1 /*
2 * Copyright 2014 The Android Open Source Project
3 *
4 * Licensed under the Apache License, Version 2.0 (the "License");
5 * you may not use this file except in compliance with the License.
6 * You may obtain a copy of the License at
7 *
8 * http://www.apache.org/licenses/LICENSE-2.0
9 *
10 * Unless required by applicable law or agreed to in writing, software
11 * distributed under the License is distributed on an "AS IS" BASIS,
12 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13 * See the License for the specific language governing permissions and
14 * limitations under the License.
15 */
16
17 #include "jit_code_cache.h"
18
19 #include <sstream>
20
21 #include <android-base/logging.h>
22
23 #include "arch/context.h"
24 #include "art_method-inl.h"
25 #include "base/enums.h"
26 #include "base/histogram-inl.h"
27 #include "base/logging.h" // For VLOG.
28 #include "base/membarrier.h"
29 #include "base/memfd.h"
30 #include "base/mem_map.h"
31 #include "base/quasi_atomic.h"
32 #include "base/stl_util.h"
33 #include "base/systrace.h"
34 #include "base/time_utils.h"
35 #include "base/utils.h"
36 #include "cha.h"
37 #include "debugger_interface.h"
38 #include "dex/dex_file_loader.h"
39 #include "dex/method_reference.h"
40 #include "entrypoints/entrypoint_utils-inl.h"
41 #include "entrypoints/runtime_asm_entrypoints.h"
42 #include "gc/accounting/bitmap-inl.h"
43 #include "gc/allocator/dlmalloc.h"
44 #include "gc/scoped_gc_critical_section.h"
45 #include "handle.h"
46 #include "handle_scope-inl.h"
47 #include "instrumentation.h"
48 #include "intern_table.h"
49 #include "jit/jit.h"
50 #include "jit/profiling_info.h"
51 #include "jit/jit_scoped_code_cache_write.h"
52 #include "linear_alloc.h"
53 #include "oat_file-inl.h"
54 #include "oat_quick_method_header.h"
55 #include "object_callbacks.h"
56 #include "profile/profile_compilation_info.h"
57 #include "scoped_thread_state_change-inl.h"
58 #include "stack.h"
59 #include "thread-current-inl.h"
60 #include "thread-inl.h"
61 #include "thread_list.h"
62
63 namespace art {
64 namespace jit {
65
66 static constexpr size_t kCodeSizeLogThreshold = 50 * KB;
67 static constexpr size_t kStackMapSizeLogThreshold = 50 * KB;
68
69 class JitCodeCache::JniStubKey {
70 public:
REQUIRES_SHARED(Locks::mutator_lock_)71 explicit JniStubKey(ArtMethod* method) REQUIRES_SHARED(Locks::mutator_lock_)
72 : shorty_(method->GetShorty()),
73 is_static_(method->IsStatic()),
74 is_fast_native_(method->IsFastNative()),
75 is_critical_native_(method->IsCriticalNative()),
76 is_synchronized_(method->IsSynchronized()) {
77 DCHECK(!(is_fast_native_ && is_critical_native_));
78 }
79
operator <(const JniStubKey & rhs) const80 bool operator<(const JniStubKey& rhs) const {
81 if (is_static_ != rhs.is_static_) {
82 return rhs.is_static_;
83 }
84 if (is_synchronized_ != rhs.is_synchronized_) {
85 return rhs.is_synchronized_;
86 }
87 if (is_fast_native_ != rhs.is_fast_native_) {
88 return rhs.is_fast_native_;
89 }
90 if (is_critical_native_ != rhs.is_critical_native_) {
91 return rhs.is_critical_native_;
92 }
93 return strcmp(shorty_, rhs.shorty_) < 0;
94 }
95
96 // Update the shorty to point to another method's shorty. Call this function when removing
97 // the method that references the old shorty from JniCodeData and not removing the entire
98 // JniCodeData; the old shorty may become a dangling pointer when that method is unloaded.
UpdateShorty(ArtMethod * method) const99 void UpdateShorty(ArtMethod* method) const REQUIRES_SHARED(Locks::mutator_lock_) {
100 const char* shorty = method->GetShorty();
101 DCHECK_STREQ(shorty_, shorty);
102 shorty_ = shorty;
103 }
104
105 private:
106 // The shorty points to a DexFile data and may need to change
107 // to point to the same shorty in a different DexFile.
108 mutable const char* shorty_;
109
110 const bool is_static_;
111 const bool is_fast_native_;
112 const bool is_critical_native_;
113 const bool is_synchronized_;
114 };
115
116 class JitCodeCache::JniStubData {
117 public:
JniStubData()118 JniStubData() : code_(nullptr), methods_() {}
119
SetCode(const void * code)120 void SetCode(const void* code) {
121 DCHECK(code != nullptr);
122 code_ = code;
123 }
124
UpdateEntryPoints(const void * entrypoint)125 void UpdateEntryPoints(const void* entrypoint) REQUIRES_SHARED(Locks::mutator_lock_) {
126 DCHECK(IsCompiled());
127 DCHECK(entrypoint == OatQuickMethodHeader::FromCodePointer(GetCode())->GetEntryPoint());
128 instrumentation::Instrumentation* instrum = Runtime::Current()->GetInstrumentation();
129 for (ArtMethod* m : GetMethods()) {
130 // Because `m` might be in the process of being deleted:
131 // - Call the dedicated method instead of the more generic UpdateMethodsCode
132 // - Check the class status without a full read barrier; use ReadBarrier::IsMarked().
133 bool can_set_entrypoint = true;
134 if (NeedsClinitCheckBeforeCall(m)) {
135 // To avoid resurrecting an unreachable object, we must not use a full read
136 // barrier but we do not want to miss updating an entrypoint under common
137 // circumstances, i.e. during a GC the class becomes visibly initialized,
138 // the method becomes hot, we compile the thunk and want to update the
139 // entrypoint while the method's declaring class field still points to the
140 // from-space class object with the old status. Therefore we read the
141 // declaring class without a read barrier and check if it's already marked.
142 // If yes, we check the status of the to-space class object as intended.
143 // Otherwise, there is no to-space object and the from-space class object
144 // contains the most recent value of the status field; even if this races
145 // with another thread doing a read barrier and updating the status, that's
146 // no different from a race with a thread that just updates the status.
147 // Such race can happen only for the zygote method pre-compilation, as we
148 // otherwise compile only thunks for methods of visibly initialized classes.
149 ObjPtr<mirror::Class> klass = m->GetDeclaringClass<kWithoutReadBarrier>();
150 ObjPtr<mirror::Class> marked = ReadBarrier::IsMarked(klass.Ptr());
151 ObjPtr<mirror::Class> checked_klass = (marked != nullptr) ? marked : klass;
152 can_set_entrypoint = checked_klass->IsVisiblyInitialized();
153 }
154 if (can_set_entrypoint) {
155 instrum->UpdateNativeMethodsCodeToJitCode(m, entrypoint);
156 }
157 }
158 }
159
GetCode() const160 const void* GetCode() const {
161 return code_;
162 }
163
IsCompiled() const164 bool IsCompiled() const {
165 return GetCode() != nullptr;
166 }
167
AddMethod(ArtMethod * method)168 void AddMethod(ArtMethod* method) {
169 if (!ContainsElement(methods_, method)) {
170 methods_.push_back(method);
171 }
172 }
173
GetMethods() const174 const std::vector<ArtMethod*>& GetMethods() const {
175 return methods_;
176 }
177
RemoveMethodsIn(const LinearAlloc & alloc)178 void RemoveMethodsIn(const LinearAlloc& alloc) REQUIRES_SHARED(Locks::mutator_lock_) {
179 auto kept_end = std::partition(
180 methods_.begin(),
181 methods_.end(),
182 [&alloc](ArtMethod* method) { return !alloc.ContainsUnsafe(method); });
183 for (auto it = kept_end; it != methods_.end(); it++) {
184 VLOG(jit) << "JIT removed (JNI) " << (*it)->PrettyMethod() << ": " << code_;
185 }
186 methods_.erase(kept_end, methods_.end());
187 }
188
RemoveMethod(ArtMethod * method)189 bool RemoveMethod(ArtMethod* method) REQUIRES_SHARED(Locks::mutator_lock_) {
190 auto it = std::find(methods_.begin(), methods_.end(), method);
191 if (it != methods_.end()) {
192 VLOG(jit) << "JIT removed (JNI) " << (*it)->PrettyMethod() << ": " << code_;
193 methods_.erase(it);
194 return true;
195 } else {
196 return false;
197 }
198 }
199
MoveObsoleteMethod(ArtMethod * old_method,ArtMethod * new_method)200 void MoveObsoleteMethod(ArtMethod* old_method, ArtMethod* new_method) {
201 std::replace(methods_.begin(), methods_.end(), old_method, new_method);
202 }
203
204 private:
205 const void* code_;
206 std::vector<ArtMethod*> methods_;
207 };
208
Create(bool used_only_for_profile_data,bool rwx_memory_allowed,bool is_zygote,std::string * error_msg)209 JitCodeCache* JitCodeCache::Create(bool used_only_for_profile_data,
210 bool rwx_memory_allowed,
211 bool is_zygote,
212 std::string* error_msg) {
213 // Register for membarrier expedited sync core if JIT will be generating code.
214 if (!used_only_for_profile_data) {
215 if (art::membarrier(art::MembarrierCommand::kRegisterPrivateExpeditedSyncCore) != 0) {
216 // MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE ensures that CPU instruction pipelines are
217 // flushed and it's used when adding code to the JIT. The memory used by the new code may
218 // have just been released and, in theory, the old code could still be in a pipeline.
219 VLOG(jit) << "Kernel does not support membarrier sync-core";
220 }
221 }
222
223 size_t initial_capacity = Runtime::Current()->GetJITOptions()->GetCodeCacheInitialCapacity();
224 // Check whether the provided max capacity in options is below 1GB.
225 size_t max_capacity = Runtime::Current()->GetJITOptions()->GetCodeCacheMaxCapacity();
226 // We need to have 32 bit offsets from method headers in code cache which point to things
227 // in the data cache. If the maps are more than 4G apart, having multiple maps wouldn't work.
228 // Ensure we're below 1 GB to be safe.
229 if (max_capacity > 1 * GB) {
230 std::ostringstream oss;
231 oss << "Maxium code cache capacity is limited to 1 GB, "
232 << PrettySize(max_capacity) << " is too big";
233 *error_msg = oss.str();
234 return nullptr;
235 }
236
237 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
238 JitMemoryRegion region;
239 if (!region.Initialize(initial_capacity,
240 max_capacity,
241 rwx_memory_allowed,
242 is_zygote,
243 error_msg)) {
244 return nullptr;
245 }
246
247 std::unique_ptr<JitCodeCache> jit_code_cache(new JitCodeCache());
248 if (is_zygote) {
249 // Zygote should never collect code to share the memory with the children.
250 jit_code_cache->garbage_collect_code_ = false;
251 jit_code_cache->shared_region_ = std::move(region);
252 } else {
253 jit_code_cache->private_region_ = std::move(region);
254 }
255
256 VLOG(jit) << "Created jit code cache: initial capacity="
257 << PrettySize(initial_capacity)
258 << ", maximum capacity="
259 << PrettySize(max_capacity);
260
261 return jit_code_cache.release();
262 }
263
JitCodeCache()264 JitCodeCache::JitCodeCache()
265 : is_weak_access_enabled_(true),
266 inline_cache_cond_("Jit inline cache condition variable", *Locks::jit_lock_),
267 zygote_map_(&shared_region_),
268 lock_cond_("Jit code cache condition variable", *Locks::jit_lock_),
269 collection_in_progress_(false),
270 last_collection_increased_code_cache_(false),
271 garbage_collect_code_(true),
272 number_of_baseline_compilations_(0),
273 number_of_optimized_compilations_(0),
274 number_of_osr_compilations_(0),
275 number_of_collections_(0),
276 histogram_stack_map_memory_use_("Memory used for stack maps", 16),
277 histogram_code_memory_use_("Memory used for compiled code", 16),
278 histogram_profiling_info_memory_use_("Memory used for profiling info", 16) {
279 }
280
~JitCodeCache()281 JitCodeCache::~JitCodeCache() {}
282
PrivateRegionContainsPc(const void * ptr) const283 bool JitCodeCache::PrivateRegionContainsPc(const void* ptr) const {
284 return private_region_.IsInExecSpace(ptr);
285 }
286
ContainsPc(const void * ptr) const287 bool JitCodeCache::ContainsPc(const void* ptr) const {
288 return PrivateRegionContainsPc(ptr) || shared_region_.IsInExecSpace(ptr);
289 }
290
ContainsMethod(ArtMethod * method)291 bool JitCodeCache::ContainsMethod(ArtMethod* method) {
292 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
293 if (UNLIKELY(method->IsNative())) {
294 auto it = jni_stubs_map_.find(JniStubKey(method));
295 if (it != jni_stubs_map_.end() &&
296 it->second.IsCompiled() &&
297 ContainsElement(it->second.GetMethods(), method)) {
298 return true;
299 }
300 } else {
301 for (const auto& it : method_code_map_) {
302 if (it.second == method) {
303 return true;
304 }
305 }
306 if (zygote_map_.ContainsMethod(method)) {
307 return true;
308 }
309 }
310 return false;
311 }
312
GetJniStubCode(ArtMethod * method)313 const void* JitCodeCache::GetJniStubCode(ArtMethod* method) {
314 DCHECK(method->IsNative());
315 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
316 auto it = jni_stubs_map_.find(JniStubKey(method));
317 if (it != jni_stubs_map_.end()) {
318 JniStubData& data = it->second;
319 if (data.IsCompiled() && ContainsElement(data.GetMethods(), method)) {
320 return data.GetCode();
321 }
322 }
323 return nullptr;
324 }
325
GetSavedEntryPointOfPreCompiledMethod(ArtMethod * method)326 const void* JitCodeCache::GetSavedEntryPointOfPreCompiledMethod(ArtMethod* method) {
327 if (method->IsPreCompiled()) {
328 const void* code_ptr = nullptr;
329 if (method->GetDeclaringClass()->IsBootStrapClassLoaded()) {
330 code_ptr = zygote_map_.GetCodeFor(method);
331 } else {
332 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
333 auto it = saved_compiled_methods_map_.find(method);
334 if (it != saved_compiled_methods_map_.end()) {
335 code_ptr = it->second;
336 }
337 }
338 if (code_ptr != nullptr) {
339 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
340 return method_header->GetEntryPoint();
341 }
342 }
343 return nullptr;
344 }
345
WaitForPotentialCollectionToComplete(Thread * self)346 bool JitCodeCache::WaitForPotentialCollectionToComplete(Thread* self) {
347 bool in_collection = false;
348 while (collection_in_progress_) {
349 in_collection = true;
350 lock_cond_.Wait(self);
351 }
352 return in_collection;
353 }
354
FromCodeToAllocation(const void * code)355 static uintptr_t FromCodeToAllocation(const void* code) {
356 size_t alignment = GetInstructionSetAlignment(kRuntimeISA);
357 return reinterpret_cast<uintptr_t>(code) - RoundUp(sizeof(OatQuickMethodHeader), alignment);
358 }
359
FromAllocationToCode(const uint8_t * alloc)360 static const void* FromAllocationToCode(const uint8_t* alloc) {
361 size_t alignment = GetInstructionSetAlignment(kRuntimeISA);
362 return reinterpret_cast<const void*>(alloc + RoundUp(sizeof(OatQuickMethodHeader), alignment));
363 }
364
GetNumberOfRoots(const uint8_t * stack_map)365 static uint32_t GetNumberOfRoots(const uint8_t* stack_map) {
366 // The length of the table is stored just before the stack map (and therefore at the end of
367 // the table itself), in order to be able to fetch it from a `stack_map` pointer.
368 return reinterpret_cast<const uint32_t*>(stack_map)[-1];
369 }
370
DCheckRootsAreValid(const std::vector<Handle<mirror::Object>> & roots,bool is_shared_region)371 static void DCheckRootsAreValid(const std::vector<Handle<mirror::Object>>& roots,
372 bool is_shared_region)
373 REQUIRES(!Locks::intern_table_lock_) REQUIRES_SHARED(Locks::mutator_lock_) {
374 if (!kIsDebugBuild) {
375 return;
376 }
377 // Put all roots in `roots_data`.
378 for (Handle<mirror::Object> object : roots) {
379 // Ensure the string is strongly interned. b/32995596
380 if (object->IsString()) {
381 ObjPtr<mirror::String> str = object->AsString();
382 ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
383 CHECK(class_linker->GetInternTable()->LookupStrong(Thread::Current(), str) != nullptr);
384 }
385 // Ensure that we don't put movable objects in the shared region.
386 if (is_shared_region) {
387 CHECK(!Runtime::Current()->GetHeap()->IsMovableObject(object.Get()));
388 }
389 }
390 }
391
GetRootTable(const void * code_ptr,uint32_t * number_of_roots=nullptr)392 static const uint8_t* GetRootTable(const void* code_ptr, uint32_t* number_of_roots = nullptr) {
393 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
394 uint8_t* data = method_header->GetOptimizedCodeInfoPtr();
395 uint32_t roots = GetNumberOfRoots(data);
396 if (number_of_roots != nullptr) {
397 *number_of_roots = roots;
398 }
399 return data - ComputeRootTableSize(roots);
400 }
401
SweepRootTables(IsMarkedVisitor * visitor)402 void JitCodeCache::SweepRootTables(IsMarkedVisitor* visitor) {
403 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
404 for (const auto& entry : method_code_map_) {
405 uint32_t number_of_roots = 0;
406 const uint8_t* root_table = GetRootTable(entry.first, &number_of_roots);
407 uint8_t* roots_data = private_region_.IsInDataSpace(root_table)
408 ? private_region_.GetWritableDataAddress(root_table)
409 : shared_region_.GetWritableDataAddress(root_table);
410 GcRoot<mirror::Object>* roots = reinterpret_cast<GcRoot<mirror::Object>*>(roots_data);
411 for (uint32_t i = 0; i < number_of_roots; ++i) {
412 // This does not need a read barrier because this is called by GC.
413 mirror::Object* object = roots[i].Read<kWithoutReadBarrier>();
414 if (object == nullptr || object == Runtime::GetWeakClassSentinel()) {
415 // entry got deleted in a previous sweep.
416 } else if (object->IsString<kDefaultVerifyFlags>()) {
417 mirror::Object* new_object = visitor->IsMarked(object);
418 // We know the string is marked because it's a strongly-interned string that
419 // is always alive. The IsMarked implementation of the CMS collector returns
420 // null for newly allocated objects, but we know those haven't moved. Therefore,
421 // only update the entry if we get a different non-null string.
422 // TODO: Do not use IsMarked for j.l.Class, and adjust once we move this method
423 // out of the weak access/creation pause. b/32167580
424 if (new_object != nullptr && new_object != object) {
425 DCHECK(new_object->IsString());
426 roots[i] = GcRoot<mirror::Object>(new_object);
427 }
428 } else {
429 Runtime::ProcessWeakClass(
430 reinterpret_cast<GcRoot<mirror::Class>*>(&roots[i]),
431 visitor,
432 Runtime::GetWeakClassSentinel());
433 }
434 }
435 }
436 // Walk over inline caches to clear entries containing unloaded classes.
437 for (auto it : profiling_infos_) {
438 ProfilingInfo* info = it.second;
439 for (size_t i = 0; i < info->number_of_inline_caches_; ++i) {
440 InlineCache* cache = &info->cache_[i];
441 for (size_t j = 0; j < InlineCache::kIndividualCacheSize; ++j) {
442 Runtime::ProcessWeakClass(&cache->classes_[j], visitor, nullptr);
443 }
444 }
445 }
446 }
447
FreeCodeAndData(const void * code_ptr)448 void JitCodeCache::FreeCodeAndData(const void* code_ptr) {
449 if (IsInZygoteExecSpace(code_ptr)) {
450 // No need to free, this is shared memory.
451 return;
452 }
453 uintptr_t allocation = FromCodeToAllocation(code_ptr);
454 const uint8_t* data = nullptr;
455 if (OatQuickMethodHeader::FromCodePointer(code_ptr)->IsOptimized()) {
456 data = GetRootTable(code_ptr);
457 } // else this is a JNI stub without any data.
458
459 FreeLocked(&private_region_, reinterpret_cast<uint8_t*>(allocation), data);
460 }
461
FreeAllMethodHeaders(const std::unordered_set<OatQuickMethodHeader * > & method_headers)462 void JitCodeCache::FreeAllMethodHeaders(
463 const std::unordered_set<OatQuickMethodHeader*>& method_headers) {
464 // We need to remove entries in method_headers from CHA dependencies
465 // first since once we do FreeCode() below, the memory can be reused
466 // so it's possible for the same method_header to start representing
467 // different compile code.
468 {
469 MutexLock mu2(Thread::Current(), *Locks::cha_lock_);
470 Runtime::Current()->GetClassLinker()->GetClassHierarchyAnalysis()
471 ->RemoveDependentsWithMethodHeaders(method_headers);
472 }
473
474 ScopedCodeCacheWrite scc(private_region_);
475 for (const OatQuickMethodHeader* method_header : method_headers) {
476 FreeCodeAndData(method_header->GetCode());
477 }
478
479 // We have potentially removed a lot of debug info. Do maintenance pass to save space.
480 RepackNativeDebugInfoForJit();
481
482 // Check that the set of compiled methods exactly matches native debug information.
483 // Does not check zygote methods since they can change concurrently.
484 if (kIsDebugBuild && !Runtime::Current()->IsZygote()) {
485 std::map<const void*, ArtMethod*> compiled_methods;
486 VisitAllMethods([&](const void* addr, ArtMethod* method) {
487 if (!IsInZygoteExecSpace(addr)) {
488 CHECK(addr != nullptr && method != nullptr);
489 compiled_methods.emplace(addr, method);
490 }
491 });
492 std::set<const void*> debug_info;
493 ForEachNativeDebugSymbol([&](const void* addr, size_t, const char* name) {
494 addr = AlignDown(addr, GetInstructionSetInstructionAlignment(kRuntimeISA)); // Thumb-bit.
495 CHECK(debug_info.emplace(addr).second) << "Duplicate debug info: " << addr << " " << name;
496 CHECK_EQ(compiled_methods.count(addr), 1u) << "Extra debug info: " << addr << " " << name;
497 });
498 if (!debug_info.empty()) { // If debug-info generation is enabled.
499 for (auto it : compiled_methods) {
500 CHECK_EQ(debug_info.count(it.first), 1u) << "No debug info: " << it.second->PrettyMethod();
501 }
502 CHECK_EQ(compiled_methods.size(), debug_info.size());
503 }
504 }
505 }
506
RemoveMethodsIn(Thread * self,const LinearAlloc & alloc)507 void JitCodeCache::RemoveMethodsIn(Thread* self, const LinearAlloc& alloc) {
508 ScopedTrace trace(__PRETTY_FUNCTION__);
509 // We use a set to first collect all method_headers whose code need to be
510 // removed. We need to free the underlying code after we remove CHA dependencies
511 // for entries in this set. And it's more efficient to iterate through
512 // the CHA dependency map just once with an unordered_set.
513 std::unordered_set<OatQuickMethodHeader*> method_headers;
514 {
515 MutexLock mu(self, *Locks::jit_lock_);
516 // We do not check if a code cache GC is in progress, as this method comes
517 // with the classlinker_classes_lock_ held, and suspending ourselves could
518 // lead to a deadlock.
519 {
520 for (auto it = jni_stubs_map_.begin(); it != jni_stubs_map_.end();) {
521 it->second.RemoveMethodsIn(alloc);
522 if (it->second.GetMethods().empty()) {
523 method_headers.insert(OatQuickMethodHeader::FromCodePointer(it->second.GetCode()));
524 it = jni_stubs_map_.erase(it);
525 } else {
526 it->first.UpdateShorty(it->second.GetMethods().front());
527 ++it;
528 }
529 }
530 for (auto it = method_code_map_.begin(); it != method_code_map_.end();) {
531 if (alloc.ContainsUnsafe(it->second)) {
532 method_headers.insert(OatQuickMethodHeader::FromCodePointer(it->first));
533 VLOG(jit) << "JIT removed " << it->second->PrettyMethod() << ": " << it->first;
534 it = method_code_map_.erase(it);
535 } else {
536 ++it;
537 }
538 }
539 }
540 for (auto it = osr_code_map_.begin(); it != osr_code_map_.end();) {
541 if (alloc.ContainsUnsafe(it->first)) {
542 // Note that the code has already been pushed to method_headers in the loop
543 // above and is going to be removed in FreeCode() below.
544 it = osr_code_map_.erase(it);
545 } else {
546 ++it;
547 }
548 }
549 for (auto it = profiling_infos_.begin(); it != profiling_infos_.end();) {
550 ProfilingInfo* info = it->second;
551 if (alloc.ContainsUnsafe(info->GetMethod())) {
552 private_region_.FreeWritableData(reinterpret_cast<uint8_t*>(info));
553 it = profiling_infos_.erase(it);
554 } else {
555 ++it;
556 }
557 }
558 FreeAllMethodHeaders(method_headers);
559 }
560 }
561
IsWeakAccessEnabled(Thread * self) const562 bool JitCodeCache::IsWeakAccessEnabled(Thread* self) const {
563 return kUseReadBarrier
564 ? self->GetWeakRefAccessEnabled()
565 : is_weak_access_enabled_.load(std::memory_order_seq_cst);
566 }
567
WaitUntilInlineCacheAccessible(Thread * self)568 void JitCodeCache::WaitUntilInlineCacheAccessible(Thread* self) {
569 if (IsWeakAccessEnabled(self)) {
570 return;
571 }
572 ScopedThreadSuspension sts(self, ThreadState::kWaitingWeakGcRootRead);
573 MutexLock mu(self, *Locks::jit_lock_);
574 while (!IsWeakAccessEnabled(self)) {
575 inline_cache_cond_.Wait(self);
576 }
577 }
578
BroadcastForInlineCacheAccess()579 void JitCodeCache::BroadcastForInlineCacheAccess() {
580 Thread* self = Thread::Current();
581 MutexLock mu(self, *Locks::jit_lock_);
582 inline_cache_cond_.Broadcast(self);
583 }
584
AllowInlineCacheAccess()585 void JitCodeCache::AllowInlineCacheAccess() {
586 DCHECK(!kUseReadBarrier);
587 is_weak_access_enabled_.store(true, std::memory_order_seq_cst);
588 BroadcastForInlineCacheAccess();
589 }
590
DisallowInlineCacheAccess()591 void JitCodeCache::DisallowInlineCacheAccess() {
592 DCHECK(!kUseReadBarrier);
593 is_weak_access_enabled_.store(false, std::memory_order_seq_cst);
594 }
595
CopyInlineCacheInto(const InlineCache & ic,StackHandleScope<InlineCache::kIndividualCacheSize> * classes)596 void JitCodeCache::CopyInlineCacheInto(
597 const InlineCache& ic,
598 /*out*/StackHandleScope<InlineCache::kIndividualCacheSize>* classes) {
599 static_assert(arraysize(ic.classes_) == InlineCache::kIndividualCacheSize);
600 DCHECK_EQ(classes->NumberOfReferences(), InlineCache::kIndividualCacheSize);
601 DCHECK_EQ(classes->RemainingSlots(), InlineCache::kIndividualCacheSize);
602 WaitUntilInlineCacheAccessible(Thread::Current());
603 // Note that we don't need to lock `lock_` here, the compiler calling
604 // this method has already ensured the inline cache will not be deleted.
605 for (const GcRoot<mirror::Class>& root : ic.classes_) {
606 mirror::Class* object = root.Read();
607 if (object != nullptr) {
608 DCHECK_NE(classes->RemainingSlots(), 0u);
609 classes->NewHandle(object);
610 }
611 }
612 }
613
ClearMethodCounter(ArtMethod * method,bool was_warm)614 static void ClearMethodCounter(ArtMethod* method, bool was_warm)
615 REQUIRES_SHARED(Locks::mutator_lock_) {
616 if (was_warm) {
617 method->SetPreviouslyWarm();
618 }
619 method->ResetCounter(Runtime::Current()->GetJITOptions()->GetWarmupThreshold());
620 // We add one sample so that the profile knows that the method was executed at least once.
621 // This is required for layout purposes.
622 method->UpdateCounter(/* new_samples= */ 1);
623 }
624
WaitForPotentialCollectionToCompleteRunnable(Thread * self)625 void JitCodeCache::WaitForPotentialCollectionToCompleteRunnable(Thread* self) {
626 while (collection_in_progress_) {
627 Locks::jit_lock_->Unlock(self);
628 {
629 ScopedThreadSuspension sts(self, ThreadState::kSuspended);
630 MutexLock mu(self, *Locks::jit_lock_);
631 WaitForPotentialCollectionToComplete(self);
632 }
633 Locks::jit_lock_->Lock(self);
634 }
635 }
636
Commit(Thread * self,JitMemoryRegion * region,ArtMethod * method,ArrayRef<const uint8_t> reserved_code,ArrayRef<const uint8_t> code,ArrayRef<const uint8_t> reserved_data,const std::vector<Handle<mirror::Object>> & roots,ArrayRef<const uint8_t> stack_map,const std::vector<uint8_t> & debug_info,bool is_full_debug_info,CompilationKind compilation_kind,bool has_should_deoptimize_flag,const ArenaSet<ArtMethod * > & cha_single_implementation_list)637 bool JitCodeCache::Commit(Thread* self,
638 JitMemoryRegion* region,
639 ArtMethod* method,
640 ArrayRef<const uint8_t> reserved_code,
641 ArrayRef<const uint8_t> code,
642 ArrayRef<const uint8_t> reserved_data,
643 const std::vector<Handle<mirror::Object>>& roots,
644 ArrayRef<const uint8_t> stack_map,
645 const std::vector<uint8_t>& debug_info,
646 bool is_full_debug_info,
647 CompilationKind compilation_kind,
648 bool has_should_deoptimize_flag,
649 const ArenaSet<ArtMethod*>& cha_single_implementation_list) {
650 DCHECK_IMPLIES(method->IsNative(), (compilation_kind != CompilationKind::kOsr));
651
652 if (!method->IsNative()) {
653 // We need to do this before grabbing the lock_ because it needs to be able to see the string
654 // InternTable. Native methods do not have roots.
655 DCheckRootsAreValid(roots, IsSharedRegion(*region));
656 }
657
658 const uint8_t* roots_data = reserved_data.data();
659 size_t root_table_size = ComputeRootTableSize(roots.size());
660 const uint8_t* stack_map_data = roots_data + root_table_size;
661
662 MutexLock mu(self, *Locks::jit_lock_);
663 // We need to make sure that there will be no jit-gcs going on and wait for any ongoing one to
664 // finish.
665 WaitForPotentialCollectionToCompleteRunnable(self);
666 const uint8_t* code_ptr = region->CommitCode(
667 reserved_code, code, stack_map_data, has_should_deoptimize_flag);
668 if (code_ptr == nullptr) {
669 return false;
670 }
671 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
672
673 // Commit roots and stack maps before updating the entry point.
674 if (!region->CommitData(reserved_data, roots, stack_map)) {
675 return false;
676 }
677
678 switch (compilation_kind) {
679 case CompilationKind::kOsr:
680 number_of_osr_compilations_++;
681 break;
682 case CompilationKind::kBaseline:
683 number_of_baseline_compilations_++;
684 break;
685 case CompilationKind::kOptimized:
686 number_of_optimized_compilations_++;
687 break;
688 }
689
690 // We need to update the debug info before the entry point gets set.
691 // At the same time we want to do under JIT lock so that debug info and JIT maps are in sync.
692 if (!debug_info.empty()) {
693 // NB: Don't allow packing of full info since it would remove non-backtrace data.
694 AddNativeDebugInfoForJit(code_ptr, debug_info, /*allow_packing=*/ !is_full_debug_info);
695 }
696
697 // We need to update the entry point in the runnable state for the instrumentation.
698 {
699 // The following needs to be guarded by cha_lock_ also. Otherwise it's possible that the
700 // compiled code is considered invalidated by some class linking, but below we still make the
701 // compiled code valid for the method. Need cha_lock_ for checking all single-implementation
702 // flags and register dependencies.
703 MutexLock cha_mu(self, *Locks::cha_lock_);
704 bool single_impl_still_valid = true;
705 for (ArtMethod* single_impl : cha_single_implementation_list) {
706 if (!single_impl->HasSingleImplementation()) {
707 // Simply discard the compiled code. Clear the counter so that it may be recompiled later.
708 // Hopefully the class hierarchy will be more stable when compilation is retried.
709 single_impl_still_valid = false;
710 ClearMethodCounter(method, /*was_warm=*/ false);
711 break;
712 }
713 }
714
715 // Discard the code if any single-implementation assumptions are now invalid.
716 if (UNLIKELY(!single_impl_still_valid)) {
717 VLOG(jit) << "JIT discarded jitted code due to invalid single-implementation assumptions.";
718 return false;
719 }
720 DCHECK(cha_single_implementation_list.empty() || !Runtime::Current()->IsJavaDebuggable())
721 << "Should not be using cha on debuggable apps/runs!";
722
723 ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
724 for (ArtMethod* single_impl : cha_single_implementation_list) {
725 class_linker->GetClassHierarchyAnalysis()->AddDependency(single_impl, method, method_header);
726 }
727
728 if (UNLIKELY(method->IsNative())) {
729 auto it = jni_stubs_map_.find(JniStubKey(method));
730 DCHECK(it != jni_stubs_map_.end())
731 << "Entry inserted in NotifyCompilationOf() should be alive.";
732 JniStubData* data = &it->second;
733 DCHECK(ContainsElement(data->GetMethods(), method))
734 << "Entry inserted in NotifyCompilationOf() should contain this method.";
735 data->SetCode(code_ptr);
736 data->UpdateEntryPoints(method_header->GetEntryPoint());
737 } else {
738 if (method->IsPreCompiled() && IsSharedRegion(*region)) {
739 zygote_map_.Put(code_ptr, method);
740 } else {
741 method_code_map_.Put(code_ptr, method);
742 }
743 if (compilation_kind == CompilationKind::kOsr) {
744 osr_code_map_.Put(method, code_ptr);
745 } else if (NeedsClinitCheckBeforeCall(method) &&
746 !method->GetDeclaringClass()->IsVisiblyInitialized()) {
747 // This situation currently only occurs in the jit-zygote mode.
748 DCHECK(!garbage_collect_code_);
749 DCHECK(method->IsPreCompiled());
750 // The shared region can easily be queried. For the private region, we
751 // use a side map.
752 if (!IsSharedRegion(*region)) {
753 saved_compiled_methods_map_.Put(method, code_ptr);
754 }
755 } else {
756 Runtime::Current()->GetInstrumentation()->UpdateMethodsCode(
757 method, method_header->GetEntryPoint());
758 }
759 }
760 if (collection_in_progress_) {
761 // We need to update the live bitmap if there is a GC to ensure it sees this new
762 // code.
763 GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(code_ptr));
764 }
765 VLOG(jit)
766 << "JIT added (kind=" << compilation_kind << ") "
767 << ArtMethod::PrettyMethod(method) << "@" << method
768 << " ccache_size=" << PrettySize(CodeCacheSizeLocked()) << ": "
769 << " dcache_size=" << PrettySize(DataCacheSizeLocked()) << ": "
770 << reinterpret_cast<const void*>(method_header->GetEntryPoint()) << ","
771 << reinterpret_cast<const void*>(method_header->GetEntryPoint() +
772 method_header->GetCodeSize());
773 }
774
775 return true;
776 }
777
CodeCacheSize()778 size_t JitCodeCache::CodeCacheSize() {
779 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
780 return CodeCacheSizeLocked();
781 }
782
RemoveMethod(ArtMethod * method,bool release_memory)783 bool JitCodeCache::RemoveMethod(ArtMethod* method, bool release_memory) {
784 // This function is used only for testing and only with non-native methods.
785 CHECK(!method->IsNative());
786
787 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
788
789 bool osr = osr_code_map_.find(method) != osr_code_map_.end();
790 bool in_cache = RemoveMethodLocked(method, release_memory);
791
792 if (!in_cache) {
793 return false;
794 }
795
796 ClearMethodCounter(method, /* was_warm= */ false);
797 Runtime::Current()->GetInstrumentation()->InitializeMethodsCode(method, /*aot_code=*/ nullptr);
798 VLOG(jit)
799 << "JIT removed (osr=" << std::boolalpha << osr << std::noboolalpha << ") "
800 << ArtMethod::PrettyMethod(method) << "@" << method
801 << " ccache_size=" << PrettySize(CodeCacheSizeLocked()) << ": "
802 << " dcache_size=" << PrettySize(DataCacheSizeLocked());
803 return true;
804 }
805
RemoveMethodLocked(ArtMethod * method,bool release_memory)806 bool JitCodeCache::RemoveMethodLocked(ArtMethod* method, bool release_memory) {
807 if (LIKELY(!method->IsNative())) {
808 auto it = profiling_infos_.find(method);
809 if (it != profiling_infos_.end()) {
810 profiling_infos_.erase(it);
811 }
812 }
813
814 bool in_cache = false;
815 ScopedCodeCacheWrite ccw(private_region_);
816 if (UNLIKELY(method->IsNative())) {
817 auto it = jni_stubs_map_.find(JniStubKey(method));
818 if (it != jni_stubs_map_.end() && it->second.RemoveMethod(method)) {
819 in_cache = true;
820 if (it->second.GetMethods().empty()) {
821 if (release_memory) {
822 FreeCodeAndData(it->second.GetCode());
823 }
824 jni_stubs_map_.erase(it);
825 } else {
826 it->first.UpdateShorty(it->second.GetMethods().front());
827 }
828 }
829 } else {
830 for (auto it = method_code_map_.begin(); it != method_code_map_.end();) {
831 if (it->second == method) {
832 in_cache = true;
833 if (release_memory) {
834 FreeCodeAndData(it->first);
835 }
836 VLOG(jit) << "JIT removed " << it->second->PrettyMethod() << ": " << it->first;
837 it = method_code_map_.erase(it);
838 } else {
839 ++it;
840 }
841 }
842
843 auto osr_it = osr_code_map_.find(method);
844 if (osr_it != osr_code_map_.end()) {
845 osr_code_map_.erase(osr_it);
846 }
847 }
848
849 return in_cache;
850 }
851
852 // This notifies the code cache that the given method has been redefined and that it should remove
853 // any cached information it has on the method. All threads must be suspended before calling this
854 // method. The compiled code for the method (if there is any) must not be in any threads call stack.
NotifyMethodRedefined(ArtMethod * method)855 void JitCodeCache::NotifyMethodRedefined(ArtMethod* method) {
856 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
857 RemoveMethodLocked(method, /* release_memory= */ true);
858 }
859
860 // This invalidates old_method. Once this function returns one can no longer use old_method to
861 // execute code unless it is fixed up. This fixup will happen later in the process of installing a
862 // class redefinition.
863 // TODO We should add some info to ArtMethod to note that 'old_method' has been invalidated and
864 // shouldn't be used since it is no longer logically in the jit code cache.
865 // TODO We should add DCHECKS that validate that the JIT is paused when this method is entered.
MoveObsoleteMethod(ArtMethod * old_method,ArtMethod * new_method)866 void JitCodeCache::MoveObsoleteMethod(ArtMethod* old_method, ArtMethod* new_method) {
867 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
868 if (old_method->IsNative()) {
869 // Update methods in jni_stubs_map_.
870 for (auto& entry : jni_stubs_map_) {
871 JniStubData& data = entry.second;
872 data.MoveObsoleteMethod(old_method, new_method);
873 }
874 return;
875 }
876 // Update method_code_map_ to point to the new method.
877 for (auto& it : method_code_map_) {
878 if (it.second == old_method) {
879 it.second = new_method;
880 }
881 }
882 // Update osr_code_map_ to point to the new method.
883 auto code_map = osr_code_map_.find(old_method);
884 if (code_map != osr_code_map_.end()) {
885 osr_code_map_.Put(new_method, code_map->second);
886 osr_code_map_.erase(old_method);
887 }
888 }
889
TransitionToDebuggable()890 void JitCodeCache::TransitionToDebuggable() {
891 // Check that none of our methods have an entrypoint in the zygote exec
892 // space (this should be taken care of by
893 // ClassLinker::UpdateEntryPointsClassVisitor.
894 {
895 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
896 if (kIsDebugBuild) {
897 for (const auto& it : method_code_map_) {
898 ArtMethod* method = it.second;
899 DCHECK(!method->IsPreCompiled());
900 DCHECK(!IsInZygoteExecSpace(method->GetEntryPointFromQuickCompiledCode()));
901 }
902 }
903 // Not strictly necessary, but this map is useless now.
904 saved_compiled_methods_map_.clear();
905 }
906 if (kIsDebugBuild) {
907 for (const auto& entry : zygote_map_) {
908 ArtMethod* method = entry.method;
909 if (method != nullptr) {
910 DCHECK(!method->IsPreCompiled());
911 DCHECK(!IsInZygoteExecSpace(method->GetEntryPointFromQuickCompiledCode()));
912 }
913 }
914 }
915 }
916
CodeCacheSizeLocked()917 size_t JitCodeCache::CodeCacheSizeLocked() {
918 return GetCurrentRegion()->GetUsedMemoryForCode();
919 }
920
DataCacheSize()921 size_t JitCodeCache::DataCacheSize() {
922 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
923 return DataCacheSizeLocked();
924 }
925
DataCacheSizeLocked()926 size_t JitCodeCache::DataCacheSizeLocked() {
927 return GetCurrentRegion()->GetUsedMemoryForData();
928 }
929
Reserve(Thread * self,JitMemoryRegion * region,size_t code_size,size_t stack_map_size,size_t number_of_roots,ArtMethod * method,ArrayRef<const uint8_t> * reserved_code,ArrayRef<const uint8_t> * reserved_data)930 bool JitCodeCache::Reserve(Thread* self,
931 JitMemoryRegion* region,
932 size_t code_size,
933 size_t stack_map_size,
934 size_t number_of_roots,
935 ArtMethod* method,
936 /*out*/ArrayRef<const uint8_t>* reserved_code,
937 /*out*/ArrayRef<const uint8_t>* reserved_data) {
938 code_size = OatQuickMethodHeader::InstructionAlignedSize() + code_size;
939 size_t data_size = RoundUp(ComputeRootTableSize(number_of_roots) + stack_map_size, sizeof(void*));
940
941 const uint8_t* code;
942 const uint8_t* data;
943 while (true) {
944 bool at_max_capacity = false;
945 {
946 ScopedThreadSuspension sts(self, ThreadState::kSuspended);
947 MutexLock mu(self, *Locks::jit_lock_);
948 WaitForPotentialCollectionToComplete(self);
949 ScopedCodeCacheWrite ccw(*region);
950 code = region->AllocateCode(code_size);
951 data = region->AllocateData(data_size);
952 at_max_capacity = IsAtMaxCapacity();
953 }
954 if (code != nullptr && data != nullptr) {
955 break;
956 }
957 Free(self, region, code, data);
958 if (at_max_capacity) {
959 VLOG(jit) << "JIT failed to allocate code of size "
960 << PrettySize(code_size)
961 << ", and data of size "
962 << PrettySize(data_size);
963 return false;
964 }
965 // Run a code cache collection and try again.
966 GarbageCollectCache(self);
967 }
968
969 *reserved_code = ArrayRef<const uint8_t>(code, code_size);
970 *reserved_data = ArrayRef<const uint8_t>(data, data_size);
971
972 MutexLock mu(self, *Locks::jit_lock_);
973 histogram_code_memory_use_.AddValue(code_size);
974 if (code_size > kCodeSizeLogThreshold) {
975 LOG(INFO) << "JIT allocated "
976 << PrettySize(code_size)
977 << " for compiled code of "
978 << ArtMethod::PrettyMethod(method);
979 }
980 histogram_stack_map_memory_use_.AddValue(data_size);
981 if (data_size > kStackMapSizeLogThreshold) {
982 LOG(INFO) << "JIT allocated "
983 << PrettySize(data_size)
984 << " for stack maps of "
985 << ArtMethod::PrettyMethod(method);
986 }
987 return true;
988 }
989
Free(Thread * self,JitMemoryRegion * region,const uint8_t * code,const uint8_t * data)990 void JitCodeCache::Free(Thread* self,
991 JitMemoryRegion* region,
992 const uint8_t* code,
993 const uint8_t* data) {
994 MutexLock mu(self, *Locks::jit_lock_);
995 ScopedCodeCacheWrite ccw(*region);
996 FreeLocked(region, code, data);
997 }
998
FreeLocked(JitMemoryRegion * region,const uint8_t * code,const uint8_t * data)999 void JitCodeCache::FreeLocked(JitMemoryRegion* region, const uint8_t* code, const uint8_t* data) {
1000 if (code != nullptr) {
1001 RemoveNativeDebugInfoForJit(reinterpret_cast<const void*>(FromAllocationToCode(code)));
1002 region->FreeCode(code);
1003 }
1004 if (data != nullptr) {
1005 region->FreeData(data);
1006 }
1007 }
1008
1009 class MarkCodeClosure final : public Closure {
1010 public:
MarkCodeClosure(JitCodeCache * code_cache,CodeCacheBitmap * bitmap,Barrier * barrier)1011 MarkCodeClosure(JitCodeCache* code_cache, CodeCacheBitmap* bitmap, Barrier* barrier)
1012 : code_cache_(code_cache), bitmap_(bitmap), barrier_(barrier) {}
1013
Run(Thread * thread)1014 void Run(Thread* thread) override REQUIRES_SHARED(Locks::mutator_lock_) {
1015 ScopedTrace trace(__PRETTY_FUNCTION__);
1016 DCHECK(thread == Thread::Current() || thread->IsSuspended());
1017 StackVisitor::WalkStack(
1018 [&](const art::StackVisitor* stack_visitor) {
1019 const OatQuickMethodHeader* method_header =
1020 stack_visitor->GetCurrentOatQuickMethodHeader();
1021 if (method_header == nullptr) {
1022 return true;
1023 }
1024 const void* code = method_header->GetCode();
1025 if (code_cache_->ContainsPc(code) && !code_cache_->IsInZygoteExecSpace(code)) {
1026 // Use the atomic set version, as multiple threads are executing this code.
1027 bitmap_->AtomicTestAndSet(FromCodeToAllocation(code));
1028 }
1029 return true;
1030 },
1031 thread,
1032 /* context= */ nullptr,
1033 art::StackVisitor::StackWalkKind::kSkipInlinedFrames);
1034
1035 if (kIsDebugBuild) {
1036 // The stack walking code queries the side instrumentation stack if it
1037 // sees an instrumentation exit pc, so the JIT code of methods in that stack
1038 // must have been seen. We check this below.
1039 for (const auto& it : *thread->GetInstrumentationStack()) {
1040 // The 'method_' in InstrumentationStackFrame is the one that has return_pc_ in
1041 // its stack frame, it is not the method owning return_pc_. We just pass null to
1042 // LookupMethodHeader: the method is only checked against in debug builds.
1043 OatQuickMethodHeader* method_header =
1044 code_cache_->LookupMethodHeader(it.second.return_pc_, /* method= */ nullptr);
1045 if (method_header != nullptr) {
1046 const void* code = method_header->GetCode();
1047 CHECK(bitmap_->Test(FromCodeToAllocation(code)));
1048 }
1049 }
1050 }
1051 barrier_->Pass(Thread::Current());
1052 }
1053
1054 private:
1055 JitCodeCache* const code_cache_;
1056 CodeCacheBitmap* const bitmap_;
1057 Barrier* const barrier_;
1058 };
1059
NotifyCollectionDone(Thread * self)1060 void JitCodeCache::NotifyCollectionDone(Thread* self) {
1061 collection_in_progress_ = false;
1062 lock_cond_.Broadcast(self);
1063 }
1064
MarkCompiledCodeOnThreadStacks(Thread * self)1065 void JitCodeCache::MarkCompiledCodeOnThreadStacks(Thread* self) {
1066 Barrier barrier(0);
1067 size_t threads_running_checkpoint = 0;
1068 MarkCodeClosure closure(this, GetLiveBitmap(), &barrier);
1069 threads_running_checkpoint = Runtime::Current()->GetThreadList()->RunCheckpoint(&closure);
1070 // Now that we have run our checkpoint, move to a suspended state and wait
1071 // for other threads to run the checkpoint.
1072 ScopedThreadSuspension sts(self, ThreadState::kSuspended);
1073 if (threads_running_checkpoint != 0) {
1074 barrier.Increment(self, threads_running_checkpoint);
1075 }
1076 }
1077
IsAtMaxCapacity() const1078 bool JitCodeCache::IsAtMaxCapacity() const {
1079 return private_region_.GetCurrentCapacity() == private_region_.GetMaxCapacity();
1080 }
1081
ShouldDoFullCollection()1082 bool JitCodeCache::ShouldDoFullCollection() {
1083 if (IsAtMaxCapacity()) {
1084 // Always do a full collection when the code cache is full.
1085 return true;
1086 } else if (private_region_.GetCurrentCapacity() < kReservedCapacity) {
1087 // Always do partial collection when the code cache size is below the reserved
1088 // capacity.
1089 return false;
1090 } else if (last_collection_increased_code_cache_) {
1091 // This time do a full collection.
1092 return true;
1093 } else {
1094 // This time do a partial collection.
1095 return false;
1096 }
1097 }
1098
GarbageCollectCache(Thread * self)1099 void JitCodeCache::GarbageCollectCache(Thread* self) {
1100 ScopedTrace trace(__FUNCTION__);
1101 // Wait for an existing collection, or let everyone know we are starting one.
1102 {
1103 ScopedThreadSuspension sts(self, ThreadState::kSuspended);
1104 MutexLock mu(self, *Locks::jit_lock_);
1105 if (!garbage_collect_code_) {
1106 private_region_.IncreaseCodeCacheCapacity();
1107 return;
1108 } else if (WaitForPotentialCollectionToComplete(self)) {
1109 return;
1110 } else {
1111 number_of_collections_++;
1112 live_bitmap_.reset(CodeCacheBitmap::Create(
1113 "code-cache-bitmap",
1114 reinterpret_cast<uintptr_t>(private_region_.GetExecPages()->Begin()),
1115 reinterpret_cast<uintptr_t>(
1116 private_region_.GetExecPages()->Begin() + private_region_.GetCurrentCapacity() / 2)));
1117 collection_in_progress_ = true;
1118 }
1119 }
1120
1121 TimingLogger logger("JIT code cache timing logger", true, VLOG_IS_ON(jit));
1122 {
1123 TimingLogger::ScopedTiming st("Code cache collection", &logger);
1124
1125 bool do_full_collection = false;
1126 {
1127 MutexLock mu(self, *Locks::jit_lock_);
1128 do_full_collection = ShouldDoFullCollection();
1129 }
1130
1131 VLOG(jit) << "Do "
1132 << (do_full_collection ? "full" : "partial")
1133 << " code cache collection, code="
1134 << PrettySize(CodeCacheSize())
1135 << ", data=" << PrettySize(DataCacheSize());
1136
1137 DoCollection(self, /* collect_profiling_info= */ do_full_collection);
1138
1139 VLOG(jit) << "After code cache collection, code="
1140 << PrettySize(CodeCacheSize())
1141 << ", data=" << PrettySize(DataCacheSize());
1142
1143 {
1144 MutexLock mu(self, *Locks::jit_lock_);
1145
1146 // Increase the code cache only when we do partial collections.
1147 // TODO: base this strategy on how full the code cache is?
1148 if (do_full_collection) {
1149 last_collection_increased_code_cache_ = false;
1150 } else {
1151 last_collection_increased_code_cache_ = true;
1152 private_region_.IncreaseCodeCacheCapacity();
1153 }
1154
1155 bool next_collection_will_be_full = ShouldDoFullCollection();
1156
1157 // Start polling the liveness of compiled code to prepare for the next full collection.
1158 if (next_collection_will_be_full) {
1159 for (auto it : profiling_infos_) {
1160 it.second->ResetCounter();
1161 }
1162
1163 // Change entry points of native methods back to the GenericJNI entrypoint.
1164 for (const auto& entry : jni_stubs_map_) {
1165 const JniStubData& data = entry.second;
1166 if (!data.IsCompiled() || IsInZygoteExecSpace(data.GetCode())) {
1167 continue;
1168 }
1169 const OatQuickMethodHeader* method_header =
1170 OatQuickMethodHeader::FromCodePointer(data.GetCode());
1171 for (ArtMethod* method : data.GetMethods()) {
1172 if (method->GetEntryPointFromQuickCompiledCode() == method_header->GetEntryPoint()) {
1173 // Don't call Instrumentation::UpdateMethodsCode(), same as for normal methods above.
1174 // Make sure a single invocation of the GenericJNI trampoline tries to recompile.
1175 method->SetHotCounter();
1176 method->SetEntryPointFromQuickCompiledCode(GetQuickGenericJniStub());
1177 }
1178 }
1179 }
1180 }
1181 live_bitmap_.reset(nullptr);
1182 NotifyCollectionDone(self);
1183 }
1184 }
1185 Runtime::Current()->GetJit()->AddTimingLogger(logger);
1186 }
1187
RemoveUnmarkedCode(Thread * self)1188 void JitCodeCache::RemoveUnmarkedCode(Thread* self) {
1189 ScopedTrace trace(__FUNCTION__);
1190 std::unordered_set<OatQuickMethodHeader*> method_headers;
1191 {
1192 MutexLock mu(self, *Locks::jit_lock_);
1193 // Iterate over all compiled code and remove entries that are not marked.
1194 for (auto it = jni_stubs_map_.begin(); it != jni_stubs_map_.end();) {
1195 JniStubData* data = &it->second;
1196 if (IsInZygoteExecSpace(data->GetCode()) ||
1197 !data->IsCompiled() ||
1198 GetLiveBitmap()->Test(FromCodeToAllocation(data->GetCode()))) {
1199 ++it;
1200 } else {
1201 method_headers.insert(OatQuickMethodHeader::FromCodePointer(data->GetCode()));
1202 for (ArtMethod* method : data->GetMethods()) {
1203 VLOG(jit) << "JIT removed (JNI) " << method->PrettyMethod() << ": " << data->GetCode();
1204 }
1205 it = jni_stubs_map_.erase(it);
1206 }
1207 }
1208 for (auto it = method_code_map_.begin(); it != method_code_map_.end();) {
1209 const void* code_ptr = it->first;
1210 uintptr_t allocation = FromCodeToAllocation(code_ptr);
1211 if (IsInZygoteExecSpace(code_ptr) || GetLiveBitmap()->Test(allocation)) {
1212 ++it;
1213 } else {
1214 OatQuickMethodHeader* header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1215 method_headers.insert(header);
1216 VLOG(jit) << "JIT removed " << it->second->PrettyMethod() << ": " << it->first;
1217 it = method_code_map_.erase(it);
1218 }
1219 }
1220 FreeAllMethodHeaders(method_headers);
1221 }
1222 }
1223
GetGarbageCollectCode()1224 bool JitCodeCache::GetGarbageCollectCode() {
1225 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1226 return garbage_collect_code_;
1227 }
1228
SetGarbageCollectCode(bool value)1229 void JitCodeCache::SetGarbageCollectCode(bool value) {
1230 Thread* self = Thread::Current();
1231 MutexLock mu(self, *Locks::jit_lock_);
1232 // Update the flag while holding the lock to ensure no thread will try to GC.
1233 garbage_collect_code_ = value;
1234 }
1235
RemoveMethodBeingCompiled(ArtMethod * method,CompilationKind kind)1236 void JitCodeCache::RemoveMethodBeingCompiled(ArtMethod* method, CompilationKind kind) {
1237 DCHECK(IsMethodBeingCompiled(method, kind));
1238 switch (kind) {
1239 case CompilationKind::kOsr:
1240 current_osr_compilations_.erase(method);
1241 break;
1242 case CompilationKind::kBaseline:
1243 current_baseline_compilations_.erase(method);
1244 break;
1245 case CompilationKind::kOptimized:
1246 current_optimized_compilations_.erase(method);
1247 break;
1248 }
1249 }
1250
AddMethodBeingCompiled(ArtMethod * method,CompilationKind kind)1251 void JitCodeCache::AddMethodBeingCompiled(ArtMethod* method, CompilationKind kind) {
1252 DCHECK(!IsMethodBeingCompiled(method, kind));
1253 switch (kind) {
1254 case CompilationKind::kOsr:
1255 current_osr_compilations_.insert(method);
1256 break;
1257 case CompilationKind::kBaseline:
1258 current_baseline_compilations_.insert(method);
1259 break;
1260 case CompilationKind::kOptimized:
1261 current_optimized_compilations_.insert(method);
1262 break;
1263 }
1264 }
1265
IsMethodBeingCompiled(ArtMethod * method,CompilationKind kind)1266 bool JitCodeCache::IsMethodBeingCompiled(ArtMethod* method, CompilationKind kind) {
1267 switch (kind) {
1268 case CompilationKind::kOsr:
1269 return ContainsElement(current_osr_compilations_, method);
1270 case CompilationKind::kBaseline:
1271 return ContainsElement(current_baseline_compilations_, method);
1272 case CompilationKind::kOptimized:
1273 return ContainsElement(current_optimized_compilations_, method);
1274 }
1275 }
1276
IsMethodBeingCompiled(ArtMethod * method)1277 bool JitCodeCache::IsMethodBeingCompiled(ArtMethod* method) {
1278 return ContainsElement(current_optimized_compilations_, method) ||
1279 ContainsElement(current_osr_compilations_, method) ||
1280 ContainsElement(current_baseline_compilations_, method);
1281 }
1282
GetProfilingInfo(ArtMethod * method,Thread * self)1283 ProfilingInfo* JitCodeCache::GetProfilingInfo(ArtMethod* method, Thread* self) {
1284 MutexLock mu(self, *Locks::jit_lock_);
1285 DCHECK(IsMethodBeingCompiled(method))
1286 << "GetProfilingInfo should only be called when the method is being compiled";
1287 auto it = profiling_infos_.find(method);
1288 if (it == profiling_infos_.end()) {
1289 return nullptr;
1290 }
1291 return it->second;
1292 }
1293
ResetHotnessCounter(ArtMethod * method,Thread * self)1294 void JitCodeCache::ResetHotnessCounter(ArtMethod* method, Thread* self) {
1295 MutexLock mu(self, *Locks::jit_lock_);
1296 auto it = profiling_infos_.find(method);
1297 DCHECK(it != profiling_infos_.end());
1298 it->second->ResetCounter();
1299 }
1300
1301
DoCollection(Thread * self,bool collect_profiling_info)1302 void JitCodeCache::DoCollection(Thread* self, bool collect_profiling_info) {
1303 ScopedTrace trace(__FUNCTION__);
1304 {
1305 MutexLock mu(self, *Locks::jit_lock_);
1306
1307 // Update to interpreter the methods that have baseline entrypoints and whose baseline
1308 // hotness count hasn't changed.
1309 // Note that these methods may be in thread stack or concurrently revived
1310 // between. That's OK, as the thread executing it will mark it.
1311 uint16_t warmup_threshold = Runtime::Current()->GetJITOptions()->GetWarmupThreshold();
1312 for (auto it : profiling_infos_) {
1313 ProfilingInfo* info = it.second;
1314 if (!info->CounterHasChanged()) {
1315 const void* entry_point = info->GetMethod()->GetEntryPointFromQuickCompiledCode();
1316 if (ContainsPc(entry_point)) {
1317 OatQuickMethodHeader* method_header =
1318 OatQuickMethodHeader::FromEntryPoint(entry_point);
1319 if (CodeInfo::IsBaseline(method_header->GetOptimizedCodeInfoPtr())) {
1320 info->GetMethod()->ResetCounter(warmup_threshold);
1321 Runtime::Current()->GetInstrumentation()->InitializeMethodsCode(
1322 info->GetMethod(), /*aot_code=*/ nullptr);
1323 }
1324 }
1325 }
1326 }
1327 // TODO: collect profiling info
1328 // TODO: collect optimized code
1329
1330 // Mark compiled code that are entrypoints of ArtMethods. Compiled code that is not
1331 // an entry point is either:
1332 // - an osr compiled code, that will be removed if not in a thread call stack.
1333 // - discarded compiled code, that will be removed if not in a thread call stack.
1334 for (const auto& entry : jni_stubs_map_) {
1335 const JniStubData& data = entry.second;
1336 const void* code_ptr = data.GetCode();
1337 if (IsInZygoteExecSpace(code_ptr)) {
1338 continue;
1339 }
1340 const OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1341 for (ArtMethod* method : data.GetMethods()) {
1342 if (method_header->GetEntryPoint() == method->GetEntryPointFromQuickCompiledCode()) {
1343 GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(code_ptr));
1344 break;
1345 }
1346 }
1347 }
1348 for (const auto& it : method_code_map_) {
1349 ArtMethod* method = it.second;
1350 const void* code_ptr = it.first;
1351 if (IsInZygoteExecSpace(code_ptr)) {
1352 continue;
1353 }
1354 const OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1355 if (method_header->GetEntryPoint() == method->GetEntryPointFromQuickCompiledCode()) {
1356 GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(code_ptr));
1357 }
1358 }
1359
1360 // Empty osr method map, as osr compiled code will be deleted (except the ones
1361 // on thread stacks).
1362 osr_code_map_.clear();
1363 }
1364
1365 // Run a checkpoint on all threads to mark the JIT compiled code they are running.
1366 MarkCompiledCodeOnThreadStacks(self);
1367
1368 // At this point, mutator threads are still running, and entrypoints of methods can
1369 // change. We do know they cannot change to a code cache entry that is not marked,
1370 // therefore we can safely remove those entries.
1371 RemoveUnmarkedCode(self);
1372
1373 if (collect_profiling_info) {
1374 // TODO: Collect unused profiling infos.
1375 }
1376 }
1377
LookupMethodHeader(uintptr_t pc,ArtMethod * method)1378 OatQuickMethodHeader* JitCodeCache::LookupMethodHeader(uintptr_t pc, ArtMethod* method) {
1379 static_assert(kRuntimeISA != InstructionSet::kThumb2, "kThumb2 cannot be a runtime ISA");
1380 if (kRuntimeISA == InstructionSet::kArm) {
1381 // On Thumb-2, the pc is offset by one.
1382 --pc;
1383 }
1384 if (!ContainsPc(reinterpret_cast<const void*>(pc))) {
1385 return nullptr;
1386 }
1387
1388 if (!kIsDebugBuild) {
1389 // Called with null `method` only from MarkCodeClosure::Run() in debug build.
1390 CHECK(method != nullptr);
1391 }
1392
1393 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1394 OatQuickMethodHeader* method_header = nullptr;
1395 ArtMethod* found_method = nullptr; // Only for DCHECK(), not for JNI stubs.
1396 if (method != nullptr && UNLIKELY(method->IsNative())) {
1397 auto it = jni_stubs_map_.find(JniStubKey(method));
1398 if (it == jni_stubs_map_.end() || !ContainsElement(it->second.GetMethods(), method)) {
1399 return nullptr;
1400 }
1401 const void* code_ptr = it->second.GetCode();
1402 method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1403 if (!method_header->Contains(pc)) {
1404 return nullptr;
1405 }
1406 } else {
1407 if (shared_region_.IsInExecSpace(reinterpret_cast<const void*>(pc))) {
1408 const void* code_ptr = zygote_map_.GetCodeFor(method, pc);
1409 if (code_ptr != nullptr) {
1410 return OatQuickMethodHeader::FromCodePointer(code_ptr);
1411 }
1412 }
1413 auto it = method_code_map_.lower_bound(reinterpret_cast<const void*>(pc));
1414 if (it != method_code_map_.begin()) {
1415 --it;
1416 const void* code_ptr = it->first;
1417 if (OatQuickMethodHeader::FromCodePointer(code_ptr)->Contains(pc)) {
1418 method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1419 found_method = it->second;
1420 }
1421 }
1422 if (method_header == nullptr && method == nullptr) {
1423 // Scan all compiled JNI stubs as well. This slow search is used only
1424 // for checks in debug build, for release builds the `method` is not null.
1425 for (auto&& entry : jni_stubs_map_) {
1426 const JniStubData& data = entry.second;
1427 if (data.IsCompiled() &&
1428 OatQuickMethodHeader::FromCodePointer(data.GetCode())->Contains(pc)) {
1429 method_header = OatQuickMethodHeader::FromCodePointer(data.GetCode());
1430 }
1431 }
1432 }
1433 if (method_header == nullptr) {
1434 return nullptr;
1435 }
1436 }
1437
1438 if (kIsDebugBuild && method != nullptr && !method->IsNative()) {
1439 DCHECK_EQ(found_method, method)
1440 << ArtMethod::PrettyMethod(method) << " "
1441 << ArtMethod::PrettyMethod(found_method) << " "
1442 << std::hex << pc;
1443 }
1444 return method_header;
1445 }
1446
LookupOsrMethodHeader(ArtMethod * method)1447 OatQuickMethodHeader* JitCodeCache::LookupOsrMethodHeader(ArtMethod* method) {
1448 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1449 auto it = osr_code_map_.find(method);
1450 if (it == osr_code_map_.end()) {
1451 return nullptr;
1452 }
1453 return OatQuickMethodHeader::FromCodePointer(it->second);
1454 }
1455
AddProfilingInfo(Thread * self,ArtMethod * method,const std::vector<uint32_t> & entries)1456 ProfilingInfo* JitCodeCache::AddProfilingInfo(Thread* self,
1457 ArtMethod* method,
1458 const std::vector<uint32_t>& entries) {
1459 DCHECK(CanAllocateProfilingInfo());
1460 ProfilingInfo* info = nullptr;
1461 {
1462 MutexLock mu(self, *Locks::jit_lock_);
1463 info = AddProfilingInfoInternal(self, method, entries);
1464 }
1465
1466 if (info == nullptr) {
1467 GarbageCollectCache(self);
1468 MutexLock mu(self, *Locks::jit_lock_);
1469 info = AddProfilingInfoInternal(self, method, entries);
1470 }
1471 return info;
1472 }
1473
AddProfilingInfoInternal(Thread * self ATTRIBUTE_UNUSED,ArtMethod * method,const std::vector<uint32_t> & entries)1474 ProfilingInfo* JitCodeCache::AddProfilingInfoInternal(Thread* self ATTRIBUTE_UNUSED,
1475 ArtMethod* method,
1476 const std::vector<uint32_t>& entries) {
1477 // Check whether some other thread has concurrently created it.
1478 auto it = profiling_infos_.find(method);
1479 if (it != profiling_infos_.end()) {
1480 return it->second;
1481 }
1482
1483 size_t profile_info_size = RoundUp(
1484 sizeof(ProfilingInfo) + sizeof(InlineCache) * entries.size(),
1485 sizeof(void*));
1486
1487 const uint8_t* data = private_region_.AllocateData(profile_info_size);
1488 if (data == nullptr) {
1489 return nullptr;
1490 }
1491 uint8_t* writable_data = private_region_.GetWritableDataAddress(data);
1492 ProfilingInfo* info = new (writable_data) ProfilingInfo(method, entries);
1493
1494 profiling_infos_.Put(method, info);
1495 histogram_profiling_info_memory_use_.AddValue(profile_info_size);
1496 return info;
1497 }
1498
MoreCore(const void * mspace,intptr_t increment)1499 void* JitCodeCache::MoreCore(const void* mspace, intptr_t increment) {
1500 return shared_region_.OwnsSpace(mspace)
1501 ? shared_region_.MoreCore(mspace, increment)
1502 : private_region_.MoreCore(mspace, increment);
1503 }
1504
GetProfiledMethods(const std::set<std::string> & dex_base_locations,std::vector<ProfileMethodInfo> & methods)1505 void JitCodeCache::GetProfiledMethods(const std::set<std::string>& dex_base_locations,
1506 std::vector<ProfileMethodInfo>& methods) {
1507 Thread* self = Thread::Current();
1508 WaitUntilInlineCacheAccessible(self);
1509 MutexLock mu(self, *Locks::jit_lock_);
1510 ScopedTrace trace(__FUNCTION__);
1511 for (auto it : profiling_infos_) {
1512 ProfilingInfo* info = it.second;
1513 ArtMethod* method = info->GetMethod();
1514 const DexFile* dex_file = method->GetDexFile();
1515 const std::string base_location = DexFileLoader::GetBaseLocation(dex_file->GetLocation());
1516 if (!ContainsElement(dex_base_locations, base_location)) {
1517 // Skip dex files which are not profiled.
1518 continue;
1519 }
1520 std::vector<ProfileMethodInfo::ProfileInlineCache> inline_caches;
1521
1522 // If the method is still baseline compiled, don't save the inline caches.
1523 // They might be incomplete and cause unnecessary deoptimizations.
1524 // If the inline cache is empty the compiler will generate a regular invoke virtual/interface.
1525 const void* entry_point = method->GetEntryPointFromQuickCompiledCode();
1526 if (ContainsPc(entry_point) &&
1527 CodeInfo::IsBaseline(
1528 OatQuickMethodHeader::FromEntryPoint(entry_point)->GetOptimizedCodeInfoPtr())) {
1529 methods.emplace_back(/*ProfileMethodInfo*/
1530 MethodReference(dex_file, method->GetDexMethodIndex()), inline_caches);
1531 continue;
1532 }
1533
1534 for (size_t i = 0; i < info->number_of_inline_caches_; ++i) {
1535 std::vector<TypeReference> profile_classes;
1536 const InlineCache& cache = info->cache_[i];
1537 ArtMethod* caller = info->GetMethod();
1538 bool is_missing_types = false;
1539 for (size_t k = 0; k < InlineCache::kIndividualCacheSize; k++) {
1540 mirror::Class* cls = cache.classes_[k].Read();
1541 if (cls == nullptr) {
1542 break;
1543 }
1544
1545 // Check if the receiver is in the boot class path or if it's in the
1546 // same class loader as the caller. If not, skip it, as there is not
1547 // much we can do during AOT.
1548 if (!cls->IsBootStrapClassLoaded() &&
1549 caller->GetClassLoader() != cls->GetClassLoader()) {
1550 is_missing_types = true;
1551 continue;
1552 }
1553
1554 const DexFile* class_dex_file = nullptr;
1555 dex::TypeIndex type_index;
1556
1557 if (cls->GetDexCache() == nullptr) {
1558 DCHECK(cls->IsArrayClass()) << cls->PrettyClass();
1559 // Make a best effort to find the type index in the method's dex file.
1560 // We could search all open dex files but that might turn expensive
1561 // and probably not worth it.
1562 class_dex_file = dex_file;
1563 type_index = cls->FindTypeIndexInOtherDexFile(*dex_file);
1564 } else {
1565 class_dex_file = &(cls->GetDexFile());
1566 type_index = cls->GetDexTypeIndex();
1567 }
1568 if (!type_index.IsValid()) {
1569 // Could be a proxy class or an array for which we couldn't find the type index.
1570 is_missing_types = true;
1571 continue;
1572 }
1573 if (ContainsElement(dex_base_locations,
1574 DexFileLoader::GetBaseLocation(class_dex_file->GetLocation()))) {
1575 // Only consider classes from the same apk (including multidex).
1576 profile_classes.emplace_back(/*ProfileMethodInfo::ProfileClassReference*/
1577 class_dex_file, type_index);
1578 } else {
1579 is_missing_types = true;
1580 }
1581 }
1582 if (!profile_classes.empty()) {
1583 inline_caches.emplace_back(/*ProfileMethodInfo::ProfileInlineCache*/
1584 cache.dex_pc_, is_missing_types, profile_classes);
1585 }
1586 }
1587 methods.emplace_back(/*ProfileMethodInfo*/
1588 MethodReference(dex_file, method->GetDexMethodIndex()), inline_caches);
1589 }
1590 }
1591
IsOsrCompiled(ArtMethod * method)1592 bool JitCodeCache::IsOsrCompiled(ArtMethod* method) {
1593 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1594 return osr_code_map_.find(method) != osr_code_map_.end();
1595 }
1596
NotifyCompilationOf(ArtMethod * method,Thread * self,CompilationKind compilation_kind,bool prejit)1597 bool JitCodeCache::NotifyCompilationOf(ArtMethod* method,
1598 Thread* self,
1599 CompilationKind compilation_kind,
1600 bool prejit) {
1601 const void* existing_entry_point = method->GetEntryPointFromQuickCompiledCode();
1602 if (compilation_kind != CompilationKind::kOsr && ContainsPc(existing_entry_point)) {
1603 OatQuickMethodHeader* method_header =
1604 OatQuickMethodHeader::FromEntryPoint(existing_entry_point);
1605 bool is_baseline = (compilation_kind == CompilationKind::kBaseline);
1606 if (CodeInfo::IsBaseline(method_header->GetOptimizedCodeInfoPtr()) == is_baseline) {
1607 VLOG(jit) << "Not compiling "
1608 << method->PrettyMethod()
1609 << " because it has already been compiled"
1610 << " kind=" << compilation_kind;
1611 return false;
1612 }
1613 }
1614
1615 if (NeedsClinitCheckBeforeCall(method) && !prejit) {
1616 // We do not need a synchronization barrier for checking the visibly initialized status
1617 // or checking the initialized status just for requesting visible initialization.
1618 ClassStatus status = method->GetDeclaringClass()
1619 ->GetStatus<kDefaultVerifyFlags, /*kWithSynchronizationBarrier=*/ false>();
1620 if (status != ClassStatus::kVisiblyInitialized) {
1621 // Unless we're pre-jitting, we currently don't save the JIT compiled code if we cannot
1622 // update the entrypoint due to needing an initialization check.
1623 if (status == ClassStatus::kInitialized) {
1624 // Request visible initialization but do not block to allow compiling other methods.
1625 // Hopefully, this will complete by the time the method becomes hot again.
1626 Runtime::Current()->GetClassLinker()->MakeInitializedClassesVisiblyInitialized(
1627 self, /*wait=*/ false);
1628 }
1629 VLOG(jit) << "Not compiling "
1630 << method->PrettyMethod()
1631 << " because it has the resolution stub";
1632 // Give it a new chance to be hot.
1633 ClearMethodCounter(method, /*was_warm=*/ false);
1634 return false;
1635 }
1636 }
1637
1638 if (compilation_kind == CompilationKind::kOsr) {
1639 MutexLock mu(self, *Locks::jit_lock_);
1640 if (osr_code_map_.find(method) != osr_code_map_.end()) {
1641 return false;
1642 }
1643 }
1644
1645 if (UNLIKELY(method->IsNative())) {
1646 MutexLock mu(self, *Locks::jit_lock_);
1647 JniStubKey key(method);
1648 auto it = jni_stubs_map_.find(key);
1649 bool new_compilation = false;
1650 if (it == jni_stubs_map_.end()) {
1651 // Create a new entry to mark the stub as being compiled.
1652 it = jni_stubs_map_.Put(key, JniStubData{});
1653 new_compilation = true;
1654 }
1655 JniStubData* data = &it->second;
1656 data->AddMethod(method);
1657 if (data->IsCompiled()) {
1658 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(data->GetCode());
1659 const void* entrypoint = method_header->GetEntryPoint();
1660 // Update also entrypoints of other methods held by the JniStubData.
1661 // We could simply update the entrypoint of `method` but if the last JIT GC has
1662 // changed these entrypoints to GenericJNI in preparation for a full GC, we may
1663 // as well change them back as this stub shall not be collected anyway and this
1664 // can avoid a few expensive GenericJNI calls.
1665 data->UpdateEntryPoints(entrypoint);
1666 if (collection_in_progress_) {
1667 if (!IsInZygoteExecSpace(data->GetCode())) {
1668 GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(data->GetCode()));
1669 }
1670 }
1671 }
1672 return new_compilation;
1673 } else {
1674 if (compilation_kind == CompilationKind::kBaseline) {
1675 DCHECK(CanAllocateProfilingInfo());
1676 bool has_profiling_info = false;
1677 {
1678 MutexLock mu(self, *Locks::jit_lock_);
1679 has_profiling_info = (profiling_infos_.find(method) != profiling_infos_.end());
1680 }
1681 if (!has_profiling_info) {
1682 if (ProfilingInfo::Create(self, method) == nullptr) {
1683 VLOG(jit) << method->PrettyMethod() << " needs a ProfilingInfo to be compiled baseline";
1684 ClearMethodCounter(method, /*was_warm=*/ false);
1685 return false;
1686 }
1687 }
1688 }
1689 MutexLock mu(self, *Locks::jit_lock_);
1690 if (IsMethodBeingCompiled(method, compilation_kind)) {
1691 return false;
1692 }
1693 AddMethodBeingCompiled(method, compilation_kind);
1694 return true;
1695 }
1696 }
1697
NotifyCompilerUse(ArtMethod * method,Thread * self)1698 ProfilingInfo* JitCodeCache::NotifyCompilerUse(ArtMethod* method, Thread* self) {
1699 MutexLock mu(self, *Locks::jit_lock_);
1700 auto it = profiling_infos_.find(method);
1701 if (it == profiling_infos_.end()) {
1702 return nullptr;
1703 }
1704 if (!it->second->IncrementInlineUse()) {
1705 // Overflow of inlining uses, just bail.
1706 return nullptr;
1707 }
1708 return it->second;
1709 }
1710
DoneCompilerUse(ArtMethod * method,Thread * self)1711 void JitCodeCache::DoneCompilerUse(ArtMethod* method, Thread* self) {
1712 MutexLock mu(self, *Locks::jit_lock_);
1713 auto it = profiling_infos_.find(method);
1714 DCHECK(it != profiling_infos_.end());
1715 it->second->DecrementInlineUse();
1716 }
1717
DoneCompiling(ArtMethod * method,Thread * self,CompilationKind compilation_kind)1718 void JitCodeCache::DoneCompiling(ArtMethod* method,
1719 Thread* self,
1720 CompilationKind compilation_kind) {
1721 DCHECK_EQ(Thread::Current(), self);
1722 MutexLock mu(self, *Locks::jit_lock_);
1723 if (UNLIKELY(method->IsNative())) {
1724 auto it = jni_stubs_map_.find(JniStubKey(method));
1725 DCHECK(it != jni_stubs_map_.end());
1726 JniStubData* data = &it->second;
1727 DCHECK(ContainsElement(data->GetMethods(), method));
1728 if (UNLIKELY(!data->IsCompiled())) {
1729 // Failed to compile; the JNI compiler never fails, but the cache may be full.
1730 jni_stubs_map_.erase(it); // Remove the entry added in NotifyCompilationOf().
1731 } // else Commit() updated entrypoints of all methods in the JniStubData.
1732 } else {
1733 RemoveMethodBeingCompiled(method, compilation_kind);
1734 }
1735 }
1736
InvalidateAllCompiledCode()1737 void JitCodeCache::InvalidateAllCompiledCode() {
1738 art::MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1739 VLOG(jit) << "Invalidating all compiled code";
1740 ClassLinker* linker = Runtime::Current()->GetClassLinker();
1741 for (auto it : method_code_map_) {
1742 ArtMethod* meth = it.second;
1743 // We were compiled, so we must be warm.
1744 ClearMethodCounter(meth, /*was_warm=*/true);
1745 if (meth->IsObsolete()) {
1746 linker->SetEntryPointsForObsoleteMethod(meth);
1747 } else {
1748 Runtime::Current()->GetInstrumentation()->InitializeMethodsCode(meth, /*aot_code=*/ nullptr);
1749 }
1750 }
1751 saved_compiled_methods_map_.clear();
1752 osr_code_map_.clear();
1753 }
1754
InvalidateCompiledCodeFor(ArtMethod * method,const OatQuickMethodHeader * header)1755 void JitCodeCache::InvalidateCompiledCodeFor(ArtMethod* method,
1756 const OatQuickMethodHeader* header) {
1757 DCHECK(!method->IsNative());
1758 const void* method_entrypoint = method->GetEntryPointFromQuickCompiledCode();
1759
1760 // Clear the method counter if we are running jitted code since we might want to jit this again in
1761 // the future.
1762 if (method_entrypoint == header->GetEntryPoint()) {
1763 // The entrypoint is the one to invalidate, so we just update it to the interpreter entry point
1764 // and clear the counter to get the method Jitted again.
1765 Runtime::Current()->GetInstrumentation()->InitializeMethodsCode(method, /*aot_code=*/ nullptr);
1766 ClearMethodCounter(method, /*was_warm=*/ true);
1767 } else {
1768 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1769 auto it = osr_code_map_.find(method);
1770 if (it != osr_code_map_.end() && OatQuickMethodHeader::FromCodePointer(it->second) == header) {
1771 // Remove the OSR method, to avoid using it again.
1772 osr_code_map_.erase(it);
1773 }
1774 }
1775
1776 // In case the method was pre-compiled, clear that information so we
1777 // can recompile it ourselves.
1778 if (method->IsPreCompiled()) {
1779 method->ClearPreCompiled();
1780 }
1781 }
1782
Dump(std::ostream & os)1783 void JitCodeCache::Dump(std::ostream& os) {
1784 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1785 os << "Current JIT code cache size (used / resident): "
1786 << GetCurrentRegion()->GetUsedMemoryForCode() / KB << "KB / "
1787 << GetCurrentRegion()->GetResidentMemoryForCode() / KB << "KB\n"
1788 << "Current JIT data cache size (used / resident): "
1789 << GetCurrentRegion()->GetUsedMemoryForData() / KB << "KB / "
1790 << GetCurrentRegion()->GetResidentMemoryForData() / KB << "KB\n";
1791 if (!Runtime::Current()->IsZygote()) {
1792 os << "Zygote JIT code cache size (at point of fork): "
1793 << shared_region_.GetUsedMemoryForCode() / KB << "KB / "
1794 << shared_region_.GetResidentMemoryForCode() / KB << "KB\n"
1795 << "Zygote JIT data cache size (at point of fork): "
1796 << shared_region_.GetUsedMemoryForData() / KB << "KB / "
1797 << shared_region_.GetResidentMemoryForData() / KB << "KB\n";
1798 }
1799 os << "Current JIT mini-debug-info size: " << PrettySize(GetJitMiniDebugInfoMemUsage()) << "\n"
1800 << "Current JIT capacity: " << PrettySize(GetCurrentRegion()->GetCurrentCapacity()) << "\n"
1801 << "Current number of JIT JNI stub entries: " << jni_stubs_map_.size() << "\n"
1802 << "Current number of JIT code cache entries: " << method_code_map_.size() << "\n"
1803 << "Total number of JIT baseline compilations: " << number_of_baseline_compilations_ << "\n"
1804 << "Total number of JIT optimized compilations: " << number_of_optimized_compilations_ << "\n"
1805 << "Total number of JIT compilations for on stack replacement: "
1806 << number_of_osr_compilations_ << "\n"
1807 << "Total number of JIT code cache collections: " << number_of_collections_ << std::endl;
1808 histogram_stack_map_memory_use_.PrintMemoryUse(os);
1809 histogram_code_memory_use_.PrintMemoryUse(os);
1810 histogram_profiling_info_memory_use_.PrintMemoryUse(os);
1811 }
1812
PostForkChildAction(bool is_system_server,bool is_zygote)1813 void JitCodeCache::PostForkChildAction(bool is_system_server, bool is_zygote) {
1814 Thread* self = Thread::Current();
1815
1816 // Remove potential tasks that have been inherited from the zygote.
1817 // We do this now and not in Jit::PostForkChildAction, as system server calls
1818 // JitCodeCache::PostForkChildAction first, and then does some code loading
1819 // that may result in new JIT tasks that we want to keep.
1820 ThreadPool* pool = Runtime::Current()->GetJit()->GetThreadPool();
1821 if (pool != nullptr) {
1822 pool->RemoveAllTasks(self);
1823 }
1824
1825 MutexLock mu(self, *Locks::jit_lock_);
1826
1827 // Reset potential writable MemMaps inherited from the zygote. We never want
1828 // to write to them.
1829 shared_region_.ResetWritableMappings();
1830
1831 if (is_zygote || Runtime::Current()->IsSafeMode()) {
1832 // Don't create a private region for a child zygote. Regions are usually map shared
1833 // (to satisfy dual-view), and we don't want children of a child zygote to inherit it.
1834 return;
1835 }
1836
1837 // Reset all statistics to be specific to this process.
1838 number_of_baseline_compilations_ = 0;
1839 number_of_optimized_compilations_ = 0;
1840 number_of_osr_compilations_ = 0;
1841 number_of_collections_ = 0;
1842 histogram_stack_map_memory_use_.Reset();
1843 histogram_code_memory_use_.Reset();
1844 histogram_profiling_info_memory_use_.Reset();
1845
1846 size_t initial_capacity = Runtime::Current()->GetJITOptions()->GetCodeCacheInitialCapacity();
1847 size_t max_capacity = Runtime::Current()->GetJITOptions()->GetCodeCacheMaxCapacity();
1848 std::string error_msg;
1849 if (!private_region_.Initialize(initial_capacity,
1850 max_capacity,
1851 /* rwx_memory_allowed= */ !is_system_server,
1852 is_zygote,
1853 &error_msg)) {
1854 LOG(WARNING) << "Could not create private region after zygote fork: " << error_msg;
1855 }
1856 }
1857
GetCurrentRegion()1858 JitMemoryRegion* JitCodeCache::GetCurrentRegion() {
1859 return Runtime::Current()->IsZygote() ? &shared_region_ : &private_region_;
1860 }
1861
VisitAllMethods(const std::function<void (const void *,ArtMethod *)> & cb)1862 void JitCodeCache::VisitAllMethods(const std::function<void(const void*, ArtMethod*)>& cb) {
1863 for (const auto& it : jni_stubs_map_) {
1864 const JniStubData& data = it.second;
1865 if (data.IsCompiled()) {
1866 for (ArtMethod* method : data.GetMethods()) {
1867 cb(data.GetCode(), method);
1868 }
1869 }
1870 }
1871 for (auto it : method_code_map_) { // Includes OSR methods.
1872 cb(it.first, it.second);
1873 }
1874 for (auto it : saved_compiled_methods_map_) {
1875 cb(it.second, it.first);
1876 }
1877 for (auto it : zygote_map_) {
1878 if (it.code_ptr != nullptr && it.method != nullptr) {
1879 cb(it.code_ptr, it.method);
1880 }
1881 }
1882 }
1883
Initialize(uint32_t number_of_methods)1884 void ZygoteMap::Initialize(uint32_t number_of_methods) {
1885 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1886 // Allocate for 40-80% capacity. This will offer OK lookup times, and termination
1887 // cases.
1888 size_t capacity = RoundUpToPowerOfTwo(number_of_methods * 100 / 80);
1889 const uint8_t* memory = region_->AllocateData(
1890 capacity * sizeof(Entry) + sizeof(ZygoteCompilationState));
1891 if (memory == nullptr) {
1892 LOG(WARNING) << "Could not allocate data for the zygote map";
1893 return;
1894 }
1895 const Entry* data = reinterpret_cast<const Entry*>(memory);
1896 region_->FillData(data, capacity, Entry { nullptr, nullptr });
1897 map_ = ArrayRef(data, capacity);
1898 compilation_state_ = reinterpret_cast<const ZygoteCompilationState*>(
1899 memory + capacity * sizeof(Entry));
1900 region_->WriteData(compilation_state_, ZygoteCompilationState::kInProgress);
1901 }
1902
GetCodeFor(ArtMethod * method,uintptr_t pc) const1903 const void* ZygoteMap::GetCodeFor(ArtMethod* method, uintptr_t pc) const {
1904 if (map_.empty()) {
1905 return nullptr;
1906 }
1907
1908 if (method == nullptr) {
1909 // Do a linear search. This should only be used in debug builds.
1910 CHECK(kIsDebugBuild);
1911 for (const Entry& entry : map_) {
1912 const void* code_ptr = entry.code_ptr;
1913 if (code_ptr != nullptr) {
1914 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1915 if (method_header->Contains(pc)) {
1916 return code_ptr;
1917 }
1918 }
1919 }
1920 return nullptr;
1921 }
1922
1923 std::hash<ArtMethod*> hf;
1924 size_t index = hf(method) & (map_.size() - 1u);
1925 size_t original_index = index;
1926 // Loop over the array: we know this loop terminates as we will either
1927 // encounter the given method, or a null entry. Both terminate the loop.
1928 // Note that the zygote may concurrently write new entries to the map. That's OK as the
1929 // map is never resized.
1930 while (true) {
1931 const Entry& entry = map_[index];
1932 if (entry.method == nullptr) {
1933 // Not compiled yet.
1934 return nullptr;
1935 }
1936 if (entry.method == method) {
1937 if (entry.code_ptr == nullptr) {
1938 // This is a race with the zygote which wrote the method, but hasn't written the
1939 // code. Just bail and wait for the next time we need the method.
1940 return nullptr;
1941 }
1942 if (pc != 0 && !OatQuickMethodHeader::FromCodePointer(entry.code_ptr)->Contains(pc)) {
1943 return nullptr;
1944 }
1945 return entry.code_ptr;
1946 }
1947 index = (index + 1) & (map_.size() - 1);
1948 DCHECK_NE(original_index, index);
1949 }
1950 }
1951
Put(const void * code,ArtMethod * method)1952 void ZygoteMap::Put(const void* code, ArtMethod* method) {
1953 if (map_.empty()) {
1954 return;
1955 }
1956 CHECK(Runtime::Current()->IsZygote());
1957 std::hash<ArtMethod*> hf;
1958 size_t index = hf(method) & (map_.size() - 1);
1959 size_t original_index = index;
1960 // Because the size of the map is bigger than the number of methods that will
1961 // be added, we are guaranteed to find a free slot in the array, and
1962 // therefore for this loop to terminate.
1963 while (true) {
1964 const Entry* entry = &map_[index];
1965 if (entry->method == nullptr) {
1966 // Note that readers can read this memory concurrently, but that's OK as
1967 // we are writing pointers.
1968 region_->WriteData(entry, Entry { method, code });
1969 break;
1970 }
1971 index = (index + 1) & (map_.size() - 1);
1972 DCHECK_NE(original_index, index);
1973 }
1974 DCHECK_EQ(GetCodeFor(method), code);
1975 }
1976
1977 } // namespace jit
1978 } // namespace art
1979