blob: 54a56f293907c3dbe7acd4c63e2134609c8b125c [file] [log] [blame]
/*
* Copyright 2014 The Android Open Source Project
*
* Licensed under the Apache License, Version 2.0 (the "License");
* you may not use this file except in compliance with the License.
* You may obtain a copy of the License at
*
* http://www.apache.org/licenses/LICENSE-2.0
*
* Unless required by applicable law or agreed to in writing, software
* distributed under the License is distributed on an "AS IS" BASIS,
* WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
* See the License for the specific language governing permissions and
* limitations under the License.
*/
#include "jit.h"
#include <dlfcn.h>
#include "art_method-inl.h"
#include "base/enums.h"
#include "base/file_utils.h"
#include "base/logging.h" // For VLOG.
#include "base/memfd.h"
#include "base/memory_tool.h"
#include "base/runtime_debug.h"
#include "base/scoped_flock.h"
#include "base/utils.h"
#include "class_root-inl.h"
#include "compilation_kind.h"
#include "debugger.h"
#include "dex/type_lookup_table.h"
#include "entrypoints/entrypoint_utils-inl.h"
#include "entrypoints/runtime_asm_entrypoints.h"
#include "gc/space/image_space.h"
#include "interpreter/interpreter.h"
#include "jit-inl.h"
#include "jit_code_cache.h"
#include "jit_create.h"
#include "jni/java_vm_ext.h"
#include "mirror/method_handle_impl.h"
#include "mirror/var_handle.h"
#include "oat/image-inl.h"
#include "oat/oat_file.h"
#include "oat/oat_file_manager.h"
#include "oat/oat_quick_method_header.h"
#include "oat/stack_map.h"
#include "profile/profile_boot_info.h"
#include "profile/profile_compilation_info.h"
#include "profile_saver.h"
#include "runtime.h"
#include "runtime_options.h"
#include "small_pattern_matcher.h"
#include "stack.h"
#include "thread-inl.h"
#include "thread_list.h"
using android::base::unique_fd;
namespace art HIDDEN {
namespace jit {
static constexpr bool kEnableOnStackReplacement = true;
// JIT compiler
JitCompilerInterface* Jit::jit_compiler_ = nullptr;
void Jit::DumpInfo(std::ostream& os) {
code_cache_->Dump(os);
cumulative_timings_.Dump(os);
MutexLock mu(Thread::Current(), lock_);
memory_use_.PrintMemoryUse(os);
}
void Jit::DumpForSigQuit(std::ostream& os) {
DumpInfo(os);
ProfileSaver::DumpInstanceInfo(os);
}
void Jit::AddTimingLogger(const TimingLogger& logger) {
cumulative_timings_.AddLogger(logger);
}
Jit::Jit(JitCodeCache* code_cache, JitOptions* options)
: code_cache_(code_cache),
options_(options),
boot_completed_lock_("Jit::boot_completed_lock_"),
cumulative_timings_("JIT timings"),
memory_use_("Memory used for compilation", 16),
lock_("JIT memory use lock"),
zygote_mapping_methods_(),
fd_methods_(-1),
fd_methods_size_(0) {}
std::unique_ptr<Jit> Jit::Create(JitCodeCache* code_cache, JitOptions* options) {
jit_compiler_ = jit_create();
std::unique_ptr<Jit> jit(new Jit(code_cache, options));
// If the code collector is enabled, check if that still holds:
// With 'perf', we want a 1-1 mapping between an address and a method.
// We aren't able to keep method pointers live during the instrumentation method entry trampoline
// so we will just disable jit-gc if we are doing that.
// JitAtFirstUse compiles the methods synchronously on mutator threads. While this should work
// in theory it is causing deadlocks in some jvmti tests related to Jit GC. Hence, disabling
// Jit GC for now (b/147208992).
if (code_cache->GetGarbageCollectCode()) {
code_cache->SetGarbageCollectCode(!jit_compiler_->GenerateDebugInfo() &&
!jit->JitAtFirstUse());
}
VLOG(jit) << "JIT created with initial_capacity="
<< PrettySize(options->GetCodeCacheInitialCapacity())
<< ", max_capacity=" << PrettySize(options->GetCodeCacheMaxCapacity())
<< ", warmup_threshold=" << options->GetWarmupThreshold()
<< ", optimize_threshold=" << options->GetOptimizeThreshold()
<< ", profile_saver_options=" << options->GetProfileSaverOptions();
// We want to know whether the compiler is compiling baseline, as this
// affects how we GC ProfilingInfos.
for (const std::string& option : Runtime::Current()->GetCompilerOptions()) {
if (option == "--baseline") {
options->SetUseBaselineCompiler();
break;
}
}
// Notify native debugger about the classes already loaded before the creation of the jit.
jit->DumpTypeInfoForLoadedTypes(Runtime::Current()->GetClassLinker());
return jit;
}
bool Jit::TryPatternMatch(ArtMethod* method_to_compile, CompilationKind compilation_kind) {
// Try to pattern match the method. Only on arm and arm64 for now as we have
// sufficiently similar calling convention between C++ and managed code.
if (kRuntimeISA == InstructionSet::kArm || kRuntimeISA == InstructionSet::kArm64) {
if (!Runtime::Current()->IsJavaDebuggable() &&
compilation_kind == CompilationKind::kBaseline &&
!method_to_compile->StillNeedsClinitCheck()) {
const void* pattern = SmallPatternMatcher::TryMatch(method_to_compile);
if (pattern != nullptr) {
VLOG(jit) << "Successfully pattern matched " << method_to_compile->PrettyMethod();
Runtime::Current()->GetInstrumentation()->UpdateMethodsCode(method_to_compile, pattern);
return true;
}
}
}
return false;
}
bool Jit::CompileMethodInternal(ArtMethod* method,
Thread* self,
CompilationKind compilation_kind,
bool prejit) {
DCHECK(Runtime::Current()->UseJitCompilation());
DCHECK(!method->IsRuntimeMethod());
// If the baseline flag was explicitly passed in the compiler options, change the compilation kind
// from optimized to baseline.
if (jit_compiler_->IsBaselineCompiler() && compilation_kind == CompilationKind::kOptimized) {
compilation_kind = CompilationKind::kBaseline;
}
// If we're asked to compile baseline, but we cannot allocate profiling infos,
// change the compilation kind to optimized.
if ((compilation_kind == CompilationKind::kBaseline) &&
!GetCodeCache()->CanAllocateProfilingInfo()) {
compilation_kind = CompilationKind::kOptimized;
}
// Don't compile the method if it has breakpoints.
if (Runtime::Current()->GetInstrumentation()->IsDeoptimized(method)) {
VLOG(jit) << "JIT not compiling " << method->PrettyMethod()
<< " due to not being safe to jit according to runtime-callbacks. For example, there"
<< " could be breakpoints in this method.";
return false;
}
if (!method->IsCompilable()) {
DCHECK(method->GetDeclaringClass()->IsObsoleteObject() ||
method->IsProxyMethod()) << method->PrettyMethod();
VLOG(jit) << "JIT not compiling " << method->PrettyMethod() << " due to method being made "
<< "obsolete while waiting for JIT task to run. This probably happened due to "
<< "concurrent structural class redefinition.";
return false;
}
// Don't compile the method if we are supposed to be deoptimized.
instrumentation::Instrumentation* instrumentation = Runtime::Current()->GetInstrumentation();
if (instrumentation->AreAllMethodsDeoptimized() || instrumentation->IsDeoptimized(method)) {
VLOG(jit) << "JIT not compiling " << method->PrettyMethod() << " due to deoptimization";
return false;
}
JitMemoryRegion* region = GetCodeCache()->GetCurrentRegion();
if ((compilation_kind == CompilationKind::kOsr) && GetCodeCache()->IsSharedRegion(*region)) {
VLOG(jit) << "JIT not osr compiling "
<< method->PrettyMethod()
<< " due to using shared region";
return false;
}
// If we get a request to compile a proxy method, we pass the actual Java method
// of that proxy method, as the compiler does not expect a proxy method.
ArtMethod* method_to_compile = method->GetInterfaceMethodIfProxy(kRuntimePointerSize);
if (TryPatternMatch(method_to_compile, compilation_kind)) {
return true;
}
if (!code_cache_->NotifyCompilationOf(method_to_compile, self, compilation_kind, prejit)) {
return false;
}
VLOG(jit) << "Compiling method "
<< ArtMethod::PrettyMethod(method_to_compile)
<< " kind=" << compilation_kind;
bool success = jit_compiler_->CompileMethod(self, region, method_to_compile, compilation_kind);
code_cache_->DoneCompiling(method_to_compile, self);
if (!success) {
VLOG(jit) << "Failed to compile method "
<< ArtMethod::PrettyMethod(method_to_compile)
<< " kind=" << compilation_kind;
}
if (kIsDebugBuild) {
if (self->IsExceptionPending()) {
mirror::Throwable* exception = self->GetException();
LOG(FATAL) << "No pending exception expected after compiling "
<< ArtMethod::PrettyMethod(method)
<< ": "
<< exception->Dump();
}
}
return success;
}
void Jit::WaitForWorkersToBeCreated() {
if (thread_pool_ != nullptr) {
thread_pool_->WaitForWorkersToBeCreated();
}
}
void Jit::DeleteThreadPool() {
Thread* self = Thread::Current();
if (thread_pool_ != nullptr) {
std::unique_ptr<JitThreadPool> pool;
{
ScopedSuspendAll ssa(__FUNCTION__);
// Clear thread_pool_ field while the threads are suspended.
// A mutator in the 'AddSamples' method will check against it.
pool = std::move(thread_pool_);
}
// When running sanitized, let all tasks finish to not leak. Otherwise just clear the queue.
if (!kRunningOnMemoryTool) {
pool->StopWorkers(self);
pool->RemoveAllTasks(self);
}
// We could just suspend all threads, but we know those threads
// will finish in a short period, so it's not worth adding a suspend logic
// here. Besides, this is only done for shutdown.
pool->Wait(self, false, false);
}
}
void Jit::StartProfileSaver(const std::string& profile_filename,
const std::vector<std::string>& code_paths,
const std::string& ref_profile_filename) {
if (options_->GetSaveProfilingInfo()) {
ProfileSaver::Start(options_->GetProfileSaverOptions(),
profile_filename,
code_cache_,
code_paths,
ref_profile_filename);
}
}
void Jit::StopProfileSaver() {
if (options_->GetSaveProfilingInfo() && ProfileSaver::IsStarted()) {
ProfileSaver::Stop(options_->DumpJitInfoOnShutdown());
}
}
bool Jit::JitAtFirstUse() {
return HotMethodThreshold() == 0;
}
bool Jit::CanInvokeCompiledCode(ArtMethod* method) {
return code_cache_->ContainsPc(method->GetEntryPointFromQuickCompiledCode());
}
Jit::~Jit() {
DCHECK_IMPLIES(options_->GetSaveProfilingInfo(), !ProfileSaver::IsStarted());
if (options_->DumpJitInfoOnShutdown()) {
DumpInfo(LOG_STREAM(INFO));
Runtime::Current()->DumpDeoptimizations(LOG_STREAM(INFO));
}
DeleteThreadPool();
if (jit_compiler_ != nullptr) {
delete jit_compiler_;
jit_compiler_ = nullptr;
}
}
void Jit::NewTypeLoadedIfUsingJit(mirror::Class* type) {
if (!Runtime::Current()->UseJitCompilation()) {
// No need to notify if we only use the JIT to save profiles.
return;
}
jit::Jit* jit = Runtime::Current()->GetJit();
if (jit->jit_compiler_->GenerateDebugInfo()) {
jit_compiler_->TypesLoaded(&type, 1);
}
}
void Jit::DumpTypeInfoForLoadedTypes(ClassLinker* linker) {
struct CollectClasses : public ClassVisitor {
bool operator()(ObjPtr<mirror::Class> klass) override REQUIRES_SHARED(Locks::mutator_lock_) {
classes_.push_back(klass.Ptr());
return true;
}
std::vector<mirror::Class*> classes_;
};
if (jit_compiler_->GenerateDebugInfo()) {
ScopedObjectAccess so(Thread::Current());
CollectClasses visitor;
linker->VisitClasses(&visitor);
jit_compiler_->TypesLoaded(visitor.classes_.data(), visitor.classes_.size());
}
}
extern "C" void art_quick_osr_stub(void** stack,
size_t stack_size_in_bytes,
const uint8_t* native_pc,
JValue* result,
const char* shorty,
Thread* self);
OsrData* Jit::PrepareForOsr(ArtMethod* method, uint32_t dex_pc, uint32_t* vregs) {
if (!kEnableOnStackReplacement) {
return nullptr;
}
// Cheap check if the method has been compiled already. That's an indicator that we should
// osr into it.
if (!GetCodeCache()->ContainsPc(method->GetEntryPointFromQuickCompiledCode())) {
return nullptr;
}
// Fetch some data before looking up for an OSR method. We don't want thread
// suspension once we hold an OSR method, as the JIT code cache could delete the OSR
// method while we are being suspended.
CodeItemDataAccessor accessor(method->DexInstructionData());
const size_t number_of_vregs = accessor.RegistersSize();
std::string method_name(VLOG_IS_ON(jit) ? method->PrettyMethod() : "");
OsrData* osr_data = nullptr;
{
ScopedAssertNoThreadSuspension sts("Holding OSR method");
const OatQuickMethodHeader* osr_method = GetCodeCache()->LookupOsrMethodHeader(method);
if (osr_method == nullptr) {
// No osr method yet, just return to the interpreter.
return nullptr;
}
CodeInfo code_info(osr_method);
// Find stack map starting at the target dex_pc.
StackMap stack_map = code_info.GetOsrStackMapForDexPc(dex_pc);
if (!stack_map.IsValid()) {
// There is no OSR stack map for this dex pc offset. Just return to the interpreter in the
// hope that the next branch has one.
return nullptr;
}
// We found a stack map, now fill the frame with dex register values from the interpreter's
// shadow frame.
DexRegisterMap vreg_map = code_info.GetDexRegisterMapOf(stack_map);
DCHECK_EQ(vreg_map.size(), number_of_vregs);
size_t frame_size = osr_method->GetFrameSizeInBytes();
// Allocate memory to put shadow frame values. The osr stub will copy that memory to
// stack.
// Note that we could pass the shadow frame to the stub, and let it copy the values there,
// but that is engineering complexity not worth the effort for something like OSR.
osr_data = reinterpret_cast<OsrData*>(malloc(sizeof(OsrData) + frame_size));
if (osr_data == nullptr) {
return nullptr;
}
memset(osr_data, 0, sizeof(OsrData) + frame_size);
osr_data->frame_size = frame_size;
// Art ABI: ArtMethod is at the bottom of the stack.
osr_data->memory[0] = method;
if (vreg_map.empty()) {
// If we don't have a dex register map, then there are no live dex registers at
// this dex pc.
} else {
for (uint16_t vreg = 0; vreg < number_of_vregs; ++vreg) {
DexRegisterLocation::Kind location = vreg_map[vreg].GetKind();
if (location == DexRegisterLocation::Kind::kNone) {
// Dex register is dead or uninitialized.
continue;
}
if (location == DexRegisterLocation::Kind::kConstant) {
// We skip constants because the compiled code knows how to handle them.
continue;
}
DCHECK_EQ(location, DexRegisterLocation::Kind::kInStack);
int32_t vreg_value = vregs[vreg];
int32_t slot_offset = vreg_map[vreg].GetStackOffsetInBytes();
DCHECK_LT(slot_offset, static_cast<int32_t>(frame_size));
DCHECK_GT(slot_offset, 0);
(reinterpret_cast<int32_t*>(osr_data->memory))[slot_offset / sizeof(int32_t)] = vreg_value;
}
}
osr_data->native_pc = stack_map.GetNativePcOffset(kRuntimeISA) +
osr_method->GetEntryPoint();
VLOG(jit) << "Jumping to "
<< method_name
<< "@"
<< std::hex << reinterpret_cast<uintptr_t>(osr_data->native_pc);
}
return osr_data;
}
bool Jit::MaybeDoOnStackReplacement(Thread* thread,
ArtMethod* method,
uint32_t dex_pc,
int32_t dex_pc_offset,
JValue* result) {
Jit* jit = Runtime::Current()->GetJit();
if (jit == nullptr) {
return false;
}
if (UNLIKELY(__builtin_frame_address(0) < thread->GetStackEnd())) {
// Don't attempt to do an OSR if we are close to the stack limit. Since
// the interpreter frames are still on stack, OSR has the potential
// to stack overflow even for a simple loop.
// b/27094810.
return false;
}
// Get the actual Java method if this method is from a proxy class. The compiler
// and the JIT code cache do not expect methods from proxy classes.
method = method->GetInterfaceMethodIfProxy(kRuntimePointerSize);
// Before allowing the jump, make sure no code is actively inspecting the method to avoid
// jumping from interpreter to OSR while e.g. single stepping. Note that we could selectively
// disable OSR when single stepping, but that's currently hard to know at this point.
// Currently, HaveLocalsChanged is not frame specific. It is possible to make it frame specific
// to allow OSR of frames that don't have any locals changed but it isn't worth the additional
// complexity.
if (Runtime::Current()->GetInstrumentation()->NeedsSlowInterpreterForMethod(thread, method) ||
Runtime::Current()->GetRuntimeCallbacks()->HaveLocalsChanged()) {
return false;
}
ShadowFrame* shadow_frame = thread->GetManagedStack()->GetTopShadowFrame();
OsrData* osr_data = jit->PrepareForOsr(method,
dex_pc + dex_pc_offset,
shadow_frame->GetVRegArgs(0));
if (osr_data == nullptr) {
return false;
}
{
thread->PopShadowFrame();
ManagedStack fragment;
thread->PushManagedStackFragment(&fragment);
(*art_quick_osr_stub)(osr_data->memory,
osr_data->frame_size,
osr_data->native_pc,
result,
method->GetShorty(),
thread);
if (UNLIKELY(thread->GetException() == Thread::GetDeoptimizationException())) {
thread->DeoptimizeWithDeoptimizationException(result);
}
thread->PopManagedStackFragment(fragment);
}
free(osr_data);
thread->PushShadowFrame(shadow_frame);
VLOG(jit) << "Done running OSR code for " << method->PrettyMethod();
return true;
}
void Jit::AddMemoryUsage(ArtMethod* method, size_t bytes) {
if (bytes > 4 * MB) {
LOG(INFO) << "Compiler allocated "
<< PrettySize(bytes)
<< " to compile "
<< ArtMethod::PrettyMethod(method);
}
MutexLock mu(Thread::Current(), lock_);
memory_use_.AddValue(bytes);
}
void Jit::NotifyZygoteCompilationDone() {
if (fd_methods_ == -1) {
return;
}
size_t offset = 0;
for (gc::space::ImageSpace* space : Runtime::Current()->GetHeap()->GetBootImageSpaces()) {
const ImageHeader& header = space->GetImageHeader();
const ImageSection& section = header.GetMethodsSection();
// Because mremap works at page boundaries, we can only handle methods
// within a page range. For methods that falls above or below the range,
// the child processes will copy their contents to their private mapping
// in `child_mapping_methods`. See `MapBootImageMethods`.
uint8_t* page_start = AlignUp(header.GetImageBegin() + section.Offset(), gPageSize);
uint8_t* page_end =
AlignDown(header.GetImageBegin() + section.Offset() + section.Size(), gPageSize);
if (page_end > page_start) {
uint64_t capacity = page_end - page_start;
memcpy(zygote_mapping_methods_.Begin() + offset, page_start, capacity);
offset += capacity;
}
}
// Do an msync to ensure we are not affected by writes still being in caches.
if (msync(zygote_mapping_methods_.Begin(), fd_methods_size_, MS_SYNC) != 0) {
PLOG(WARNING) << "Failed to sync boot image methods memory";
code_cache_->GetZygoteMap()->SetCompilationState(ZygoteCompilationState::kNotifiedFailure);
return;
}
// We don't need the shared mapping anymore, and we need to drop it in case
// the file hasn't been sealed writable.
zygote_mapping_methods_ = MemMap::Invalid();
// Seal writes now. Zygote and children will map the memory private in order
// to write to it.
if (fcntl(fd_methods_, F_ADD_SEALS, F_SEAL_SEAL | F_SEAL_WRITE) == -1) {
PLOG(WARNING) << "Failed to seal boot image methods file descriptor";
code_cache_->GetZygoteMap()->SetCompilationState(ZygoteCompilationState::kNotifiedFailure);
return;
}
std::string error_str;
MemMap child_mapping_methods = MemMap::MapFile(
fd_methods_size_,
PROT_READ | PROT_WRITE,
MAP_PRIVATE,
fd_methods_,
/* start= */ 0,
/* low_4gb= */ false,
"boot-image-methods",
&error_str);
if (!child_mapping_methods.IsValid()) {
LOG(WARNING) << "Failed to create child mapping of boot image methods: " << error_str;
code_cache_->GetZygoteMap()->SetCompilationState(ZygoteCompilationState::kNotifiedFailure);
return;
}
// Ensure the contents are the same as before: there was a window between
// the memcpy and the sealing where other processes could have changed the
// contents.
// Note this would not be needed if we could have used F_SEAL_FUTURE_WRITE,
// see b/143833776.
offset = 0;
for (gc::space::ImageSpace* space : Runtime::Current()->GetHeap()->GetBootImageSpaces()) {
const ImageHeader& header = space->GetImageHeader();
const ImageSection& section = header.GetMethodsSection();
// Because mremap works at page boundaries, we can only handle methods
// within a page range. For methods that falls above or below the range,
// the child processes will copy their contents to their private mapping
// in `child_mapping_methods`. See `MapBootImageMethods`.
uint8_t* page_start = AlignUp(header.GetImageBegin() + section.Offset(), gPageSize);
uint8_t* page_end =
AlignDown(header.GetImageBegin() + section.Offset() + section.Size(), gPageSize);
if (page_end > page_start) {
uint64_t capacity = page_end - page_start;
if (memcmp(child_mapping_methods.Begin() + offset, page_start, capacity) != 0) {
LOG(WARNING) << "Contents differ in boot image methods data";
code_cache_->GetZygoteMap()->SetCompilationState(
ZygoteCompilationState::kNotifiedFailure);
return;
}
offset += capacity;
}
}
// Future spawned processes don't need the fd anymore.
fd_methods_.reset();
// In order to have the zygote and children share the memory, we also remap
// the memory into the zygote process.
offset = 0;
for (gc::space::ImageSpace* space : Runtime::Current()->GetHeap()->GetBootImageSpaces()) {
const ImageHeader& header = space->GetImageHeader();
const ImageSection& section = header.GetMethodsSection();
// Because mremap works at page boundaries, we can only handle methods
// within a page range. For methods that falls above or below the range,
// the child processes will copy their contents to their private mapping
// in `child_mapping_methods`. See `MapBootImageMethods`.
uint8_t* page_start = AlignUp(header.GetImageBegin() + section.Offset(), gPageSize);
uint8_t* page_end =
AlignDown(header.GetImageBegin() + section.Offset() + section.Size(), gPageSize);
if (page_end > page_start) {
uint64_t capacity = page_end - page_start;
if (mremap(child_mapping_methods.Begin() + offset,
capacity,
capacity,
MREMAP_FIXED | MREMAP_MAYMOVE,
page_start) == MAP_FAILED) {
// Failing to remap is safe as the process will just use the old
// contents.
PLOG(WARNING) << "Failed mremap of boot image methods of " << space->GetImageFilename();
}
offset += capacity;
}
}
LOG(INFO) << "Successfully notified child processes on sharing boot image methods";
// Mark that compilation of boot classpath is done, and memory can now be
// shared. Other processes will pick up this information.
code_cache_->GetZygoteMap()->SetCompilationState(ZygoteCompilationState::kNotifiedOk);
// The private mapping created for this process has been mremaped. We can
// reset it.
child_mapping_methods.Reset();
}
class JitCompileTask final : public Task {
public:
enum class TaskKind {
kCompile,
kPreCompile,
};
JitCompileTask(ArtMethod* method,
TaskKind task_kind,
CompilationKind compilation_kind)
: method_(method),
kind_(task_kind),
compilation_kind_(compilation_kind) {
}
void Run(Thread* self) override {
{
ScopedObjectAccess soa(self);
switch (kind_) {
case TaskKind::kCompile:
case TaskKind::kPreCompile: {
Runtime::Current()->GetJit()->CompileMethodInternal(
method_,
self,
compilation_kind_,
/* prejit= */ (kind_ == TaskKind::kPreCompile));
break;
}
}
}
ProfileSaver::NotifyJitActivity();
}
void Finalize() override {
JitThreadPool* thread_pool = Runtime::Current()->GetJit()->GetThreadPool();
if (thread_pool != nullptr) {
thread_pool->Remove(this);
}
delete this;
}
ArtMethod* GetArtMethod() const {
return method_;
}
CompilationKind GetCompilationKind() const {
return compilation_kind_;
}
private:
ArtMethod* const method_;
const TaskKind kind_;
const CompilationKind compilation_kind_;
DISALLOW_IMPLICIT_CONSTRUCTORS(JitCompileTask);
};
static std::string GetProfileFile(const std::string& dex_location) {
// Hardcoded assumption where the profile file is.
// TODO(ngeoffray): this is brittle and we would need to change change if we
// wanted to do more eager JITting of methods in a profile. This is
// currently only for system server.
return dex_location + ".prof";
}
static std::string GetBootProfileFile(const std::string& profile) {
// The boot profile can be found next to the compilation profile, with a
// different extension.
return ReplaceFileExtension(profile, "bprof");
}
/**
* A JIT task to run after all profile compilation is done.
*/
class JitDoneCompilingProfileTask final : public SelfDeletingTask {
public:
explicit JitDoneCompilingProfileTask(const std::vector<const DexFile*>& dex_files)
: dex_files_(dex_files) {}
void Run([[maybe_unused]] Thread* self) override {
// Madvise DONTNEED dex files now that we're done compiling methods.
for (const DexFile* dex_file : dex_files_) {
if (IsAddressKnownBackedByFileOrShared(dex_file->Begin())) {
int result = madvise(const_cast<uint8_t*>(AlignDown(dex_file->Begin(), gPageSize)),
RoundUp(dex_file->Size(), gPageSize),
MADV_DONTNEED);
if (result == -1) {
PLOG(WARNING) << "Madvise failed";
}
}
}
}
private:
std::vector<const DexFile*> dex_files_;
DISALLOW_COPY_AND_ASSIGN(JitDoneCompilingProfileTask);
};
class JitZygoteDoneCompilingTask final : public SelfDeletingTask {
public:
JitZygoteDoneCompilingTask() {}
void Run([[maybe_unused]] Thread* self) override {
DCHECK(Runtime::Current()->IsZygote());
Runtime::Current()->GetJit()->GetCodeCache()->GetZygoteMap()->SetCompilationState(
ZygoteCompilationState::kDone);
}
private:
DISALLOW_COPY_AND_ASSIGN(JitZygoteDoneCompilingTask);
};
/**
* A JIT task to run Java verification of boot classpath classes that were not
* verified at compile-time.
*/
class ZygoteVerificationTask final : public Task {
public:
ZygoteVerificationTask() {}
void Run(Thread* self) override {
// We are going to load class and run verification, which may also need to load
// classes. If the thread cannot load classes (typically when the runtime is
// debuggable), then just return.
if (!self->CanLoadClasses()) {
return;
}
Runtime* runtime = Runtime::Current();
ClassLinker* linker = runtime->GetClassLinker();
const std::vector<const DexFile*>& boot_class_path =
runtime->GetClassLinker()->GetBootClassPath();
ScopedObjectAccess soa(self);
StackHandleScope<1> hs(self);
MutableHandle<mirror::Class> klass = hs.NewHandle<mirror::Class>(nullptr);
uint64_t start_ns = ThreadCpuNanoTime();
uint64_t number_of_classes = 0;
for (const DexFile* dex_file : boot_class_path) {
for (uint32_t i = 0; i < dex_file->NumClassDefs(); ++i) {
const dex::ClassDef& class_def = dex_file->GetClassDef(i);
const char* descriptor = dex_file->GetClassDescriptor(class_def);
klass.Assign(linker->LookupResolvedType(descriptor, /* class_loader= */ nullptr));
if (klass == nullptr) {
// Class not loaded yet.
DCHECK(!self->IsExceptionPending());
continue;
}
if (klass->IsVerified()) {
continue;
}
if (linker->VerifyClass(self, /* verifier_deps= */ nullptr, klass) ==
verifier::FailureKind::kHardFailure) {
CHECK(self->IsExceptionPending());
LOG(WARNING) << "Methods in the boot classpath failed to verify: "
<< self->GetException()->Dump();
self->ClearException();
} else {
++number_of_classes;
}
CHECK(!self->IsExceptionPending());
}
}
LOG(INFO) << "Background verification of "
<< number_of_classes
<< " classes from boot classpath took "
<< PrettyDuration(ThreadCpuNanoTime() - start_ns);
}
};
class ZygoteTask final : public Task {
public:
ZygoteTask() {}
void Run(Thread* self) override {
Runtime* runtime = Runtime::Current();
uint32_t added_to_queue = 0;
for (gc::space::ImageSpace* space : Runtime::Current()->GetHeap()->GetBootImageSpaces()) {
const std::vector<const DexFile*>& boot_class_path =
runtime->GetClassLinker()->GetBootClassPath();
ScopedNullHandle<mirror::ClassLoader> null_handle;
// We avoid doing compilation at boot for the secondary zygote, as apps forked from it are not
// critical for boot.
if (Runtime::Current()->IsPrimaryZygote()) {
for (const std::string& profile_file : space->GetProfileFiles()) {
std::string boot_profile = GetBootProfileFile(profile_file);
LOG(INFO) << "JIT Zygote looking at boot profile " << boot_profile;
// We add to the queue for zygote so that we can fork processes in-between compilations.
added_to_queue += runtime->GetJit()->CompileMethodsFromBootProfile(
self, boot_class_path, boot_profile, null_handle, /* add_to_queue= */ true);
}
}
for (const std::string& profile_file : space->GetProfileFiles()) {
LOG(INFO) << "JIT Zygote looking at profile " << profile_file;
added_to_queue += runtime->GetJit()->CompileMethodsFromProfile(
self, boot_class_path, profile_file, null_handle, /* add_to_queue= */ true);
}
}
DCHECK(runtime->GetJit()->InZygoteUsingJit());
runtime->GetJit()->AddPostBootTask(self, new JitZygoteDoneCompilingTask());
JitCodeCache* code_cache = runtime->GetJit()->GetCodeCache();
code_cache->GetZygoteMap()->Initialize(added_to_queue);
}
void Finalize() override {
delete this;
}
private:
DISALLOW_COPY_AND_ASSIGN(ZygoteTask);
};
class JitProfileTask final : public Task {
public:
JitProfileTask(const std::vector<std::unique_ptr<const DexFile>>& dex_files,
jobject class_loader) {
ScopedObjectAccess soa(Thread::Current());
StackHandleScope<1> hs(soa.Self());
Handle<mirror::ClassLoader> h_loader(hs.NewHandle(
soa.Decode<mirror::ClassLoader>(class_loader)));
ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
for (const auto& dex_file : dex_files) {
dex_files_.push_back(dex_file.get());
// Register the dex file so that we can guarantee it doesn't get deleted
// while reading it during the task.
class_linker->RegisterDexFile(*dex_file.get(), h_loader.Get());
}
// We also create our own global ref to use this class loader later.
class_loader_ = soa.Vm()->AddGlobalRef(soa.Self(), h_loader.Get());
}
void Run(Thread* self) override {
ScopedObjectAccess soa(self);
StackHandleScope<1> hs(self);
Handle<mirror::ClassLoader> loader = hs.NewHandle<mirror::ClassLoader>(
soa.Decode<mirror::ClassLoader>(class_loader_));
std::string profile = GetProfileFile(dex_files_[0]->GetLocation());
std::string boot_profile = GetBootProfileFile(profile);
Jit* jit = Runtime::Current()->GetJit();
jit->CompileMethodsFromBootProfile(
self,
dex_files_,
boot_profile,
loader,
/* add_to_queue= */ false);
jit->CompileMethodsFromProfile(
self,
dex_files_,
profile,
loader,
/* add_to_queue= */ true);
}
void Finalize() override {
delete this;
}
~JitProfileTask() {
ScopedObjectAccess soa(Thread::Current());
soa.Vm()->DeleteGlobalRef(soa.Self(), class_loader_);
}
private:
std::vector<const DexFile*> dex_files_;
jobject class_loader_;
DISALLOW_COPY_AND_ASSIGN(JitProfileTask);
};
static void CopyIfDifferent(void* s1, const void* s2, size_t n) {
if (memcmp(s1, s2, n) != 0) {
memcpy(s1, s2, n);
}
}
void Jit::MapBootImageMethods() {
if (Runtime::Current()->IsJavaDebuggable()) {
LOG(INFO) << "Not mapping boot image methods due to process being debuggable";
return;
}
CHECK_NE(fd_methods_.get(), -1);
if (!code_cache_->GetZygoteMap()->CanMapBootImageMethods()) {
LOG(WARNING) << "Not mapping boot image methods due to error from zygote";
// We don't need the fd anymore.
fd_methods_.reset();
return;
}
std::string error_str;
MemMap child_mapping_methods = MemMap::MapFile(
fd_methods_size_,
PROT_READ | PROT_WRITE,
MAP_PRIVATE,
fd_methods_,
/* start= */ 0,
/* low_4gb= */ false,
"boot-image-methods",
&error_str);
// We don't need the fd anymore.
fd_methods_.reset();
if (!child_mapping_methods.IsValid()) {
LOG(WARNING) << "Failed to create child mapping of boot image methods: " << error_str;
return;
}
// We are going to mremap the child mapping into the image:
//
// ImageSection ChildMappingMethods
//
// section start --> -----------
// | |
// | |
// page_start --> | | <----- -----------
// | | | |
// | | | |
// | | | |
// | | | |
// | | | |
// | | | |
// | | | |
// page_end --> | | <----- -----------
// | |
// section end --> -----------
//
size_t offset = 0;
for (gc::space::ImageSpace* space : Runtime::Current()->GetHeap()->GetBootImageSpaces()) {
const ImageHeader& header = space->GetImageHeader();
const ImageSection& section = header.GetMethodsSection();
uint8_t* page_start = AlignUp(header.GetImageBegin() + section.Offset(), gPageSize);
uint8_t* page_end =
AlignDown(header.GetImageBegin() + section.Offset() + section.Size(), gPageSize);
if (page_end <= page_start) {
// Section doesn't contain one aligned entire page.
continue;
}
uint64_t capacity = page_end - page_start;
// Walk over methods in the boot image, and check for:
// 1) methods whose class is not initialized in the process, but are in the
// zygote process. For such methods, we need their entrypoints to be stubs
// that do the initialization check.
// 2) native methods whose data pointer is different than the one in the
// zygote. Such methods may have had custom native implementation provided
// by JNI RegisterNatives.
header.VisitPackedArtMethods([&](ArtMethod& method) NO_THREAD_SAFETY_ANALYSIS {
// Methods in the boot image should never have their single
// implementation flag set (and therefore never have a `data_` pointing
// to an ArtMethod for single implementation).
CHECK(method.IsIntrinsic() || !method.HasSingleImplementationFlag());
if (method.IsRuntimeMethod()) {
return;
}
// Pointer to the method we're currently using.
uint8_t* pointer = reinterpret_cast<uint8_t*>(&method);
// The data pointer of that method that we want to keep.
uint8_t* data_pointer = pointer + ArtMethod::DataOffset(kRuntimePointerSize).Int32Value();
if (method.IsNative() && data_pointer >= page_start && data_pointer < page_end) {
// The data pointer of the ArtMethod in the shared memory we are going to remap into our
// own mapping. This is the data that we will see after the remap.
uint8_t* new_data_pointer =
child_mapping_methods.Begin() + offset + (data_pointer - page_start);
CopyIfDifferent(new_data_pointer, data_pointer, sizeof(void*));
}
// The entrypoint of the method we're currently using and that we want to
// keep.
uint8_t* entry_point_pointer = pointer +
ArtMethod::EntryPointFromQuickCompiledCodeOffset(kRuntimePointerSize).Int32Value();
if (!method.GetDeclaringClassUnchecked()->IsVisiblyInitialized() &&
method.IsStatic() &&
!method.IsConstructor() &&
entry_point_pointer >= page_start &&
entry_point_pointer < page_end) {
// The entry point of the ArtMethod in the shared memory we are going to remap into our
// own mapping. This is the entrypoint that we will see after the remap.
uint8_t* new_entry_point_pointer =
child_mapping_methods.Begin() + offset + (entry_point_pointer - page_start);
CopyIfDifferent(new_entry_point_pointer, entry_point_pointer, sizeof(void*));
}
}, space->Begin(), kRuntimePointerSize);
// Map the memory in the boot image range.
if (mremap(child_mapping_methods.Begin() + offset,
capacity,
capacity,
MREMAP_FIXED | MREMAP_MAYMOVE,
page_start) == MAP_FAILED) {
PLOG(WARNING) << "Fail to mremap boot image methods for " << space->GetImageFilename();
}
offset += capacity;
}
// The private mapping created for this process has been mremaped. We can
// reset it.
child_mapping_methods.Reset();
LOG(INFO) << "Successfully mapped boot image methods";
}
bool Jit::InZygoteUsingJit() {
Runtime* runtime = Runtime::Current();
return runtime->IsZygote() && runtime->HasImageWithProfile() && runtime->UseJitCompilation();
}
void Jit::CreateThreadPool() {
// There is a DCHECK in the 'AddSamples' method to ensure the tread pool
// is not null when we instrument.
thread_pool_.reset(JitThreadPool::Create("Jit thread pool", 1));
Runtime* runtime = Runtime::Current();
thread_pool_->SetPthreadPriority(
runtime->IsZygote()
? options_->GetZygoteThreadPoolPthreadPriority()
: options_->GetThreadPoolPthreadPriority());
Start();
if (runtime->IsZygote()) {
// To speed up class lookups, generate a type lookup table for
// dex files not backed by oat file.
for (const DexFile* dex_file : runtime->GetClassLinker()->GetBootClassPath()) {
if (dex_file->GetOatDexFile() == nullptr) {
TypeLookupTable type_lookup_table = TypeLookupTable::Create(*dex_file);
type_lookup_tables_.push_back(
std::make_unique<art::OatDexFile>(std::move(type_lookup_table)));
dex_file->SetOatDexFile(type_lookup_tables_.back().get());
}
}
// Add a task that will verify boot classpath jars that were not
// pre-compiled.
thread_pool_->AddTask(Thread::Current(), new ZygoteVerificationTask());
}
if (InZygoteUsingJit()) {
// If we have an image with a profile, request a JIT task to
// compile all methods in that profile.
thread_pool_->AddTask(Thread::Current(), new ZygoteTask());
// And create mappings to share boot image methods memory from the zygote to
// child processes.
// Compute the total capacity required for the boot image methods.
uint64_t total_capacity = 0;
for (gc::space::ImageSpace* space : Runtime::Current()->GetHeap()->GetBootImageSpaces()) {
const ImageHeader& header = space->GetImageHeader();
const ImageSection& section = header.GetMethodsSection();
// Mappings need to be at the page level.
uint8_t* page_start = AlignUp(header.GetImageBegin() + section.Offset(), gPageSize);
uint8_t* page_end =
AlignDown(header.GetImageBegin() + section.Offset() + section.Size(), gPageSize);
if (page_end > page_start) {
total_capacity += (page_end - page_start);
}
}
// Create the child and zygote mappings to the boot image methods.
if (total_capacity > 0) {
// Start with '/boot' and end with '.art' to match the pattern recognized
// by android_os_Debug.cpp for boot images.
const char* name = "/boot-image-methods.art";
unique_fd mem_fd =
unique_fd(art::memfd_create(name, /* flags= */ MFD_ALLOW_SEALING | MFD_CLOEXEC));
if (mem_fd.get() == -1) {
PLOG(WARNING) << "Could not create boot image methods file descriptor";
return;
}
if (ftruncate(mem_fd.get(), total_capacity) != 0) {
PLOG(WARNING) << "Failed to truncate boot image methods file to " << total_capacity;
return;
}
std::string error_str;
// Create the shared mapping eagerly, as this prevents other processes
// from adding the writable seal.
zygote_mapping_methods_ = MemMap::MapFile(
total_capacity,
PROT_READ | PROT_WRITE,
MAP_SHARED,
mem_fd,
/* start= */ 0,
/* low_4gb= */ false,
"boot-image-methods",
&error_str);
if (!zygote_mapping_methods_.IsValid()) {
LOG(WARNING) << "Failed to create zygote mapping of boot image methods: " << error_str;
return;
}
if (zygote_mapping_methods_.MadviseDontFork() != 0) {
LOG(WARNING) << "Failed to madvise dont fork boot image methods";
zygote_mapping_methods_ = MemMap();
return;
}
// We should use the F_SEAL_FUTURE_WRITE flag, but this has unexpected
// behavior on private mappings after fork (the mapping becomes shared between
// parent and children), see b/143833776.
// We will seal the write once we are done writing to the shared mapping.
if (fcntl(mem_fd, F_ADD_SEALS, F_SEAL_SHRINK | F_SEAL_GROW) == -1) {
PLOG(WARNING) << "Failed to seal boot image methods file descriptor";
zygote_mapping_methods_ = MemMap();
return;
}
fd_methods_ = unique_fd(mem_fd.release());
fd_methods_size_ = total_capacity;
}
}
}
void Jit::RegisterDexFiles(const std::vector<std::unique_ptr<const DexFile>>& dex_files,
jobject class_loader) {
if (dex_files.empty()) {
return;
}
Runtime* runtime = Runtime::Current();
// If the runtime is debuggable, don't bother precompiling methods.
// If system server is being profiled, don't precompile as we are going to use
// the JIT to count hotness. Note that --count-hotness-in-compiled-code is
// only forced when we also profile the boot classpath, see
// AndroidRuntime.cpp.
if (runtime->IsSystemServer() &&
UseJitCompilation() &&
options_->UseProfiledJitCompilation() &&
runtime->HasImageWithProfile() &&
!runtime->IsSystemServerProfiled() &&
!runtime->IsJavaDebuggable()) {
// Note: this precompilation is currently not running in production because:
// - UseProfiledJitCompilation() is not set by default.
// - System server dex files are registered *before* we set the runtime as
// system server (though we are in the system server process).
thread_pool_->AddTask(Thread::Current(), new JitProfileTask(dex_files, class_loader));
}
}
void Jit::AddCompileTask(Thread* self,
ArtMethod* method,
CompilationKind compilation_kind) {
thread_pool_->AddTask(self, method, compilation_kind);
}
bool Jit::CompileMethodFromProfile(Thread* self,
ClassLinker* class_linker,
uint32_t method_idx,
Handle<mirror::DexCache> dex_cache,
Handle<mirror::ClassLoader> class_loader,
bool add_to_queue,
bool compile_after_boot) {
ArtMethod* method = class_linker->ResolveMethodWithoutInvokeType(
method_idx, dex_cache, class_loader);
if (method == nullptr) {
self->ClearException();
return false;
}
if (!method->IsCompilable() || !method->IsInvokable()) {
return false;
}
if (method->IsPreCompiled()) {
// Already seen by another profile.
return false;
}
CompilationKind compilation_kind = CompilationKind::kOptimized;
const void* entry_point = method->GetEntryPointFromQuickCompiledCode();
if (class_linker->IsQuickToInterpreterBridge(entry_point) ||
class_linker->IsQuickGenericJniStub(entry_point) ||
class_linker->IsNterpEntryPoint(entry_point) ||
// We explicitly check for the resolution stub, and not the resolution trampoline.
// The trampoline is for methods backed by a .oat file that has a compiled version of
// the method.
(entry_point == GetQuickResolutionStub())) {
VLOG(jit) << "JIT Zygote processing method " << ArtMethod::PrettyMethod(method)
<< " from profile";
method->SetPreCompiled();
if (!add_to_queue) {
CompileMethodInternal(method, self, compilation_kind, /* prejit= */ true);
} else {
Task* task = new JitCompileTask(
method, JitCompileTask::TaskKind::kPreCompile, compilation_kind);
if (compile_after_boot) {
AddPostBootTask(self, task);
} else {
thread_pool_->AddTask(self, task);
}
return true;
}
}
return false;
}
uint32_t Jit::CompileMethodsFromBootProfile(
Thread* self,
const std::vector<const DexFile*>& dex_files,
const std::string& profile_file,
Handle<mirror::ClassLoader> class_loader,
bool add_to_queue) {
unix_file::FdFile profile(profile_file, O_RDONLY, true);
if (profile.Fd() == -1) {
PLOG(WARNING) << "No boot profile: " << profile_file;
return 0u;
}
ProfileBootInfo profile_info;
if (!profile_info.Load(profile.Fd(), dex_files)) {
LOG(ERROR) << "Could not load profile file: " << profile_file;
return 0u;
}
ScopedObjectAccess soa(self);
VariableSizedHandleScope handles(self);
std::vector<Handle<mirror::DexCache>> dex_caches;
ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
for (const DexFile* dex_file : profile_info.GetDexFiles()) {
dex_caches.push_back(handles.NewHandle(class_linker->FindDexCache(self, *dex_file)));
}
uint32_t added_to_queue = 0;
for (const std::pair<uint32_t, uint32_t>& pair : profile_info.GetMethods()) {
if (CompileMethodFromProfile(self,
class_linker,
pair.second,
dex_caches[pair.first],
class_loader,
add_to_queue,
/*compile_after_boot=*/false)) {
++added_to_queue;
}
}
return added_to_queue;
}
uint32_t Jit::CompileMethodsFromProfile(
Thread* self,
const std::vector<const DexFile*>& dex_files,
const std::string& profile_file,
Handle<mirror::ClassLoader> class_loader,
bool add_to_queue) {
if (profile_file.empty()) {
LOG(WARNING) << "Expected a profile file in JIT zygote mode";
return 0u;
}
// We don't generate boot profiles on device, therefore we don't
// need to lock the file.
unix_file::FdFile profile(profile_file, O_RDONLY, true);
if (profile.Fd() == -1) {
PLOG(WARNING) << "No profile: " << profile_file;
return 0u;
}
ProfileCompilationInfo profile_info(/* for_boot_image= */ class_loader.IsNull());
if (!profile_info.Load(profile.Fd())) {
LOG(ERROR) << "Could not load profile file";
return 0u;
}
ScopedObjectAccess soa(self);
StackHandleScope<1> hs(self);
MutableHandle<mirror::DexCache> dex_cache = hs.NewHandle<mirror::DexCache>(nullptr);
ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
uint32_t added_to_queue = 0u;
for (const DexFile* dex_file : dex_files) {
std::set<dex::TypeIndex> class_types;
std::set<uint16_t> all_methods;
if (!profile_info.GetClassesAndMethods(*dex_file,
&class_types,
&all_methods,
&all_methods,
&all_methods)) {
// This means the profile file did not reference the dex file, which is the case
// if there's no classes and methods of that dex file in the profile.
continue;
}
dex_cache.Assign(class_linker->FindDexCache(self, *dex_file));
CHECK(dex_cache != nullptr) << "Could not find dex cache for " << dex_file->GetLocation();
for (uint16_t method_idx : all_methods) {
if (CompileMethodFromProfile(self,
class_linker,
method_idx,
dex_cache,
class_loader,
add_to_queue,
/*compile_after_boot=*/true)) {
++added_to_queue;
}
}
}
// Add a task to run when all compilation is done.
AddPostBootTask(self, new JitDoneCompilingProfileTask(dex_files));
return added_to_queue;
}
bool Jit::IgnoreSamplesForMethod(ArtMethod* method) REQUIRES_SHARED(Locks::mutator_lock_) {
if (method->IsClassInitializer() || !method->IsCompilable()) {
// We do not want to compile such methods.
return true;
}
if (method->IsNative()) {
ObjPtr<mirror::Class> klass = method->GetDeclaringClass();
if (klass == GetClassRoot<mirror::MethodHandle>() ||
klass == GetClassRoot<mirror::VarHandle>()) {
// MethodHandle and VarHandle invocation methods are required to throw an
// UnsupportedOperationException if invoked reflectively. We achieve this by having native
// implementations that raise the exception. We need to disable JIT compilation of these JNI
// methods as it can lead to transitioning between JIT compiled JNI stubs and generic JNI
// stubs. Since these stubs have different stack representations we can then crash in stack
// walking (b/78151261).
return true;
}
}
return false;
}
void Jit::EnqueueOptimizedCompilation(ArtMethod* method, Thread* self) {
// Note the hotness counter will be reset by the compiled code.
if (thread_pool_ == nullptr) {
return;
}
// We arrive here after a baseline compiled code has reached its baseline
// hotness threshold. If we're not only using the baseline compiler, enqueue a compilation
// task that will compile optimize the method.
if (!options_->UseBaselineCompiler()) {
AddCompileTask(self, method, CompilationKind::kOptimized);
}
}
class ScopedSetRuntimeThread {
public:
explicit ScopedSetRuntimeThread(Thread* self)
: self_(self), was_runtime_thread_(self_->IsRuntimeThread()) {
self_->SetIsRuntimeThread(true);
}
~ScopedSetRuntimeThread() {
self_->SetIsRuntimeThread(was_runtime_thread_);
}
private:
Thread* self_;
bool was_runtime_thread_;
};
void Jit::MethodEntered(Thread* self, ArtMethod* method) {
Runtime* runtime = Runtime::Current();
if (UNLIKELY(runtime->UseJitCompilation() && JitAtFirstUse())) {
ArtMethod* np_method = method->GetInterfaceMethodIfProxy(kRuntimePointerSize);
if (np_method->IsCompilable()) {
CompileMethod(method, self, CompilationKind::kOptimized, /* prejit= */ false);
}
return;
}
AddSamples(self, method);
}
void Jit::WaitForCompilationToFinish(Thread* self) {
if (thread_pool_ != nullptr) {
thread_pool_->Wait(self, false, false);
}
}
void Jit::Stop() {
Thread* self = Thread::Current();
// TODO(ngeoffray): change API to not require calling WaitForCompilationToFinish twice.
WaitForCompilationToFinish(self);
GetThreadPool()->StopWorkers(self);
WaitForCompilationToFinish(self);
}
void Jit::Start() {
GetThreadPool()->StartWorkers(Thread::Current());
}
ScopedJitSuspend::ScopedJitSuspend() {
jit::Jit* jit = Runtime::Current()->GetJit();
was_on_ = (jit != nullptr) && (jit->GetThreadPool() != nullptr);
if (was_on_) {
jit->Stop();
}
}
ScopedJitSuspend::~ScopedJitSuspend() {
if (was_on_) {
DCHECK(Runtime::Current()->GetJit() != nullptr);
DCHECK(Runtime::Current()->GetJit()->GetThreadPool() != nullptr);
Runtime::Current()->GetJit()->Start();
}
}
static void* RunPollingThread(void* arg) {
Jit* jit = reinterpret_cast<Jit*>(arg);
do {
sleep(10);
} while (!jit->GetCodeCache()->GetZygoteMap()->IsCompilationNotified());
// We will suspend other threads: we can only do that if we're attached to the
// runtime.
Runtime* runtime = Runtime::Current();
bool thread_attached = runtime->AttachCurrentThread(
"BootImagePollingThread",
/* as_daemon= */ true,
/* thread_group= */ nullptr,
/* create_peer= */ false);
CHECK(thread_attached);
{
// Prevent other threads from running while we are remapping the boot image
// ArtMethod's. Native threads might still be running, but they cannot
// change the contents of ArtMethod's.
ScopedSuspendAll ssa(__FUNCTION__);
runtime->GetJit()->MapBootImageMethods();
}
Runtime::Current()->DetachCurrentThread();
return nullptr;
}
void Jit::PostForkChildAction(bool is_system_server, bool is_zygote) {
// Clear the potential boot tasks inherited from the zygote.
{
MutexLock mu(Thread::Current(), boot_completed_lock_);
tasks_after_boot_.clear();
}
Runtime* const runtime = Runtime::Current();
// Check if we'll need to remap the boot image methods.
if (!is_zygote && fd_methods_ != -1) {
// Create a thread that will poll the status of zygote compilation, and map
// the private mapping of boot image methods.
// For child zygote, we instead query IsCompilationNotified() post zygote fork.
zygote_mapping_methods_.ResetInForkedProcess();
pthread_t polling_thread;
pthread_attr_t attr;
CHECK_PTHREAD_CALL(pthread_attr_init, (&attr), "new thread");
CHECK_PTHREAD_CALL(pthread_attr_setdetachstate, (&attr, PTHREAD_CREATE_DETACHED),
"PTHREAD_CREATE_DETACHED");
CHECK_PTHREAD_CALL(
pthread_create,
(&polling_thread, &attr, RunPollingThread, reinterpret_cast<void*>(this)),
"Methods maps thread");
}
if (is_zygote || runtime->IsSafeMode()) {
// Delete the thread pool, we are not going to JIT.
thread_pool_.reset(nullptr);
return;
}
// At this point, the compiler options have been adjusted to the particular configuration
// of the forked child. Parse them again.
jit_compiler_->ParseCompilerOptions();
// Adjust the status of code cache collection: the status from zygote was to not collect.
// JitAtFirstUse compiles the methods synchronously on mutator threads. While this should work
// in theory it is causing deadlocks in some jvmti tests related to Jit GC. Hence, disabling
// Jit GC for now (b/147208992).
code_cache_->SetGarbageCollectCode(
!jit_compiler_->GenerateDebugInfo() &&
!JitAtFirstUse());
if (is_system_server && runtime->HasImageWithProfile()) {
// Disable garbage collection: we don't want it to delete methods we're compiling
// through boot and system server profiles.
// TODO(ngeoffray): Fix this so we still collect deoptimized and unused code.
code_cache_->SetGarbageCollectCode(false);
}
// We do this here instead of PostZygoteFork, as NativeDebugInfoPostFork only
// applies to a child.
NativeDebugInfoPostFork();
}
void Jit::PreZygoteFork() {
if (thread_pool_ == nullptr) {
return;
}
thread_pool_->DeleteThreads();
NativeDebugInfoPreFork();
}
void Jit::PostZygoteFork() {
Runtime* runtime = Runtime::Current();
if (thread_pool_ == nullptr) {
// If this is a child zygote, check if we need to remap the boot image
// methods.
if (runtime->IsZygote() &&
fd_methods_ != -1 &&
code_cache_->GetZygoteMap()->IsCompilationNotified()) {
ScopedSuspendAll ssa(__FUNCTION__);
MapBootImageMethods();
}
return;
}
if (runtime->IsZygote() && code_cache_->GetZygoteMap()->IsCompilationDoneButNotNotified()) {
// Copy the boot image methods data to the mappings we created to share
// with the children. We do this here as we are the only thread running and
// we don't risk other threads concurrently updating the ArtMethod's.
CHECK_EQ(GetTaskCount(), 1);
NotifyZygoteCompilationDone();
CHECK(code_cache_->GetZygoteMap()->IsCompilationNotified());
}
thread_pool_->CreateThreads();
thread_pool_->SetPthreadPriority(
runtime->IsZygote()
? options_->GetZygoteThreadPoolPthreadPriority()
: options_->GetThreadPoolPthreadPriority());
}
void Jit::AddPostBootTask(Thread* self, Task* task) {
MutexLock mu(self, boot_completed_lock_);
if (boot_completed_) {
thread_pool_->AddTask(self, task);
} else {
tasks_after_boot_.push_back(task);
}
}
void Jit::BootCompleted() {
Thread* self = Thread::Current();
std::deque<Task*> tasks;
{
MutexLock mu(self, boot_completed_lock_);
tasks = std::move(tasks_after_boot_);
boot_completed_ = true;
}
for (Task* task : tasks) {
thread_pool_->AddTask(self, task);
}
}
bool Jit::CanEncodeMethod(ArtMethod* method, bool is_for_shared_region) const {
return !is_for_shared_region ||
Runtime::Current()->GetHeap()->ObjectIsInBootImageSpace(method->GetDeclaringClass());
}
bool Jit::CanEncodeClass(ObjPtr<mirror::Class> cls, bool is_for_shared_region) const {
return !is_for_shared_region || Runtime::Current()->GetHeap()->ObjectIsInBootImageSpace(cls);
}
bool Jit::CanEncodeString(ObjPtr<mirror::String> string, bool is_for_shared_region) const {
return !is_for_shared_region || Runtime::Current()->GetHeap()->ObjectIsInBootImageSpace(string);
}
bool Jit::CanAssumeInitialized(ObjPtr<mirror::Class> cls, bool is_for_shared_region) const {
if (!is_for_shared_region) {
return cls->IsInitialized();
} else {
// Look up the class status in the oat file.
const DexFile& dex_file = *cls->GetDexCache()->GetDexFile();
const OatDexFile* oat_dex_file = dex_file.GetOatDexFile();
// In case we run without an image there won't be a backing oat file.
if (oat_dex_file == nullptr || oat_dex_file->GetOatFile() == nullptr) {
return false;
}
uint16_t class_def_index = cls->GetDexClassDefIndex();
return oat_dex_file->GetOatClass(class_def_index).GetStatus() >= ClassStatus::kInitialized;
}
}
void Jit::MaybeEnqueueCompilation(ArtMethod* method, Thread* self) {
if (thread_pool_ == nullptr) {
return;
}
if (JitAtFirstUse()) {
// Tests might request JIT on first use (compiled synchronously in the interpreter).
return;
}
if (!UseJitCompilation()) {
return;
}
if (IgnoreSamplesForMethod(method)) {
return;
}
if (GetCodeCache()->ContainsPc(method->GetEntryPointFromQuickCompiledCode())) {
if (!method->IsNative() && !code_cache_->IsOsrCompiled(method)) {
// If we already have compiled code for it, nterp may be stuck in a loop.
// Compile OSR.
AddCompileTask(self, method, CompilationKind::kOsr);
}
return;
}
// Check if we have precompiled this method.
if (UNLIKELY(method->IsPreCompiled())) {
if (!method->StillNeedsClinitCheck()) {
const void* entry_point = code_cache_->GetSavedEntryPointOfPreCompiledMethod(method);
if (entry_point != nullptr) {
Runtime::Current()->GetInstrumentation()->UpdateMethodsCode(method, entry_point);
}
}
return;
}
static constexpr size_t kIndividualSharedMethodHotnessThreshold = 0x3f;
if (method->IsMemorySharedMethod()) {
MutexLock mu(self, lock_);
auto it = shared_method_counters_.find(method);
if (it == shared_method_counters_.end()) {
shared_method_counters_[method] = kIndividualSharedMethodHotnessThreshold;
return;
} else if (it->second != 0) {
DCHECK_LE(it->second, kIndividualSharedMethodHotnessThreshold);
shared_method_counters_[method] = it->second - 1;
return;
} else {
shared_method_counters_[method] = kIndividualSharedMethodHotnessThreshold;
}
}
if (!method->IsNative() && GetCodeCache()->CanAllocateProfilingInfo()) {
AddCompileTask(self, method, CompilationKind::kBaseline);
} else {
AddCompileTask(self, method, CompilationKind::kOptimized);
}
}
bool Jit::CompileMethod(ArtMethod* method,
Thread* self,
CompilationKind compilation_kind,
bool prejit) {
// Fake being in a runtime thread so that class-load behavior will be the same as normal jit.
ScopedSetRuntimeThread ssrt(self);
// TODO(ngeoffray): For JIT at first use, use kPreCompile. Currently we don't due to
// conflicts with jitzygote optimizations.
return CompileMethodInternal(method, self, compilation_kind, prejit);
}
size_t JitThreadPool::GetTaskCount(Thread* self) {
MutexLock mu(self, task_queue_lock_);
return generic_queue_.size() +
baseline_queue_.size() +
optimized_queue_.size() +
osr_queue_.size();
}
void JitThreadPool::RemoveAllTasks(Thread* self) {
// The ThreadPool is responsible for calling Finalize (which usually deletes
// the task memory) on all the tasks.
Task* task = nullptr;
do {
{
MutexLock mu(self, task_queue_lock_);
if (generic_queue_.empty()) {
break;
}
task = generic_queue_.front();
generic_queue_.pop_front();
}
task->Finalize();
} while (true);
MutexLock mu(self, task_queue_lock_);
baseline_queue_.clear();
optimized_queue_.clear();
osr_queue_.clear();
}
JitThreadPool::~JitThreadPool() {
DeleteThreads();
RemoveAllTasks(Thread::Current());
}
void JitThreadPool::AddTask(Thread* self, Task* task) {
MutexLock mu(self, task_queue_lock_);
// We don't want to enqueue any new tasks when thread pool has stopped. This simplifies
// the implementation of redefinition feature in jvmti.
if (!started_) {
task->Finalize();
return;
}
generic_queue_.push_back(task);
// If we have any waiters, signal one.
if (waiting_count_ != 0) {
task_queue_condition_.Signal(self);
}
}
void JitThreadPool::AddTask(Thread* self, ArtMethod* method, CompilationKind kind) {
MutexLock mu(self, task_queue_lock_);
// We don't want to enqueue any new tasks when thread pool has stopped. This simplifies
// the implementation of redefinition feature in jvmti.
if (!started_) {
return;
}
switch (kind) {
case CompilationKind::kOsr:
if (ContainsElement(osr_enqueued_methods_, method)) {
return;
}
osr_enqueued_methods_.insert(method);
osr_queue_.push_back(method);
break;
case CompilationKind::kBaseline:
if (ContainsElement(baseline_enqueued_methods_, method)) {
return;
}
baseline_enqueued_methods_.insert(method);
baseline_queue_.push_back(method);
break;
case CompilationKind::kOptimized:
if (ContainsElement(optimized_enqueued_methods_, method)) {
return;
}
optimized_enqueued_methods_.insert(method);
optimized_queue_.push_back(method);
break;
}
// If we have any waiters, signal one.
if (waiting_count_ != 0) {
task_queue_condition_.Signal(self);
}
}
Task* JitThreadPool::TryGetTaskLocked() {
if (!started_) {
return nullptr;
}
// Fetch generic tasks first.
if (!generic_queue_.empty()) {
Task* task = generic_queue_.front();
generic_queue_.pop_front();
return task;
}
// OSR requests second, then baseline and finally optimized.
Task* task = FetchFrom(osr_queue_, CompilationKind::kOsr);
if (task == nullptr) {
task = FetchFrom(baseline_queue_, CompilationKind::kBaseline);
if (task == nullptr) {
task = FetchFrom(optimized_queue_, CompilationKind::kOptimized);
}
}
return task;
}
Task* JitThreadPool::FetchFrom(std::deque<ArtMethod*>& methods, CompilationKind kind) {
if (!methods.empty()) {
ArtMethod* method = methods.front();
methods.pop_front();
JitCompileTask* task = new JitCompileTask(method, JitCompileTask::TaskKind::kCompile, kind);
current_compilations_.insert(task);
return task;
}
return nullptr;
}
void JitThreadPool::Remove(JitCompileTask* task) {
MutexLock mu(Thread::Current(), task_queue_lock_);
current_compilations_.erase(task);
switch (task->GetCompilationKind()) {
case CompilationKind::kOsr: {
osr_enqueued_methods_.erase(task->GetArtMethod());
break;
}
case CompilationKind::kBaseline: {
baseline_enqueued_methods_.erase(task->GetArtMethod());
break;
}
case CompilationKind::kOptimized: {
optimized_enqueued_methods_.erase(task->GetArtMethod());
break;
}
}
}
void Jit::VisitRoots(RootVisitor* visitor) {
if (thread_pool_ != nullptr) {
thread_pool_->VisitRoots(visitor);
}
}
void JitThreadPool::VisitRoots(RootVisitor* visitor) {
if (Runtime::Current()->GetHeap()->IsPerformingUffdCompaction()) {
// In case of userfaultfd compaction, ArtMethods are updated concurrently
// via linear-alloc.
return;
}
// Fetch all ArtMethod first, to avoid holding `task_queue_lock_` for too
// long.
std::vector<ArtMethod*> methods;
{
MutexLock mu(Thread::Current(), task_queue_lock_);
// We don't look at `generic_queue_` because it contains:
// - Generic tasks like `ZygoteVerificationTask` which don't hold any root.
// - `JitCompileTask` for precompiled methods, which we know are live, being
// part of the boot classpath or system server classpath.
methods.insert(methods.end(), osr_queue_.begin(), osr_queue_.end());
methods.insert(methods.end(), baseline_queue_.begin(), baseline_queue_.end());
methods.insert(methods.end(), optimized_queue_.begin(), optimized_queue_.end());
for (JitCompileTask* task : current_compilations_) {
methods.push_back(task->GetArtMethod());
}
}
UnbufferedRootVisitor root_visitor(visitor, RootInfo(kRootStickyClass));
for (ArtMethod* method : methods) {
method->VisitRoots(root_visitor, kRuntimePointerSize);
}
}
} // namespace jit
} // namespace art