Profile-guided optimization of builtins

Design doc:
https://docs.google.com/document/d/1szInbXZfaErWW70d30hJsOLL0Es-l5_g8d2rXm1ZBqI/edit?usp=sharing

V8 can already collect data about how many times each basic block in the
builtins is run. This change enables using that data for profile-guided
optimization. New comments in BUILD.gn describe how to use this feature.

A few implementation details worth mentioning, which aren't covered in
the design doc:

- BasicBlockProfilerData currently contains an array of RPO numbers.
  However, this array is always just [0, 1, 2, 3, ...], so this change
  removes that array. A new DCHECK in BasicBlockInstrumentor::Instrument
  ensures that the removal is valid.

- RPO numbers, while useful for printing data that matches with the
  stringified schedule, are not useful for matching profiling data with
  blocks that haven't been scheduled yet. This change adds a new array
  of block IDs in BasicBlockProfilerData, so that block counters can be
  used for PGO.

- Basic block counters need to be written to a file so that they can be
  provided to a subsequent run of mksnapshot, but the design doc doesn't
  specify the transfer format or what file is used. In this change, I
  propose using the existing v8.log file for that purpose. Block count
  records look like this:

  block,TestLessThanHandler,37,29405

  This line indicates that block ID 37 in TestLessThanHandler was run
  29405 times. If multiple lines refer to the same block, the reader
  adds them all together. I like this format because it's easy to use:
  - V8 already has robust logic for creating the log file, naming it to
    avoid conflicts in multi-process situations, etc.
  - Line order doesn't matter, and interleaved writes from various
    logging sources are fine, given that V8 writes each line atomically.
  - Combining multiple sources of profiling data is as simple as
    concatenating their v8.log files together.

- It is a good idea to avoid making any changes based on profiling data
  if the function being compiled doesn't match the one that was
  profiled, since it is common to use profiling data downloaded from a
  central lab which is updated only periodically. To check whether a
  function matches, I propose using a hash of the Graph state right
  before scheduling. This might be stricter than necessary, as some
  changes to the function might be small enough that the profile data is
  still relevant, but I'd rather err on the side of not making incorrect
  changes. This hash is also written to the v8.log file, in a line that
  looks like this:

  builtin_hash,LdaZeroHandler,3387822046

Bug: v8:10470
Change-Id: I429e5ce5efa94e01e7489deb3996012cf860cf13
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2220765
Commit-Queue: Seth Brenith <seth.brenith@microsoft.com>
Reviewed-by: Ross McIlroy <rmcilroy@chromium.org>
Reviewed-by: Tobias Tebbi <tebbi@chromium.org>
Cr-Commit-Position: refs/heads/master@{#69008}
This commit is contained in:
Seth Brenith 2020-07-16 09:37:08 -07:00 committed by Commit Bot
parent 6f356f799b
commit 922983dfd3
21 changed files with 490 additions and 70 deletions

View File

@ -144,6 +144,22 @@ declare_args() {
# used builtins. # used builtins.
v8_enable_builtins_profiling_verbose = false v8_enable_builtins_profiling_verbose = false
# Provides the given V8 log file as an input to mksnapshot, where it can be
# used for profile-guided optimization of builtins.
#
# To do profile-guided optimizations of builtins:
# 1. Build with v8_enable_builtins_profiling = true
# 2. Run your chosen workload with the --turbo-profiling-log-builtins flag.
# For Chrome, the invocation might look like this:
# chrome --no-sandbox --disable-extensions
# --js-flags="--turbo-profiling-log-builtins --logfile=path/to/v8.log"
# "http://localhost/test-suite"
# 3. Optionally repeat step 2 for additional workloads, and concatenate all of
# the resulting log files into a single file.
# 4. Build again with v8_builtins_profiling_log_file set to the file created
# in steps 2-3.
v8_builtins_profiling_log_file = ""
# Enables various testing features. # Enables various testing features.
v8_enable_test_features = "" v8_enable_test_features = ""
@ -1473,6 +1489,12 @@ template("run_mksnapshot") {
if (v8_enable_builtins_profiling_verbose) { if (v8_enable_builtins_profiling_verbose) {
args += [ "--turbo-profiling-verbose" ] args += [ "--turbo-profiling-verbose" ]
} }
if (v8_builtins_profiling_log_file != "") {
args += [
"--turbo-profiling-log-file",
v8_builtins_profiling_log_file,
]
}
# This is needed to distinguish between generating code for the simulator # This is needed to distinguish between generating code for the simulator
# and cross-compiling. The latter may need to run code on the host with the # and cross-compiling. The latter may need to run code on the host with the
@ -1728,6 +1750,8 @@ v8_source_set("v8_initializers") {
"src/builtins/builtins-wasm-gen.h", "src/builtins/builtins-wasm-gen.h",
"src/builtins/growable-fixed-array-gen.cc", "src/builtins/growable-fixed-array-gen.cc",
"src/builtins/growable-fixed-array-gen.h", "src/builtins/growable-fixed-array-gen.h",
"src/builtins/profile-data-reader.cc",
"src/builtins/profile-data-reader.h",
"src/builtins/setup-builtins-internal.cc", "src/builtins/setup-builtins-internal.cc",
"src/codegen/code-stub-assembler.cc", "src/codegen/code-stub-assembler.cc",
"src/codegen/code-stub-assembler.h", "src/codegen/code-stub-assembler.h",
@ -1869,6 +1893,7 @@ v8_header_set("v8_shared_internal_headers") {
v8_compiler_sources = [ v8_compiler_sources = [
### gcmole(all) ### ### gcmole(all) ###
"src/builtins/profile-data-reader.h",
"src/compiler/access-builder.cc", "src/compiler/access-builder.cc",
"src/compiler/access-builder.h", "src/compiler/access-builder.h",
"src/compiler/access-info.cc", "src/compiler/access-info.cc",
@ -2261,6 +2286,7 @@ v8_source_set("v8_base_without_compiler") {
"src/builtins/builtins.h", "src/builtins/builtins.h",
"src/builtins/constants-table-builder.cc", "src/builtins/constants-table-builder.cc",
"src/builtins/constants-table-builder.h", "src/builtins/constants-table-builder.h",
"src/builtins/profile-data-reader.h",
"src/codegen/assembler-arch.h", "src/codegen/assembler-arch.h",
"src/codegen/assembler-inl.h", "src/codegen/assembler-inl.h",
"src/codegen/assembler.cc", "src/codegen/assembler.cc",

View File

@ -0,0 +1,121 @@
// Copyright 2020 the V8 project authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
#include "src/builtins/profile-data-reader.h"
#include <fstream>
#include <iostream>
#include <unordered_map>
#include "src/base/lazy-instance.h"
#include "src/flags/flags.h"
#include "src/utils/utils.h"
namespace v8 {
namespace internal {
namespace {
class ProfileDataFromFileInternal : public ProfileDataFromFile {
public:
bool hash_has_value() const { return hash_has_value_; }
void set_hash(int hash) {
hash_ = hash;
hash_has_value_ = true;
}
void AddCountToBlock(size_t block_id, uint32_t count) {
if (block_counts_by_id_.size() <= block_id) {
// std::vector initializes new data to zero when resizing.
block_counts_by_id_.resize(block_id + 1);
}
block_counts_by_id_[block_id] += count;
}
private:
bool hash_has_value_ = false;
};
const std::unordered_map<std::string, ProfileDataFromFileInternal>&
EnsureInitProfileData() {
static base::LeakyObject<
std::unordered_map<std::string, ProfileDataFromFileInternal>>
data;
static bool initialized = false;
if (initialized) return *data.get();
initialized = true;
const char* filename = FLAG_turbo_profiling_log_file;
if (filename == nullptr) return *data.get();
std::ifstream file(filename);
CHECK_WITH_MSG(file.good(), "Can't read log file");
for (std::string line; std::getline(file, line);) {
std::string token;
std::istringstream line_stream(line);
if (!std::getline(line_stream, token, ',')) continue;
if (token == ProfileDataFromFileConstants::kBlockCounterMarker) {
// Any line starting with kBlockCounterMarker is a block usage count.
// As defined by Logger::BasicBlockCounterEvent, the format is:
// literal kBlockCounterMarker , builtin_name , block_id , usage_count
std::string builtin_name;
CHECK(std::getline(line_stream, builtin_name, ','));
CHECK(std::getline(line_stream, token, ','));
char* end = nullptr;
uint32_t id = static_cast<uint32_t>(strtoul(token.c_str(), &end, 0));
CHECK(errno == 0 && end != token.c_str());
std::getline(line_stream, token, ',');
CHECK(line_stream.eof());
uint32_t count = static_cast<uint32_t>(strtoul(token.c_str(), &end, 0));
CHECK(errno == 0 && end != token.c_str());
ProfileDataFromFileInternal& counters_and_hash =
(*data.get())[builtin_name];
// We allow concatenating data from several Isolates, so we might see the
// same block multiple times. Just sum them all.
counters_and_hash.AddCountToBlock(id, count);
} else if (token == ProfileDataFromFileConstants::kBuiltinHashMarker) {
// Any line starting with kBuiltinHashMarker is a function hash record.
// As defined by Logger::BuiltinHashEvent, the format is:
// literal kBuiltinHashMarker , builtin_name , hash
std::string builtin_name;
CHECK(std::getline(line_stream, builtin_name, ','));
std::getline(line_stream, token, ',');
CHECK(line_stream.eof());
char* end = nullptr;
int hash = static_cast<int>(strtol(token.c_str(), &end, 0));
CHECK(errno == 0 && end != token.c_str());
ProfileDataFromFileInternal& counters_and_hash =
(*data.get())[builtin_name];
// We allow concatenating data from several Isolates, but expect them all
// to be running the same build. Any file with mismatched hashes for a
// function is considered ill-formed.
CHECK_IMPLIES(counters_and_hash.hash_has_value(),
counters_and_hash.hash() == hash);
counters_and_hash.set_hash(hash);
}
}
for (const auto& pair : *data.get()) {
// Every function is required to have a hash in the log.
CHECK(pair.second.hash_has_value());
}
if (data.get()->size() == 0) {
PrintF(
"No basic block counters were found in log file.\n"
"Did you build with v8_enable_builtins_profiling=true\n"
"and run with --turbo-profiling-log-builtins?\n");
}
return *data.get();
}
} // namespace
const ProfileDataFromFile* ProfileDataFromFile::TryRead(const char* name) {
const auto& data = EnsureInitProfileData();
auto it = data.find(name);
return it == data.end() ? nullptr : &it->second;
}
} // namespace internal
} // namespace v8

View File

@ -0,0 +1,62 @@
// Copyright 2020 the V8 project authors. All rights reserved.
// Use of this source code is governed by a BSD-style license that can be
// found in the LICENSE file.
#ifndef V8_BUILTINS_PROFILE_DATA_READER_H_
#define V8_BUILTINS_PROFILE_DATA_READER_H_
#include <cstddef>
#include <cstdint>
#include <vector>
namespace v8 {
namespace internal {
class ProfileDataFromFile {
public:
// A hash of the function's Graph before scheduling. Allows us to avoid using
// profiling data if the function has been changed.
int hash() const { return hash_; }
// Returns how many times the block with the given ID was executed during
// profiling.
uint32_t GetCounter(size_t block_id) const {
// The profile data is allowed to omit blocks which were never hit, so be
// careful to avoid out-of-bounds access.
return block_id < block_counts_by_id_.size() ? block_counts_by_id_[block_id]
: 0;
}
// Load basic block profiling data for the builtin with the given name, if
// such data exists. The returned vector is indexed by block ID, and its
// values are the number of times each block was executed while profiling.
static const ProfileDataFromFile* TryRead(const char* name);
protected:
int hash_ = 0;
// How many times each block was executed, indexed by block ID. This vector
// may be shorter than the total number of blocks; any omitted block should be
// treated as a zero.
std::vector<uint32_t> block_counts_by_id_;
};
// The following strings can't be static members of ProfileDataFromFile until
// C++ 17; see https://stackoverflow.com/q/8016780/839379 . So for now we use a
// namespace.
namespace ProfileDataFromFileConstants {
// Any line in a v8.log beginning with this string represents a basic block
// counter.
static constexpr char kBlockCounterMarker[] = "block";
// Any line in a v8.log beginning with this string represents the hash of the
// function Graph for a builtin.
static constexpr char kBuiltinHashMarker[] = "builtin_hash";
} // namespace ProfileDataFromFileConstants
} // namespace internal
} // namespace v8
#endif // V8_BUILTINS_PROFILE_DATA_READER_H_

View File

@ -3,6 +3,7 @@
// found in the LICENSE file. // found in the LICENSE file.
#include "src/builtins/builtins.h" #include "src/builtins/builtins.h"
#include "src/builtins/profile-data-reader.h"
#include "src/codegen/assembler-inl.h" #include "src/codegen/assembler-inl.h"
#include "src/codegen/interface-descriptors.h" #include "src/codegen/interface-descriptors.h"
#include "src/codegen/macro-assembler-inl.h" #include "src/codegen/macro-assembler-inl.h"
@ -165,7 +166,8 @@ Code BuildWithCodeStubAssemblerJS(Isolate* isolate, int32_t builtin_index,
PoisoningMitigationLevel::kDontPoison, builtin_index); PoisoningMitigationLevel::kDontPoison, builtin_index);
generator(&state); generator(&state);
Handle<Code> code = compiler::CodeAssembler::GenerateCode( Handle<Code> code = compiler::CodeAssembler::GenerateCode(
&state, BuiltinAssemblerOptions(isolate, builtin_index)); &state, BuiltinAssemblerOptions(isolate, builtin_index),
ProfileDataFromFile::TryRead(name));
return *code; return *code;
} }
@ -189,7 +191,8 @@ Code BuildWithCodeStubAssemblerCS(Isolate* isolate, int32_t builtin_index,
PoisoningMitigationLevel::kDontPoison, builtin_index); PoisoningMitigationLevel::kDontPoison, builtin_index);
generator(&state); generator(&state);
Handle<Code> code = compiler::CodeAssembler::GenerateCode( Handle<Code> code = compiler::CodeAssembler::GenerateCode(
&state, BuiltinAssemblerOptions(isolate, builtin_index)); &state, BuiltinAssemblerOptions(isolate, builtin_index),
ProfileDataFromFile::TryRead(name));
return *code; return *code;
} }

View File

@ -98,7 +98,9 @@ BasicBlockProfilerData* BasicBlockInstrumentor::Instrument(
for (BasicBlockVector::iterator it = blocks->begin(); block_number < n_blocks; for (BasicBlockVector::iterator it = blocks->begin(); block_number < n_blocks;
++it, ++block_number) { ++it, ++block_number) {
BasicBlock* block = (*it); BasicBlock* block = (*it);
data->SetBlockRpoNumber(block_number, block->rpo_number()); // Iteration is already in reverse post-order.
DCHECK_EQ(block->rpo_number(), block_number);
data->SetBlockId(block_number, block->id().ToInt());
// It is unnecessary to wire effect and control deps for load and store // It is unnecessary to wire effect and control deps for load and store
// since this happens after scheduling. // since this happens after scheduling.
// Construct increment operation. // Construct increment operation.

View File

@ -175,8 +175,9 @@ PoisoningMitigationLevel CodeAssembler::poisoning_level() const {
} }
// static // static
Handle<Code> CodeAssembler::GenerateCode(CodeAssemblerState* state, Handle<Code> CodeAssembler::GenerateCode(
const AssemblerOptions& options) { CodeAssemblerState* state, const AssemblerOptions& options,
const ProfileDataFromFile* profile_data) {
DCHECK(!state->code_generated_); DCHECK(!state->code_generated_);
RawMachineAssembler* rasm = state->raw_assembler_.get(); RawMachineAssembler* rasm = state->raw_assembler_.get();
@ -184,11 +185,12 @@ Handle<Code> CodeAssembler::GenerateCode(CodeAssemblerState* state,
Handle<Code> code; Handle<Code> code;
Graph* graph = rasm->ExportForOptimization(); Graph* graph = rasm->ExportForOptimization();
code = Pipeline::GenerateCodeForCodeStub( code =
rasm->isolate(), rasm->call_descriptor(), graph, state->jsgraph_, Pipeline::GenerateCodeForCodeStub(
rasm->source_positions(), state->kind_, state->name_, rasm->isolate(), rasm->call_descriptor(), graph, state->jsgraph_,
state->builtin_index_, rasm->poisoning_level(), options) rasm->source_positions(), state->kind_, state->name_,
.ToHandleChecked(); state->builtin_index_, rasm->poisoning_level(), options, profile_data)
.ToHandleChecked();
state->code_generated_ = true; state->code_generated_ = true;
return code; return code;

View File

@ -68,6 +68,7 @@ class JSFinalizationRegistry;
class JSWeakMap; class JSWeakMap;
class JSWeakRef; class JSWeakRef;
class JSWeakSet; class JSWeakSet;
class ProfileDataFromFile;
class PromiseCapability; class PromiseCapability;
class PromiseFulfillReactionJobTask; class PromiseFulfillReactionJobTask;
class PromiseReaction; class PromiseReaction;
@ -375,7 +376,8 @@ class V8_EXPORT_PRIVATE CodeAssembler {
~CodeAssembler(); ~CodeAssembler();
static Handle<Code> GenerateCode(CodeAssemblerState* state, static Handle<Code> GenerateCode(CodeAssemblerState* state,
const AssemblerOptions& options); const AssemblerOptions& options,
const ProfileDataFromFile* profile_data);
bool Is64() const; bool Is64() const;
bool Is32() const; bool Is32() const;

View File

@ -11,6 +11,7 @@
#include "src/base/optional.h" #include "src/base/optional.h"
#include "src/base/platform/elapsed-timer.h" #include "src/base/platform/elapsed-timer.h"
#include "src/builtins/profile-data-reader.h"
#include "src/codegen/assembler-inl.h" #include "src/codegen/assembler-inl.h"
#include "src/codegen/compiler.h" #include "src/codegen/compiler.h"
#include "src/codegen/optimized-compilation-info.h" #include "src/codegen/optimized-compilation-info.h"
@ -215,7 +216,8 @@ class PipelineData {
JSGraph* jsgraph, Schedule* schedule, JSGraph* jsgraph, Schedule* schedule,
SourcePositionTable* source_positions, SourcePositionTable* source_positions,
NodeOriginTable* node_origins, JumpOptimizationInfo* jump_opt, NodeOriginTable* node_origins, JumpOptimizationInfo* jump_opt,
const AssemblerOptions& assembler_options) const AssemblerOptions& assembler_options,
const ProfileDataFromFile* profile_data)
: isolate_(isolate), : isolate_(isolate),
wasm_engine_(isolate_->wasm_engine()), wasm_engine_(isolate_->wasm_engine()),
allocator_(allocator), allocator_(allocator),
@ -236,7 +238,8 @@ class PipelineData {
kRegisterAllocationZoneName), kRegisterAllocationZoneName),
register_allocation_zone_(register_allocation_zone_scope_.zone()), register_allocation_zone_(register_allocation_zone_scope_.zone()),
jump_optimization_info_(jump_opt), jump_optimization_info_(jump_opt),
assembler_options_(assembler_options) { assembler_options_(assembler_options),
profile_data_(profile_data) {
if (jsgraph) { if (jsgraph) {
jsgraph_ = jsgraph; jsgraph_ = jsgraph;
simplified_ = jsgraph->simplified(); simplified_ = jsgraph->simplified();
@ -533,6 +536,11 @@ class PipelineData {
return roots_relative_addressing_enabled_; return roots_relative_addressing_enabled_;
} }
const ProfileDataFromFile* profile_data() const { return profile_data_; }
void set_profile_data(const ProfileDataFromFile* profile_data) {
profile_data_ = profile_data;
}
// RuntimeCallStats that is only available during job execution but not // RuntimeCallStats that is only available during job execution but not
// finalization. // finalization.
// TODO(delphick): Currently even during execution this can be nullptr, due to // TODO(delphick): Currently even during execution this can be nullptr, due to
@ -614,6 +622,7 @@ class PipelineData {
size_t max_pushed_argument_count_ = 0; size_t max_pushed_argument_count_ = 0;
RuntimeCallStats* runtime_call_stats_ = nullptr; RuntimeCallStats* runtime_call_stats_ = nullptr;
const ProfileDataFromFile* profile_data_ = nullptr;
DISALLOW_COPY_AND_ASSIGN(PipelineData); DISALLOW_COPY_AND_ASSIGN(PipelineData);
}; };
@ -1243,7 +1252,7 @@ class WasmHeapStubCompilationJob final : public OptimizedCompilationJob {
graph_(graph), graph_(graph),
data_(&zone_stats_, &info_, isolate, wasm_engine->allocator(), graph_, data_(&zone_stats_, &info_, isolate, wasm_engine->allocator(), graph_,
nullptr, nullptr, source_positions, nullptr, nullptr, source_positions,
zone_->New<NodeOriginTable>(graph_), nullptr, options), zone_->New<NodeOriginTable>(graph_), nullptr, options, nullptr),
pipeline_(&data_), pipeline_(&data_),
wasm_engine_(wasm_engine) {} wasm_engine_(wasm_engine) {}
@ -1728,7 +1737,7 @@ struct EffectControlLinearizationPhase {
// 'floating' allocation regions.) // 'floating' allocation regions.)
Schedule* schedule = Scheduler::ComputeSchedule( Schedule* schedule = Scheduler::ComputeSchedule(
temp_zone, data->graph(), Scheduler::kTempSchedule, temp_zone, data->graph(), Scheduler::kTempSchedule,
&data->info()->tick_counter()); &data->info()->tick_counter(), data->profile_data());
TraceScheduleAndVerify(data->info(), data, schedule, TraceScheduleAndVerify(data->info(), data, schedule,
"effect linearization schedule"); "effect linearization schedule");
@ -2028,7 +2037,7 @@ struct ComputeSchedulePhase {
temp_zone, data->graph(), temp_zone, data->graph(),
data->info()->splitting() ? Scheduler::kSplitNodes data->info()->splitting() ? Scheduler::kSplitNodes
: Scheduler::kNoFlags, : Scheduler::kNoFlags,
&data->info()->tick_counter()); &data->info()->tick_counter(), data->profile_data());
data->set_schedule(schedule); data->set_schedule(schedule);
} }
}; };
@ -2326,9 +2335,9 @@ struct PrintGraphPhase {
AccountingAllocator allocator; AccountingAllocator allocator;
Schedule* schedule = data->schedule(); Schedule* schedule = data->schedule();
if (schedule == nullptr) { if (schedule == nullptr) {
schedule = Scheduler::ComputeSchedule(temp_zone, data->graph(), schedule = Scheduler::ComputeSchedule(
Scheduler::kNoFlags, temp_zone, data->graph(), Scheduler::kNoFlags,
&info->tick_counter()); &info->tick_counter(), data->profile_data());
} }
AllowHandleDereference allow_deref; AllowHandleDereference allow_deref;
@ -2633,11 +2642,80 @@ bool PipelineImpl::OptimizeGraphForMidTier(Linkage* linkage) {
return SelectInstructions(linkage); return SelectInstructions(linkage);
} }
namespace {
// Compute a hash of the given graph, in a way that should provide the same
// result in multiple runs of mksnapshot, meaning the hash cannot depend on any
// external pointer values or uncompressed heap constants. This hash can be used
// to reject profiling data if the builtin's current code doesn't match the
// version that was profiled. Hash collisions are not catastrophic; in the worst
// case, we just defer some blocks that ideally shouldn't be deferred. The
// result value is in the valid Smi range.
int HashGraphForPGO(Graph* graph) {
AccountingAllocator allocator;
Zone local_zone(&allocator, ZONE_NAME);
constexpr NodeId kUnassigned = static_cast<NodeId>(-1);
constexpr byte kUnvisited = 0;
constexpr byte kOnStack = 1;
constexpr byte kVisited = 2;
// Do a depth-first post-order traversal of the graph. For every node, hash:
//
// - the node's traversal number
// - the opcode
// - the number of inputs
// - each input node's traversal number
//
// What's a traversal number? We can't use node IDs because they're not stable
// build-to-build, so we assign a new number for each node as it is visited.
ZoneVector<byte> state(graph->NodeCount(), kUnvisited, &local_zone);
ZoneVector<NodeId> traversal_numbers(graph->NodeCount(), kUnassigned,
&local_zone);
ZoneStack<Node*> stack(&local_zone);
NodeId visited_count = 0;
size_t hash = 0;
stack.push(graph->end());
state[graph->end()->id()] = kOnStack;
traversal_numbers[graph->end()->id()] = visited_count++;
while (!stack.empty()) {
Node* n = stack.top();
bool pop = true;
for (Node* const i : n->inputs()) {
if (state[i->id()] == kUnvisited) {
state[i->id()] = kOnStack;
traversal_numbers[i->id()] = visited_count++;
stack.push(i);
pop = false;
break;
}
}
if (pop) {
state[n->id()] = kVisited;
stack.pop();
hash = base::hash_combine(hash, traversal_numbers[n->id()], n->opcode(),
n->InputCount());
for (Node* const i : n->inputs()) {
DCHECK(traversal_numbers[i->id()] != kUnassigned);
hash = base::hash_combine(hash, traversal_numbers[i->id()]);
}
}
}
return Smi(IntToSmi(static_cast<int>(hash))).value();
}
} // namespace
MaybeHandle<Code> Pipeline::GenerateCodeForCodeStub( MaybeHandle<Code> Pipeline::GenerateCodeForCodeStub(
Isolate* isolate, CallDescriptor* call_descriptor, Graph* graph, Isolate* isolate, CallDescriptor* call_descriptor, Graph* graph,
JSGraph* jsgraph, SourcePositionTable* source_positions, Code::Kind kind, JSGraph* jsgraph, SourcePositionTable* source_positions, Code::Kind kind,
const char* debug_name, int32_t builtin_index, const char* debug_name, int32_t builtin_index,
PoisoningMitigationLevel poisoning_level, const AssemblerOptions& options) { PoisoningMitigationLevel poisoning_level, const AssemblerOptions& options,
const ProfileDataFromFile* profile_data) {
OptimizedCompilationInfo info(CStrVector(debug_name), graph->zone(), kind); OptimizedCompilationInfo info(CStrVector(debug_name), graph->zone(), kind);
info.set_builtin_index(builtin_index); info.set_builtin_index(builtin_index);
@ -2654,7 +2732,8 @@ MaybeHandle<Code> Pipeline::GenerateCodeForCodeStub(
!FLAG_turbo_profiling; !FLAG_turbo_profiling;
PipelineData data(&zone_stats, &info, isolate, isolate->allocator(), graph, PipelineData data(&zone_stats, &info, isolate, isolate->allocator(), graph,
jsgraph, nullptr, source_positions, &node_origins, jsgraph, nullptr, source_positions, &node_origins,
should_optimize_jumps ? &jump_opt : nullptr, options); should_optimize_jumps ? &jump_opt : nullptr, options,
profile_data);
PipelineJobScope scope(&data, isolate->counters()->runtime_call_stats()); PipelineJobScope scope(&data, isolate->counters()->runtime_call_stats());
RuntimeCallTimerScope timer_scope(isolate, RuntimeCallTimerScope timer_scope(isolate,
RuntimeCallCounterId::kOptimizeCode); RuntimeCallCounterId::kOptimizeCode);
@ -2699,6 +2778,19 @@ MaybeHandle<Code> Pipeline::GenerateCodeForCodeStub(
true); true);
pipeline.Run<VerifyGraphPhase>(true); pipeline.Run<VerifyGraphPhase>(true);
int graph_hash_before_scheduling = 0;
if (FLAG_turbo_profiling || profile_data != nullptr) {
graph_hash_before_scheduling = HashGraphForPGO(data.graph());
}
if (profile_data != nullptr &&
profile_data->hash() != graph_hash_before_scheduling) {
PrintF("Rejected profile data for %s due to function change\n", debug_name);
profile_data = nullptr;
data.set_profile_data(profile_data);
}
pipeline.ComputeScheduledGraph(); pipeline.ComputeScheduledGraph();
DCHECK_NOT_NULL(data.schedule()); DCHECK_NOT_NULL(data.schedule());
@ -2708,13 +2800,18 @@ MaybeHandle<Code> Pipeline::GenerateCodeForCodeStub(
PipelineData second_data(&zone_stats, &info, isolate, isolate->allocator(), PipelineData second_data(&zone_stats, &info, isolate, isolate->allocator(),
data.graph(), data.jsgraph(), data.schedule(), data.graph(), data.jsgraph(), data.schedule(),
data.source_positions(), data.node_origins(), data.source_positions(), data.node_origins(),
data.jump_optimization_info(), options); data.jump_optimization_info(), options,
profile_data);
PipelineJobScope second_scope(&second_data, PipelineJobScope second_scope(&second_data,
isolate->counters()->runtime_call_stats()); isolate->counters()->runtime_call_stats());
second_data.set_verify_graph(FLAG_verify_csa); second_data.set_verify_graph(FLAG_verify_csa);
PipelineImpl second_pipeline(&second_data); PipelineImpl second_pipeline(&second_data);
second_pipeline.SelectInstructionsAndAssemble(call_descriptor); second_pipeline.SelectInstructionsAndAssemble(call_descriptor);
if (FLAG_turbo_profiling) {
info.profiler_data()->SetHash(graph_hash_before_scheduling);
}
Handle<Code> code; Handle<Code> code;
if (jump_opt.is_optimizable()) { if (jump_opt.is_optimizable()) {
jump_opt.set_optimizing(); jump_opt.set_optimizing();
@ -2882,7 +2979,7 @@ MaybeHandle<Code> Pipeline::GenerateCodeForTesting(
NodeOriginTable* node_positions = info->zone()->New<NodeOriginTable>(graph); NodeOriginTable* node_positions = info->zone()->New<NodeOriginTable>(graph);
PipelineData data(&zone_stats, info, isolate, isolate->allocator(), graph, PipelineData data(&zone_stats, info, isolate, isolate->allocator(), graph,
nullptr, schedule, nullptr, node_positions, nullptr, nullptr, schedule, nullptr, node_positions, nullptr,
options); options, nullptr);
std::unique_ptr<PipelineStatistics> pipeline_statistics; std::unique_ptr<PipelineStatistics> pipeline_statistics;
if (FLAG_turbo_stats || FLAG_turbo_stats_nvp) { if (FLAG_turbo_stats || FLAG_turbo_stats_nvp) {
pipeline_statistics.reset(new PipelineStatistics( pipeline_statistics.reset(new PipelineStatistics(

View File

@ -19,6 +19,7 @@ namespace internal {
struct AssemblerOptions; struct AssemblerOptions;
class OptimizedCompilationInfo; class OptimizedCompilationInfo;
class OptimizedCompilationJob; class OptimizedCompilationJob;
class ProfileDataFromFile;
class RegisterConfiguration; class RegisterConfiguration;
namespace wasm { namespace wasm {
@ -78,8 +79,8 @@ class Pipeline : public AllStatic {
Isolate* isolate, CallDescriptor* call_descriptor, Graph* graph, Isolate* isolate, CallDescriptor* call_descriptor, Graph* graph,
JSGraph* jsgraph, SourcePositionTable* source_positions, Code::Kind kind, JSGraph* jsgraph, SourcePositionTable* source_positions, Code::Kind kind,
const char* debug_name, int32_t builtin_index, const char* debug_name, int32_t builtin_index,
PoisoningMitigationLevel poisoning_level, PoisoningMitigationLevel poisoning_level, const AssemblerOptions& options,
const AssemblerOptions& options); const ProfileDataFromFile* profile_data);
// --------------------------------------------------------------------------- // ---------------------------------------------------------------------------
// The following methods are for testing purposes only. Avoid production use. // The following methods are for testing purposes only. Avoid production use.

View File

@ -7,6 +7,7 @@
#include <iomanip> #include <iomanip>
#include "src/base/iterator.h" #include "src/base/iterator.h"
#include "src/builtins/profile-data-reader.h"
#include "src/codegen/tick-counter.h" #include "src/codegen/tick-counter.h"
#include "src/compiler/common-operator.h" #include "src/compiler/common-operator.h"
#include "src/compiler/control-equivalence.h" #include "src/compiler/control-equivalence.h"
@ -27,7 +28,8 @@ namespace compiler {
} while (false) } while (false)
Scheduler::Scheduler(Zone* zone, Graph* graph, Schedule* schedule, Flags flags, Scheduler::Scheduler(Zone* zone, Graph* graph, Schedule* schedule, Flags flags,
size_t node_count_hint, TickCounter* tick_counter) size_t node_count_hint, TickCounter* tick_counter,
const ProfileDataFromFile* profile_data)
: zone_(zone), : zone_(zone),
graph_(graph), graph_(graph),
schedule_(schedule), schedule_(schedule),
@ -36,13 +38,15 @@ Scheduler::Scheduler(Zone* zone, Graph* graph, Schedule* schedule, Flags flags,
schedule_root_nodes_(zone), schedule_root_nodes_(zone),
schedule_queue_(zone), schedule_queue_(zone),
node_data_(zone), node_data_(zone),
tick_counter_(tick_counter) { tick_counter_(tick_counter),
profile_data_(profile_data) {
node_data_.reserve(node_count_hint); node_data_.reserve(node_count_hint);
node_data_.resize(graph->NodeCount(), DefaultSchedulerData()); node_data_.resize(graph->NodeCount(), DefaultSchedulerData());
} }
Schedule* Scheduler::ComputeSchedule(Zone* zone, Graph* graph, Flags flags, Schedule* Scheduler::ComputeSchedule(Zone* zone, Graph* graph, Flags flags,
TickCounter* tick_counter) { TickCounter* tick_counter,
const ProfileDataFromFile* profile_data) {
Zone* schedule_zone = Zone* schedule_zone =
(flags & Scheduler::kTempSchedule) ? zone : graph->zone(); (flags & Scheduler::kTempSchedule) ? zone : graph->zone();
@ -54,7 +58,7 @@ Schedule* Scheduler::ComputeSchedule(Zone* zone, Graph* graph, Flags flags,
Schedule* schedule = Schedule* schedule =
schedule_zone->New<Schedule>(schedule_zone, node_count_hint); schedule_zone->New<Schedule>(schedule_zone, node_count_hint);
Scheduler scheduler(zone, graph, schedule, flags, node_count_hint, Scheduler scheduler(zone, graph, schedule, flags, node_count_hint,
tick_counter); tick_counter, profile_data);
scheduler.BuildCFG(); scheduler.BuildCFG();
scheduler.ComputeSpecialRPONumbering(); scheduler.ComputeSpecialRPONumbering();
@ -472,9 +476,38 @@ class CFGBuilder : public ZoneObject {
CollectSuccessorBlocks(branch, successor_blocks, CollectSuccessorBlocks(branch, successor_blocks,
arraysize(successor_blocks)); arraysize(successor_blocks));
BranchHint hint_from_profile = BranchHint::kNone;
if (const ProfileDataFromFile* profile_data = scheduler_->profile_data()) {
uint32_t block_zero_count =
profile_data->GetCounter(successor_blocks[0]->id().ToSize());
uint32_t block_one_count =
profile_data->GetCounter(successor_blocks[1]->id().ToSize());
// If a branch is visited a non-trivial number of times and substantially
// more often than its alternative, then mark it as likely.
constexpr uint32_t kMinimumCount = 100000;
constexpr uint32_t kThresholdRatio = 4000;
if (block_zero_count > kMinimumCount &&
block_zero_count / kThresholdRatio > block_one_count) {
hint_from_profile = BranchHint::kTrue;
} else if (block_one_count > kMinimumCount &&
block_one_count / kThresholdRatio > block_zero_count) {
hint_from_profile = BranchHint::kFalse;
}
}
// Consider branch hints. // Consider branch hints.
switch (BranchHintOf(branch->op())) { switch (hint_from_profile) {
case BranchHint::kNone: case BranchHint::kNone:
switch (BranchHintOf(branch->op())) {
case BranchHint::kNone:
break;
case BranchHint::kTrue:
successor_blocks[1]->set_deferred(true);
break;
case BranchHint::kFalse:
successor_blocks[0]->set_deferred(true);
break;
}
break; break;
case BranchHint::kTrue: case BranchHint::kTrue:
successor_blocks[1]->set_deferred(true); successor_blocks[1]->set_deferred(true);
@ -484,6 +517,12 @@ class CFGBuilder : public ZoneObject {
break; break;
} }
if (hint_from_profile != BranchHint::kNone &&
BranchHintOf(branch->op()) != BranchHint::kNone &&
hint_from_profile != BranchHintOf(branch->op())) {
PrintF("Warning: profiling data overrode manual branch hint.\n");
}
if (branch == component_entry_) { if (branch == component_entry_) {
TraceConnect(branch, component_start_, successor_blocks[0]); TraceConnect(branch, component_start_, successor_blocks[0]);
TraceConnect(branch, component_start_, successor_blocks[1]); TraceConnect(branch, component_start_, successor_blocks[1]);

View File

@ -16,6 +16,7 @@
namespace v8 { namespace v8 {
namespace internal { namespace internal {
class ProfileDataFromFile;
class TickCounter; class TickCounter;
namespace compiler { namespace compiler {
@ -37,7 +38,8 @@ class V8_EXPORT_PRIVATE Scheduler {
// The complete scheduling algorithm. Creates a new schedule and places all // The complete scheduling algorithm. Creates a new schedule and places all
// nodes from the graph into it. // nodes from the graph into it.
static Schedule* ComputeSchedule(Zone* temp_zone, Graph* graph, Flags flags, static Schedule* ComputeSchedule(Zone* temp_zone, Graph* graph, Flags flags,
TickCounter* tick_counter); TickCounter* tick_counter,
const ProfileDataFromFile* profile_data);
// Compute the RPO of blocks in an existing schedule. // Compute the RPO of blocks in an existing schedule.
static BasicBlockVector* ComputeSpecialRPO(Zone* zone, Schedule* schedule); static BasicBlockVector* ComputeSpecialRPO(Zone* zone, Schedule* schedule);
@ -45,6 +47,8 @@ class V8_EXPORT_PRIVATE Scheduler {
// Computes the dominator tree on an existing schedule that has RPO computed. // Computes the dominator tree on an existing schedule that has RPO computed.
static void GenerateDominatorTree(Schedule* schedule); static void GenerateDominatorTree(Schedule* schedule);
const ProfileDataFromFile* profile_data() const { return profile_data_; }
private: private:
// Placement of a node changes during scheduling. The placement state // Placement of a node changes during scheduling. The placement state
// transitions over time while the scheduler is choosing a position: // transitions over time while the scheduler is choosing a position:
@ -85,9 +89,11 @@ class V8_EXPORT_PRIVATE Scheduler {
SpecialRPONumberer* special_rpo_; // Special RPO numbering of blocks. SpecialRPONumberer* special_rpo_; // Special RPO numbering of blocks.
ControlEquivalence* equivalence_; // Control dependence equivalence. ControlEquivalence* equivalence_; // Control dependence equivalence.
TickCounter* const tick_counter_; TickCounter* const tick_counter_;
const ProfileDataFromFile* profile_data_;
Scheduler(Zone* zone, Graph* graph, Schedule* schedule, Flags flags, Scheduler(Zone* zone, Graph* graph, Schedule* schedule, Flags flags,
size_t node_count_hint_, TickCounter* tick_counter); size_t node_count_hint_, TickCounter* tick_counter,
const ProfileDataFromFile* profile_data);
inline SchedulerData DefaultSchedulerData(); inline SchedulerData DefaultSchedulerData();
inline SchedulerData* GetData(Node* node); inline SchedulerData* GetData(Node* node);

View File

@ -18,7 +18,7 @@ namespace internal {
DEFINE_LAZY_LEAKY_OBJECT_GETTER(BasicBlockProfiler, BasicBlockProfiler::Get) DEFINE_LAZY_LEAKY_OBJECT_GETTER(BasicBlockProfiler, BasicBlockProfiler::Get)
BasicBlockProfilerData::BasicBlockProfilerData(size_t n_blocks) BasicBlockProfilerData::BasicBlockProfilerData(size_t n_blocks)
: block_rpo_numbers_(n_blocks), counts_(n_blocks, 0) {} : block_ids_(n_blocks), counts_(n_blocks, 0) {}
void BasicBlockProfilerData::SetCode(const std::ostringstream& os) { void BasicBlockProfilerData::SetCode(const std::ostringstream& os) {
code_ = os.str(); code_ = os.str();
@ -32,12 +32,13 @@ void BasicBlockProfilerData::SetSchedule(const std::ostringstream& os) {
schedule_ = os.str(); schedule_ = os.str();
} }
void BasicBlockProfilerData::SetBlockRpoNumber(size_t offset, void BasicBlockProfilerData::SetBlockId(size_t offset, int32_t id) {
int32_t block_rpo) {
DCHECK(offset < n_blocks()); DCHECK(offset < n_blocks());
block_rpo_numbers_[offset] = block_rpo; block_ids_[offset] = id;
} }
void BasicBlockProfilerData::SetHash(int hash) { hash_ = hash; }
void BasicBlockProfilerData::ResetCounts() { void BasicBlockProfilerData::ResetCounts() {
for (size_t i = 0; i < n_blocks(); ++i) { for (size_t i = 0; i < n_blocks(); ++i) {
counts_[i] = 0; counts_[i] = 0;
@ -58,7 +59,7 @@ Handle<String> CopyStringToJSHeap(const std::string& source, Isolate* isolate) {
AllocationType::kOld); AllocationType::kOld);
} }
// Size of entries in both block_rpo_numbers and counts. // Size of entries in both block_ids and counts.
constexpr int kBasicBlockSlotSize = kInt32Size; constexpr int kBasicBlockSlotSize = kInt32Size;
} // namespace } // namespace
@ -71,11 +72,12 @@ BasicBlockProfilerData::BasicBlockProfilerData(
for (int i = 0; i < counts->length() / kBasicBlockSlotSize; ++i) { for (int i = 0; i < counts->length() / kBasicBlockSlotSize; ++i) {
counts_.push_back(counts->get_uint32(i)); counts_.push_back(counts->get_uint32(i));
} }
Handle<ByteArray> rpo_numbers(js_heap_data->block_rpo_numbers(), isolate); Handle<ByteArray> block_ids(js_heap_data->block_ids(), isolate);
for (int i = 0; i < rpo_numbers->length() / kBasicBlockSlotSize; ++i) { for (int i = 0; i < block_ids->length() / kBasicBlockSlotSize; ++i) {
block_rpo_numbers_.push_back(rpo_numbers->get_int(i)); block_ids_.push_back(block_ids->get_int(i));
} }
CHECK_EQ(block_rpo_numbers_.size(), counts_.size()); CHECK_EQ(block_ids_.size(), counts_.size());
hash_ = js_heap_data->hash();
} }
BasicBlockProfilerData::BasicBlockProfilerData( BasicBlockProfilerData::BasicBlockProfilerData(
@ -87,11 +89,11 @@ BasicBlockProfilerData::BasicBlockProfilerData(
for (int i = 0; i < counts.length() / kBasicBlockSlotSize; ++i) { for (int i = 0; i < counts.length() / kBasicBlockSlotSize; ++i) {
counts_.push_back(counts.get_uint32(i)); counts_.push_back(counts.get_uint32(i));
} }
ByteArray rpo_numbers(js_heap_data.block_rpo_numbers()); ByteArray block_ids(js_heap_data.block_ids());
for (int i = 0; i < rpo_numbers.length() / kBasicBlockSlotSize; ++i) { for (int i = 0; i < block_ids.length() / kBasicBlockSlotSize; ++i) {
block_rpo_numbers_.push_back(rpo_numbers.get_int(i)); block_ids_.push_back(block_ids.get_int(i));
} }
CHECK_EQ(block_rpo_numbers_.size(), counts_.size()); CHECK_EQ(block_ids_.size(), counts_.size());
} }
Handle<OnHeapBasicBlockProfilerData> BasicBlockProfilerData::CopyToJSHeap( Handle<OnHeapBasicBlockProfilerData> BasicBlockProfilerData::CopyToJSHeap(
@ -100,10 +102,10 @@ Handle<OnHeapBasicBlockProfilerData> BasicBlockProfilerData::CopyToJSHeap(
CHECK(array_size_in_bytes >= 0 && CHECK(array_size_in_bytes >= 0 &&
static_cast<size_t>(array_size_in_bytes) / kBasicBlockSlotSize == static_cast<size_t>(array_size_in_bytes) / kBasicBlockSlotSize ==
n_blocks()); // Overflow n_blocks()); // Overflow
Handle<ByteArray> block_rpo_numbers = isolate->factory()->NewByteArray( Handle<ByteArray> block_ids = isolate->factory()->NewByteArray(
array_size_in_bytes, AllocationType::kOld); array_size_in_bytes, AllocationType::kOld);
for (int i = 0; i < static_cast<int>(n_blocks()); ++i) { for (int i = 0; i < static_cast<int>(n_blocks()); ++i) {
block_rpo_numbers->set_int(i, block_rpo_numbers_[i]); block_ids->set_int(i, block_ids_[i]);
} }
Handle<ByteArray> counts = isolate->factory()->NewByteArray( Handle<ByteArray> counts = isolate->factory()->NewByteArray(
array_size_in_bytes, AllocationType::kOld); array_size_in_bytes, AllocationType::kOld);
@ -115,7 +117,7 @@ Handle<OnHeapBasicBlockProfilerData> BasicBlockProfilerData::CopyToJSHeap(
Handle<String> code = CopyStringToJSHeap(code_, isolate); Handle<String> code = CopyStringToJSHeap(code_, isolate);
return isolate->factory()->NewOnHeapBasicBlockProfilerData( return isolate->factory()->NewOnHeapBasicBlockProfilerData(
block_rpo_numbers, counts, name, schedule, code, AllocationType::kOld); block_ids, counts, name, schedule, code, hash_, AllocationType::kOld);
} }
void BasicBlockProfiler::ResetCounts(Isolate* isolate) { void BasicBlockProfiler::ResetCounts(Isolate* isolate) {
@ -147,11 +149,18 @@ void BasicBlockProfiler::Print(std::ostream& os, Isolate* isolate) {
HandleScope scope(isolate); HandleScope scope(isolate);
Handle<ArrayList> list(isolate->heap()->basic_block_profiling_data(), Handle<ArrayList> list(isolate->heap()->basic_block_profiling_data(),
isolate); isolate);
std::unordered_set<std::string> builtin_names;
for (int i = 0; i < list->Length(); ++i) { for (int i = 0; i < list->Length(); ++i) {
BasicBlockProfilerData data( BasicBlockProfilerData data(
handle(OnHeapBasicBlockProfilerData::cast(list->Get(i)), isolate), handle(OnHeapBasicBlockProfilerData::cast(list->Get(i)), isolate),
isolate); isolate);
// Print data for builtins to both stdout and the log file, if logging is
// enabled.
os << data; os << data;
data.Log(isolate);
// Ensure that all builtin names are unique; otherwise profile-guided
// optimization might get confused.
CHECK(builtin_names.insert(data.function_name_).second);
} }
os << "---- End Profiling Data ----" << std::endl; os << "---- End Profiling Data ----" << std::endl;
} }
@ -171,6 +180,20 @@ std::vector<bool> BasicBlockProfiler::GetCoverageBitmap(Isolate* isolate) {
return out; return out;
} }
void BasicBlockProfilerData::Log(Isolate* isolate) {
bool any_nonzero_counter = false;
for (size_t i = 0; i < n_blocks(); ++i) {
if (counts_[i] > 0) {
any_nonzero_counter = true;
isolate->logger()->BasicBlockCounterEvent(function_name_.c_str(),
block_ids_[i], counts_[i]);
}
}
if (any_nonzero_counter) {
isolate->logger()->BuiltinHashEvent(function_name_.c_str(), hash_);
}
}
std::ostream& operator<<(std::ostream& os, const BasicBlockProfilerData& d) { std::ostream& operator<<(std::ostream& os, const BasicBlockProfilerData& d) {
int block_count_sum = std::accumulate(d.counts_.begin(), d.counts_.end(), 0); int block_count_sum = std::accumulate(d.counts_.begin(), d.counts_.end(), 0);
if (block_count_sum == 0) return os; if (block_count_sum == 0) return os;
@ -184,17 +207,17 @@ std::ostream& operator<<(std::ostream& os, const BasicBlockProfilerData& d) {
os << d.schedule_.c_str() << std::endl; os << d.schedule_.c_str() << std::endl;
} }
os << "block counts for " << name << ":" << std::endl; os << "block counts for " << name << ":" << std::endl;
std::vector<std::pair<int32_t, uint32_t>> pairs; std::vector<std::pair<size_t, uint32_t>> pairs;
pairs.reserve(d.n_blocks()); pairs.reserve(d.n_blocks());
for (size_t i = 0; i < d.n_blocks(); ++i) { for (size_t i = 0; i < d.n_blocks(); ++i) {
pairs.push_back(std::make_pair(d.block_rpo_numbers_[i], d.counts_[i])); pairs.push_back(std::make_pair(i, d.counts_[i]));
} }
std::sort(pairs.begin(), pairs.end(), std::sort(
[=](std::pair<int32_t, uint32_t> left, pairs.begin(), pairs.end(),
std::pair<int32_t, uint32_t> right) { [=](std::pair<size_t, uint32_t> left, std::pair<size_t, uint32_t> right) {
if (right.second == left.second) return left.first < right.first; if (right.second == left.second) return left.first < right.first;
return right.second < left.second; return right.second < left.second;
}); });
for (auto it : pairs) { for (auto it : pairs) {
if (it.second == 0) break; if (it.second == 0) break;
os << "block B" << it.first << " : " << it.second << std::endl; os << "block B" << it.first << " : " << it.second << std::endl;

View File

@ -28,21 +28,24 @@ class BasicBlockProfilerData {
OnHeapBasicBlockProfilerData js_heap_data); OnHeapBasicBlockProfilerData js_heap_data);
size_t n_blocks() const { size_t n_blocks() const {
DCHECK_EQ(block_rpo_numbers_.size(), counts_.size()); DCHECK_EQ(block_ids_.size(), counts_.size());
return block_rpo_numbers_.size(); return block_ids_.size();
} }
const uint32_t* counts() const { return &counts_[0]; } const uint32_t* counts() const { return &counts_[0]; }
void SetCode(const std::ostringstream& os); void SetCode(const std::ostringstream& os);
void SetFunctionName(std::unique_ptr<char[]> name); void SetFunctionName(std::unique_ptr<char[]> name);
void SetSchedule(const std::ostringstream& os); void SetSchedule(const std::ostringstream& os);
void SetBlockRpoNumber(size_t offset, int32_t block_rpo); void SetBlockId(size_t offset, int32_t id);
void SetHash(int hash);
// Copy the data from this object into an equivalent object stored on the JS // Copy the data from this object into an equivalent object stored on the JS
// heap, so that it can survive snapshotting and relocation. This must // heap, so that it can survive snapshotting and relocation. This must
// happen on the main thread during finalization of the compilation. // happen on the main thread during finalization of the compilation.
Handle<OnHeapBasicBlockProfilerData> CopyToJSHeap(Isolate* isolate); Handle<OnHeapBasicBlockProfilerData> CopyToJSHeap(Isolate* isolate);
void Log(Isolate* isolate);
private: private:
friend class BasicBlockProfiler; friend class BasicBlockProfiler;
friend std::ostream& operator<<(std::ostream& os, friend std::ostream& operator<<(std::ostream& os,
@ -50,11 +53,13 @@ class BasicBlockProfilerData {
V8_EXPORT_PRIVATE void ResetCounts(); V8_EXPORT_PRIVATE void ResetCounts();
std::vector<int32_t> block_rpo_numbers_; // These vectors are indexed by reverse post-order block number.
std::vector<int32_t> block_ids_;
std::vector<uint32_t> counts_; std::vector<uint32_t> counts_;
std::string function_name_; std::string function_name_;
std::string schedule_; std::string schedule_;
std::string code_; std::string code_;
int hash_;
DISALLOW_COPY_AND_ASSIGN(BasicBlockProfilerData); DISALLOW_COPY_AND_ASSIGN(BasicBlockProfilerData);
}; };

View File

@ -640,6 +640,9 @@ DEFINE_BOOL(turbo_profiling_verbose, false,
"enable basic block profiling in TurboFan, and include each " "enable basic block profiling in TurboFan, and include each "
"function's schedule and disassembly in the output") "function's schedule and disassembly in the output")
DEFINE_IMPLICATION(turbo_profiling_verbose, turbo_profiling) DEFINE_IMPLICATION(turbo_profiling_verbose, turbo_profiling)
DEFINE_BOOL(turbo_profiling_log_builtins, false,
"emit data about basic block usage in builtins to v8.log (requires "
"that V8 was built with v8_enable_builtins_profiling=true)")
DEFINE_BOOL(turbo_verify_allocation, DEBUG_BOOL, DEFINE_BOOL(turbo_verify_allocation, DEBUG_BOOL,
"verify register allocation in TurboFan") "verify register allocation in TurboFan")
DEFINE_BOOL(turbo_move_optimization, true, "optimize gap moves in TurboFan") DEFINE_BOOL(turbo_move_optimization, true, "optimize gap moves in TurboFan")
@ -1449,6 +1452,9 @@ DEFINE_BOOL(target_is_simulator, false,
"Instruct mksnapshot that the target is meant to run in the " "Instruct mksnapshot that the target is meant to run in the "
"simulator and it can generate simulator-specific instructions. " "simulator and it can generate simulator-specific instructions. "
"(mksnapshot only)") "(mksnapshot only)")
DEFINE_STRING(turbo_profiling_log_file, nullptr,
"Path of the input file containing basic block counters for "
"builtins. (mksnapshot only)")
// //
// Minor mark compact collector flags. // Minor mark compact collector flags.

View File

@ -9,6 +9,7 @@
#include "src/builtins/builtins-constructor-gen.h" #include "src/builtins/builtins-constructor-gen.h"
#include "src/builtins/builtins-iterator-gen.h" #include "src/builtins/builtins-iterator-gen.h"
#include "src/builtins/profile-data-reader.h"
#include "src/codegen/code-factory.h" #include "src/codegen/code-factory.h"
#include "src/debug/debug.h" #include "src/debug/debug.h"
#include "src/ic/accessor-assembler.h" #include "src/ic/accessor-assembler.h"
@ -3127,7 +3128,8 @@ Handle<Code> GenerateBytecodeHandler(Isolate* isolate, const char* debug_name,
#undef CALL_GENERATOR #undef CALL_GENERATOR
} }
Handle<Code> code = compiler::CodeAssembler::GenerateCode(&state, options); Handle<Code> code = compiler::CodeAssembler::GenerateCode(
&state, options, ProfileDataFromFile::TryRead(debug_name));
#ifdef ENABLE_DISASSEMBLER #ifdef ENABLE_DISASSEMBLER
if (FLAG_trace_ignition_codegen) { if (FLAG_trace_ignition_codegen) {

View File

@ -37,7 +37,8 @@ class Log {
FLAG_log_suspect || FLAG_ll_prof || FLAG_perf_basic_prof || FLAG_log_suspect || FLAG_ll_prof || FLAG_perf_basic_prof ||
FLAG_perf_prof || FLAG_log_source_code || FLAG_gdbjit || FLAG_perf_prof || FLAG_log_source_code || FLAG_gdbjit ||
FLAG_log_internal_timer_events || FLAG_prof_cpp || FLAG_trace_ic || FLAG_log_internal_timer_events || FLAG_prof_cpp || FLAG_trace_ic ||
FLAG_log_function_events || FLAG_trace_zone_stats; FLAG_log_function_events || FLAG_trace_zone_stats ||
FLAG_turbo_profiling_log_builtins;
} }
// Frees all resources acquired in Initialize and Open... functions. // Frees all resources acquired in Initialize and Open... functions.

View File

@ -10,6 +10,7 @@
#include "src/api/api-inl.h" #include "src/api/api-inl.h"
#include "src/base/platform/platform.h" #include "src/base/platform/platform.h"
#include "src/builtins/profile-data-reader.h"
#include "src/codegen/bailout-reason.h" #include "src/codegen/bailout-reason.h"
#include "src/codegen/macro-assembler.h" #include "src/codegen/macro-assembler.h"
#include "src/codegen/source-position-table.h" #include "src/codegen/source-position-table.h"
@ -34,11 +35,10 @@
#include "src/strings/unicode-inl.h" #include "src/strings/unicode-inl.h"
#include "src/tracing/tracing-category-observer.h" #include "src/tracing/tracing-category-observer.h"
#include "src/utils/memcopy.h" #include "src/utils/memcopy.h"
#include "src/utils/version.h"
#include "src/wasm/wasm-code-manager.h" #include "src/wasm/wasm-code-manager.h"
#include "src/wasm/wasm-objects-inl.h" #include "src/wasm/wasm-objects-inl.h"
#include "src/utils/version.h"
namespace v8 { namespace v8 {
namespace internal { namespace internal {
@ -1072,6 +1072,23 @@ void Logger::TimerEvent(Logger::StartEnd se, const char* name) {
msg.WriteToLogFile(); msg.WriteToLogFile();
} }
void Logger::BasicBlockCounterEvent(const char* name, int block_id,
uint32_t count) {
if (!log_->IsEnabled() || !FLAG_turbo_profiling_log_builtins) return;
Log::MessageBuilder msg(log_.get());
msg << ProfileDataFromFileConstants::kBlockCounterMarker << kNext << name
<< kNext << block_id << kNext << count;
msg.WriteToLogFile();
}
void Logger::BuiltinHashEvent(const char* name, int hash) {
if (!log_->IsEnabled() || !FLAG_turbo_profiling_log_builtins) return;
Log::MessageBuilder msg(log_.get());
msg << ProfileDataFromFileConstants::kBuiltinHashMarker << kNext << name
<< kNext << hash;
msg.WriteToLogFile();
}
bool Logger::is_logging() { bool Logger::is_logging() {
// Disable logging while the CPU profiler is running. // Disable logging while the CPU profiler is running.
if (isolate_->is_profiling()) return false; if (isolate_->is_profiling()) return false;

View File

@ -246,6 +246,10 @@ class Logger : public CodeEventListener {
V8_EXPORT_PRIVATE void TimerEvent(StartEnd se, const char* name); V8_EXPORT_PRIVATE void TimerEvent(StartEnd se, const char* name);
void BasicBlockCounterEvent(const char* name, int block_id, uint32_t count);
void BuiltinHashEvent(const char* name, int hash);
static void EnterExternal(Isolate* isolate); static void EnterExternal(Isolate* isolate);
static void LeaveExternal(Isolate* isolate); static void LeaveExternal(Isolate* isolate);

View File

@ -81,9 +81,10 @@ extern class UncompiledDataWithPreparseData extends UncompiledData {
@export @export
class OnHeapBasicBlockProfilerData extends HeapObject { class OnHeapBasicBlockProfilerData extends HeapObject {
block_rpo_numbers: ByteArray; // Stored as 4-byte ints block_ids: ByteArray; // Stored as 4-byte ints
counts: ByteArray; // Stored as 4-byte ints counts: ByteArray; // Stored as 4-byte ints
name: String; name: String;
schedule: String; schedule: String;
code: String; code: String;
hash: Smi;
} }

View File

@ -63,7 +63,7 @@ class CodeAssemblerTester {
if (state_.InsideBlock()) { if (state_.InsideBlock()) {
CodeAssembler(&state_).Unreachable(); CodeAssembler(&state_).Unreachable();
} }
return CodeAssembler::GenerateCode(&state_, options); return CodeAssembler::GenerateCode(&state_, options, nullptr);
} }
Handle<Code> GenerateCodeCloseAndEscape() { Handle<Code> GenerateCodeCloseAndEscape() {

View File

@ -40,7 +40,7 @@ class SchedulerTest : public TestWithIsolateAndZone {
} }
Schedule* schedule = Scheduler::ComputeSchedule( Schedule* schedule = Scheduler::ComputeSchedule(
zone(), graph(), Scheduler::kSplitNodes, tick_counter()); zone(), graph(), Scheduler::kSplitNodes, tick_counter(), nullptr);
if (FLAG_trace_turbo_scheduler) { if (FLAG_trace_turbo_scheduler) {
StdoutStream{} << *schedule << std::endl; StdoutStream{} << *schedule << std::endl;
@ -92,7 +92,7 @@ TEST_F(SchedulerTest, BuildScheduleEmpty) {
graph()->SetStart(graph()->NewNode(common()->Start(0))); graph()->SetStart(graph()->NewNode(common()->Start(0)));
graph()->SetEnd(graph()->NewNode(common()->End(1), graph()->start())); graph()->SetEnd(graph()->NewNode(common()->End(1), graph()->start()));
USE(Scheduler::ComputeSchedule(zone(), graph(), Scheduler::kNoFlags, USE(Scheduler::ComputeSchedule(zone(), graph(), Scheduler::kNoFlags,
tick_counter())); tick_counter(), nullptr));
} }
@ -107,7 +107,7 @@ TEST_F(SchedulerTest, BuildScheduleOneParameter) {
graph()->SetEnd(graph()->NewNode(common()->End(1), ret)); graph()->SetEnd(graph()->NewNode(common()->End(1), ret));
USE(Scheduler::ComputeSchedule(zone(), graph(), Scheduler::kNoFlags, USE(Scheduler::ComputeSchedule(zone(), graph(), Scheduler::kNoFlags,
tick_counter())); tick_counter(), nullptr));
} }