src/ia32/deoptimizer-ia32.cc

// Copyright 2012 the V8 project authors. All rights reserved.
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are
// met:
//
//     * Redistributions of source code must retain the above copyright
//       notice, this list of conditions and the following disclaimer.
//     * Redistributions in binary form must reproduce the above
//       copyright notice, this list of conditions and the following
//       disclaimer in the documentation and/or other materials provided
//       with the distribution.
//     * Neither the name of Google Inc. nor the names of its
//       contributors may be used to endorse or promote products derived
//       from this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

#include "v8.h"

#if V8_TARGET_ARCH_IA32

#include "codegen.h"
#include "deoptimizer.h"
#include "full-codegen.h"
#include "safepoint-table.h"

namespace v8 {
namespace internal {

const int Deoptimizer::table_entry_size_ = 10;


int Deoptimizer::patch_size() {
  return Assembler::kCallInstructionLength;
}


void Deoptimizer::EnsureRelocSpaceForLazyDeoptimization(Handle<Code> code) {
  Isolate* isolate = code->GetIsolate();
  HandleScope scope(isolate);

  // Compute the size of relocation information needed for the code
  // patching in Deoptimizer::DeoptimizeFunction.
  int min_reloc_size = 0;
  int prev_pc_offset = 0;
  DeoptimizationInputData* deopt_data =
      DeoptimizationInputData::cast(code->deoptimization_data());
  for (int i = 0; i < deopt_data->DeoptCount(); i++) {
    int pc_offset = deopt_data->Pc(i)->value();
    if (pc_offset == -1) continue;
    ASSERT_GE(pc_offset, prev_pc_offset);
    int pc_delta = pc_offset - prev_pc_offset;
    // We use RUNTIME_ENTRY reloc info which has a size of 2 bytes
    // if encodable with small pc delta encoding and up to 6 bytes
    // otherwise.
    if (pc_delta <= RelocInfo::kMaxSmallPCDelta) {
      min_reloc_size += 2;
    } else {
      min_reloc_size += 6;
    }
    prev_pc_offset = pc_offset;
  }

  // If the relocation information is not big enough we create a new
  // relocation info object that is padded with comments to make it
  // big enough for lazy doptimization.
  int reloc_length = code->relocation_info()->length();
  if (min_reloc_size > reloc_length) {
    int comment_reloc_size = RelocInfo::kMinRelocCommentSize;
    // Padding needed.
    int min_padding = min_reloc_size - reloc_length;
    // Number of comments needed to take up at least that much space.
    int additional_comments =
        (min_padding + comment_reloc_size - 1) / comment_reloc_size;
    // Actual padding size.
    int padding = additional_comments * comment_reloc_size;
    // Allocate new relocation info and copy old relocation to the end
    // of the new relocation info array because relocation info is
    // written and read backwards.
    Factory* factory = isolate->factory();
    Handle<ByteArray> new_reloc =
        factory->NewByteArray(reloc_length + padding, TENURED);
    OS::MemCopy(new_reloc->GetDataStartAddress() + padding,
                code->relocation_info()->GetDataStartAddress(),
                reloc_length);
    // Create a relocation writer to write the comments in the padding
    // space. Use position 0 for everything to ensure short encoding.
    RelocInfoWriter reloc_info_writer(
        new_reloc->GetDataStartAddress() + padding, 0);
    intptr_t comment_string
        = reinterpret_cast<intptr_t>(RelocInfo::kFillerCommentString);
    RelocInfo rinfo(0, RelocInfo::COMMENT, comment_string, NULL);
    for (int i = 0; i < additional_comments; ++i) {
#ifdef DEBUG
      byte* pos_before = reloc_info_writer.pos();
#endif
      reloc_info_writer.Write(&rinfo);
      ASSERT(RelocInfo::kMinRelocCommentSize ==
             pos_before - reloc_info_writer.pos());
    }
    // Replace relocation information on the code object.
    code->set_relocation_info(*new_reloc);
  }
}


void Deoptimizer::PatchCodeForDeoptimization(Isolate* isolate, Code* code) {
  Address code_start_address = code->instruction_start();
  // We will overwrite the code's relocation info in-place. Relocation info
  // is written backward. The relocation info is the payload of a byte
  // array.  Later on we will slide this to the start of the byte array and
  // create a filler object in the remaining space.
  ByteArray* reloc_info = code->relocation_info();
  Address reloc_end_address = reloc_info->address() + reloc_info->Size();
  RelocInfoWriter reloc_info_writer(reloc_end_address, code_start_address);

  // For each LLazyBailout instruction insert a call to the corresponding
  // deoptimization entry.

  // Since the call is a relative encoding, write new
  // reloc info.  We do not need any of the existing reloc info because the
  // existing code will not be used again (we zap it in debug builds).
  //
  // Emit call to lazy deoptimization at all lazy deopt points.
  DeoptimizationInputData* deopt_data =
      DeoptimizationInputData::cast(code->deoptimization_data());
#ifdef DEBUG
  Address prev_call_address = NULL;
#endif
  for (int i = 0; i < deopt_data->DeoptCount(); i++) {
    if (deopt_data->Pc(i)->value() == -1) continue;
    // Patch lazy deoptimization entry.
    Address call_address = code_start_address + deopt_data->Pc(i)->value();
    CodePatcher patcher(call_address, patch_size());
    Address deopt_entry = GetDeoptimizationEntry(isolate, i, LAZY);
    patcher.masm()->call(deopt_entry, RelocInfo::NONE32);
    // We use RUNTIME_ENTRY for deoptimization bailouts.
    RelocInfo rinfo(call_address + 1,  // 1 after the call opcode.
                    RelocInfo::RUNTIME_ENTRY,
                    reinterpret_cast<intptr_t>(deopt_entry),
                    NULL);
    reloc_info_writer.Write(&rinfo);
    ASSERT_GE(reloc_info_writer.pos(),
              reloc_info->address() + ByteArray::kHeaderSize);
    ASSERT(prev_call_address == NULL ||
           call_address >= prev_call_address + patch_size());
    ASSERT(call_address + patch_size() <= code->instruction_end());
#ifdef DEBUG
    prev_call_address = call_address;
#endif
  }

  // Move the relocation info to the beginning of the byte array.
  int new_reloc_size = reloc_end_address - reloc_info_writer.pos();
  OS::MemMove(
      code->relocation_start(), reloc_info_writer.pos(), new_reloc_size);

  // The relocation info is in place, update the size.
  reloc_info->set_length(new_reloc_size);

  // Handle the junk part after the new relocation info. We will create
  // a non-live object in the extra space at the end of the former reloc info.
  Address junk_address = reloc_info->address() + reloc_info->Size();
  ASSERT(junk_address <= reloc_end_address);
  isolate->heap()->CreateFillerObjectAt(junk_address,
                                        reloc_end_address - junk_address);
}


static const byte kJnsInstruction = 0x79;
static const byte kJnsOffset = 0x11;
static const byte kCallInstruction = 0xe8;
static const byte kNopByteOne = 0x66;
static const byte kNopByteTwo = 0x90;

// The back edge bookkeeping code matches the pattern:
//
//     sub <profiling_counter>, <delta>
//     jns ok
//     call <interrupt stub>
//   ok:
//
// The patched back edge looks like this:
//
//     sub <profiling_counter>, <delta>  ;; Not changed
//     nop
//     nop
//     call <on-stack replacment>
//   ok:

void Deoptimizer::PatchInterruptCodeAt(Code* unoptimized_code,
                                       Address pc_after,
                                       Code* replacement_code) {
  // Turn the jump into nops.
  Address call_target_address = pc_after - kIntSize;
  *(call_target_address - 3) = kNopByteOne;
  *(call_target_address - 2) = kNopByteTwo;
  // Replace the call address.
  Assembler::set_target_address_at(call_target_address,
                                   replacement_code->entry());

  unoptimized_code->GetHeap()->incremental_marking()->RecordCodeTargetPatch(
      unoptimized_code, call_target_address, replacement_code);
}


void Deoptimizer::RevertInterruptCodeAt(Code* unoptimized_code,
                                        Address pc_after,
                                        Code* interrupt_code) {
  // Restore the original jump.
  Address call_target_address = pc_after - kIntSize;
  *(call_target_address - 3) = kJnsInstruction;
  *(call_target_address - 2) = kJnsOffset;
  // Restore the original call address.
  Assembler::set_target_address_at(call_target_address,
                                   interrupt_code->entry());

  interrupt_code->GetHeap()->incremental_marking()->RecordCodeTargetPatch(
      unoptimized_code, call_target_address, interrupt_code);
}


#ifdef DEBUG
Deoptimizer::InterruptPatchState Deoptimizer::GetInterruptPatchState(
    Isolate* isolate,
    Code* unoptimized_code,
    Address pc_after) {
  Address call_target_address = pc_after - kIntSize;
  ASSERT_EQ(kCallInstruction, *(call_target_address - 1));
  if (*(call_target_address - 3) == kNopByteOne) {
    ASSERT_EQ(kNopByteTwo,      *(call_target_address - 2));
    Code* osr_builtin =
        isolate->builtins()->builtin(Builtins::kOnStackReplacement);
    ASSERT_EQ(osr_builtin->entry(),
              Assembler::target_address_at(call_target_address));
    return PATCHED_FOR_OSR;
  } else {
    // Get the interrupt stub code object to match against from cache.
    Code* interrupt_builtin =
        isolate->builtins()->builtin(Builtins::kInterruptCheck);
    ASSERT_EQ(interrupt_builtin->entry(),
              Assembler::target_address_at(call_target_address));
    ASSERT_EQ(kJnsInstruction,  *(call_target_address - 3));
    ASSERT_EQ(kJnsOffset,       *(call_target_address - 2));
    return NOT_PATCHED;
  }
}
#endif  // DEBUG


static int LookupBailoutId(DeoptimizationInputData* data, BailoutId ast_id) {
  ByteArray* translations = data->TranslationByteArray();
  int length = data->DeoptCount();
  for (int i = 0; i < length; i++) {
    if (data->AstId(i) == ast_id) {
      TranslationIterator it(translations,  data->TranslationIndex(i)->value());
      int value = it.Next();
      ASSERT(Translation::BEGIN == static_cast<Translation::Opcode>(value));
      // Read the number of frames.
      value = it.Next();
      if (value == 1) return i;
    }
  }
  UNREACHABLE();
  return -1;
}


void Deoptimizer::DoComputeOsrOutputFrame() {
  DeoptimizationInputData* data = DeoptimizationInputData::cast(
      compiled_code_->deoptimization_data());
  unsigned ast_id = data->OsrAstId()->value();
  // TODO(kasperl): This should not be the bailout_id_. It should be
  // the ast id. Confusing.
  ASSERT(bailout_id_ == ast_id);

  int bailout_id = LookupBailoutId(data, BailoutId(ast_id));
  unsigned translation_index = data->TranslationIndex(bailout_id)->value();
  ByteArray* translations = data->TranslationByteArray();

  TranslationIterator iterator(translations, translation_index);
  Translation::Opcode opcode =
      static_cast<Translation::Opcode>(iterator.Next());
  ASSERT(Translation::BEGIN == opcode);
  USE(opcode);
  int count = iterator.Next();
  iterator.Next();  // Drop JS frames count.
  ASSERT(count == 1);
  USE(count);

  opcode = static_cast<Translation::Opcode>(iterator.Next());
  USE(opcode);
  ASSERT(Translation::JS_FRAME == opcode);
  unsigned node_id = iterator.Next();
  USE(node_id);
  ASSERT(node_id == ast_id);
  int closure_id = iterator.Next();
  USE(closure_id);
  ASSERT_EQ(Translation::kSelfLiteralId, closure_id);
  unsigned height = iterator.Next();
  unsigned height_in_bytes = height * kPointerSize;
  USE(height_in_bytes);

  unsigned fixed_size = ComputeFixedSize(function_);
  unsigned input_frame_size = input_->GetFrameSize();
  ASSERT(fixed_size + height_in_bytes == input_frame_size);

  unsigned stack_slot_size = compiled_code_->stack_slots() * kPointerSize;
  unsigned outgoing_height = data->ArgumentsStackHeight(bailout_id)->value();
  unsigned outgoing_size = outgoing_height * kPointerSize;
  unsigned output_frame_size = fixed_size + stack_slot_size + outgoing_size;
  ASSERT(outgoing_size == 0);  // OSR does not happen in the middle of a call.

  if (FLAG_trace_osr) {
    PrintF("[on-stack replacement: begin 0x%08" V8PRIxPTR " ",
           reinterpret_cast<intptr_t>(function_));
    PrintFunctionName();
    PrintF(" => node=%u, frame=%d->%d, ebp:esp=0x%08x:0x%08x]\n",
           ast_id,
           input_frame_size,
           output_frame_size,
           input_->GetRegister(ebp.code()),
           input_->GetRegister(esp.code()));
  }

  // There's only one output frame in the OSR case.
  output_count_ = 1;
  output_ = new FrameDescription*[1];
  output_[0] = new(output_frame_size) FrameDescription(
      output_frame_size, function_);
  output_[0]->SetFrameType(StackFrame::JAVA_SCRIPT);

  // Clear the incoming parameters in the optimized frame to avoid
  // confusing the garbage collector.
  unsigned output_offset = output_frame_size - kPointerSize;
  int parameter_count = function_->shared()->formal_parameter_count() + 1;
  for (int i = 0; i < parameter_count; ++i) {
    output_[0]->SetFrameSlot(output_offset, 0);
    output_offset -= kPointerSize;
  }

  // Translate the incoming parameters. This may overwrite some of the
  // incoming argument slots we've just cleared.
  int input_offset = input_frame_size - kPointerSize;
  bool ok = true;
  int limit = input_offset - (parameter_count * kPointerSize);
  while (ok && input_offset > limit) {
    ok = DoOsrTranslateCommand(&iterator, &input_offset);
  }

  // There are no translation commands for the caller's pc and fp, the
  // context, and the function.  Set them up explicitly.
  for (int i =  StandardFrameConstants::kCallerPCOffset;
       ok && i >=  StandardFrameConstants::kMarkerOffset;
       i -= kPointerSize) {
    uint32_t input_value = input_->GetFrameSlot(input_offset);
    if (FLAG_trace_osr) {
      const char* name = "UNKNOWN";
      switch (i) {
        case StandardFrameConstants::kCallerPCOffset:
          name = "caller's pc";
          break;
        case StandardFrameConstants::kCallerFPOffset:
          name = "fp";
          break;
        case StandardFrameConstants::kContextOffset:
          name = "context";
          break;
        case StandardFrameConstants::kMarkerOffset:
          name = "function";
          break;
      }
      PrintF("    [sp + %d] <- 0x%08x ; [sp + %d] (fixed part - %s)\n",
             output_offset,
             input_value,
             input_offset,
             name);
    }
    output_[0]->SetFrameSlot(output_offset, input_->GetFrameSlot(input_offset));
    input_offset -= kPointerSize;
    output_offset -= kPointerSize;
  }

  // All OSR stack frames are dynamically aligned to an 8-byte boundary.
  int frame_pointer = input_->GetRegister(ebp.code());
  if ((frame_pointer & kPointerSize) != 0) {
    frame_pointer -= kPointerSize;
    has_alignment_padding_ = 1;
  }

  int32_t alignment_state = (has_alignment_padding_ == 1) ?
    kAlignmentPaddingPushed :
    kNoAlignmentPadding;
  if (FLAG_trace_osr) {
    PrintF("    [sp + %d] <- 0x%08x ; (alignment state)\n",
           output_offset,
           alignment_state);
  }
  output_[0]->SetFrameSlot(output_offset, alignment_state);
  output_offset -= kPointerSize;

  // Translate the rest of the frame.
  while (ok && input_offset >= 0) {
    ok = DoOsrTranslateCommand(&iterator, &input_offset);
  }

  // If translation of any command failed, continue using the input frame.
  if (!ok) {
    delete output_[0];
    output_[0] = input_;
    output_[0]->SetPc(reinterpret_cast<uint32_t>(from_));
  } else {
    // Set up the frame pointer and the context pointer.
    output_[0]->SetRegister(ebp.code(), frame_pointer);
    output_[0]->SetRegister(esi.code(), input_->GetRegister(esi.code()));

    unsigned pc_offset = data->OsrPcOffset()->value();
    uint32_t pc = reinterpret_cast<uint32_t>(
        compiled_code_->entry() + pc_offset);
    output_[0]->SetPc(pc);
  }
  Code* continuation =
      function_->GetIsolate()->builtins()->builtin(Builtins::kNotifyOSR);
  output_[0]->SetContinuation(
      reinterpret_cast<uint32_t>(continuation->entry()));

  if (FLAG_trace_osr) {
    PrintF("[on-stack replacement translation %s: 0x%08" V8PRIxPTR " ",
           ok ? "finished" : "aborted",
           reinterpret_cast<intptr_t>(function_));
    PrintFunctionName();
    PrintF(" => pc=0x%0x]\n", output_[0]->GetPc());
  }
}


void Deoptimizer::FillInputFrame(Address tos, JavaScriptFrame* frame) {
  // Set the register values. The values are not important as there are no
  // callee saved registers in JavaScript frames, so all registers are
  // spilled. Registers ebp and esp are set to the correct values though.

  for (int i = 0; i < Register::kNumRegisters; i++) {
    input_->SetRegister(i, i * 4);
  }
  input_->SetRegister(esp.code(), reinterpret_cast<intptr_t>(frame->sp()));
  input_->SetRegister(ebp.code(), reinterpret_cast<intptr_t>(frame->fp()));
  for (int i = 0; i < DoubleRegister::NumAllocatableRegisters(); i++) {
    input_->SetDoubleRegister(i, 0.0);
  }

  // Fill the frame content from the actual data on the frame.
  for (unsigned i = 0; i < input_->GetFrameSize(); i += kPointerSize) {
    input_->SetFrameSlot(i, Memory::uint32_at(tos + i));
  }
}


void Deoptimizer::SetPlatformCompiledStubRegisters(
    FrameDescription* output_frame, CodeStubInterfaceDescriptor* descriptor) {
  intptr_t handler =
      reinterpret_cast<intptr_t>(descriptor->deoptimization_handler_);
  int params = descriptor->register_param_count_;
  if (descriptor->stack_parameter_count_ != NULL) {
    params++;
  }
  output_frame->SetRegister(eax.code(), params);
  output_frame->SetRegister(ebx.code(), handler);
}


void Deoptimizer::CopyDoubleRegisters(FrameDescription* output_frame) {
  for (int i = 0; i < XMMRegister::kNumAllocatableRegisters; ++i) {
    double double_value = input_->GetDoubleRegister(i);
    output_frame->SetDoubleRegister(i, double_value);
  }
}


bool Deoptimizer::HasAlignmentPadding(JSFunction* function) {
  int parameter_count = function->shared()->formal_parameter_count() + 1;
  unsigned input_frame_size = input_->GetFrameSize();
  unsigned alignment_state_offset =
      input_frame_size - parameter_count * kPointerSize -
      StandardFrameConstants::kFixedFrameSize -
      kPointerSize;
  ASSERT(JavaScriptFrameConstants::kDynamicAlignmentStateOffset ==
      JavaScriptFrameConstants::kLocal0Offset);
  int32_t alignment_state = input_->GetFrameSlot(alignment_state_offset);
  return (alignment_state == kAlignmentPaddingPushed);
}


#define __ masm()->

void Deoptimizer::EntryGenerator::Generate() {
  GeneratePrologue();

  // Save all general purpose registers before messing with them.
  const int kNumberOfRegisters = Register::kNumRegisters;

  const int kDoubleRegsSize = kDoubleSize *
                              XMMRegister::kNumAllocatableRegisters;
  __ sub(esp, Immediate(kDoubleRegsSize));
  if (CpuFeatures::IsSupported(SSE2)) {
    CpuFeatureScope scope(masm(), SSE2);
    for (int i = 0; i < XMMRegister::kNumAllocatableRegisters; ++i) {
      XMMRegister xmm_reg = XMMRegister::FromAllocationIndex(i);
      int offset = i * kDoubleSize;
      __ movdbl(Operand(esp, offset), xmm_reg);
    }
  }

  __ pushad();

  const int kSavedRegistersAreaSize = kNumberOfRegisters * kPointerSize +
                                      kDoubleRegsSize;

  // Get the bailout id from the stack.
  __ mov(ebx, Operand(esp, kSavedRegistersAreaSize));

  // Get the address of the location in the code object
  // and compute the fp-to-sp delta in register edx.
  __ mov(ecx, Operand(esp, kSavedRegistersAreaSize + 1 * kPointerSize));
  __ lea(edx, Operand(esp, kSavedRegistersAreaSize + 2 * kPointerSize));

  __ sub(edx, ebp);
  __ neg(edx);

  // Allocate a new deoptimizer object.
  __ PrepareCallCFunction(6, eax);
  __ mov(eax, Operand(ebp, JavaScriptFrameConstants::kFunctionOffset));
  __ mov(Operand(esp, 0 * kPointerSize), eax);  // Function.
  __ mov(Operand(esp, 1 * kPointerSize), Immediate(type()));  // Bailout type.
  __ mov(Operand(esp, 2 * kPointerSize), ebx);  // Bailout id.
  __ mov(Operand(esp, 3 * kPointerSize), ecx);  // Code address or 0.
  __ mov(Operand(esp, 4 * kPointerSize), edx);  // Fp-to-sp delta.
  __ mov(Operand(esp, 5 * kPointerSize),
         Immediate(ExternalReference::isolate_address(isolate())));
  {
    AllowExternalCallThatCantCauseGC scope(masm());
    __ CallCFunction(ExternalReference::new_deoptimizer_function(isolate()), 6);
  }

  // Preserve deoptimizer object in register eax and get the input
  // frame descriptor pointer.
  __ mov(ebx, Operand(eax, Deoptimizer::input_offset()));

  // Fill in the input registers.
  for (int i = kNumberOfRegisters - 1; i >= 0; i--) {
    int offset = (i * kPointerSize) + FrameDescription::registers_offset();
    __ pop(Operand(ebx, offset));
  }

  int double_regs_offset = FrameDescription::double_registers_offset();
  if (CpuFeatures::IsSupported(SSE2)) {
    CpuFeatureScope scope(masm(), SSE2);
    // Fill in the double input registers.
    for (int i = 0; i < XMMRegister::kNumAllocatableRegisters; ++i) {
      int dst_offset = i * kDoubleSize + double_regs_offset;
      int src_offset = i * kDoubleSize;
      __ movdbl(xmm0, Operand(esp, src_offset));
      __ movdbl(Operand(ebx, dst_offset), xmm0);
    }
  }

  // Clear FPU all exceptions.
  // TODO(ulan): Find out why the TOP register is not zero here in some cases,
  // and check that the generated code never deoptimizes with unbalanced stack.
  __ fnclex();

  // Remove the bailout id, return address and the double registers.
  __ add(esp, Immediate(kDoubleRegsSize + 2 * kPointerSize));

  // Compute a pointer to the unwinding limit in register ecx; that is
  // the first stack slot not part of the input frame.
  __ mov(ecx, Operand(ebx, FrameDescription::frame_size_offset()));
  __ add(ecx, esp);

  // Unwind the stack down to - but not including - the unwinding
  // limit and copy the contents of the activation frame to the input
  // frame description.
  __ lea(edx, Operand(ebx, FrameDescription::frame_content_offset()));
  Label pop_loop_header;
  __ jmp(&pop_loop_header);
  Label pop_loop;
  __ bind(&pop_loop);
  __ pop(Operand(edx, 0));
  __ add(edx, Immediate(sizeof(uint32_t)));
  __ bind(&pop_loop_header);
  __ cmp(ecx, esp);
  __ j(not_equal, &pop_loop);

  // Compute the output frame in the deoptimizer.
  __ push(eax);
  __ PrepareCallCFunction(1, ebx);
  __ mov(Operand(esp, 0 * kPointerSize), eax);
  {
    AllowExternalCallThatCantCauseGC scope(masm());
    __ CallCFunction(
        ExternalReference::compute_output_frames_function(isolate()), 1);
  }
  __ pop(eax);

  if (type() != OSR) {
    // If frame was dynamically aligned, pop padding.
    Label no_padding;
    __ cmp(Operand(eax, Deoptimizer::has_alignment_padding_offset()),
           Immediate(0));
    __ j(equal, &no_padding);
    __ pop(ecx);
    if (FLAG_debug_code) {
      __ cmp(ecx, Immediate(kAlignmentZapValue));
      __ Assert(equal, kAlignmentMarkerExpected);
    }
    __ bind(&no_padding);
  } else {
    // If frame needs dynamic alignment push padding.
    Label no_padding;
    __ cmp(Operand(eax, Deoptimizer::has_alignment_padding_offset()),
           Immediate(0));
    __ j(equal, &no_padding);
    __ push(Immediate(kAlignmentZapValue));
    __ bind(&no_padding);
  }

  // Replace the current frame with the output frames.
  Label outer_push_loop, inner_push_loop,
      outer_loop_header, inner_loop_header;
  // Outer loop state: eax = current FrameDescription**, edx = one past the
  // last FrameDescription**.
  __ mov(edx, Operand(eax, Deoptimizer::output_count_offset()));
  __ mov(eax, Operand(eax, Deoptimizer::output_offset()));
  __ lea(edx, Operand(eax, edx, times_4, 0));
  __ jmp(&outer_loop_header);
  __ bind(&outer_push_loop);
  // Inner loop state: ebx = current FrameDescription*, ecx = loop index.
  __ mov(ebx, Operand(eax, 0));
  __ mov(ecx, Operand(ebx, FrameDescription::frame_size_offset()));
  __ jmp(&inner_loop_header);
  __ bind(&inner_push_loop);
  __ sub(ecx, Immediate(sizeof(uint32_t)));
  __ push(Operand(ebx, ecx, times_1, FrameDescription::frame_content_offset()));
  __ bind(&inner_loop_header);
  __ test(ecx, ecx);
  __ j(not_zero, &inner_push_loop);
  __ add(eax, Immediate(kPointerSize));
  __ bind(&outer_loop_header);
  __ cmp(eax, edx);
  __ j(below, &outer_push_loop);

  // In case of OSR or a failed STUB, we have to restore the XMM registers.
  if (CpuFeatures::IsSupported(SSE2)) {
    CpuFeatureScope scope(masm(), SSE2);
    for (int i = 0; i < XMMRegister::kNumAllocatableRegisters; ++i) {
      XMMRegister xmm_reg = XMMRegister::FromAllocationIndex(i);
      int src_offset = i * kDoubleSize + double_regs_offset;
      __ movdbl(xmm_reg, Operand(ebx, src_offset));
    }
  }

  // Push state, pc, and continuation from the last output frame.
  if (type() != OSR) {
    __ push(Operand(ebx, FrameDescription::state_offset()));
  }
  __ push(Operand(ebx, FrameDescription::pc_offset()));
  __ push(Operand(ebx, FrameDescription::continuation_offset()));


  // Push the registers from the last output frame.
  for (int i = 0; i < kNumberOfRegisters; i++) {
    int offset = (i * kPointerSize) + FrameDescription::registers_offset();
    __ push(Operand(ebx, offset));
  }

  // Restore the registers from the stack.
  __ popad();

  // Return to the continuation point.
  __ ret(0);
}


void Deoptimizer::TableEntryGenerator::GeneratePrologue() {
  // Create a sequence of deoptimization entries.
  Label done;
  for (int i = 0; i < count(); i++) {
    int start = masm()->pc_offset();
    USE(start);
    __ push_imm32(i);
    __ jmp(&done);
    ASSERT(masm()->pc_offset() - start == table_entry_size_);
  }
  __ bind(&done);
}


void FrameDescription::SetCallerPc(unsigned offset, intptr_t value) {
  SetFrameSlot(offset, value);
}


void FrameDescription::SetCallerFp(unsigned offset, intptr_t value) {
  SetFrameSlot(offset, value);
}


#undef __


} }  // namespace v8::internal

#endif  // V8_TARGET_ARCH_IA32