);
      break;
    case SimdOpInstr::kFloat32x4Shuffle:
    case SimdOpInstr::kInt32x4Shuffle:
      __ shufps(value, value, Immediate(instr->mask()));
      break;
    case SimdOpInstr::kFloat32x4Splat:
      // Convert to Float32.
      __ cvtsd2ss(value, value);
      // Splat across all lanes.
      __ shufps(value, value, Immediate(0x00));
      break;
    case SimdOpInstr::kFloat32x4ToFloat64x2:
      __ cvtps2pd(value, value);
      break;
    case SimdOpInstr::kFloat64x2ToFloat32x4:
      __ cvtpd2ps(value, value);
      break;
    case SimdOpInstr::kInt32x4ToFloat32x4:
    case SimdOpInstr::kFloat32x4ToInt32x4:
      // TODO(dartbug.com/30949) these operations are essentially nop and should
      // not generate any code. They should be removed from the graph before
      // code generation.
      break;
    case SimdOpInstr::kFloat64x2GetX:
      // NOP.
      break;
    case SimdOpInstr::kFloat64x2GetY:
      __ shufpd(value, value, Immediate(0x33));
      break;
    case SimdOpInstr::kFloat64x2Splat:
      __ shufpd(value, value, Immediate(0x0));
      break;
    default:
      UNREACHABLE();
      break;
  }
}

DEFINE_EMIT(SimdGetSignMask, (Register out, XmmRegister value)) {
  switch (instr->kind()) {
    case SimdOpInstr::kFloat32x4GetSignMask:
    case SimdOpInstr::kInt32x4GetSignMask:
      __ movmskps(out, value);
      break;
    case SimdOpInstr::kFloat64x2GetSignMask:
      __ movmskpd(out, value);
      break;
    default:
      UNREACHABLE();
      break;
  }
}

DEFINE_EMIT(
    Float32x4Constructor,
    (SameAsFirstInput, XmmRegister v0, XmmRegister, XmmRegister, XmmRegister)) {
  // TODO(dartbug.com/30949) avoid transfer through memory. SSE4.1 has
  // insertps, with SSE2 this instruction can be implemented through unpcklps.
  const XmmRegister out = v0;
  __ SubImmediate(RSP, Immediate(kSimd128Size));
  for (intptr_t i = 0; i < 4; i++) {
    __ cvtsd2ss(out, instr->locs()->in(i).fpu_reg());
    __ movss(Address(RSP, i * kFloatSize), out);
  }
  __ movups(out, Address(RSP, 0));
  __ AddImmediate(RSP, Immediate(kSimd128Size));
}

DEFINE_EMIT(Float32x4Zero, (XmmRegister value)) {
  __ xorps(value, value);
}

DEFINE_EMIT(Float64x2Zero, (XmmRegister value)) {
  __ xorpd(value, value);
}

DEFINE_EMIT(Float32x4Clamp,
            (SameAsFirstInput,
             XmmRegister value,
             XmmRegister lower,
             XmmRegister upper)) {
  __ minps(value, upper);
  __ maxps(value, lower);
}

DEFINE_EMIT(Int32x4Constructor,
            (XmmRegister result, Register, Register, Register, Register)) {
  // TODO(dartbug.com/30949) avoid transfer through memory.
  __ SubImmediate(RSP, Immediate(kSimd128Size));
  for (intptr_t i = 0; i < 4; i++) {
    __ movl(Address(RSP, i * kInt32Size), instr->locs()->in(i).reg());
  }
  __ movups(result, Address(RSP, 0));
  __ AddImmediate(RSP, Immediate(kSimd128Size));
}

DEFINE_EMIT(Int32x4BoolConstructor,
            (XmmRegister result,
             Register,
             Register,
             Register,
             Register,
             Temp<Register> temp)) {
  // TODO(dartbug.com/30949) avoid transfer through memory.
  __ SubImmediate(RSP, Immediate(kSimd128Size));
  for (intptr_t i = 0; i < 4; i++) {
    Label done, load_false;
    __ xorq(temp, temp);
    __ CompareObject(instr->locs()->in(i).reg(), Bool::True());
    __ setcc(EQUAL, ByteRegisterOf(temp));
    __ negl(temp);  // temp = input ? -1 : 0
    __ movl(Address(RSP, kInt32Size * i), temp);
  }
  __ movups(result, Address(RSP, 0));
  __ AddImmediate(RSP, Immediate(kSimd128Size));
}

static void EmitToBoolean(FlowGraphCompiler* compiler, Register out) {
  ASSERT_BOOL_FALSE_FOLLOWS_BOOL_TRUE();
  __ testl(out, out);
  __ setcc(ZERO, ByteRegisterOf(out));
  __ movzxb(out, out);
  __ movq(out, Address(THR, out, TIMES_8, Thread::bool_true_offset()));
}

DEFINE_EMIT(Int32x4GetFlagZorW,
            (Register out, XmmRegister value, Temp<XmmRegister> temp)) {
  __ movhlps(temp, value);  // extract upper half.
  __ movq(out, temp);
  if (instr->kind() == SimdOpInstr::kInt32x4GetFlagW) {
    __ shrq(out, Immediate(32));  // extract upper 32bits.
  }
  EmitToBoolean(compiler, out);
}

DEFINE_EMIT(Int32x4GetFlagXorY, (Register out, XmmRegister value)) {
  __ movq(out, value);
  if (instr->kind() == SimdOpInstr::kInt32x4GetFlagY) {
    __ shrq(out, Immediate(32));  // extract upper 32bits.
  }
  EmitToBoolean(compiler, out);
}

DEFINE_EMIT(
    Int32x4WithFlag,
    (SameAsFirstInput, XmmRegister mask, Register flag, Temp<Register> temp)) {
  // TODO(dartbug.com/30949) avoid transfer through memory.
  COMPILE_ASSERT(
      SimdOpInstr::kInt32x4WithFlagY == (SimdOpInstr::kInt32x4WithFlagX + 1) &&
      SimdOpInstr::kInt32x4WithFlagZ == (SimdOpInstr::kInt32x4WithFlagX + 2) &&
      SimdOpInstr::kInt32x4WithFlagW == (SimdOpInstr::kInt32x4WithFlagX + 3));
  const intptr_t lane_index = instr->kind() - SimdOpInstr::kInt32x4WithFlagX;
  ASSERT(0 <= lane_index && lane_index < 4);
  __ SubImmediate(RSP, Immediate(kSimd128Size));
  __ movups(Address(RSP, 0), mask);

  // temp = flag == true ? -1 : 0
  __ xorq(temp, temp);
  __ CompareObject(flag, Bool::True());
  __ setcc(EQUAL, ByteRegisterOf(temp));
  __ negl(temp);

  __ movl(Address(RSP, lane_index * kInt32Size), temp);
  __ movups(mask, Address(RSP, 0));
  __ AddImmediate(RSP, Immediate(kSimd128Size));
}

DEFINE_EMIT(Int32x4Select,
            (SameAsFirstInput,
             XmmRegister mask,
             XmmRegister trueValue,
             XmmRegister falseValue,
             Temp<XmmRegister> temp)) {
  // Copy mask.
  __ movaps(temp, mask);
  // Invert it.
  __ notps(temp);
  // mask = mask & trueValue.
  __ andps(mask, trueValue);
  // temp = temp & falseValue.
  __ andps(temp, falseValue);
  // out = mask | temp.
  __ orps(mask, temp);
}

// Map SimdOpInstr::Kind-s to corresponding emit functions. Uses the following
// format:
//
//     CASE(OpA) CASE(OpB) ____(Emitter) - Emitter is used to emit OpA and OpB.
//     SIMPLE(OpA) - Emitter with name OpA is used to emit OpA.
//
#define SIMD_OP_VARIANTS(CASE, ____, SIMPLE)                                   \
  SIMD_OP_SIMPLE_BINARY(CASE)                                                  \
  CASE(Float32x4Scale)                                                         \
  CASE(Float32x4ShuffleMix)                                                    \
  CASE(Int32x4ShuffleMix)                                                      \
  CASE(Float64x2Constructor)                                                   \
  CASE(Float64x2Scale)                                                         \
  CASE(Float64x2WithX)                                                         \
  CASE(Float64x2WithY)                                                         \
  CASE(Float32x4WithX)                                                         \
  CASE(Float32x4WithY)                                                         \
  CASE(Float32x4WithZ)                                                         \
  CASE(Float32x4WithW)                                                         \
  ____(SimdBinaryOp)                                                           \
  SIMD_OP_SIMPLE_UNARY(CASE)                                                   \
  CASE(Float32x4ShuffleX)                                                      \
  CASE(Float32x4ShuffleY)                                                      \
  CASE(Float32x4ShuffleZ)                                                      \
  CASE(Float32x4ShuffleW)                                                      \
  CASE(Float32x4Shuffle)                                                       \
  CASE(Int32x4Shuffle)                                                         \
  CASE(Float32x4Splat)                                                         \
  CASE(Float32x4ToFloat64x2)                                                   \
  CASE(Float64x2ToFloat32x4)                                                   \
  CASE(Int32x4ToFloat32x4)                                                     \
  CASE(Float32x4ToInt32x4)                                                     \
  CASE(Float64x2GetX)                                                          \
  CASE(Float64x2GetY)                                                          \
  CASE(Float64x2Splat)                                                         \
  ____(SimdUnaryOp)                                                            \
  CASE(Float32x4GetSignMask)                                                   \
  CASE(Int32x4GetSignMask)                                                     \
  CASE(Float64x2GetSignMask)                                                   \
  ____(SimdGetSignMask)                                                        \
  SIMPLE(Float32x4Constructor)                                                 \
  SIMPLE(Int32x4Constructor)                                                   \
  SIMPLE(Int32x4BoolConstructor)                                               \
  SIMPLE(Float32x4Zero)                                                        \
  SIMPLE(Float64x2Zero)                                                        \
  SIMPLE(Float32x4Clamp)                                                       \
  CASE(Int32x4GetFlagX)                                                        \
  CASE(Int32x4GetFlagY)                                                        \
  ____(Int32x4GetFlagXorY)                                                     \
  CASE(Int32x4GetFlagZ)                                                        \
  CASE(Int32x4GetFlagW)                                                        \
  ____(Int32x4GetFlagZorW)                                                     \
  CASE(Int32x4WithFlagX)                                                       \
  CASE(Int32x4WithFlagY)                                                       \
  CASE(Int32x4WithFlagZ)                                                       \
  CASE(Int32x4WithFlagW)                                                       \
  ____(Int32x4WithFlag)                                                        \
  SIMPLE(Int32x4Select)

LocationSummary* SimdOpInstr::MakeLocationSummary(Zone* zone, bool opt) const {
  switch (kind()) {
#define CASE(Name, ...) case k##Name:
#define EMIT(Name)                                                             \
  return MakeLocationSummaryFromEmitter(zone, this, &Emit##Name);
#define SIMPLE(Name) CASE(Name) EMIT(Name)
    SIMD_OP_VARIANTS(CASE, EMIT, SIMPLE)
#undef CASE
#undef EMIT
#undef SIMPLE
    case kIllegalSimdOp:
      break;
  }
  UNREACHABLE();
  return NULL;
}

void SimdOpInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  switch (kind()) {
#define CASE(Name, ...) case k##Name:
#define EMIT(Name)                                                             \
  InvokeEmitter(compiler, this, &Emit##Name);                                  \
  break;
#define SIMPLE(Name) CASE(Name) EMIT(Name)
    SIMD_OP_VARIANTS(CASE, EMIT, SIMPLE)
#undef CASE
#undef EMIT
#undef SIMPLE
    case kIllegalSimdOp:
      UNREACHABLE();
      break;
  }
}

#undef DEFINE_EMIT

LocationSummary* MathUnaryInstr::MakeLocationSummary(Zone* zone,
                                                     bool opt) const {
  ASSERT((kind() == MathUnaryInstr::kSqrt) ||
         (kind() == MathUnaryInstr::kDoubleSquare));
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  if (kind() == MathUnaryInstr::kDoubleSquare) {
    summary->set_out(0, Location::SameAsFirstInput());
  } else {
    summary->set_out(0, Location::RequiresFpuRegister());
  }
  return summary;
}

void MathUnaryInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  if (kind() == MathUnaryInstr::kSqrt) {
    __ sqrtsd(locs()->out(0).fpu_reg(), locs()->in(0).fpu_reg());
  } else if (kind() == MathUnaryInstr::kDoubleSquare) {
    XmmRegister value_reg = locs()->in(0).fpu_reg();
    __ mulsd(value_reg, value_reg);
    ASSERT(value_reg == locs()->out(0).fpu_reg());
  } else {
    UNREACHABLE();
  }
}

LocationSummary* CaseInsensitiveCompareUC16Instr::MakeLocationSummary(
    Zone* zone,
    bool opt) const {
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, InputCount(), kNumTemps, LocationSummary::kCall);
  summary->set_in(0, Location::RegisterLocation(CallingConventions::kArg1Reg));
  summary->set_in(1, Location::RegisterLocation(CallingConventions::kArg2Reg));
  summary->set_in(2, Location::RegisterLocation(CallingConventions::kArg3Reg));
  summary->set_in(3, Location::RegisterLocation(CallingConventions::kArg4Reg));
  summary->set_out(0, Location::RegisterLocation(RAX));
  return summary;
}

void CaseInsensitiveCompareUC16Instr::EmitNativeCode(
    FlowGraphCompiler* compiler) {
  // Save RSP. R13 is chosen because it is callee saved so we do not need to
  // back it up before calling into the runtime.
  static const Register kSavedSPReg = R13;
  __ movq(kSavedSPReg, RSP);
  __ ReserveAlignedFrameSpace(0);

  // Call the function. Parameters are already in their correct spots.
  __ CallRuntime(TargetFunction(), TargetFunction().argument_count());

  // Restore RSP.
  __ movq(RSP, kSavedSPReg);
}

LocationSummary* UnarySmiOpInstr::MakeLocationSummary(Zone* zone,
                                                      bool opt) const {
  const intptr_t kNumInputs = 1;
  return LocationSummary::Make(zone, kNumInputs, Location::SameAsFirstInput(),
                               LocationSummary::kNoCall);
}

void UnarySmiOpInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register value = locs()->in(0).reg();
  ASSERT(value == locs()->out(0).reg());
  switch (op_kind()) {
    case Token::kNEGATE: {
      Label* deopt = compiler->AddDeoptStub(deopt_id(), ICData::kDeoptUnaryOp);
      __ negq(value);
      __ j(OVERFLOW, deopt);
      break;
    }
    case Token::kBIT_NOT:
      __ notq(value);
      // Remove inverted smi-tag.
      __ AndImmediate(value, Immediate(~kSmiTagMask));
      break;
    default:
      UNREACHABLE();
  }
}

LocationSummary* UnaryDoubleOpInstr::MakeLocationSummary(Zone* zone,
                                                         bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_out(0, Location::SameAsFirstInput());
  return summary;
}

void UnaryDoubleOpInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister value = locs()->in(0).fpu_reg();
  ASSERT(locs()->out(0).fpu_reg() == value);
  __ DoubleNegate(value);
}

LocationSummary* MathMinMaxInstr::MakeLocationSummary(Zone* zone,
                                                      bool opt) const {
  if (result_cid() == kDoubleCid) {
    const intptr_t kNumInputs = 2;
    const intptr_t kNumTemps = 1;
    LocationSummary* summary = new (zone)
        LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
    summary->set_in(0, Location::RequiresFpuRegister());
    summary->set_in(1, Location::RequiresFpuRegister());
    // Reuse the left register so that code can be made shorter.
    summary->set_out(0, Location::SameAsFirstInput());
    summary->set_temp(0, Location::RequiresRegister());
    return summary;
  }
  ASSERT(result_cid() == kSmiCid);
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  summary->set_in(1, Location::RequiresRegister());
  // Reuse the left register so that code can be made shorter.
  summary->set_out(0, Location::SameAsFirstInput());
  return summary;
}

void MathMinMaxInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  ASSERT((op_kind() == MethodRecognizer::kMathMin) ||
         (op_kind() == MethodRecognizer::kMathMax));
  const intptr_t is_min = (op_kind() == MethodRecognizer::kMathMin);
  if (result_cid() == kDoubleCid) {
    Label done, returns_nan, are_equal;
    XmmRegister left = locs()->in(0).fpu_reg();
    XmmRegister right = locs()->in(1).fpu_reg();
    XmmRegister result = locs()->out(0).fpu_reg();
    Register temp = locs()->temp(0).reg();
    __ comisd(left, right);
    __ j(PARITY_EVEN, &returns_nan, Assembler::kNearJump);
    __ j(EQUAL, &are_equal, Assembler::kNearJump);
    const Condition double_condition =
        is_min ? TokenKindToDoubleCondition(Token::kLT)
               : TokenKindToDoubleCondition(Token::kGT);
    ASSERT(left == result);
    __ j(double_condition, &done, Assembler::kNearJump);
    __ movsd(result, right);
    __ jmp(&done, Assembler::kNearJump);

    __ Bind(&returns_nan);
    static double kNaN = NAN;
    __ LoadImmediate(temp, Immediate(reinterpret_cast<intptr_t>(&kNaN)));
    __ movsd(result, Address(temp, 0));
    __ jmp(&done, Assembler::kNearJump);

    __ Bind(&are_equal);
    Label left_is_negative;
    // Check for negative zero: -0.0 is equal 0.0 but min or max must return
    // -0.0 or 0.0 respectively.
    // Check for negative left value (get the sign bit):
    // - min -> left is negative ? left : right.
    // - max -> left is negative ? right : left
    // Check the sign bit.
    __ movmskpd(temp, left);
    __ testq(temp, Immediate(1));
    if (is_min) {
      ASSERT(left == result);
      __ j(NOT_ZERO, &done, Assembler::kNearJump);  // Negative -> return left.
    } else {
      ASSERT(left == result);
      __ j(ZERO, &done, Assembler::kNearJump);  // Positive -> return left.
    }
    __ movsd(result, right);
    __ Bind(&done);
    return;
  }

  ASSERT(result_cid() == kSmiCid);
  Register left = locs()->in(0).reg();
  Register right = locs()->in(1).reg();
  Register result = locs()->out(0).reg();
  __ cmpq(left, right);
  ASSERT(result == left);
  if (is_min) {
    __ cmovgeq(result, right);
  } else {
    __ cmovlessq(result, right);
  }
}

LocationSummary* Int32ToDoubleInstr::MakeLocationSummary(Zone* zone,
                                                         bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* result = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  result->set_in(0, Location::RequiresRegister());
  result->set_out(0, Location::RequiresFpuRegister());
  return result;
}

void Int32ToDoubleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register value = locs()->in(0).reg();
  FpuRegister result = locs()->out(0).fpu_reg();
  __ cvtsi2sdl(result, value);
}

LocationSummary* SmiToDoubleInstr::MakeLocationSummary(Zone* zone,
                                                       bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* result = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  result->set_in(0, Location::WritableRegister());
  result->set_out(0, Location::RequiresFpuRegister());
  return result;
}

void SmiToDoubleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register value = locs()->in(0).reg();
  FpuRegister result = locs()->out(0).fpu_reg();
  __ SmiUntag(value);
  __ cvtsi2sdq(result, value);
}

LocationSummary* MintToDoubleInstr::MakeLocationSummary(Zone* zone,
                                                        bool opt) const {
  UNIMPLEMENTED();
  return NULL;
}

void MintToDoubleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  UNIMPLEMENTED();
}

LocationSummary* DoubleToIntegerInstr::MakeLocationSummary(Zone* zone,
                                                           bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 1;
  LocationSummary* result = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kCall);
  result->set_in(0, Location::RegisterLocation(RCX));
  result->set_out(0, Location::RegisterLocation(RAX));
  result->set_temp(0, Location::RegisterLocation(RBX));
  return result;
}

void DoubleToIntegerInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register result = locs()->out(0).reg();
  Register value_obj = locs()->in(0).reg();
  Register temp = locs()->temp(0).reg();
  XmmRegister value_double = XMM0;
  ASSERT(result == RAX);
  ASSERT(result != value_obj);
  ASSERT(result != temp);
  __ movsd(value_double, FieldAddress(value_obj, Double::value_offset()));
  __ cvttsd2siq(result, value_double);
  // Overflow is signalled with minint.
  Label do_call, done;
  // Check for overflow and that it fits into Smi.
  __ movq(temp, result);
  __ shlq(temp, Immediate(1));
  __ j(OVERFLOW, &do_call, Assembler::kNearJump);
  __ SmiTag(result);
  __ jmp(&done);
  __ Bind(&do_call);
  __ pushq(value_obj);
  ASSERT(instance_call()->HasICData());
  const ICData& ic_data = *instance_call()->ic_data();
  ASSERT(ic_data.NumberOfChecksIs(1));
  const Function& target = Function::ZoneHandle(ic_data.GetTargetAt(0));
  const int kTypeArgsLen = 0;
  const int kNumberOfArguments = 1;
  const Array& kNoArgumentNames = Object::null_array();
  ArgumentsInfo args_info(kTypeArgsLen, kNumberOfArguments, kNoArgumentNames);
  compiler->GenerateStaticCall(deopt_id(), instance_call()->token_pos(), target,
                               args_info, locs(), ICData::Handle(),
                               ICData::kStatic);
  __ Bind(&done);
}

LocationSummary* DoubleToSmiInstr::MakeLocationSummary(Zone* zone,
                                                       bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 1;
  LocationSummary* result = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  result->set_in(0, Location::RequiresFpuRegister());
  result->set_out(0, Location::RequiresRegister());
  result->set_temp(0, Location::RequiresRegister());
  return result;
}

void DoubleToSmiInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Label* deopt = compiler->AddDeoptStub(deopt_id(), ICData::kDeoptDoubleToSmi);
  Register result = locs()->out(0).reg();
  XmmRegister value = locs()->in(0).fpu_reg();
  Register temp = locs()->temp(0).reg();

  __ cvttsd2siq(result, value);
  // Overflow is signalled with minint.
  Label do_call, done;
  // Check for overflow and that it fits into Smi.
  __ movq(temp, result);
  __ shlq(temp, Immediate(1));
  __ j(OVERFLOW, deopt);
  __ SmiTag(result);
}

LocationSummary* DoubleToDoubleInstr::MakeLocationSummary(Zone* zone,
                                                          bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* result = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  result->set_in(0, Location::RequiresFpuRegister());
  result->set_out(0, Location::RequiresFpuRegister());
  return result;
}

void DoubleToDoubleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister value = locs()->in(0).fpu_reg();
  XmmRegister result = locs()->out(0).fpu_reg();
  switch (recognized_kind()) {
    case MethodRecognizer::kDoubleTruncate:
      __ roundsd(result, value, Assembler::kRoundToZero);
      break;
    case MethodRecognizer::kDoubleFloor:
      __ roundsd(result, value, Assembler::kRoundDown);
      break;
    case MethodRecognizer::kDoubleCeil:
      __ roundsd(result, value, Assembler::kRoundUp);
      break;
    default:
      UNREACHABLE();
  }
}

LocationSummary* DoubleToFloatInstr::MakeLocationSummary(Zone* zone,
                                                         bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* result = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  result->set_in(0, Location::RequiresFpuRegister());
  result->set_out(0, Location::SameAsFirstInput());
  return result;
}

void DoubleToFloatInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  __ cvtsd2ss(locs()->out(0).fpu_reg(), locs()->in(0).fpu_reg());
}

LocationSummary* FloatToDoubleInstr::MakeLocationSummary(Zone* zone,
                                                         bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* result = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  result->set_in(0, Location::RequiresFpuRegister());
  result->set_out(0, Location::SameAsFirstInput());
  return result;
}

void FloatToDoubleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  __ cvtss2sd(locs()->out(0).fpu_reg(), locs()->in(0).fpu_reg());
}

LocationSummary* InvokeMathCFunctionInstr::MakeLocationSummary(Zone* zone,
                                                               bool opt) const {
  // Calling convention on x64 uses XMM0 and XMM1 to pass the first two
  // double arguments and XMM0 to return the result. Unfortunately
  // currently we can't specify these registers because ParallelMoveResolver
  // assumes that XMM0 is free at all times.
  // TODO(vegorov): allow XMM0 to be used.
  ASSERT((InputCount() == 1) || (InputCount() == 2));
  const intptr_t kNumTemps =
      (recognized_kind() == MethodRecognizer::kMathDoublePow) ? 3 : 1;
  LocationSummary* result = new (zone)
      LocationSummary(zone, InputCount(), kNumTemps, LocationSummary::kCall);
  ASSERT(R13 != CALLEE_SAVED_TEMP);
  ASSERT(((1 << R13) & CallingConventions::kCalleeSaveCpuRegisters) != 0);
  result->set_temp(0, Location::RegisterLocation(R13));
  result->set_in(0, Location::FpuRegisterLocation(XMM2));
  if (InputCount() == 2) {
    result->set_in(1, Location::FpuRegisterLocation(XMM1));
  }
  if (recognized_kind() == MethodRecognizer::kMathDoublePow) {
    // Temp index 1.
    result->set_temp(1, Location::RegisterLocation(RAX));
    // Temp index 2.
    result->set_temp(2, Location::FpuRegisterLocation(XMM4));
  }
  result->set_out(0, Location::FpuRegisterLocation(XMM3));
  return result;
}

// Pseudo code:
// if (exponent == 0.0) return 1.0;
// // Speed up simple cases.
// if (exponent == 1.0) return base;
// if (exponent == 2.0) return base * base;
// if (exponent == 3.0) return base * base * base;
// if (base == 1.0) return 1.0;
// if (base.isNaN || exponent.isNaN) {
//    return double.NAN;
// }
// if (base != -Infinity && exponent == 0.5) {
//   if (base == 0.0) return 0.0;
//   return sqrt(value);
// }
// TODO(srdjan): Move into a stub?
static void InvokeDoublePow(FlowGraphCompiler* compiler,
                            InvokeMathCFunctionInstr* instr) {
  ASSERT(instr->recognized_kind() == MethodRecognizer::kMathDoublePow);
  const intptr_t kInputCount = 2;
  ASSERT(instr->InputCount() == kInputCount);
  LocationSummary* locs = instr->locs();

  XmmRegister base = locs->in(0).fpu_reg();
  XmmRegister exp = locs->in(1).fpu_reg();
  XmmRegister result = locs->out(0).fpu_reg();
  Register temp = locs->temp(InvokeMathCFunctionInstr::kObjectTempIndex).reg();
  XmmRegister zero_temp =
      locs->temp(InvokeMathCFunctionInstr::kDoubleTempIndex).fpu_reg();

  __ xorps(zero_temp, zero_temp);
  __ LoadObject(temp, Double::ZoneHandle(Double::NewCanonical(1)));
  __ movsd(result, FieldAddress(temp, Double::value_offset()));

  Label check_base, skip_call;
  // exponent == 0.0 -> return 1.0;
  __ comisd(exp, zero_temp);
  __ j(PARITY_EVEN, &check_base, Assembler::kNearJump);
  __ j(EQUAL, &skip_call);  // 'result' is 1.0.

  // exponent == 1.0 ?
  __ comisd(exp, result);
  Label return_base;
  __ j(EQUAL, &return_base, Assembler::kNearJump);

  // exponent == 2.0 ?
  __ LoadObject(temp, Double::ZoneHandle(Double::NewCanonical(2.0)));
  __ movsd(XMM0, FieldAddress(temp, Double::value_offset()));
  __ comisd(exp, XMM0);
  Label return_base_times_2;
  __ j(EQUAL, &return_base_times_2, Assembler::kNearJump);

  // exponent == 3.0 ?
  __ LoadObject(temp, Double::ZoneHandle(Double::NewCanonical(3.0)));
  __ movsd(XMM0, FieldAddress(temp, Double::value_offset()));
  __ comisd(exp, XMM0);
  __ j(NOT_EQUAL, &check_base);

  // Base times 3.
  __ movsd(result, base);
  __ mulsd(result, base);
  __ mulsd(result, base);
  __ jmp(&skip_call);

  __ Bind(&return_base);
  __ movsd(result, base);
  __ jmp(&skip_call);

  __ Bind(&return_base_times_2);
  __ movsd(result, base);
  __ mulsd(result, base);
  __ jmp(&skip_call);

  __ Bind(&check_base);
  // Note: 'exp' could be NaN.

  Label return_nan;
  // base == 1.0 -> return 1.0;
  __ comisd(base, result);
  __ j(PARITY_EVEN, &return_nan, Assembler::kNearJump);
  __ j(EQUAL, &skip_call, Assembler::kNearJump);
  // Note: 'base' could be NaN.
  __ comisd(exp, base);
  // Neither 'exp' nor 'base' is NaN.
  Label try_sqrt;
  __ j(PARITY_ODD, &try_sqrt, Assembler::kNearJump);
  // Return NaN.
  __ Bind(&return_nan);
  __ LoadObject(temp, Double::ZoneHandle(Double::NewCanonical(NAN)));
  __ movsd(result, FieldAddress(temp, Double::value_offset()));
  __ jmp(&skip_call);

  Label do_pow, return_zero;
  __ Bind(&try_sqrt);
  // Before calling pow, check if we could use sqrt instead of pow.
  __ LoadObject(temp, Double::ZoneHandle(Double::NewCanonical(kNegInfinity)));
  __ movsd(result, FieldAddress(temp, Double::value_offset()));
  // base == -Infinity -> call pow;
  __ comisd(base, result);
  __ j(EQUAL, &do_pow, Assembler::kNearJump);

  // exponent == 0.5 ?
  __ LoadObject(temp, Double::ZoneHandle(Double::NewCanonical(0.5)));
  __ movsd(result, FieldAddress(temp, Double::value_offset()));
  __ comisd(exp, result);
  __ j(NOT_EQUAL, &do_pow, Assembler::kNearJump);

  // base == 0 -> return 0;
  __ comisd(base, zero_temp);
  __ j(EQUAL, &return_zero, Assembler::kNearJump);

  __ sqrtsd(result, base);
  __ jmp(&skip_call, Assembler::kNearJump);

  __ Bind(&return_zero);
  __ movsd(result, zero_temp);
  __ jmp(&skip_call);

  __ Bind(&do_pow);

  // Save RSP.
  __ movq(locs->temp(InvokeMathCFunctionInstr::kSavedSpTempIndex).reg(), RSP);
  __ ReserveAlignedFrameSpace(0);
  __ movaps(XMM0, locs->in(0).fpu_reg());
  ASSERT(locs->in(1).fpu_reg() == XMM1);

  __ CallRuntime(instr->TargetFunction(), kInputCount);
  __ movaps(locs->out(0).fpu_reg(), XMM0);
  // Restore RSP.
  __ movq(RSP, locs->temp(InvokeMathCFunctionInstr::kSavedSpTempIndex).reg());
  __ Bind(&skip_call);
}

void InvokeMathCFunctionInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  if (recognized_kind() == MethodRecognizer::kMathDoublePow) {
    InvokeDoublePow(compiler, this);
    return;
  }
  // Save RSP.
  __ movq(locs()->temp(kSavedSpTempIndex).reg(), RSP);
  __ ReserveAlignedFrameSpace(0);
  __ movaps(XMM0, locs()->in(0).fpu_reg());
  if (InputCount() == 2) {
    ASSERT(locs()->in(1).fpu_reg() == XMM1);
  }

  __ CallRuntime(TargetFunction(), InputCount());
  __ movaps(locs()->out(0).fpu_reg(), XMM0);
  // Restore RSP.
  __ movq(RSP, locs()->temp(kSavedSpTempIndex).reg());
}

LocationSummary* ExtractNthOutputInstr::MakeLocationSummary(Zone* zone,
                                                            bool opt) const {
  // Only use this instruction in optimized code.
  ASSERT(opt);
  const intptr_t kNumInputs = 1;
  LocationSummary* summary =
      new (zone) LocationSummary(zone, kNumInputs, 0, LocationSummary::kNoCall);
  if (representation() == kUnboxedDouble) {
    if (index() == 0) {
      summary->set_in(
          0, Location::Pair(Location::RequiresFpuRegister(), Location::Any()));
    } else {
      ASSERT(index() == 1);
      summary->set_in(
          0, Location::Pair(Location::Any(), Location::RequiresFpuRegister()));
    }
    summary->set_out(0, Location::RequiresFpuRegister());
  } else {
    ASSERT(representation() == kTagged);
    if (index() == 0) {
      summary->set_in(
          0, Location::Pair(Location::RequiresRegister(), Location::Any()));
    } else {
      ASSERT(index() == 1);
      summary->set_in(
          0, Location::Pair(Location::Any(), Location::RequiresRegister()));
    }
    summary->set_out(0, Location::RequiresRegister());
  }
  return summary;
}

void ExtractNthOutputInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  ASSERT(locs()->in(0).IsPairLocation());
  PairLocation* pair = locs()->in(0).AsPairLocation();
  Location in_loc = pair->At(index());
  if (representation() == kUnboxedDouble) {
    XmmRegister out = locs()->out(0).fpu_reg();
    XmmRegister in = in_loc.fpu_reg();
    __ movaps(out, in);
  } else {
    ASSERT(representation() == kTagged);
    Register out = locs()->out(0).reg();
    Register in = in_loc.reg();
    __ movq(out, in);
  }
}

LocationSummary* TruncDivModInstr::MakeLocationSummary(Zone* zone,
                                                       bool opt) const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  // Both inputs must be writable because they will be untagged.
  summary->set_in(0, Location::RegisterLocation(RAX));
  summary->set_in(1, Location::WritableRegister());
  summary->set_out(0, Location::Pair(Location::RegisterLocation(RAX),
                                     Location::RegisterLocation(RDX)));
  return summary;
}

void TruncDivModInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  ASSERT(CanDeoptimize());
  Label* deopt = compiler->AddDeoptStub(deopt_id(), ICData::kDeoptBinarySmiOp);
  Register left = locs()->in(0).reg();
  Register right = locs()->in(1).reg();
  ASSERT(locs()->out(0).IsPairLocation());
  PairLocation* pair = locs()->out(0).AsPairLocation();
  Register result1 = pair->At(0).reg();
  Register result2 = pair->At(1).reg();
  Label not_32bit, done;
  Register temp = RDX;
  ASSERT(left == RAX);
  ASSERT((right != RDX) && (right != RAX));
  ASSERT(result1 == RAX);
  ASSERT(result2 == RDX);
  if (RangeUtils::CanBeZero(divisor_range())) {
    // Handle divide by zero in runtime.
    __ testq(right, right);
    __ j(ZERO, deopt);
  }
  // Check if both operands fit into 32bits as idiv with 64bit operands
  // requires twice as many cycles and has much higher latency.
  // We are checking this before untagging them to avoid corner case
  // dividing INT_MAX by -1 that raises exception because quotient is
  // too large for 32bit register.
  __ movsxd(temp, left);
  __ cmpq(temp, left);
  __ j(NOT_EQUAL, &not_32bit);
  __ movsxd(temp, right);
  __ cmpq(temp, right);
  __ j(NOT_EQUAL, &not_32bit);

  // Both operands are 31bit smis. Divide using 32bit idiv.
  __ SmiUntag(left);
  __ SmiUntag(right);
  __ cdq();
  __ idivl(right);
  __ movsxd(RAX, RAX);
  __ movsxd(RDX, RDX);
  __ jmp(&done);

  // Divide using 64bit idiv.
  __ Bind(&not_32bit);
  __ SmiUntag(left);
  __ SmiUntag(right);
  __ cqo();         // Sign extend RAX -> RDX:RAX.
  __ idivq(right);  //  RAX: quotient, RDX: remainder.
  // Check the corner case of dividing the 'MIN_SMI' with -1, in which
  // case we cannot tag the result.
  __ CompareImmediate(RAX, Immediate(0x4000000000000000));
  __ j(EQUAL, deopt);
  __ Bind(&done);

  // Modulo correction (RDX).
  //  res = left % right;
  //  if (res < 0) {
  //    if (right < 0) {
  //      res = res - right;
  //    } else {
  //      res = res + right;
  //    }
  //  }
  Label all_done;
  __ cmpq(RDX, Immediate(0));
  __ j(GREATER_EQUAL, &all_done, Assembler::kNearJump);
  // Result is negative, adjust it.
  if ((divisor_range() == NULL) || divisor_range()->Overlaps(-1, 1)) {
    Label subtract;
    __ cmpq(right, Immediate(0));
    __ j(LESS, &subtract, Assembler::kNearJump);
    __ addq(RDX, right);
    __ jmp(&all_done, Assembler::kNearJump);
    __ Bind(&subtract);
    __ subq(RDX, right);
  } else if (divisor_range()->IsPositive()) {
    // Right is positive.
    __ addq(RDX, right);
  } else {
    // Right is negative.
    __ subq(RDX, right);
  }
  __ Bind(&all_done);

  __ SmiTag(RAX);
  __ SmiTag(RDX);
  // Note that the result of an integer division/modulo of two
  // in-range arguments, cannot create out-of-range result.
}

LocationSummary* PolymorphicInstanceCallInstr::MakeLocationSummary(
    Zone* zone,
    bool opt) const {
  return MakeCallSummary(zone);
}

LocationSummary* BranchInstr::MakeLocationSummary(Zone* zone, bool opt) const {
  comparison()->InitializeLocationSummary(zone, opt);
  // Branches don't produce a result.
  comparison()->locs()->set_out(0, Location::NoLocation());
  return comparison()->locs();
}

void BranchInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  comparison()->EmitBranchCode(compiler, this);
}

LocationSummary* CheckClassInstr::MakeLocationSummary(Zone* zone,
                                                      bool opt) const {
  const intptr_t kNumInputs = 1;
  const bool need_mask_temp = IsBitTest();
  const intptr_t kNumTemps = !IsNullCheck() ? (need_mask_temp ? 2 : 1) : 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  if (!IsNullCheck()) {
    summary->set_temp(0, Location::RequiresRegister());
    if (need_mask_temp) {
      summary->set_temp(1, Location::RequiresRegister());
    }
  }
  return summary;
}

void CheckClassInstr::EmitNullCheck(FlowGraphCompiler* compiler, Label* deopt) {
  __ CompareObject(locs()->in(0).reg(), Object::null_object());
  Condition cond = IsDeoptIfNull() ? EQUAL : NOT_EQUAL;
  __ j(cond, deopt);
}

void CheckClassInstr::EmitBitTest(FlowGraphCompiler* compiler,
                                  intptr_t min,
                                  intptr_t max,
                                  intptr_t mask,
                                  Label* deopt) {
  Register biased_cid = locs()->temp(0).reg();
  __ subq(biased_cid, Immediate(min));
  __ cmpq(biased_cid, Immediate(max - min));
  __ j(ABOVE, deopt);

  Register mask_reg = locs()->temp(1).reg();
  __ movq(mask_reg, Immediate(mask));
  __ btq(mask_reg, biased_cid);
  __ j(NOT_CARRY, deopt);
}

int CheckClassInstr::EmitCheckCid(FlowGraphCompiler* compiler,
                                  int bias,
                                  intptr_t cid_start,
                                  intptr_t cid_end,
                                  bool is_last,
                                  Label* is_ok,
                                  Label* deopt,
                                  bool use_near_jump) {
  Register biased_cid = locs()->temp(0).reg();
  Condition no_match, match;
  if (cid_start == cid_end) {
    __ cmpl(biased_cid, Immediate(cid_start - bias));
    no_match = NOT_EQUAL;
    match = EQUAL;
  } else {
    // For class ID ranges use a subtract followed by an unsigned
    // comparison to check both ends of the ranges with one comparison.
    __ addl(biased_cid, Immediate(bias - cid_start));
    bias = cid_start;
    __ cmpl(biased_cid, Immediate(cid_end - cid_start));
    no_match = ABOVE;
    match = BELOW_EQUAL;
  }

  if (is_last) {
    __ j(no_match, deopt);
  } else {
    if (use_near_jump) {
      __ j(match, is_ok, Assembler::kNearJump);
    } else {
      __ j(match, is_ok);
    }
  }
  return bias;
}

LocationSummary* CheckSmiInstr::MakeLocationSummary(Zone* zone,
                                                    bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  return summary;
}

void CheckSmiInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register value = locs()->in(0).reg();
  Label* deopt = compiler->AddDeoptStub(deopt_id(), ICData::kDeoptCheckSmi,
                                        licm_hoisted_ ? ICData::kHoisted : 0);
  __ BranchIfNotSmi(value, deopt);
}

LocationSummary* CheckClassIdInstr::MakeLocationSummary(Zone* zone,
                                                        bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, cids_.IsSingleCid() ? Location::RequiresRegister()
                                         : Location::WritableRegister());
  return summary;
}

void CheckClassIdInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register value = locs()->in(0).reg();
  Label* deopt = compiler->AddDeoptStub(deopt_id(), ICData::kDeoptCheckClass);
  if (cids_.IsSingleCid()) {
    __ CompareImmediate(value, Immediate(Smi::RawValue(cids_.cid_start)));
    __ j(NOT_ZERO, deopt);
  } else {
    __ AddImmediate(value, Immediate(-Smi::RawValue(cids_.cid_start)));
    __ cmpq(value, Immediate(Smi::RawValue(cids_.Extent())));
    __ j(ABOVE, deopt);
  }
}

LocationSummary* CheckArrayBoundInstr::MakeLocationSummary(Zone* zone,
                                                           bool opt) const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  locs->set_in(kLengthPos, Location::RegisterOrSmiConstant(length()));
  locs->set_in(kIndexPos, Location::RegisterOrSmiConstant(index()));
  return locs;
}

void CheckArrayBoundInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  uint32_t flags = generalized_ ? ICData::kGeneralized : 0;
  flags |= licm_hoisted_ ? ICData::kHoisted : 0;
  Label* deopt =
      compiler->AddDeoptStub(deopt_id(), ICData::kDeoptCheckArrayBound, flags);

  Location length_loc = locs()->in(kLengthPos);
  Location index_loc = locs()->in(kIndexPos);

  if (length_loc.IsConstant() && index_loc.IsConstant()) {
    ASSERT((Smi::Cast(length_loc.constant()).Value() <=
            Smi::Cast(index_loc.constant()).Value()) ||
           (Smi::Cast(index_loc.constant()).Value() < 0));
    // Unconditionally deoptimize for constant bounds checks because they
    // only occur only when index is out-of-bounds.
    __ jmp(deopt);
    return;
  }

  const intptr_t index_cid = index()->Type()->ToCid();
  if (index_loc.IsConstant()) {
    Register length = length_loc.reg();
    const Smi& index = Smi::Cast(index_loc.constant());
    __ CompareImmediate(length,
                        Immediate(reinterpret_cast<int64_t>(index.raw())));
    __ j(BELOW_EQUAL, deopt);
  } else if (length_loc.IsConstant()) {
    const Smi& length = Smi::Cast(length_loc.constant());
    Register index = index_loc.reg();
    if (index_cid != kSmiCid) {
      __ BranchIfNotSmi(index, deopt);
    }
    if (length.Value() == Smi::kMaxValue) {
      __ testq(index, index);
      __ j(NEGATIVE, deopt);
    } else {
      __ CompareImmediate(index,
                          Immediate(reinterpret_cast<int64_t>(length.raw())));
      __ j(ABOVE_EQUAL, deopt);
    }
  } else {
    Register length = length_loc.reg();
    Register index = index_loc.reg();
    if (index_cid != kSmiCid) {
      __ BranchIfNotSmi(index, deopt);
    }
    __ cmpq(index, length);
    __ j(ABOVE_EQUAL, deopt);
  }
}

template <typename OperandType>
static void EmitInt64Arithmetic(FlowGraphCompiler* compiler,
                                Token::Kind op_kind,
                                Register left,
                                const OperandType& right,
                                Label* deopt) {
  switch (op_kind) {
    case Token::kADD:
      __ addq(left, right);
      break;
    case Token::kSUB:
      __ subq(left, right);
      break;
    case Token::kBIT_AND:
      __ andq(left, right);
      break;
    case Token::kBIT_OR:
      __ orq(left, right);
      break;
    case Token::kBIT_XOR:
      __ xorq(left, right);
      break;
    case Token::kMUL:
      __ imulq(left, right);
      break;
    default:
      UNREACHABLE();
  }
  if (deopt != NULL) __ j(OVERFLOW, deopt);
}

LocationSummary* BinaryInt64OpInstr::MakeLocationSummary(Zone* zone,
                                                         bool opt) const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  summary->set_in(1, Location::RequiresRegister());
  summary->set_out(0, Location::SameAsFirstInput());
  return summary;
}

void BinaryInt64OpInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  const Register left = locs()->in(0).reg();
  const Register right = locs()->in(1).reg();
  const Register out = locs()->out(0).reg();

  ASSERT(out == left);

  Label* deopt = NULL;
  if (CanDeoptimize()) {
    deopt = compiler->AddDeoptStub(deopt_id(), ICData::kDeoptBinaryInt64Op);
  }

  EmitInt64Arithmetic(compiler, op_kind(), left, right, deopt);
}

LocationSummary* UnaryInt64OpInstr::MakeLocationSummary(Zone* zone,
                                                        bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  summary->set_out(0, Location::SameAsFirstInput());
  return summary;
}

void UnaryInt64OpInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  ASSERT(op_kind() == Token::kBIT_NOT);
  const Register left = locs()->in(0).reg();
  const Register out = locs()->out(0).reg();
  ASSERT(out == left);
  __ notq(left);
}

LocationSummary* ShiftInt64OpInstr::MakeLocationSummary(Zone* zone,
                                                        bool opt) const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = can_overflow() ? 1 : 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  summary->set_in(1, Location::FixedRegisterOrSmiConstant(right(), RCX));
  if (kNumTemps > 0) {
    summary->set_temp(0, Location::RequiresRegister());
  }
  summary->set_out(0, Location::SameAsFirstInput());
  return summary;
}

void ShiftInt64OpInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  const Register left = locs()->in(0).reg();
  const Register out = locs()->out(0).reg();
  ASSERT(left == out);

  Label* deopt = NULL;
  if (CanDeoptimize()) {
    deopt = compiler->AddDeoptStub(deopt_id(), ICData::kDeoptBinaryInt64Op);
  }
  if (locs()->in(1).IsConstant()) {
    // Code for a constant shift amount.
    ASSERT(locs()->in(1).constant().IsSmi());
    const int64_t shift = Smi::Cast(locs()->in(1).constant()).Value();
    ASSERT(shift >= 0);
    switch (op_kind()) {
      case Token::kSHR:
        __ sarq(left,
                Immediate(Utils::Minimum<int64_t>(shift, kBitsPerWord - 1)));
        break;
      case Token::kSHL: {
        ASSERT(shift < 64);
        if (can_overflow()) {
          // Check for overflow.
          Register temp = locs()->temp(0).reg();
          __ movq(temp, left);
          __ shlq(left, Immediate(shift));
          __ sarq(left, Immediate(shift));
          __ cmpq(left, temp);
          __ j(NOT_EQUAL, deopt);  // Overflow.
        }
        // Shift for result now we know there is no overflow.
        __ shlq(left, Immediate(shift));
        break;
      }
      default:
        UNREACHABLE();
    }
  } else {
    // Code for a variable shift amount.
    // Deoptimize if shift count is > 63 or negative.
    // Sarq and shlq instructions mask the count to 6 bits.
    __ SmiUntag(RCX);
    if (!IsShiftCountInRange()) {
      __ cmpq(RCX, Immediate(kMintShiftCountLimit));
      __ j(ABOVE, deopt);
    }
    Label done, large_shift;
    switch (op_kind()) {
      case Token::kSHR: {
        __ sarq(left, RCX);
        break;
      }
      case Token::kSHL: {
        if (can_overflow()) {
          // Check for overflow.
          Register temp = locs()->temp(0).reg();
          __ movq(temp, left);
          __ shlq(left, RCX);
          __ sarq(left, RCX);
          __ cmpq(left, temp);
          __ j(NOT_EQUAL, deopt);  // Overflow.
        }
        // Shift for result now we know there is no overflow.
        __ shlq(left, RCX);
        break;
      }
      default:
        UNREACHABLE();
    }
  }
}

CompileType BinaryUint32OpInstr::ComputeType() const {
  return CompileType::FromCid(kSmiCid);
}

CompileType ShiftUint32OpInstr::ComputeType() const {
  return CompileType::FromCid(kSmiCid);
}

CompileType UnaryUint32OpInstr::ComputeType() const {
  return CompileType::FromCid(kSmiCid);
}

LocationSummary* BinaryUint32OpInstr::MakeLocationSummary(Zone* zone,
                                                          bool opt) const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  summary->set_in(1, Location::RequiresRegister());
  summary->set_out(0, Location::SameAsFirstInput());
  return summary;
}

template <typename OperandType>
static void EmitIntegerArithmetic(FlowGraphCompiler* compiler,
                                  Token::Kind op_kind,
                                  Register left,
                                  const OperandType& right,
                                  Label* deopt) {
  switch (op_kind) {
    case Token::kADD:
      __ addl(left, right);
      break;
    case Token::kSUB:
      __ subl(left, right);
      break;
    case Token::kBIT_AND:
      __ andl(left, right);
      break;
    case Token::kBIT_OR:
      __ orl(left, right);
      break;
    case Token::kBIT_XOR:
      __ xorl(left, right);
      break;
    case Token::kMUL:
      __ imull(left, right);
      break;
    default:
      UNREACHABLE();
  }
  if (deopt != NULL) __ j(OVERFLOW, deopt);
}

void BinaryUint32OpInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register left = locs()->in(0).reg();
  Register right = locs()->in(1).reg();
  Register out = locs()->out(0).reg();
  ASSERT(out == left);
  switch (op_kind()) {
    case Token::kBIT_AND:
    case Token::kBIT_OR:
    case Token::kBIT_XOR:
    case Token::kADD:
    case Token::kSUB:
    case Token::kMUL:
      EmitIntegerArithmetic(compiler, op_kind(), left, right, NULL);
      return;
    default:
      UNREACHABLE();
  }
}

LocationSummary* ShiftUint32OpInstr::MakeLocationSummary(Zone* zone,
                                                         bool opt) const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  summary->set_in(1, Location::FixedRegisterOrSmiConstant(right(), RCX));
  summary->set_out(0, Location::SameAsFirstInput());
  return summary;
}

void ShiftUint32OpInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  const intptr_t kShifterLimit = 31;

  Register left = locs()->in(0).reg();
  Register out = locs()->out(0).reg();
  ASSERT(left == out);

  Label* deopt =
      compiler->AddDeoptStub(deopt_id(), ICData::kDeoptBinaryInt64Op);

  if (locs()->in(1).IsConstant()) {
    // Shifter is constant.

    const Object& constant = locs()->in(1).constant();
    ASSERT(constant.IsSmi());
    const intptr_t shift_value = Smi::Cast(constant).Value();

    // Do the shift: (shift_value > 0) && (shift_value <= kShifterLimit).
    switch (op_kind()) {
      case Token::kSHR:
        __ shrl(left, Immediate(shift_value));
        break;
      case Token::kSHL:
        __ shll(left, Immediate(shift_value));
        break;
      default:
        UNREACHABLE();
    }
    return;
  }

  // Non constant shift value.

  Register shifter = locs()->in(1).reg();
  ASSERT(shifter == RCX);

  Label done;
  Label zero;

  // TODO(johnmccutchan): Use range information to avoid these checks.
  __ SmiUntag(shifter);
  __ cmpq(shifter, Immediate(0));
  // If shift value is < 0, deoptimize.
  __ j(NEGATIVE, deopt);
  __ cmpq(shifter, Immediate(kShifterLimit));
  // If shift value is >= 32, return zero.
  __ j(ABOVE, &zero);

  // Do the shift.
  switch (op_kind()) {
    case Token::kSHR:
      __ shrl(left, shifter);
      __ jmp(&done);
      break;
    case Token::kSHL:
      __ shll(left, shifter);
      __ jmp(&done);
      break;
    default:
      UNREACHABLE();
  }

  __ Bind(&zero);
  // Shift was greater than 31 bits, just return zero.
  __ xorq(left, left);

  // Exit path.
  __ Bind(&done);
}

DEFINE_BACKEND(UnaryUint32Op, (SameAsFirstInput, Register value)) {
  __ notl(value);
}

DEFINE_UNIMPLEMENTED_INSTRUCTION(BinaryInt32OpInstr)

LocationSummary* UnboxedIntConverterInstr::MakeLocationSummary(Zone* zone,
                                                               bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  if (from() == kUnboxedInt64) {
    ASSERT((to() == kUnboxedUint32) || (to() == kUnboxedInt32));
    summary->set_in(0, Location::RequiresRegister());
    summary->set_out(0, Location::SameAsFirstInput());
  } else if (to() == kUnboxedInt64) {
    ASSERT((from() == kUnboxedInt32) || (from() == kUnboxedUint32));
    summary->set_in(0, Location::RequiresRegister());
    summary->set_out(0, Location::SameAsFirstInput());
  } else {
    ASSERT((to() == kUnboxedUint32) || (to() == kUnboxedInt32));
    ASSERT((from() == kUnboxedUint32) || (from() == kUnboxedInt32));
    summary->set_in(0, Location::RequiresRegister());
    summary->set_out(0, Location::SameAsFirstInput());
  }
  return summary;
}

void UnboxedIntConverterInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  if (from() == kUnboxedInt32 && to() == kUnboxedUint32) {
    const Register value = locs()->in(0).reg();
    const Register out = locs()->out(0).reg();
    // Representations are bitwise equivalent but we want to normalize
    // upperbits for safety reasons.
    // TODO(vegorov) if we ensure that we never use upperbits we could
    // avoid this.
    __ movl(out, value);
  } else if (from() == kUnboxedUint32 && to() == kUnboxedInt32) {
    // Representations are bitwise equivalent.
    const Register value = locs()->in(0).reg();
    const Register out = locs()->out(0).reg();
    __ movsxd(out, value);
    if (CanDeoptimize()) {
      Label* deopt =
          compiler->AddDeoptStub(deopt_id(), ICData::kDeoptUnboxInteger);
      __ testl(out, out);
      __ j(NEGATIVE, deopt);
    }
  } else if (from() == kUnboxedInt64) {
    ASSERT((to() == kUnboxedUint32) || (to() == kUnboxedInt32));
    const Register value = locs()->in(0).reg();
    const Register out = locs()->out(0).reg();
    if (!CanDeoptimize()) {
      // Copy low.
      __ movl(out, value);
    } else {
      Label* deopt =
          compiler->AddDeoptStub(deopt_id(), ICData::kDeoptUnboxInteger);
      // Sign extend.
      __ movsxd(out, value);
      // Compare with original value.
      __ cmpq(out, value);
      // Value cannot be held in Int32, deopt.
      __ j(NOT_EQUAL, deopt);
    }
  } else if (to() == kUnboxedInt64) {
    ASSERT((from() == kUnboxedUint32) || (from() == kUnboxedInt32));
    const Register value = locs()->in(0).reg();
    const Register out = locs()->out(0).reg();
    if (from() == kUnboxedUint32) {
      // Zero extend.
      __ movl(out, value);
    } else {
      // Sign extend.
      ASSERT(from() == kUnboxedInt32);
      __ movsxd(out, value);
    }
  } else {
    UNREACHABLE();
  }
}

LocationSummary* ThrowInstr::MakeLocationSummary(Zone* zone, bool opt) const {
  return new (zone) LocationSummary(zone, 0, 0, LocationSummary::kCall);
}

void ThrowInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  compiler->GenerateRuntimeCall(token_pos(), deopt_id(), kThrowRuntimeEntry, 1,
                                locs());
  __ int3();
}

LocationSummary* ReThrowInstr::MakeLocationSummary(Zone* zone, bool opt) const {
  return new (zone) LocationSummary(zone, 0, 0, LocationSummary::kCall);
}

void ReThrowInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  compiler->SetNeedsStackTrace(catch_try_index());
  compiler->GenerateRuntimeCall(token_pos(), deopt_id(), kReThrowRuntimeEntry,
                                2, locs());
  __ int3();
}

LocationSummary* StopInstr::MakeLocationSummary(Zone* zone, bool opt) const {
  return new (zone) LocationSummary(zone, 0, 0, LocationSummary::kNoCall);
}

void StopInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  __ Stop(message());
}

void GraphEntryInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  if (!compiler->CanFallThroughTo(normal_entry())) {
    __ jmp(compiler->GetJumpLabel(normal_entry()));
  }
}

LocationSummary* GotoInstr::MakeLocationSummary(Zone* zone, bool opt) const {
  return new (zone) LocationSummary(zone, 0, 0, LocationSummary::kNoCall);
}

void GotoInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  if (!compiler->is_optimizing()) {
    if (FLAG_reorder_basic_blocks) {
      compiler->EmitEdgeCounter(block()->preorder_number());
    }
    // Add a deoptimization descriptor for deoptimizing instructions that
    // may be inserted before this instruction.
    compiler->AddCurrentDescriptor(RawPcDescriptors::kDeopt, GetDeoptId(),
                                   TokenPosition::kNoSource);
  }
  if (HasParallelMove()) {
    compiler->parallel_move_resolver()->EmitNativeCode(parallel_move());
  }

  // We can fall through if the successor is the next block in the list.
  // Otherwise, we need a jump.
  if (!compiler->CanFallThroughTo(successor())) {
    __ jmp(compiler->GetJumpLabel(successor()));
  }
}

LocationSummary* IndirectGotoInstr::MakeLocationSummary(Zone* zone,
                                                        bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 1;

  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);

  summary->set_in(0, Location::RequiresRegister());
  summary->set_temp(0, Location::RequiresRegister());

  return summary;
}

void IndirectGotoInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register offset_reg = locs()->in(0).reg();
  Register target_address_reg = locs()->temp(0).reg();

  {
    const intptr_t kRIPRelativeLeaqSize = 7;
    const intptr_t entry_to_rip_offset = __ CodeSize() + kRIPRelativeLeaqSize;
    __ leaq(target_address_reg,
            Address::AddressRIPRelative(-entry_to_rip_offset));
    ASSERT(__ CodeSize() == entry_to_rip_offset);
  }

  // Load from [current frame pointer] + kPcMarkerSlotFromFp.

  // Calculate the final absolute address.
  if (offset()->definition()->representation() == kTagged) {
    __ SmiUntag(offset_reg);
  }
  __ addq(target_address_reg, offset_reg);

  // Jump to the absolute address.
  __ jmp(target_address_reg);
}

LocationSummary* StrictCompareInstr::MakeLocationSummary(Zone* zone,
                                                         bool opt) const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  if (needs_number_check()) {
    LocationSummary* locs = new (zone)
        LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kCall);
    locs->set_in(0, Location::RegisterLocation(RAX));
    locs->set_in(1, Location::RegisterLocation(RCX));
    locs->set_out(0, Location::RegisterLocation(RAX));
    return locs;
  }
  LocationSummary* locs = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kNoCall);
  locs->set_in(0, Location::RegisterOrConstant(left()));
  // Only one of the inputs can be a constant. Choose register if the first one
  // is a constant.
  locs->set_in(1, locs->in(0).IsConstant()
                      ? Location::RequiresRegister()
                      : Location::RegisterOrConstant(right()));
  locs->set_out(0, Location::RequiresRegister());
  return locs;
}

Condition StrictCompareInstr::EmitComparisonCode(FlowGraphCompiler* compiler,
                                                 BranchLabels labels) {
  Location left = locs()->in(0);
  Location right = locs()->in(1);
  ASSERT(!left.IsConstant() || !right.IsConstant());
  Condition true_condition;
  if (left.IsConstant()) {
    true_condition = compiler->EmitEqualityRegConstCompare(
        right.reg(), left.constant(), needs_number_check(), token_pos(),
        deopt_id_);
  } else if (right.IsConstant()) {
    true_condition = compiler->EmitEqualityRegConstCompare(
        left.reg(), right.constant(), needs_number_check(), token_pos(),
        deopt_id_);
  } else {
    true_condition = compiler->EmitEqualityRegRegCompare(
        left.reg(), right.reg(), needs_number_check(), token_pos(), deopt_id_);
  }
  if (kind() != Token::kEQ_STRICT) {
    ASSERT(kind() == Token::kNE_STRICT);
    true_condition = NegateCondition(true_condition);
  }
  return true_condition;
}

LocationSummary* ClosureCallInstr::MakeLocationSummary(Zone* zone,
                                                       bool opt) const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary = new (zone)
      LocationSummary(zone, kNumInputs, kNumTemps, LocationSummary::kCall);
  summary->set_in(0, Location::RegisterLocation(RAX));  // Function.
  summary->set_out(0, Location::RegisterLocation(RAX));
  return summary;
}

void ClosureCallInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  // Arguments descriptor is expected in R10.
  const intptr_t argument_count = ArgumentCount();  // Includes type args.
  const Array& arguments_descriptor =
      Array::ZoneHandle(Z, GetArgumentsDescriptor());
  __ LoadObject(R10, arguments_descriptor);

  // Function in RAX.
  ASSERT(locs()->in(0).reg() == RAX);
  __ movq(CODE_REG, FieldAddress(RAX, Function::code_offset()));
  __ movq(RCX, FieldAddress(RAX, Function::entry_point_offset()));

  // RAX: Function.
  // R10: Arguments descriptor array.
  // RBX: Smi 0 (no IC data; the lazy-compile stub expects a GC-safe value).
  __ xorq(RBX, RBX);
  __ call(RCX);
  compiler->RecordSafepoint(locs());
  compiler->EmitCatchEntryState();
  // Marks either the continuation point in unoptimized code or the
  // deoptimization point in optimized code, after call.
  const intptr_t deopt_id_after = Thread::ToDeoptAfter(deopt_id());
  if (compiler->is_optimizing()) {
    compiler->AddDeoptIndexAtCall(deopt_id_after);
  }
  // Add deoptimization continuation point after the call and before the
  // arguments are removed.
  // In optimized code this descriptor is needed for exception handling.
  compiler->AddCurrentDescriptor(RawPcDescriptors::kDeopt, deopt_id_after,
                                 token_pos());
  __ Drop(argument_count);
}

LocationSummary* BooleanNegateInstr::MakeLocationSummary(Zone* zone,
                                                         bool opt) const {
  return LocationSummary::Make(zone, 1, Location::RequiresRegister(),
                               LocationSummary::kNoCall);
}

void BooleanNegateInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register value = locs()->in(0).reg();
  Register result = locs()->out(0).reg();

  Label done;
  __ LoadObject(result, Bool::True());
  __ CompareRegisters(result, value);
  __ j(NOT_EQUAL, &done, Assembler::kNearJump);
  __ LoadObject(result, Bool::False());
  __ Bind(&done);
}

LocationSummary* AllocateObjectInstr::MakeLocationSummary(Zone* zone,
                                                          bool opt) const {
  return MakeCallSummary(zone);
}

void AllocateObjectInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  const Code& stub = Code::ZoneHandle(
      compiler->zone(), StubCode::GetAllocationStubForClass(cls()));
  const StubEntry stub_entry(stub);
  compiler->GenerateCall(token_pos(), stub_entry, RawPcDescriptors::kOther,
                         locs());
  compiler->AddStubCallTarget(stub);
  __ Drop(ArgumentCount());  // Discard arguments.
}

void DebugStepCheckInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  ASSERT(!compiler->is_optimizing());
  __ CallPatchable(*StubCode::DebugStepCheck_entry());
  compiler->AddCurrentDescriptor(stub_kind_, deopt_id_, token_pos());
  compiler->RecordSafepoint(locs());
}

}  // namespace dart

#undef __

#endif  // defined(TARGET_ARCH_X64) && !defined(DART_PRECOMPILED_RUNTIME)