AuroraMiddleware/v8
1
0
Fork 0
Code Issues 2 Pull Requests Releases Wiki Activity

ARM: Implement OSR infrastructure.

Browse Source 
Review URL: http://codereview.chromium.org/6460034

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@6758 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
This commit is contained in:
karlklose@chromium.org 2011-02-13 10:24:39 +00:00
parent 3f4701df7f
commit dd68fcc7ab
4 changed files with 201 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

40
src/arm/builtins-arm.cc
View File

@ -1156,12 +1156,48 @@ void Builtins::Generate_NotifyLazyDeoptimized(MacroAssembler* masm) {
void Builtins::Generate_NotifyOSR(MacroAssembler* masm) {
__ stop("builtins-arm.cc: NotifyOSR");
// For now, we are relying on the fact that Runtime::NotifyOSR
// doesn't do any garbage collection which allows us to save/restore
// the registers without worrying about which of them contain
// pointers. This seems a bit fragile.
__ stm(db_w, sp, kJSCallerSaved | kCalleeSaved | lr.bit() | fp.bit());
__ EnterInternalFrame();
__ CallRuntime(Runtime::kNotifyOSR, 0);
__ LeaveInternalFrame();
__ ldm(ia_w, sp, kJSCallerSaved | kCalleeSaved | lr.bit() | fp.bit());
__ Ret();
}
void Builtins::Generate_OnStackReplacement(MacroAssembler* masm) {
__ stop("builtins-arm.cc: OnStackReplacement");
// Probe the CPU to set the supported features, because this builtin
// may be called before the initialization performs CPU setup.
CpuFeatures::Probe(false);
// Lookup the function in the JavaScript frame and push it as an
// argument to the on-stack replacement function.
__ ldr(r0, MemOperand(fp, JavaScriptFrameConstants::kFunctionOffset));
__ EnterInternalFrame();
__ push(r0);
__ CallRuntime(Runtime::kCompileForOnStackReplacement, 1);
__ LeaveInternalFrame();
// If the result was -1 it means that we couldn't optimize the
// function. Just return and continue in the unoptimized version.
Label skip;
__ cmp(r0, Operand(Smi::FromInt(-1)));
__ b(ne, &skip);
__ Ret();
__ bind(&skip);
// Untag the AST id and push it on the stack.
__ SmiUntag(r0);
__ push(r0);
// Generate the code for doing the frame-to-frame translation using
// the deoptimizer infrastructure.
Deoptimizer::EntryGenerator generator(masm, Deoptimizer::OSR);
generator.Generate();
}

152
src/arm/deoptimizer-arm.cc
View File

@ -135,8 +135,145 @@ void Deoptimizer::RevertStackCheckCodeAt(Address pc_after,
}
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();
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++) {
uint32_t input_value = input_->GetFrameSlot(input_offset);
if (FLAG_trace_osr) {
PrintF(" [sp + %d] <- 0x%08x ; [sp + %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<uint32_t>(from_));
} else {
// Setup the frame pointer and the context pointer.
output_[0]->SetRegister(fp.code(), input_->GetRegister(fp.code()));
output_[0]->SetRegister(cp.code(), input_->GetRegister(cp.code()));
unsigned pc_offset = data->OsrPcOffset()->value();
uint32_t pc = reinterpret_cast<uint32_t>(
optimized_code_->entry() + pc_offset);
output_[0]->SetPc(pc);
}
Code* continuation = Builtins::builtin(Builtins::NotifyOSR);
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));
function->PrintName();
PrintF(" => pc=0x%0x]\n", output_[0]->GetPc());
}
}
@ -318,7 +455,6 @@ void Deoptimizer::DoComputeFrame(TranslationIterator* iterator,
// easily ported.
void Deoptimizer::EntryGenerator::Generate() {
GeneratePrologue();
// TOS: bailout-id; TOS+1: return address if not EAGER.
CpuFeatures::Scope scope(VFP3);
// Save all general purpose registers before messing with them.
const int kNumberOfRegisters = Register::kNumRegisters;
@ -353,6 +489,10 @@ void Deoptimizer::EntryGenerator::Generate() {
__ mov(r3, Operand(0));
// Correct one word for bailout id.
__ add(r4, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
} else if (type() == OSR) {
__ mov(r3, lr);
// Correct one word for bailout id.
__ add(r4, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
} else {
__ mov(r3, lr);
// Correct two words for bailout id and return address.
@ -375,7 +515,6 @@ void Deoptimizer::EntryGenerator::Generate() {
// frame descriptor pointer to r1 (deoptimizer->input_);
__ ldr(r1, MemOperand(r0, Deoptimizer::input_offset()));
// Copy core registers into FrameDescription::registers_[kNumRegisters].
ASSERT(Register::kNumRegisters == kNumberOfRegisters);
for (int i = 0; i < kNumberOfRegisters; i++) {
@ -396,7 +535,7 @@ void Deoptimizer::EntryGenerator::Generate() {
// Remove the bailout id, eventually return address, and the saved registers
// from the stack.
if (type() == EAGER) {
if (type() == EAGER || type() == OSR) {
__ add(sp, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
} else {
__ add(sp, sp, Operand(kSavedRegistersAreaSize + (2 * kPointerSize)));
@ -450,11 +589,6 @@ void Deoptimizer::EntryGenerator::Generate() {
__ cmp(r0, r1);
__ b(lt, &outer_push_loop);
// In case of OSR, we have to restore the XMM registers.
if (type() == OSR) {
UNIMPLEMENTED();
}
// Push state, pc, and continuation from the last output frame.
if (type() != OSR) {
__ ldr(r6, MemOperand(r2, FrameDescription::state_offset()));

14
src/arm/lithium-codegen-arm.cc
View File

@ -3912,7 +3912,19 @@ void LCodeGen::DoStackCheck(LStackCheck* instr) {
void LCodeGen::DoOsrEntry(LOsrEntry* instr) {
Abort("DoOsrEntry unimplemented.");
// This is a pseudo-instruction that ensures that the environment here is
// properly registered for deoptimization and records the assembler's PC
// offset.
LEnvironment* environment = instr->environment();
environment->SetSpilledRegisters(instr->SpilledRegisterArray(),
instr->SpilledDoubleRegisterArray());
// If the environment were already registered, we would have no way of
// backpatching it with the spill slot operands.
ASSERT(!environment->HasBeenRegistered());
RegisterEnvironmentForDeoptimization(environment);
ASSERT(osr_pc_offset_ == -1);
osr_pc_offset_ = masm()->pc_offset();
}

14
src/deoptimizer.cc
View File

@ -663,7 +663,7 @@ bool Deoptimizer::DoOsrTranslateCommand(TranslationIterator* iterator,
case Translation::REGISTER: {
int output_reg = iterator->Next();
if (FLAG_trace_osr) {
PrintF(" %s <- 0x%08" V8PRIxPTR " ; [esp + %d]\n",
PrintF(" %s <- 0x%08" V8PRIxPTR " ; [sp + %d]\n",
converter.NameOfCPURegister(output_reg),
input_value,
*input_offset);
@ -690,7 +690,7 @@ bool Deoptimizer::DoOsrTranslateCommand(TranslationIterator* iterator,
return false;
}
if (FLAG_trace_osr) {
PrintF(" %s <- %d (int32) ; [esp + %d]\n",
PrintF(" %s <- %d (int32) ; [sp + %d]\n",
converter.NameOfCPURegister(output_reg),
int32_value,
*input_offset);
@ -706,7 +706,7 @@ bool Deoptimizer::DoOsrTranslateCommand(TranslationIterator* iterator,
int output_reg = iterator->Next();
double double_value = input_object->Number();
if (FLAG_trace_osr) {
PrintF(" %s <- %g (double) ; [esp + %d]\n",
PrintF(" %s <- %g (double) ; [sp + %d]\n",
DoubleRegister::AllocationIndexToString(output_reg),
double_value,
*input_offset);
@ -720,7 +720,7 @@ bool Deoptimizer::DoOsrTranslateCommand(TranslationIterator* iterator,
unsigned output_offset =
output->GetOffsetFromSlotIndex(this, output_index);
if (FLAG_trace_osr) {
PrintF(" [esp + %d] <- 0x%08" V8PRIxPTR " ; [esp + %d]\n",
PrintF(" [sp + %d] <- 0x%08" V8PRIxPTR " ; [sp + %d]\n",
output_offset,
input_value,
*input_offset);
@ -749,7 +749,7 @@ bool Deoptimizer::DoOsrTranslateCommand(TranslationIterator* iterator,
return false;
}
if (FLAG_trace_osr) {
PrintF(" [esp + %d] <- %d (int32) ; [esp + %d]\n",
PrintF(" [sp + %d] <- %d (int32) ; [sp + %d]\n",
output_offset,
int32_value,
*input_offset);
@ -773,12 +773,12 @@ bool Deoptimizer::DoOsrTranslateCommand(TranslationIterator* iterator,
int32_t lower = static_cast<int32_t>(int_value);
int32_t upper = static_cast<int32_t>(int_value >> kBitsPerInt);
if (FLAG_trace_osr) {
PrintF(" [esp + %d] <- 0x%08x (upper bits of %g) ; [esp + %d]\n",
PrintF(" [sp + %d] <- 0x%08x (upper bits of %g) ; [sp + %d]\n",
output_offset + kUpperOffset,
upper,
double_value,
*input_offset);
PrintF(" [esp + %d] <- 0x%08x (lower bits of %g) ; [esp + %d]\n",
PrintF(" [sp + %d] <- 0x%08x (lower bits of %g) ; [sp + %d]\n",
output_offset + kLowerOffset,
lower,
double_value,