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

MIPS: crankshaft implementation

Browse Source 
BUG=
TEST=

Review URL: http://codereview.chromium.org/7934002
Patch from Paul Lind <plind44@gmail.com>.

git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@9828 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
This commit is contained in:
kmillikin@chromium.org 2011-10-28 08:14:46 +00:00
parent b5b3345799
commit f077a41b12
11 changed files with 10530 additions and 225 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
src/SConscript
View File

@ -172,6 +172,9 @@ SOURCES = {
mips/frames-mips.cc
mips/full-codegen-mips.cc
mips/ic-mips.cc
mips/lithium-codegen-mips.cc
mips/lithium-gap-resolver-mips.cc
mips/lithium-mips.cc
mips/macro-assembler-mips.cc
mips/regexp-macro-assembler-mips.cc
mips/stub-cache-mips.cc

15
src/deoptimizer.h
View File

@ -369,7 +369,20 @@ class FrameDescription {
}
double GetDoubleFrameSlot(unsigned offset) {
return *reinterpret_cast<double*>(GetFrameSlotPointer(offset));
intptr_t* ptr = GetFrameSlotPointer(offset);
#if V8_TARGET_ARCH_MIPS
// Prevent gcc from using load-double (mips ldc1) on (possibly)
// non-64-bit aligned double. Uses two lwc1 instructions.
union conversion {
double d;
uint32_t u[2];
} c;
c.u[0] = *reinterpret_cast<uint32_t*>(ptr);
c.u[1] = *(reinterpret_cast<uint32_t*>(ptr) + 1);
return c.d;
#else
return *reinterpret_cast<double*>(ptr);
#endif
}
void SetFrameSlot(unsigned offset, intptr_t value) {

9
src/flag-definitions.h
View File

@ -114,11 +114,7 @@ DEFINE_bool(clever_optimizations,
"Optimize object size, Array shift, DOM strings and string +")
// Flags for Crankshaft.
#ifdef V8_TARGET_ARCH_MIPS
DEFINE_bool(crankshaft, false, "use crankshaft")
#else
DEFINE_bool(crankshaft, true, "use crankshaft")
#endif
DEFINE_bool(crankshaft, true, "use crankshaft")
DEFINE_string(hydrogen_filter, "", "hydrogen use/trace filter")
DEFINE_bool(use_hydrogen, true, "use generated hydrogen for compilation")
DEFINE_bool(build_lithium, true, "use lithium chunk builder")
@ -326,7 +322,8 @@ DEFINE_bool(strict_mode, true, "allow strict mode directives")
// simulator-arm.cc and simulator-mips.cc
DEFINE_bool(trace_sim, false, "Trace simulator execution")
DEFINE_bool(check_icache, false, "Check icache flushes in ARM simulator")
DEFINE_bool(check_icache, false,
"Check icache flushes in ARM and MIPS simulator")
DEFINE_int(stop_sim_at, 0, "Simulator stop after x number of instructions")
DEFINE_int(sim_stack_alignment, 8,
"Stack alingment in bytes in simulator (4 or 8, 8 is default)")

85
src/mips/builtins-mips.cc
View File

@ -1176,24 +1176,93 @@ void Builtins::Generate_LazyRecompile(MacroAssembler* masm) {
}
// These functions are called from C++ but cannot be used in live code.
static void Generate_NotifyDeoptimizedHelper(MacroAssembler* masm,
Deoptimizer::BailoutType type) {
{
FrameScope scope(masm, StackFrame::INTERNAL);
// Pass the function and deoptimization type to the runtime system.
__ li(a0, Operand(Smi::FromInt(static_cast<int>(type))));
__ push(a0);
__ CallRuntime(Runtime::kNotifyDeoptimized, 1);
}
// Get the full codegen state from the stack and untag it -> t2.
__ lw(t2, MemOperand(sp, 0 * kPointerSize));
__ SmiUntag(t2);
// Switch on the state.
Label with_tos_register, unknown_state;
__ Branch(&with_tos_register,
ne, t2, Operand(FullCodeGenerator::NO_REGISTERS));
__ Addu(sp, sp, Operand(1 * kPointerSize)); // Remove state.
__ Ret();
__ bind(&with_tos_register);
__ lw(v0, MemOperand(sp, 1 * kPointerSize));
__ Branch(&unknown_state, ne, t2, Operand(FullCodeGenerator::TOS_REG));
__ Addu(sp, sp, Operand(2 * kPointerSize)); // Remove state.
__ Ret();
__ bind(&unknown_state);
__ stop("no cases left");
}
void Builtins::Generate_NotifyDeoptimized(MacroAssembler* masm) {
__ Abort("Call to unimplemented function in builtins-mips.cc");
Generate_NotifyDeoptimizedHelper(masm, Deoptimizer::EAGER);
}
void Builtins::Generate_NotifyLazyDeoptimized(MacroAssembler* masm) {
__ Abort("Call to unimplemented function in builtins-mips.cc");
Generate_NotifyDeoptimizedHelper(masm, Deoptimizer::LAZY);
}
void Builtins::Generate_NotifyOSR(MacroAssembler* masm) {
__ Abort("Call to unimplemented function in builtins-mips.cc");
// 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.
RegList saved_regs =
(kJSCallerSaved | kCalleeSaved | ra.bit() | fp.bit()) & ~sp.bit();
__ MultiPush(saved_regs);
{
FrameScope scope(masm, StackFrame::INTERNAL);
__ CallRuntime(Runtime::kNotifyOSR, 0);
}
__ MultiPop(saved_regs);
__ Ret();
}
void Builtins::Generate_OnStackReplacement(MacroAssembler* masm) {
__ Abort("Call to unimplemented function in builtins-mips.cc");
CpuFeatures::TryForceFeatureScope scope(VFP3);
if (!CpuFeatures::IsSupported(FPU)) {
__ Abort("Unreachable code: Cannot optimize without FPU support.");
return;
}
// Lookup the function in the JavaScript frame and push it as an
// argument to the on-stack replacement function.
__ lw(a0, MemOperand(fp, JavaScriptFrameConstants::kFunctionOffset));
{
FrameScope scope(masm, StackFrame::INTERNAL);
__ push(a0);
__ CallRuntime(Runtime::kCompileForOnStackReplacement, 1);
}
// If the result was -1 it means that we couldn't optimize the
// function. Just return and continue in the unoptimized version.
__ Ret(eq, v0, Operand(Smi::FromInt(-1)));
// Untag the AST id and push it on the stack.
__ SmiUntag(v0);
__ push(v0);
// Generate the code for doing the frame-to-frame translation using
// the deoptimizer infrastructure.
Deoptimizer::EntryGenerator generator(masm, Deoptimizer::OSR);
generator.Generate();
}
@ -1395,8 +1464,7 @@ void Builtins::Generate_FunctionApply(MacroAssembler* masm) {
const int kFunctionOffset = 4 * kPointerSize;
{
FrameScope scope(masm, StackFrame::INTERNAL);
FrameScope frame_scope(masm, StackFrame::INTERNAL);
__ lw(a0, MemOperand(fp, kFunctionOffset)); // Get the function.
__ push(a0);
__ lw(a0, MemOperand(fp, kArgsOffset)); // Get the args array.
@ -1530,8 +1598,7 @@ void Builtins::Generate_FunctionApply(MacroAssembler* masm) {
__ InvokeFunction(a1, actual, CALL_FUNCTION,
NullCallWrapper(), CALL_AS_METHOD);
scope.GenerateLeaveFrame();
frame_scope.GenerateLeaveFrame();
__ Ret(USE_DELAY_SLOT);
__ Addu(sp, sp, Operand(3 * kPointerSize)); // In delay slot.

735
src/mips/deoptimizer-mips.cc
View File

@ -32,24 +32,112 @@
#include "full-codegen.h"
#include "safepoint-table.h"
// Note: this file was taken from the X64 version. ARM has a partially working
// lithium implementation, but for now it is not ported to mips.
namespace v8 {
namespace internal {
const int Deoptimizer::table_entry_size_ = 10;
const int Deoptimizer::table_entry_size_ = 32;
int Deoptimizer::patch_size() {
const int kCallInstructionSizeInWords = 3;
const int kCallInstructionSizeInWords = 4;
return kCallInstructionSizeInWords * Assembler::kInstrSize;
}
void Deoptimizer::EnsureRelocSpaceForLazyDeoptimization(Handle<Code> code) {
// Nothing to do. No new relocation information is written for lazy
// deoptimization on MIPS.
}
void Deoptimizer::DeoptimizeFunction(JSFunction* function) {
UNIMPLEMENTED();
HandleScope scope;
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 an 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();
// Check that we did not shoot past next safepoint.
CHECK(pc_offset >= last_pc_offset);
#ifdef DEBUG
// Destroy the code which is not supposed to be run again.
int instructions = (pc_offset - last_pc_offset) / Assembler::kInstrSize;
CodePatcher destroyer(code->instruction_start() + last_pc_offset,
instructions);
for (int x = 0; x < instructions; x++) {
destroyer.masm()->break_(0);
}
#endif
last_pc_offset = pc_offset;
if (deoptimization_index != Safepoint::kNoDeoptimizationIndex) {
Address deoptimization_entry = Deoptimizer::GetDeoptimizationEntry(
deoptimization_index, Deoptimizer::LAZY);
last_pc_offset += gap_code_size;
int call_size_in_bytes = MacroAssembler::CallSize(deoptimization_entry,
RelocInfo::NONE);
int call_size_in_words = call_size_in_bytes / Assembler::kInstrSize;
ASSERT(call_size_in_bytes % Assembler::kInstrSize == 0);
ASSERT(call_size_in_bytes <= patch_size());
CodePatcher patcher(code->instruction_start() + last_pc_offset,
call_size_in_words);
patcher.masm()->Call(deoptimization_entry, RelocInfo::NONE);
last_pc_offset += call_size_in_bytes;
}
}
#ifdef DEBUG
// Destroy the code which is not supposed to be run again.
int instructions =
(code->safepoint_table_offset() - last_pc_offset) / Assembler::kInstrSize;
CodePatcher destroyer(code->instruction_start() + last_pc_offset,
instructions);
for (int x = 0; x < instructions; x++) {
destroyer.masm()->break_(0);
}
#endif
Isolate* isolate = code->GetIsolate();
// Add the deoptimizing code to the list.
DeoptimizingCodeListNode* node = new DeoptimizingCodeListNode(code);
DeoptimizerData* data = isolate->deoptimizer_data();
node->set_next(data->deoptimizing_code_list_);
data->deoptimizing_code_list_ = node;
// We might be in the middle of incremental marking with compaction.
// Tell collector to treat this code object in a special way and
// ignore all slots that might have been recorded on it.
isolate->heap()->mark_compact_collector()->InvalidateCode(code);
// 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(" / %x]\n", reinterpret_cast<uint32_t>(function));
#ifdef DEBUG
if (FLAG_print_code) {
code->PrintLn();
}
#endif
}
}
@ -57,7 +145,42 @@ void Deoptimizer::PatchStackCheckCodeAt(Code* unoptimized_code,
Address pc_after,
Code* check_code,
Code* replacement_code) {
UNIMPLEMENTED();
const int kInstrSize = Assembler::kInstrSize;
// This structure comes from FullCodeGenerator::EmitStackCheck.
// The call of the stack guard check has the following form:
// sltu at, sp, t0
// beq at, zero_reg, ok
// lui t9, <stack guard address> upper
// ori t9, <stack guard address> lower
// jalr t9
// nop
// ----- pc_after points here
ASSERT(Assembler::IsBeq(Assembler::instr_at(pc_after - 5 * kInstrSize)));
// Replace the sltu instruction with load-imm 1 to at, so beq is not taken.
CodePatcher patcher(pc_after - 6 * kInstrSize, 1);
patcher.masm()->addiu(at, zero_reg, 1);
// Replace the stack check address in the load-immediate (lui/ori pair)
// with the entry address of the replacement code.
ASSERT(reinterpret_cast<uint32_t>(
Assembler::target_address_at(pc_after - 4 * kInstrSize)) ==
reinterpret_cast<uint32_t>(check_code->entry()));
Assembler::set_target_address_at(pc_after - 4 * kInstrSize,
replacement_code->entry());
// We patched the code to the following form:
// addiu at, zero_reg, 1
// beq at, zero_reg, ok ;; Not changed
// lui t9, <on-stack replacement address> upper
// ori t9, <on-stack replacement address> lower
// jalr t9 ;; Not changed
// nop ;; Not changed
// ----- pc_after points here
unoptimized_code->GetHeap()->incremental_marking()->RecordCodeTargetPatch(
unoptimized_code, pc_after - 4 * kInstrSize, replacement_code);
}
@ -65,34 +188,618 @@ void Deoptimizer::RevertStackCheckCodeAt(Code* unoptimized_code,
Address pc_after,
Code* check_code,
Code* replacement_code) {
UNIMPLEMENTED();
// Exact opposite of the function above.
const int kInstrSize = Assembler::kInstrSize;
ASSERT(Assembler::IsAddImmediate(
Assembler::instr_at(pc_after - 6 * kInstrSize)));
ASSERT(Assembler::IsBeq(Assembler::instr_at(pc_after - 5 * kInstrSize)));
// Restore the sltu instruction so beq can be taken again.
CodePatcher patcher(pc_after - 6 * kInstrSize, 1);
patcher.masm()->sltu(at, sp, t0);
// Replace the on-stack replacement address in the load-immediate (lui/ori
// pair) with the entry address of the normal stack-check code.
ASSERT(reinterpret_cast<uint32_t>(
Assembler::target_address_at(pc_after - 4 * kInstrSize)) ==
reinterpret_cast<uint32_t>(replacement_code->entry()));
Assembler::set_target_address_at(pc_after - 4 * kInstrSize,
check_code->entry());
check_code->GetHeap()->incremental_marking()->RecordCodeTargetPatch(
unoptimized_code, pc_after - 4 * kInstrSize, check_code);
}
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_);
#ifdef DEBUG
output_[0]->SetKind(Code::OPTIMIZED_FUNCTION);
#endif
// 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;
}
// 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 = isolate_->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));
function->PrintName();
PrintF(" => pc=0x%0x]\n", output_[0]->GetPc());
}
}
// This code is very similar to ia32/arm code, but relies on register names
// (fp, sp) and how the frame is laid out.
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);
#ifdef DEBUG
output_frame->SetKind(Code::FUNCTION);
#endif
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.
uint32_t top_address;
if (is_bottommost) {
// 2 = context and function in the frame.
top_address =
input_->GetRegister(fp.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%08x: [top + %d] <- 0x%08x ; 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(fp.code()) == fp_value);
output_frame->SetFp(fp_value);
if (is_topmost) {
output_frame->SetRegister(fp.code(), fp_value);
}
if (FLAG_trace_deopt) {
PrintF(" 0x%08x: [top + %d] <- 0x%08x ; caller's fp\n",
fp_value, output_offset, value);
}
// For the bottommost output frame the context can be gotten from the input
// frame. For all subsequent output frames it can be gotten from the function
// so long as we don't inline functions that need local contexts.
output_offset -= kPointerSize;
input_offset -= kPointerSize;
if (is_bottommost) {
value = input_->GetFrameSlot(input_offset);
} else {
value = reinterpret_cast<intptr_t>(function->context());
}
output_frame->SetFrameSlot(output_offset, value);
if (is_topmost) {
output_frame->SetRegister(cp.code(), value);
}
if (FLAG_trace_deopt) {
PrintF(" 0x%08x: [top + %d] <- 0x%08x ; 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<uint32_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%08x: [top + %d] <- 0x%08x ; 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);
uint32_t pc_value = reinterpret_cast<uint32_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 && bailout_type_ != DEBUGGER) {
Builtins* builtins = isolate_->builtins();
Code* continuation = (bailout_type_ == EAGER)
? builtins->builtin(Builtins::kNotifyDeoptimized)
: builtins->builtin(Builtins::kNotifyLazyDeoptimized);
output_frame->SetContinuation(
reinterpret_cast<uint32_t>(continuation->entry()));
}
}
void Deoptimizer::FillInputFrame(Address tos, JavaScriptFrame* frame) {
UNIMPLEMENTED();
// 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 fp and sp are set to the correct values though.
for (int i = 0; i < Register::kNumRegisters; i++) {
input_->SetRegister(i, i * 4);
}
input_->SetRegister(sp.code(), reinterpret_cast<intptr_t>(frame->sp()));
input_->SetRegister(fp.code(), reinterpret_cast<intptr_t>(frame->fp()));
for (int i = 0; i < DoubleRegister::kNumAllocatableRegisters; 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));
}
}
#define __ masm()->
// This code tries to be close to ia32 code so that any changes can be
// easily ported.
void Deoptimizer::EntryGenerator::Generate() {
UNIMPLEMENTED();
GeneratePrologue();
Isolate* isolate = masm()->isolate();
CpuFeatures::Scope scope(FPU);
// Unlike on ARM we don't save all the registers, just the useful ones.
// For the rest, there are gaps on the stack, so the offsets remain the same.
const int kNumberOfRegisters = Register::kNumRegisters;
RegList restored_regs = kJSCallerSaved | kCalleeSaved;
RegList saved_regs = restored_regs | sp.bit() | ra.bit();
const int kDoubleRegsSize =
kDoubleSize * FPURegister::kNumAllocatableRegisters;
// Save all FPU registers before messing with them.
__ Subu(sp, sp, Operand(kDoubleRegsSize));
for (int i = 0; i < FPURegister::kNumAllocatableRegisters; ++i) {
FPURegister fpu_reg = FPURegister::FromAllocationIndex(i);
int offset = i * kDoubleSize;
__ sdc1(fpu_reg, MemOperand(sp, offset));
}
// Push saved_regs (needed to populate FrameDescription::registers_).
// Leave gaps for other registers.
__ Subu(sp, sp, kNumberOfRegisters * kPointerSize);
for (int16_t i = kNumberOfRegisters - 1; i >= 0; i--) {
if ((saved_regs & (1 << i)) != 0) {
__ sw(ToRegister(i), MemOperand(sp, kPointerSize * i));
}
}
const int kSavedRegistersAreaSize =
(kNumberOfRegisters * kPointerSize) + kDoubleRegsSize;
// Get the bailout id from the stack.
__ lw(a2, MemOperand(sp, kSavedRegistersAreaSize));
// Get the address of the location in the code object if possible (a3) (return
// address for lazy deoptimization) and compute the fp-to-sp delta in
// register t0.
if (type() == EAGER) {
__ mov(a3, zero_reg);
// Correct one word for bailout id.
__ Addu(t0, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
} else if (type() == OSR) {
__ mov(a3, ra);
// Correct one word for bailout id.
__ Addu(t0, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
} else {
__ mov(a3, ra);
// Correct two words for bailout id and return address.
__ Addu(t0, sp, Operand(kSavedRegistersAreaSize + (2 * kPointerSize)));
}
__ Subu(t0, fp, t0);
// Allocate a new deoptimizer object.
// Pass four arguments in a0 to a3 and fifth & sixth arguments on stack.
__ PrepareCallCFunction(6, t1);
__ lw(a0, MemOperand(fp, JavaScriptFrameConstants::kFunctionOffset));
__ li(a1, Operand(type())); // bailout type,
// a2: bailout id already loaded.
// a3: code address or 0 already loaded.
__ sw(t0, CFunctionArgumentOperand(5)); // Fp-to-sp delta.
__ li(t1, Operand(ExternalReference::isolate_address()));
__ sw(t1, CFunctionArgumentOperand(6)); // Isolate.
// Call Deoptimizer::New().
{
AllowExternalCallThatCantCauseGC scope(masm());
__ CallCFunction(ExternalReference::new_deoptimizer_function(isolate), 6);
}
// Preserve "deoptimizer" object in register v0 and get the input
// frame descriptor pointer to a1 (deoptimizer->input_);
// Move deopt-obj to a0 for call to Deoptimizer::ComputeOutputFrames() below.
__ mov(a0, v0);
__ lw(a1, MemOperand(v0, Deoptimizer::input_offset()));
// Copy core registers into FrameDescription::registers_[kNumRegisters].
ASSERT(Register::kNumRegisters == kNumberOfRegisters);
for (int i = 0; i < kNumberOfRegisters; i++) {
int offset = (i * kPointerSize) + FrameDescription::registers_offset();
if ((saved_regs & (1 << i)) != 0) {
__ lw(a2, MemOperand(sp, i * kPointerSize));
__ sw(a2, MemOperand(a1, offset));
} else if (FLAG_debug_code) {
__ li(a2, kDebugZapValue);
__ sw(a2, MemOperand(a1, offset));
}
}
// Copy FPU registers to
// double_registers_[DoubleRegister::kNumAllocatableRegisters]
int double_regs_offset = FrameDescription::double_registers_offset();
for (int i = 0; i < FPURegister::kNumAllocatableRegisters; ++i) {
int dst_offset = i * kDoubleSize + double_regs_offset;
int src_offset = i * kDoubleSize + kNumberOfRegisters * kPointerSize;
__ ldc1(f0, MemOperand(sp, src_offset));
__ sdc1(f0, MemOperand(a1, dst_offset));
}
// Remove the bailout id, eventually return address, and the saved registers
// from the stack.
if (type() == EAGER || type() == OSR) {
__ Addu(sp, sp, Operand(kSavedRegistersAreaSize + (1 * kPointerSize)));
} else {
__ Addu(sp, sp, Operand(kSavedRegistersAreaSize + (2 * kPointerSize)));
}
// Compute a pointer to the unwinding limit in register a2; that is
// the first stack slot not part of the input frame.
__ lw(a2, MemOperand(a1, FrameDescription::frame_size_offset()));
__ Addu(a2, a2, sp);
// Unwind the stack down to - but not including - the unwinding
// limit and copy the contents of the activation frame to the input
// frame description.
__ Addu(a3, a1, Operand(FrameDescription::frame_content_offset()));
Label pop_loop;
__ bind(&pop_loop);
__ pop(t0);
__ sw(t0, MemOperand(a3, 0));
__ Branch(USE_DELAY_SLOT, &pop_loop, ne, a2, Operand(sp));
__ addiu(a3, a3, sizeof(uint32_t)); // In delay slot.
// Compute the output frame in the deoptimizer.
__ push(a0); // Preserve deoptimizer object across call.
// a0: deoptimizer object; a1: scratch.
__ PrepareCallCFunction(1, a1);
// Call Deoptimizer::ComputeOutputFrames().
{
AllowExternalCallThatCantCauseGC scope(masm());
__ CallCFunction(
ExternalReference::compute_output_frames_function(isolate), 1);
}
__ pop(a0); // Restore deoptimizer object (class Deoptimizer).
// Replace the current (input) frame with the output frames.
Label outer_push_loop, inner_push_loop;
// Outer loop state: a0 = current "FrameDescription** output_",
// a1 = one past the last FrameDescription**.
__ lw(a1, MemOperand(a0, Deoptimizer::output_count_offset()));
__ lw(a0, MemOperand(a0, Deoptimizer::output_offset())); // a0 is output_.
__ sll(a1, a1, kPointerSizeLog2); // Count to offset.
__ addu(a1, a0, a1); // a1 = one past the last FrameDescription**.
__ bind(&outer_push_loop);
// Inner loop state: a2 = current FrameDescription*, a3 = loop index.
__ lw(a2, MemOperand(a0, 0)); // output_[ix]
__ lw(a3, MemOperand(a2, FrameDescription::frame_size_offset()));
__ bind(&inner_push_loop);
__ Subu(a3, a3, Operand(sizeof(uint32_t)));
__ Addu(t2, a2, Operand(a3));
__ lw(t3, MemOperand(t2, FrameDescription::frame_content_offset()));
__ push(t3);
__ Branch(&inner_push_loop, ne, a3, Operand(zero_reg));
__ Addu(a0, a0, Operand(kPointerSize));
__ Branch(&outer_push_loop, lt, a0, Operand(a1));
// Push state, pc, and continuation from the last output frame.
if (type() != OSR) {
__ lw(t2, MemOperand(a2, FrameDescription::state_offset()));
__ push(t2);
}
__ lw(t2, MemOperand(a2, FrameDescription::pc_offset()));
__ push(t2);
__ lw(t2, MemOperand(a2, FrameDescription::continuation_offset()));
__ push(t2);
// Technically restoring 'at' should work unless zero_reg is also restored
// but it's safer to check for this.
ASSERT(!(at.bit() & restored_regs));
// Restore the registers from the last output frame.
__ mov(at, a2);
for (int i = kNumberOfRegisters - 1; i >= 0; i--) {
int offset = (i * kPointerSize) + FrameDescription::registers_offset();
if ((restored_regs & (1 << i)) != 0) {
__ lw(ToRegister(i), MemOperand(at, offset));
}
}
// Set up the roots register.
ExternalReference roots_array_start =
ExternalReference::roots_array_start(isolate);
__ li(roots, Operand(roots_array_start));
__ pop(at); // Get continuation, leave pc on stack.
__ pop(ra);
__ Jump(at);
__ stop("Unreachable.");
}
void Deoptimizer::TableEntryGenerator::GeneratePrologue() {
UNIMPLEMENTED();
Assembler::BlockTrampolinePoolScope block_trampoline_pool(masm());
// Create a sequence of deoptimization entries. Note that any
// registers may be still live.
Label done;
for (int i = 0; i < count(); i++) {
int start = masm()->pc_offset();
USE(start);
if (type() != EAGER) {
// Emulate ia32 like call by pushing return address to stack.
__ push(ra);
}
__ li(at, Operand(i));
__ push(at);
__ Branch(&done);
// Pad the rest of the code.
while (table_entry_size_ > (masm()->pc_offset() - start)) {
__ nop();
}
ASSERT_EQ(table_entry_size_, masm()->pc_offset() - start);
}
__ bind(&done);
}
#undef __
} } // namespace v8::internal

4628
src/mips/lithium-codegen-mips.cc Normal file
View File

File diff suppressed because it is too large Load Diff

383
src/mips/lithium-codegen-mips.h
View File

@ -29,35 +29,398 @@
#define V8_MIPS_LITHIUM_CODEGEN_MIPS_H_
#include "mips/lithium-mips.h"
#include "mips/lithium-gap-resolver-mips.h"
#include "deoptimizer.h"
#include "safepoint-table.h"
#include "scopes.h"
// Note: this file was taken from the X64 version. ARM has a partially working
// lithium implementation, but for now it is not ported to mips.
namespace v8 {
namespace internal {
// Forward declarations.
class LDeferredCode;
class SafepointGenerator;
class LCodeGen BASE_EMBEDDED {
public:
LCodeGen(LChunk* chunk, MacroAssembler* assembler, CompilationInfo* info) { }
LCodeGen(LChunk* chunk, MacroAssembler* assembler, CompilationInfo* info)
: chunk_(chunk),
masm_(assembler),
info_(info),
current_block_(-1),
current_instruction_(-1),
instructions_(chunk->instructions()),
deoptimizations_(4),
deopt_jump_table_(4),
deoptimization_literals_(8),
inlined_function_count_(0),
scope_(info->scope()),
status_(UNUSED),
deferred_(8),
osr_pc_offset_(-1),
resolver_(this),
expected_safepoint_kind_(Safepoint::kSimple) {
PopulateDeoptimizationLiteralsWithInlinedFunctions();
}
// Simple accessors.
MacroAssembler* masm() const { return masm_; }
CompilationInfo* info() const { return info_; }
Isolate* isolate() const { return info_->isolate(); }
Factory* factory() const { return isolate()->factory(); }
Heap* heap() const { return isolate()->heap(); }
// Support for converting LOperands to assembler types.
// LOperand must be a register.
Register ToRegister(LOperand* op) const;
// LOperand is loaded into scratch, unless already a register.
Register EmitLoadRegister(LOperand* op, Register scratch);
// LOperand must be a double register.
DoubleRegister ToDoubleRegister(LOperand* op) const;
// LOperand is loaded into dbl_scratch, unless already a double register.
DoubleRegister EmitLoadDoubleRegister(LOperand* op,
FloatRegister flt_scratch,
DoubleRegister dbl_scratch);
int ToInteger32(LConstantOperand* op) const;
double ToDouble(LConstantOperand* op) const;
Operand ToOperand(LOperand* op);
MemOperand ToMemOperand(LOperand* op) const;
// Returns a MemOperand pointing to the high word of a DoubleStackSlot.
MemOperand ToHighMemOperand(LOperand* op) const;
// Try to generate code for the entire chunk, but it may fail if the
// chunk contains constructs we cannot handle. Returns true if the
// code generation attempt succeeded.
bool GenerateCode() {
UNIMPLEMENTED();
return false;
}
bool GenerateCode();
// Finish the code by setting stack height, safepoint, and bailout
// information on it.
void FinishCode(Handle<Code> code) { UNIMPLEMENTED(); }
void FinishCode(Handle<Code> code);
// Deferred code support.
template<int T>
void DoDeferredBinaryOpStub(LTemplateInstruction<1, 2, T>* instr,
Token::Value op);
void DoDeferredNumberTagD(LNumberTagD* instr);
void DoDeferredNumberTagI(LNumberTagI* instr);
void DoDeferredTaggedToI(LTaggedToI* instr);
void DoDeferredMathAbsTaggedHeapNumber(LUnaryMathOperation* instr);
void DoDeferredStackCheck(LStackCheck* instr);
void DoDeferredStringCharCodeAt(LStringCharCodeAt* instr);
void DoDeferredStringCharFromCode(LStringCharFromCode* instr);
void DoDeferredLInstanceOfKnownGlobal(LInstanceOfKnownGlobal* instr,
Label* map_check);
// Parallel move support.
void DoParallelMove(LParallelMove* move);
void DoGap(LGap* instr);
// Emit frame translation commands for an environment.
void WriteTranslation(LEnvironment* environment, Translation* translation);
// Declare methods that deal with the individual node types.
#define DECLARE_DO(type) void Do##type(L##type* node);
LITHIUM_CONCRETE_INSTRUCTION_LIST(DECLARE_DO)
#undef DECLARE_DO
private:
enum Status {
UNUSED,
GENERATING,
DONE,
ABORTED
};
bool is_unused() const { return status_ == UNUSED; }
bool is_generating() const { return status_ == GENERATING; }
bool is_done() const { return status_ == DONE; }
bool is_aborted() const { return status_ == ABORTED; }
StrictModeFlag strict_mode_flag() const {
return info()->strict_mode_flag();
}
LChunk* chunk() const { return chunk_; }
Scope* scope() const { return scope_; }
HGraph* graph() const { return chunk_->graph(); }
Register scratch0() { return lithiumScratchReg; }
Register scratch1() { return lithiumScratchReg2; }
DoubleRegister double_scratch0() { return lithiumScratchDouble; }
int GetNextEmittedBlock(int block);
LInstruction* GetNextInstruction();
void EmitClassOfTest(Label* if_true,
Label* if_false,
Handle<String> class_name,
Register input,
Register temporary,
Register temporary2);
int GetStackSlotCount() const { return chunk()->spill_slot_count(); }
int GetParameterCount() const { return scope()->num_parameters(); }
void Abort(const char* format, ...);
void Comment(const char* format, ...);
void AddDeferredCode(LDeferredCode* code) { deferred_.Add(code); }
// Code generation passes. Returns true if code generation should
// continue.
bool GeneratePrologue();
bool GenerateBody();
bool GenerateDeferredCode();
bool GenerateDeoptJumpTable();
bool GenerateSafepointTable();
enum SafepointMode {
RECORD_SIMPLE_SAFEPOINT,
RECORD_SAFEPOINT_WITH_REGISTERS_AND_NO_ARGUMENTS
};
void CallCode(Handle<Code> code,
RelocInfo::Mode mode,
LInstruction* instr);
void CallCodeGeneric(Handle<Code> code,
RelocInfo::Mode mode,
LInstruction* instr,
SafepointMode safepoint_mode);
void CallRuntime(const Runtime::Function* function,
int num_arguments,
LInstruction* instr);
void CallRuntime(Runtime::FunctionId id,
int num_arguments,
LInstruction* instr) {
const Runtime::Function* function = Runtime::FunctionForId(id);
CallRuntime(function, num_arguments, instr);
}
void CallRuntimeFromDeferred(Runtime::FunctionId id,
int argc,
LInstruction* instr);
// Generate a direct call to a known function. Expects the function
// to be in a1.
void CallKnownFunction(Handle<JSFunction> function,
int arity,
LInstruction* instr,
CallKind call_kind);
void LoadHeapObject(Register result, Handle<HeapObject> object);
void RegisterLazyDeoptimization(LInstruction* instr,
SafepointMode safepoint_mode);
void RegisterEnvironmentForDeoptimization(LEnvironment* environment);
void DeoptimizeIf(Condition cc,
LEnvironment* environment,
Register src1,
const Operand& src2);
void AddToTranslation(Translation* translation,
LOperand* op,
bool is_tagged);
void PopulateDeoptimizationData(Handle<Code> code);
int DefineDeoptimizationLiteral(Handle<Object> literal);
void PopulateDeoptimizationLiteralsWithInlinedFunctions();
Register ToRegister(int index) const;
DoubleRegister ToDoubleRegister(int index) const;
// Specific math operations - used from DoUnaryMathOperation.
void EmitIntegerMathAbs(LUnaryMathOperation* instr);
void DoMathAbs(LUnaryMathOperation* instr);
void DoMathFloor(LUnaryMathOperation* instr);
void DoMathRound(LUnaryMathOperation* instr);
void DoMathSqrt(LUnaryMathOperation* instr);
void DoMathPowHalf(LUnaryMathOperation* instr);
void DoMathLog(LUnaryMathOperation* instr);
void DoMathCos(LUnaryMathOperation* instr);
void DoMathSin(LUnaryMathOperation* instr);
// Support for recording safepoint and position information.
void RecordSafepoint(LPointerMap* pointers,
Safepoint::Kind kind,
int arguments,
int deoptimization_index);
void RecordSafepoint(LPointerMap* pointers, int deoptimization_index);
void RecordSafepoint(int deoptimization_index);
void RecordSafepointWithRegisters(LPointerMap* pointers,
int arguments,
int deoptimization_index);
void RecordSafepointWithRegistersAndDoubles(LPointerMap* pointers,
int arguments,
int deoptimization_index);
void RecordPosition(int position);
int LastSafepointEnd() {
return static_cast<int>(safepoints_.GetPcAfterGap());
}
static Condition TokenToCondition(Token::Value op, bool is_unsigned);
void EmitGoto(int block);
void EmitBranch(int left_block,
int right_block,
Condition cc,
Register src1,
const Operand& src2);
void EmitBranchF(int left_block,
int right_block,
Condition cc,
FPURegister src1,
FPURegister src2);
void EmitCmpI(LOperand* left, LOperand* right);
void EmitNumberUntagD(Register input,
DoubleRegister result,
bool deoptimize_on_undefined,
LEnvironment* env);
// Emits optimized code for typeof x == "y". Modifies input register.
// Returns the condition on which a final split to
// true and false label should be made, to optimize fallthrough.
// Returns two registers in cmp1 and cmp2 that can be used in the
// Branch instruction after EmitTypeofIs.
Condition EmitTypeofIs(Label* true_label,
Label* false_label,
Register input,
Handle<String> type_name,
Register& cmp1,
Operand& cmp2);
// Emits optimized code for %_IsObject(x). Preserves input register.
// Returns the condition on which a final split to
// true and false label should be made, to optimize fallthrough.
Condition EmitIsObject(Register input,
Register temp1,
Label* is_not_object,
Label* is_object);
// Emits optimized code for %_IsConstructCall().
// Caller should branch on equal condition.
void EmitIsConstructCall(Register temp1, Register temp2);
void EmitLoadFieldOrConstantFunction(Register result,
Register object,
Handle<Map> type,
Handle<String> name);
struct JumpTableEntry {
explicit inline JumpTableEntry(Address entry)
: label(),
address(entry) { }
Label label;
Address address;
};
LChunk* const chunk_;
MacroAssembler* const masm_;
CompilationInfo* const info_;
int current_block_;
int current_instruction_;
const ZoneList<LInstruction*>* instructions_;
ZoneList<LEnvironment*> deoptimizations_;
ZoneList<JumpTableEntry> deopt_jump_table_;
ZoneList<Handle<Object> > deoptimization_literals_;
int inlined_function_count_;
Scope* const scope_;
Status status_;
TranslationBuffer translations_;
ZoneList<LDeferredCode*> deferred_;
int osr_pc_offset_;
// Builder that keeps track of safepoints in the code. The table
// itself is emitted at the end of the generated code.
SafepointTableBuilder safepoints_;
// Compiler from a set of parallel moves to a sequential list of moves.
LGapResolver resolver_;
Safepoint::Kind expected_safepoint_kind_;
class PushSafepointRegistersScope BASE_EMBEDDED {
public:
PushSafepointRegistersScope(LCodeGen* codegen,
Safepoint::Kind kind)
: codegen_(codegen) {
ASSERT(codegen_->expected_safepoint_kind_ == Safepoint::kSimple);
codegen_->expected_safepoint_kind_ = kind;
switch (codegen_->expected_safepoint_kind_) {
case Safepoint::kWithRegisters:
codegen_->masm_->PushSafepointRegisters();
break;
case Safepoint::kWithRegistersAndDoubles:
codegen_->masm_->PushSafepointRegistersAndDoubles();
break;
default:
UNREACHABLE();
}
}
~PushSafepointRegistersScope() {
Safepoint::Kind kind = codegen_->expected_safepoint_kind_;
ASSERT((kind & Safepoint::kWithRegisters) != 0);
switch (kind) {
case Safepoint::kWithRegisters:
codegen_->masm_->PopSafepointRegisters();
break;
case Safepoint::kWithRegistersAndDoubles:
codegen_->masm_->PopSafepointRegistersAndDoubles();
break;
default:
UNREACHABLE();
}
codegen_->expected_safepoint_kind_ = Safepoint::kSimple;
}
private:
LCodeGen* codegen_;
};
friend class LDeferredCode;
friend class LEnvironment;
friend class SafepointGenerator;
DISALLOW_COPY_AND_ASSIGN(LCodeGen);
};
class LDeferredCode: public ZoneObject {
public:
explicit LDeferredCode(LCodeGen* codegen)
: codegen_(codegen),
external_exit_(NULL),
instruction_index_(codegen->current_instruction_) {
codegen->AddDeferredCode(this);
}
virtual ~LDeferredCode() { }
virtual void Generate() = 0;
virtual LInstruction* instr() = 0;
void SetExit(Label *exit) { external_exit_ = exit; }
Label* entry() { return &entry_; }
Label* exit() { return external_exit_ != NULL ? external_exit_ : &exit_; }
int instruction_index() const { return instruction_index_; }
protected:
LCodeGen* codegen() const { return codegen_; }
MacroAssembler* masm() const { return codegen_->masm(); }
private:
LCodeGen* codegen_;
Label entry_;
Label exit_;
Label* external_exit_;
int instruction_index_;
};
} } // namespace v8::internal

309
src/mips/lithium-gap-resolver-mips.cc Normal file
View File

@ -0,0 +1,309 @@
// Copyright 2011 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"
#include "mips/lithium-gap-resolver-mips.h"
#include "mips/lithium-codegen-mips.h"
namespace v8 {
namespace internal {
static const Register kSavedValueRegister = lithiumScratchReg;
static const DoubleRegister kSavedDoubleValueRegister = lithiumScratchDouble;
LGapResolver::LGapResolver(LCodeGen* owner)
: cgen_(owner),
moves_(32),
root_index_(0),
in_cycle_(false),
saved_destination_(NULL) {}
void LGapResolver::Resolve(LParallelMove* parallel_move) {
ASSERT(moves_.is_empty());
// Build up a worklist of moves.
BuildInitialMoveList(parallel_move);
for (int i = 0; i < moves_.length(); ++i) {
LMoveOperands move = moves_[i];
// Skip constants to perform them last. They don't block other moves
// and skipping such moves with register destinations keeps those
// registers free for the whole algorithm.
if (!move.IsEliminated() && !move.source()->IsConstantOperand()) {
root_index_ = i; // Any cycle is found when by reaching this move again.
PerformMove(i);
if (in_cycle_) {
RestoreValue();
}
}
}
// Perform the moves with constant sources.
for (int i = 0; i < moves_.length(); ++i) {
if (!moves_[i].IsEliminated()) {
ASSERT(moves_[i].source()->IsConstantOperand());
EmitMove(i);
}
}
moves_.Rewind(0);
}
void LGapResolver::BuildInitialMoveList(LParallelMove* parallel_move) {
// Perform a linear sweep of the moves to add them to the initial list of
// moves to perform, ignoring any move that is redundant (the source is
// the same as the destination, the destination is ignored and
// unallocated, or the move was already eliminated).
const ZoneList<LMoveOperands>* moves = parallel_move->move_operands();
for (int i = 0; i < moves->length(); ++i) {
LMoveOperands move = moves->at(i);
if (!move.IsRedundant()) moves_.Add(move);
}
Verify();
}
void LGapResolver::PerformMove(int index) {
// Each call to this function performs a move and deletes it from the move
// graph. We first recursively perform any move blocking this one. We
// mark a move as "pending" on entry to PerformMove in order to detect
// cycles in the move graph.
// We can only find a cycle, when doing a depth-first traversal of moves,
// be encountering the starting move again. So by spilling the source of
// the starting move, we break the cycle. All moves are then unblocked,
// and the starting move is completed by writing the spilled value to
// its destination. All other moves from the spilled source have been
// completed prior to breaking the cycle.
// An additional complication is that moves to MemOperands with large
// offsets (more than 1K or 4K) require us to spill this spilled value to
// the stack, to free up the register.
ASSERT(!moves_[index].IsPending());
ASSERT(!moves_[index].IsRedundant());
// Clear this move's destination to indicate a pending move. The actual
// destination is saved in a stack allocated local. Multiple moves can
// be pending because this function is recursive.
ASSERT(moves_[index].source() != NULL); // Or else it will look eliminated.
LOperand* destination = moves_[index].destination();
moves_[index].set_destination(NULL);
// Perform a depth-first traversal of the move graph to resolve
// dependencies. Any unperformed, unpending move with a source the same
// as this one's destination blocks this one so recursively perform all
// such moves.
for (int i = 0; i < moves_.length(); ++i) {
LMoveOperands other_move = moves_[i];
if (other_move.Blocks(destination) && !other_move.IsPending()) {
PerformMove(i);
// If there is a blocking, pending move it must be moves_[root_index_]
// and all other moves with the same source as moves_[root_index_] are
// sucessfully executed (because they are cycle-free) by this loop.
}
}
// We are about to resolve this move and don't need it marked as
// pending, so restore its destination.
moves_[index].set_destination(destination);
// The move may be blocked on a pending move, which must be the starting move.
// In this case, we have a cycle, and we save the source of this move to
// a scratch register to break it.
LMoveOperands other_move = moves_[root_index_];
if (other_move.Blocks(destination)) {
ASSERT(other_move.IsPending());
BreakCycle(index);
return;
}
// This move is no longer blocked.
EmitMove(index);
}
void LGapResolver::Verify() {
#ifdef ENABLE_SLOW_ASSERTS
// No operand should be the destination for more than one move.
for (int i = 0; i < moves_.length(); ++i) {
LOperand* destination = moves_[i].destination();
for (int j = i + 1; j < moves_.length(); ++j) {
SLOW_ASSERT(!destination->Equals(moves_[j].destination()));
}
}
#endif
}
#define __ ACCESS_MASM(cgen_->masm())
void LGapResolver::BreakCycle(int index) {
// We save in a register the value that should end up in the source of
// moves_[root_index]. After performing all moves in the tree rooted
// in that move, we save the value to that source.
ASSERT(moves_[index].destination()->Equals(moves_[root_index_].source()));
ASSERT(!in_cycle_);
in_cycle_ = true;
LOperand* source = moves_[index].source();
saved_destination_ = moves_[index].destination();
if (source->IsRegister()) {
__ mov(kSavedValueRegister, cgen_->ToRegister(source));
} else if (source->IsStackSlot()) {
__ lw(kSavedValueRegister, cgen_->ToMemOperand(source));
} else if (source->IsDoubleRegister()) {
__ mov_d(kSavedDoubleValueRegister, cgen_->ToDoubleRegister(source));
} else if (source->IsDoubleStackSlot()) {
__ ldc1(kSavedDoubleValueRegister, cgen_->ToMemOperand(source));
} else {
UNREACHABLE();
}
// This move will be done by restoring the saved value to the destination.
moves_[index].Eliminate();
}
void LGapResolver::RestoreValue() {
ASSERT(in_cycle_);
ASSERT(saved_destination_ != NULL);
// Spilled value is in kSavedValueRegister or kSavedDoubleValueRegister.
if (saved_destination_->IsRegister()) {
__ mov(cgen_->ToRegister(saved_destination_), kSavedValueRegister);
} else if (saved_destination_->IsStackSlot()) {
__ sw(kSavedValueRegister, cgen_->ToMemOperand(saved_destination_));
} else if (saved_destination_->IsDoubleRegister()) {
__ mov_d(cgen_->ToDoubleRegister(saved_destination_),
kSavedDoubleValueRegister);
} else if (saved_destination_->IsDoubleStackSlot()) {
__ sdc1(kSavedDoubleValueRegister,
cgen_->ToMemOperand(saved_destination_));
} else {
UNREACHABLE();
}
in_cycle_ = false;
saved_destination_ = NULL;
}
void LGapResolver::EmitMove(int index) {
LOperand* source = moves_[index].source();
LOperand* destination = moves_[index].destination();
// Dispatch on the source and destination operand kinds. Not all
// combinations are possible.
if (source->IsRegister()) {
Register source_register = cgen_->ToRegister(source);
if (destination->IsRegister()) {
__ mov(cgen_->ToRegister(destination), source_register);
} else {
ASSERT(destination->IsStackSlot());
__ sw(source_register, cgen_->ToMemOperand(destination));
}
} else if (source->IsStackSlot()) {
MemOperand source_operand = cgen_->ToMemOperand(source);
if (destination->IsRegister()) {
__ lw(cgen_->ToRegister(destination), source_operand);
} else {
ASSERT(destination->IsStackSlot());
MemOperand destination_operand = cgen_->ToMemOperand(destination);
if (in_cycle_) {
if (!destination_operand.OffsetIsInt16Encodable()) {
// 'at' is overwritten while saving the value to the destination.
// Therefore we can't use 'at'. It is OK if the read from the source
// destroys 'at', since that happens before the value is read.
// This uses only a single reg of the double reg-pair.
__ lwc1(kSavedDoubleValueRegister, source_operand);
__ swc1(kSavedDoubleValueRegister, destination_operand);
} else {
__ lw(at, source_operand);
__ sw(at, destination_operand);
}
} else {
__ lw(kSavedValueRegister, source_operand);
__ sw(kSavedValueRegister, destination_operand);
}
}
} else if (source->IsConstantOperand()) {
Operand source_operand = cgen_->ToOperand(source);
if (destination->IsRegister()) {
__ li(cgen_->ToRegister(destination), source_operand);
} else {
ASSERT(destination->IsStackSlot());
ASSERT(!in_cycle_); // Constant moves happen after all cycles are gone.
MemOperand destination_operand = cgen_->ToMemOperand(destination);
__ li(kSavedValueRegister, source_operand);
__ sw(kSavedValueRegister, cgen_->ToMemOperand(destination));
}
} else if (source->IsDoubleRegister()) {
DoubleRegister source_register = cgen_->ToDoubleRegister(source);
if (destination->IsDoubleRegister()) {
__ mov_d(cgen_->ToDoubleRegister(destination), source_register);
} else {
ASSERT(destination->IsDoubleStackSlot());
MemOperand destination_operand = cgen_->ToMemOperand(destination);
__ sdc1(source_register, destination_operand);
}
} else if (source->IsDoubleStackSlot()) {
MemOperand source_operand = cgen_->ToMemOperand(source);
if (destination->IsDoubleRegister()) {
__ ldc1(cgen_->ToDoubleRegister(destination), source_operand);
} else {
ASSERT(destination->IsDoubleStackSlot());
MemOperand destination_operand = cgen_->ToMemOperand(destination);
if (in_cycle_) {
// kSavedDoubleValueRegister was used to break the cycle,
// but kSavedValueRegister is free.
MemOperand source_high_operand =
cgen_->ToHighMemOperand(source);
MemOperand destination_high_operand =
cgen_->ToHighMemOperand(destination);
__ lw(kSavedValueRegister, source_operand);
__ sw(kSavedValueRegister, destination_operand);
__ lw(kSavedValueRegister, source_high_operand);
__ sw(kSavedValueRegister, destination_high_operand);
} else {
__ ldc1(kSavedDoubleValueRegister, source_operand);
__ sdc1(kSavedDoubleValueRegister, destination_operand);
}
}
} else {
UNREACHABLE();
}
moves_[index].Eliminate();
}
#undef __
} } // namespace v8::internal

84
src/mips/lithium-gap-resolver-mips.h Normal file
View File

@ -0,0 +1,84 @@
// Copyright 2011 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.
#ifndef V8_MIPS_LITHIUM_GAP_RESOLVER_MIPS_H_
#define V8_MIPS_LITHIUM_GAP_RESOLVER_MIPS_H_
#include "v8.h"
#include "lithium.h"
namespace v8 {
namespace internal {
class LCodeGen;
class LGapResolver;
class LGapResolver BASE_EMBEDDED {
public:
explicit LGapResolver(LCodeGen* owner);
// Resolve a set of parallel moves, emitting assembler instructions.
void Resolve(LParallelMove* parallel_move);
private:
// Build the initial list of moves.
void BuildInitialMoveList(LParallelMove* parallel_move);
// Perform the move at the moves_ index in question (possibly requiring
// other moves to satisfy dependencies).
void PerformMove(int index);
// If a cycle is found in the series of moves, save the blocking value to
// a scratch register. The cycle must be found by hitting the root of the
// depth-first search.
void BreakCycle(int index);
// After a cycle has been resolved, restore the value from the scratch
// register to its proper destination.
void RestoreValue();
// Emit a move and remove it from the move graph.
void EmitMove(int index);
// Verify the move list before performing moves.
void Verify();
LCodeGen* cgen_;
// List of moves not yet resolved.
ZoneList<LMoveOperands> moves_;
int root_index_;
bool in_cycle_;
LOperand* saved_destination_;
};
} } // namespace v8::internal
#endif // V8_MIPS_LITHIUM_GAP_RESOLVER_MIPS_H_

2218
src/mips/lithium-mips.cc Normal file
View File

File diff suppressed because it is too large Load Diff

2286
src/mips/lithium-mips.h
View File

File diff suppressed because it is too large Load Diff