// 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)

#include "vm/intermediate_language.h"

#include "lib/error.h"
#include "vm/dart_entry.h"
#include "vm/flow_graph_compiler.h"
#include "vm/locations.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()->

namespace dart {

DECLARE_FLAG(int, optimization_counter_threshold);
DECLARE_FLAG(bool, propagate_ic_data);
DECLARE_FLAG(bool, throw_on_javascript_int_overflow);
DECLARE_FLAG(bool, use_osr);

// 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() {
  LocationSummary* result = new LocationSummary(0, 0, LocationSummary::kCall);
  result->set_out(Location::RegisterLocation(RAX));
  return result;
}


LocationSummary* PushArgumentInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps= 0;
  LocationSummary* locs =
      new LocationSummary(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() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs =
      new LocationSummary(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 defined(DEBUG)
  // TODO(srdjan): Fix for functions with finally clause.
  // A finally clause may leave a previously pushed return value if it
  // has its own return instruction. Method that have finally are currently
  // not optimized.
  if (!compiler->HasFinally()) {
    __ 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);
    __ cmpq(RDI, Immediate(fp_sp_dist));
    __ j(EQUAL, &done, Assembler::kNearJump);
    __ int3();
    __ Bind(&done);
  }
#endif
  __ LeaveFrame();
  __ ret();

  // Generate 8 bytes of NOPs so that the debugger can patch the
  // return pattern with a call to the debug stub.
  // Note that the nop(8) byte pattern is not recognized by the debugger.
  __ nop(1);
  __ nop(1);
  __ nop(1);
  __ nop(1);
  __ nop(1);
  __ nop(1);
  __ nop(1);
  __ nop(1);
  compiler->AddCurrentDescriptor(PcDescriptors::kReturn,
                                 Isolate::kNoDeoptId,
                                 token_pos());
}


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;
    default:
      UNIMPLEMENTED();
      return EQUAL;
  }
}


static bool BindsToSmiConstant(Value* val, intptr_t* smi_value) {
  if (!val->BindsToConstant()) {
    return false;
  }

  const Object& bound_constant = val->BoundConstant();
  if (!bound_constant.IsSmi()) {
    return false;
  }

  *smi_value = Smi::Cast(bound_constant).Value();
  return true;
}


// 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));
}


bool IfThenElseInstr::IsSupported() {
  return true;
}


bool IfThenElseInstr::Supports(ComparisonInstr* comparison,
                               Value* v1,
                               Value* v2) {
  if (!(comparison->IsStrictCompare() &&
        !comparison->AsStrictCompare()->needs_number_check()) &&
      !(comparison->IsEqualityCompare() &&
        (comparison->AsEqualityCompare()->operation_cid() == kSmiCid))) {
    return false;
  }

  intptr_t v1_value, v2_value;

  if (!BindsToSmiConstant(v1, &v1_value) ||
      !BindsToSmiConstant(v2, &v2_value)) {
    return false;
  }

  return false;
}


LocationSummary* IfThenElseInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  locs->set_in(0, Location::RegisterOrConstant(left()));
  locs->set_in(1, Location::RegisterOrConstant(right()));
  // TODO(vegorov): support byte register constraints in the register allocator.
  locs->set_out(Location::RegisterLocation(RDX));
  return locs;
}


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

  Location left = locs()->in(0);
  Location right = locs()->in(1);
  if (left.IsConstant() && right.IsConstant()) {
    // TODO(srdjan): Determine why this instruction was not eliminated.
    bool result = (left.constant().raw() == right.constant().raw());
    if ((kind_ == Token::kNE_STRICT) || (kind_ == Token::kNE)) {
      result = !result;
    }
    __ movq(locs()->out().reg(),
            Immediate(reinterpret_cast<int64_t>(
                Smi::New(result ? if_true_ : if_false_))));
    return;
  }

  ASSERT(!left.IsConstant() || !right.IsConstant());

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

  // Compare left and right. For now only equality comparison is supported.
  // TODO(vegorov): reuse code from the other comparison instructions instead of
  // generating it inline here.
  if (left.IsConstant()) {
    __ CompareObject(right.reg(), left.constant());
  } else if (right.IsConstant()) {
    __ CompareObject(left.reg(), right.constant());
  } else {
    __ cmpq(left.reg(), right.reg());
  }

  Condition true_condition =
      ((kind_ == Token::kEQ_STRICT) || (kind_ == Token::kEQ)) ? EQUAL
                                                              : NOT_EQUAL;

  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 {
    __ subq(RDX, Immediate(1));
    __ andq(RDX, Immediate(
        Smi::RawValue(true_value) - Smi::RawValue(false_value)));
    if (false_value != 0) {
      __ addq(RDX, Immediate(Smi::RawValue(false_value)));
    }
  }
}


LocationSummary* LoadLocalInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 0;
  return LocationSummary::Make(kNumInputs,
                               Location::RequiresRegister(),
                               LocationSummary::kNoCall);
}


void LoadLocalInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register result = locs()->out().reg();
  __ movq(result, Address(RBP, local().index() * kWordSize));
}


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


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


LocationSummary* ConstantInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 0;
  return LocationSummary::Make(kNumInputs,
                               Location::RequiresRegister(),
                               LocationSummary::kNoCall);
}


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


LocationSummary* AssertAssignableInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 3;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kCall);
  summary->set_in(0, Location::RegisterLocation(RAX));  // Value.
  summary->set_in(1, Location::RegisterLocation(RCX));  // Instantiator.
  summary->set_in(2, Location::RegisterLocation(RDX));  // Type arguments.
  summary->set_out(Location::RegisterLocation(RAX));
  return summary;
}


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


static void EmitAssertBoolean(Register reg,
                              intptr_t 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;
  __ CompareObject(reg, Bool::True());
  __ j(EQUAL, &done, Assembler::kNearJump);
  __ CompareObject(reg, Bool::False());
  __ j(EQUAL, &done, Assembler::kNearJump);

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


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

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


static Condition TokenKindToSmiCondition(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() const {
  const intptr_t kNumInputs = 2;
  if (operation_cid() == kDoubleCid) {
    const intptr_t kNumTemps =  0;
    LocationSummary* locs =
        new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
    locs->set_in(0, Location::RequiresFpuRegister());
    locs->set_in(1, Location::RequiresFpuRegister());
    locs->set_out(Location::RequiresRegister());
    return locs;
  }
  if (operation_cid() == kSmiCid) {
    const intptr_t kNumTemps = 0;
    LocationSummary* locs =
        new LocationSummary(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(Location::RequiresRegister());
    return locs;
  }
  if (IsCheckedStrictEqual()) {
    const intptr_t kNumTemps = 1;
    LocationSummary* locs =
        new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
    locs->set_in(0, Location::RequiresRegister());
    locs->set_in(1, Location::RequiresRegister());
    locs->set_temp(0, Location::RequiresRegister());
    locs->set_out(Location::RequiresRegister());
    return locs;
  }
  if (IsPolymorphic()) {
    const intptr_t kNumTemps = 1;
    LocationSummary* locs =
        new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kCall);
    locs->set_in(0, Location::RegisterLocation(RCX));
    locs->set_in(1, Location::RegisterLocation(RDX));
    locs->set_temp(0, Location::RegisterLocation(RBX));
    locs->set_out(Location::RegisterLocation(RAX));
    return locs;
  }
  const intptr_t kNumTemps = 1;
  LocationSummary* locs =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kCall);
  locs->set_in(0, Location::RegisterLocation(RCX));
  locs->set_in(1, Location::RegisterLocation(RDX));
  locs->set_temp(0, Location::RegisterLocation(RBX));
  locs->set_out(Location::RegisterLocation(RAX));
  return locs;
}


static void EmitEqualityAsInstanceCall(FlowGraphCompiler* compiler,
                                       intptr_t deopt_id,
                                       intptr_t token_pos,
                                       Token::Kind kind,
                                       LocationSummary* locs,
                                       const ICData& original_ic_data) {
  if (!compiler->is_optimizing()) {
    compiler->AddCurrentDescriptor(PcDescriptors::kDeopt,
                                   deopt_id,
                                   token_pos);
  }
  const int kNumberOfArguments = 2;
  const Array& kNoArgumentNames = Object::null_array();
  const int kNumArgumentsChecked = 2;

  const Immediate& raw_null =
      Immediate(reinterpret_cast<intptr_t>(Object::null()));
  Label check_identity;
  __ cmpq(Address(RSP, 0 * kWordSize), raw_null);
  __ j(EQUAL, &check_identity);
  __ cmpq(Address(RSP, 1 * kWordSize), raw_null);
  __ j(EQUAL, &check_identity);

  ICData& equality_ic_data = ICData::ZoneHandle(original_ic_data.raw());
  if (compiler->is_optimizing() && FLAG_propagate_ic_data) {
    ASSERT(!original_ic_data.IsNull());
    if (original_ic_data.NumberOfChecks() == 0) {
      // IC call for reoptimization populates original ICData.
      equality_ic_data = original_ic_data.raw();
    } else {
      // Megamorphic call.
      equality_ic_data = original_ic_data.AsUnaryClassChecks();
    }
  } else {
    const Array& arguments_descriptor =
        Array::Handle(ArgumentsDescriptor::New(kNumberOfArguments,
                                               kNoArgumentNames));
    equality_ic_data = ICData::New(compiler->parsed_function().function(),
                                   Symbols::EqualOperator(),
                                   arguments_descriptor,
                                   deopt_id,
                                   kNumArgumentsChecked);
  }
  compiler->GenerateInstanceCall(deopt_id,
                                 token_pos,
                                 kNumberOfArguments,
                                 kNoArgumentNames,
                                 locs,
                                 equality_ic_data);
  Label check_ne;
  __ jmp(&check_ne);

  __ Bind(&check_identity);
  Label equality_done;
  if (compiler->is_optimizing()) {
    // No need to update IC data.
    Label is_true;
    __ popq(RAX);
    __ popq(RDX);
    __ cmpq(RAX, RDX);
    __ j(EQUAL, &is_true);
    __ LoadObject(RAX, (kind == Token::kEQ) ? Bool::False() : Bool::True());
    __ jmp(&equality_done);
    __ Bind(&is_true);
    __ LoadObject(RAX, (kind == Token::kEQ) ? Bool::True() : Bool::False());
    if (kind == Token::kNE) {
      // Skip not-equal result conversion.
      __ jmp(&equality_done);
    }
  } else {
    // Call stub, load IC data in register. The stub will update ICData if
    // necessary.
    Register ic_data_reg = locs->temp(0).reg();
    ASSERT(ic_data_reg == RBX);  // Stub depends on it.
    __ LoadObject(ic_data_reg, equality_ic_data);
    compiler->GenerateCall(token_pos,
                           &StubCode::EqualityWithNullArgLabel(),
                           PcDescriptors::kRuntimeCall,
                           locs);
    __ Drop(2);
  }
  __ Bind(&check_ne);
  if (kind == Token::kNE) {
    Label true_label, done;
    // Negate the condition: true label returns false and vice versa.
    __ CompareObject(RAX, Bool::True());
    __ j(EQUAL, &true_label, Assembler::kNearJump);
    __ LoadObject(RAX, Bool::True());
    __ jmp(&done, Assembler::kNearJump);
    __ Bind(&true_label);
    __ LoadObject(RAX, Bool::False());
    __ Bind(&done);
  }
  __ Bind(&equality_done);
}


static void LoadValueCid(FlowGraphCompiler* compiler,
                         Register value_cid_reg,
                         Register value_reg,
                         Label* value_is_smi = NULL) {
  Label done;
  if (value_is_smi == NULL) {
    __ movq(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 void EmitEqualityAsPolymorphicCall(FlowGraphCompiler* compiler,
                                          const ICData& orig_ic_data,
                                          LocationSummary* locs,
                                          BranchInstr* branch,
                                          Token::Kind kind,
                                          intptr_t deopt_id,
                                          intptr_t token_pos) {
  ASSERT((kind == Token::kEQ) || (kind == Token::kNE));
  const ICData& ic_data = ICData::Handle(orig_ic_data.AsUnaryClassChecks());
  ASSERT(ic_data.NumberOfChecks() > 0);
  ASSERT(ic_data.num_args_tested() == 1);
  Label* deopt = compiler->AddDeoptStub(deopt_id, kDeoptEquality);
  Register left = locs->in(0).reg();
  Register right = locs->in(1).reg();
  Register temp = locs->temp(0).reg();
  LoadValueCid(compiler, temp, left,
               (ic_data.GetReceiverClassIdAt(0) == kSmiCid) ? NULL : deopt);
  // 'temp' contains class-id of the left argument.
  ObjectStore* object_store = Isolate::Current()->object_store();
  Condition cond = TokenKindToSmiCondition(kind);
  Label done;
  const intptr_t len = ic_data.NumberOfChecks();
  for (intptr_t i = 0; i < len; i++) {
    // Assert that the Smi is at position 0, if at all.
    ASSERT((ic_data.GetReceiverClassIdAt(i) != kSmiCid) || (i == 0));
    Label next_test;
    __ cmpq(temp, Immediate(ic_data.GetReceiverClassIdAt(i)));
    if (i < len - 1) {
      __ j(NOT_EQUAL, &next_test);
    } else {
      __ j(NOT_EQUAL, deopt);
    }
    const Function& target = Function::ZoneHandle(ic_data.GetTargetAt(i));
    if (target.Owner() == object_store->object_class()) {
      // Object.== is same as ===.
      __ Drop(2);
      __ cmpq(left, right);
      if (branch != NULL) {
        branch->EmitBranchOnCondition(compiler, cond);
      } else {
        // This case should be rare.
        Register result = locs->out().reg();
        Label load_true;
        __ j(cond, &load_true, Assembler::kNearJump);
        __ LoadObject(result, Bool::False());
        __ jmp(&done);
        __ Bind(&load_true);
        __ LoadObject(result, Bool::True());
      }
    } else {
      const int kNumberOfArguments = 2;
      const Array& kNoArgumentNames = Object::null_array();
      compiler->GenerateStaticCall(deopt_id,
                                   token_pos,
                                   target,
                                   kNumberOfArguments,
                                   kNoArgumentNames,
                                   locs);
      if (branch == NULL) {
        if (kind == Token::kNE) {
          Label false_label;
          __ CompareObject(RAX, Bool::True());
          __ j(EQUAL, &false_label, Assembler::kNearJump);
          __ LoadObject(RAX, Bool::True());
          __ jmp(&done);
          __ Bind(&false_label);
          __ LoadObject(RAX, Bool::False());
        }
      } else {
        if (branch->is_checked()) {
          EmitAssertBoolean(RAX, token_pos, deopt_id, locs, compiler);
        }
        __ CompareObject(RAX, Bool::True());
        branch->EmitBranchOnCondition(compiler, cond);
      }
    }
    if (i < len - 1) {
      __ jmp(&done);
      __ Bind(&next_test);
    }
  }
  __ Bind(&done);
}


// Emit code when ICData's targets are all Object == (which is ===).
static void EmitCheckedStrictEqual(FlowGraphCompiler* compiler,
                                   const ICData& orig_ic_data,
                                   const LocationSummary& locs,
                                   Token::Kind kind,
                                   BranchInstr* branch,
                                   intptr_t deopt_id) {
  ASSERT((kind == Token::kEQ) || (kind == Token::kNE));
  Register left = locs.in(0).reg();
  Register right = locs.in(1).reg();
  Register temp = locs.temp(0).reg();
  Label* deopt = compiler->AddDeoptStub(deopt_id, kDeoptEquality);
  __ testq(left, Immediate(kSmiTagMask));
  __ j(ZERO, deopt);
  // 'left' is not Smi.
  const Immediate& raw_null =
      Immediate(reinterpret_cast<intptr_t>(Object::null()));
  Label identity_compare;
  __ cmpq(right, raw_null);
  __ j(EQUAL, &identity_compare);
  __ cmpq(left, raw_null);
  __ j(EQUAL, &identity_compare);

  __ LoadClassId(temp, left);
  const ICData& ic_data = ICData::Handle(orig_ic_data.AsUnaryClassChecks());
  const intptr_t len = ic_data.NumberOfChecks();
  for (intptr_t i = 0; i < len; i++) {
    __ cmpq(temp, Immediate(ic_data.GetReceiverClassIdAt(i)));
    if (i == (len - 1)) {
      __ j(NOT_EQUAL, deopt);
    } else {
      __ j(EQUAL, &identity_compare);
    }
  }
  __ Bind(&identity_compare);
  __ cmpq(left, right);
  if (branch == NULL) {
    Label done, is_equal;
    Register result = locs.out().reg();
    __ j(EQUAL, &is_equal, Assembler::kNearJump);
    // Not equal.
    __ LoadObject(result, (kind == Token::kEQ) ? Bool::False() : Bool::True());
    __ jmp(&done, Assembler::kNearJump);
    __ Bind(&is_equal);
    __ LoadObject(result, (kind == Token::kEQ) ? Bool::True() : Bool::False());
    __ Bind(&done);
  } else {
    Condition cond = TokenKindToSmiCondition(kind);
    branch->EmitBranchOnCondition(compiler, cond);
  }
}


// First test if receiver is NULL, in which case === is applied.
// If type feedback was provided (lists of <class-id, target>), do a
// type by type check (either === or static call to the operator.
static void EmitGenericEqualityCompare(FlowGraphCompiler* compiler,
                                       LocationSummary* locs,
                                       Token::Kind kind,
                                       BranchInstr* branch,
                                       const ICData& ic_data,
                                       intptr_t deopt_id,
                                       intptr_t token_pos) {
  ASSERT((kind == Token::kEQ) || (kind == Token::kNE));
  ASSERT(!ic_data.IsNull() && (ic_data.NumberOfChecks() > 0));
  Register left = locs->in(0).reg();
  Register right = locs->in(1).reg();
  const Immediate& raw_null =
      Immediate(reinterpret_cast<intptr_t>(Object::null()));
  Label done, identity_compare, non_null_compare;
  __ cmpq(right, raw_null);
  __ j(EQUAL, &identity_compare, Assembler::kNearJump);
  __ cmpq(left, raw_null);
  __ j(NOT_EQUAL, &non_null_compare, Assembler::kNearJump);
  // Comparison with NULL is "===".
  __ Bind(&identity_compare);
  __ cmpq(left, right);
  Condition cond = TokenKindToSmiCondition(kind);
  if (branch != NULL) {
    branch->EmitBranchOnCondition(compiler, cond);
  } else {
    Register result = locs->out().reg();
    Label load_true;
    __ j(cond, &load_true, Assembler::kNearJump);
    __ LoadObject(result, Bool::False());
    __ jmp(&done);
    __ Bind(&load_true);
    __ LoadObject(result, Bool::True());
  }
  __ jmp(&done);
  __ Bind(&non_null_compare);  // Receiver is not null.
  __ pushq(left);
  __ pushq(right);
  EmitEqualityAsPolymorphicCall(compiler, ic_data, locs, branch, kind,
                                deopt_id, token_pos);
  __ 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 EmitSmiComparisonOp(FlowGraphCompiler* compiler,
                                const LocationSummary& locs,
                                Token::Kind kind,
                                BranchInstr* branch) {
  Location left = locs.in(0);
  Location right = locs.in(1);
  ASSERT(!left.IsConstant() || !right.IsConstant());

  Condition true_condition = TokenKindToSmiCondition(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());
  }

  if (branch != NULL) {
    branch->EmitBranchOnCondition(compiler, true_condition);
  } else {
    Register result = locs.out().reg();
    Label done, is_true;
    __ j(true_condition, &is_true);
    __ LoadObject(result, Bool::False());
    __ jmp(&done);
    __ Bind(&is_true);
    __ LoadObject(result, Bool::True());
    __ Bind(&done);
  }
}


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 void EmitDoubleComparisonOp(FlowGraphCompiler* compiler,
                                   const LocationSummary& locs,
                                   Token::Kind kind,
                                   BranchInstr* branch) {
  XmmRegister left = locs.in(0).fpu_reg();
  XmmRegister right = locs.in(1).fpu_reg();

  Condition true_condition = TokenKindToDoubleCondition(kind);
  if (branch != NULL) {
    compiler->EmitDoubleCompareBranch(
        true_condition, left, right, branch);
  } else {
    compiler->EmitDoubleCompareBool(
        true_condition, left, right, locs.out().reg());
  }
}


void EqualityCompareInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  ASSERT((kind() == Token::kEQ) || (kind() == Token::kNE));
  BranchInstr* kNoBranch = NULL;
  if (operation_cid() == kSmiCid) {
    // Deoptimizes if both arguments not Smi.
    EmitSmiComparisonOp(compiler, *locs(), kind(), kNoBranch);
    return;
  }
  if (operation_cid() == kDoubleCid) {
    // Deoptimizes if both arguments are Smi, or if none is Double or Smi.
    EmitDoubleComparisonOp(compiler, *locs(), kind(), kNoBranch);
    return;
  }
  if (IsCheckedStrictEqual()) {
    EmitCheckedStrictEqual(compiler, *ic_data(), *locs(), kind(), kNoBranch,
                           deopt_id());
    return;
  }
  if (IsPolymorphic()) {
    EmitGenericEqualityCompare(compiler, locs(), kind(), kNoBranch, *ic_data(),
                               deopt_id(), token_pos());
    return;
  }
  Register left = locs()->in(0).reg();
  Register right = locs()->in(1).reg();
  __ pushq(left);
  __ pushq(right);
  EmitEqualityAsInstanceCall(compiler,
                             deopt_id(),
                             token_pos(),
                             kind(),
                             locs(),
                             *ic_data());
  ASSERT(locs()->out().reg() == RAX);
}


void EqualityCompareInstr::EmitBranchCode(FlowGraphCompiler* compiler,
                                          BranchInstr* branch) {
  ASSERT((kind() == Token::kNE) || (kind() == Token::kEQ));
  if (operation_cid() == kSmiCid) {
    // Deoptimizes if both arguments not Smi.
    EmitSmiComparisonOp(compiler, *locs(), kind(), branch);
    return;
  }
  if (operation_cid() == kDoubleCid) {
    // Deoptimizes if both arguments are Smi, or if none is Double or Smi.
    EmitDoubleComparisonOp(compiler, *locs(), kind(), branch);
    return;
  }
  if (IsCheckedStrictEqual()) {
    EmitCheckedStrictEqual(compiler, *ic_data(), *locs(), kind(), branch,
                           deopt_id());
    return;
  }
  if (IsPolymorphic()) {
    EmitGenericEqualityCompare(compiler, locs(), kind(), branch, *ic_data(),
                               deopt_id(), token_pos());
    return;
  }
  Register left = locs()->in(0).reg();
  Register right = locs()->in(1).reg();
  __ pushq(left);
  __ pushq(right);
  EmitEqualityAsInstanceCall(compiler,
                             deopt_id(),
                             token_pos(),
                             Token::kEQ,  // kNE reverse occurs at branch.
                             locs(),
                             *ic_data());
  if (branch->is_checked()) {
    EmitAssertBoolean(RAX, token_pos(), deopt_id(), locs(), compiler);
  }
  Condition branch_condition = (kind() == Token::kNE) ? NOT_EQUAL : EQUAL;
  __ CompareObject(RAX, Bool::True());
  branch->EmitBranchOnCondition(compiler, branch_condition);
}


LocationSummary* RelationalOpInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  if (operation_cid() == kDoubleCid) {
    LocationSummary* summary =
        new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
    summary->set_in(0, Location::RequiresFpuRegister());
    summary->set_in(1, Location::RequiresFpuRegister());
    summary->set_out(Location::RequiresRegister());
    return summary;
  } else if (operation_cid() == kSmiCid) {
    LocationSummary* summary =
        new LocationSummary(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(Location::RequiresRegister());
    return summary;
  }
  LocationSummary* locs =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kCall);
  // Pick arbitrary fixed input registers because this is a call.
  locs->set_in(0, Location::RegisterLocation(RAX));
  locs->set_in(1, Location::RegisterLocation(RCX));
  locs->set_out(Location::RegisterLocation(RAX));
  return locs;
}


void RelationalOpInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  if (operation_cid() == kSmiCid) {
    EmitSmiComparisonOp(compiler, *locs(), kind(), NULL);
    return;
  }
  if (operation_cid() == kDoubleCid) {
    EmitDoubleComparisonOp(compiler, *locs(), kind(), NULL);
    return;
  }

  // Push arguments for the call.
  // TODO(fschneider): Split this instruction into different types to avoid
  // explicitly pushing arguments to the call here.
  Register left = locs()->in(0).reg();
  Register right = locs()->in(1).reg();
  __ pushq(left);
  __ pushq(right);
  if (HasICData() && (ic_data()->NumberOfChecks() > 0)) {
    Label* deopt = compiler->AddDeoptStub(deopt_id(), kDeoptRelationalOp);

    // Load class into RDI. Since this is a call, any register except
    // the fixed input registers would be ok.
    ASSERT((left != RDI) && (right != RDI));
    LoadValueCid(compiler, RDI, left);
    const intptr_t kNumArguments = 2;
    compiler->EmitTestAndCall(ICData::Handle(ic_data()->AsUnaryClassChecks()),
                              RDI,  // Class id register.
                              kNumArguments,
                              Object::null_array(),  // No named arguments.
                              deopt,  // Deoptimize target.
                              deopt_id(),
                              token_pos(),
                              locs());
    return;
  }
  const String& function_name =
      String::ZoneHandle(Symbols::New(Token::Str(kind())));
  if (!compiler->is_optimizing()) {
    compiler->AddCurrentDescriptor(PcDescriptors::kDeopt,
                                   deopt_id(),
                                   token_pos());
  }
  const intptr_t kNumArguments = 2;
  const intptr_t kNumArgsChecked = 2;  // Type-feedback.
  ICData& relational_ic_data = ICData::ZoneHandle(ic_data()->raw());
  if (compiler->is_optimizing() && FLAG_propagate_ic_data) {
    ASSERT(!ic_data()->IsNull());
    if (ic_data()->NumberOfChecks() == 0) {
      // IC call for reoptimization populates original ICData.
      relational_ic_data = ic_data()->raw();
    } else {
      // Megamorphic call.
      relational_ic_data = ic_data()->AsUnaryClassChecks();
    }
  } else {
    const Array& arguments_descriptor =
        Array::Handle(ArgumentsDescriptor::New(kNumArguments,
                                               Object::null_array()));
    relational_ic_data = ICData::New(compiler->parsed_function().function(),
                                     function_name,
                                     arguments_descriptor,
                                     deopt_id(),
                                     kNumArgsChecked);
  }
  compiler->GenerateInstanceCall(deopt_id(),
                                 token_pos(),
                                 kNumArguments,
                                 Object::null_array(),  // No optional args.
                                 locs(),
                                 relational_ic_data);
}


void RelationalOpInstr::EmitBranchCode(FlowGraphCompiler* compiler,
                                       BranchInstr* branch) {
  if (operation_cid() == kSmiCid) {
    EmitSmiComparisonOp(compiler, *locs(), kind(), branch);
    return;
  }
  if (operation_cid() == kDoubleCid) {
    EmitDoubleComparisonOp(compiler, *locs(), kind(), branch);
    return;
  }
  EmitNativeCode(compiler);
  __ CompareObject(RAX, Bool::True());
  branch->EmitBranchOnCondition(compiler, EQUAL);
}


LocationSummary* NativeCallInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 0;
  const intptr_t kNumTemps = 3;
  LocationSummary* locs =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kCall);
  locs->set_temp(0, Location::RegisterLocation(RAX));
  locs->set_temp(1, Location::RegisterLocation(RBX));
  locs->set_temp(2, Location::RegisterLocation(R10));
  locs->set_out(Location::RegisterLocation(RAX));
  return locs;
}


void NativeCallInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  ASSERT(locs()->temp(0).reg() == RAX);
  ASSERT(locs()->temp(1).reg() == RBX);
  ASSERT(locs()->temp(2).reg() == R10);
  Register result = locs()->out().reg();

  // Push the result place holder initialized to NULL.
  __ PushObject(Object::ZoneHandle());
  // Pass a pointer to the first argument in RAX.
  if (!function().HasOptionalParameters()) {
    __ leaq(RAX, Address(RBP, (kParamEndSlotFromFp +
                               function().NumParameters()) * kWordSize));
  } else {
    __ leaq(RAX,
            Address(RBP, kFirstLocalSlotFromFp * kWordSize));
  }
  __ movq(RBX, Immediate(reinterpret_cast<uword>(native_c_function())));
  __ movq(R10, Immediate(NativeArguments::ComputeArgcTag(function())));
  compiler->GenerateCall(token_pos(),
                         &StubCode::CallNativeCFunctionLabel(),
                         PcDescriptors::kOther,
                         locs());
  __ popq(result);
}


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 = FlowGraphCompiler::ElementSizeFor(cid);
  const intptr_t data_offset = FlowGraphCompiler::DataOffsetFor(cid);
  const int64_t disp = smi_const.AsInt64Value() * scale + data_offset;
  return Utils::IsInt(32, disp);
}


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


void StringFromCharCodeInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register char_code = locs()->in(0).reg();
  Register result = locs()->out().reg();
  __ movq(result,
          Immediate(reinterpret_cast<uword>(Symbols::PredefinedAddress())));
  __ movq(result, Address(result,
                          char_code,
                          TIMES_HALF_WORD_SIZE,  // Char code is a smi.
                          Symbols::kNullCharCodeSymbolOffset * kWordSize));
}


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


void LoadUntaggedInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register object = locs()->in(0).reg();
  Register result = locs()->out().reg();
  __ movq(result, FieldAddress(object, offset()));
}


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


void LoadClassIdInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register object = locs()->in(0).reg();
  Register result = locs()->out().reg();
  Label load, done;
  __ testq(object, Immediate(kSmiTagMask));
  __ j(NOT_ZERO, &load, Assembler::kNearJump);
  __ movq(result, Immediate(Smi::RawValue(kSmiCid)));
  __ jmp(&done);
  __ Bind(&load);
  __ LoadClassId(result, object);
  __ SmiTag(result);
  __ Bind(&done);
}


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

    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 kTypedDataInt32ArrayCid:
    case kTypedDataUint32ArrayCid:
      return kTagged;
    case kTypedDataFloat32ArrayCid:
    case kTypedDataFloat64ArrayCid:
      return kUnboxedDouble;
    case kTypedDataFloat32x4ArrayCid:
      return kUnboxedFloat32x4;
    default:
      UNIMPLEMENTED();
      return kTagged;
  }
}


LocationSummary* LoadIndexedInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs =
      new LocationSummary(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()->value())
                      : Location::WritableRegister());
  } else {
    locs->set_in(1, CanBeImmediateIndex(index(), class_id())
                      ? Location::Constant(
                          index()->definition()->AsConstant()->value())
                      : Location::RequiresRegister());
  }
  if (representation() == kUnboxedDouble) {
    locs->set_out(Location::RequiresFpuRegister());
  } else {
    locs->set_out(Location::RequiresRegister());
  }
  return locs;
}


void LoadIndexedInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register array = locs()->in(0).reg();
  Location index = locs()->in(1);

  const bool is_external =
      (this->array()->definition()->representation() == kUntagged);
  Address element_address(kNoRegister, 0);

  if (is_external) {
    element_address = index.IsRegister()
        ? FlowGraphCompiler::ExternalElementAddressForRegIndex(
            index_scale(), array, index.reg())
        : FlowGraphCompiler::ExternalElementAddressForIntIndex(
            index_scale(), array, Smi::Cast(index.constant()).Value());
  } else {
    ASSERT(this->array()->definition()->representation() == kTagged);
    element_address = index.IsRegister()
        ? FlowGraphCompiler::ElementAddressForRegIndex(
            class_id(), index_scale(), array, index.reg())
        : FlowGraphCompiler::ElementAddressForIntIndex(
            class_id(), index_scale(), array,
            Smi::Cast(index.constant()).Value());
  }

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

    XmmRegister result = locs()->out().fpu_reg();
    if (class_id() == kTypedDataFloat32ArrayCid) {
      // Load single precision float.
      __ movss(result, element_address);
      // Promote to double.
      __ cvtss2sd(result, locs()->out().fpu_reg());
    } else if (class_id() == kTypedDataFloat64ArrayCid) {
      __ movsd(result, element_address);
    } else {
      ASSERT(class_id() == kTypedDataFloat32x4ArrayCid);
      __ movups(result, element_address);
    }
    return;
  }

  if ((index_scale() == 1) && index.IsRegister()) {
    __ SmiUntag(index.reg());
  }
  Register result = locs()->out().reg();
  switch (class_id()) {
    case kTypedDataInt8ArrayCid:
      __ movsxb(result, element_address);
      __ SmiTag(result);
      break;
    case kTypedDataUint8ArrayCid:
    case kTypedDataUint8ClampedArrayCid:
    case kExternalTypedDataUint8ArrayCid:
    case kExternalTypedDataUint8ClampedArrayCid:
    case kOneByteStringCid:
      __ movzxb(result, element_address);
      __ SmiTag(result);
      break;
    case kTypedDataInt16ArrayCid:
      __ movsxw(result, element_address);
      __ SmiTag(result);
      break;
    case kTypedDataUint16ArrayCid:
    case kTwoByteStringCid:
      __ movzxw(result, element_address);
      __ SmiTag(result);
      break;
    case kTypedDataInt32ArrayCid:
      __ movsxd(result, element_address);
      __ SmiTag(result);
      break;
    case kTypedDataUint32ArrayCid:
      __ movl(result, element_address);
      __ SmiTag(result);
      break;
    default:
      ASSERT((class_id() == kArrayCid) || (class_id() == kImmutableArrayCid));
      __ movq(result, element_address);
      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:
    case kTypedDataInt32ArrayCid:
    case kTypedDataUint32ArrayCid:
      return kTagged;
    case kTypedDataFloat32ArrayCid:
    case kTypedDataFloat64ArrayCid:
      return kUnboxedDouble;
    case kTypedDataFloat32x4ArrayCid:
      return kUnboxedFloat32x4;
    default:
      UNIMPLEMENTED();
      return kTagged;
  }
}


LocationSummary* StoreIndexedInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 3;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs =
      new LocationSummary(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()->value())
                      : Location::WritableRegister());
  } else {
    locs->set_in(1, CanBeImmediateIndex(index(), class_id())
                      ? Location::Constant(
                          index()->definition()->AsConstant()->value())
                      : 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:
    case kTypedDataInt32ArrayCid:
    case kTypedDataUint32ArrayCid:
      // Writable register because the value must be untagged before storing.
      locs->set_in(2, Location::WritableRegister());
      break;
    case kTypedDataFloat32ArrayCid:
      // Need temp register for float-to-double conversion.
      locs->AddTemp(Location::RequiresFpuRegister());
      // Fall through.
    case kTypedDataFloat64ArrayCid:
      // TODO(srdjan): Support Float64 constants.
      locs->set_in(2, Location::RequiresFpuRegister());
      break;
    case kTypedDataFloat32x4ArrayCid:
      locs->set_in(2, Location::RequiresFpuRegister());
      break;
    default:
      UNREACHABLE();
      return NULL;
  }
  return locs;
}


void StoreIndexedInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register array = locs()->in(0).reg();
  Location index = locs()->in(1);

  const bool is_external =
      (this->array()->definition()->representation() == kUntagged);
  Address element_address(kNoRegister, 0);
  if (is_external) {
    element_address = index.IsRegister()
        ? FlowGraphCompiler::ExternalElementAddressForRegIndex(
            index_scale(), array, index.reg())
        : FlowGraphCompiler::ExternalElementAddressForIntIndex(
            index_scale(), array, Smi::Cast(index.constant()).Value());
  } else {
    ASSERT(this->array()->definition()->representation() == kTagged);
    element_address = index.IsRegister()
        ? FlowGraphCompiler::ElementAddressForRegIndex(
            class_id(), index_scale(), array, index.reg())
        : FlowGraphCompiler::ElementAddressForIntIndex(
            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();
        __ StoreObject(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);
        __ cmpq(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);
        __ movq(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();
      __ SmiUntag(value);
      __ movl(element_address, value);
        break;
    }
    case kTypedDataFloat32ArrayCid:
      // Convert to single precision.
      __ cvtsd2ss(locs()->temp(0).fpu_reg(), locs()->in(2).fpu_reg());
      // Store.
      __ movss(element_address, locs()->temp(0).fpu_reg());
      break;
    case kTypedDataFloat64ArrayCid:
      __ movsd(element_address, locs()->in(2).fpu_reg());
      break;
    case kTypedDataFloat32x4ArrayCid:
      __ movups(element_address, locs()->in(2).fpu_reg());
      break;
    default:
      UNREACHABLE();
  }
}


LocationSummary* GuardFieldInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, 0, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  if ((value()->Type()->ToCid() == kDynamicCid) &&
      (field().guarded_cid() != kSmiCid)) {
    summary->AddTemp(Location::RequiresRegister());
  }
  if (field().guarded_cid() == kIllegalCid) {
    summary->AddTemp(Location::RequiresRegister());
  }
  return summary;
}


void GuardFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  const intptr_t field_cid = field().guarded_cid();
  const intptr_t nullability = field().is_nullable() ? kNullCid : kIllegalCid;

  if (field_cid == kDynamicCid) {
    ASSERT(!compiler->is_optimizing());
    return;  // Nothing to emit.
  }

  const intptr_t value_cid = value()->Type()->ToCid();

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

  Register value_cid_reg = ((value_cid == kDynamicCid) &&
      (field_cid != kSmiCid)) ? locs()->temp(0).reg() : kNoRegister;

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

  Label ok, fail_label;

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

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

  const bool ok_is_fall_through = (deopt != NULL);

  if (!compiler->is_optimizing() || (field_cid == kIllegalCid)) {
    if (!compiler->is_optimizing()) {
      // Currently we can't have different location summaries for optimized
      // and non-optimized code. So instead we manually pick up a register
      // that is known to be free because we know how non-optimizing compiler
      // allocates registers.
      field_reg = RBX;
      ASSERT((field_reg != value_reg) && (field_reg != value_cid_reg));
    }

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

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

    if (value_cid == kDynamicCid) {
      if (value_cid_reg == kNoRegister) {
        ASSERT(!compiler->is_optimizing());
        value_cid_reg = RDX;
        ASSERT((value_cid_reg != value_reg) && (field_reg != value_cid_reg));
      }

      LoadValueCid(compiler, value_cid_reg, value_reg);

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

    __ cmpq(field_cid_operand, Immediate(kIllegalCid));
    __ j(NOT_EQUAL, fail);

    if (value_cid == kDynamicCid) {
      __ movq(field_cid_operand, value_cid_reg);
      __ movq(field_nullability_operand, value_cid_reg);
    } else {
      __ movq(field_cid_operand, Immediate(value_cid));
      __ movq(field_nullability_operand, Immediate(value_cid));
    }

    if (!ok_is_fall_through) {
      __ jmp(&ok);
    }
  } else {
    if (value_cid == kDynamicCid) {
      // Field's guarded class id is fixed but 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);
        __ cmpq(value_cid_reg, Immediate(field_cid));
      }

      if (field().is_nullable() && (field_cid != kNullCid)) {
        __ j(EQUAL, &ok);
        const Immediate& raw_null =
            Immediate(reinterpret_cast<intptr_t>(Object::null()));
        __ cmpq(value_reg, raw_null);
      }

      if (ok_is_fall_through) {
        __ j(NOT_EQUAL, fail);
      } else {
        __ j(EQUAL, &ok);
      }
    } else {
      // Both value's and field's class id is known.
      if ((value_cid != field_cid) && (value_cid != nullability)) {
        if (ok_is_fall_through) {
          __ jmp(fail);
        }
      } else {
        // Nothing to emit.
        ASSERT(!compiler->is_optimizing());
        return;
      }
    }
  }

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

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

    __ pushq(field_reg);
    __ pushq(value_reg);
    __ CallRuntime(kUpdateFieldCidRuntimeEntry);
    __ Drop(2);  // Drop the field and the value.
  }

  __ Bind(&ok);
}


LocationSummary* StoreInstanceFieldInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  summary->set_in(1, ShouldEmitStoreBarrier()
                       ? Location::WritableRegister()
                       : Location::RegisterOrConstant(value()));
  return summary;
}


void StoreInstanceFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register instance_reg = locs()->in(0).reg();
  if (ShouldEmitStoreBarrier()) {
    Register value_reg = locs()->in(1).reg();
    __ StoreIntoObject(instance_reg,
                       FieldAddress(instance_reg, field().Offset()),
                       value_reg,
                       CanValueBeSmi());
  } else {
    if (locs()->in(1).IsConstant()) {
      __ StoreObject(FieldAddress(instance_reg, field().Offset()),
                     locs()->in(1).constant());
    } else {
      Register value_reg = locs()->in(1).reg();
      __ StoreIntoObjectNoBarrier(instance_reg,
          FieldAddress(instance_reg, field().Offset()), value_reg);
    }
  }
}


LocationSummary* LoadStaticFieldInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  summary->set_out(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().reg();
  __ movq(result, FieldAddress(field, Field::value_offset()));
}


LocationSummary* StoreStaticFieldInstr::MakeLocationSummary() const {
  LocationSummary* locs = new LocationSummary(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());
  if (this->value()->NeedsStoreBuffer()) {
    __ StoreIntoObject(temp,
        FieldAddress(temp, Field::value_offset()), value, CanValueBeSmi());
  } else {
    __ StoreIntoObjectNoBarrier(
        temp, FieldAddress(temp, Field::value_offset()), value);
  }
}


LocationSummary* InstanceOfInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 3;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kCall);
  summary->set_in(0, Location::RegisterLocation(RAX));
  summary->set_in(1, Location::RegisterLocation(RCX));
  summary->set_in(2, Location::RegisterLocation(RDX));
  summary->set_out(Location::RegisterLocation(RAX));
  return summary;
}


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

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


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


void CreateArrayInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  // Allocate the array.  R10 = length, RBX = element type.
  ASSERT(locs()->in(0).reg() == RBX);
  __ movq(R10, Immediate(Smi::RawValue(num_elements())));
  compiler->GenerateCall(token_pos(),
                         &StubCode::AllocateArrayLabel(),
                         PcDescriptors::kOther,
                         locs());
  ASSERT(locs()->out().reg() == RAX);
}


LocationSummary*
AllocateObjectWithBoundsCheckInstr::MakeLocationSummary() const {
  return MakeCallSummary();
}


void AllocateObjectWithBoundsCheckInstr::EmitNativeCode(
    FlowGraphCompiler* compiler) {
  compiler->GenerateCallRuntime(token_pos(),
                                deopt_id(),
                                kAllocateObjectWithBoundsCheckRuntimeEntry,
                                locs());
  __ Drop(3);
  ASSERT(locs()->out().reg() == RAX);
  __ popq(RAX);  // Pop new instance.
}


LocationSummary* LoadFieldInstr::MakeLocationSummary() const {
  return LocationSummary::Make(1,
                               Location::RequiresRegister(),
                               LocationSummary::kNoCall);
}


void LoadFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register instance_reg = locs()->in(0).reg();
  Register result_reg = locs()->out().reg();

  __ movq(result_reg, FieldAddress(instance_reg, offset_in_bytes()));
}


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


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

  // 'instantiator_reg' is the instantiator AbstractTypeArguments object
  // (or null).
  // A runtime call to instantiate the type is required.
  __ PushObject(Object::ZoneHandle());  // Make room for the result.
  __ PushObject(type());
  __ pushq(instantiator_reg);  // Push instantiator type arguments.
  compiler->GenerateCallRuntime(token_pos(),
                                deopt_id(),
                                kInstantiateTypeRuntimeEntry,
                                locs());
  __ Drop(2);  // Drop instantiator and uninstantiated type.
  __ popq(result_reg);  // Pop instantiated type.
  ASSERT(instantiator_reg == result_reg);
}


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


void InstantiateTypeArgumentsInstr::EmitNativeCode(
    FlowGraphCompiler* compiler) {
  Register instantiator_reg = locs()->in(0).reg();
  Register result_reg = locs()->out().reg();

  // 'instantiator_reg' is the instantiator AbstractTypeArguments object
  // (or null).
  ASSERT(!type_arguments().IsUninstantiatedIdentity() &&
         !type_arguments().CanShareInstantiatorTypeArguments(
             instantiator_class()));
  // If the instantiator is 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().IsRawInstantiatedRaw(len)) {
    const Immediate& raw_null =
        Immediate(reinterpret_cast<intptr_t>(Object::null()));
    __ cmpq(instantiator_reg, raw_null);
    __ j(EQUAL, &type_arguments_instantiated, Assembler::kNearJump);
  }
  // Instantiate non-null type arguments.
  // A runtime call to instantiate the type arguments is required.
  __ PushObject(Object::ZoneHandle());  // Make room for the result.
  __ PushObject(type_arguments());
  __ pushq(instantiator_reg);  // Push instantiator type arguments.
  compiler->GenerateCallRuntime(token_pos(),
                                deopt_id(),
                                kInstantiateTypeArgumentsRuntimeEntry,
                                locs());
  __ Drop(2);  // Drop instantiator and uninstantiated type arguments.
  __ popq(result_reg);  // Pop instantiated type arguments.
  __ Bind(&type_arguments_instantiated);
  ASSERT(instantiator_reg == result_reg);
}


LocationSummary*
ExtractConstructorTypeArgumentsInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  locs->set_in(0, Location::RequiresRegister());
  locs->set_out(Location::SameAsFirstInput());
  return locs;
}


void ExtractConstructorTypeArgumentsInstr::EmitNativeCode(
    FlowGraphCompiler* compiler) {
  Register instantiator_reg = locs()->in(0).reg();
  Register result_reg = locs()->out().reg();
  ASSERT(instantiator_reg == result_reg);

  // instantiator_reg is the instantiator type argument vector, i.e. an
  // AbstractTypeArguments object (or null).
  ASSERT(!type_arguments().IsUninstantiatedIdentity() &&
         !type_arguments().CanShareInstantiatorTypeArguments(
             instantiator_class()));
  // If the instantiator is 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;
  ASSERT(type_arguments().IsRawInstantiatedRaw(type_arguments().Length()));
  const Immediate& raw_null =
      Immediate(reinterpret_cast<intptr_t>(Object::null()));
  __ cmpq(instantiator_reg, raw_null);
  __ j(EQUAL, &type_arguments_instantiated, Assembler::kNearJump);
  // Instantiate non-null type arguments.
  // In the non-factory case, we rely on the allocation stub to
  // instantiate the type arguments.
  __ LoadObject(result_reg, type_arguments());
  // result_reg: uninstantiated type arguments.

  __ Bind(&type_arguments_instantiated);
  // result_reg: uninstantiated or instantiated type arguments.
}


LocationSummary*
ExtractConstructorInstantiatorInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  locs->set_in(0, Location::RequiresRegister());
  locs->set_out(Location::SameAsFirstInput());
  return locs;
}


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

  // instantiator_reg is the instantiator AbstractTypeArguments object
  // (or null).
  ASSERT(!type_arguments().IsUninstantiatedIdentity() &&
         !type_arguments().CanShareInstantiatorTypeArguments(
             instantiator_class()));

  // If the instantiator is null and if the type argument vector
  // instantiated from null becomes a vector of dynamic, then use null as
  // the type arguments and do not pass the instantiator.
  ASSERT(type_arguments().IsRawInstantiatedRaw(type_arguments().Length()));
  const Immediate& raw_null =
      Immediate(reinterpret_cast<intptr_t>(Object::null()));
  Label instantiator_not_null;
  __ cmpq(instantiator_reg, raw_null);
  __ j(NOT_EQUAL, &instantiator_not_null, Assembler::kNearJump);
  // Null was used in VisitExtractConstructorTypeArguments as the
  // instantiated type arguments, no proper instantiator needed.
  __ movq(instantiator_reg,
          Immediate(Smi::RawValue(StubCode::kNoInstantiator)));
  __ Bind(&instantiator_not_null);
  // instantiator_reg: instantiator or kNoInstantiator.
}


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


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

  __ movq(R10, Immediate(num_context_variables()));
  const ExternalLabel label("alloc_context",
                            StubCode::AllocateContextEntryPoint());
  compiler->GenerateCall(token_pos(),
                         &label,
                         PcDescriptors::kOther,
                         locs());
}


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


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

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


LocationSummary* CatchEntryInstr::MakeLocationSummary() const {
  return LocationSummary::Make(0,
                               Location::NoLocation(),
                               LocationSummary::kNoCall);
}


// Restore stack and initialize the two exception variables:
// exception and stack trace variables.
void CatchEntryInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  // 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));

  ASSERT(!exception_var().is_captured());
  ASSERT(!stacktrace_var().is_captured());
  __ movq(Address(RBP, exception_var().index() * kWordSize),
          kExceptionObjectReg);
  __ movq(Address(RBP, stacktrace_var().index() * kWordSize),
          kStackTraceObjectReg);
}


LocationSummary* CheckStackOverflowInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 0;
  const intptr_t kNumTemps = 1;
  LocationSummary* summary =
      new LocationSummary(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) {
    __ 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);
    compiler->pending_deoptimization_env_ = instruction_->env();
    compiler->GenerateCallRuntime(instruction_->token_pos(),
                                  instruction_->deopt_id(),
                                  kStackOverflowRuntimeEntry,
                                  instruction_->locs());

    if (FLAG_use_osr && !compiler->is_optimizing() && instruction_->in_loop()) {
      // In unoptimized code, record loop stack checks as possible OSR entries.
      compiler->AddCurrentDescriptor(PcDescriptors::kOsrEntry,
                                     instruction_->deopt_id(),
                                     0);  // No token position.
    }
    compiler->pending_deoptimization_env_ = NULL;
    compiler->RestoreLiveRegisters(instruction_->locs());
    __ jmp(exit_label());
  }

 private:
  CheckStackOverflowInstr* instruction_;
};


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.
  __ movq(temp, Immediate(Isolate::Current()->stack_limit_address()));
  __ cmpq(RSP, Address(temp, 0));
  __ 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());
    intptr_t threshold =
        FLAG_optimization_counter_threshold * (loop_depth() + 1);
    __ cmpq(FieldAddress(temp, Function::usage_counter_offset()),
            Immediate(threshold));
    __ j(GREATER_EQUAL, slow_path->entry_label());
  }
  __ Bind(slow_path->exit_label());
}


static void EmitJavascriptOverflowCheck(FlowGraphCompiler* compiler,
                                        Range* range,
                                        Label* overflow,
                                        Register result) {
  if (!range->IsWithin(-0x20000000000000LL, 0x20000000000000LL)) {
    ASSERT(overflow != NULL);
    __ cmpq(result, Immediate(-0x20000000000000LL));
    __ j(LESS, overflow);
    __ cmpq(result, Immediate(0x20000000000000LL));
    __ j(GREATER, overflow);
  }
}


static void EmitSmiShiftLeft(FlowGraphCompiler* compiler,
                             BinarySmiOpInstr* shift_left) {
  const bool is_truncating = shift_left->is_truncating();
  const LocationSummary& locs = *shift_left->locs();
  Register left = locs.in(0).reg();
  Register result = locs.out().reg();
  ASSERT(left == result);
  Label* deopt = shift_left->CanDeoptimize() ?
      compiler->AddDeoptStub(shift_left->deopt_id(), 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();
    if (value == 0) {
      // No code needed.
    } else if ((value < 0) || (value >= kCountLimit)) {
      // This condition may not be known earlier in some cases because
      // of constant propagation, inlining, etc.
      if ((value >=kCountLimit) && is_truncating) {
        __ xorq(result, result);
      } else {
        // Result is Mint or exception.
        __ jmp(deopt);
      }
    } else {
      if (!is_truncating) {
        // 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));
    }
    if (FLAG_throw_on_javascript_int_overflow) {
      EmitJavascriptOverflowCheck(compiler, shift_left->range(), deopt, result);
    }
    return;
  }

  // Right (locs.in(1)) is not constant.
  Register right = locs.in(1).reg();
  Range* right_range = shift_left->right()->definition()->range();
  if (shift_left->left()->BindsToConstant() && !is_truncating) {
    // 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) {
        __ cmpq(right, Immediate(0));
        __ j(NEGATIVE, deopt);
        return;
      }
      const intptr_t max_right = kSmiBits - Utils::HighestBit(left_int);
      const bool right_needs_check =
          (right_range == NULL) ||
          !right_range->IsWithin(0, max_right - 1);
      if (right_needs_check) {
        __ cmpq(right,
          Immediate(reinterpret_cast<int64_t>(Smi::New(max_right))));
        __ j(ABOVE_EQUAL, deopt);
      }
      __ SmiUntag(right);
      __ shlq(left, right);
    }
    if (FLAG_throw_on_javascript_int_overflow) {
      EmitJavascriptOverflowCheck(compiler, shift_left->range(), deopt, result);
    }
    return;
  }

  const bool right_needs_check =
      (right_range == NULL) || !right_range->IsWithin(0, (Smi::kBits - 1));
  ASSERT(right == RCX);  // Count must be in RCX
  if (is_truncating) {
    if (right_needs_check) {
      const bool right_may_be_negative =
          (right_range == NULL) ||
          !right_range->IsWithin(0, RangeBoundary::kPlusInfinity);
      if (right_may_be_negative) {
        ASSERT(shift_left->CanDeoptimize());
        __ cmpq(right, Immediate(0));
        __ j(NEGATIVE, deopt);
      }
      Label done, is_not_zero;
      __ cmpq(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());
      __ cmpq(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);
  }
  if (FLAG_throw_on_javascript_int_overflow) {
    EmitJavascriptOverflowCheck(compiler, shift_left->range(), deopt, result);
  }
}


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


LocationSummary* BinarySmiOpInstr::MakeLocationSummary() 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) &&
      CanBeImmediate(right_constant->value())) {
    const intptr_t kNumTemps = 0;
    LocationSummary* summary =
        new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
    summary->set_in(0, Location::RequiresRegister());
    summary->set_in(1, Location::Constant(right_constant->value()));
    summary->set_out(Location::SameAsFirstInput());
    return summary;
  }

  if (op_kind() == Token::kTRUNCDIV) {
    const intptr_t kNumTemps = 1;
    LocationSummary* summary =
        new LocationSummary(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->value()));
      summary->set_temp(0, Location::RequiresRegister());
      summary->set_out(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(Location::SameAsFirstInput());
      // Will be used for sign extension and division.
      summary->set_temp(0, Location::RegisterLocation(RDX));
    }
    return summary;
  } else if (op_kind() == Token::kSHR) {
    const intptr_t kNumTemps = 0;
    LocationSummary* summary =
        new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
    summary->set_in(0, Location::RequiresRegister());
    summary->set_in(1, Location::FixedRegisterOrSmiConstant(right(), RCX));
    summary->set_out(Location::SameAsFirstInput());
    return summary;
  } else if (op_kind() == Token::kSHL) {
    const intptr_t kNumTemps = 0;
    LocationSummary* summary =
        new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
    summary->set_in(0, Location::RequiresRegister());
    summary->set_in(1, Location::FixedRegisterOrSmiConstant(right(), RCX));
    if (!is_truncating()) {
      summary->AddTemp(Location::RequiresRegister());
    }
    summary->set_out(Location::SameAsFirstInput());
    return summary;
  } else {
    const intptr_t kNumTemps = 0;
    LocationSummary* summary =
        new LocationSummary(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(Location::SameAsFirstInput());
    return summary;
  }
}

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

  ASSERT(!is_truncating());
  Register left = locs()->in(0).reg();
  Register result = locs()->out().reg();
  ASSERT(left == result);
  Label* deopt = NULL;
  if (CanDeoptimize()) {
    deopt = compiler->AddDeoptStub(deopt_id(),
                                   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: {
        __ addq(left, Immediate(imm));
        if (deopt != NULL) __ j(OVERFLOW, deopt);
        break;
      }
      case Token::kSUB: {
        __ subq(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();
        if (value == 2) {
          __ shlq(left, Immediate(1));
        } else {
          __ imulq(left, Immediate(value));
        }
        if (deopt != NULL) __ j(OVERFLOW, deopt);
        break;
      }
      case Token::kTRUNCDIV: {
        const intptr_t value = Smi::Cast(constant).Value();
        if (value == 1) {
          // Do nothing.
          break;
        } else if (value == -1) {
          // Check the corner case of dividing the 'MIN_SMI' with -1, in which
          // case we cannot negate the result.
          __ cmpq(left, Immediate(0x8000000000000000));
          __ j(EQUAL, deopt);
          __ negq(left);
          break;
        }

        ASSERT((value != 0) && 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.
        __ andq(left, Immediate(imm));
        break;
      }
      case Token::kBIT_OR: {
        // No overflow check.
        __ orq(left, Immediate(imm));
        break;
      }
      case Token::kBIT_XOR: {
        // No overflow check.
        __ xorq(left, Immediate(imm));
        break;
      }

      case Token::kSHR: {
        // sarq operation masks the count to 6 bits.
        const intptr_t kCountLimit = 0x3F;
        intptr_t value = Smi::Cast(constant).Value();

        if (value == 0) {
          // TODO(vegorov): should be handled outside.
          break;
        } else if (value < 0) {
          // TODO(vegorov): should be handled outside.
          __ jmp(deopt);
          break;
        }

        value = value + kSmiTagSize;
        if (value >= kCountLimit) value = kCountLimit;

        __ sarq(left, Immediate(value));
        __ SmiTag(left);
        break;
      }

      default:
        UNREACHABLE();
        break;
    }
    if (FLAG_throw_on_javascript_int_overflow) {
      EmitJavascriptOverflowCheck(compiler, range(), deopt, result);
    }
    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;
    }
    if (FLAG_throw_on_javascript_int_overflow) {
      EmitJavascriptOverflowCheck(compiler, range(), deopt, result);
    }
    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);

      // 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.
      __ cmpq(result, Immediate(0x4000000000000000));
      __ j(EQUAL, deopt);
      __ Bind(&done);
      __ SmiTag(result);
      break;
    }
    case Token::kSHR: {
      if (CanDeoptimize()) {
        __ cmpq(right, Immediate(0));
        __ j(LESS, deopt);
      }
      __ SmiUntag(right);
      // sarq operation masks the count to 6 bits.
      const intptr_t kCountLimit = 0x3F;
      Range* right_range = this->right()->definition()->range();
      if ((right_range == NULL) ||
          !right_range->IsWithin(RangeBoundary::kMinusInfinity, kCountLimit)) {
        __ cmpq(right, Immediate(kCountLimit));
        Label count_ok;
        __ j(LESS, &count_ok, Assembler::kNearJump);
        __ movq(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::kMOD: {
      // TODO(srdjan): Implement.
      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;
  }
  if (FLAG_throw_on_javascript_int_overflow) {
    EmitJavascriptOverflowCheck(compiler, range(), deopt, result);
  }
}


LocationSummary* CheckEitherNonSmiInstr::MakeLocationSummary() 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_cid != kSmiCid) && (right_cid != kSmiCid);
  const intptr_t kNumTemps = need_temp ? 1 : 0;
  LocationSummary* summary =
    new LocationSummary(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(), kDeoptBinaryDoubleOp);
  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 (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* BoxDoubleInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs,
                          kNumTemps,
                          LocationSummary::kCallOnSlowPath);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_out(Location::RequiresRegister());
  return summary;
}


class BoxDoubleSlowPath : public SlowPathCode {
 public:
  explicit BoxDoubleSlowPath(BoxDoubleInstr* instruction)
      : instruction_(instruction) { }

  virtual void EmitNativeCode(FlowGraphCompiler* compiler) {
    __ Comment("BoxDoubleSlowPath");
    __ Bind(entry_label());
    const Class& double_class = compiler->double_class();
    const Code& stub =
        Code::Handle(StubCode::GetAllocationStubForClass(double_class));
    const ExternalLabel label(double_class.ToCString(), stub.EntryPoint());

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

    compiler->SaveLiveRegisters(locs);
    compiler->GenerateCall(Scanner::kDummyTokenIndex,  // No token position.
                           &label,
                           PcDescriptors::kOther,
                           locs);
    __ MoveRegister(locs->out().reg(), RAX);
    compiler->RestoreLiveRegisters(locs);

    __ jmp(exit_label());
  }

 private:
  BoxDoubleInstr* instruction_;
};


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

  Register out_reg = locs()->out().reg();
  XmmRegister value = locs()->in(0).fpu_reg();

  __ TryAllocate(compiler->double_class(),
                 slow_path->entry_label(),
                 Assembler::kFarJump,
                 out_reg);
  __ Bind(slow_path->exit_label());
  __ movsd(FieldAddress(out_reg, Double::value_offset()), value);
}


LocationSummary* UnboxDoubleInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  const bool needs_writable_input = (value()->Type()->ToCid() != kDoubleCid);
  summary->set_in(0, needs_writable_input
                     ? Location::WritableRegister()
                     : Location::RequiresRegister());
  summary->set_out(Location::RequiresFpuRegister());
  return summary;
}


void UnboxDoubleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  const intptr_t value_cid = value()->Type()->ToCid();
  const Register value = locs()->in(0).reg();
  const XmmRegister result = locs()->out().fpu_reg();

  if (value_cid == kDoubleCid) {
    __ movsd(result, FieldAddress(value, Double::value_offset()));
  } else if (value_cid == kSmiCid) {
    __ SmiUntag(value);  // Untag input before conversion.
    __ cvtsi2sd(result, value);
  } else {
    Label* deopt = compiler->AddDeoptStub(deopt_id_, kDeoptBinaryDoubleOp);
    Label is_smi, done;
    __ testq(value, Immediate(kSmiTagMask));
    __ j(ZERO, &is_smi);
    __ CompareClassId(value, kDoubleCid);
    __ j(NOT_EQUAL, deopt);
    __ movsd(result, FieldAddress(value, Double::value_offset()));
    __ jmp(&done);
    __ Bind(&is_smi);
    __ SmiUntag(value);
    __ cvtsi2sd(result, value);
    __ Bind(&done);
  }
}


LocationSummary* BoxFloat32x4Instr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs,
                          kNumTemps,
                          LocationSummary::kCallOnSlowPath);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_out(Location::RequiresRegister());
  return summary;
}


class BoxFloat32x4SlowPath : public SlowPathCode {
 public:
  explicit BoxFloat32x4SlowPath(BoxFloat32x4Instr* instruction)
      : instruction_(instruction) { }

  virtual void EmitNativeCode(FlowGraphCompiler* compiler) {
    __ Comment("BoxFloat32x4SlowPath");
    __ Bind(entry_label());
    const Class& float32x4_class = compiler->float32x4_class();
    const Code& stub =
        Code::Handle(StubCode::GetAllocationStubForClass(float32x4_class));
    const ExternalLabel label(float32x4_class.ToCString(), stub.EntryPoint());

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

    compiler->SaveLiveRegisters(locs);
    compiler->GenerateCall(Scanner::kDummyTokenIndex,  // No token position.
                           &label,
                           PcDescriptors::kOther,
                           locs);
    __ MoveRegister(locs->out().reg(), RAX);
    compiler->RestoreLiveRegisters(locs);

    __ jmp(exit_label());
  }

 private:
  BoxFloat32x4Instr* instruction_;
};


void BoxFloat32x4Instr::EmitNativeCode(FlowGraphCompiler* compiler) {
  BoxFloat32x4SlowPath* slow_path = new BoxFloat32x4SlowPath(this);
  compiler->AddSlowPathCode(slow_path);

  Register out_reg = locs()->out().reg();
  XmmRegister value = locs()->in(0).fpu_reg();

  __ TryAllocate(compiler->float32x4_class(),
                 slow_path->entry_label(),
                 Assembler::kFarJump,
                 out_reg);
  __ Bind(slow_path->exit_label());
  __ movups(FieldAddress(out_reg, Float32x4::value_offset()), value);
}


LocationSummary* UnboxFloat32x4Instr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  return LocationSummary::Make(kNumInputs,
                               Location::RequiresFpuRegister(),
                               LocationSummary::kNoCall);
}


void UnboxFloat32x4Instr::EmitNativeCode(FlowGraphCompiler* compiler) {
  const intptr_t value_cid = value()->Type()->ToCid();
  const Register value = locs()->in(0).reg();
  const XmmRegister result = locs()->out().fpu_reg();

  if (value_cid != kFloat32x4Cid) {
    Label* deopt = compiler->AddDeoptStub(deopt_id_, kDeoptCheckClass);
    __ testq(value, Immediate(kSmiTagMask));
    __ j(ZERO, deopt);
    __ CompareClassId(value, kFloat32x4Cid);
    __ j(NOT_EQUAL, deopt);
  }
  __ movups(result, FieldAddress(value, Float32x4::value_offset()));
}


LocationSummary* BoxUint32x4Instr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs,
                          kNumTemps,
                          LocationSummary::kCallOnSlowPath);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_out(Location::RequiresRegister());
  return summary;
}


class BoxUint32x4SlowPath : public SlowPathCode {
 public:
  explicit BoxUint32x4SlowPath(BoxUint32x4Instr* instruction)
      : instruction_(instruction) { }

  virtual void EmitNativeCode(FlowGraphCompiler* compiler) {
    __ Comment("BoxUint32x4SlowPath");
    __ Bind(entry_label());
    const Class& uint32x4_class = compiler->uint32x4_class();
    const Code& stub =
        Code::Handle(StubCode::GetAllocationStubForClass(uint32x4_class));
    const ExternalLabel label(uint32x4_class.ToCString(), stub.EntryPoint());

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

    compiler->SaveLiveRegisters(locs);
    compiler->GenerateCall(Scanner::kDummyTokenIndex,  // No token position.
                           &label,
                           PcDescriptors::kOther,
                           locs);
    __ MoveRegister(locs->out().reg(), RAX);
    compiler->RestoreLiveRegisters(locs);

    __ jmp(exit_label());
  }

 private:
  BoxUint32x4Instr* instruction_;
};


void BoxUint32x4Instr::EmitNativeCode(FlowGraphCompiler* compiler) {
  BoxUint32x4SlowPath* slow_path = new BoxUint32x4SlowPath(this);
  compiler->AddSlowPathCode(slow_path);

  Register out_reg = locs()->out().reg();
  XmmRegister value = locs()->in(0).fpu_reg();

  __ TryAllocate(compiler->uint32x4_class(),
                 slow_path->entry_label(),
                 Assembler::kFarJump,
                 out_reg);
  __ Bind(slow_path->exit_label());
  __ movups(FieldAddress(out_reg, Uint32x4::value_offset()), value);
}


LocationSummary* UnboxUint32x4Instr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  summary->set_out(Location::RequiresFpuRegister());
  return summary;
}


void UnboxUint32x4Instr::EmitNativeCode(FlowGraphCompiler* compiler) {
  const intptr_t value_cid = value()->Type()->ToCid();
  const Register value = locs()->in(0).reg();
  const XmmRegister result = locs()->out().fpu_reg();

  if (value_cid != kUint32x4Cid) {
    Label* deopt = compiler->AddDeoptStub(deopt_id_, kDeoptCheckClass);
    __ testq(value, Immediate(kSmiTagMask));
    __ j(ZERO, deopt);
    __ CompareClassId(value, kUint32x4Cid);
    __ j(NOT_EQUAL, deopt);
  }
  __ movups(result, FieldAddress(value, Uint32x4::value_offset()));
}


LocationSummary* BinaryDoubleOpInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_in(1, Location::RequiresFpuRegister());
  summary->set_out(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().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* BinaryFloat32x4OpInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_in(1, Location::RequiresFpuRegister());
  summary->set_out(Location::SameAsFirstInput());
  return summary;
}


void BinaryFloat32x4OpInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister left = locs()->in(0).fpu_reg();
  XmmRegister right = locs()->in(1).fpu_reg();

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

  switch (op_kind()) {
    case Token::kADD: __ addps(left, right); break;
    case Token::kSUB: __ subps(left, right); break;
    case Token::kMUL: __ mulps(left, right); break;
    case Token::kDIV: __ divps(left, right); break;
    default: UNREACHABLE();
  }
}


LocationSummary* Float32x4ShuffleInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_out(Location::SameAsFirstInput());
  return summary;
}


void Float32x4ShuffleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister value = locs()->in(0).fpu_reg();

  ASSERT(locs()->out().fpu_reg() == value);

  switch (op_kind()) {
    case MethodRecognizer::kFloat32x4ShuffleX:
      __ shufps(value, value, Immediate(0x00));
      __ cvtss2sd(value, value);
      break;
    case MethodRecognizer::kFloat32x4ShuffleY:
      __ shufps(value, value, Immediate(0x55));
      __ cvtss2sd(value, value);
      break;
    case MethodRecognizer::kFloat32x4ShuffleZ:
      __ shufps(value, value, Immediate(0xAA));
      __ cvtss2sd(value, value);
      break;
    case MethodRecognizer::kFloat32x4ShuffleW:
      __ shufps(value, value, Immediate(0xFF));
      __ cvtss2sd(value, value);
      break;
    case MethodRecognizer::kFloat32x4Shuffle:
      __ shufps(value, value, Immediate(mask_));
      break;
    default: UNREACHABLE();
  }
}


LocationSummary* Float32x4ConstructorInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 4;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_in(1, Location::RequiresFpuRegister());
  summary->set_in(2, Location::RequiresFpuRegister());
  summary->set_in(3, Location::RequiresFpuRegister());
  summary->set_out(Location::SameAsFirstInput());
  return summary;
}


void Float32x4ConstructorInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister v0 = locs()->in(0).fpu_reg();
  XmmRegister v1 = locs()->in(1).fpu_reg();
  XmmRegister v2 = locs()->in(2).fpu_reg();
  XmmRegister v3 = locs()->in(3).fpu_reg();
  ASSERT(v0 == locs()->out().fpu_reg());
  __ subq(RSP, Immediate(16));
  __ cvtsd2ss(v0, v0);
  __ movss(Address(RSP, 0), v0);
  __ movsd(v0, v1);
  __ cvtsd2ss(v0, v0);
  __ movss(Address(RSP, 4), v0);
  __ movsd(v0, v2);
  __ cvtsd2ss(v0, v0);
  __ movss(Address(RSP, 8), v0);
  __ movsd(v0, v3);
  __ cvtsd2ss(v0, v0);
  __ movss(Address(RSP, 12), v0);
  __ movups(v0, Address(RSP, 0));
  __ addq(RSP, Immediate(16));
}


LocationSummary* Float32x4ZeroInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 0;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_out(Location::RequiresFpuRegister());
  return summary;
}


void Float32x4ZeroInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister value = locs()->out().fpu_reg();
  __ xorps(value, value);
}


LocationSummary* Float32x4SplatInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_out(Location::SameAsFirstInput());
  return summary;
}


void Float32x4SplatInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister value = locs()->out().fpu_reg();
  ASSERT(locs()->in(0).fpu_reg() == locs()->out().fpu_reg());
  // Convert to Float32.
  __ cvtsd2ss(value, value);
  // Splat across all lanes.
  __ shufps(value, value, Immediate(0x00));
}


LocationSummary* Float32x4ComparisonInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_in(1, Location::RequiresFpuRegister());
  summary->set_out(Location::SameAsFirstInput());
  return summary;
}


void Float32x4ComparisonInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister left = locs()->in(0).fpu_reg();
  XmmRegister right = locs()->in(1).fpu_reg();

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

  switch (op_kind()) {
    case MethodRecognizer::kFloat32x4Equal:
      __ cmppseq(left, right);
      break;
    case MethodRecognizer::kFloat32x4NotEqual:
      __ cmppsneq(left, right);
      break;
    case MethodRecognizer::kFloat32x4GreaterThan:
      __ cmppsnle(left, right);
      break;
    case MethodRecognizer::kFloat32x4GreaterThanOrEqual:
      __ cmppsnlt(left, right);
      break;
    case MethodRecognizer::kFloat32x4LessThan:
      __ cmppslt(left, right);
      break;
    case MethodRecognizer::kFloat32x4LessThanOrEqual:
      __ cmppsle(left, right);
      break;

    default: UNREACHABLE();
  }
}


LocationSummary* Float32x4MinMaxInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_in(1, Location::RequiresFpuRegister());
  summary->set_out(Location::SameAsFirstInput());
  return summary;
}


void Float32x4MinMaxInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister left = locs()->in(0).fpu_reg();
  XmmRegister right = locs()->in(1).fpu_reg();

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

  switch (op_kind()) {
    case MethodRecognizer::kFloat32x4Min:
      __ minps(left, right);
      break;
    case MethodRecognizer::kFloat32x4Max:
      __ maxps(left, right);
      break;
    default: UNREACHABLE();
  }
}


LocationSummary* Float32x4ScaleInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_in(1, Location::RequiresFpuRegister());
  summary->set_out(Location::SameAsFirstInput());
  return summary;
}


void Float32x4ScaleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister left = locs()->in(0).fpu_reg();
  XmmRegister right = locs()->in(1).fpu_reg();

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

  switch (op_kind()) {
    case MethodRecognizer::kFloat32x4Scale:
      __ cvtsd2ss(left, left);
      __ shufps(left, left, Immediate(0x00));
      __ mulps(left, right);
      break;
    default: UNREACHABLE();
  }
}


LocationSummary* Float32x4SqrtInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_out(Location::SameAsFirstInput());
  return summary;
}


void Float32x4SqrtInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister left = locs()->in(0).fpu_reg();

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

  switch (op_kind()) {
    case MethodRecognizer::kFloat32x4Sqrt:
      __ sqrtps(left);
      break;
    case MethodRecognizer::kFloat32x4Reciprocal:
      __ reciprocalps(left);
      break;
    case MethodRecognizer::kFloat32x4ReciprocalSqrt:
      __ rsqrtps(left);
      break;
    default: UNREACHABLE();
  }
}


LocationSummary* Float32x4ZeroArgInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_out(Location::SameAsFirstInput());
  return summary;
}


void Float32x4ZeroArgInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister left = locs()->in(0).fpu_reg();

  ASSERT(locs()->out().fpu_reg() == left);
  switch (op_kind()) {
    case MethodRecognizer::kFloat32x4Negate:
      __ negateps(left);
      break;
    case MethodRecognizer::kFloat32x4Absolute:
      __ absps(left);
      break;
    default: UNREACHABLE();
  }
}


LocationSummary* Float32x4ClampInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 3;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_in(1, Location::RequiresFpuRegister());
  summary->set_in(2, Location::RequiresFpuRegister());
  summary->set_out(Location::SameAsFirstInput());
  return summary;
}


void Float32x4ClampInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister left = locs()->in(0).fpu_reg();
  XmmRegister lower = locs()->in(1).fpu_reg();
  XmmRegister upper = locs()->in(2).fpu_reg();
  ASSERT(locs()->out().fpu_reg() == left);
  __ minps(left, upper);
  __ maxps(left, lower);
}


LocationSummary* Float32x4WithInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_in(1, Location::RequiresFpuRegister());
  summary->set_out(Location::SameAsFirstInput());
  return summary;
}


void Float32x4WithInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister replacement = locs()->in(0).fpu_reg();
  XmmRegister value = locs()->in(1).fpu_reg();

  ASSERT(locs()->out().fpu_reg() == replacement);

  switch (op_kind()) {
    case MethodRecognizer::kFloat32x4WithX:
      __ cvtsd2ss(replacement, replacement);
      __ subq(RSP, Immediate(16));
      // Move value to stack.
      __ movups(Address(RSP, 0), value);
      // Write over X value.
      __ movss(Address(RSP, 0), replacement);
      // Move updated value into output register.
      __ movups(replacement, Address(RSP, 0));
      __ addq(RSP, Immediate(16));
      break;
    case MethodRecognizer::kFloat32x4WithY:
      __ cvtsd2ss(replacement, replacement);
      __ subq(RSP, Immediate(16));
      // Move value to stack.
      __ movups(Address(RSP, 0), value);
      // Write over Y value.
      __ movss(Address(RSP, 4), replacement);
      // Move updated value into output register.
      __ movups(replacement, Address(RSP, 0));
      __ addq(RSP, Immediate(16));
      break;
    case MethodRecognizer::kFloat32x4WithZ:
      __ cvtsd2ss(replacement, replacement);
      __ subq(RSP, Immediate(16));
      // Move value to stack.
      __ movups(Address(RSP, 0), value);
      // Write over Z value.
      __ movss(Address(RSP, 8), replacement);
      // Move updated value into output register.
      __ movups(replacement, Address(RSP, 0));
      __ addq(RSP, Immediate(16));
      break;
    case MethodRecognizer::kFloat32x4WithW:
      __ cvtsd2ss(replacement, replacement);
      __ subq(RSP, Immediate(16));
      // Move value to stack.
      __ movups(Address(RSP, 0), value);
      // Write over W value.
      __ movss(Address(RSP, 12), replacement);
      // Move updated value into output register.
      __ movups(replacement, Address(RSP, 0));
      __ addq(RSP, Immediate(16));
      break;
    default: UNREACHABLE();
  }
}


LocationSummary* Float32x4ToUint32x4Instr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_out(Location::SameAsFirstInput());
  return summary;
}


void Float32x4ToUint32x4Instr::EmitNativeCode(FlowGraphCompiler* compiler) {
  // NOP.
}


LocationSummary* Float32x4TwoArgShuffleInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_in(1, Location::RequiresFpuRegister());
  summary->set_out(Location::SameAsFirstInput());
  return summary;
}


void Float32x4TwoArgShuffleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister left = locs()->in(0).fpu_reg();
  XmmRegister right = locs()->in(1).fpu_reg();

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

  switch (op_kind()) {
    case MethodRecognizer::kFloat32x4WithZWInXY:
      __ movhlps(left, right);
    break;
    case MethodRecognizer::kFloat32x4InterleaveXY:
      __ unpcklps(left, right);
    break;
    case MethodRecognizer::kFloat32x4InterleaveZW:
      __ unpckhps(left, right);
    break;
    case MethodRecognizer::kFloat32x4InterleaveXYPairs:
      __ unpcklpd(left, right);
    break;
    case MethodRecognizer::kFloat32x4InterleaveZWPairs:
      __ unpckhpd(left, right);
    break;
    default: UNREACHABLE();
  }
}


LocationSummary* Uint32x4BoolConstructorInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 4;
  const intptr_t kNumTemps = 1;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  summary->set_in(1, Location::RequiresRegister());
  summary->set_in(2, Location::RequiresRegister());
  summary->set_in(3, Location::RequiresRegister());
  summary->set_temp(0, Location::RequiresRegister());
  summary->set_out(Location::RequiresFpuRegister());
  return summary;
}


void Uint32x4BoolConstructorInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register v0 = locs()->in(0).reg();
  Register v1 = locs()->in(1).reg();
  Register v2 = locs()->in(2).reg();
  Register v3 = locs()->in(3).reg();
  Register temp = locs()->temp(0).reg();
  XmmRegister result = locs()->out().fpu_reg();
  Label x_false, x_done;
  Label y_false, y_done;
  Label z_false, z_done;
  Label w_false, w_done;
  __ subq(RSP, Immediate(16));

  __ CompareObject(v0, Bool::True());
  __ j(NOT_EQUAL, &x_false);
  __ movq(temp, Immediate(0xFFFFFFFF));
  __ jmp(&x_done);
  __ Bind(&x_false);
  __ movq(temp, Immediate(0x0));
  __ Bind(&x_done);
  __ movl(Address(RSP, 0), temp);

  __ CompareObject(v1, Bool::True());
  __ j(NOT_EQUAL, &y_false);
  __ movq(temp, Immediate(0xFFFFFFFF));
  __ jmp(&y_done);
  __ Bind(&y_false);
  __ movq(temp, Immediate(0x0));
  __ Bind(&y_done);
  __ movl(Address(RSP, 4), temp);

  __ CompareObject(v2, Bool::True());
  __ j(NOT_EQUAL, &z_false);
  __ movq(temp, Immediate(0xFFFFFFFF));
  __ jmp(&z_done);
  __ Bind(&z_false);
  __ movq(temp, Immediate(0x0));
  __ Bind(&z_done);
  __ movl(Address(RSP, 8), temp);

  __ CompareObject(v3, Bool::True());
  __ j(NOT_EQUAL, &w_false);
  __ movq(temp, Immediate(0xFFFFFFFF));
  __ jmp(&w_done);
  __ Bind(&w_false);
  __ movq(temp, Immediate(0x0));
  __ Bind(&w_done);
  __ movl(Address(RSP, 12), temp);

  __ movups(result, Address(RSP, 0));
  __ addq(RSP, Immediate(16));
}


LocationSummary* Uint32x4GetFlagInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_out(Location::RequiresRegister());
  return summary;
}


void Uint32x4GetFlagInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister value = locs()->in(0).fpu_reg();
  Register result = locs()->out().reg();
  Label done;
  Label non_zero;
  __ subq(RSP, Immediate(16));
  // Move value to stack.
  __ movups(Address(RSP, 0), value);
  switch (op_kind()) {
    case MethodRecognizer::kUint32x4GetFlagX:
      __ movl(result, Address(RSP, 0));
      break;
    case MethodRecognizer::kUint32x4GetFlagY:
      __ movl(result, Address(RSP, 4));
      break;
    case MethodRecognizer::kUint32x4GetFlagZ:
      __ movl(result, Address(RSP, 8));
      break;
    case MethodRecognizer::kUint32x4GetFlagW:
      __ movl(result, Address(RSP, 12));
      break;
    default: UNREACHABLE();
  }
  __ addq(RSP, Immediate(16));
  __ testl(result, result);
  __ j(NOT_ZERO, &non_zero, Assembler::kNearJump);
  __ LoadObject(result, Bool::False());
  __ jmp(&done);
  __ Bind(&non_zero);
  __ LoadObject(result, Bool::True());
  __ Bind(&done);
}


LocationSummary* Uint32x4SelectInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 3;
  const intptr_t kNumTemps = 1;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_in(1, Location::RequiresFpuRegister());
  summary->set_in(2, Location::RequiresFpuRegister());
  summary->set_temp(0, Location::RequiresFpuRegister());
  summary->set_out(Location::SameAsFirstInput());
  return summary;
}


void Uint32x4SelectInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister mask = locs()->in(0).fpu_reg();
  XmmRegister trueValue = locs()->in(1).fpu_reg();
  XmmRegister falseValue = locs()->in(2).fpu_reg();
  XmmRegister out = locs()->out().fpu_reg();
  XmmRegister temp = locs()->temp(0).fpu_reg();
  ASSERT(out == mask);
  // 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);
}


LocationSummary* Uint32x4SetFlagInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 1;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_in(1, Location::RequiresRegister());
  summary->set_temp(0, Location::RequiresRegister());
  summary->set_out(Location::SameAsFirstInput());
  return summary;
}


void Uint32x4SetFlagInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister mask = locs()->in(0).fpu_reg();
  Register flag = locs()->in(1).reg();
  Register temp = locs()->temp(0).reg();
  ASSERT(mask == locs()->out().fpu_reg());
  __ subq(RSP, Immediate(16));
  // Copy mask to stack.
  __ movups(Address(RSP, 0), mask);
  Label falsePath, exitPath;
  __ CompareObject(flag, Bool::True());
  __ j(NOT_EQUAL, &falsePath);
  switch (op_kind()) {
    case MethodRecognizer::kUint32x4WithFlagX:
      __ movq(temp, Immediate(0xFFFFFFFF));
      __ movl(Address(RSP, 0), temp);
      __ jmp(&exitPath);
      __ Bind(&falsePath);
      __ movq(temp, Immediate(0x0));
      __ movl(Address(RSP, 0), temp);
    break;
    case MethodRecognizer::kUint32x4WithFlagY:
      __ movq(temp, Immediate(0xFFFFFFFF));
      __ movl(Address(RSP, 4), temp);
      __ jmp(&exitPath);
      __ Bind(&falsePath);
      __ movq(temp, Immediate(0x0));
      __ movl(Address(RSP, 4), temp);
    break;
    case MethodRecognizer::kUint32x4WithFlagZ:
      __ movq(temp, Immediate(0xFFFFFFFF));
      __ movl(Address(RSP, 8), temp);
      __ jmp(&exitPath);
      __ Bind(&falsePath);
      __ movq(temp, Immediate(0x0));
      __ movl(Address(RSP, 8), temp);
    break;
    case MethodRecognizer::kUint32x4WithFlagW:
      __ movq(temp, Immediate(0xFFFFFFFF));
      __ movl(Address(RSP, 12), temp);
      __ jmp(&exitPath);
      __ Bind(&falsePath);
      __ movq(temp, Immediate(0x0));
      __ movl(Address(RSP, 12), temp);
    break;
    default: UNREACHABLE();
  }
  __ Bind(&exitPath);
  // Copy mask back to register.
  __ movups(mask, Address(RSP, 0));
  __ addq(RSP, Immediate(16));
}


LocationSummary* Uint32x4ToFloat32x4Instr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_out(Location::SameAsFirstInput());
  return summary;
}


void Uint32x4ToFloat32x4Instr::EmitNativeCode(FlowGraphCompiler* compiler) {
  // NOP.
}


LocationSummary* BinaryUint32x4OpInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_in(1, Location::RequiresFpuRegister());
  summary->set_out(Location::SameAsFirstInput());
  return summary;
}


void BinaryUint32x4OpInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister left = locs()->in(0).fpu_reg();
  XmmRegister right = locs()->in(1).fpu_reg();
  ASSERT(left == locs()->out().fpu_reg());
  switch (op_kind()) {
    case Token::kBIT_AND: {
      __ andps(left, right);
      break;
    }
    case Token::kBIT_OR: {
      __ orps(left, right);
      break;
    }
    case Token::kBIT_XOR: {
      __ xorps(left, right);
      break;
    }
    default: UNREACHABLE();
  }
}


LocationSummary* MathSqrtInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresFpuRegister());
  summary->set_out(Location::RequiresFpuRegister());
  return summary;
}


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


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


LocationSummary* MathMinMaxInstr::MakeLocationSummary() const {
  if (result_cid() == kDoubleCid) {
    const intptr_t kNumInputs = 2;
    const intptr_t kNumTemps = 1;
    LocationSummary* summary =
        new LocationSummary(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(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 LocationSummary(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(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().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;
    __ movq(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().reg();
  __ cmpq(left, right);
  ASSERT(result == left);
  if (is_min) {
    __ cmovgeq(result, right);
  } else {
    __ cmovlessq(result, right);
  }
}


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


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


void SmiToDoubleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register value = locs()->in(0).reg();
  FpuRegister result = locs()->out().fpu_reg();
  __ SmiUntag(value);
  __ cvtsi2sd(result, value);
}


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


void DoubleToIntegerInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register result = locs()->out().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);
  if (FLAG_throw_on_javascript_int_overflow) {
    EmitJavascriptOverflowCheck(compiler, range(), &do_call, result);
  }
  __ jmp(&done);
  __ Bind(&do_call);
  ASSERT(instance_call()->HasICData());
  const ICData& ic_data = *instance_call()->ic_data();
  ASSERT((ic_data.NumberOfChecks() == 1));
  const Function& target = Function::ZoneHandle(ic_data.GetTargetAt(0));

  const intptr_t kNumberOfArguments = 1;
  __ pushq(value_obj);
  compiler->GenerateStaticCall(deopt_id(),
                               instance_call()->token_pos(),
                               target,
                               kNumberOfArguments,
                               Object::null_array(),  // No argument names.
                               locs());
  __ Bind(&done);
}


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


void DoubleToSmiInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Label* deopt = compiler->AddDeoptStub(deopt_id(), kDeoptDoubleToSmi);
  Register result = locs()->out().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);
  if (FLAG_throw_on_javascript_int_overflow) {
    EmitJavascriptOverflowCheck(compiler, range(), deopt, result);
  }
}


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


void DoubleToDoubleInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  XmmRegister value = locs()->in(0).fpu_reg();
  XmmRegister result = locs()->out().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* InvokeMathCFunctionInstr::MakeLocationSummary() 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 = 0;
  LocationSummary* result =
      new LocationSummary(InputCount(), kNumTemps, LocationSummary::kCall);
  result->set_in(0, Location::FpuRegisterLocation(XMM1));
  if (InputCount() == 2) {
    result->set_in(1, Location::FpuRegisterLocation(XMM2));
  }
  result->set_out(Location::FpuRegisterLocation(XMM1));
  return result;
}


void InvokeMathCFunctionInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  ASSERT(locs()->in(0).fpu_reg() == XMM1);
  __ EnterFrame(0);
  __ ReserveAlignedFrameSpace(0);
  __ movaps(XMM0, locs()->in(0).fpu_reg());
  if (InputCount() == 2) {
    ASSERT(locs()->in(1).fpu_reg() == XMM2);
    __ movaps(XMM1, locs()->in(1).fpu_reg());
  }
  // For pow-function return NaN if exponent is NaN.
  Label do_call, skip_call;
  if (recognized_kind() == MethodRecognizer::kDoublePow) {
    XmmRegister exp = locs()->in(1).fpu_reg();
    __ comisd(exp, exp);
    __ j(PARITY_ODD, &do_call, Assembler::kNearJump);  // NaN -> false;
    // Exponent is NaN, return NaN.
    __ movaps(locs()->out().fpu_reg(), exp);
    __ jmp(&skip_call, Assembler::kNearJump);
  }
  __ Bind(&do_call);
  __ CallRuntime(TargetFunction());
  __ movaps(locs()->out().fpu_reg(), XMM0);
  __ Bind(&skip_call);
  __ leave();
}


LocationSummary* PolymorphicInstanceCallInstr::MakeLocationSummary() const {
  return MakeCallSummary();
}


void PolymorphicInstanceCallInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Label* deopt = compiler->AddDeoptStub(deopt_id(),
                                        kDeoptPolymorphicInstanceCallTestFail);
  if (ic_data().NumberOfChecks() == 0) {
    __ jmp(deopt);
    return;
  }
  ASSERT(ic_data().num_args_tested() == 1);
  if (!with_checks()) {
    ASSERT(ic_data().HasOneTarget());
    const Function& target = Function::ZoneHandle(ic_data().GetTargetAt(0));
    compiler->GenerateStaticCall(deopt_id(),
                                 instance_call()->token_pos(),
                                 target,
                                 instance_call()->ArgumentCount(),
                                 instance_call()->argument_names(),
                                 locs());
    return;
  }

  // Load receiver into RAX.
  __ movq(RAX,
      Address(RSP, (instance_call()->ArgumentCount() - 1) * kWordSize));
  LoadValueCid(compiler, RDI, RAX,
               (ic_data().GetReceiverClassIdAt(0) == kSmiCid) ? NULL : deopt);
  compiler->EmitTestAndCall(ic_data(),
                            RDI,  // Class id register.
                            instance_call()->ArgumentCount(),
                            instance_call()->argument_names(),
                            deopt,
                            deopt_id(),
                            instance_call()->token_pos(),
                            locs());
}


LocationSummary* BranchInstr::MakeLocationSummary() const {
  UNREACHABLE();
  return NULL;
}


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


LocationSummary* CheckClassInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  summary->set_in(0, Location::RequiresRegister());
  if (!null_check()) {
    summary->AddTemp(Location::RequiresRegister());
  }
  return summary;
}


void CheckClassInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  if (null_check()) {
    Label* deopt = compiler->AddDeoptStub(deopt_id(),
                                          kDeoptCheckClass);
    const Immediate& raw_null =
        Immediate(reinterpret_cast<intptr_t>(Object::null()));
    __ cmpq(locs()->in(0).reg(), raw_null);
    __ j(EQUAL, deopt);
    return;
  }

  ASSERT((unary_checks().GetReceiverClassIdAt(0) != kSmiCid) ||
         (unary_checks().NumberOfChecks() > 1));
  Register value = locs()->in(0).reg();
  Register temp = locs()->temp(0).reg();
  Label* deopt = compiler->AddDeoptStub(deopt_id(),
                                        kDeoptCheckClass);
  Label is_ok;
  intptr_t cix = 0;
  if (unary_checks().GetReceiverClassIdAt(cix) == kSmiCid) {
    __ testq(value, Immediate(kSmiTagMask));
    __ j(ZERO, &is_ok);
    cix++;  // Skip first check.
  } else {
    __ testq(value, Immediate(kSmiTagMask));
    __ j(ZERO, deopt);
  }
  __ LoadClassId(temp, value);
  const intptr_t num_checks = unary_checks().NumberOfChecks();
  const bool use_near_jump = num_checks < 5;
  for (intptr_t i = cix; i < num_checks; i++) {
    ASSERT(unary_checks().GetReceiverClassIdAt(i) != kSmiCid);
    __ cmpl(temp, Immediate(unary_checks().GetReceiverClassIdAt(i)));
    if (i == (num_checks - 1)) {
      __ j(NOT_EQUAL, deopt);
    } else {
      if (use_near_jump) {
        __ j(EQUAL, &is_ok, Assembler::kNearJump);
      } else {
        __ j(EQUAL, &is_ok);
      }
    }
  }
  __ Bind(&is_ok);
}


LocationSummary* CheckSmiInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 1;
  const intptr_t kNumTemps = 0;
  LocationSummary* summary =
      new LocationSummary(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(),
                                        kDeoptCheckSmi);
  __ testq(value, Immediate(kSmiTagMask));
  __ j(NOT_ZERO, deopt);
}


LocationSummary* CheckArrayBoundInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs =
      new LocationSummary(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) {
  Label* deopt = compiler->AddDeoptStub(deopt_id(), kDeoptCheckArrayBound);

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

  if (length_loc.IsConstant() && index_loc.IsConstant()) {
    // TODO(srdjan): remove this code once failures are fixed.
    if ((Smi::Cast(length_loc.constant()).Value() >
         Smi::Cast(index_loc.constant()).Value()) &&
        (Smi::Cast(index_loc.constant()).Value() >= 0)) {
      // This CheckArrayBoundInstr should have been eliminated.
      return;
    }
    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;
  }

  if (index_loc.IsConstant()) {
    Register length = length_loc.reg();
    const Smi& index = Smi::Cast(index_loc.constant());
    __ cmpq(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();
    __ cmpq(index, Immediate(reinterpret_cast<int64_t>(length.raw())));
    __ j(ABOVE_EQUAL, deopt);
  } else {
    Register length = length_loc.reg();
    Register index = index_loc.reg();
    __ cmpq(index, length);
    __ j(ABOVE_EQUAL, deopt);
  }
}


LocationSummary* UnboxIntegerInstr::MakeLocationSummary() const {
  UNIMPLEMENTED();
  return NULL;
}


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


LocationSummary* BoxIntegerInstr::MakeLocationSummary() const {
  UNIMPLEMENTED();
  return NULL;
}


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


LocationSummary* BinaryMintOpInstr::MakeLocationSummary() const {
  UNIMPLEMENTED();
  return NULL;
}


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


LocationSummary* UnaryMintOpInstr::MakeLocationSummary() const {
  UNIMPLEMENTED();
  return NULL;
}


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


LocationSummary* ShiftMintOpInstr::MakeLocationSummary() const {
  UNIMPLEMENTED();
  return NULL;
}


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


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


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


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


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


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


void GotoInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  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()));
  }
}


void ControlInstruction::EmitBranchOnValue(FlowGraphCompiler* compiler,
                                           bool value) {
  if (value && !compiler->CanFallThroughTo(true_successor())) {
    __ jmp(compiler->GetJumpLabel(true_successor()));
  } else if (!value && !compiler->CanFallThroughTo(false_successor())) {
    __ jmp(compiler->GetJumpLabel(false_successor()));
  }
}


void ControlInstruction::EmitBranchOnCondition(FlowGraphCompiler* compiler,
                                               Condition true_condition) {
  if (compiler->CanFallThroughTo(false_successor())) {
    // If the next block is the false successor we will fall through to it.
    __ j(true_condition, compiler->GetJumpLabel(true_successor()));
  } else {
    // If the next block is the true successor we negate comparison and fall
    // through to it.
    Condition false_condition = NegateCondition(true_condition);
    __ j(false_condition, compiler->GetJumpLabel(false_successor()));

    // Fall through or jump to the true successor.
    if (!compiler->CanFallThroughTo(true_successor())) {
      __ jmp(compiler->GetJumpLabel(true_successor()));
    }
  }
}


LocationSummary* CurrentContextInstr::MakeLocationSummary() const {
  return LocationSummary::Make(0,
                               Location::RequiresRegister(),
                               LocationSummary::kNoCall);
}


void CurrentContextInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  __ MoveRegister(locs()->out().reg(), CTX);
}


LocationSummary* StrictCompareInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  locs->set_in(0, Location::RegisterOrConstant(left()));
  locs->set_in(1, Location::RegisterOrConstant(right()));
  locs->set_out(Location::RequiresRegister());
  return locs;
}


// Special code for numbers (compare values instead of references.)
void StrictCompareInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  ASSERT(kind() == Token::kEQ_STRICT || kind() == Token::kNE_STRICT);
  Location left = locs()->in(0);
  Location right = locs()->in(1);
  if (left.IsConstant() && right.IsConstant()) {
    // TODO(vegorov): should be eliminated earlier by constant propagation.
    const bool result = (kind() == Token::kEQ_STRICT) ?
        left.constant().raw() == right.constant().raw() :
        left.constant().raw() != right.constant().raw();
    __ LoadObject(locs()->out().reg(), result ? Bool::True() : Bool::False());
    return;
  }
  if (left.IsConstant()) {
    compiler->EmitEqualityRegConstCompare(right.reg(),
                                          left.constant(),
                                          needs_number_check(),
                                          token_pos());
  } else if (right.IsConstant()) {
    compiler->EmitEqualityRegConstCompare(left.reg(),
                                          right.constant(),
                                          needs_number_check(),
                                          token_pos());
  } else {
    compiler->EmitEqualityRegRegCompare(left.reg(),
                                       right.reg(),
                                       needs_number_check(),
                                       token_pos());
  }

  Register result = locs()->out().reg();
  Label load_true, done;
  Condition true_condition = (kind() == Token::kEQ_STRICT) ? EQUAL : NOT_EQUAL;
  __ j(true_condition, &load_true, Assembler::kNearJump);
  __ LoadObject(result, Bool::False());
  __ jmp(&done, Assembler::kNearJump);
  __ Bind(&load_true);
  __ LoadObject(result, Bool::True());
  __ Bind(&done);
}


void StrictCompareInstr::EmitBranchCode(FlowGraphCompiler* compiler,
                                        BranchInstr* branch) {
  ASSERT(kind() == Token::kEQ_STRICT || kind() == Token::kNE_STRICT);
  Location left = locs()->in(0);
  Location right = locs()->in(1);
  if (left.IsConstant() && right.IsConstant()) {
    // TODO(vegorov): should be eliminated earlier by constant propagation.
    const bool result = (kind() == Token::kEQ_STRICT) ?
        left.constant().raw() == right.constant().raw() :
        left.constant().raw() != right.constant().raw();
    branch->EmitBranchOnValue(compiler, result);
    return;
  }
  if (left.IsConstant()) {
    compiler->EmitEqualityRegConstCompare(right.reg(),
                                          left.constant(),
                                          needs_number_check(),
                                          token_pos());
  } else if (right.IsConstant()) {
    compiler->EmitEqualityRegConstCompare(left.reg(),
                                          right.constant(),
                                          needs_number_check(),
                                          token_pos());
  } else {
    compiler->EmitEqualityRegRegCompare(left.reg(),
                                        right.reg(),
                                        needs_number_check(),
                                        token_pos());
  }

  Condition true_condition = (kind() == Token::kEQ_STRICT) ? EQUAL : NOT_EQUAL;
  branch->EmitBranchOnCondition(compiler, true_condition);
}


LocationSummary* ClosureCallInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 0;
  const intptr_t kNumTemps = 1;
  LocationSummary* result =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kCall);
  result->set_out(Location::RegisterLocation(RAX));
  result->set_temp(0, Location::RegisterLocation(R10));  // Arg. descriptor.
  return result;
}


void ClosureCallInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  // The arguments to the stub include the closure, as does the arguments
  // descriptor.
  Register temp_reg = locs()->temp(0).reg();
  int argument_count = ArgumentCount();
  const Array& arguments_descriptor =
      Array::ZoneHandle(ArgumentsDescriptor::New(argument_count,
                                                 argument_names()));
  __ LoadObject(temp_reg, arguments_descriptor);
  ASSERT(temp_reg == R10);
  compiler->GenerateDartCall(deopt_id(),
                             token_pos(),
                             &StubCode::CallClosureFunctionLabel(),
                             PcDescriptors::kClosureCall,
                             locs());
  __ Drop(argument_count);
}


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


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

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


LocationSummary* StoreVMFieldInstr::MakeLocationSummary() const {
  const intptr_t kNumInputs = 2;
  const intptr_t kNumTemps = 0;
  LocationSummary* locs =
      new LocationSummary(kNumInputs, kNumTemps, LocationSummary::kNoCall);
  locs->set_in(0, value()->NeedsStoreBuffer() ? Location::WritableRegister()
                                              : Location::RequiresRegister());
  locs->set_in(1, Location::RequiresRegister());
  return locs;
}


void StoreVMFieldInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  Register value_reg = locs()->in(0).reg();
  Register dest_reg = locs()->in(1).reg();

  if (value()->NeedsStoreBuffer()) {
    __ StoreIntoObject(dest_reg, FieldAddress(dest_reg, offset_in_bytes()),
                       value_reg);
  } else {
    __ StoreIntoObjectNoBarrier(
        dest_reg, FieldAddress(dest_reg, offset_in_bytes()), value_reg);
  }
}


LocationSummary* AllocateObjectInstr::MakeLocationSummary() const {
  return MakeCallSummary();
}


void AllocateObjectInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  const Code& stub = Code::Handle(StubCode::GetAllocationStubForClass(cls()));
  const ExternalLabel label(cls().ToCString(), stub.EntryPoint());
  compiler->GenerateCall(token_pos(),
                         &label,
                         PcDescriptors::kOther,
                         locs());
  __ Drop(ArgumentCount());  // Discard arguments.
}


LocationSummary* CreateClosureInstr::MakeLocationSummary() const {
  return MakeCallSummary();
}


void CreateClosureInstr::EmitNativeCode(FlowGraphCompiler* compiler) {
  const Function& closure_function = function();
  ASSERT(!closure_function.IsImplicitStaticClosureFunction());
  const Code& stub = Code::Handle(
      StubCode::GetAllocationStubForClosure(closure_function));
  const ExternalLabel label(closure_function.ToCString(), stub.EntryPoint());
  compiler->GenerateCall(token_pos(),
                         &label,
                         PcDescriptors::kOther,
                         locs());
  __ Drop(2);  // Discard type arguments and receiver.
}

}  // namespace dart

#undef __

#endif  // defined TARGET_ARCH_X64