// 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_ARM.
#if defined(TARGET_ARCH_ARM)

#include "vm/flow_graph_compiler.h"

#include "lib/error.h"
#include "vm/ast_printer.h"
#include "vm/dart_entry.h"
#include "vm/deopt_instructions.h"
#include "vm/il_printer.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"

namespace dart {

DEFINE_FLAG(bool, trap_on_deoptimization, false, "Trap on deoptimization.");
DECLARE_FLAG(int, optimization_counter_threshold);
DECLARE_FLAG(int, reoptimization_counter_threshold);
DECLARE_FLAG(bool, print_ast);
DECLARE_FLAG(bool, print_scopes);
DECLARE_FLAG(bool, enable_type_checks);
DECLARE_FLAG(bool, eliminate_type_checks);


FlowGraphCompiler::~FlowGraphCompiler() {
  // BlockInfos are zone-allocated, so their destructors are not called.
  // Verify the labels explicitly here.
  for (int i = 0; i < block_info_.length(); ++i) {
    ASSERT(!block_info_[i]->jump_label()->IsLinked());
  }
}


bool FlowGraphCompiler::SupportsUnboxedMints() {
  return false;
}


RawDeoptInfo* CompilerDeoptInfo::CreateDeoptInfo(FlowGraphCompiler* compiler,
                                                 DeoptInfoBuilder* builder) {
  if (deopt_env_ == NULL) return DeoptInfo::null();

  intptr_t stack_height = compiler->StackSize();
  AllocateIncomingParametersRecursive(deopt_env_, &stack_height);

  intptr_t slot_ix = 0;
  Environment* current = deopt_env_;

  // Emit all kMaterializeObject instructions describing objects to be
  // materialized on the deoptimization as a prefix to the deoptimization info.
  EmitMaterializations(deopt_env_, builder);

  // The real frame starts here.
  builder->MarkFrameStart();

  // Current PP, FP, and PC.
  builder->AddPp(current->function(), slot_ix++);
  builder->AddCallerFp(slot_ix++);
  builder->AddReturnAddress(current->function(), deopt_id(), slot_ix++);

  // Callee's PC marker is not used anymore. Pass Function::null() to set to 0.
  builder->AddPcMarker(Function::Handle(), slot_ix++);

  // Emit all values that are needed for materialization as a part of the
  // expression stack for the bottom-most frame. This guarantees that GC
  // will be able to find them during materialization.
  slot_ix = builder->EmitMaterializationArguments(slot_ix);

  // For the innermost environment, set outgoing arguments and the locals.
  for (intptr_t i = current->Length() - 1;
       i >= current->fixed_parameter_count();
       i--) {
    builder->AddCopy(current->ValueAt(i), current->LocationAt(i), slot_ix++);
  }

  Environment* previous = current;
  current = current->outer();
  while (current != NULL) {
    // PP, FP, and PC.
    builder->AddPp(current->function(), slot_ix++);
    builder->AddCallerFp(slot_ix++);

    // For any outer environment the deopt id is that of the call instruction
    // which is recorded in the outer environment.
    builder->AddReturnAddress(current->function(),
                              Isolate::ToDeoptAfter(current->deopt_id()),
                              slot_ix++);

    // PC marker.
    builder->AddPcMarker(previous->function(), slot_ix++);

    // The values of outgoing arguments can be changed from the inlined call so
    // we must read them from the previous environment.
    for (intptr_t i = previous->fixed_parameter_count() - 1; i >= 0; i--) {
      builder->AddCopy(previous->ValueAt(i),
                       previous->LocationAt(i),
                       slot_ix++);
    }

    // Set the locals, note that outgoing arguments are not in the environment.
    for (intptr_t i = current->Length() - 1;
         i >= current->fixed_parameter_count();
         i--) {
      builder->AddCopy(current->ValueAt(i),
                       current->LocationAt(i),
                       slot_ix++);
    }

    // Iterate on the outer environment.
    previous = current;
    current = current->outer();
  }
  // The previous pointer is now the outermost environment.
  ASSERT(previous != NULL);

  // For the outermost environment, set caller PC, caller PP, and caller FP.
  builder->AddCallerPp(slot_ix++);
  builder->AddCallerFp(slot_ix++);
  builder->AddCallerPc(slot_ix++);

  // PC marker.
  builder->AddPcMarker(previous->function(), slot_ix++);

  // For the outermost environment, set the incoming arguments.
  for (intptr_t i = previous->fixed_parameter_count() - 1; i >= 0; i--) {
    builder->AddCopy(previous->ValueAt(i), previous->LocationAt(i), slot_ix++);
  }

  const DeoptInfo& deopt_info = DeoptInfo::Handle(builder->CreateDeoptInfo());
  return deopt_info.raw();
}


void CompilerDeoptInfoWithStub::GenerateCode(FlowGraphCompiler* compiler,
                                             intptr_t stub_ix) {
  // Calls do not need stubs, they share a deoptimization trampoline.
  ASSERT(reason() != kDeoptAtCall);
  Assembler* assem = compiler->assembler();
#define __ assem->
  __ Comment("Deopt stub for id %"Pd"", deopt_id());
  __ Bind(entry_label());
  if (FLAG_trap_on_deoptimization) __ bkpt(0);

  ASSERT(deopt_env() != NULL);

  __ BranchLink(&StubCode::DeoptimizeLabel());
  set_pc_offset(assem->CodeSize());
  __ bkpt(0);  // TODO(regis): Remove breakpoint to save space.
#undef __
}


#define __ assembler()->


// Fall through if bool_register contains null.
void FlowGraphCompiler::GenerateBoolToJump(Register bool_register,
                                           Label* is_true,
                                           Label* is_false) {
  Label fall_through;
  __ CompareImmediate(bool_register,
                      reinterpret_cast<intptr_t>(Object::null()));
  __ b(&fall_through, EQ);
  __ CompareObject(bool_register, Bool::True());
  __ b(is_true, EQ);
  __ b(is_false);
  __ Bind(&fall_through);
}


// R0: instance (must be preserved).
// R1: instantiator type arguments (if used).
RawSubtypeTestCache* FlowGraphCompiler::GenerateCallSubtypeTestStub(
    TypeTestStubKind test_kind,
    Register instance_reg,
    Register type_arguments_reg,
    Register temp_reg,
    Label* is_instance_lbl,
    Label* is_not_instance_lbl) {
  ASSERT(instance_reg == R0);
  ASSERT(temp_reg == kNoRegister);  // Unused on ARM.
  const SubtypeTestCache& type_test_cache =
      SubtypeTestCache::ZoneHandle(SubtypeTestCache::New());
  __ LoadObject(R2, type_test_cache);
  if (test_kind == kTestTypeOneArg) {
    ASSERT(type_arguments_reg == kNoRegister);
    __ LoadImmediate(R1, reinterpret_cast<intptr_t>(Object::null()));
    __ BranchLink(&StubCode::Subtype1TestCacheLabel());
  } else if (test_kind == kTestTypeTwoArgs) {
    ASSERT(type_arguments_reg == kNoRegister);
    __ LoadImmediate(R1, reinterpret_cast<intptr_t>(Object::null()));
    __ BranchLink(&StubCode::Subtype2TestCacheLabel());
  } else if (test_kind == kTestTypeThreeArgs) {
    ASSERT(type_arguments_reg == R1);
    __ BranchLink(&StubCode::Subtype3TestCacheLabel());
  } else {
    UNREACHABLE();
  }
  // Result is in R1: null -> not found, otherwise Bool::True or Bool::False.
  GenerateBoolToJump(R1, is_instance_lbl, is_not_instance_lbl);
  return type_test_cache.raw();
}


// Jumps to labels 'is_instance' or 'is_not_instance' respectively, if
// type test is conclusive, otherwise fallthrough if a type test could not
// be completed.
// R0: instance being type checked (preserved).
// Clobbers R2.
RawSubtypeTestCache*
FlowGraphCompiler::GenerateInstantiatedTypeWithArgumentsTest(
    intptr_t token_pos,
    const AbstractType& type,
    Label* is_instance_lbl,
    Label* is_not_instance_lbl) {
  __ Comment("InstantiatedTypeWithArgumentsTest");
  ASSERT(type.IsInstantiated());
  const Class& type_class = Class::ZoneHandle(type.type_class());
  ASSERT(type_class.HasTypeArguments() || type_class.IsSignatureClass());
  const Register kInstanceReg = R0;
  Error& malformed_error = Error::Handle();
  const Type& int_type = Type::Handle(Type::IntType());
  const bool smi_is_ok = int_type.IsSubtypeOf(type, &malformed_error);
  // Malformed type should have been handled at graph construction time.
  ASSERT(smi_is_ok || malformed_error.IsNull());
  __ tst(kInstanceReg, ShifterOperand(kSmiTagMask));
  if (smi_is_ok) {
    __ b(is_instance_lbl, EQ);
  } else {
    __ b(is_not_instance_lbl, EQ);
  }
  const AbstractTypeArguments& type_arguments =
      AbstractTypeArguments::ZoneHandle(type.arguments());
  const bool is_raw_type = type_arguments.IsNull() ||
      type_arguments.IsRaw(type_arguments.Length());
  // Signature class is an instantiated parameterized type.
  if (!type_class.IsSignatureClass()) {
    if (is_raw_type) {
      const Register kClassIdReg = R2;
      // dynamic type argument, check only classes.
      __ LoadClassId(kClassIdReg, kInstanceReg);
      __ CompareImmediate(kClassIdReg, type_class.id());
      __ b(is_instance_lbl, EQ);
      // List is a very common case.
      if (IsListClass(type_class)) {
        GenerateListTypeCheck(kClassIdReg, is_instance_lbl);
      }
      return GenerateSubtype1TestCacheLookup(
          token_pos, type_class, is_instance_lbl, is_not_instance_lbl);
    }
    // If one type argument only, check if type argument is Object or dynamic.
    if (type_arguments.Length() == 1) {
      const AbstractType& tp_argument = AbstractType::ZoneHandle(
          type_arguments.TypeAt(0));
      ASSERT(!tp_argument.IsMalformed());
      if (tp_argument.IsType()) {
        ASSERT(tp_argument.HasResolvedTypeClass());
        // Check if type argument is dynamic or Object.
        const Type& object_type = Type::Handle(Type::ObjectType());
        if (object_type.IsSubtypeOf(tp_argument, NULL)) {
          // Instance class test only necessary.
          return GenerateSubtype1TestCacheLookup(
              token_pos, type_class, is_instance_lbl, is_not_instance_lbl);
        }
      }
    }
  }
  // Regular subtype test cache involving instance's type arguments.
  const Register kTypeArgumentsReg = kNoRegister;
  const Register kTempReg = kNoRegister;
  // R0: instance (must be preserved).
  return GenerateCallSubtypeTestStub(kTestTypeTwoArgs,
                                     kInstanceReg,
                                     kTypeArgumentsReg,
                                     kTempReg,
                                     is_instance_lbl,
                                     is_not_instance_lbl);
}


void FlowGraphCompiler::CheckClassIds(Register class_id_reg,
                                      const GrowableArray<intptr_t>& class_ids,
                                      Label* is_equal_lbl,
                                      Label* is_not_equal_lbl) {
  for (intptr_t i = 0; i < class_ids.length(); i++) {
    __ CompareImmediate(class_id_reg, class_ids[i]);
    __ b(is_equal_lbl, EQ);
  }
  __ b(is_not_equal_lbl);
}


// Testing against an instantiated type with no arguments, without
// SubtypeTestCache.
// R0: instance being type checked (preserved).
// Clobbers R2, R3.
// Returns true if there is a fallthrough.
bool FlowGraphCompiler::GenerateInstantiatedTypeNoArgumentsTest(
    intptr_t token_pos,
    const AbstractType& type,
    Label* is_instance_lbl,
    Label* is_not_instance_lbl) {
  __ Comment("InstantiatedTypeNoArgumentsTest");
  ASSERT(type.IsInstantiated());
  const Class& type_class = Class::Handle(type.type_class());
  ASSERT(!type_class.HasTypeArguments());

  const Register kInstanceReg = R0;
  __ tst(kInstanceReg, ShifterOperand(kSmiTagMask));
  // If instance is Smi, check directly.
  const Class& smi_class = Class::Handle(Smi::Class());
  if (smi_class.IsSubtypeOf(TypeArguments::Handle(),
                            type_class,
                            TypeArguments::Handle(),
                            NULL)) {
    __ b(is_instance_lbl, EQ);
  } else {
    __ b(is_not_instance_lbl, EQ);
  }
  // Compare if the classes are equal.
  const Register kClassIdReg = R2;
  __ LoadClassId(kClassIdReg, kInstanceReg);
  __ CompareImmediate(kClassIdReg, type_class.id());
  __ b(is_instance_lbl, EQ);
  // See ClassFinalizer::ResolveSuperTypeAndInterfaces for list of restricted
  // interfaces.
  // Bool interface can be implemented only by core class Bool.
  if (type.IsBoolType()) {
    __ CompareImmediate(kClassIdReg, kBoolCid);
    __ b(is_instance_lbl, EQ);
    __ b(is_not_instance_lbl);
    return false;
  }
  if (type.IsFunctionType()) {
    // Check if instance is a closure.
    __ LoadClassById(R3, kClassIdReg);
    __ ldr(R3, FieldAddress(R3, Class::signature_function_offset()));
    __ CompareImmediate(R3, reinterpret_cast<int32_t>(Object::null()));
    __ b(is_instance_lbl, NE);
  }
  // Custom checking for numbers (Smi, Mint, Bigint and Double).
  // Note that instance is not Smi (checked above).
  if (type.IsSubtypeOf(Type::Handle(Type::Number()), NULL)) {
    GenerateNumberTypeCheck(
        kClassIdReg, type, is_instance_lbl, is_not_instance_lbl);
    return false;
  }
  if (type.IsStringType()) {
    GenerateStringTypeCheck(kClassIdReg, is_instance_lbl, is_not_instance_lbl);
    return false;
  }
  // Otherwise fallthrough.
  return true;
}


// Uses SubtypeTestCache to store instance class and result.
// R0: instance to test.
// Clobbers R1-R5.
// Immediate class test already done.
// TODO(srdjan): Implement a quicker subtype check, as type test
// arrays can grow too high, but they may be useful when optimizing
// code (type-feedback).
RawSubtypeTestCache* FlowGraphCompiler::GenerateSubtype1TestCacheLookup(
    intptr_t token_pos,
    const Class& type_class,
    Label* is_instance_lbl,
    Label* is_not_instance_lbl) {
  __ Comment("Subtype1TestCacheLookup");
  const Register kInstanceReg = R0;
  __ LoadClass(R1, kInstanceReg, R2);
  // R1: instance class.
  // Check immediate superclass equality.
  __ ldr(R2, FieldAddress(R1, Class::super_type_offset()));
  __ ldr(R2, FieldAddress(R2, Type::type_class_offset()));
  __ CompareObject(R2, type_class);
  __ b(is_instance_lbl, EQ);

  const Register kTypeArgumentsReg = kNoRegister;
  const Register kTempReg = kNoRegister;
  return GenerateCallSubtypeTestStub(kTestTypeOneArg,
                                     kInstanceReg,
                                     kTypeArgumentsReg,
                                     kTempReg,
                                     is_instance_lbl,
                                     is_not_instance_lbl);
}


// Generates inlined check if 'type' is a type parameter or type itself
// R0: instance (preserved).
RawSubtypeTestCache* FlowGraphCompiler::GenerateUninstantiatedTypeTest(
    intptr_t token_pos,
    const AbstractType& type,
    Label* is_instance_lbl,
    Label* is_not_instance_lbl) {
  __ Comment("UninstantiatedTypeTest");
  ASSERT(!type.IsInstantiated());
  // Skip check if destination is a dynamic type.
  if (type.IsTypeParameter()) {
    const TypeParameter& type_param = TypeParameter::Cast(type);
    // Load instantiator (or null) and instantiator type arguments on stack.
    __ ldr(R1, Address(SP, 0));  // Get instantiator type arguments.
    // R1: instantiator type arguments.
    // Check if type argument is dynamic.
    __ CompareImmediate(R1, reinterpret_cast<intptr_t>(Object::null()));
    __ b(is_instance_lbl, EQ);
    // Can handle only type arguments that are instances of TypeArguments.
    // (runtime checks canonicalize type arguments).
    Label fall_through;
    __ CompareClassId(R1, kTypeArgumentsCid, R2);
    __ b(&fall_through, NE);
    __ ldr(R2,
        FieldAddress(R1, TypeArguments::type_at_offset(type_param.index())));
    // R2: concrete type of type.
    // Check if type argument is dynamic.
    __ CompareObject(R2, Type::ZoneHandle(Type::DynamicType()));
    __ b(is_instance_lbl, EQ);
    __ CompareImmediate(R2, reinterpret_cast<intptr_t>(Object::null()));
    __ b(is_instance_lbl, EQ);
    const Type& object_type = Type::ZoneHandle(Type::ObjectType());
    __ CompareObject(R2, object_type);
    __ b(is_instance_lbl, EQ);

    // For Smi check quickly against int and num interfaces.
    Label not_smi;
    __ tst(R0, ShifterOperand(kSmiTagMask));  // Value is Smi?
    __ b(&not_smi, NE);
    __ CompareObject(R2, Type::ZoneHandle(Type::IntType()));
    __ b(is_instance_lbl, EQ);
    __ CompareObject(R2, Type::ZoneHandle(Type::Number()));
    __ b(is_instance_lbl, EQ);
    // Smi must be handled in runtime.
    __ b(&fall_through);

    __ Bind(&not_smi);
    // R1: instantiator type arguments.
    // R0: instance.
    const Register kInstanceReg = R0;
    const Register kTypeArgumentsReg = R1;
    const Register kTempReg = kNoRegister;
    const SubtypeTestCache& type_test_cache =
        SubtypeTestCache::ZoneHandle(
            GenerateCallSubtypeTestStub(kTestTypeThreeArgs,
                                        kInstanceReg,
                                        kTypeArgumentsReg,
                                        kTempReg,
                                        is_instance_lbl,
                                        is_not_instance_lbl));
    __ Bind(&fall_through);
    return type_test_cache.raw();
  }
  if (type.IsType()) {
    const Register kInstanceReg = R0;
    const Register kTypeArgumentsReg = R1;
    __ tst(kInstanceReg, ShifterOperand(kSmiTagMask));  // Is instance Smi?
    __ b(is_not_instance_lbl, EQ);
    __ ldr(kTypeArgumentsReg, Address(SP, 0));  // Instantiator type args.
    // Uninstantiated type class is known at compile time, but the type
    // arguments are determined at runtime by the instantiator.
    const Register kTempReg = kNoRegister;
    return GenerateCallSubtypeTestStub(kTestTypeThreeArgs,
                                       kInstanceReg,
                                       kTypeArgumentsReg,
                                       kTempReg,
                                       is_instance_lbl,
                                       is_not_instance_lbl);
  }
  return SubtypeTestCache::null();
}


// Inputs:
// - R0: instance being type checked (preserved).
// - R1: optional instantiator type arguments (preserved).
// Clobbers R2, R3.
// Returns:
// - preserved instance in R0 and optional instantiator type arguments in R1.
// Note that this inlined code must be followed by the runtime_call code, as it
// may fall through to it. Otherwise, this inline code will jump to the label
// is_instance or to the label is_not_instance.
RawSubtypeTestCache* FlowGraphCompiler::GenerateInlineInstanceof(
    intptr_t token_pos,
    const AbstractType& type,
    Label* is_instance_lbl,
    Label* is_not_instance_lbl) {
  __ Comment("InlineInstanceof");
  if (type.IsVoidType()) {
    // A non-null value is returned from a void function, which will result in a
    // type error. A null value is handled prior to executing this inline code.
    return SubtypeTestCache::null();
  }
  if (TypeCheckAsClassEquality(type)) {
    const intptr_t type_cid = Class::Handle(type.type_class()).id();
    const Register kInstanceReg = R0;
    __ tst(kInstanceReg, ShifterOperand(kSmiTagMask));
    if (type_cid == kSmiCid) {
      __ b(is_instance_lbl, EQ);
    } else {
      __ b(is_not_instance_lbl, EQ);
      __ CompareClassId(kInstanceReg, type_cid, R3);
      __ b(is_instance_lbl, EQ);
    }
    __ b(is_not_instance_lbl);
    return SubtypeTestCache::null();
  }
  if (type.IsInstantiated()) {
    const Class& type_class = Class::ZoneHandle(type.type_class());
    // A class equality check is only applicable with a dst type of a
    // non-parameterized class, non-signature class, or with a raw dst type of
    // a parameterized class.
    if (type_class.IsSignatureClass() || type_class.HasTypeArguments()) {
      return GenerateInstantiatedTypeWithArgumentsTest(token_pos,
                                                       type,
                                                       is_instance_lbl,
                                                       is_not_instance_lbl);
      // Fall through to runtime call.
    }
    const bool has_fall_through =
        GenerateInstantiatedTypeNoArgumentsTest(token_pos,
                                                type,
                                                is_instance_lbl,
                                                is_not_instance_lbl);
    if (has_fall_through) {
      // If test non-conclusive so far, try the inlined type-test cache.
      // 'type' is known at compile time.
      return GenerateSubtype1TestCacheLookup(
          token_pos, type_class, is_instance_lbl, is_not_instance_lbl);
    } else {
      return SubtypeTestCache::null();
    }
  }
  return GenerateUninstantiatedTypeTest(token_pos,
                                        type,
                                        is_instance_lbl,
                                        is_not_instance_lbl);
}


// If instanceof type test cannot be performed successfully at compile time and
// therefore eliminated, optimize it by adding inlined tests for:
// - NULL -> return false.
// - Smi -> compile time subtype check (only if dst class is not parameterized).
// - Class equality (only if class is not parameterized).
// Inputs:
// - R0: object.
// - R1: instantiator type arguments or raw_null.
// - R2: instantiator or raw_null.
// Returns:
// - true or false in R0.
void FlowGraphCompiler::GenerateInstanceOf(intptr_t token_pos,
                                           intptr_t deopt_id,
                                           const AbstractType& type,
                                           bool negate_result,
                                           LocationSummary* locs) {
  ASSERT(type.IsFinalized() && !type.IsMalformed());

  // Preserve instantiator (R2) and its type arguments (R1).
  __ PushList((1 << R1) | (1 << R2));

  Label is_instance, is_not_instance;
  // If type is instantiated and non-parameterized, we can inline code
  // checking whether the tested instance is a Smi.
  if (type.IsInstantiated()) {
    // A null object is only an instance of Object and dynamic, which has
    // already been checked above (if the type is instantiated). So we can
    // return false here if the instance is null (and if the type is
    // instantiated).
    // We can only inline this null check if the type is instantiated at compile
    // time, since an uninstantiated type at compile time could be Object or
    // dynamic at run time.
    __ CompareImmediate(R0, reinterpret_cast<int32_t>(Object::null()));
    __ b(&is_not_instance, EQ);
  }

  // Generate inline instanceof test.
  SubtypeTestCache& test_cache = SubtypeTestCache::ZoneHandle();
  test_cache = GenerateInlineInstanceof(token_pos, type,
                                        &is_instance, &is_not_instance);

  // test_cache is null if there is no fall-through.
  Label done;
  if (!test_cache.IsNull()) {
    // Generate runtime call.
    // Load instantiator (R2) and its type arguments (R1).
    __ ldm(IA, SP,  (1 << R1) | (1 << R2));
    __ PushObject(Object::ZoneHandle());  // Make room for the result.
    __ Push(R0);  // Push the instance.
    __ PushObject(type);  // Push the type.
    // Push instantiator (R2) and its type arguments (R1).
    __ PushList((1 << R1) | (1 << R2));
    __ LoadObject(R0, test_cache);
    __ Push(R0);
    GenerateCallRuntime(token_pos, deopt_id, kInstanceofRuntimeEntry, locs);
    // Pop the parameters supplied to the runtime entry. The result of the
    // instanceof runtime call will be left as the result of the operation.
    __ Drop(5);
    if (negate_result) {
      __ Pop(R1);
      __ LoadObject(R0, Bool::True());
      __ cmp(R1, ShifterOperand(R0));
      __ b(&done, NE);
      __ LoadObject(R0, Bool::False());
    } else {
      __ Pop(R0);
    }
    __ b(&done);
  }
  __ Bind(&is_not_instance);
  __ LoadObject(R0, negate_result ? Bool::True() : Bool::False());
  __ b(&done);

  __ Bind(&is_instance);
  __ LoadObject(R0, negate_result ? Bool::False() : Bool::True());
  __ Bind(&done);
  // Remove instantiator (R2) and its type arguments (R1).
  __ Drop(2);
}


// Optimize assignable type check by adding inlined tests for:
// - NULL -> return NULL.
// - Smi -> compile time subtype check (only if dst class is not parameterized).
// - Class equality (only if class is not parameterized).
// Inputs:
// - R0: instance being type checked.
// - R1: instantiator type arguments or raw_null.
// - R2: instantiator or raw_null.
// Returns:
// - object in R0 for successful assignable check (or throws TypeError).
// Performance notes: positive checks must be quick, negative checks can be slow
// as they throw an exception.
void FlowGraphCompiler::GenerateAssertAssignable(intptr_t token_pos,
                                                 intptr_t deopt_id,
                                                 const AbstractType& dst_type,
                                                 const String& dst_name,
                                                 LocationSummary* locs) {
  ASSERT(token_pos >= 0);
  ASSERT(!dst_type.IsNull());
  ASSERT(dst_type.IsFinalized());
  // Assignable check is skipped in FlowGraphBuilder, not here.
  ASSERT(dst_type.IsMalformed() ||
         (!dst_type.IsDynamicType() && !dst_type.IsObjectType()));
  // Preserve instantiator (R2) and its type arguments (R1).
  __ PushList((1 << R1) | (1 << R2));
  // A null object is always assignable and is returned as result.
  Label is_assignable, runtime_call;
  __ CompareImmediate(R0, reinterpret_cast<int32_t>(Object::null()));
  __ b(&is_assignable, EQ);

  if (!FLAG_eliminate_type_checks || dst_type.IsMalformed()) {
    // If type checks are not eliminated during the graph building then
    // a transition sentinel can be seen here.
    __ CompareObject(R0, Object::transition_sentinel());
    __ b(&is_assignable, EQ);
  }

  // Generate throw new TypeError() if the type is malformed.
  if (dst_type.IsMalformed()) {
    const Error& error = Error::Handle(dst_type.malformed_error());
    const String& error_message = String::ZoneHandle(
        Symbols::New(error.ToErrorCString()));
    __ PushObject(Object::ZoneHandle());  // Make room for the result.
    __ Push(R0);  // Push the source object.
    __ PushObject(dst_name);  // Push the name of the destination.
    __ PushObject(error_message);
    GenerateCallRuntime(token_pos,
                        deopt_id,
                        kMalformedTypeErrorRuntimeEntry,
                        locs);
    // We should never return here.
    __ bkpt(0);

    __ Bind(&is_assignable);  // For a null object.
    // Restore instantiator (R2) and its type arguments (R1).
    __ PopList((1 << R1) | (1 << R2));
    return;
  }

  // Generate inline type check, linking to runtime call if not assignable.
  SubtypeTestCache& test_cache = SubtypeTestCache::ZoneHandle();
  test_cache = GenerateInlineInstanceof(token_pos, dst_type,
                                        &is_assignable, &runtime_call);

  __ Bind(&runtime_call);
  // Load instantiator (R2) and its type arguments (R1).
  __ ldm(IA, SP,  (1 << R1) | (1 << R2));
  __ PushObject(Object::ZoneHandle());  // Make room for the result.
  __ Push(R0);  // Push the source object.
  __ PushObject(dst_type);  // Push the type of the destination.
  // Push instantiator (R2) and its type arguments (R1).
  __ PushList((1 << R1) | (1 << R2));
  __ PushObject(dst_name);  // Push the name of the destination.
  __ LoadObject(R0, test_cache);
  __ Push(R0);
  GenerateCallRuntime(token_pos, deopt_id, kTypeCheckRuntimeEntry, locs);
  // Pop the parameters supplied to the runtime entry. The result of the
  // type check runtime call is the checked value.
  __ Drop(6);
  __ Pop(R0);

  __ Bind(&is_assignable);
  // Restore instantiator (R2) and its type arguments (R1).
  __ PopList((1 << R1) | (1 << R2));
}


void FlowGraphCompiler::EmitInstructionPrologue(Instruction* instr) {
  if (!is_optimizing()) {
    if (FLAG_enable_type_checks && instr->IsAssertAssignable()) {
      AssertAssignableInstr* assert = instr->AsAssertAssignable();
      AddCurrentDescriptor(PcDescriptors::kDeopt,
                           assert->deopt_id(),
                           assert->token_pos());
    } else if (instr->IsGuardField() ||
               instr->CanBecomeDeoptimizationTarget()) {
      AddCurrentDescriptor(PcDescriptors::kDeopt,
                           instr->deopt_id(),
                           Scanner::kDummyTokenIndex);
    }
    AllocateRegistersLocally(instr);
  } else if (instr->MayThrow()  &&
             (CurrentTryIndex() != CatchClauseNode::kInvalidTryIndex)) {
    // Optimized try-block: Sync locals to fixed stack locations.
    EmitTrySync(instr, CurrentTryIndex());
  }
}


void FlowGraphCompiler::EmitTrySyncMove(intptr_t dest_offset,
                                        Location loc,
                                        bool* push_emitted) {
  if (loc.IsConstant()) {
    if (!*push_emitted) {
      __ Push(R0);
      *push_emitted = true;
    }
    __ LoadObject(R0, loc.constant());
    __ StoreToOffset(kWord, R0, FP, dest_offset);
  } else if (loc.IsRegister()) {
    if (*push_emitted && (loc.reg() == R0)) {
      __ ldr(R0, Address(SP, 0));
      __ StoreToOffset(kWord, R0, FP, dest_offset);
    } else {
      __ StoreToOffset(kWord, loc.reg(), FP, dest_offset);
    }
  } else {
    const intptr_t src_offset = loc.ToStackSlotOffset();
    if (src_offset != dest_offset) {
      if (!*push_emitted) {
        __ Push(R0);
        *push_emitted = true;
      }
      __ LoadFromOffset(kWord, R0, FP, src_offset);
      __ StoreToOffset(kWord, R0, FP, dest_offset);
    }
  }
}


void FlowGraphCompiler::EmitTrySync(Instruction* instr, intptr_t try_index) {
  ASSERT(is_optimizing());
  Environment* env = instr->env();
  CatchBlockEntryInstr* catch_block =
      flow_graph().graph_entry()->GetCatchEntry(try_index);
  const GrowableArray<Definition*>* idefs = catch_block->initial_definitions();
  // Parameters.
  intptr_t i = 0;
  bool push_emitted = false;
  const intptr_t num_non_copied_params = flow_graph().num_non_copied_params();
  const intptr_t param_base =
      kParamEndSlotFromFp + num_non_copied_params;
  for (; i < num_non_copied_params; ++i) {
    if ((*idefs)[i]->IsConstant()) continue;  // Common constants
    Location loc = env->LocationAt(i);
    EmitTrySyncMove((param_base - i) * kWordSize, loc, &push_emitted);
  }

  // Process locals. Skip exception_var and stacktrace_var.
  CatchEntryInstr* catch_entry = catch_block->next()->AsCatchEntry();
  intptr_t local_base = kFirstLocalSlotFromFp + num_non_copied_params;
  intptr_t ex_idx = local_base - catch_entry->exception_var().index();
  intptr_t st_idx = local_base - catch_entry->stacktrace_var().index();
  for (; i < flow_graph().variable_count(); ++i) {
    if (i == ex_idx || i == st_idx) continue;
    if ((*idefs)[i]->IsConstant()) continue;
    Location loc = env->LocationAt(i);
    EmitTrySyncMove((local_base - i) * kWordSize, loc, &push_emitted);
    // Update safepoint bitmap to indicate that the target location
    // now contains a pointer.
    instr->locs()->stack_bitmap()->Set(i - num_non_copied_params, true);
  }
  if (push_emitted) {
    __ Pop(R0);
  }
}


void FlowGraphCompiler::EmitInstructionEpilogue(Instruction* instr) {
  if (is_optimizing()) return;
  Definition* defn = instr->AsDefinition();
  if ((defn != NULL) && defn->is_used()) {
    __ Push(defn->locs()->out().reg());
  }
}


// Input parameters:
//   R4: arguments descriptor array.
void FlowGraphCompiler::CopyParameters() {
  __ Comment("Copy parameters");
  const Function& function = parsed_function().function();
  LocalScope* scope = parsed_function().node_sequence()->scope();
  const int num_fixed_params = function.num_fixed_parameters();
  const int num_opt_pos_params = function.NumOptionalPositionalParameters();
  const int num_opt_named_params = function.NumOptionalNamedParameters();
  const int num_params =
      num_fixed_params + num_opt_pos_params + num_opt_named_params;
  ASSERT(function.NumParameters() == num_params);
  ASSERT(parsed_function().first_parameter_index() == kFirstLocalSlotFromFp);

  // Check that min_num_pos_args <= num_pos_args <= max_num_pos_args,
  // where num_pos_args is the number of positional arguments passed in.
  const int min_num_pos_args = num_fixed_params;
  const int max_num_pos_args = num_fixed_params + num_opt_pos_params;

  __ ldr(R8, FieldAddress(R4, ArgumentsDescriptor::positional_count_offset()));
  // Check that min_num_pos_args <= num_pos_args.
  Label wrong_num_arguments;
  __ CompareImmediate(R8, Smi::RawValue(min_num_pos_args));
  __ b(&wrong_num_arguments, LT);
  // Check that num_pos_args <= max_num_pos_args.
  __ CompareImmediate(R8, Smi::RawValue(max_num_pos_args));
  __ b(&wrong_num_arguments, GT);

  // Copy positional arguments.
  // Argument i passed at fp[kParamEndSlotFromFp + num_args - i] is copied
  // to fp[kFirstLocalSlotFromFp - i].

  __ ldr(R7, FieldAddress(R4, ArgumentsDescriptor::count_offset()));
  // Since R7 and R8 are Smi, use LSL 1 instead of LSL 2.
  // Let R7 point to the last passed positional argument, i.e. to
  // fp[kParamEndSlotFromFp + num_args - (num_pos_args - 1)].
  __ sub(R7, R7, ShifterOperand(R8));
  __ add(R7, FP, ShifterOperand(R7, LSL, 1));
  __ add(R7, R7, ShifterOperand((kParamEndSlotFromFp + 1) * kWordSize));

  // Let R6 point to the last copied positional argument, i.e. to
  // fp[kFirstLocalSlotFromFp - (num_pos_args - 1)].
  __ AddImmediate(R6, FP, (kFirstLocalSlotFromFp + 1) * kWordSize);
  __ sub(R6, R6, ShifterOperand(R8, LSL, 1));  // R8 is a Smi.
  __ SmiUntag(R8);
  Label loop, loop_condition;
  __ b(&loop_condition);
  // We do not use the final allocation index of the variable here, i.e.
  // scope->VariableAt(i)->index(), because captured variables still need
  // to be copied to the context that is not yet allocated.
  const Address argument_addr(R7, R8, LSL, 2);
  const Address copy_addr(R6, R8, LSL, 2);
  __ Bind(&loop);
  __ ldr(IP, argument_addr);
  __ str(IP, copy_addr);
  __ Bind(&loop_condition);
  __ subs(R8, R8, ShifterOperand(1));
  __ b(&loop, PL);

  // Copy or initialize optional named arguments.
  Label all_arguments_processed;
#ifdef DEBUG
    const bool check_correct_named_args = true;
#else
    const bool check_correct_named_args = function.IsClosureFunction();
#endif
  if (num_opt_named_params > 0) {
    // Start by alphabetically sorting the names of the optional parameters.
    LocalVariable** opt_param = new LocalVariable*[num_opt_named_params];
    int* opt_param_position = new int[num_opt_named_params];
    for (int pos = num_fixed_params; pos < num_params; pos++) {
      LocalVariable* parameter = scope->VariableAt(pos);
      const String& opt_param_name = parameter->name();
      int i = pos - num_fixed_params;
      while (--i >= 0) {
        LocalVariable* param_i = opt_param[i];
        const intptr_t result = opt_param_name.CompareTo(param_i->name());
        ASSERT(result != 0);
        if (result > 0) break;
        opt_param[i + 1] = opt_param[i];
        opt_param_position[i + 1] = opt_param_position[i];
      }
      opt_param[i + 1] = parameter;
      opt_param_position[i + 1] = pos;
    }
    // Generate code handling each optional parameter in alphabetical order.
    __ ldr(R7, FieldAddress(R4, ArgumentsDescriptor::count_offset()));
    __ ldr(R8,
           FieldAddress(R4, ArgumentsDescriptor::positional_count_offset()));
    __ SmiUntag(R8);
    // Let R7 point to the first passed argument, i.e. to
    // fp[kParamEndSlotFromFp + num_args - 0]; num_args (R7) is Smi.
    __ add(R7, FP, ShifterOperand(R7, LSL, 1));
    __ AddImmediate(R7, R7, kParamEndSlotFromFp * kWordSize);
    // Let R6 point to the entry of the first named argument.
    __ add(R6, R4, ShifterOperand(
        ArgumentsDescriptor::first_named_entry_offset() - kHeapObjectTag));
    for (int i = 0; i < num_opt_named_params; i++) {
      Label load_default_value, assign_optional_parameter;
      const int param_pos = opt_param_position[i];
      // Check if this named parameter was passed in.
      // Load R5 with the name of the argument.
      __ ldr(R5, Address(R6, ArgumentsDescriptor::name_offset()));
      ASSERT(opt_param[i]->name().IsSymbol());
      __ CompareObject(R5, opt_param[i]->name());
      __ b(&load_default_value, NE);
      // Load R5 with passed-in argument at provided arg_pos, i.e. at
      // fp[kParamEndSlotFromFp + num_args - arg_pos].
      __ ldr(R5, Address(R6, ArgumentsDescriptor::position_offset()));
      // R5 is arg_pos as Smi.
      // Point to next named entry.
      __ add(R6, R6, ShifterOperand(ArgumentsDescriptor::named_entry_size()));
      __ rsb(R5, R5, ShifterOperand(0));
      Address argument_addr(R7, R5, LSL, 1);  // R5 is a negative Smi.
      __ ldr(R5, argument_addr);
      __ b(&assign_optional_parameter);
      __ Bind(&load_default_value);
      // Load R5 with default argument.
      const Object& value = Object::ZoneHandle(
          parsed_function().default_parameter_values().At(
              param_pos - num_fixed_params));
      __ LoadObject(R5, value);
      __ Bind(&assign_optional_parameter);
      // Assign R5 to fp[kFirstLocalSlotFromFp - param_pos].
      // We do not use the final allocation index of the variable here, i.e.
      // scope->VariableAt(i)->index(), because captured variables still need
      // to be copied to the context that is not yet allocated.
      const intptr_t computed_param_pos = kFirstLocalSlotFromFp - param_pos;
      const Address param_addr(FP, computed_param_pos * kWordSize);
      __ str(R5, param_addr);
    }
    delete[] opt_param;
    delete[] opt_param_position;
    if (check_correct_named_args) {
      // Check that R6 now points to the null terminator in the arguments
      // descriptor.
      __ ldr(R5, Address(R6, 0));
      __ CompareImmediate(R5, reinterpret_cast<int32_t>(Object::null()));
      __ b(&all_arguments_processed, EQ);
    }
  } else {
    ASSERT(num_opt_pos_params > 0);
    __ ldr(R8,
           FieldAddress(R4, ArgumentsDescriptor::positional_count_offset()));
    __ SmiUntag(R8);
    for (int i = 0; i < num_opt_pos_params; i++) {
      Label next_parameter;
      // Handle this optional positional parameter only if k or fewer positional
      // arguments have been passed, where k is param_pos, the position of this
      // optional parameter in the formal parameter list.
      const int param_pos = num_fixed_params + i;
      __ CompareImmediate(R8, param_pos);
      __ b(&next_parameter, GT);
      // Load R5 with default argument.
      const Object& value = Object::ZoneHandle(
          parsed_function().default_parameter_values().At(i));
      __ LoadObject(R5, value);
      // Assign R5 to fp[kFirstLocalSlotFromFp - param_pos].
      // We do not use the final allocation index of the variable here, i.e.
      // scope->VariableAt(i)->index(), because captured variables still need
      // to be copied to the context that is not yet allocated.
      const intptr_t computed_param_pos = kFirstLocalSlotFromFp - param_pos;
      const Address param_addr(FP, computed_param_pos * kWordSize);
      __ str(R5, param_addr);
      __ Bind(&next_parameter);
    }
    if (check_correct_named_args) {
      __ ldr(R7, FieldAddress(R4, ArgumentsDescriptor::count_offset()));
      __ SmiUntag(R7);
      // Check that R8 equals R7, i.e. no named arguments passed.
      __ cmp(R8, ShifterOperand(R7));
      __ b(&all_arguments_processed, EQ);
    }
  }

  __ Bind(&wrong_num_arguments);
  if (function.IsClosureFunction()) {
    // Invoke noSuchMethod function passing "call" as the original name.
    const int kNumArgsChecked = 1;
    const ICData& ic_data = ICData::ZoneHandle(
        ICData::New(function, Symbols::Call(), Object::empty_array(),
                    Isolate::kNoDeoptId, kNumArgsChecked));
    __ LoadObject(R5, ic_data);
    __ LeaveDartFrame();  // The arguments are still on the stack.
    __ Branch(&StubCode::CallNoSuchMethodFunctionLabel());
    // The noSuchMethod call may return to the caller, but not here.
    __ bkpt(0);
  } else if (check_correct_named_args) {
    __ Stop("Wrong arguments");
  }

  __ Bind(&all_arguments_processed);
  // Nullify originally passed arguments only after they have been copied and
  // checked, otherwise noSuchMethod would not see their original values.
  // This step can be skipped in case we decide that formal parameters are
  // implicitly final, since garbage collecting the unmodified value is not
  // an issue anymore.

  // R4 : arguments descriptor array.
  __ ldr(R8, FieldAddress(R4, ArgumentsDescriptor::count_offset()));
  __ SmiUntag(R8);
  __ add(R7, FP, ShifterOperand((kParamEndSlotFromFp + 1) * kWordSize));
  const Address original_argument_addr(R7, R8, LSL, 2);
  __ LoadImmediate(IP, reinterpret_cast<intptr_t>(Object::null()));
  Label null_args_loop, null_args_loop_condition;
  __ b(&null_args_loop_condition);
  __ Bind(&null_args_loop);
  __ str(IP, original_argument_addr);
  __ Bind(&null_args_loop_condition);
  __ subs(R8, R8, ShifterOperand(1));
  __ b(&null_args_loop, PL);
}


void FlowGraphCompiler::GenerateInlinedGetter(intptr_t offset) {
  // LR: return address.
  // SP: receiver.
  // Sequence node has one return node, its input is load field node.
  __ ldr(R0, Address(SP, 0 * kWordSize));
  __ LoadFromOffset(kWord, R0, R0, offset - kHeapObjectTag);
  __ Ret();
}


void FlowGraphCompiler::GenerateInlinedSetter(intptr_t offset) {
  // LR: return address.
  // SP+1: receiver.
  // SP+0: value.
  // Sequence node has one store node and one return NULL node.
  __ ldr(R0, Address(SP, 1 * kWordSize));  // Receiver.
  __ ldr(R1, Address(SP, 0 * kWordSize));  // Value.
  __ StoreIntoObject(R0, FieldAddress(R0, offset), R1);
  __ LoadImmediate(R0, reinterpret_cast<intptr_t>(Object::null()));
  __ Ret();
}


void FlowGraphCompiler::EmitFrameEntry() {
  const Function& function = parsed_function().function();
  if (CanOptimizeFunction() &&
      function.is_optimizable() &&
      (!is_optimizing() || may_reoptimize())) {
    const Register function_reg = R6;

    // The pool pointer is not setup before entering the Dart frame.
    // Preserve PP of caller.
    __ mov(R7, ShifterOperand(PP));
    // Temporarily setup pool pointer for this dart function.
    __ LoadPoolPointer();
    // Load function object from object pool.
    __ LoadObject(function_reg, function);  // Uses PP.
    // Restore PP of caller.
    __ mov(PP, ShifterOperand(R7));

    // Patch point is after the eventually inlined function object.
    AddCurrentDescriptor(PcDescriptors::kEntryPatch,
                         Isolate::kNoDeoptId,
                         0);  // No token position.
    intptr_t threshold = FLAG_optimization_counter_threshold;
    __ ldr(R7, FieldAddress(function_reg,
                            Function::usage_counter_offset()));
    if (is_optimizing()) {
      // Reoptimization of an optimized function is triggered by counting in
      // IC stubs, but not at the entry of the function.
      threshold = FLAG_reoptimization_counter_threshold;
    } else {
      __ add(R7, R7, ShifterOperand(1));
      __ str(R7, FieldAddress(function_reg,
                              Function::usage_counter_offset()));
    }
    __ CompareImmediate(R7, threshold);
    ASSERT(function_reg == R6);
    __ Branch(&StubCode::OptimizeFunctionLabel(), GE);
  } else if (!flow_graph().IsCompiledForOsr()) {
    AddCurrentDescriptor(PcDescriptors::kEntryPatch,
                         Isolate::kNoDeoptId,
                         0);  // No token position.
  }
  __ Comment("Enter frame");
  if (flow_graph().IsCompiledForOsr()) {
    intptr_t extra_slots = StackSize()
        - flow_graph().num_stack_locals()
        - flow_graph().num_copied_params();
    ASSERT(extra_slots >= 0);
    __ EnterOsrFrame(extra_slots * kWordSize);
  } else {
    ASSERT(StackSize() >= 0);
    __ EnterDartFrame(StackSize() * kWordSize);
  }
}


// Input parameters:
//   LR: return address.
//   SP: address of last argument.
//   FP: caller's frame pointer.
//   PP: caller's pool pointer.
//   R5: ic-data.
//   R4: arguments descriptor array.
void FlowGraphCompiler::CompileGraph() {
  InitCompiler();
  if (TryIntrinsify()) {
    // Although this intrinsified code will never be patched, it must satisfy
    // CodePatcher::CodeIsPatchable, which verifies that this code has a minimum
    // code size.
    __ bkpt(0);
    __ Branch(&StubCode::FixCallersTargetLabel());
    return;
  }

  EmitFrameEntry();

  const Function& function = parsed_function().function();

  const int num_fixed_params = function.num_fixed_parameters();
  const int num_copied_params = parsed_function().num_copied_params();
  const int num_locals = parsed_function().num_stack_locals();

  // We check the number of passed arguments when we have to copy them due to
  // the presence of optional parameters.
  // No such checking code is generated if only fixed parameters are declared,
  // unless we are in debug mode or unless we are compiling a closure.
  if (num_copied_params == 0) {
#ifdef DEBUG
    ASSERT(!parsed_function().function().HasOptionalParameters());
    const bool check_arguments = !flow_graph().IsCompiledForOsr();
#else
    const bool check_arguments =
        function.IsClosureFunction() && !flow_graph().IsCompiledForOsr();
#endif
    if (check_arguments) {
      __ Comment("Check argument count");
      // Check that exactly num_fixed arguments are passed in.
      Label correct_num_arguments, wrong_num_arguments;
      __ ldr(R0, FieldAddress(R4, ArgumentsDescriptor::count_offset()));
      __ CompareImmediate(R0, Smi::RawValue(num_fixed_params));
      __ b(&wrong_num_arguments, NE);
      __ ldr(R1, FieldAddress(R4,
                              ArgumentsDescriptor::positional_count_offset()));
      __ cmp(R0, ShifterOperand(R1));
      __ b(&correct_num_arguments, EQ);
      __ Bind(&wrong_num_arguments);
      if (function.IsClosureFunction()) {
        // Invoke noSuchMethod function passing the original function name.
        // For closure functions, use "call" as the original name.
        const String& name =
            String::Handle(function.IsClosureFunction()
                             ? Symbols::Call().raw()
                             : function.name());
        const int kNumArgsChecked = 1;
        const ICData& ic_data = ICData::ZoneHandle(
            ICData::New(function, name, Object::empty_array(),
                        Isolate::kNoDeoptId, kNumArgsChecked));
        __ LoadObject(R5, ic_data);
        __ LeaveDartFrame();  // The arguments are still on the stack.
        __ Branch(&StubCode::CallNoSuchMethodFunctionLabel());
        // The noSuchMethod call may return to the caller, but not here.
        __ bkpt(0);
      } else {
        __ Stop("Wrong number of arguments");
      }
      __ Bind(&correct_num_arguments);
    }
  } else if (!flow_graph().IsCompiledForOsr()) {
    CopyParameters();
  }

  // In unoptimized code, initialize (non-argument) stack allocated slots to
  // null.
  if (!is_optimizing() && (num_locals > 0)) {
    __ Comment("Initialize spill slots");
    const intptr_t slot_base = parsed_function().first_stack_local_index();
    __ LoadImmediate(R0, reinterpret_cast<intptr_t>(Object::null()));
    for (intptr_t i = 0; i < num_locals; ++i) {
      // Subtract index i (locals lie at lower addresses than FP).
      __ str(R0, Address(FP, (slot_base - i) * kWordSize));
    }
  }

  if (FLAG_print_scopes) {
    // Print the function scope (again) after generating the prologue in order
    // to see annotations such as allocation indices of locals.
    if (FLAG_print_ast) {
      // Second printing.
      OS::Print("Annotated ");
    }
    AstPrinter::PrintFunctionScope(parsed_function());
  }

  VisitBlocks();

  __ bkpt(0);
  GenerateDeferredCode();
  // Emit function patching code. This will be swapped with the first 3
  // instructions at entry point.
  AddCurrentDescriptor(PcDescriptors::kPatchCode,
                       Isolate::kNoDeoptId,
                       0);  // No token position.
  __ BranchPatchable(&StubCode::FixCallersTargetLabel());
  AddCurrentDescriptor(PcDescriptors::kLazyDeoptJump,
                       Isolate::kNoDeoptId,
                       0);  // No token position.
  __ Branch(&StubCode::DeoptimizeLazyLabel());
}


void FlowGraphCompiler::GenerateCall(intptr_t token_pos,
                                     const ExternalLabel* label,
                                     PcDescriptors::Kind kind,
                                     LocationSummary* locs) {
  __ BranchLinkPatchable(label);
  AddCurrentDescriptor(kind, Isolate::kNoDeoptId, token_pos);
  RecordSafepoint(locs);
}


void FlowGraphCompiler::GenerateDartCall(intptr_t deopt_id,
                                         intptr_t token_pos,
                                         const ExternalLabel* label,
                                         PcDescriptors::Kind kind,
                                         LocationSummary* locs) {
  __ BranchLinkPatchable(label);
  AddCurrentDescriptor(kind, deopt_id, token_pos);
  RecordSafepoint(locs);
  // Marks either the continuation point in unoptimized code or the
  // deoptimization point in optimized code, after call.
  const intptr_t deopt_id_after = Isolate::ToDeoptAfter(deopt_id);
  if (is_optimizing()) {
    AddDeoptIndexAtCall(deopt_id_after, token_pos);
  } else {
    // Add deoptimization continuation point after the call and before the
    // arguments are removed.
    AddCurrentDescriptor(PcDescriptors::kDeopt, deopt_id_after, token_pos);
  }
}


void FlowGraphCompiler::GenerateCallRuntime(intptr_t token_pos,
                                            intptr_t deopt_id,
                                            const RuntimeEntry& entry,
                                            LocationSummary* locs) {
  __ CallRuntime(entry);
  AddCurrentDescriptor(PcDescriptors::kOther, deopt_id, token_pos);
  RecordSafepoint(locs);
  if (deopt_id != Isolate::kNoDeoptId) {
    // Marks either the continuation point in unoptimized code or the
    // deoptimization point in optimized code, after call.
    const intptr_t deopt_id_after = Isolate::ToDeoptAfter(deopt_id);
    if (is_optimizing()) {
      AddDeoptIndexAtCall(deopt_id_after, token_pos);
    } else {
      // Add deoptimization continuation point after the call and before the
      // arguments are removed.
      AddCurrentDescriptor(PcDescriptors::kDeopt,
                           deopt_id_after,
                           token_pos);
    }
  }
}


void FlowGraphCompiler::EmitOptimizedInstanceCall(
    ExternalLabel* target_label,
    const ICData& ic_data,
    intptr_t argument_count,
    intptr_t deopt_id,
    intptr_t token_pos,
    LocationSummary* locs) {
  // Each ICData propagated from unoptimized to optimized code contains the
  // function that corresponds to the Dart function of that IC call. Due
  // to inlining in optimized code, that function may not correspond to the
  // top-level function (parsed_function().function()) which could be
  // reoptimized and which counter needs to be incremented.
  // Pass the function explicitly, it is used in IC stub.

  __ LoadObject(R6, parsed_function().function());
  __ LoadObject(R5, ic_data);
  GenerateDartCall(deopt_id,
                   token_pos,
                   target_label,
                   PcDescriptors::kIcCall,
                   locs);
  __ Drop(argument_count);
}


void FlowGraphCompiler::EmitInstanceCall(ExternalLabel* target_label,
                                         const ICData& ic_data,
                                         intptr_t argument_count,
                                         intptr_t deopt_id,
                                         intptr_t token_pos,
                                         LocationSummary* locs) {
  __ LoadObject(R5, ic_data);
  GenerateDartCall(deopt_id,
                   token_pos,
                   target_label,
                   PcDescriptors::kIcCall,
                   locs);
  __ Drop(argument_count);
}


void FlowGraphCompiler::EmitMegamorphicInstanceCall(
    const ICData& ic_data,
    intptr_t argument_count,
    intptr_t deopt_id,
    intptr_t token_pos,
    LocationSummary* locs) {
  MegamorphicCacheTable* table = Isolate::Current()->megamorphic_cache_table();
  const String& name = String::Handle(ic_data.target_name());
  const Array& arguments_descriptor =
      Array::ZoneHandle(ic_data.arguments_descriptor());
  ASSERT(!arguments_descriptor.IsNull());
  const MegamorphicCache& cache =
      MegamorphicCache::ZoneHandle(table->Lookup(name, arguments_descriptor));
  Label not_smi, load_cache;
  __ LoadFromOffset(kWord, R0, SP, (argument_count - 1) * kWordSize);
  __ tst(R0, ShifterOperand(kSmiTagMask));
  __ b(&not_smi, NE);
  __ mov(R0, ShifterOperand(Smi::RawValue(kSmiCid)));
  __ b(&load_cache);

  __ Bind(&not_smi);
  __ LoadClassId(R0, R0);
  __ SmiTag(R0);

  // R0: class ID of the receiver (smi).
  __ Bind(&load_cache);
  __ LoadObject(R1, cache);
  __ ldr(R2, FieldAddress(R1, MegamorphicCache::buckets_offset()));
  __ ldr(R1, FieldAddress(R1, MegamorphicCache::mask_offset()));
  // R2: cache buckets array.
  // R1: mask.
  __ mov(R3, ShifterOperand(R0));

  Label loop, update, call_target_function;
  __ b(&loop);

  __ Bind(&update);
  __ add(R3, R3, ShifterOperand(Smi::RawValue(1)));
  __ Bind(&loop);
  __ and_(R3, R3, ShifterOperand(R1));
  const intptr_t base = Array::data_offset();
  // R3 is smi tagged, but table entries are two words, so LSL 2.
  __ add(IP, R2, ShifterOperand(R3, LSL, 2));
  __ ldr(R4, FieldAddress(IP, base));

  ASSERT(kIllegalCid == 0);
  __ tst(R4, ShifterOperand(R4));
  __ b(&call_target_function, EQ);
  __ cmp(R4, ShifterOperand(R0));
  __ b(&update, NE);

  __ Bind(&call_target_function);
  // Call the target found in the cache.  For a class id match, this is a
  // proper target for the given name and arguments descriptor.  If the
  // illegal class id was found, the target is a cache miss handler that can
  // be invoked as a normal Dart function.
  __ add(IP, R2, ShifterOperand(R3, LSL, 2));
  __ ldr(R0, FieldAddress(IP, base + kWordSize));
  __ ldr(R0, FieldAddress(R0, Function::code_offset()));
  __ ldr(R0, FieldAddress(R0, Code::instructions_offset()));
  __ LoadObject(R5, ic_data);
  __ LoadObject(R4, arguments_descriptor);
  __ AddImmediate(R0, Instructions::HeaderSize() - kHeapObjectTag);
  __ blx(R0);
  AddCurrentDescriptor(PcDescriptors::kOther, Isolate::kNoDeoptId, token_pos);
  RecordSafepoint(locs);
  AddDeoptIndexAtCall(Isolate::ToDeoptAfter(deopt_id), token_pos);
  __ Drop(argument_count);
}


void FlowGraphCompiler::EmitOptimizedStaticCall(
    const Function& function,
    const Array& arguments_descriptor,
    intptr_t argument_count,
    intptr_t deopt_id,
    intptr_t token_pos,
    LocationSummary* locs) {
  __ LoadObject(R4, arguments_descriptor);
  // Do not use the code from the function, but let the code be patched so that
  // we can record the outgoing edges to other code.
  GenerateDartCall(deopt_id,
                   token_pos,
                   &StubCode::CallStaticFunctionLabel(),
                   PcDescriptors::kOptStaticCall,
                   locs);
  AddStaticCallTarget(function);
  __ Drop(argument_count);
}


void FlowGraphCompiler::EmitEqualityRegConstCompare(Register reg,
                                                    const Object& obj,
                                                    bool needs_number_check,
                                                    intptr_t token_pos) {
  if (needs_number_check &&
      (obj.IsMint() || obj.IsDouble() || obj.IsBigint())) {
    __ Push(reg);
    __ PushObject(obj);
    if (is_optimizing()) {
      __ BranchLink(&StubCode::OptimizedIdenticalWithNumberCheckLabel());
    } else {
      __ BranchLink(&StubCode::UnoptimizedIdenticalWithNumberCheckLabel());
    }
    AddCurrentDescriptor(PcDescriptors::kRuntimeCall,
                         Isolate::kNoDeoptId,
                         token_pos);
    __ Drop(1);  // Discard constant.
    __ Pop(reg);  // Restore 'reg'.
    return;
  }

  __ CompareObject(reg, obj);
}


void FlowGraphCompiler::EmitEqualityRegRegCompare(Register left,
                                                  Register right,
                                                  bool needs_number_check,
                                                  intptr_t token_pos) {
  if (needs_number_check) {
    __ Push(left);
    __ Push(right);
    if (is_optimizing()) {
      __ BranchLink(&StubCode::OptimizedIdenticalWithNumberCheckLabel());
    } else {
      __ BranchLink(&StubCode::UnoptimizedIdenticalWithNumberCheckLabel());
    }
    AddCurrentDescriptor(PcDescriptors::kRuntimeCall,
                         Isolate::kNoDeoptId,
                         token_pos);
    // Stub returns result in flags (result of a cmpl, we need ZF computed).
    __ Pop(right);
    __ Pop(left);
  } else {
    __ cmp(left, ShifterOperand(right));
  }
}


// Implement equality spec: if any of the arguments is null do identity check.
// Fallthrough calls super equality.
void FlowGraphCompiler::EmitSuperEqualityCallPrologue(Register result,
                                                      Label* skip_call) {
  Label check_identity, fall_through;
  __ LoadImmediate(IP, reinterpret_cast<intptr_t>(Object::null()));
  __ ldr(result, Address(SP, 0 * kWordSize));  // Load right operand.
  __ cmp(result, ShifterOperand(IP));  // Is right null?
  __ ldr(result, Address(SP, 1 * kWordSize));  // Load left operand.
  __ b(&check_identity, EQ);  // Branch if right (IP) is null; left in result.
  __ cmp(result, ShifterOperand(IP));  // Right is non-null; is left null?
  __ b(&fall_through, NE);
  // Right is non-null, left is null. We could return false, but we save code
  // by falling through with an IP different than null.
  __ mov(IP, ShifterOperand(0));
  __ Bind(&check_identity);
  __ cmp(result, ShifterOperand(IP));
  __ LoadObject(result, Bool::True(), EQ);
  __ LoadObject(result, Bool::False(), NE);
  __ Drop(2);
  __ b(skip_call);
  __ Bind(&fall_through);
}


void FlowGraphCompiler::SaveLiveRegisters(LocationSummary* locs) {
  // TODO(vegorov): consider saving only caller save (volatile) registers.
  const intptr_t fpu_regs_count = locs->live_registers()->fpu_regs_count();
  if (fpu_regs_count > 0) {
    __ AddImmediate(SP, -(fpu_regs_count * kFpuRegisterSize));
    // Store fpu registers with the lowest register number at the lowest
    // address.
    intptr_t offset = 0;
    for (intptr_t reg_idx = 0; reg_idx < kNumberOfFpuRegisters; ++reg_idx) {
      QRegister fpu_reg = static_cast<QRegister>(reg_idx);
      if (locs->live_registers()->ContainsFpuRegister(fpu_reg)) {
        DRegister d1 = EvenDRegisterOf(fpu_reg);
        DRegister d2 = OddDRegisterOf(fpu_reg);
        // TOOD(regis): merge stores using vstmd instruction.
        __ vstrd(d1, Address(SP, offset));
        __ vstrd(d2, Address(SP, offset + 2 * kWordSize));
        offset += kFpuRegisterSize;
      }
    }
    ASSERT(offset == (fpu_regs_count * kFpuRegisterSize));
  }

  // Store general purpose registers with the lowest register number at the
  // lowest address.
  const intptr_t cpu_registers = locs->live_registers()->cpu_registers();
  ASSERT((cpu_registers & ~kAllCpuRegistersList) == 0);
  if (cpu_registers != 0) {
    __ PushList(cpu_registers);
  }
}


void FlowGraphCompiler::RestoreLiveRegisters(LocationSummary* locs) {
  // General purpose registers have the lowest register number at the
  // lowest address.
  const intptr_t cpu_registers = locs->live_registers()->cpu_registers();
  ASSERT((cpu_registers & ~kAllCpuRegistersList) == 0);
  if (cpu_registers != 0) {
    __ PopList(cpu_registers);
  }

  const intptr_t fpu_regs_count = locs->live_registers()->fpu_regs_count();
  if (fpu_regs_count > 0) {
    // Fpu registers have the lowest register number at the lowest address.
    intptr_t offset = 0;
    for (intptr_t reg_idx = 0; reg_idx < kNumberOfFpuRegisters; ++reg_idx) {
      QRegister fpu_reg = static_cast<QRegister>(reg_idx);
      if (locs->live_registers()->ContainsFpuRegister(fpu_reg)) {
        DRegister d1 = EvenDRegisterOf(fpu_reg);
        DRegister d2 = OddDRegisterOf(fpu_reg);
        // TOOD(regis): merge loads using vldmd instruction.
        __ vldrd(d1, Address(SP, offset));
        __ vldrd(d2, Address(SP, offset + 2 * kWordSize));
        offset += kFpuRegisterSize;
      }
    }
    ASSERT(offset == (fpu_regs_count * kFpuRegisterSize));
    __ AddImmediate(SP, offset);
  }
}


void FlowGraphCompiler::EmitTestAndCall(const ICData& ic_data,
                                        Register class_id_reg,
                                        intptr_t argument_count,
                                        const Array& argument_names,
                                        Label* deopt,
                                        intptr_t deopt_id,
                                        intptr_t token_index,
                                        LocationSummary* locs) {
  ASSERT(is_optimizing());
  ASSERT(!ic_data.IsNull() && (ic_data.NumberOfChecks() > 0));
  Label match_found;
  const intptr_t len = ic_data.NumberOfChecks();
  GrowableArray<CidTarget> sorted(len);
  SortICDataByCount(ic_data, &sorted);
  ASSERT(class_id_reg != R4);
  ASSERT(len > 0);  // Why bother otherwise.
  const Array& arguments_descriptor =
      Array::ZoneHandle(ArgumentsDescriptor::New(argument_count,
                                                 argument_names));
  __ LoadObject(R4, arguments_descriptor);
  for (intptr_t i = 0; i < len; i++) {
    const bool is_last_check = (i == (len - 1));
    Label next_test;
    assembler()->CompareImmediate(class_id_reg, sorted[i].cid);
    if (is_last_check) {
      assembler()->b(deopt, NE);
    } else {
      assembler()->b(&next_test, NE);
    }
    // Do not use the code from the function, but let the code be patched so
    // that we can record the outgoing edges to other code.
    GenerateDartCall(deopt_id,
                     token_index,
                     &StubCode::CallStaticFunctionLabel(),
                     PcDescriptors::kOptStaticCall,
                     locs);
    const Function& function = *sorted[i].target;
    AddStaticCallTarget(function);
    __ Drop(argument_count);
    if (!is_last_check) {
      assembler()->b(&match_found);
    }
    assembler()->Bind(&next_test);
  }
  assembler()->Bind(&match_found);
}


void FlowGraphCompiler::EmitDoubleCompareBranch(Condition true_condition,
                                                FpuRegister left,
                                                FpuRegister right,
                                                BranchInstr* branch) {
  ASSERT(branch != NULL);
  DRegister dleft = EvenDRegisterOf(left);
  DRegister dright = EvenDRegisterOf(right);
  assembler()->vcmpd(dleft, dright);
  assembler()->vmstat();
  BlockEntryInstr* nan_result = (true_condition == NE) ?
      branch->true_successor() : branch->false_successor();
  assembler()->b(GetJumpLabel(nan_result), VS);
  branch->EmitBranchOnCondition(this, true_condition);
}


void FlowGraphCompiler::EmitDoubleCompareBool(Condition true_condition,
                                              FpuRegister left,
                                              FpuRegister right,
                                              Register result) {
  DRegister dleft = EvenDRegisterOf(left);
  DRegister dright = EvenDRegisterOf(right);
  assembler()->vcmpd(dleft, dright);
  assembler()->vmstat();
  assembler()->LoadObject(result, Bool::False());
  Label done;
  assembler()->b(&done, VS);  // NaN -> false.
  assembler()->LoadObject(result, Bool::True(), true_condition);
  assembler()->Bind(&done);
}


// Do not implement or use this function.
FieldAddress FlowGraphCompiler::ElementAddressForIntIndex(intptr_t cid,
                                                          intptr_t index_scale,
                                                          Register array,
                                                          intptr_t index) {
  UNREACHABLE();
  return FieldAddress(array, index);
}


// Do not implement or use this function.
FieldAddress FlowGraphCompiler::ElementAddressForRegIndex(intptr_t cid,
                                                          intptr_t index_scale,
                                                          Register array,
                                                          Register index) {
  UNREACHABLE();  // No register indexed with offset addressing mode on ARM.
  return FieldAddress(array, index);
}


Address FlowGraphCompiler::ExternalElementAddressForIntIndex(
    intptr_t index_scale,
    Register array,
    intptr_t index) {
  UNREACHABLE();
  return FieldAddress(array, index);
}


Address FlowGraphCompiler::ExternalElementAddressForRegIndex(
    intptr_t index_scale,
    Register array,
    Register index) {
  UNREACHABLE();
  return FieldAddress(array, index);
}


#undef __
#define __ compiler_->assembler()->


void ParallelMoveResolver::EmitMove(int index) {
  MoveOperands* move = moves_[index];
  const Location source = move->src();
  const Location destination = move->dest();

  if (source.IsRegister()) {
    if (destination.IsRegister()) {
      __ mov(destination.reg(), ShifterOperand(source.reg()));
    } else {
      ASSERT(destination.IsStackSlot());
      const intptr_t dest_offset = destination.ToStackSlotOffset();
      __ StoreToOffset(kWord, source.reg(), FP, dest_offset);
    }
  } else if (source.IsStackSlot()) {
    if (destination.IsRegister()) {
      const intptr_t source_offset = source.ToStackSlotOffset();
      __ LoadFromOffset(kWord, destination.reg(), FP, source_offset);
    } else {
      ASSERT(destination.IsStackSlot());
      const intptr_t source_offset = source.ToStackSlotOffset();
      const intptr_t dest_offset = destination.ToStackSlotOffset();
      __ LoadFromOffset(kWord, TMP, FP, source_offset);
      __ StoreToOffset(kWord, TMP, FP, dest_offset);
    }
  } else if (source.IsFpuRegister()) {
    if (destination.IsFpuRegister()) {
      __ vmovq(destination.fpu_reg(), source.fpu_reg());
    } else {
      if (destination.IsDoubleStackSlot()) {
        const intptr_t dest_offset = destination.ToStackSlotOffset();
        DRegister src = EvenDRegisterOf(source.fpu_reg());
        __ StoreDToOffset(src, FP, dest_offset);
      } else {
        ASSERT(destination.IsQuadStackSlot());
        const intptr_t dest_offset = destination.ToStackSlotOffset();
        DRegister dsrc0 = EvenDRegisterOf(source.fpu_reg());
        DRegister dsrc1 = OddDRegisterOf(source.fpu_reg());
        // TODO(zra): Write and use {Load,Store}Q{From,To}Offset(), which can
        // use a single vld1/vst1 instruction.
        __ StoreDToOffset(dsrc0, FP, dest_offset);
        __ StoreDToOffset(dsrc1, FP, dest_offset + 2*kWordSize);
      }
    }
  } else if (source.IsDoubleStackSlot()) {
    if (destination.IsFpuRegister()) {
      const intptr_t dest_offset = source.ToStackSlotOffset();
      DRegister dst = EvenDRegisterOf(destination.fpu_reg());
      __ LoadDFromOffset(dst, FP, dest_offset);
    } else {
      ASSERT(destination.IsDoubleStackSlot());
      const intptr_t source_offset = source.ToStackSlotOffset();
      const intptr_t dest_offset = destination.ToStackSlotOffset();
      __ LoadDFromOffset(DTMP, FP, source_offset);
      __ StoreDToOffset(DTMP, FP, dest_offset);
    }
  } else if (source.IsQuadStackSlot()) {
    if (destination.IsFpuRegister()) {
      const intptr_t dest_offset = source.ToStackSlotOffset();
      DRegister dst0 = EvenDRegisterOf(destination.fpu_reg());
      DRegister dst1 = OddDRegisterOf(destination.fpu_reg());
      __ LoadDFromOffset(dst0, FP, dest_offset);
      __ LoadDFromOffset(dst1, FP, dest_offset + 2*kWordSize);
    } else {
      ASSERT(destination.IsQuadStackSlot());
      const intptr_t source_offset = source.ToStackSlotOffset();
      const intptr_t dest_offset = destination.ToStackSlotOffset();
      DRegister dtmp0 = DTMP;
      DRegister dtmp1 = OddDRegisterOf(QTMP);
      __ LoadDFromOffset(dtmp0, FP, source_offset);
      __ LoadDFromOffset(dtmp1, FP, source_offset + 2*kWordSize);
      __ StoreDToOffset(dtmp0, FP, dest_offset);
      __ StoreDToOffset(dtmp1, FP, dest_offset + 2*kWordSize);
    }
  } else {
    ASSERT(source.IsConstant());
    if (destination.IsRegister()) {
      const Object& constant = source.constant();
      __ LoadObject(destination.reg(), constant);
    } else {
      ASSERT(destination.IsStackSlot());
      const intptr_t dest_offset = destination.ToStackSlotOffset();
      __ LoadObject(TMP, source.constant());
      __ StoreToOffset(kWord, TMP, FP, dest_offset);
    }
  }

  move->Eliminate();
}


void ParallelMoveResolver::EmitSwap(int index) {
  MoveOperands* move = moves_[index];
  const Location source = move->src();
  const Location destination = move->dest();

  if (source.IsRegister() && destination.IsRegister()) {
    ASSERT(source.reg() != IP);
    ASSERT(destination.reg() != IP);
    __ mov(IP, ShifterOperand(source.reg()));
    __ mov(source.reg(), ShifterOperand(destination.reg()));
    __ mov(destination.reg(), ShifterOperand(IP));
  } else if (source.IsRegister() && destination.IsStackSlot()) {
    Exchange(source.reg(), destination.ToStackSlotOffset());
  } else if (source.IsStackSlot() && destination.IsRegister()) {
    Exchange(destination.reg(), source.ToStackSlotOffset());
  } else if (source.IsStackSlot() && destination.IsStackSlot()) {
    Exchange(source.ToStackSlotOffset(), destination.ToStackSlotOffset());
  } else if (source.IsFpuRegister() && destination.IsFpuRegister()) {
    DRegister dst = EvenDRegisterOf(destination.fpu_reg());
    DRegister src = EvenDRegisterOf(source.fpu_reg());
    __ vmovd(DTMP, src);
    __ vmovd(src, dst);
    __ vmovd(dst, DTMP);
  } else if (source.IsFpuRegister() || destination.IsFpuRegister()) {
    ASSERT(destination.IsDoubleStackSlot() ||
           destination.IsQuadStackSlot() ||
           source.IsDoubleStackSlot() ||
           source.IsQuadStackSlot());
    bool double_width = destination.IsDoubleStackSlot() ||
                        source.IsDoubleStackSlot();
    QRegister qreg = source.IsFpuRegister() ? source.fpu_reg()
                                            : destination.fpu_reg();
    DRegister reg = EvenDRegisterOf(qreg);
    const intptr_t slot_offset = source.IsFpuRegister()
        ? destination.ToStackSlotOffset()
        : source.ToStackSlotOffset();

    if (double_width) {
      __ LoadDFromOffset(DTMP, FP, slot_offset);
      __ StoreDToOffset(reg, FP, slot_offset);
      __ vmovd(reg, DTMP);
    } else {
      UNIMPLEMENTED();
    }
  } else if (source.IsDoubleStackSlot() && destination.IsDoubleStackSlot()) {
    const intptr_t source_offset = source.ToStackSlotOffset();
    const intptr_t dest_offset = destination.ToStackSlotOffset();

    ScratchFpuRegisterScope ensure_scratch(this, QTMP);
    DRegister scratch = EvenDRegisterOf(ensure_scratch.reg());
    __ LoadDFromOffset(DTMP, FP, source_offset);
    __ LoadDFromOffset(scratch, FP, dest_offset);
    __ StoreDToOffset(DTMP, FP, dest_offset);
    __ StoreDToOffset(scratch, FP, source_offset);
  } else if (source.IsQuadStackSlot() && destination.IsQuadStackSlot()) {
    UNIMPLEMENTED();
  } else {
    UNREACHABLE();
  }

  // The swap of source and destination has executed a move from source to
  // destination.
  move->Eliminate();

  // Any unperformed (including pending) move with a source of either
  // this move's source or destination needs to have their source
  // changed to reflect the state of affairs after the swap.
  for (int i = 0; i < moves_.length(); ++i) {
    const MoveOperands& other_move = *moves_[i];
    if (other_move.Blocks(source)) {
      moves_[i]->set_src(destination);
    } else if (other_move.Blocks(destination)) {
      moves_[i]->set_src(source);
    }
  }
}


void ParallelMoveResolver::MoveMemoryToMemory(const Address& dst,
                                              const Address& src) {
  __ ldr(IP, src);
  __ str(IP, dst);
}


void ParallelMoveResolver::StoreObject(const Address& dst, const Object& obj) {
  __ LoadObject(IP, obj);
  __ str(IP, dst);
}


// Do not call or implement this function. Instead, use the form below that
// uses an offset from the frame pointer instead of an Address.
void ParallelMoveResolver::Exchange(Register reg, const Address& mem) {
  UNREACHABLE();
}


// Do not call or implement this function. Instead, use the form below that
// uses offsets from the frame pointer instead of Addresses.
void ParallelMoveResolver::Exchange(const Address& mem1, const Address& mem2) {
  UNREACHABLE();
}


void ParallelMoveResolver::Exchange(Register reg, intptr_t stack_offset) {
  __ mov(TMP, ShifterOperand(reg));
  __ LoadFromOffset(kWord, reg, FP, stack_offset);
  __ StoreToOffset(kWord, TMP, FP, stack_offset);
}


void ParallelMoveResolver::Exchange(intptr_t stack_offset1,
                                    intptr_t stack_offset2) {
  ScratchRegisterScope ensure_scratch(this, IP);
  __ LoadFromOffset(kWord, ensure_scratch.reg(), FP, stack_offset1);
  __ LoadFromOffset(kWord, TMP, FP, stack_offset2);
  __ StoreToOffset(kWord, ensure_scratch.reg(), FP, stack_offset2);
  __ StoreToOffset(kWord, TMP, FP, stack_offset1);
}


void ParallelMoveResolver::SpillScratch(Register reg) {
  __ Push(reg);
}


void ParallelMoveResolver::RestoreScratch(Register reg) {
  __ Pop(reg);
}


void ParallelMoveResolver::SpillFpuScratch(FpuRegister reg) {
  DRegister dreg = EvenDRegisterOf(reg);
  __ vstrd(dreg, Address(SP, -kDoubleSize, Address::PreIndex));
}


void ParallelMoveResolver::RestoreFpuScratch(FpuRegister reg) {
  DRegister dreg = EvenDRegisterOf(reg);
  __ vldrd(dreg, Address(SP, kDoubleSize, Address::PostIndex));
}


#undef __

}  // namespace dart

#endif  // defined TARGET_ARCH_ARM