Commit 77fffe22 authored by ricow@chromium.org's avatar ricow@chromium.org

X64: Implement DoComputeOsrOutputFrame and fix Generate_OnStackReplacement.

The existing version does not have a correct implementation of
Generate_OnStackReplacement since we do not correctly find the loop
depth. This is mainly because EmitStackCheck in full-codegen-x64.cc
did not write it and partly due to us reading at the wrong offset
(which had no effect when it was not written in the first place, we
simply got a random number from the next instruction).

The DoComputeOsrOoutputFrame is very similar to the ia32 version.



Review URL: http://codereview.chromium.org/6581028

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@6931 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
parent bd2fb513
......@@ -135,11 +135,8 @@ DEFINE_bool(deoptimize_uncommon_cases, true, "deoptimize uncommon cases")
DEFINE_bool(polymorphic_inlining, true, "polymorphic inlining")
DEFINE_bool(aggressive_loop_invariant_motion, true,
"aggressive motion of instructions out of loops")
#ifdef V8_TARGET_ARCH_X64
DEFINE_bool(use_osr, false, "use on-stack replacement")
#else
DEFINE_bool(use_osr, true, "use on-stack replacement")
#endif
DEFINE_bool(trace_osr, false, "trace on-stack replacement")
DEFINE_int(stress_runs, 0, "number of stress runs")
DEFINE_bool(optimize_closures, true, "optimize closures")
......
......@@ -224,7 +224,7 @@ void Deoptimizer::PatchStackCheckCodeAt(Address pc_after,
// ok:
//
ASSERT(*(call_target_address - 3) == 0x73 && // jae
*(call_target_address - 2) == 0x05 && // offset
*(call_target_address - 2) == 0x07 && // offset
*(call_target_address - 1) == 0xe8); // call
*(call_target_address - 3) = 0x90; // nop
*(call_target_address - 2) = 0x90; // nop
......@@ -245,14 +245,154 @@ void Deoptimizer::RevertStackCheckCodeAt(Address pc_after,
*(call_target_address - 2) == 0x90 && // nop
*(call_target_address - 1) == 0xe8); // call
*(call_target_address - 3) = 0x73; // jae
*(call_target_address - 2) = 0x05; // offset
*(call_target_address - 2) = 0x07; // offset
Assembler::set_target_address_at(call_target_address,
check_code->entry());
}
static int LookupBailoutId(DeoptimizationInputData* data, unsigned ast_id) {
ByteArray* translations = data->TranslationByteArray();
int length = data->DeoptCount();
for (int i = 0; i < length; i++) {
if (static_cast<unsigned>(data->AstId(i)->value()) == 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() {
UNIMPLEMENTED();
DeoptimizationInputData* data = DeoptimizationInputData::cast(
optimized_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, 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();
ASSERT(count == 1);
USE(count);
opcode = static_cast<Translation::Opcode>(iterator.Next());
USE(opcode);
ASSERT(Translation::FRAME == opcode);
unsigned node_id = iterator.Next();
USE(node_id);
ASSERT(node_id == ast_id);
JSFunction* function = JSFunction::cast(ComputeLiteral(iterator.Next()));
USE(function);
ASSERT(function == function_);
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 = optimized_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_));
function_->PrintName();
PrintF(" => node=%u, frame=%d->%d]\n",
ast_id,
input_frame_size,
output_frame_size);
}
// 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_);
// 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 = 0; ok && i < 4; i++) {
intptr_t input_value = input_->GetFrameSlot(input_offset);
if (FLAG_trace_osr) {
PrintF(" [esp + %d] <- 0x%08" V8PRIxPTR " ; [esp + %d] (fixed part)\n",
output_offset,
input_value,
input_offset);
}
output_[0]->SetFrameSlot(output_offset, input_->GetFrameSlot(input_offset));
input_offset -= kPointerSize;
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<intptr_t>(from_));
} else {
// Setup the frame pointer and the context pointer.
output_[0]->SetRegister(rbp.code(), input_->GetRegister(rbp.code()));
output_[0]->SetRegister(rsi.code(), input_->GetRegister(rsi.code()));
unsigned pc_offset = data->OsrPcOffset()->value();
intptr_t pc = reinterpret_cast<intptr_t>(
optimized_code_->entry() + pc_offset);
output_[0]->SetPc(pc);
}
Code* continuation = Builtins::builtin(Builtins::NotifyOSR);
output_[0]->SetContinuation(
reinterpret_cast<intptr_t>(continuation->entry()));
if (FLAG_trace_osr) {
PrintF("[on-stack replacement translation %s: 0x%08" V8PRIxPTR " ",
ok ? "finished" : "aborted",
reinterpret_cast<intptr_t>(function));
function->PrintName();
PrintF(" => pc=0x%0" V8PRIxPTR "]\n", output_[0]->GetPc());
}
}
......
......@@ -269,6 +269,13 @@ void FullCodeGenerator::EmitStackCheck(IterationStatement* stmt) {
// the deoptimization input data found in the optimized code.
RecordStackCheck(stmt->OsrEntryId());
// Loop stack checks can be patched to perform on-stack replacement. In
// order to decide whether or not to perform OSR we embed the loop depth
// in a test instruction after the call so we can extract it from the OSR
// builtin.
ASSERT(loop_depth() > 0);
__ testl(rax, Immediate(Min(loop_depth(), Code::kMaxLoopNestingMarker)));
__ bind(&ok);
PrepareForBailoutForId(stmt->EntryId(), NO_REGISTERS);
// Record a mapping of the OSR id to this PC. This is used if the OSR
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment