Add support for deoptimization on x64. ...

src/x64/deoptimizer-x64.cc

Index: src/x64/deoptimizer-x64.cc
===================================================================
--- src/x64/deoptimizer-x64.cc	(revision 6418)
+++ src/x64/deoptimizer-x64.cc	(working copy)
@@ -41,8 +41,68 @@
 int Deoptimizer::table_entry_size_ = 10;
 
 void Deoptimizer::DeoptimizeFunction(JSFunction* function) {
-  // UNIMPLEMENTED, for now just return.
-  return;
+  AssertNoAllocation no_allocation;
+
+  if (!function->IsOptimized()) return;
+
+  // Get the optimized code.
+  Code* code = function->code();
+
+  // Invalidate the relocation information, as it will become invalid by the
+  // code patching below, and is not needed any more.
+  code->InvalidateRelocation();
+
+  // For each return after a safepoint insert a absolute call to the
+  // corresponding deoptimization entry.
+  unsigned last_pc_offset = 0;
+  SafepointTable table(function->code());
+  for (unsigned i = 0; i < table.length(); i++) {
+    unsigned pc_offset = table.GetPcOffset(i);
+    SafepointEntry safepoint_entry = table.GetEntry(i);
+    int deoptimization_index = safepoint_entry.deoptimization_index();
+    int gap_code_size = safepoint_entry.gap_code_size();
+#ifdef DEBUG
+    // Destroy the code which is not supposed to run again.
+    unsigned instructions = pc_offset - last_pc_offset;
+    CodePatcher destroyer(code->instruction_start() + last_pc_offset,
+                          instructions);
+    for (unsigned i = 0; i < instructions; i++) {
+      destroyer.masm()->int3();
+    }
+#endif
+    last_pc_offset = pc_offset;
+    if (deoptimization_index != Safepoint::kNoDeoptimizationIndex) {
+      CodePatcher patcher(
+          code->instruction_start() + pc_offset + gap_code_size,
+          Assembler::kCallInstructionLength);
+      patcher.masm()->Call(GetDeoptimizationEntry(deoptimization_index, LAZY),
+                           RelocInfo::NONE);
+      last_pc_offset += gap_code_size + Assembler::kCallInstructionLength;
+    }
+  }
+#ifdef DEBUG
+  // Destroy the code which is not supposed to run again.
+  unsigned instructions = code->safepoint_table_start() - last_pc_offset;
+  CodePatcher destroyer(code->instruction_start() + last_pc_offset,
+                        instructions);
+  for (unsigned i = 0; i < instructions; i++) {
+    destroyer.masm()->int3();
+  }
+#endif
+
+  // Add the deoptimizing code to the list.
+  DeoptimizingCodeListNode* node = new DeoptimizingCodeListNode(code);
+  node->set_next(deoptimizing_code_list_);
+  deoptimizing_code_list_ = node;
+
+  // Set the code for the function to non-optimized version.
+  function->ReplaceCode(function->shared()->code());
+
+  if (FLAG_trace_deopt) {
+    PrintF("[forced deoptimization: ");
+    function->PrintName();
+    PrintF(" / %" V8PRIxPTR "]\n", reinterpret_cast<intptr_t>(function));
+  }
 }
 
 
@@ -64,20 +124,387 @@
 
 void Deoptimizer::DoComputeFrame(TranslationIterator* iterator,
                                  int frame_index) {
-  UNIMPLEMENTED();
+  // Read the ast node id, function, and frame height for this output frame.
+  Translation::Opcode opcode =
+      static_cast<Translation::Opcode>(iterator->Next());
+  USE(opcode);
+  ASSERT(Translation::FRAME == opcode);
+  int node_id = iterator->Next();
+  JSFunction* function = JSFunction::cast(ComputeLiteral(iterator->Next()));
+  unsigned height = iterator->Next();
+  unsigned height_in_bytes = height * kPointerSize;
+  if (FLAG_trace_deopt) {
+    PrintF("  translating ");
+    function->PrintName();
+    PrintF(" => node=%d, height=%d\n", node_id, height_in_bytes);
+  }
+
+  // The 'fixed' part of the frame consists of the incoming parameters and
+  // the part described by JavaScriptFrameConstants.
+  unsigned fixed_frame_size = ComputeFixedSize(function);
+  unsigned input_frame_size = input_->GetFrameSize();
+  unsigned output_frame_size = height_in_bytes + fixed_frame_size;
+
+  // Allocate and store the output frame description.
+  FrameDescription* output_frame =
+      new(output_frame_size) FrameDescription(output_frame_size, function);
+
+  bool is_bottommost = (0 == frame_index);
+  bool is_topmost = (output_count_ - 1 == frame_index);
+  ASSERT(frame_index >= 0 && frame_index < output_count_);
+  ASSERT(output_[frame_index] == NULL);
+  output_[frame_index] = output_frame;
+
+  // The top address for the bottommost output frame can be computed from
+  // the input frame pointer and the output frame's height.  For all
+  // subsequent output frames, it can be computed from the previous one's
+  // top address and the current frame's size.
+  intptr_t top_address;
+  if (is_bottommost) {
+    // 2 = context and function in the frame.
+    top_address =
+        input_->GetRegister(rbp.code()) - (2 * kPointerSize) - height_in_bytes;
+  } else {
+    top_address = output_[frame_index - 1]->GetTop() - output_frame_size;
+  }
+  output_frame->SetTop(top_address);
+
+  // Compute the incoming parameter translation.
+  int parameter_count = function->shared()->formal_parameter_count() + 1;
+  unsigned output_offset = output_frame_size;
+  unsigned input_offset = input_frame_size;
+  for (int i = 0; i < parameter_count; ++i) {
+    output_offset -= kPointerSize;
+    DoTranslateCommand(iterator, frame_index, output_offset);
+  }
+  input_offset -= (parameter_count * kPointerSize);
+
+  // There are no translation commands for the caller's pc and fp, the
+  // context, and the function.  Synthesize their values and set them up
+  // explicitly.
+  //
+  // The caller's pc for the bottommost output frame is the same as in the
+  // input frame.  For all subsequent output frames, it can be read from the
+  // previous one.  This frame's pc can be computed from the non-optimized
+  // function code and AST id of the bailout.
+  output_offset -= kPointerSize;
+  input_offset -= kPointerSize;
+  intptr_t value;
+  if (is_bottommost) {
+    value = input_->GetFrameSlot(input_offset);
+  } else {
+    value = output_[frame_index - 1]->GetPc();
+  }
+  output_frame->SetFrameSlot(output_offset, value);
+  if (FLAG_trace_deopt) {
+    PrintF("    0x%08" V8PRIxPTR ": [top + %d] <- 0x%08"
+           V8PRIxPTR  " ; caller's pc\n",
+           top_address + output_offset, output_offset, value);
+  }
+
+  // The caller's frame pointer for the bottommost output frame is the same
+  // as in the input frame.  For all subsequent output frames, it can be
+  // read from the previous one.  Also compute and set this frame's frame
+  // pointer.
+  output_offset -= kPointerSize;
+  input_offset -= kPointerSize;
+  if (is_bottommost) {
+    value = input_->GetFrameSlot(input_offset);
+  } else {
+    value = output_[frame_index - 1]->GetFp();
+  }
+  output_frame->SetFrameSlot(output_offset, value);
+  intptr_t fp_value = top_address + output_offset;
+  ASSERT(!is_bottommost || input_->GetRegister(rbp.code()) == fp_value);
+  output_frame->SetFp(fp_value);
+  if (is_topmost) output_frame->SetRegister(rbp.code(), fp_value);
+  if (FLAG_trace_deopt) {
+    PrintF("    0x%08" V8PRIxPTR ": [top + %d] <- 0x%08"
+           V8PRIxPTR " ; caller's fp\n",
+           fp_value, output_offset, value);
+  }
+
+  // The context can be gotten from the function so long as we don't
+  // optimize functions that need local contexts.
+  output_offset -= kPointerSize;
+  input_offset -= kPointerSize;
+  value = reinterpret_cast<intptr_t>(function->context());
+  // The context for the bottommost output frame should also agree with the
+  // input frame.
+  ASSERT(!is_bottommost || input_->GetFrameSlot(input_offset) == value);
+  output_frame->SetFrameSlot(output_offset, value);
+  if (is_topmost) output_frame->SetRegister(rsi.code(), value);
+  if (FLAG_trace_deopt) {
+    PrintF("    0x%08" V8PRIxPTR ": [top + %d] <- 0x%08"
+           V8PRIxPTR "; context\n",
+           top_address + output_offset, output_offset, value);
+  }
+
+  // The function was mentioned explicitly in the BEGIN_FRAME.
+  output_offset -= kPointerSize;
+  input_offset -= kPointerSize;
+  value = reinterpret_cast<intptr_t>(function);
+  // The function for the bottommost output frame should also agree with the
+  // input frame.
+  ASSERT(!is_bottommost || input_->GetFrameSlot(input_offset) == value);
+  output_frame->SetFrameSlot(output_offset, value);
+  if (FLAG_trace_deopt) {
+    PrintF("    0x%08" V8PRIxPTR ": [top + %d] <- 0x%08"
+           V8PRIxPTR "; function\n",
+           top_address + output_offset, output_offset, value);
+  }
+
+  // Translate the rest of the frame.
+  for (unsigned i = 0; i < height; ++i) {
+    output_offset -= kPointerSize;
+    DoTranslateCommand(iterator, frame_index, output_offset);
+  }
+  ASSERT(0 == output_offset);
+
+  // Compute this frame's PC, state, and continuation.
+  Code* non_optimized_code = function->shared()->code();
+  FixedArray* raw_data = non_optimized_code->deoptimization_data();
+  DeoptimizationOutputData* data = DeoptimizationOutputData::cast(raw_data);
+  Address start = non_optimized_code->instruction_start();
+  unsigned pc_and_state = GetOutputInfo(data, node_id, function->shared());
+  unsigned pc_offset = FullCodeGenerator::PcField::decode(pc_and_state);
+  intptr_t pc_value = reinterpret_cast<intptr_t>(start + pc_offset);
+  output_frame->SetPc(pc_value);
+
+  FullCodeGenerator::State state =
+      FullCodeGenerator::StateField::decode(pc_and_state);
+  output_frame->SetState(Smi::FromInt(state));
+
+  // Set the continuation for the topmost frame.
+  if (is_topmost) {
+    Code* continuation = (bailout_type_ == EAGER)
+        ? Builtins::builtin(Builtins::NotifyDeoptimized)
+        : Builtins::builtin(Builtins::NotifyLazyDeoptimized);
+    output_frame->SetContinuation(
+        reinterpret_cast<intptr_t>(continuation->entry()));
+  }
+
+  if (output_count_ - 1 == frame_index) iterator->Done();
 }
 
 
+#define __ masm()->
+
 void Deoptimizer::EntryGenerator::Generate() {
-  // UNIMPLEMENTED, for now just return.
-  return;
+  GeneratePrologue();
+  CpuFeatures::Scope scope(SSE2);
+
+  // Save all general purpose registers before messing with them.
+  const int kNumberOfRegisters = Register::kNumRegisters;
+
+  const int kDoubleRegsSize = kDoubleSize *
+                              XMMRegister::kNumAllocatableRegisters;
+  __ subq(rsp, Immediate(kDoubleRegsSize));
+
+  for (int i = 0; i < XMMRegister::kNumAllocatableRegisters; ++i) {
+    XMMRegister xmm_reg = XMMRegister::FromAllocationIndex(i);
+    int offset = i * kDoubleSize;
+    __ movsd(Operand(rsp, offset), xmm_reg);
+  }
+
+  // We push all registers onto the stack, even though we do not need
+  // to restore all later.
+  for (int i = 0; i < kNumberOfRegisters; i++) {
+    Register r = Register::toRegister(i);
+    __ push(r);
+  }
+
+  const int kSavedRegistersAreaSize = kNumberOfRegisters * kPointerSize +
+                                      kDoubleRegsSize;
+
+  // When calling new_deoptimizer_function we need to pass the last argument
+  // on the stack on windows and in r8 on windows. The remaining arguments are

[Kevin Millikin (Chromium), 2011/01/24 13:52:21]

Comment mentions "on windows" and "on windows".

[Rico, 2011/01/25 07:48:16]

Done.

+  // all passed in registers (different ones on linux and windows though).
+
+#ifdef _WIN64
+  Register arg4 = r9;
+  Register arg3 = r8;
+  Register arg2 = rdx;
+  Register arg1 = rcx;
+#else
+  Register arg4 = rcx;
+  Register arg3 = rdx;
+  Register arg2 = rsi;
+  Register arg1 = rdi;
+#endif
+
+  // We use this to keep the value of the fifth argument temporarily.
+  // Unfortunately we can't store it directly in r8 (used for passing
+  // this on linux), since it is another parameter passing register on windows.
+  Register arg5 = r11;
+
+  // Get the bailout id from the stack.
+  __ movq(arg3, Operand(rsp, kSavedRegistersAreaSize));
+
+  // Get the address of the location in the code object if possible
+  // and compute the fp-to-sp delta in register arg5.
+  if (type() == EAGER) {
+    __ Set(arg4, 0);
+    __ lea(arg5, Operand(rsp, kSavedRegistersAreaSize + 1 * kPointerSize));
+  } else {
+    __ movq(arg4, Operand(rsp, kSavedRegistersAreaSize + 1 * kPointerSize));
+    __ lea(arg5, Operand(rsp, kSavedRegistersAreaSize + 2 * kPointerSize));
+  }
+
+  __ subq(arg5, rbp);
+  __ neg(arg5);
+
+  // Allocate a new deoptimizer object.
+  __ PrepareCallCFunction(5);
+  __ movq(rax, Operand(rbp, JavaScriptFrameConstants::kFunctionOffset));
+  __ movq(arg1, rax);
+  __ movq(arg2, Immediate(type()));
+  // Args 3 and 4 are already in the right registers.
+
+  // On windows put the argument on the stack (PrepareCallCFunction have
+  // created space for this). On linux pass the argument in r8
+#ifdef _WIN64
+  __ movq(Operand(rsp, 0 * kPointerSize), arg5);
+#else
+  __ movq(r8, arg5);
+#endif
+
+  __ CallCFunction(ExternalReference::new_deoptimizer_function(), 5);
+    // Preserve deoptimizer object in register rax and get the input

[Lasse Reichstein, 2011/01/24 13:24:21]

Indentation.

[Rico, 2011/01/24 13:40:14]

Done.

+  // frame descriptor pointer.
+  __ movq(rbx, Operand(rax, Deoptimizer::input_offset()));
+
+  // Fill in the input registers.
+  for (int i = 0; i < kNumberOfRegisters; i++) {
+    int offset = (i * kPointerSize) + FrameDescription::registers_offset();
+    __ movq(rcx, Operand(rsp, (kNumberOfRegisters - 1 - i) * kPointerSize));

[Kevin Millikin (Chromium), 2011/01/24 13:52:21]

I wonder why we don't count i down from kNumberOfRegisters to zero and

__ pop(Operand(rbx, offset))

for all of them?

[Rico, 2011/01/25 07:48:16]

Done.

+    __ movq(Operand(rbx, offset), rcx);
+  }
+
+    // Fill in the double input registers.
+  int double_regs_offset = FrameDescription::double_registers_offset();
+  for (int i = 0; i < XMMRegister::kNumAllocatableRegisters; ++i) {
+    int dst_offset = i * kDoubleSize + double_regs_offset;
+    int src_offset = i * kDoubleSize + kNumberOfRegisters * kPointerSize;

[Lasse Reichstein, 2011/01/24 13:24:21]

If the values are at the end of the stack, you could pop them directly to a
memory location.

[Rico, 2011/01/24 13:40:14]

They are not at the end, the general purpose registers are on top of these. In
addition, this makes the code very similar to the ia32 version.

[Kevin Millikin (Chromium), 2011/01/24 13:52:21]

Agree with lrn.

[Rico, 2011/01/25 07:48:16]

Done.

+    __ movsd(xmm0, Operand(rsp, src_offset));
+    __ movsd(Operand(rbx, dst_offset), xmm0);
+  }
+
+  // Remove the bailout id and the general purpose registers from the stack.
+  if (type() == EAGER) {
+    __ addq(rsp, Immediate(kSavedRegistersAreaSize + kPointerSize));
+  } else {
+    __ addq(rsp, Immediate(kSavedRegistersAreaSize + 2 * kPointerSize));
+  }
+
+  // Compute a pointer to the unwinding limit in register ecx; that is
+  // the first stack slot not part of the input frame.
+  __ movq(rcx, Operand(rbx, FrameDescription::frame_size_offset()));
+  __ addq(rcx, rsp);
+
+  // 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(rdx, Operand(rbx, FrameDescription::frame_content_offset()));
+  Label pop_loop;
+  __ bind(&pop_loop);
+  __ pop(Operand(rdx, 0));
+  __ addq(rdx, Immediate(sizeof(intptr_t)));
+  __ cmpq(rcx, rsp);
+  __ j(not_equal, &pop_loop);
+
+  // Compute the output frame in the deoptimizer.
+  __ push(rax);
+  __ PrepareCallCFunction(1);
+  __ movq(arg1, rax);
+  __ CallCFunction(ExternalReference::compute_output_frames_function(), 1);
+  __ pop(rax);
+
+  // Replace the current frame with the output frames.
+  Label outer_push_loop, inner_push_loop;
+  // Outer loop state: rax = current FrameDescription**, rdx = one past the
+  // last FrameDescription**.
+  __ movl(rdx, Operand(rax, Deoptimizer::output_count_offset()));
+  __ movq(rax, Operand(rax, Deoptimizer::output_offset()));
+  __ lea(rdx, Operand(rax, rdx, times_8, 0));
+  __ bind(&outer_push_loop);
+  // Inner loop state: rbx = current FrameDescription*, rcx = loop index.
+  __ movq(rbx, Operand(rax, 0));
+  __ movq(rcx, Operand(rbx, FrameDescription::frame_size_offset()));
+  __ bind(&inner_push_loop);
+  __ subq(rcx, Immediate(sizeof(intptr_t)));
+  __ push(Operand(rbx, rcx, times_1, FrameDescription::frame_content_offset()));
+  __ testq(rcx, rcx);
+  __ j(not_zero, &inner_push_loop);
+  __ addq(rax, Immediate(kPointerSize));
+  __ cmpq(rax, rdx);
+  __ j(below, &outer_push_loop);
+
+  // In case of OSR, we have to restore the XMM registers.
+  if (type() == OSR) {
+    for (int i = 0; i < XMMRegister::kNumAllocatableRegisters; ++i) {
+      XMMRegister xmm_reg = XMMRegister::FromAllocationIndex(i);
+      int src_offset = i * kDoubleSize + double_regs_offset;
+      __ movsd(xmm_reg, Operand(rbx, src_offset));
+    }
+  }
+
+  // Push state, pc, and continuation from the last output frame.
+  if (type() != OSR) {
+    __ push(Operand(rbx, FrameDescription::state_offset()));
+  }
+  __ push(Operand(rbx, FrameDescription::pc_offset()));
+  __ push(Operand(rbx, 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(rbx, offset));
+  }
+
+  // Restore the registers from the stack.
+  for (int i = kNumberOfRegisters - 1; i >= 0 ; i--) {
+    Register r = Register::toRegister(i);
+    // Do not restore rsp, simply pop the value into the next register
+    // and overwrite this immediately.
+    if (r.is(rsp)) {
+      ASSERT(i > 0);

[Lasse Reichstein, 2011/01/24 13:24:21]

Why not just 
 if (r.is(rsp)) {
   ASSERT(i > 0);
   r = Register::toRegister(i - 1);
 }
instead of unrolling the loop once.

[Rico, 2011/01/24 13:40:14]

Done.

+      i--;
+      r = Register::toRegister(i);
+      __ pop(r);
+    }
+    __ pop(r);
+  }
+
+  // Set up the roots register.
+  ExternalReference roots_address = ExternalReference::roots_address();
+  __ movq(r13, roots_address);
+
+  __ movq(kSmiConstantRegister,
+         reinterpret_cast<uint64_t>(Smi::FromInt(kSmiConstantRegisterValue)),

[Lasse Reichstein, 2011/01/24 13:24:21]

Indentation.

[Rico, 2011/01/24 13:40:14]

Done.

+         RelocInfo::NONE);
+
+  // Return to the continuation point.
+  __ ret(0);
 }
 
 
 void Deoptimizer::TableEntryGenerator::GeneratePrologue() {
-  UNIMPLEMENTED();
+  // 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);
 }
 
+#undef __
+
+
 } }  // namespace v8::internal
 
 #endif  // V8_TARGET_ARCH_X64