/* * 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_code_cache.h" #include #include #include "arch/context.h" #include "art_method-inl.h" #include "base/enums.h" #include "base/histogram-inl.h" #include "base/logging.h" // For VLOG. #include "base/membarrier.h" #include "base/memfd.h" #include "base/mem_map.h" #include "base/quasi_atomic.h" #include "base/stl_util.h" #include "base/systrace.h" #include "base/time_utils.h" #include "base/utils.h" #include "cha.h" #include "debugger_interface.h" #include "dex/dex_file_loader.h" #include "dex/method_reference.h" #include "entrypoints/entrypoint_utils-inl.h" #include "entrypoints/runtime_asm_entrypoints.h" #include "gc/accounting/bitmap-inl.h" #include "gc/allocator/dlmalloc.h" #include "gc/scoped_gc_critical_section.h" #include "handle.h" #include "handle_scope-inl.h" #include "instrumentation.h" #include "intern_table.h" #include "jit/jit.h" #include "jit/profiling_info.h" #include "jit/jit_scoped_code_cache_write.h" #include "linear_alloc.h" #include "oat_file-inl.h" #include "oat_quick_method_header.h" #include "object_callbacks.h" #include "profile/profile_compilation_info.h" #include "scoped_thread_state_change-inl.h" #include "stack.h" #include "thread-current-inl.h" #include "thread_list.h" namespace art { namespace jit { static constexpr size_t kCodeSizeLogThreshold = 50 * KB; static constexpr size_t kStackMapSizeLogThreshold = 50 * KB; class JitCodeCache::JniStubKey { public: explicit JniStubKey(ArtMethod* method) REQUIRES_SHARED(Locks::mutator_lock_) : shorty_(method->GetShorty()), is_static_(method->IsStatic()), is_fast_native_(method->IsFastNative()), is_critical_native_(method->IsCriticalNative()), is_synchronized_(method->IsSynchronized()) { DCHECK(!(is_fast_native_ && is_critical_native_)); } bool operator<(const JniStubKey& rhs) const { if (is_static_ != rhs.is_static_) { return rhs.is_static_; } if (is_synchronized_ != rhs.is_synchronized_) { return rhs.is_synchronized_; } if (is_fast_native_ != rhs.is_fast_native_) { return rhs.is_fast_native_; } if (is_critical_native_ != rhs.is_critical_native_) { return rhs.is_critical_native_; } return strcmp(shorty_, rhs.shorty_) < 0; } // Update the shorty to point to another method's shorty. Call this function when removing // the method that references the old shorty from JniCodeData and not removing the entire // JniCodeData; the old shorty may become a dangling pointer when that method is unloaded. void UpdateShorty(ArtMethod* method) const REQUIRES_SHARED(Locks::mutator_lock_) { const char* shorty = method->GetShorty(); DCHECK_STREQ(shorty_, shorty); shorty_ = shorty; } private: // The shorty points to a DexFile data and may need to change // to point to the same shorty in a different DexFile. mutable const char* shorty_; const bool is_static_; const bool is_fast_native_; const bool is_critical_native_; const bool is_synchronized_; }; class JitCodeCache::JniStubData { public: JniStubData() : code_(nullptr), methods_() {} void SetCode(const void* code) { DCHECK(code != nullptr); code_ = code; } void UpdateEntryPoints(const void* entrypoint) REQUIRES_SHARED(Locks::mutator_lock_) { DCHECK(IsCompiled()); DCHECK(entrypoint == OatQuickMethodHeader::FromCodePointer(GetCode())->GetEntryPoint()); instrumentation::Instrumentation* instrum = Runtime::Current()->GetInstrumentation(); for (ArtMethod* m : GetMethods()) { // Because `m` might be in the process of being deleted: // - Call the dedicated method instead of the more generic UpdateMethodsCode // - Check the class status without a full read barrier; use ReadBarrier::IsMarked(). bool can_set_entrypoint = true; if (NeedsClinitCheckBeforeCall(m)) { // To avoid resurrecting an unreachable object, we must not use a full read // barrier but we do not want to miss updating an entrypoint under common // circumstances, i.e. during a GC the class becomes visibly initialized, // the method becomes hot, we compile the thunk and want to update the // entrypoint while the method's declaring class field still points to the // from-space class object with the old status. Therefore we read the // declaring class without a read barrier and check if it's already marked. // If yes, we check the status of the to-space class object as intended. // Otherwise, there is no to-space object and the from-space class object // contains the most recent value of the status field; even if this races // with another thread doing a read barrier and updating the status, that's // no different from a race with a thread that just updates the status. // Such race can happen only for the zygote method pre-compilation, as we // otherwise compile only thunks for methods of visibly initialized classes. ObjPtr klass = m->GetDeclaringClass(); ObjPtr marked = ReadBarrier::IsMarked(klass.Ptr()); ObjPtr checked_klass = (marked != nullptr) ? marked : klass; can_set_entrypoint = checked_klass->IsVisiblyInitialized(); } if (can_set_entrypoint) { instrum->UpdateNativeMethodsCodeToJitCode(m, entrypoint); } } } const void* GetCode() const { return code_; } bool IsCompiled() const { return GetCode() != nullptr; } void AddMethod(ArtMethod* method) { if (!ContainsElement(methods_, method)) { methods_.push_back(method); } } const std::vector& GetMethods() const { return methods_; } void RemoveMethodsIn(const LinearAlloc& alloc) REQUIRES_SHARED(Locks::mutator_lock_) { auto kept_end = std::partition( methods_.begin(), methods_.end(), [&alloc](ArtMethod* method) { return !alloc.ContainsUnsafe(method); }); for (auto it = kept_end; it != methods_.end(); it++) { VLOG(jit) << "JIT removed (JNI) " << (*it)->PrettyMethod() << ": " << code_; } methods_.erase(kept_end, methods_.end()); } bool RemoveMethod(ArtMethod* method) REQUIRES_SHARED(Locks::mutator_lock_) { auto it = std::find(methods_.begin(), methods_.end(), method); if (it != methods_.end()) { VLOG(jit) << "JIT removed (JNI) " << (*it)->PrettyMethod() << ": " << code_; methods_.erase(it); return true; } else { return false; } } void MoveObsoleteMethod(ArtMethod* old_method, ArtMethod* new_method) { std::replace(methods_.begin(), methods_.end(), old_method, new_method); } private: const void* code_; std::vector methods_; }; JitCodeCache* JitCodeCache::Create(bool used_only_for_profile_data, bool rwx_memory_allowed, bool is_zygote, std::string* error_msg) { // Register for membarrier expedited sync core if JIT will be generating code. if (!used_only_for_profile_data) { if (art::membarrier(art::MembarrierCommand::kRegisterPrivateExpeditedSyncCore) != 0) { // MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE ensures that CPU instruction pipelines are // flushed and it's used when adding code to the JIT. The memory used by the new code may // have just been released and, in theory, the old code could still be in a pipeline. VLOG(jit) << "Kernel does not support membarrier sync-core"; } } size_t initial_capacity = Runtime::Current()->GetJITOptions()->GetCodeCacheInitialCapacity(); // Check whether the provided max capacity in options is below 1GB. size_t max_capacity = Runtime::Current()->GetJITOptions()->GetCodeCacheMaxCapacity(); // We need to have 32 bit offsets from method headers in code cache which point to things // in the data cache. If the maps are more than 4G apart, having multiple maps wouldn't work. // Ensure we're below 1 GB to be safe. if (max_capacity > 1 * GB) { std::ostringstream oss; oss << "Maxium code cache capacity is limited to 1 GB, " << PrettySize(max_capacity) << " is too big"; *error_msg = oss.str(); return nullptr; } MutexLock mu(Thread::Current(), *Locks::jit_lock_); JitMemoryRegion region; if (!region.Initialize(initial_capacity, max_capacity, rwx_memory_allowed, is_zygote, error_msg)) { return nullptr; } std::unique_ptr jit_code_cache(new JitCodeCache()); if (is_zygote) { // Zygote should never collect code to share the memory with the children. jit_code_cache->garbage_collect_code_ = false; jit_code_cache->shared_region_ = std::move(region); } else { jit_code_cache->private_region_ = std::move(region); } VLOG(jit) << "Created jit code cache: initial capacity=" << PrettySize(initial_capacity) << ", maximum capacity=" << PrettySize(max_capacity); return jit_code_cache.release(); } JitCodeCache::JitCodeCache() : is_weak_access_enabled_(true), inline_cache_cond_("Jit inline cache condition variable", *Locks::jit_lock_), zygote_map_(&shared_region_), lock_cond_("Jit code cache condition variable", *Locks::jit_lock_), collection_in_progress_(false), last_collection_increased_code_cache_(false), garbage_collect_code_(true), number_of_baseline_compilations_(0), number_of_optimized_compilations_(0), number_of_osr_compilations_(0), number_of_collections_(0), histogram_stack_map_memory_use_("Memory used for stack maps", 16), histogram_code_memory_use_("Memory used for compiled code", 16), histogram_profiling_info_memory_use_("Memory used for profiling info", 16) { } JitCodeCache::~JitCodeCache() {} bool JitCodeCache::PrivateRegionContainsPc(const void* ptr) const { return private_region_.IsInExecSpace(ptr); } bool JitCodeCache::ContainsPc(const void* ptr) const { return PrivateRegionContainsPc(ptr) || shared_region_.IsInExecSpace(ptr); } bool JitCodeCache::ContainsMethod(ArtMethod* method) { MutexLock mu(Thread::Current(), *Locks::jit_lock_); if (UNLIKELY(method->IsNative())) { auto it = jni_stubs_map_.find(JniStubKey(method)); if (it != jni_stubs_map_.end() && it->second.IsCompiled() && ContainsElement(it->second.GetMethods(), method)) { return true; } } else { for (const auto& it : method_code_map_) { if (it.second == method) { return true; } } if (zygote_map_.ContainsMethod(method)) { return true; } } return false; } const void* JitCodeCache::GetJniStubCode(ArtMethod* method) { DCHECK(method->IsNative()); MutexLock mu(Thread::Current(), *Locks::jit_lock_); auto it = jni_stubs_map_.find(JniStubKey(method)); if (it != jni_stubs_map_.end()) { JniStubData& data = it->second; if (data.IsCompiled() && ContainsElement(data.GetMethods(), method)) { return data.GetCode(); } } return nullptr; } const void* JitCodeCache::GetSavedEntryPointOfPreCompiledMethod(ArtMethod* method) { if (method->IsPreCompiled()) { const void* code_ptr = nullptr; if (method->GetDeclaringClass()->GetClassLoader() == nullptr) { code_ptr = zygote_map_.GetCodeFor(method); } else { MutexLock mu(Thread::Current(), *Locks::jit_lock_); auto it = saved_compiled_methods_map_.find(method); if (it != saved_compiled_methods_map_.end()) { code_ptr = it->second; } } if (code_ptr != nullptr) { OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr); return method_header->GetEntryPoint(); } } return nullptr; } bool JitCodeCache::WaitForPotentialCollectionToComplete(Thread* self) { bool in_collection = false; while (collection_in_progress_) { in_collection = true; lock_cond_.Wait(self); } return in_collection; } static uintptr_t FromCodeToAllocation(const void* code) { size_t alignment = GetInstructionSetAlignment(kRuntimeISA); return reinterpret_cast(code) - RoundUp(sizeof(OatQuickMethodHeader), alignment); } static const void* FromAllocationToCode(const uint8_t* alloc) { size_t alignment = GetInstructionSetAlignment(kRuntimeISA); return reinterpret_cast(alloc + RoundUp(sizeof(OatQuickMethodHeader), alignment)); } static uint32_t GetNumberOfRoots(const uint8_t* stack_map) { // The length of the table is stored just before the stack map (and therefore at the end of // the table itself), in order to be able to fetch it from a `stack_map` pointer. return reinterpret_cast(stack_map)[-1]; } static void DCheckRootsAreValid(const std::vector>& roots, bool is_shared_region) REQUIRES(!Locks::intern_table_lock_) REQUIRES_SHARED(Locks::mutator_lock_) { if (!kIsDebugBuild) { return; } // Put all roots in `roots_data`. for (Handle object : roots) { // Ensure the string is strongly interned. b/32995596 if (object->IsString()) { ObjPtr str = object->AsString(); ClassLinker* class_linker = Runtime::Current()->GetClassLinker(); CHECK(class_linker->GetInternTable()->LookupStrong(Thread::Current(), str) != nullptr); } // Ensure that we don't put movable objects in the shared region. if (is_shared_region) { CHECK(!Runtime::Current()->GetHeap()->IsMovableObject(object.Get())); } } } static const uint8_t* GetRootTable(const void* code_ptr, uint32_t* number_of_roots = nullptr) { OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr); uint8_t* data = method_header->GetOptimizedCodeInfoPtr(); uint32_t roots = GetNumberOfRoots(data); if (number_of_roots != nullptr) { *number_of_roots = roots; } return data - ComputeRootTableSize(roots); } void JitCodeCache::SweepRootTables(IsMarkedVisitor* visitor) { MutexLock mu(Thread::Current(), *Locks::jit_lock_); for (const auto& entry : method_code_map_) { uint32_t number_of_roots = 0; const uint8_t* root_table = GetRootTable(entry.first, &number_of_roots); uint8_t* roots_data = private_region_.IsInDataSpace(root_table) ? private_region_.GetWritableDataAddress(root_table) : shared_region_.GetWritableDataAddress(root_table); GcRoot* roots = reinterpret_cast*>(roots_data); for (uint32_t i = 0; i < number_of_roots; ++i) { // This does not need a read barrier because this is called by GC. mirror::Object* object = roots[i].Read(); if (object == nullptr || object == Runtime::GetWeakClassSentinel()) { // entry got deleted in a previous sweep. } else if (object->IsString()) { mirror::Object* new_object = visitor->IsMarked(object); // We know the string is marked because it's a strongly-interned string that // is always alive. The IsMarked implementation of the CMS collector returns // null for newly allocated objects, but we know those haven't moved. Therefore, // only update the entry if we get a different non-null string. // TODO: Do not use IsMarked for j.l.Class, and adjust once we move this method // out of the weak access/creation pause. b/32167580 if (new_object != nullptr && new_object != object) { DCHECK(new_object->IsString()); roots[i] = GcRoot(new_object); } } else { Runtime::ProcessWeakClass( reinterpret_cast*>(&roots[i]), visitor, Runtime::GetWeakClassSentinel()); } } } // Walk over inline caches to clear entries containing unloaded classes. for (auto it : profiling_infos_) { ProfilingInfo* info = it.second; for (size_t i = 0; i < info->number_of_inline_caches_; ++i) { InlineCache* cache = &info->cache_[i]; for (size_t j = 0; j < InlineCache::kIndividualCacheSize; ++j) { Runtime::ProcessWeakClass(&cache->classes_[j], visitor, nullptr); } } } } void JitCodeCache::FreeCodeAndData(const void* code_ptr) { if (IsInZygoteExecSpace(code_ptr)) { // No need to free, this is shared memory. return; } uintptr_t allocation = FromCodeToAllocation(code_ptr); const uint8_t* data = nullptr; if (OatQuickMethodHeader::FromCodePointer(code_ptr)->IsOptimized()) { data = GetRootTable(code_ptr); } // else this is a JNI stub without any data. FreeLocked(&private_region_, reinterpret_cast(allocation), data); } void JitCodeCache::FreeAllMethodHeaders( const std::unordered_set& method_headers) { // We need to remove entries in method_headers from CHA dependencies // first since once we do FreeCode() below, the memory can be reused // so it's possible for the same method_header to start representing // different compile code. { MutexLock mu2(Thread::Current(), *Locks::cha_lock_); Runtime::Current()->GetClassLinker()->GetClassHierarchyAnalysis() ->RemoveDependentsWithMethodHeaders(method_headers); } ScopedCodeCacheWrite scc(private_region_); for (const OatQuickMethodHeader* method_header : method_headers) { FreeCodeAndData(method_header->GetCode()); } // We have potentially removed a lot of debug info. Do maintenance pass to save space. RepackNativeDebugInfoForJit(); // Check that the set of compiled methods exactly matches native debug information. // Does not check zygote methods since they can change concurrently. if (kIsDebugBuild && !Runtime::Current()->IsZygote()) { std::map compiled_methods; VisitAllMethods([&](const void* addr, ArtMethod* method) { if (!IsInZygoteExecSpace(addr)) { CHECK(addr != nullptr && method != nullptr); compiled_methods.emplace(addr, method); } }); std::set debug_info; ForEachNativeDebugSymbol([&](const void* addr, size_t, const char* name) { addr = AlignDown(addr, GetInstructionSetInstructionAlignment(kRuntimeISA)); // Thumb-bit. CHECK(debug_info.emplace(addr).second) << "Duplicate debug info: " << addr << " " << name; CHECK_EQ(compiled_methods.count(addr), 1u) << "Extra debug info: " << addr << " " << name; }); if (!debug_info.empty()) { // If debug-info generation is enabled. for (auto it : compiled_methods) { CHECK_EQ(debug_info.count(it.first), 1u) << "No debug info: " << it.second->PrettyMethod(); } CHECK_EQ(compiled_methods.size(), debug_info.size()); } } } void JitCodeCache::RemoveMethodsIn(Thread* self, const LinearAlloc& alloc) { ScopedTrace trace(__PRETTY_FUNCTION__); // We use a set to first collect all method_headers whose code need to be // removed. We need to free the underlying code after we remove CHA dependencies // for entries in this set. And it's more efficient to iterate through // the CHA dependency map just once with an unordered_set. std::unordered_set method_headers; { MutexLock mu(self, *Locks::jit_lock_); // We do not check if a code cache GC is in progress, as this method comes // with the classlinker_classes_lock_ held, and suspending ourselves could // lead to a deadlock. { for (auto it = jni_stubs_map_.begin(); it != jni_stubs_map_.end();) { it->second.RemoveMethodsIn(alloc); if (it->second.GetMethods().empty()) { method_headers.insert(OatQuickMethodHeader::FromCodePointer(it->second.GetCode())); it = jni_stubs_map_.erase(it); } else { it->first.UpdateShorty(it->second.GetMethods().front()); ++it; } } for (auto it = method_code_map_.begin(); it != method_code_map_.end();) { if (alloc.ContainsUnsafe(it->second)) { method_headers.insert(OatQuickMethodHeader::FromCodePointer(it->first)); VLOG(jit) << "JIT removed " << it->second->PrettyMethod() << ": " << it->first; it = method_code_map_.erase(it); } else { ++it; } } } for (auto it = osr_code_map_.begin(); it != osr_code_map_.end();) { if (alloc.ContainsUnsafe(it->first)) { // Note that the code has already been pushed to method_headers in the loop // above and is going to be removed in FreeCode() below. it = osr_code_map_.erase(it); } else { ++it; } } for (auto it = profiling_infos_.begin(); it != profiling_infos_.end();) { ProfilingInfo* info = it->second; if (alloc.ContainsUnsafe(info->GetMethod())) { private_region_.FreeWritableData(reinterpret_cast(info)); it = profiling_infos_.erase(it); } else { ++it; } } FreeAllMethodHeaders(method_headers); } } bool JitCodeCache::IsWeakAccessEnabled(Thread* self) const { return kUseReadBarrier ? self->GetWeakRefAccessEnabled() : is_weak_access_enabled_.load(std::memory_order_seq_cst); } void JitCodeCache::WaitUntilInlineCacheAccessible(Thread* self) { if (IsWeakAccessEnabled(self)) { return; } ScopedThreadSuspension sts(self, kWaitingWeakGcRootRead); MutexLock mu(self, *Locks::jit_lock_); while (!IsWeakAccessEnabled(self)) { inline_cache_cond_.Wait(self); } } void JitCodeCache::BroadcastForInlineCacheAccess() { Thread* self = Thread::Current(); MutexLock mu(self, *Locks::jit_lock_); inline_cache_cond_.Broadcast(self); } void JitCodeCache::AllowInlineCacheAccess() { DCHECK(!kUseReadBarrier); is_weak_access_enabled_.store(true, std::memory_order_seq_cst); BroadcastForInlineCacheAccess(); } void JitCodeCache::DisallowInlineCacheAccess() { DCHECK(!kUseReadBarrier); is_weak_access_enabled_.store(false, std::memory_order_seq_cst); } void JitCodeCache::CopyInlineCacheInto( const InlineCache& ic, /*out*/StackHandleScope* classes) { static_assert(arraysize(ic.classes_) == InlineCache::kIndividualCacheSize); DCHECK_EQ(classes->NumberOfReferences(), InlineCache::kIndividualCacheSize); DCHECK_EQ(classes->RemainingSlots(), InlineCache::kIndividualCacheSize); WaitUntilInlineCacheAccessible(Thread::Current()); // Note that we don't need to lock `lock_` here, the compiler calling // this method has already ensured the inline cache will not be deleted. for (const GcRoot& root : ic.classes_) { mirror::Class* object = root.Read(); if (object != nullptr) { DCHECK_NE(classes->RemainingSlots(), 0u); classes->NewHandle(object); } } } static void ClearMethodCounter(ArtMethod* method, bool was_warm) REQUIRES_SHARED(Locks::mutator_lock_) { if (was_warm) { method->SetPreviouslyWarm(); } // We reset the counter to 1 so that the profile knows that the method was executed at least once. // This is required for layout purposes. // We also need to make sure we'll pass the warmup threshold again, so we set to 0 if // the warmup threshold is 1. uint16_t jit_warmup_threshold = Runtime::Current()->GetJITOptions()->GetWarmupThreshold(); method->SetCounter(std::min(jit_warmup_threshold - 1, 1)); } void JitCodeCache::WaitForPotentialCollectionToCompleteRunnable(Thread* self) { while (collection_in_progress_) { Locks::jit_lock_->Unlock(self); { ScopedThreadSuspension sts(self, kSuspended); MutexLock mu(self, *Locks::jit_lock_); WaitForPotentialCollectionToComplete(self); } Locks::jit_lock_->Lock(self); } } bool JitCodeCache::Commit(Thread* self, JitMemoryRegion* region, ArtMethod* method, ArrayRef reserved_code, ArrayRef code, ArrayRef reserved_data, const std::vector>& roots, ArrayRef stack_map, const std::vector& debug_info, bool is_full_debug_info, CompilationKind compilation_kind, bool has_should_deoptimize_flag, const ArenaSet& cha_single_implementation_list) { DCHECK(!method->IsNative() || (compilation_kind != CompilationKind::kOsr)); if (!method->IsNative()) { // We need to do this before grabbing the lock_ because it needs to be able to see the string // InternTable. Native methods do not have roots. DCheckRootsAreValid(roots, IsSharedRegion(*region)); } const uint8_t* roots_data = reserved_data.data(); size_t root_table_size = ComputeRootTableSize(roots.size()); const uint8_t* stack_map_data = roots_data + root_table_size; MutexLock mu(self, *Locks::jit_lock_); // We need to make sure that there will be no jit-gcs going on and wait for any ongoing one to // finish. WaitForPotentialCollectionToCompleteRunnable(self); const uint8_t* code_ptr = region->CommitCode( reserved_code, code, stack_map_data, has_should_deoptimize_flag); if (code_ptr == nullptr) { return false; } OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr); // Commit roots and stack maps before updating the entry point. if (!region->CommitData(reserved_data, roots, stack_map)) { return false; } switch (compilation_kind) { case CompilationKind::kOsr: number_of_osr_compilations_++; break; case CompilationKind::kBaseline: number_of_baseline_compilations_++; break; case CompilationKind::kOptimized: number_of_optimized_compilations_++; break; } // We need to update the debug info before the entry point gets set. // At the same time we want to do under JIT lock so that debug info and JIT maps are in sync. if (!debug_info.empty()) { // NB: Don't allow packing of full info since it would remove non-backtrace data. AddNativeDebugInfoForJit(code_ptr, debug_info, /*allow_packing=*/ !is_full_debug_info); } // We need to update the entry point in the runnable state for the instrumentation. { // The following needs to be guarded by cha_lock_ also. Otherwise it's possible that the // compiled code is considered invalidated by some class linking, but below we still make the // compiled code valid for the method. Need cha_lock_ for checking all single-implementation // flags and register dependencies. MutexLock cha_mu(self, *Locks::cha_lock_); bool single_impl_still_valid = true; for (ArtMethod* single_impl : cha_single_implementation_list) { if (!single_impl->HasSingleImplementation()) { // Simply discard the compiled code. Clear the counter so that it may be recompiled later. // Hopefully the class hierarchy will be more stable when compilation is retried. single_impl_still_valid = false; ClearMethodCounter(method, /*was_warm=*/ false); break; } } // Discard the code if any single-implementation assumptions are now invalid. if (UNLIKELY(!single_impl_still_valid)) { VLOG(jit) << "JIT discarded jitted code due to invalid single-implementation assumptions."; return false; } DCHECK(cha_single_implementation_list.empty() || !Runtime::Current()->IsJavaDebuggable()) << "Should not be using cha on debuggable apps/runs!"; ClassLinker* class_linker = Runtime::Current()->GetClassLinker(); for (ArtMethod* single_impl : cha_single_implementation_list) { class_linker->GetClassHierarchyAnalysis()->AddDependency(single_impl, method, method_header); } if (UNLIKELY(method->IsNative())) { auto it = jni_stubs_map_.find(JniStubKey(method)); DCHECK(it != jni_stubs_map_.end()) << "Entry inserted in NotifyCompilationOf() should be alive."; JniStubData* data = &it->second; DCHECK(ContainsElement(data->GetMethods(), method)) << "Entry inserted in NotifyCompilationOf() should contain this method."; data->SetCode(code_ptr); data->UpdateEntryPoints(method_header->GetEntryPoint()); } else { if (method->IsPreCompiled() && IsSharedRegion(*region)) { zygote_map_.Put(code_ptr, method); } else { method_code_map_.Put(code_ptr, method); } if (compilation_kind == CompilationKind::kOsr) { osr_code_map_.Put(method, code_ptr); } else if (NeedsClinitCheckBeforeCall(method) && !method->GetDeclaringClass()->IsVisiblyInitialized()) { // This situation currently only occurs in the jit-zygote mode. DCHECK(!garbage_collect_code_); DCHECK(method->IsPreCompiled()); // The shared region can easily be queried. For the private region, we // use a side map. if (!IsSharedRegion(*region)) { saved_compiled_methods_map_.Put(method, code_ptr); } } else { Runtime::Current()->GetInstrumentation()->UpdateMethodsCode( method, method_header->GetEntryPoint()); } } if (collection_in_progress_) { // We need to update the live bitmap if there is a GC to ensure it sees this new // code. GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(code_ptr)); } VLOG(jit) << "JIT added (kind=" << compilation_kind << ") " << ArtMethod::PrettyMethod(method) << "@" << method << " ccache_size=" << PrettySize(CodeCacheSizeLocked()) << ": " << " dcache_size=" << PrettySize(DataCacheSizeLocked()) << ": " << reinterpret_cast(method_header->GetEntryPoint()) << "," << reinterpret_cast(method_header->GetEntryPoint() + method_header->GetCodeSize()); } return true; } size_t JitCodeCache::CodeCacheSize() { MutexLock mu(Thread::Current(), *Locks::jit_lock_); return CodeCacheSizeLocked(); } bool JitCodeCache::RemoveMethod(ArtMethod* method, bool release_memory) { // This function is used only for testing and only with non-native methods. CHECK(!method->IsNative()); MutexLock mu(Thread::Current(), *Locks::jit_lock_); bool osr = osr_code_map_.find(method) != osr_code_map_.end(); bool in_cache = RemoveMethodLocked(method, release_memory); if (!in_cache) { return false; } method->SetCounter(0); Runtime::Current()->GetInstrumentation()->UpdateMethodsCode( method, GetQuickToInterpreterBridge()); VLOG(jit) << "JIT removed (osr=" << std::boolalpha << osr << std::noboolalpha << ") " << ArtMethod::PrettyMethod(method) << "@" << method << " ccache_size=" << PrettySize(CodeCacheSizeLocked()) << ": " << " dcache_size=" << PrettySize(DataCacheSizeLocked()); return true; } bool JitCodeCache::RemoveMethodLocked(ArtMethod* method, bool release_memory) { if (LIKELY(!method->IsNative())) { auto it = profiling_infos_.find(method); if (it != profiling_infos_.end()) { profiling_infos_.erase(it); } } bool in_cache = false; ScopedCodeCacheWrite ccw(private_region_); if (UNLIKELY(method->IsNative())) { auto it = jni_stubs_map_.find(JniStubKey(method)); if (it != jni_stubs_map_.end() && it->second.RemoveMethod(method)) { in_cache = true; if (it->second.GetMethods().empty()) { if (release_memory) { FreeCodeAndData(it->second.GetCode()); } jni_stubs_map_.erase(it); } else { it->first.UpdateShorty(it->second.GetMethods().front()); } } } else { for (auto it = method_code_map_.begin(); it != method_code_map_.end();) { if (it->second == method) { in_cache = true; if (release_memory) { FreeCodeAndData(it->first); } VLOG(jit) << "JIT removed " << it->second->PrettyMethod() << ": " << it->first; it = method_code_map_.erase(it); } else { ++it; } } auto osr_it = osr_code_map_.find(method); if (osr_it != osr_code_map_.end()) { osr_code_map_.erase(osr_it); } } return in_cache; } // This notifies the code cache that the given method has been redefined and that it should remove // any cached information it has on the method. All threads must be suspended before calling this // method. The compiled code for the method (if there is any) must not be in any threads call stack. void JitCodeCache::NotifyMethodRedefined(ArtMethod* method) { MutexLock mu(Thread::Current(), *Locks::jit_lock_); RemoveMethodLocked(method, /* release_memory= */ true); } // This invalidates old_method. Once this function returns one can no longer use old_method to // execute code unless it is fixed up. This fixup will happen later in the process of installing a // class redefinition. // TODO We should add some info to ArtMethod to note that 'old_method' has been invalidated and // shouldn't be used since it is no longer logically in the jit code cache. // TODO We should add DCHECKS that validate that the JIT is paused when this method is entered. void JitCodeCache::MoveObsoleteMethod(ArtMethod* old_method, ArtMethod* new_method) { MutexLock mu(Thread::Current(), *Locks::jit_lock_); if (old_method->IsNative()) { // Update methods in jni_stubs_map_. for (auto& entry : jni_stubs_map_) { JniStubData& data = entry.second; data.MoveObsoleteMethod(old_method, new_method); } return; } // Update method_code_map_ to point to the new method. for (auto& it : method_code_map_) { if (it.second == old_method) { it.second = new_method; } } // Update osr_code_map_ to point to the new method. auto code_map = osr_code_map_.find(old_method); if (code_map != osr_code_map_.end()) { osr_code_map_.Put(new_method, code_map->second); osr_code_map_.erase(old_method); } } void JitCodeCache::TransitionToDebuggable() { // Check that none of our methods have an entrypoint in the zygote exec // space (this should be taken care of by // ClassLinker::UpdateEntryPointsClassVisitor. { MutexLock mu(Thread::Current(), *Locks::jit_lock_); if (kIsDebugBuild) { for (const auto& it : method_code_map_) { ArtMethod* method = it.second; DCHECK(!method->IsPreCompiled()); DCHECK(!IsInZygoteExecSpace(method->GetEntryPointFromQuickCompiledCode())); } } // Not strictly necessary, but this map is useless now. saved_compiled_methods_map_.clear(); } if (kIsDebugBuild) { for (const auto& entry : zygote_map_) { ArtMethod* method = entry.method; if (method != nullptr) { DCHECK(!method->IsPreCompiled()); DCHECK(!IsInZygoteExecSpace(method->GetEntryPointFromQuickCompiledCode())); } } } } size_t JitCodeCache::CodeCacheSizeLocked() { return GetCurrentRegion()->GetUsedMemoryForCode(); } size_t JitCodeCache::DataCacheSize() { MutexLock mu(Thread::Current(), *Locks::jit_lock_); return DataCacheSizeLocked(); } size_t JitCodeCache::DataCacheSizeLocked() { return GetCurrentRegion()->GetUsedMemoryForData(); } bool JitCodeCache::Reserve(Thread* self, JitMemoryRegion* region, size_t code_size, size_t stack_map_size, size_t number_of_roots, ArtMethod* method, /*out*/ArrayRef* reserved_code, /*out*/ArrayRef* reserved_data) { code_size = OatQuickMethodHeader::InstructionAlignedSize() + code_size; size_t data_size = RoundUp(ComputeRootTableSize(number_of_roots) + stack_map_size, sizeof(void*)); const uint8_t* code; const uint8_t* data; while (true) { bool at_max_capacity = false; { ScopedThreadSuspension sts(self, kSuspended); MutexLock mu(self, *Locks::jit_lock_); WaitForPotentialCollectionToComplete(self); ScopedCodeCacheWrite ccw(*region); code = region->AllocateCode(code_size); data = region->AllocateData(data_size); at_max_capacity = IsAtMaxCapacity(); } if (code != nullptr && data != nullptr) { break; } Free(self, region, code, data); if (at_max_capacity) { VLOG(jit) << "JIT failed to allocate code of size " << PrettySize(code_size) << ", and data of size " << PrettySize(data_size); return false; } // Run a code cache collection and try again. GarbageCollectCache(self); } *reserved_code = ArrayRef(code, code_size); *reserved_data = ArrayRef(data, data_size); MutexLock mu(self, *Locks::jit_lock_); histogram_code_memory_use_.AddValue(code_size); if (code_size > kCodeSizeLogThreshold) { LOG(INFO) << "JIT allocated " << PrettySize(code_size) << " for compiled code of " << ArtMethod::PrettyMethod(method); } histogram_stack_map_memory_use_.AddValue(data_size); if (data_size > kStackMapSizeLogThreshold) { LOG(INFO) << "JIT allocated " << PrettySize(data_size) << " for stack maps of " << ArtMethod::PrettyMethod(method); } return true; } void JitCodeCache::Free(Thread* self, JitMemoryRegion* region, const uint8_t* code, const uint8_t* data) { MutexLock mu(self, *Locks::jit_lock_); ScopedCodeCacheWrite ccw(*region); FreeLocked(region, code, data); } void JitCodeCache::FreeLocked(JitMemoryRegion* region, const uint8_t* code, const uint8_t* data) { if (code != nullptr) { RemoveNativeDebugInfoForJit(reinterpret_cast(FromAllocationToCode(code))); region->FreeCode(code); } if (data != nullptr) { region->FreeData(data); } } class MarkCodeClosure final : public Closure { public: MarkCodeClosure(JitCodeCache* code_cache, CodeCacheBitmap* bitmap, Barrier* barrier) : code_cache_(code_cache), bitmap_(bitmap), barrier_(barrier) {} void Run(Thread* thread) override REQUIRES_SHARED(Locks::mutator_lock_) { ScopedTrace trace(__PRETTY_FUNCTION__); DCHECK(thread == Thread::Current() || thread->IsSuspended()); StackVisitor::WalkStack( [&](const art::StackVisitor* stack_visitor) { const OatQuickMethodHeader* method_header = stack_visitor->GetCurrentOatQuickMethodHeader(); if (method_header == nullptr) { return true; } const void* code = method_header->GetCode(); if (code_cache_->ContainsPc(code) && !code_cache_->IsInZygoteExecSpace(code)) { // Use the atomic set version, as multiple threads are executing this code. bitmap_->AtomicTestAndSet(FromCodeToAllocation(code)); } return true; }, thread, /* context= */ nullptr, art::StackVisitor::StackWalkKind::kSkipInlinedFrames); if (kIsDebugBuild) { // The stack walking code queries the side instrumentation stack if it // sees an instrumentation exit pc, so the JIT code of methods in that stack // must have been seen. We check this below. for (const auto& it : *thread->GetInstrumentationStack()) { // The 'method_' in InstrumentationStackFrame is the one that has return_pc_ in // its stack frame, it is not the method owning return_pc_. We just pass null to // LookupMethodHeader: the method is only checked against in debug builds. OatQuickMethodHeader* method_header = code_cache_->LookupMethodHeader(it.second.return_pc_, /* method= */ nullptr); if (method_header != nullptr) { const void* code = method_header->GetCode(); CHECK(bitmap_->Test(FromCodeToAllocation(code))); } } } barrier_->Pass(Thread::Current()); } private: JitCodeCache* const code_cache_; CodeCacheBitmap* const bitmap_; Barrier* const barrier_; }; void JitCodeCache::NotifyCollectionDone(Thread* self) { collection_in_progress_ = false; lock_cond_.Broadcast(self); } void JitCodeCache::MarkCompiledCodeOnThreadStacks(Thread* self) { Barrier barrier(0); size_t threads_running_checkpoint = 0; MarkCodeClosure closure(this, GetLiveBitmap(), &barrier); threads_running_checkpoint = Runtime::Current()->GetThreadList()->RunCheckpoint(&closure); // Now that we have run our checkpoint, move to a suspended state and wait // for other threads to run the checkpoint. ScopedThreadSuspension sts(self, kSuspended); if (threads_running_checkpoint != 0) { barrier.Increment(self, threads_running_checkpoint); } } bool JitCodeCache::IsAtMaxCapacity() const { return private_region_.GetCurrentCapacity() == private_region_.GetMaxCapacity(); } bool JitCodeCache::ShouldDoFullCollection() { if (IsAtMaxCapacity()) { // Always do a full collection when the code cache is full. return true; } else if (private_region_.GetCurrentCapacity() < kReservedCapacity) { // Always do partial collection when the code cache size is below the reserved // capacity. return false; } else if (last_collection_increased_code_cache_) { // This time do a full collection. return true; } else { // This time do a partial collection. return false; } } void JitCodeCache::GarbageCollectCache(Thread* self) { ScopedTrace trace(__FUNCTION__); // Wait for an existing collection, or let everyone know we are starting one. { ScopedThreadSuspension sts(self, kSuspended); MutexLock mu(self, *Locks::jit_lock_); if (!garbage_collect_code_) { private_region_.IncreaseCodeCacheCapacity(); return; } else if (WaitForPotentialCollectionToComplete(self)) { return; } else { number_of_collections_++; live_bitmap_.reset(CodeCacheBitmap::Create( "code-cache-bitmap", reinterpret_cast(private_region_.GetExecPages()->Begin()), reinterpret_cast( private_region_.GetExecPages()->Begin() + private_region_.GetCurrentCapacity() / 2))); collection_in_progress_ = true; } } TimingLogger logger("JIT code cache timing logger", true, VLOG_IS_ON(jit)); { TimingLogger::ScopedTiming st("Code cache collection", &logger); bool do_full_collection = false; { MutexLock mu(self, *Locks::jit_lock_); do_full_collection = ShouldDoFullCollection(); } VLOG(jit) << "Do " << (do_full_collection ? "full" : "partial") << " code cache collection, code=" << PrettySize(CodeCacheSize()) << ", data=" << PrettySize(DataCacheSize()); DoCollection(self, /* collect_profiling_info= */ do_full_collection); VLOG(jit) << "After code cache collection, code=" << PrettySize(CodeCacheSize()) << ", data=" << PrettySize(DataCacheSize()); { MutexLock mu(self, *Locks::jit_lock_); // Increase the code cache only when we do partial collections. // TODO: base this strategy on how full the code cache is? if (do_full_collection) { last_collection_increased_code_cache_ = false; } else { last_collection_increased_code_cache_ = true; private_region_.IncreaseCodeCacheCapacity(); } bool next_collection_will_be_full = ShouldDoFullCollection(); // Start polling the liveness of compiled code to prepare for the next full collection. if (next_collection_will_be_full) { for (auto it : profiling_infos_) { it.second->SetBaselineHotnessCount(0); } // Change entry points of native methods back to the GenericJNI entrypoint. for (const auto& entry : jni_stubs_map_) { const JniStubData& data = entry.second; if (!data.IsCompiled() || IsInZygoteExecSpace(data.GetCode())) { continue; } // Make sure a single invocation of the GenericJNI trampoline tries to recompile. uint16_t new_counter = Runtime::Current()->GetJit()->HotMethodThreshold() - 1u; const OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(data.GetCode()); for (ArtMethod* method : data.GetMethods()) { if (method->GetEntryPointFromQuickCompiledCode() == method_header->GetEntryPoint()) { // Don't call Instrumentation::UpdateMethodsCode(), same as for normal methods above. method->SetCounter(new_counter); method->SetEntryPointFromQuickCompiledCode(GetQuickGenericJniStub()); } } } } live_bitmap_.reset(nullptr); NotifyCollectionDone(self); } } Runtime::Current()->GetJit()->AddTimingLogger(logger); } void JitCodeCache::RemoveUnmarkedCode(Thread* self) { ScopedTrace trace(__FUNCTION__); std::unordered_set method_headers; { MutexLock mu(self, *Locks::jit_lock_); // Iterate over all compiled code and remove entries that are not marked. for (auto it = jni_stubs_map_.begin(); it != jni_stubs_map_.end();) { JniStubData* data = &it->second; if (IsInZygoteExecSpace(data->GetCode()) || !data->IsCompiled() || GetLiveBitmap()->Test(FromCodeToAllocation(data->GetCode()))) { ++it; } else { method_headers.insert(OatQuickMethodHeader::FromCodePointer(data->GetCode())); for (ArtMethod* method : data->GetMethods()) { VLOG(jit) << "JIT removed (JNI) " << method->PrettyMethod() << ": " << data->GetCode(); } it = jni_stubs_map_.erase(it); } } for (auto it = method_code_map_.begin(); it != method_code_map_.end();) { const void* code_ptr = it->first; uintptr_t allocation = FromCodeToAllocation(code_ptr); if (IsInZygoteExecSpace(code_ptr) || GetLiveBitmap()->Test(allocation)) { ++it; } else { OatQuickMethodHeader* header = OatQuickMethodHeader::FromCodePointer(code_ptr); method_headers.insert(header); VLOG(jit) << "JIT removed " << it->second->PrettyMethod() << ": " << it->first; it = method_code_map_.erase(it); } } FreeAllMethodHeaders(method_headers); } } bool JitCodeCache::GetGarbageCollectCode() { MutexLock mu(Thread::Current(), *Locks::jit_lock_); return garbage_collect_code_; } void JitCodeCache::SetGarbageCollectCode(bool value) { Thread* self = Thread::Current(); MutexLock mu(self, *Locks::jit_lock_); // Update the flag while holding the lock to ensure no thread will try to GC. garbage_collect_code_ = value; } void JitCodeCache::RemoveMethodBeingCompiled(ArtMethod* method, CompilationKind kind) { DCHECK(IsMethodBeingCompiled(method, kind)); switch (kind) { case CompilationKind::kOsr: current_osr_compilations_.erase(method); break; case CompilationKind::kBaseline: current_baseline_compilations_.erase(method); break; case CompilationKind::kOptimized: current_optimized_compilations_.erase(method); break; } } void JitCodeCache::AddMethodBeingCompiled(ArtMethod* method, CompilationKind kind) { DCHECK(!IsMethodBeingCompiled(method, kind)); switch (kind) { case CompilationKind::kOsr: current_osr_compilations_.insert(method); break; case CompilationKind::kBaseline: current_baseline_compilations_.insert(method); break; case CompilationKind::kOptimized: current_optimized_compilations_.insert(method); break; } } bool JitCodeCache::IsMethodBeingCompiled(ArtMethod* method, CompilationKind kind) { switch (kind) { case CompilationKind::kOsr: return ContainsElement(current_osr_compilations_, method); case CompilationKind::kBaseline: return ContainsElement(current_baseline_compilations_, method); case CompilationKind::kOptimized: return ContainsElement(current_optimized_compilations_, method); } } bool JitCodeCache::IsMethodBeingCompiled(ArtMethod* method) { return ContainsElement(current_optimized_compilations_, method) || ContainsElement(current_osr_compilations_, method) || ContainsElement(current_baseline_compilations_, method); } void JitCodeCache::DoCollection(Thread* self, bool collect_profiling_info) { ScopedTrace trace(__FUNCTION__); { MutexLock mu(self, *Locks::jit_lock_); // Update to interpreter the methods that have baseline entrypoints and whose baseline // hotness count is zero. // Note that these methods may be in thread stack or concurrently revived // between. That's OK, as the thread executing it will mark it. for (auto it : profiling_infos_) { ProfilingInfo* info = it.second; if (info->GetBaselineHotnessCount() == 0) { const void* entry_point = info->GetMethod()->GetEntryPointFromQuickCompiledCode(); if (ContainsPc(entry_point)) { OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromEntryPoint(entry_point); if (CodeInfo::IsBaseline(method_header->GetOptimizedCodeInfoPtr())) { info->GetMethod()->SetEntryPointFromQuickCompiledCode(GetQuickToInterpreterBridge()); } } } } // TODO: collect profiling info // TODO: collect optimized code // Mark compiled code that are entrypoints of ArtMethods. Compiled code that is not // an entry point is either: // - an osr compiled code, that will be removed if not in a thread call stack. // - discarded compiled code, that will be removed if not in a thread call stack. for (const auto& entry : jni_stubs_map_) { const JniStubData& data = entry.second; const void* code_ptr = data.GetCode(); if (IsInZygoteExecSpace(code_ptr)) { continue; } const OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr); for (ArtMethod* method : data.GetMethods()) { if (method_header->GetEntryPoint() == method->GetEntryPointFromQuickCompiledCode()) { GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(code_ptr)); break; } } } for (const auto& it : method_code_map_) { ArtMethod* method = it.second; const void* code_ptr = it.first; if (IsInZygoteExecSpace(code_ptr)) { continue; } const OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr); if (method_header->GetEntryPoint() == method->GetEntryPointFromQuickCompiledCode()) { GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(code_ptr)); } } // Empty osr method map, as osr compiled code will be deleted (except the ones // on thread stacks). osr_code_map_.clear(); } // Run a checkpoint on all threads to mark the JIT compiled code they are running. MarkCompiledCodeOnThreadStacks(self); // At this point, mutator threads are still running, and entrypoints of methods can // change. We do know they cannot change to a code cache entry that is not marked, // therefore we can safely remove those entries. RemoveUnmarkedCode(self); if (collect_profiling_info) { // TODO: Collect unused profiling infos. } } OatQuickMethodHeader* JitCodeCache::LookupMethodHeader(uintptr_t pc, ArtMethod* method) { static_assert(kRuntimeISA != InstructionSet::kThumb2, "kThumb2 cannot be a runtime ISA"); if (kRuntimeISA == InstructionSet::kArm) { // On Thumb-2, the pc is offset by one. --pc; } if (!ContainsPc(reinterpret_cast(pc))) { return nullptr; } if (!kIsDebugBuild) { // Called with null `method` only from MarkCodeClosure::Run() in debug build. CHECK(method != nullptr); } MutexLock mu(Thread::Current(), *Locks::jit_lock_); OatQuickMethodHeader* method_header = nullptr; ArtMethod* found_method = nullptr; // Only for DCHECK(), not for JNI stubs. if (method != nullptr && UNLIKELY(method->IsNative())) { auto it = jni_stubs_map_.find(JniStubKey(method)); if (it == jni_stubs_map_.end() || !ContainsElement(it->second.GetMethods(), method)) { return nullptr; } const void* code_ptr = it->second.GetCode(); method_header = OatQuickMethodHeader::FromCodePointer(code_ptr); if (!method_header->Contains(pc)) { return nullptr; } } else { if (shared_region_.IsInExecSpace(reinterpret_cast(pc))) { const void* code_ptr = zygote_map_.GetCodeFor(method, pc); if (code_ptr != nullptr) { return OatQuickMethodHeader::FromCodePointer(code_ptr); } } auto it = method_code_map_.lower_bound(reinterpret_cast(pc)); if (it != method_code_map_.begin()) { --it; const void* code_ptr = it->first; if (OatQuickMethodHeader::FromCodePointer(code_ptr)->Contains(pc)) { method_header = OatQuickMethodHeader::FromCodePointer(code_ptr); found_method = it->second; } } if (method_header == nullptr && method == nullptr) { // Scan all compiled JNI stubs as well. This slow search is used only // for checks in debug build, for release builds the `method` is not null. for (auto&& entry : jni_stubs_map_) { const JniStubData& data = entry.second; if (data.IsCompiled() && OatQuickMethodHeader::FromCodePointer(data.GetCode())->Contains(pc)) { method_header = OatQuickMethodHeader::FromCodePointer(data.GetCode()); } } } if (method_header == nullptr) { return nullptr; } } if (kIsDebugBuild && method != nullptr && !method->IsNative()) { DCHECK_EQ(found_method, method) << ArtMethod::PrettyMethod(method) << " " << ArtMethod::PrettyMethod(found_method) << " " << std::hex << pc; } return method_header; } OatQuickMethodHeader* JitCodeCache::LookupOsrMethodHeader(ArtMethod* method) { MutexLock mu(Thread::Current(), *Locks::jit_lock_); auto it = osr_code_map_.find(method); if (it == osr_code_map_.end()) { return nullptr; } return OatQuickMethodHeader::FromCodePointer(it->second); } ProfilingInfo* JitCodeCache::AddProfilingInfo(Thread* self, ArtMethod* method, const std::vector& entries) { DCHECK(CanAllocateProfilingInfo()); ProfilingInfo* info = nullptr; { MutexLock mu(self, *Locks::jit_lock_); info = AddProfilingInfoInternal(self, method, entries); } if (info == nullptr) { GarbageCollectCache(self); MutexLock mu(self, *Locks::jit_lock_); info = AddProfilingInfoInternal(self, method, entries); } return info; } ProfilingInfo* JitCodeCache::AddProfilingInfoInternal(Thread* self ATTRIBUTE_UNUSED, ArtMethod* method, const std::vector& entries) { // Check whether some other thread has concurrently created it. auto it = profiling_infos_.find(method); if (it != profiling_infos_.end()) { return it->second; } size_t profile_info_size = RoundUp( sizeof(ProfilingInfo) + sizeof(InlineCache) * entries.size(), sizeof(void*)); const uint8_t* data = private_region_.AllocateData(profile_info_size); if (data == nullptr) { return nullptr; } uint8_t* writable_data = private_region_.GetWritableDataAddress(data); ProfilingInfo* info = new (writable_data) ProfilingInfo(method, entries); profiling_infos_.Put(method, info); histogram_profiling_info_memory_use_.AddValue(profile_info_size); return info; } void* JitCodeCache::MoreCore(const void* mspace, intptr_t increment) { return shared_region_.OwnsSpace(mspace) ? shared_region_.MoreCore(mspace, increment) : private_region_.MoreCore(mspace, increment); } void JitCodeCache::GetProfiledMethods(const std::set& dex_base_locations, std::vector& methods) { Thread* self = Thread::Current(); WaitUntilInlineCacheAccessible(self); MutexLock mu(self, *Locks::jit_lock_); ScopedTrace trace(__FUNCTION__); uint16_t jit_compile_threshold = Runtime::Current()->GetJITOptions()->GetCompileThreshold(); for (auto it : profiling_infos_) { ProfilingInfo* info = it.second; ArtMethod* method = info->GetMethod(); const DexFile* dex_file = method->GetDexFile(); const std::string base_location = DexFileLoader::GetBaseLocation(dex_file->GetLocation()); if (!ContainsElement(dex_base_locations, base_location)) { // Skip dex files which are not profiled. continue; } std::vector inline_caches; // If the method didn't reach the compilation threshold don't save the inline caches. // They might be incomplete and cause unnecessary deoptimizations. // If the inline cache is empty the compiler will generate a regular invoke virtual/interface. if (method->GetCounter() < jit_compile_threshold) { methods.emplace_back(/*ProfileMethodInfo*/ MethodReference(dex_file, method->GetDexMethodIndex()), inline_caches); continue; } for (size_t i = 0; i < info->number_of_inline_caches_; ++i) { std::vector profile_classes; const InlineCache& cache = info->cache_[i]; ArtMethod* caller = info->GetMethod(); bool is_missing_types = false; for (size_t k = 0; k < InlineCache::kIndividualCacheSize; k++) { mirror::Class* cls = cache.classes_[k].Read(); if (cls == nullptr) { break; } // Check if the receiver is in the boot class path or if it's in the // same class loader as the caller. If not, skip it, as there is not // much we can do during AOT. if (!cls->IsBootStrapClassLoaded() && caller->GetClassLoader() != cls->GetClassLoader()) { is_missing_types = true; continue; } const DexFile* class_dex_file = nullptr; dex::TypeIndex type_index; if (cls->GetDexCache() == nullptr) { DCHECK(cls->IsArrayClass()) << cls->PrettyClass(); // Make a best effort to find the type index in the method's dex file. // We could search all open dex files but that might turn expensive // and probably not worth it. class_dex_file = dex_file; type_index = cls->FindTypeIndexInOtherDexFile(*dex_file); } else { class_dex_file = &(cls->GetDexFile()); type_index = cls->GetDexTypeIndex(); } if (!type_index.IsValid()) { // Could be a proxy class or an array for which we couldn't find the type index. is_missing_types = true; continue; } if (ContainsElement(dex_base_locations, DexFileLoader::GetBaseLocation(class_dex_file->GetLocation()))) { // Only consider classes from the same apk (including multidex). profile_classes.emplace_back(/*ProfileMethodInfo::ProfileClassReference*/ class_dex_file, type_index); } else { is_missing_types = true; } } if (!profile_classes.empty()) { inline_caches.emplace_back(/*ProfileMethodInfo::ProfileInlineCache*/ cache.dex_pc_, is_missing_types, profile_classes); } } methods.emplace_back(/*ProfileMethodInfo*/ MethodReference(dex_file, method->GetDexMethodIndex()), inline_caches); } } bool JitCodeCache::IsOsrCompiled(ArtMethod* method) { MutexLock mu(Thread::Current(), *Locks::jit_lock_); return osr_code_map_.find(method) != osr_code_map_.end(); } bool JitCodeCache::NotifyCompilationOf(ArtMethod* method, Thread* self, CompilationKind compilation_kind, bool prejit) { const void* existing_entry_point = method->GetEntryPointFromQuickCompiledCode(); if (compilation_kind != CompilationKind::kOsr && ContainsPc(existing_entry_point)) { OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromEntryPoint(existing_entry_point); bool is_baseline = (compilation_kind == CompilationKind::kBaseline); if (CodeInfo::IsBaseline(method_header->GetOptimizedCodeInfoPtr()) == is_baseline) { VLOG(jit) << "Not compiling " << method->PrettyMethod() << " because it has already been compiled" << " kind=" << compilation_kind; return false; } } if (NeedsClinitCheckBeforeCall(method) && !prejit) { // We do not need a synchronization barrier for checking the visibly initialized status // or checking the initialized status just for requesting visible initialization. ClassStatus status = method->GetDeclaringClass() ->GetStatus(); if (status != ClassStatus::kVisiblyInitialized) { // Unless we're pre-jitting, we currently don't save the JIT compiled code if we cannot // update the entrypoint due to needing an initialization check. if (status == ClassStatus::kInitialized) { // Request visible initialization but do not block to allow compiling other methods. // Hopefully, this will complete by the time the method becomes hot again. Runtime::Current()->GetClassLinker()->MakeInitializedClassesVisiblyInitialized( self, /*wait=*/ false); } VLOG(jit) << "Not compiling " << method->PrettyMethod() << " because it has the resolution stub"; // Give it a new chance to be hot. ClearMethodCounter(method, /*was_warm=*/ false); return false; } } if (compilation_kind == CompilationKind::kOsr) { MutexLock mu(self, *Locks::jit_lock_); if (osr_code_map_.find(method) != osr_code_map_.end()) { return false; } } if (UNLIKELY(method->IsNative())) { MutexLock mu(self, *Locks::jit_lock_); JniStubKey key(method); auto it = jni_stubs_map_.find(key); bool new_compilation = false; if (it == jni_stubs_map_.end()) { // Create a new entry to mark the stub as being compiled. it = jni_stubs_map_.Put(key, JniStubData{}); new_compilation = true; } JniStubData* data = &it->second; data->AddMethod(method); if (data->IsCompiled()) { OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(data->GetCode()); const void* entrypoint = method_header->GetEntryPoint(); // Update also entrypoints of other methods held by the JniStubData. // We could simply update the entrypoint of `method` but if the last JIT GC has // changed these entrypoints to GenericJNI in preparation for a full GC, we may // as well change them back as this stub shall not be collected anyway and this // can avoid a few expensive GenericJNI calls. data->UpdateEntryPoints(entrypoint); if (collection_in_progress_) { if (!IsInZygoteExecSpace(data->GetCode())) { GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(data->GetCode())); } } } return new_compilation; } else { if (CanAllocateProfilingInfo() && (compilation_kind == CompilationKind::kBaseline)) { bool has_profiling_info = false; { MutexLock mu(self, *Locks::jit_lock_); has_profiling_info = (profiling_infos_.find(method) != profiling_infos_.end()); } if (!has_profiling_info) { if (ProfilingInfo::Create(self, method) == nullptr) { VLOG(jit) << method->PrettyMethod() << " needs a ProfilingInfo to be compiled baseline"; ClearMethodCounter(method, /*was_warm=*/ false); return false; } } } MutexLock mu(self, *Locks::jit_lock_); if (IsMethodBeingCompiled(method, compilation_kind)) { return false; } AddMethodBeingCompiled(method, compilation_kind); return true; } } ProfilingInfo* JitCodeCache::NotifyCompilerUse(ArtMethod* method, Thread* self) { MutexLock mu(self, *Locks::jit_lock_); auto it = profiling_infos_.find(method); if (it == profiling_infos_.end()) { return nullptr; } if (!it->second->IncrementInlineUse()) { // Overflow of inlining uses, just bail. return nullptr; } return it->second; } void JitCodeCache::DoneCompilerUse(ArtMethod* method, Thread* self) { MutexLock mu(self, *Locks::jit_lock_); auto it = profiling_infos_.find(method); DCHECK(it != profiling_infos_.end()); it->second->DecrementInlineUse(); } void JitCodeCache::DoneCompiling(ArtMethod* method, Thread* self, CompilationKind compilation_kind) { DCHECK_EQ(Thread::Current(), self); MutexLock mu(self, *Locks::jit_lock_); if (UNLIKELY(method->IsNative())) { auto it = jni_stubs_map_.find(JniStubKey(method)); DCHECK(it != jni_stubs_map_.end()); JniStubData* data = &it->second; DCHECK(ContainsElement(data->GetMethods(), method)); if (UNLIKELY(!data->IsCompiled())) { // Failed to compile; the JNI compiler never fails, but the cache may be full. jni_stubs_map_.erase(it); // Remove the entry added in NotifyCompilationOf(). } // else Commit() updated entrypoints of all methods in the JniStubData. } else { RemoveMethodBeingCompiled(method, compilation_kind); } } void JitCodeCache::InvalidateAllCompiledCode() { art::MutexLock mu(Thread::Current(), *Locks::jit_lock_); VLOG(jit) << "Invalidating all compiled code"; ClassLinker* linker = Runtime::Current()->GetClassLinker(); for (auto it : method_code_map_) { ArtMethod* meth = it.second; // We were compiled, so we must be warm. ClearMethodCounter(meth, /*was_warm=*/true); if (meth->IsObsolete()) { linker->SetEntryPointsForObsoleteMethod(meth); } else { linker->SetEntryPointsToInterpreter(meth); } } saved_compiled_methods_map_.clear(); osr_code_map_.clear(); } void JitCodeCache::InvalidateCompiledCodeFor(ArtMethod* method, const OatQuickMethodHeader* header) { DCHECK(!method->IsNative()); const void* method_entrypoint = method->GetEntryPointFromQuickCompiledCode(); // Clear the method counter if we are running jitted code since we might want to jit this again in // the future. if (method_entrypoint == header->GetEntryPoint()) { // The entrypoint is the one to invalidate, so we just update it to the interpreter entry point // and clear the counter to get the method Jitted again. Runtime::Current()->GetInstrumentation()->UpdateMethodsCode( method, GetQuickToInterpreterBridge()); ClearMethodCounter(method, /*was_warm=*/ true); } else { MutexLock mu(Thread::Current(), *Locks::jit_lock_); auto it = osr_code_map_.find(method); if (it != osr_code_map_.end() && OatQuickMethodHeader::FromCodePointer(it->second) == header) { // Remove the OSR method, to avoid using it again. osr_code_map_.erase(it); } } // In case the method was pre-compiled, clear that information so we // can recompile it ourselves. if (method->IsPreCompiled()) { method->ClearPreCompiled(); } } void JitCodeCache::Dump(std::ostream& os) { MutexLock mu(Thread::Current(), *Locks::jit_lock_); os << "Current JIT code cache size (used / resident): " << GetCurrentRegion()->GetUsedMemoryForCode() / KB << "KB / " << GetCurrentRegion()->GetResidentMemoryForCode() / KB << "KB\n" << "Current JIT data cache size (used / resident): " << GetCurrentRegion()->GetUsedMemoryForData() / KB << "KB / " << GetCurrentRegion()->GetResidentMemoryForData() / KB << "KB\n"; if (!Runtime::Current()->IsZygote()) { os << "Zygote JIT code cache size (at point of fork): " << shared_region_.GetUsedMemoryForCode() / KB << "KB / " << shared_region_.GetResidentMemoryForCode() / KB << "KB\n" << "Zygote JIT data cache size (at point of fork): " << shared_region_.GetUsedMemoryForData() / KB << "KB / " << shared_region_.GetResidentMemoryForData() / KB << "KB\n"; } os << "Current JIT mini-debug-info size: " << PrettySize(GetJitMiniDebugInfoMemUsage()) << "\n" << "Current JIT capacity: " << PrettySize(GetCurrentRegion()->GetCurrentCapacity()) << "\n" << "Current number of JIT JNI stub entries: " << jni_stubs_map_.size() << "\n" << "Current number of JIT code cache entries: " << method_code_map_.size() << "\n" << "Total number of JIT baseline compilations: " << number_of_baseline_compilations_ << "\n" << "Total number of JIT optimized compilations: " << number_of_optimized_compilations_ << "\n" << "Total number of JIT compilations for on stack replacement: " << number_of_osr_compilations_ << "\n" << "Total number of JIT code cache collections: " << number_of_collections_ << std::endl; histogram_stack_map_memory_use_.PrintMemoryUse(os); histogram_code_memory_use_.PrintMemoryUse(os); histogram_profiling_info_memory_use_.PrintMemoryUse(os); } void JitCodeCache::PostForkChildAction(bool is_system_server, bool is_zygote) { Thread* self = Thread::Current(); // Remove potential tasks that have been inherited from the zygote. // We do this now and not in Jit::PostForkChildAction, as system server calls // JitCodeCache::PostForkChildAction first, and then does some code loading // that may result in new JIT tasks that we want to keep. ThreadPool* pool = Runtime::Current()->GetJit()->GetThreadPool(); if (pool != nullptr) { pool->RemoveAllTasks(self); } MutexLock mu(self, *Locks::jit_lock_); // Reset potential writable MemMaps inherited from the zygote. We never want // to write to them. shared_region_.ResetWritableMappings(); if (is_zygote || Runtime::Current()->IsSafeMode()) { // Don't create a private region for a child zygote. Regions are usually map shared // (to satisfy dual-view), and we don't want children of a child zygote to inherit it. return; } // Reset all statistics to be specific to this process. number_of_baseline_compilations_ = 0; number_of_optimized_compilations_ = 0; number_of_osr_compilations_ = 0; number_of_collections_ = 0; histogram_stack_map_memory_use_.Reset(); histogram_code_memory_use_.Reset(); histogram_profiling_info_memory_use_.Reset(); size_t initial_capacity = Runtime::Current()->GetJITOptions()->GetCodeCacheInitialCapacity(); size_t max_capacity = Runtime::Current()->GetJITOptions()->GetCodeCacheMaxCapacity(); std::string error_msg; if (!private_region_.Initialize(initial_capacity, max_capacity, /* rwx_memory_allowed= */ !is_system_server, is_zygote, &error_msg)) { LOG(WARNING) << "Could not create private region after zygote fork: " << error_msg; } } JitMemoryRegion* JitCodeCache::GetCurrentRegion() { return Runtime::Current()->IsZygote() ? &shared_region_ : &private_region_; } void JitCodeCache::VisitAllMethods(const std::function& cb) { for (const auto& it : jni_stubs_map_) { const JniStubData& data = it.second; if (data.IsCompiled()) { for (ArtMethod* method : data.GetMethods()) { cb(data.GetCode(), method); } } } for (auto it : method_code_map_) { // Includes OSR methods. cb(it.first, it.second); } for (auto it : saved_compiled_methods_map_) { cb(it.second, it.first); } for (auto it : zygote_map_) { if (it.code_ptr != nullptr && it.method != nullptr) { cb(it.code_ptr, it.method); } } } void ZygoteMap::Initialize(uint32_t number_of_methods) { MutexLock mu(Thread::Current(), *Locks::jit_lock_); // Allocate for 40-80% capacity. This will offer OK lookup times, and termination // cases. size_t capacity = RoundUpToPowerOfTwo(number_of_methods * 100 / 80); const uint8_t* memory = region_->AllocateData( capacity * sizeof(Entry) + sizeof(ZygoteCompilationState)); if (memory == nullptr) { LOG(WARNING) << "Could not allocate data for the zygote map"; return; } const Entry* data = reinterpret_cast(memory); region_->FillData(data, capacity, Entry { nullptr, nullptr }); map_ = ArrayRef(data, capacity); compilation_state_ = reinterpret_cast( memory + capacity * sizeof(Entry)); region_->WriteData(compilation_state_, ZygoteCompilationState::kInProgress); } const void* ZygoteMap::GetCodeFor(ArtMethod* method, uintptr_t pc) const { if (map_.empty()) { return nullptr; } if (method == nullptr) { // Do a linear search. This should only be used in debug builds. CHECK(kIsDebugBuild); for (const Entry& entry : map_) { const void* code_ptr = entry.code_ptr; if (code_ptr != nullptr) { OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr); if (method_header->Contains(pc)) { return code_ptr; } } } return nullptr; } std::hash hf; size_t index = hf(method) & (map_.size() - 1u); size_t original_index = index; // Loop over the array: we know this loop terminates as we will either // encounter the given method, or a null entry. Both terminate the loop. // Note that the zygote may concurrently write new entries to the map. That's OK as the // map is never resized. while (true) { const Entry& entry = map_[index]; if (entry.method == nullptr) { // Not compiled yet. return nullptr; } if (entry.method == method) { if (entry.code_ptr == nullptr) { // This is a race with the zygote which wrote the method, but hasn't written the // code. Just bail and wait for the next time we need the method. return nullptr; } if (pc != 0 && !OatQuickMethodHeader::FromCodePointer(entry.code_ptr)->Contains(pc)) { return nullptr; } return entry.code_ptr; } index = (index + 1) & (map_.size() - 1); DCHECK_NE(original_index, index); } } void ZygoteMap::Put(const void* code, ArtMethod* method) { if (map_.empty()) { return; } CHECK(Runtime::Current()->IsZygote()); std::hash hf; size_t index = hf(method) & (map_.size() - 1); size_t original_index = index; // Because the size of the map is bigger than the number of methods that will // be added, we are guaranteed to find a free slot in the array, and // therefore for this loop to terminate. while (true) { const Entry* entry = &map_[index]; if (entry->method == nullptr) { // Note that readers can read this memory concurrently, but that's OK as // we are writing pointers. region_->WriteData(entry, Entry { method, code }); break; } index = (index + 1) & (map_.size() - 1); DCHECK_NE(original_index, index); } DCHECK_EQ(GetCodeFor(method), code); } } // namespace jit } // namespace art