v8/src/serialize.cc

1629 lines
60 KiB
C++
Raw Normal View History

// Copyright 2012 the V8 project authors. All rights reserved.
// Redistribution and use in source and binary forms, with or without
// modification, are permitted provided that the following conditions are
// met:
//
// * Redistributions of source code must retain the above copyright
// notice, this list of conditions and the following disclaimer.
// * Redistributions in binary form must reproduce the above
// copyright notice, this list of conditions and the following
// disclaimer in the documentation and/or other materials provided
// with the distribution.
// * Neither the name of Google Inc. nor the names of its
// contributors may be used to endorse or promote products derived
// from this software without specific prior written permission.
//
// THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
// "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
// LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
// A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
// OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
// SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
// LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
// DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
// THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
// (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
// OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
#include "v8.h"
#include "accessors.h"
#include "api.h"
#include "bootstrapper.h"
#include "execution.h"
#include "global-handles.h"
#include "ic-inl.h"
#include "natives.h"
#include "platform.h"
#include "runtime.h"
#include "serialize.h"
#include "snapshot.h"
#include "stub-cache.h"
#include "v8threads.h"
namespace v8 {
namespace internal {
// -----------------------------------------------------------------------------
// Coding of external references.
// The encoding of an external reference. The type is in the high word.
// The id is in the low word.
static uint32_t EncodeExternal(TypeCode type, uint16_t id) {
return static_cast<uint32_t>(type) << 16 | id;
}
static int* GetInternalPointer(StatsCounter* counter) {
// All counters refer to dummy_counter, if deserializing happens without
// setting up counters.
static int dummy_counter = 0;
return counter->Enabled() ? counter->GetInternalPointer() : &dummy_counter;
}
ExternalReferenceTable* ExternalReferenceTable::instance(Isolate* isolate) {
ExternalReferenceTable* external_reference_table =
isolate->external_reference_table();
if (external_reference_table == NULL) {
external_reference_table = new ExternalReferenceTable(isolate);
isolate->set_external_reference_table(external_reference_table);
}
return external_reference_table;
}
void ExternalReferenceTable::AddFromId(TypeCode type,
uint16_t id,
const char* name,
Isolate* isolate) {
Address address;
switch (type) {
case C_BUILTIN: {
ExternalReference ref(static_cast<Builtins::CFunctionId>(id), isolate);
address = ref.address();
break;
}
case BUILTIN: {
ExternalReference ref(static_cast<Builtins::Name>(id), isolate);
address = ref.address();
break;
}
case RUNTIME_FUNCTION: {
ExternalReference ref(static_cast<Runtime::FunctionId>(id), isolate);
address = ref.address();
break;
}
case IC_UTILITY: {
ExternalReference ref(IC_Utility(static_cast<IC::UtilityId>(id)),
isolate);
address = ref.address();
break;
}
default:
UNREACHABLE();
return;
}
Add(address, type, id, name);
}
void ExternalReferenceTable::Add(Address address,
TypeCode type,
uint16_t id,
const char* name) {
ASSERT_NE(NULL, address);
ExternalReferenceEntry entry;
entry.address = address;
entry.code = EncodeExternal(type, id);
entry.name = name;
ASSERT_NE(0, entry.code);
refs_.Add(entry);
if (id > max_id_[type]) max_id_[type] = id;
}
void ExternalReferenceTable::PopulateTable(Isolate* isolate) {
for (int type_code = 0; type_code < kTypeCodeCount; type_code++) {
max_id_[type_code] = 0;
}
// The following populates all of the different type of external references
// into the ExternalReferenceTable.
//
// NOTE: This function was originally 100k of code. It has since been
// rewritten to be mostly table driven, as the callback macro style tends to
// very easily cause code bloat. Please be careful in the future when adding
// new references.
struct RefTableEntry {
TypeCode type;
uint16_t id;
const char* name;
};
static const RefTableEntry ref_table[] = {
// Builtins
#define DEF_ENTRY_C(name, ignored) \
{ C_BUILTIN, \
Builtins::c_##name, \
"Builtins::" #name },
BUILTIN_LIST_C(DEF_ENTRY_C)
#undef DEF_ENTRY_C
#define DEF_ENTRY_C(name, ignored) \
{ BUILTIN, \
Builtins::k##name, \
"Builtins::" #name },
#define DEF_ENTRY_A(name, kind, state, extra) DEF_ENTRY_C(name, ignored)
BUILTIN_LIST_C(DEF_ENTRY_C)
BUILTIN_LIST_A(DEF_ENTRY_A)
BUILTIN_LIST_DEBUG_A(DEF_ENTRY_A)
#undef DEF_ENTRY_C
#undef DEF_ENTRY_A
// Runtime functions
#define RUNTIME_ENTRY(name, nargs, ressize) \
{ RUNTIME_FUNCTION, \
Runtime::k##name, \
"Runtime::" #name },
RUNTIME_FUNCTION_LIST(RUNTIME_ENTRY)
#undef RUNTIME_ENTRY
// IC utilities
#define IC_ENTRY(name) \
{ IC_UTILITY, \
IC::k##name, \
"IC::" #name },
IC_UTIL_LIST(IC_ENTRY)
#undef IC_ENTRY
}; // end of ref_table[].
for (size_t i = 0; i < ARRAY_SIZE(ref_table); ++i) {
AddFromId(ref_table[i].type,
ref_table[i].id,
ref_table[i].name,
isolate);
}
#ifdef ENABLE_DEBUGGER_SUPPORT
// Debug addresses
Add(Debug_Address(Debug::k_after_break_target_address).address(isolate),
DEBUG_ADDRESS,
Debug::k_after_break_target_address << kDebugIdShift,
"Debug::after_break_target_address()");
Add(Debug_Address(Debug::k_debug_break_slot_address).address(isolate),
DEBUG_ADDRESS,
Debug::k_debug_break_slot_address << kDebugIdShift,
"Debug::debug_break_slot_address()");
Add(Debug_Address(Debug::k_debug_break_return_address).address(isolate),
DEBUG_ADDRESS,
Debug::k_debug_break_return_address << kDebugIdShift,
"Debug::debug_break_return_address()");
Add(Debug_Address(Debug::k_restarter_frame_function_pointer).address(isolate),
DEBUG_ADDRESS,
Debug::k_restarter_frame_function_pointer << kDebugIdShift,
"Debug::restarter_frame_function_pointer_address()");
#endif
// Stat counters
struct StatsRefTableEntry {
StatsCounter* (Counters::*counter)();
uint16_t id;
const char* name;
};
const StatsRefTableEntry stats_ref_table[] = {
#define COUNTER_ENTRY(name, caption) \
{ &Counters::name, \
Counters::k_##name, \
"Counters::" #name },
STATS_COUNTER_LIST_1(COUNTER_ENTRY)
STATS_COUNTER_LIST_2(COUNTER_ENTRY)
#undef COUNTER_ENTRY
}; // end of stats_ref_table[].
Counters* counters = isolate->counters();
for (size_t i = 0; i < ARRAY_SIZE(stats_ref_table); ++i) {
Add(reinterpret_cast<Address>(GetInternalPointer(
(counters->*(stats_ref_table[i].counter))())),
STATS_COUNTER,
stats_ref_table[i].id,
stats_ref_table[i].name);
}
// Top addresses
const char* AddressNames[] = {
#define BUILD_NAME_LITERAL(CamelName, hacker_name) \
"Isolate::" #hacker_name "_address",
FOR_EACH_ISOLATE_ADDRESS_NAME(BUILD_NAME_LITERAL)
NULL
#undef BUILD_NAME_LITERAL
};
for (uint16_t i = 0; i < Isolate::kIsolateAddressCount; ++i) {
Add(isolate->get_address_from_id((Isolate::AddressId)i),
TOP_ADDRESS, i, AddressNames[i]);
}
// Accessors
#define ACCESSOR_DESCRIPTOR_DECLARATION(name) \
Add((Address)&Accessors::name, \
ACCESSOR, \
Accessors::k##name, \
"Accessors::" #name);
ACCESSOR_DESCRIPTOR_LIST(ACCESSOR_DESCRIPTOR_DECLARATION)
#undef ACCESSOR_DESCRIPTOR_DECLARATION
StubCache* stub_cache = isolate->stub_cache();
// Stub cache tables
Add(stub_cache->key_reference(StubCache::kPrimary).address(),
STUB_CACHE_TABLE,
1,
"StubCache::primary_->key");
Add(stub_cache->value_reference(StubCache::kPrimary).address(),
STUB_CACHE_TABLE,
2,
"StubCache::primary_->value");
Add(stub_cache->map_reference(StubCache::kPrimary).address(),
STUB_CACHE_TABLE,
3,
"StubCache::primary_->map");
Add(stub_cache->key_reference(StubCache::kSecondary).address(),
STUB_CACHE_TABLE,
4,
"StubCache::secondary_->key");
Add(stub_cache->value_reference(StubCache::kSecondary).address(),
STUB_CACHE_TABLE,
5,
"StubCache::secondary_->value");
Add(stub_cache->map_reference(StubCache::kSecondary).address(),
STUB_CACHE_TABLE,
6,
"StubCache::secondary_->map");
// Runtime entries
Add(ExternalReference::perform_gc_function(isolate).address(),
RUNTIME_ENTRY,
1,
"Runtime::PerformGC");
Add(ExternalReference::fill_heap_number_with_random_function(
isolate).address(),
RUNTIME_ENTRY,
2,
"V8::FillHeapNumberWithRandom");
Add(ExternalReference::random_uint32_function(isolate).address(),
RUNTIME_ENTRY,
3,
"V8::Random");
Add(ExternalReference::delete_handle_scope_extensions(isolate).address(),
RUNTIME_ENTRY,
4,
"HandleScope::DeleteExtensions");
Add(ExternalReference::
incremental_marking_record_write_function(isolate).address(),
RUNTIME_ENTRY,
5,
"IncrementalMarking::RecordWrite");
Add(ExternalReference::store_buffer_overflow_function(isolate).address(),
RUNTIME_ENTRY,
6,
"StoreBuffer::StoreBufferOverflow");
Add(ExternalReference::
incremental_evacuation_record_write_function(isolate).address(),
RUNTIME_ENTRY,
7,
"IncrementalMarking::RecordWrite");
// Miscellaneous
Add(ExternalReference::roots_array_start(isolate).address(),
UNCLASSIFIED,
3,
"Heap::roots_array_start()");
Add(ExternalReference::address_of_stack_limit(isolate).address(),
UNCLASSIFIED,
4,
"StackGuard::address_of_jslimit()");
Add(ExternalReference::address_of_real_stack_limit(isolate).address(),
UNCLASSIFIED,
5,
"StackGuard::address_of_real_jslimit()");
#ifndef V8_INTERPRETED_REGEXP
Add(ExternalReference::address_of_regexp_stack_limit(isolate).address(),
UNCLASSIFIED,
6,
"RegExpStack::limit_address()");
Add(ExternalReference::address_of_regexp_stack_memory_address(
isolate).address(),
UNCLASSIFIED,
7,
"RegExpStack::memory_address()");
Add(ExternalReference::address_of_regexp_stack_memory_size(isolate).address(),
UNCLASSIFIED,
8,
"RegExpStack::memory_size()");
Add(ExternalReference::address_of_static_offsets_vector(isolate).address(),
UNCLASSIFIED,
9,
"OffsetsVector::static_offsets_vector");
#endif // V8_INTERPRETED_REGEXP
Add(ExternalReference::new_space_start(isolate).address(),
UNCLASSIFIED,
10,
"Heap::NewSpaceStart()");
Add(ExternalReference::new_space_mask(isolate).address(),
UNCLASSIFIED,
11,
"Heap::NewSpaceMask()");
Add(ExternalReference::heap_always_allocate_scope_depth(isolate).address(),
UNCLASSIFIED,
12,
"Heap::always_allocate_scope_depth()");
Add(ExternalReference::new_space_allocation_limit_address(isolate).address(),
UNCLASSIFIED,
14,
"Heap::NewSpaceAllocationLimitAddress()");
Add(ExternalReference::new_space_allocation_top_address(isolate).address(),
UNCLASSIFIED,
15,
"Heap::NewSpaceAllocationTopAddress()");
#ifdef ENABLE_DEBUGGER_SUPPORT
Add(ExternalReference::debug_break(isolate).address(),
UNCLASSIFIED,
16,
"Debug::Break()");
Add(ExternalReference::debug_step_in_fp_address(isolate).address(),
UNCLASSIFIED,
17,
"Debug::step_in_fp_addr()");
#endif
Add(ExternalReference::double_fp_operation(Token::ADD, isolate).address(),
UNCLASSIFIED,
18,
"add_two_doubles");
Add(ExternalReference::double_fp_operation(Token::SUB, isolate).address(),
UNCLASSIFIED,
19,
"sub_two_doubles");
Add(ExternalReference::double_fp_operation(Token::MUL, isolate).address(),
UNCLASSIFIED,
20,
"mul_two_doubles");
Add(ExternalReference::double_fp_operation(Token::DIV, isolate).address(),
UNCLASSIFIED,
21,
"div_two_doubles");
Add(ExternalReference::double_fp_operation(Token::MOD, isolate).address(),
UNCLASSIFIED,
22,
"mod_two_doubles");
Add(ExternalReference::compare_doubles(isolate).address(),
UNCLASSIFIED,
23,
"compare_doubles");
#ifndef V8_INTERPRETED_REGEXP
Add(ExternalReference::re_case_insensitive_compare_uc16(isolate).address(),
UNCLASSIFIED,
24,
"NativeRegExpMacroAssembler::CaseInsensitiveCompareUC16()");
Add(ExternalReference::re_check_stack_guard_state(isolate).address(),
UNCLASSIFIED,
25,
"RegExpMacroAssembler*::CheckStackGuardState()");
Add(ExternalReference::re_grow_stack(isolate).address(),
UNCLASSIFIED,
26,
"NativeRegExpMacroAssembler::GrowStack()");
Add(ExternalReference::re_word_character_map().address(),
UNCLASSIFIED,
27,
"NativeRegExpMacroAssembler::word_character_map");
#endif // V8_INTERPRETED_REGEXP
// Keyed lookup cache.
Add(ExternalReference::keyed_lookup_cache_keys(isolate).address(),
UNCLASSIFIED,
28,
"KeyedLookupCache::keys()");
Add(ExternalReference::keyed_lookup_cache_field_offsets(isolate).address(),
UNCLASSIFIED,
29,
"KeyedLookupCache::field_offsets()");
Add(ExternalReference::transcendental_cache_array_address(isolate).address(),
UNCLASSIFIED,
30,
"TranscendentalCache::caches()");
Add(ExternalReference::handle_scope_next_address().address(),
UNCLASSIFIED,
31,
"HandleScope::next");
Add(ExternalReference::handle_scope_limit_address().address(),
UNCLASSIFIED,
32,
"HandleScope::limit");
Add(ExternalReference::handle_scope_level_address().address(),
UNCLASSIFIED,
33,
"HandleScope::level");
Add(ExternalReference::new_deoptimizer_function(isolate).address(),
UNCLASSIFIED,
34,
"Deoptimizer::New()");
Add(ExternalReference::compute_output_frames_function(isolate).address(),
UNCLASSIFIED,
35,
"Deoptimizer::ComputeOutputFrames()");
Add(ExternalReference::address_of_min_int().address(),
UNCLASSIFIED,
36,
"LDoubleConstant::min_int");
Add(ExternalReference::address_of_one_half().address(),
UNCLASSIFIED,
37,
"LDoubleConstant::one_half");
Add(ExternalReference::isolate_address().address(),
UNCLASSIFIED,
38,
"isolate");
Add(ExternalReference::address_of_minus_zero().address(),
UNCLASSIFIED,
39,
"LDoubleConstant::minus_zero");
Add(ExternalReference::address_of_negative_infinity().address(),
UNCLASSIFIED,
40,
"LDoubleConstant::negative_infinity");
Add(ExternalReference::power_double_double_function(isolate).address(),
UNCLASSIFIED,
41,
"power_double_double_function");
Add(ExternalReference::power_double_int_function(isolate).address(),
UNCLASSIFIED,
42,
"power_double_int_function");
Add(ExternalReference::store_buffer_top(isolate).address(),
UNCLASSIFIED,
43,
"store_buffer_top");
Add(ExternalReference::address_of_canonical_non_hole_nan().address(),
UNCLASSIFIED,
44,
"canonical_nan");
Add(ExternalReference::address_of_the_hole_nan().address(),
UNCLASSIFIED,
45,
"the_hole_nan");
Add(ExternalReference::get_date_field_function(isolate).address(),
UNCLASSIFIED,
46,
"JSDate::GetField");
Add(ExternalReference::date_cache_stamp(isolate).address(),
UNCLASSIFIED,
47,
"date_cache_stamp");
Add(ExternalReference::address_of_pending_message_obj(isolate).address(),
UNCLASSIFIED,
48,
"address_of_pending_message_obj");
Add(ExternalReference::address_of_has_pending_message(isolate).address(),
UNCLASSIFIED,
49,
"address_of_has_pending_message");
Add(ExternalReference::address_of_pending_message_script(isolate).address(),
UNCLASSIFIED,
50,
"pending_message_script");
}
ExternalReferenceEncoder::ExternalReferenceEncoder()
: encodings_(Match),
isolate_(Isolate::Current()) {
ExternalReferenceTable* external_references =
ExternalReferenceTable::instance(isolate_);
for (int i = 0; i < external_references->size(); ++i) {
Put(external_references->address(i), i);
}
}
uint32_t ExternalReferenceEncoder::Encode(Address key) const {
int index = IndexOf(key);
ASSERT(key == NULL || index >= 0);
return index >=0 ?
ExternalReferenceTable::instance(isolate_)->code(index) : 0;
}
const char* ExternalReferenceEncoder::NameOfAddress(Address key) const {
int index = IndexOf(key);
return index >= 0 ?
ExternalReferenceTable::instance(isolate_)->name(index) : NULL;
}
int ExternalReferenceEncoder::IndexOf(Address key) const {
if (key == NULL) return -1;
HashMap::Entry* entry =
const_cast<HashMap&>(encodings_).Lookup(key, Hash(key), false);
return entry == NULL
? -1
: static_cast<int>(reinterpret_cast<intptr_t>(entry->value));
}
void ExternalReferenceEncoder::Put(Address key, int index) {
HashMap::Entry* entry = encodings_.Lookup(key, Hash(key), true);
entry->value = reinterpret_cast<void*>(index);
}
ExternalReferenceDecoder::ExternalReferenceDecoder()
: encodings_(NewArray<Address*>(kTypeCodeCount)),
isolate_(Isolate::Current()) {
ExternalReferenceTable* external_references =
ExternalReferenceTable::instance(isolate_);
for (int type = kFirstTypeCode; type < kTypeCodeCount; ++type) {
int max = external_references->max_id(type) + 1;
encodings_[type] = NewArray<Address>(max + 1);
}
for (int i = 0; i < external_references->size(); ++i) {
Put(external_references->code(i), external_references->address(i));
}
}
ExternalReferenceDecoder::~ExternalReferenceDecoder() {
for (int type = kFirstTypeCode; type < kTypeCodeCount; ++type) {
DeleteArray(encodings_[type]);
}
DeleteArray(encodings_);
}
bool Serializer::serialization_enabled_ = false;
bool Serializer::too_late_to_enable_now_ = false;
Deserializer::Deserializer(SnapshotByteSource* source)
: isolate_(NULL),
source_(source),
external_reference_decoder_(NULL) {
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
for (int i = 0; i < LAST_SPACE + 1; i++) {
reservations_[i] = kUninitializedReservation;
}
}
void Deserializer::Deserialize() {
isolate_ = Isolate::Current();
ASSERT(isolate_ != NULL);
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
isolate_->heap()->ReserveSpace(reservations_, &high_water_[0]);
// No active threads.
ASSERT_EQ(NULL, isolate_->thread_manager()->FirstThreadStateInUse());
// No active handles.
ASSERT(isolate_->handle_scope_implementer()->blocks()->is_empty());
ASSERT_EQ(NULL, external_reference_decoder_);
external_reference_decoder_ = new ExternalReferenceDecoder();
isolate_->heap()->IterateStrongRoots(this, VISIT_ONLY_STRONG);
isolate_->heap()->RepairFreeListsAfterBoot();
isolate_->heap()->IterateWeakRoots(this, VISIT_ALL);
isolate_->heap()->set_native_contexts_list(
isolate_->heap()->undefined_value());
// Update data pointers to the external strings containing natives sources.
for (int i = 0; i < Natives::GetBuiltinsCount(); i++) {
Object* source = isolate_->heap()->natives_source_cache()->get(i);
if (!source->IsUndefined()) {
ExternalAsciiString::cast(source)->update_data_cache();
}
}
// Issue code events for newly deserialized code objects.
LOG_CODE_EVENT(isolate_, LogCodeObjects());
LOG_CODE_EVENT(isolate_, LogCompiledFunctions());
}
void Deserializer::DeserializePartial(Object** root) {
isolate_ = Isolate::Current();
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
for (int i = NEW_SPACE; i < kNumberOfSpaces; i++) {
ASSERT(reservations_[i] != kUninitializedReservation);
}
isolate_->heap()->ReserveSpace(reservations_, &high_water_[0]);
if (external_reference_decoder_ == NULL) {
external_reference_decoder_ = new ExternalReferenceDecoder();
}
// Keep track of the code space start and end pointers in case new
// code objects were unserialized
OldSpace* code_space = isolate_->heap()->code_space();
Address start_address = code_space->top();
VisitPointer(root);
// There's no code deserialized here. If this assert fires
// then that's changed and logging should be added to notify
// the profiler et al of the new code.
CHECK_EQ(start_address, code_space->top());
}
Deserializer::~Deserializer() {
ASSERT(source_->AtEOF());
if (external_reference_decoder_) {
delete external_reference_decoder_;
external_reference_decoder_ = NULL;
}
}
// This is called on the roots. It is the driver of the deserialization
// process. It is also called on the body of each function.
void Deserializer::VisitPointers(Object** start, Object** end) {
// The space must be new space. Any other space would cause ReadChunk to try
// to update the remembered using NULL as the address.
ReadChunk(start, end, NEW_SPACE, NULL);
}
// This routine writes the new object into the pointer provided and then
// returns true if the new object was in young space and false otherwise.
// The reason for this strange interface is that otherwise the object is
// written very late, which means the FreeSpace map is not set up by the
// time we need to use it to mark the space at the end of a page free.
void Deserializer::ReadObject(int space_number,
Object** write_back) {
int size = source_->GetInt() << kObjectAlignmentBits;
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
Address address = Allocate(space_number, size);
*write_back = HeapObject::FromAddress(address);
Object** current = reinterpret_cast<Object**>(address);
Object** limit = current + (size >> kPointerSizeLog2);
if (FLAG_log_snapshot_positions) {
LOG(isolate_, SnapshotPositionEvent(address, source_->position()));
}
ReadChunk(current, limit, space_number, address);
#ifdef DEBUG
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
bool is_codespace = (space_number == CODE_SPACE);
ASSERT(HeapObject::FromAddress(address)->IsCode() == is_codespace);
#endif
}
void Deserializer::ReadChunk(Object** current,
Object** limit,
int source_space,
Address current_object_address) {
Isolate* const isolate = isolate_;
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
// Write barrier support costs around 1% in startup time. In fact there
// are no new space objects in current boot snapshots, so it's not needed,
// but that may change.
bool write_barrier_needed = (current_object_address != NULL &&
source_space != NEW_SPACE &&
source_space != CELL_SPACE &&
source_space != CODE_SPACE &&
source_space != OLD_DATA_SPACE);
while (current < limit) {
int data = source_->Get();
switch (data) {
#define CASE_STATEMENT(where, how, within, space_number) \
case where + how + within + space_number: \
ASSERT((where & ~kPointedToMask) == 0); \
ASSERT((how & ~kHowToCodeMask) == 0); \
ASSERT((within & ~kWhereToPointMask) == 0); \
ASSERT((space_number & ~kSpaceMask) == 0);
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
#define CASE_BODY(where, how, within, space_number_if_any) \
{ \
bool emit_write_barrier = false; \
bool current_was_incremented = false; \
int space_number = space_number_if_any == kAnyOldSpace ? \
(data & kSpaceMask) : space_number_if_any; \
if (where == kNewObject && how == kPlain && within == kStartOfObject) {\
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
ReadObject(space_number, current); \
emit_write_barrier = (space_number == NEW_SPACE); \
} else { \
Object* new_object = NULL; /* May not be a real Object pointer. */ \
if (where == kNewObject) { \
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
ReadObject(space_number, &new_object); \
} else if (where == kRootArray) { \
int root_id = source_->GetInt(); \
new_object = isolate->heap()->roots_array_start()[root_id]; \
emit_write_barrier = isolate->heap()->InNewSpace(new_object); \
} else if (where == kPartialSnapshotCache) { \
int cache_index = source_->GetInt(); \
new_object = isolate->serialize_partial_snapshot_cache() \
[cache_index]; \
emit_write_barrier = isolate->heap()->InNewSpace(new_object); \
} else if (where == kExternalReference) { \
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
int skip = source_->GetInt(); \
current = reinterpret_cast<Object**>(reinterpret_cast<Address>( \
current) + skip); \
int reference_id = source_->GetInt(); \
Address address = external_reference_decoder_-> \
Decode(reference_id); \
new_object = reinterpret_cast<Object*>(address); \
} else if (where == kBackref) { \
emit_write_barrier = (space_number == NEW_SPACE); \
new_object = GetAddressFromEnd(data & kSpaceMask); \
} else { \
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
ASSERT(where == kBackrefWithSkip); \
int skip = source_->GetInt(); \
current = reinterpret_cast<Object**>( \
reinterpret_cast<Address>(current) + skip); \
emit_write_barrier = (space_number == NEW_SPACE); \
new_object = GetAddressFromEnd(data & kSpaceMask); \
} \
if (within == kInnerPointer) { \
if (space_number != CODE_SPACE || new_object->IsCode()) { \
Code* new_code_object = reinterpret_cast<Code*>(new_object); \
new_object = reinterpret_cast<Object*>( \
new_code_object->instruction_start()); \
} else { \
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
ASSERT(space_number == CODE_SPACE); \
JSGlobalPropertyCell* cell = \
JSGlobalPropertyCell::cast(new_object); \
new_object = reinterpret_cast<Object*>( \
cell->ValueAddress()); \
} \
} \
if (how == kFromCode) { \
Address location_of_branch_data = \
reinterpret_cast<Address>(current); \
Assembler::deserialization_set_special_target_at( \
location_of_branch_data, \
reinterpret_cast<Address>(new_object)); \
location_of_branch_data += Assembler::kSpecialTargetSize; \
current = reinterpret_cast<Object**>(location_of_branch_data); \
current_was_incremented = true; \
} else { \
*current = new_object; \
} \
} \
if (emit_write_barrier && write_barrier_needed) { \
Address current_address = reinterpret_cast<Address>(current); \
isolate->heap()->RecordWrite( \
current_object_address, \
static_cast<int>(current_address - current_object_address)); \
} \
if (!current_was_incremented) { \
current++; \
} \
break; \
} \
// This generates a case and a body for the new space (which has to do extra
// write barrier handling) and handles the other spaces with 8 fall-through
// cases and one body.
#define ALL_SPACES(where, how, within) \
CASE_STATEMENT(where, how, within, NEW_SPACE) \
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
CASE_BODY(where, how, within, NEW_SPACE) \
CASE_STATEMENT(where, how, within, OLD_DATA_SPACE) \
CASE_STATEMENT(where, how, within, OLD_POINTER_SPACE) \
CASE_STATEMENT(where, how, within, CODE_SPACE) \
CASE_STATEMENT(where, how, within, CELL_SPACE) \
CASE_STATEMENT(where, how, within, MAP_SPACE) \
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
CASE_BODY(where, how, within, kAnyOldSpace)
#define FOUR_CASES(byte_code) \
case byte_code: \
case byte_code + 1: \
case byte_code + 2: \
case byte_code + 3:
#define SIXTEEN_CASES(byte_code) \
FOUR_CASES(byte_code) \
FOUR_CASES(byte_code + 4) \
FOUR_CASES(byte_code + 8) \
FOUR_CASES(byte_code + 12)
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
#define COMMON_RAW_LENGTHS(f) \
f(1) \
f(2) \
f(3) \
f(4) \
f(5) \
f(6) \
f(7) \
f(8) \
f(9) \
f(10) \
f(11) \
f(12) \
f(13) \
f(14) \
f(15) \
f(16) \
f(17) \
f(18) \
f(19) \
f(20) \
f(21) \
f(22) \
f(23) \
f(24) \
f(25) \
f(26) \
f(27) \
f(28) \
f(29) \
f(30) \
f(31)
// We generate 15 cases and bodies that process special tags that combine
// the raw data tag and the length into one byte.
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
#define RAW_CASE(index) \
case kRawData + index: { \
byte* raw_data_out = reinterpret_cast<byte*>(current); \
source_->CopyRaw(raw_data_out, index * kPointerSize); \
current = \
reinterpret_cast<Object**>(raw_data_out + index * kPointerSize); \
break; \
}
COMMON_RAW_LENGTHS(RAW_CASE)
#undef RAW_CASE
// Deserialize a chunk of raw data that doesn't have one of the popular
// lengths.
case kRawData: {
int size = source_->GetInt();
byte* raw_data_out = reinterpret_cast<byte*>(current);
source_->CopyRaw(raw_data_out, size);
break;
}
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
SIXTEEN_CASES(kRootArrayConstants + kNoSkipDistance)
SIXTEEN_CASES(kRootArrayConstants + kNoSkipDistance + 16) {
int root_id = RootArrayConstantFromByteCode(data);
Object* object = isolate->heap()->roots_array_start()[root_id];
ASSERT(!isolate->heap()->InNewSpace(object));
*current++ = object;
break;
}
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
SIXTEEN_CASES(kRootArrayConstants + kHasSkipDistance)
SIXTEEN_CASES(kRootArrayConstants + kHasSkipDistance + 16) {
int root_id = RootArrayConstantFromByteCode(data);
int skip = source_->GetInt();
current = reinterpret_cast<Object**>(
reinterpret_cast<intptr_t>(current) + skip);
Object* object = isolate->heap()->roots_array_start()[root_id];
ASSERT(!isolate->heap()->InNewSpace(object));
*current++ = object;
break;
}
case kRepeat: {
int repeats = source_->GetInt();
Object* object = current[-1];
ASSERT(!isolate->heap()->InNewSpace(object));
for (int i = 0; i < repeats; i++) current[i] = object;
current += repeats;
break;
}
STATIC_ASSERT(kRootArrayNumberOfConstantEncodings ==
Heap::kOldSpaceRoots);
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
STATIC_ASSERT(kMaxRepeats == 13);
case kConstantRepeat:
FOUR_CASES(kConstantRepeat + 1)
FOUR_CASES(kConstantRepeat + 5)
FOUR_CASES(kConstantRepeat + 9) {
int repeats = RepeatsForCode(data);
Object* object = current[-1];
ASSERT(!isolate->heap()->InNewSpace(object));
for (int i = 0; i < repeats; i++) current[i] = object;
current += repeats;
break;
}
// Deserialize a new object and write a pointer to it to the current
// object.
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
ALL_SPACES(kNewObject, kPlain, kStartOfObject)
// Support for direct instruction pointers in functions. It's an inner
// pointer because it points at the entry point, not at the start of the
// code object.
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
CASE_STATEMENT(kNewObject, kPlain, kInnerPointer, CODE_SPACE)
CASE_BODY(kNewObject, kPlain, kInnerPointer, CODE_SPACE)
// Deserialize a new code object and write a pointer to its first
// instruction to the current code object.
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
ALL_SPACES(kNewObject, kFromCode, kInnerPointer)
// Find a recently deserialized object using its offset from the current
// allocation point and write a pointer to it to the current object.
ALL_SPACES(kBackref, kPlain, kStartOfObject)
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
ALL_SPACES(kBackrefWithSkip, kPlain, kStartOfObject)
#if V8_TARGET_ARCH_MIPS
// Deserialize a new object from pointer found in code and write
// a pointer to it to the current object. Required only for MIPS, and
// omitted on the other architectures because it is fully unrolled and
// would cause bloat.
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
ALL_SPACES(kNewObject, kFromCode, kStartOfObject)
// Find a recently deserialized code object using its offset from the
// current allocation point and write a pointer to it to the current
// object. Required only for MIPS.
ALL_SPACES(kBackref, kFromCode, kStartOfObject)
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
ALL_SPACES(kBackrefWithSkip, kFromCode, kStartOfObject)
#endif
// Find a recently deserialized code object using its offset from the
// current allocation point and write a pointer to its first instruction
// to the current code object or the instruction pointer in a function
// object.
ALL_SPACES(kBackref, kFromCode, kInnerPointer)
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
ALL_SPACES(kBackrefWithSkip, kFromCode, kInnerPointer)
ALL_SPACES(kBackref, kPlain, kInnerPointer)
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
ALL_SPACES(kBackrefWithSkip, kPlain, kInnerPointer)
// Find an object in the roots array and write a pointer to it to the
// current object.
CASE_STATEMENT(kRootArray, kPlain, kStartOfObject, 0)
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
CASE_BODY(kRootArray, kPlain, kStartOfObject, 0)
// Find an object in the partial snapshots cache and write a pointer to it
// to the current object.
CASE_STATEMENT(kPartialSnapshotCache, kPlain, kStartOfObject, 0)
CASE_BODY(kPartialSnapshotCache,
kPlain,
kStartOfObject,
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
0)
// Find an code entry in the partial snapshots cache and
// write a pointer to it to the current object.
CASE_STATEMENT(kPartialSnapshotCache, kPlain, kInnerPointer, 0)
CASE_BODY(kPartialSnapshotCache,
kPlain,
kInnerPointer,
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
0)
// Find an external reference and write a pointer to it to the current
// object.
CASE_STATEMENT(kExternalReference, kPlain, kStartOfObject, 0)
CASE_BODY(kExternalReference,
kPlain,
kStartOfObject,
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
0)
// Find an external reference and write a pointer to it in the current
// code object.
CASE_STATEMENT(kExternalReference, kFromCode, kStartOfObject, 0)
CASE_BODY(kExternalReference,
kFromCode,
kStartOfObject,
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
0)
#undef CASE_STATEMENT
#undef CASE_BODY
#undef ALL_SPACES
case kSkip: {
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
int size = source_->GetInt();
current = reinterpret_cast<Object**>(
reinterpret_cast<intptr_t>(current) + size);
break;
}
case kNativesStringResource: {
int index = source_->Get();
Vector<const char> source_vector = Natives::GetRawScriptSource(index);
NativesExternalStringResource* resource =
new NativesExternalStringResource(isolate->bootstrapper(),
source_vector.start(),
source_vector.length());
*current++ = reinterpret_cast<Object*>(resource);
break;
}
case kSynchronize: {
// If we get here then that indicates that you have a mismatch between
// the number of GC roots when serializing and deserializing.
UNREACHABLE();
}
default:
UNREACHABLE();
}
}
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
ASSERT_EQ(limit, current);
}
void SnapshotByteSink::PutInt(uintptr_t integer, const char* description) {
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
ASSERT(integer < 1 << 22);
integer <<= 2;
int bytes = 1;
if (integer > 0xff) bytes = 2;
if (integer > 0xffff) bytes = 3;
integer |= bytes;
Put(static_cast<int>(integer & 0xff), "IntPart1");
if (bytes > 1) Put(static_cast<int>((integer >> 8) & 0xff), "IntPart2");
if (bytes > 2) Put(static_cast<int>((integer >> 16) & 0xff), "IntPart3");
}
Serializer::Serializer(SnapshotByteSink* sink)
: sink_(sink),
current_root_index_(0),
external_reference_encoder_(new ExternalReferenceEncoder),
root_index_wave_front_(0) {
isolate_ = Isolate::Current();
// The serializer is meant to be used only to generate initial heap images
// from a context in which there is only one isolate.
ASSERT(isolate_->IsDefaultIsolate());
for (int i = 0; i <= LAST_SPACE; i++) {
fullness_[i] = 0;
}
}
Serializer::~Serializer() {
delete external_reference_encoder_;
}
void StartupSerializer::SerializeStrongReferences() {
Isolate* isolate = Isolate::Current();
// No active threads.
CHECK_EQ(NULL, Isolate::Current()->thread_manager()->FirstThreadStateInUse());
// No active or weak handles.
CHECK(isolate->handle_scope_implementer()->blocks()->is_empty());
CHECK_EQ(0, isolate->global_handles()->NumberOfWeakHandles());
// We don't support serializing installed extensions.
CHECK(!isolate->has_installed_extensions());
HEAP->IterateStrongRoots(this, VISIT_ONLY_STRONG);
}
void PartialSerializer::Serialize(Object** object) {
this->VisitPointer(object);
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
Pad();
}
void Serializer::VisitPointers(Object** start, Object** end) {
Isolate* isolate = Isolate::Current();
for (Object** current = start; current < end; current++) {
if (start == isolate->heap()->roots_array_start()) {
root_index_wave_front_ =
Max(root_index_wave_front_, static_cast<intptr_t>(current - start));
}
if (reinterpret_cast<Address>(current) ==
isolate->heap()->store_buffer()->TopAddress()) {
sink_->Put(kSkip, "Skip");
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
sink_->PutInt(kPointerSize, "SkipOneWord");
} else if ((*current)->IsSmi()) {
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
sink_->Put(kRawData + 1, "Smi");
for (int i = 0; i < kPointerSize; i++) {
sink_->Put(reinterpret_cast<byte*>(current)[i], "Byte");
}
} else {
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
SerializeObject(*current, kPlain, kStartOfObject, 0);
}
}
}
// This ensures that the partial snapshot cache keeps things alive during GC and
// tracks their movement. When it is called during serialization of the startup
// snapshot nothing happens. When the partial (context) snapshot is created,
// this array is populated with the pointers that the partial snapshot will
// need. As that happens we emit serialized objects to the startup snapshot
// that correspond to the elements of this cache array. On deserialization we
// therefore need to visit the cache array. This fills it up with pointers to
// deserialized objects.
void SerializerDeserializer::Iterate(ObjectVisitor* visitor) {
if (Serializer::enabled()) return;
Isolate* isolate = Isolate::Current();
for (int i = 0; ; i++) {
if (isolate->serialize_partial_snapshot_cache_length() <= i) {
// Extend the array ready to get a value from the visitor when
// deserializing.
isolate->PushToPartialSnapshotCache(Smi::FromInt(0));
}
Object** cache = isolate->serialize_partial_snapshot_cache();
visitor->VisitPointers(&cache[i], &cache[i + 1]);
// Sentinel is the undefined object, which is a root so it will not normally
// be found in the cache.
if (cache[i] == isolate->heap()->undefined_value()) {
break;
}
}
}
int PartialSerializer::PartialSnapshotCacheIndex(HeapObject* heap_object) {
Isolate* isolate = Isolate::Current();
for (int i = 0;
i < isolate->serialize_partial_snapshot_cache_length();
i++) {
Object* entry = isolate->serialize_partial_snapshot_cache()[i];
if (entry == heap_object) return i;
}
// We didn't find the object in the cache. So we add it to the cache and
// then visit the pointer so that it becomes part of the startup snapshot
// and we can refer to it from the partial snapshot.
int length = isolate->serialize_partial_snapshot_cache_length();
isolate->PushToPartialSnapshotCache(heap_object);
startup_serializer_->VisitPointer(reinterpret_cast<Object**>(&heap_object));
// We don't recurse from the startup snapshot generator into the partial
// snapshot generator.
ASSERT(length == isolate->serialize_partial_snapshot_cache_length() - 1);
return length;
}
int Serializer::RootIndex(HeapObject* heap_object, HowToCode from) {
Heap* heap = HEAP;
if (heap->InNewSpace(heap_object)) return kInvalidRootIndex;
for (int i = 0; i < root_index_wave_front_; i++) {
Object* root = heap->roots_array_start()[i];
if (!root->IsSmi() && root == heap_object) {
#if V8_TARGET_ARCH_MIPS
if (from == kFromCode) {
// In order to avoid code bloat in the deserializer we don't have
// support for the encoding that specifies a particular root should
// be written into the lui/ori instructions on MIPS. Therefore we
// should not generate such serialization data for MIPS.
return kInvalidRootIndex;
}
#endif
return i;
}
}
return kInvalidRootIndex;
}
// Encode the location of an already deserialized object in order to write its
// location into a later object. We can encode the location as an offset from
// the start of the deserialized objects or as an offset backwards from the
// current allocation pointer.
void Serializer::SerializeReferenceToPreviousObject(
int space,
int address,
HowToCode how_to_code,
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
WhereToPoint where_to_point,
int skip) {
int offset = CurrentAllocationAddress(space) - address;
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
// Shift out the bits that are always 0.
offset >>= kObjectAlignmentBits;
if (skip == 0) {
sink_->Put(kBackref + how_to_code + where_to_point + space, "BackRefSer");
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
} else {
sink_->Put(kBackrefWithSkip + how_to_code + where_to_point + space,
"BackRefSerWithSkip");
sink_->PutInt(skip, "BackRefSkipDistance");
}
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
sink_->PutInt(offset, "offset");
}
void StartupSerializer::SerializeObject(
Object* o,
HowToCode how_to_code,
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
WhereToPoint where_to_point,
int skip) {
CHECK(o->IsHeapObject());
HeapObject* heap_object = HeapObject::cast(o);
int root_index;
if ((root_index = RootIndex(heap_object, how_to_code)) != kInvalidRootIndex) {
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
PutRoot(root_index, heap_object, how_to_code, where_to_point, skip);
return;
}
if (address_mapper_.IsMapped(heap_object)) {
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
int space = SpaceOfObject(heap_object);
int address = address_mapper_.MappedTo(heap_object);
SerializeReferenceToPreviousObject(space,
address,
how_to_code,
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
where_to_point,
skip);
} else {
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
if (skip != 0) {
sink_->Put(kSkip, "FlushPendingSkip");
sink_->PutInt(skip, "SkipDistance");
}
// Object has not yet been serialized. Serialize it here.
ObjectSerializer object_serializer(this,
heap_object,
sink_,
how_to_code,
where_to_point);
object_serializer.Serialize();
}
}
void StartupSerializer::SerializeWeakReferences() {
// This phase comes right after the partial serialization (of the snapshot).
// After we have done the partial serialization the partial snapshot cache
// will contain some references needed to decode the partial snapshot. We
// add one entry with 'undefined' which is the sentinel that the deserializer
// uses to know it is done deserializing the array.
Isolate* isolate = Isolate::Current();
Object* undefined = isolate->heap()->undefined_value();
VisitPointer(&undefined);
HEAP->IterateWeakRoots(this, VISIT_ALL);
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
Pad();
}
void Serializer::PutRoot(int root_index,
HeapObject* object,
SerializerDeserializer::HowToCode how_to_code,
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
SerializerDeserializer::WhereToPoint where_to_point,
int skip) {
if (how_to_code == kPlain &&
where_to_point == kStartOfObject &&
root_index < kRootArrayNumberOfConstantEncodings &&
!HEAP->InNewSpace(object)) {
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
if (skip == 0) {
sink_->Put(kRootArrayConstants + kNoSkipDistance + root_index,
"RootConstant");
} else {
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
sink_->Put(kRootArrayConstants + kHasSkipDistance + root_index,
"RootConstant");
sink_->PutInt(skip, "SkipInPutRoot");
}
} else {
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
if (skip != 0) {
sink_->Put(kSkip, "SkipFromPutRoot");
sink_->PutInt(skip, "SkipFromPutRootDistance");
}
sink_->Put(kRootArray + how_to_code + where_to_point, "RootSerialization");
sink_->PutInt(root_index, "root_index");
}
}
void PartialSerializer::SerializeObject(
Object* o,
HowToCode how_to_code,
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
WhereToPoint where_to_point,
int skip) {
CHECK(o->IsHeapObject());
HeapObject* heap_object = HeapObject::cast(o);
if (heap_object->IsMap()) {
// The code-caches link to context-specific code objects, which
// the startup and context serializes cannot currently handle.
ASSERT(Map::cast(heap_object)->code_cache() ==
heap_object->GetHeap()->empty_fixed_array());
}
int root_index;
if ((root_index = RootIndex(heap_object, how_to_code)) != kInvalidRootIndex) {
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
PutRoot(root_index, heap_object, how_to_code, where_to_point, skip);
return;
}
if (ShouldBeInThePartialSnapshotCache(heap_object)) {
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
if (skip != 0) {
sink_->Put(kSkip, "SkipFromSerializeObject");
sink_->PutInt(skip, "SkipDistanceFromSerializeObject");
}
int cache_index = PartialSnapshotCacheIndex(heap_object);
sink_->Put(kPartialSnapshotCache + how_to_code + where_to_point,
"PartialSnapshotCache");
sink_->PutInt(cache_index, "partial_snapshot_cache_index");
return;
}
// Pointers from the partial snapshot to the objects in the startup snapshot
// should go through the root array or through the partial snapshot cache.
// If this is not the case you may have to add something to the root array.
ASSERT(!startup_serializer_->address_mapper()->IsMapped(heap_object));
// All the symbols that the partial snapshot needs should be either in the
// root table or in the partial snapshot cache.
ASSERT(!heap_object->IsSymbol());
if (address_mapper_.IsMapped(heap_object)) {
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
int space = SpaceOfObject(heap_object);
int address = address_mapper_.MappedTo(heap_object);
SerializeReferenceToPreviousObject(space,
address,
how_to_code,
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
where_to_point,
skip);
} else {
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
if (skip != 0) {
sink_->Put(kSkip, "SkipFromSerializeObject");
sink_->PutInt(skip, "SkipDistanceFromSerializeObject");
}
// Object has not yet been serialized. Serialize it here.
ObjectSerializer serializer(this,
heap_object,
sink_,
how_to_code,
where_to_point);
serializer.Serialize();
}
}
void Serializer::ObjectSerializer::Serialize() {
int space = Serializer::SpaceOfObject(object_);
int size = object_->Size();
sink_->Put(kNewObject + reference_representation_ + space,
"ObjectSerialization");
sink_->PutInt(size >> kObjectAlignmentBits, "Size in words");
LOG(i::Isolate::Current(),
SnapshotPositionEvent(object_->address(), sink_->Position()));
// Mark this object as already serialized.
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
int offset = serializer_->Allocate(space, size);
serializer_->address_mapper()->AddMapping(object_, offset);
// Serialize the map (first word of the object).
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
serializer_->SerializeObject(object_->map(), kPlain, kStartOfObject, 0);
// Serialize the rest of the object.
CHECK_EQ(0, bytes_processed_so_far_);
bytes_processed_so_far_ = kPointerSize;
object_->IterateBody(object_->map()->instance_type(), size, this);
OutputRawData(object_->address() + size);
}
void Serializer::ObjectSerializer::VisitPointers(Object** start,
Object** end) {
Object** current = start;
while (current < end) {
while (current < end && (*current)->IsSmi()) current++;
if (current < end) OutputRawData(reinterpret_cast<Address>(current));
while (current < end && !(*current)->IsSmi()) {
HeapObject* current_contents = HeapObject::cast(*current);
int root_index = serializer_->RootIndex(current_contents, kPlain);
// Repeats are not subject to the write barrier so there are only some
// objects that can be used in a repeat encoding. These are the early
// ones in the root array that are never in new space.
if (current != start &&
root_index != kInvalidRootIndex &&
root_index < kRootArrayNumberOfConstantEncodings &&
current_contents == current[-1]) {
ASSERT(!HEAP->InNewSpace(current_contents));
int repeat_count = 1;
while (current < end - 1 && current[repeat_count] == current_contents) {
repeat_count++;
}
current += repeat_count;
bytes_processed_so_far_ += repeat_count * kPointerSize;
if (repeat_count > kMaxRepeats) {
sink_->Put(kRepeat, "SerializeRepeats");
sink_->PutInt(repeat_count, "SerializeRepeats");
} else {
sink_->Put(CodeForRepeats(repeat_count), "SerializeRepeats");
}
} else {
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
serializer_->SerializeObject(
current_contents, kPlain, kStartOfObject, 0);
bytes_processed_so_far_ += kPointerSize;
current++;
}
}
}
}
void Serializer::ObjectSerializer::VisitEmbeddedPointer(RelocInfo* rinfo) {
Object** current = rinfo->target_object_address();
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
int skip = OutputRawData(rinfo->target_address_address(),
kCanReturnSkipInsteadOfSkipping);
HowToCode representation = rinfo->IsCodedSpecially() ? kFromCode : kPlain;
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
serializer_->SerializeObject(*current, representation, kStartOfObject, skip);
bytes_processed_so_far_ += rinfo->target_address_size();
}
void Serializer::ObjectSerializer::VisitExternalReferences(Address* start,
Address* end) {
Address references_start = reinterpret_cast<Address>(start);
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
int skip = OutputRawData(references_start, kCanReturnSkipInsteadOfSkipping);
for (Address* current = start; current < end; current++) {
sink_->Put(kExternalReference + kPlain + kStartOfObject, "ExternalRef");
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
sink_->PutInt(skip, "SkipB4ExternalRef");
skip = 0;
int reference_id = serializer_->EncodeExternalReference(*current);
sink_->PutInt(reference_id, "reference id");
}
bytes_processed_so_far_ += static_cast<int>((end - start) * kPointerSize);
}
void Serializer::ObjectSerializer::VisitExternalReference(RelocInfo* rinfo) {
Address references_start = rinfo->target_address_address();
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
int skip = OutputRawData(references_start, kCanReturnSkipInsteadOfSkipping);
Address* current = rinfo->target_reference_address();
int representation = rinfo->IsCodedSpecially() ?
kFromCode + kStartOfObject : kPlain + kStartOfObject;
sink_->Put(kExternalReference + representation, "ExternalRef");
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
sink_->PutInt(skip, "SkipB4ExternalRef");
int reference_id = serializer_->EncodeExternalReference(*current);
sink_->PutInt(reference_id, "reference id");
bytes_processed_so_far_ += rinfo->target_address_size();
}
void Serializer::ObjectSerializer::VisitRuntimeEntry(RelocInfo* rinfo) {
Address target_start = rinfo->target_address_address();
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
int skip = OutputRawData(target_start, kCanReturnSkipInsteadOfSkipping);
Address target = rinfo->target_address();
uint32_t encoding = serializer_->EncodeExternalReference(target);
CHECK(target == NULL ? encoding == 0 : encoding != 0);
int representation;
// Can't use a ternary operator because of gcc.
if (rinfo->IsCodedSpecially()) {
representation = kStartOfObject + kFromCode;
} else {
representation = kStartOfObject + kPlain;
}
sink_->Put(kExternalReference + representation, "ExternalReference");
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
sink_->PutInt(skip, "SkipB4ExternalRef");
sink_->PutInt(encoding, "reference id");
bytes_processed_so_far_ += rinfo->target_address_size();
}
void Serializer::ObjectSerializer::VisitCodeTarget(RelocInfo* rinfo) {
CHECK(RelocInfo::IsCodeTarget(rinfo->rmode()));
Address target_start = rinfo->target_address_address();
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
int skip = OutputRawData(target_start, kCanReturnSkipInsteadOfSkipping);
Code* target = Code::GetCodeFromTargetAddress(rinfo->target_address());
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
serializer_->SerializeObject(target, kFromCode, kInnerPointer, skip);
bytes_processed_so_far_ += rinfo->target_address_size();
}
void Serializer::ObjectSerializer::VisitCodeEntry(Address entry_address) {
Code* target = Code::cast(Code::GetObjectFromEntryAddress(entry_address));
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
int skip = OutputRawData(entry_address, kCanReturnSkipInsteadOfSkipping);
serializer_->SerializeObject(target, kPlain, kInnerPointer, skip);
bytes_processed_so_far_ += kPointerSize;
}
void Serializer::ObjectSerializer::VisitGlobalPropertyCell(RelocInfo* rinfo) {
ASSERT(rinfo->rmode() == RelocInfo::GLOBAL_PROPERTY_CELL);
JSGlobalPropertyCell* cell =
JSGlobalPropertyCell::cast(rinfo->target_cell());
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
int skip = OutputRawData(rinfo->pc(), kCanReturnSkipInsteadOfSkipping);
serializer_->SerializeObject(cell, kPlain, kInnerPointer, skip);
}
void Serializer::ObjectSerializer::VisitExternalAsciiString(
v8::String::ExternalAsciiStringResource** resource_pointer) {
Address references_start = reinterpret_cast<Address>(resource_pointer);
OutputRawData(references_start);
for (int i = 0; i < Natives::GetBuiltinsCount(); i++) {
Object* source = HEAP->natives_source_cache()->get(i);
if (!source->IsUndefined()) {
ExternalAsciiString* string = ExternalAsciiString::cast(source);
typedef v8::String::ExternalAsciiStringResource Resource;
const Resource* resource = string->resource();
if (resource == *resource_pointer) {
sink_->Put(kNativesStringResource, "NativesStringResource");
sink_->PutSection(i, "NativesStringResourceEnd");
bytes_processed_so_far_ += sizeof(resource);
return;
}
}
}
// One of the strings in the natives cache should match the resource. We
// can't serialize any other kinds of external strings.
UNREACHABLE();
}
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
int Serializer::ObjectSerializer::OutputRawData(
Address up_to, Serializer::ObjectSerializer::ReturnSkip return_skip) {
Address object_start = object_->address();
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
Address base = object_start + bytes_processed_so_far_;
int up_to_offset = static_cast<int>(up_to - object_start);
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
int to_skip = up_to_offset - bytes_processed_so_far_;
int bytes_to_output = to_skip;
bytes_processed_so_far_ += to_skip;
// This assert will fail if the reloc info gives us the target_address_address
// locations in a non-ascending order. Luckily that doesn't happen.
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
ASSERT(to_skip >= 0);
bool outputting_code = false;
if (to_skip != 0 && code_object_ && !code_has_been_output_) {
// Output the code all at once and fix later.
bytes_to_output = object_->Size() + to_skip - bytes_processed_so_far_;
outputting_code = true;
code_has_been_output_ = true;
}
if (bytes_to_output != 0 &&
(!code_object_ || outputting_code)) {
#define RAW_CASE(index) \
if (!outputting_code && bytes_to_output == index * kPointerSize && \
index * kPointerSize == to_skip) { \
sink_->PutSection(kRawData + index, "RawDataFixed"); \
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
to_skip = 0; /* This insn already skips. */ \
} else /* NOLINT */
COMMON_RAW_LENGTHS(RAW_CASE)
#undef RAW_CASE
{ /* NOLINT */
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
// We always end up here if we are outputting the code of a code object.
sink_->Put(kRawData, "RawData");
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
sink_->PutInt(bytes_to_output, "length");
}
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
for (int i = 0; i < bytes_to_output; i++) {
unsigned int data = base[i];
sink_->PutSection(data, "Byte");
}
}
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
if (to_skip != 0 && return_skip == kIgnoringReturn) {
sink_->Put(kSkip, "Skip");
sink_->PutInt(to_skip, "SkipDistance");
to_skip = 0;
}
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
return to_skip;
}
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
int Serializer::SpaceOfObject(HeapObject* object) {
for (int i = FIRST_SPACE; i <= LAST_SPACE; i++) {
AllocationSpace s = static_cast<AllocationSpace>(i);
if (HEAP->InSpace(object, s)) {
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
ASSERT(i < kNumberOfSpaces);
return i;
}
}
UNREACHABLE();
return 0;
}
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
int Serializer::Allocate(int space, int size) {
CHECK(space >= 0 && space < kNumberOfSpaces);
int allocation_address = fullness_[space];
fullness_[space] = allocation_address + size;
return allocation_address;
}
int Serializer::SpaceAreaSize(int space) {
if (space == CODE_SPACE) {
return isolate_->memory_allocator()->CodePageAreaSize();
} else {
return Page::kPageSize - Page::kObjectStartOffset;
}
}
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
void Serializer::Pad() {
// The non-branching GetInt will read up to 3 bytes too far, so we need
// to pad the snapshot to make sure we don't read over the end.
for (unsigned i = 0; i < sizeof(int32_t) - 1; i++) {
sink_->Put(kNop, "Padding");
}
}
bool SnapshotByteSource::AtEOF() {
if (0u + length_ - position_ > 2 * sizeof(uint32_t)) return false;
Refactoring of snapshots. This simplifies and improves the speed of deserializing code. The current startup time improvement for V8 is around 6%, but code deserialization is speeded up disproportionately, and we will soon have more code in the snapshot. * Removed support for deserializing into large object space. The regular pages are 1Mbyte now and that is plenty. This is a big simplification. * Instead of reserving space for the snapshot we actually allocate it now. This removes some special casing from the memory management and simplifies deserialization since we are just bumping a pointer rather than calling the normal allocation routines during deserialization. * Record in the snapshot how much we need to boot up and allocate it instead of just assuming that allocations in a new VM will always be linear. * In the snapshot we always address an object as a negative offset from the current allocation point. We used to sometimes address from the start of the deserialized data, but this is less useful now that we have good support for roots and repetitions in the deserialization data. * Code objects were previously deserialized (like other objects) by alternating raw data (deserialized with memcpy) and pointers (to external references, other objects, etc.). Now we deserialize code objects with a single memcpy, followed by a series of skips and pointers that partially overwrite the code we memcopied out of the snapshot. The skips are sometimes merged into the following instruction in the deserialization data to reduce dispatch time. * Integers in the snapshot were stored in a variable length format that gives a compact representation for small positive integers. This is still the case, but the new encoding can be decoded without branches or conditional instructions, which is faster on a modern CPU. Review URL: https://chromiumcodereview.appspot.com/10918067 git-svn-id: http://v8.googlecode.com/svn/branches/bleeding_edge@12505 ce2b1a6d-e550-0410-aec6-3dcde31c8c00
2012-09-14 11:16:56 +00:00
for (int x = position_; x < length_; x++) {
if (data_[x] != SerializerDeserializer::nop()) return false;
}
return true;
}
} } // namespace v8::internal