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_list.h"
61
62 namespace art {
63 namespace jit {
64
65 static constexpr size_t kCodeSizeLogThreshold = 50 * KB;
66 static constexpr size_t kStackMapSizeLogThreshold = 50 * KB;
67
68 class JitCodeCache::JniStubKey {
69 public:
REQUIRES_SHARED(Locks::mutator_lock_)70 explicit JniStubKey(ArtMethod* method) REQUIRES_SHARED(Locks::mutator_lock_)
71 : shorty_(method->GetShorty()),
72 is_static_(method->IsStatic()),
73 is_fast_native_(method->IsFastNative()),
74 is_critical_native_(method->IsCriticalNative()),
75 is_synchronized_(method->IsSynchronized()) {
76 DCHECK(!(is_fast_native_ && is_critical_native_));
77 }
78
operator <(const JniStubKey & rhs) const79 bool operator<(const JniStubKey& rhs) const {
80 if (is_static_ != rhs.is_static_) {
81 return rhs.is_static_;
82 }
83 if (is_synchronized_ != rhs.is_synchronized_) {
84 return rhs.is_synchronized_;
85 }
86 if (is_fast_native_ != rhs.is_fast_native_) {
87 return rhs.is_fast_native_;
88 }
89 if (is_critical_native_ != rhs.is_critical_native_) {
90 return rhs.is_critical_native_;
91 }
92 return strcmp(shorty_, rhs.shorty_) < 0;
93 }
94
95 // Update the shorty to point to another method's shorty. Call this function when removing
96 // the method that references the old shorty from JniCodeData and not removing the entire
97 // JniCodeData; the old shorty may become a dangling pointer when that method is unloaded.
UpdateShorty(ArtMethod * method) const98 void UpdateShorty(ArtMethod* method) const REQUIRES_SHARED(Locks::mutator_lock_) {
99 const char* shorty = method->GetShorty();
100 DCHECK_STREQ(shorty_, shorty);
101 shorty_ = shorty;
102 }
103
104 private:
105 // The shorty points to a DexFile data and may need to change
106 // to point to the same shorty in a different DexFile.
107 mutable const char* shorty_;
108
109 const bool is_static_;
110 const bool is_fast_native_;
111 const bool is_critical_native_;
112 const bool is_synchronized_;
113 };
114
115 class JitCodeCache::JniStubData {
116 public:
JniStubData()117 JniStubData() : code_(nullptr), methods_() {}
118
SetCode(const void * code)119 void SetCode(const void* code) {
120 DCHECK(code != nullptr);
121 code_ = code;
122 }
123
UpdateEntryPoints(const void * entrypoint)124 void UpdateEntryPoints(const void* entrypoint) REQUIRES_SHARED(Locks::mutator_lock_) {
125 DCHECK(IsCompiled());
126 DCHECK(entrypoint == OatQuickMethodHeader::FromCodePointer(GetCode())->GetEntryPoint());
127 instrumentation::Instrumentation* instrum = Runtime::Current()->GetInstrumentation();
128 for (ArtMethod* m : GetMethods()) {
129 // Because `m` might be in the process of being deleted:
130 // - Call the dedicated method instead of the more generic UpdateMethodsCode
131 // - Check the class status without a full read barrier; use ReadBarrier::IsMarked().
132 bool can_set_entrypoint = true;
133 if (NeedsClinitCheckBeforeCall(m)) {
134 // To avoid resurrecting an unreachable object, we must not use a full read
135 // barrier but we do not want to miss updating an entrypoint under common
136 // circumstances, i.e. during a GC the class becomes visibly initialized,
137 // the method becomes hot, we compile the thunk and want to update the
138 // entrypoint while the method's declaring class field still points to the
139 // from-space class object with the old status. Therefore we read the
140 // declaring class without a read barrier and check if it's already marked.
141 // If yes, we check the status of the to-space class object as intended.
142 // Otherwise, there is no to-space object and the from-space class object
143 // contains the most recent value of the status field; even if this races
144 // with another thread doing a read barrier and updating the status, that's
145 // no different from a race with a thread that just updates the status.
146 // Such race can happen only for the zygote method pre-compilation, as we
147 // otherwise compile only thunks for methods of visibly initialized classes.
148 ObjPtr<mirror::Class> klass = m->GetDeclaringClass<kWithoutReadBarrier>();
149 ObjPtr<mirror::Class> marked = ReadBarrier::IsMarked(klass.Ptr());
150 ObjPtr<mirror::Class> checked_klass = (marked != nullptr) ? marked : klass;
151 can_set_entrypoint = checked_klass->IsVisiblyInitialized();
152 }
153 if (can_set_entrypoint) {
154 instrum->UpdateNativeMethodsCodeToJitCode(m, entrypoint);
155 }
156 }
157 }
158
GetCode() const159 const void* GetCode() const {
160 return code_;
161 }
162
IsCompiled() const163 bool IsCompiled() const {
164 return GetCode() != nullptr;
165 }
166
AddMethod(ArtMethod * method)167 void AddMethod(ArtMethod* method) {
168 if (!ContainsElement(methods_, method)) {
169 methods_.push_back(method);
170 }
171 }
172
GetMethods() const173 const std::vector<ArtMethod*>& GetMethods() const {
174 return methods_;
175 }
176
RemoveMethodsIn(const LinearAlloc & alloc)177 void RemoveMethodsIn(const LinearAlloc& alloc) REQUIRES_SHARED(Locks::mutator_lock_) {
178 auto kept_end = std::partition(
179 methods_.begin(),
180 methods_.end(),
181 [&alloc](ArtMethod* method) { return !alloc.ContainsUnsafe(method); });
182 for (auto it = kept_end; it != methods_.end(); it++) {
183 VLOG(jit) << "JIT removed (JNI) " << (*it)->PrettyMethod() << ": " << code_;
184 }
185 methods_.erase(kept_end, methods_.end());
186 }
187
RemoveMethod(ArtMethod * method)188 bool RemoveMethod(ArtMethod* method) REQUIRES_SHARED(Locks::mutator_lock_) {
189 auto it = std::find(methods_.begin(), methods_.end(), method);
190 if (it != methods_.end()) {
191 VLOG(jit) << "JIT removed (JNI) " << (*it)->PrettyMethod() << ": " << code_;
192 methods_.erase(it);
193 return true;
194 } else {
195 return false;
196 }
197 }
198
MoveObsoleteMethod(ArtMethod * old_method,ArtMethod * new_method)199 void MoveObsoleteMethod(ArtMethod* old_method, ArtMethod* new_method) {
200 std::replace(methods_.begin(), methods_.end(), old_method, new_method);
201 }
202
203 private:
204 const void* code_;
205 std::vector<ArtMethod*> methods_;
206 };
207
Create(bool used_only_for_profile_data,bool rwx_memory_allowed,bool is_zygote,std::string * error_msg)208 JitCodeCache* JitCodeCache::Create(bool used_only_for_profile_data,
209 bool rwx_memory_allowed,
210 bool is_zygote,
211 std::string* error_msg) {
212 // Register for membarrier expedited sync core if JIT will be generating code.
213 if (!used_only_for_profile_data) {
214 if (art::membarrier(art::MembarrierCommand::kRegisterPrivateExpeditedSyncCore) != 0) {
215 // MEMBARRIER_CMD_PRIVATE_EXPEDITED_SYNC_CORE ensures that CPU instruction pipelines are
216 // flushed and it's used when adding code to the JIT. The memory used by the new code may
217 // have just been released and, in theory, the old code could still be in a pipeline.
218 VLOG(jit) << "Kernel does not support membarrier sync-core";
219 }
220 }
221
222 size_t initial_capacity = Runtime::Current()->GetJITOptions()->GetCodeCacheInitialCapacity();
223 // Check whether the provided max capacity in options is below 1GB.
224 size_t max_capacity = Runtime::Current()->GetJITOptions()->GetCodeCacheMaxCapacity();
225 // We need to have 32 bit offsets from method headers in code cache which point to things
226 // in the data cache. If the maps are more than 4G apart, having multiple maps wouldn't work.
227 // Ensure we're below 1 GB to be safe.
228 if (max_capacity > 1 * GB) {
229 std::ostringstream oss;
230 oss << "Maxium code cache capacity is limited to 1 GB, "
231 << PrettySize(max_capacity) << " is too big";
232 *error_msg = oss.str();
233 return nullptr;
234 }
235
236 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
237 JitMemoryRegion region;
238 if (!region.Initialize(initial_capacity,
239 max_capacity,
240 rwx_memory_allowed,
241 is_zygote,
242 error_msg)) {
243 return nullptr;
244 }
245
246 std::unique_ptr<JitCodeCache> jit_code_cache(new JitCodeCache());
247 if (is_zygote) {
248 // Zygote should never collect code to share the memory with the children.
249 jit_code_cache->garbage_collect_code_ = false;
250 jit_code_cache->shared_region_ = std::move(region);
251 } else {
252 jit_code_cache->private_region_ = std::move(region);
253 }
254
255 VLOG(jit) << "Created jit code cache: initial capacity="
256 << PrettySize(initial_capacity)
257 << ", maximum capacity="
258 << PrettySize(max_capacity);
259
260 return jit_code_cache.release();
261 }
262
JitCodeCache()263 JitCodeCache::JitCodeCache()
264 : is_weak_access_enabled_(true),
265 inline_cache_cond_("Jit inline cache condition variable", *Locks::jit_lock_),
266 zygote_map_(&shared_region_),
267 lock_cond_("Jit code cache condition variable", *Locks::jit_lock_),
268 collection_in_progress_(false),
269 last_collection_increased_code_cache_(false),
270 garbage_collect_code_(true),
271 number_of_baseline_compilations_(0),
272 number_of_optimized_compilations_(0),
273 number_of_osr_compilations_(0),
274 number_of_collections_(0),
275 histogram_stack_map_memory_use_("Memory used for stack maps", 16),
276 histogram_code_memory_use_("Memory used for compiled code", 16),
277 histogram_profiling_info_memory_use_("Memory used for profiling info", 16) {
278 }
279
~JitCodeCache()280 JitCodeCache::~JitCodeCache() {}
281
PrivateRegionContainsPc(const void * ptr) const282 bool JitCodeCache::PrivateRegionContainsPc(const void* ptr) const {
283 return private_region_.IsInExecSpace(ptr);
284 }
285
ContainsPc(const void * ptr) const286 bool JitCodeCache::ContainsPc(const void* ptr) const {
287 return PrivateRegionContainsPc(ptr) || shared_region_.IsInExecSpace(ptr);
288 }
289
ContainsMethod(ArtMethod * method)290 bool JitCodeCache::ContainsMethod(ArtMethod* method) {
291 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
292 if (UNLIKELY(method->IsNative())) {
293 auto it = jni_stubs_map_.find(JniStubKey(method));
294 if (it != jni_stubs_map_.end() &&
295 it->second.IsCompiled() &&
296 ContainsElement(it->second.GetMethods(), method)) {
297 return true;
298 }
299 } else {
300 for (const auto& it : method_code_map_) {
301 if (it.second == method) {
302 return true;
303 }
304 }
305 if (zygote_map_.ContainsMethod(method)) {
306 return true;
307 }
308 }
309 return false;
310 }
311
GetJniStubCode(ArtMethod * method)312 const void* JitCodeCache::GetJniStubCode(ArtMethod* method) {
313 DCHECK(method->IsNative());
314 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
315 auto it = jni_stubs_map_.find(JniStubKey(method));
316 if (it != jni_stubs_map_.end()) {
317 JniStubData& data = it->second;
318 if (data.IsCompiled() && ContainsElement(data.GetMethods(), method)) {
319 return data.GetCode();
320 }
321 }
322 return nullptr;
323 }
324
GetSavedEntryPointOfPreCompiledMethod(ArtMethod * method)325 const void* JitCodeCache::GetSavedEntryPointOfPreCompiledMethod(ArtMethod* method) {
326 if (method->IsPreCompiled()) {
327 const void* code_ptr = nullptr;
328 if (method->GetDeclaringClass()->GetClassLoader() == nullptr) {
329 code_ptr = zygote_map_.GetCodeFor(method);
330 } else {
331 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
332 auto it = saved_compiled_methods_map_.find(method);
333 if (it != saved_compiled_methods_map_.end()) {
334 code_ptr = it->second;
335 }
336 }
337 if (code_ptr != nullptr) {
338 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
339 return method_header->GetEntryPoint();
340 }
341 }
342 return nullptr;
343 }
344
WaitForPotentialCollectionToComplete(Thread * self)345 bool JitCodeCache::WaitForPotentialCollectionToComplete(Thread* self) {
346 bool in_collection = false;
347 while (collection_in_progress_) {
348 in_collection = true;
349 lock_cond_.Wait(self);
350 }
351 return in_collection;
352 }
353
FromCodeToAllocation(const void * code)354 static uintptr_t FromCodeToAllocation(const void* code) {
355 size_t alignment = GetInstructionSetAlignment(kRuntimeISA);
356 return reinterpret_cast<uintptr_t>(code) - RoundUp(sizeof(OatQuickMethodHeader), alignment);
357 }
358
FromAllocationToCode(const uint8_t * alloc)359 static const void* FromAllocationToCode(const uint8_t* alloc) {
360 size_t alignment = GetInstructionSetAlignment(kRuntimeISA);
361 return reinterpret_cast<const void*>(alloc + RoundUp(sizeof(OatQuickMethodHeader), alignment));
362 }
363
GetNumberOfRoots(const uint8_t * stack_map)364 static uint32_t GetNumberOfRoots(const uint8_t* stack_map) {
365 // The length of the table is stored just before the stack map (and therefore at the end of
366 // the table itself), in order to be able to fetch it from a `stack_map` pointer.
367 return reinterpret_cast<const uint32_t*>(stack_map)[-1];
368 }
369
DCheckRootsAreValid(const std::vector<Handle<mirror::Object>> & roots,bool is_shared_region)370 static void DCheckRootsAreValid(const std::vector<Handle<mirror::Object>>& roots,
371 bool is_shared_region)
372 REQUIRES(!Locks::intern_table_lock_) REQUIRES_SHARED(Locks::mutator_lock_) {
373 if (!kIsDebugBuild) {
374 return;
375 }
376 // Put all roots in `roots_data`.
377 for (Handle<mirror::Object> object : roots) {
378 // Ensure the string is strongly interned. b/32995596
379 if (object->IsString()) {
380 ObjPtr<mirror::String> str = object->AsString();
381 ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
382 CHECK(class_linker->GetInternTable()->LookupStrong(Thread::Current(), str) != nullptr);
383 }
384 // Ensure that we don't put movable objects in the shared region.
385 if (is_shared_region) {
386 CHECK(!Runtime::Current()->GetHeap()->IsMovableObject(object.Get()));
387 }
388 }
389 }
390
GetRootTable(const void * code_ptr,uint32_t * number_of_roots=nullptr)391 static const uint8_t* GetRootTable(const void* code_ptr, uint32_t* number_of_roots = nullptr) {
392 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
393 uint8_t* data = method_header->GetOptimizedCodeInfoPtr();
394 uint32_t roots = GetNumberOfRoots(data);
395 if (number_of_roots != nullptr) {
396 *number_of_roots = roots;
397 }
398 return data - ComputeRootTableSize(roots);
399 }
400
SweepRootTables(IsMarkedVisitor * visitor)401 void JitCodeCache::SweepRootTables(IsMarkedVisitor* visitor) {
402 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
403 for (const auto& entry : method_code_map_) {
404 uint32_t number_of_roots = 0;
405 const uint8_t* root_table = GetRootTable(entry.first, &number_of_roots);
406 uint8_t* roots_data = private_region_.IsInDataSpace(root_table)
407 ? private_region_.GetWritableDataAddress(root_table)
408 : shared_region_.GetWritableDataAddress(root_table);
409 GcRoot<mirror::Object>* roots = reinterpret_cast<GcRoot<mirror::Object>*>(roots_data);
410 for (uint32_t i = 0; i < number_of_roots; ++i) {
411 // This does not need a read barrier because this is called by GC.
412 mirror::Object* object = roots[i].Read<kWithoutReadBarrier>();
413 if (object == nullptr || object == Runtime::GetWeakClassSentinel()) {
414 // entry got deleted in a previous sweep.
415 } else if (object->IsString<kDefaultVerifyFlags>()) {
416 mirror::Object* new_object = visitor->IsMarked(object);
417 // We know the string is marked because it's a strongly-interned string that
418 // is always alive. The IsMarked implementation of the CMS collector returns
419 // null for newly allocated objects, but we know those haven't moved. Therefore,
420 // only update the entry if we get a different non-null string.
421 // TODO: Do not use IsMarked for j.l.Class, and adjust once we move this method
422 // out of the weak access/creation pause. b/32167580
423 if (new_object != nullptr && new_object != object) {
424 DCHECK(new_object->IsString());
425 roots[i] = GcRoot<mirror::Object>(new_object);
426 }
427 } else {
428 Runtime::ProcessWeakClass(
429 reinterpret_cast<GcRoot<mirror::Class>*>(&roots[i]),
430 visitor,
431 Runtime::GetWeakClassSentinel());
432 }
433 }
434 }
435 // Walk over inline caches to clear entries containing unloaded classes.
436 for (auto it : profiling_infos_) {
437 ProfilingInfo* info = it.second;
438 for (size_t i = 0; i < info->number_of_inline_caches_; ++i) {
439 InlineCache* cache = &info->cache_[i];
440 for (size_t j = 0; j < InlineCache::kIndividualCacheSize; ++j) {
441 Runtime::ProcessWeakClass(&cache->classes_[j], visitor, nullptr);
442 }
443 }
444 }
445 }
446
FreeCodeAndData(const void * code_ptr)447 void JitCodeCache::FreeCodeAndData(const void* code_ptr) {
448 if (IsInZygoteExecSpace(code_ptr)) {
449 // No need to free, this is shared memory.
450 return;
451 }
452 uintptr_t allocation = FromCodeToAllocation(code_ptr);
453 const uint8_t* data = nullptr;
454 if (OatQuickMethodHeader::FromCodePointer(code_ptr)->IsOptimized()) {
455 data = GetRootTable(code_ptr);
456 } // else this is a JNI stub without any data.
457
458 FreeLocked(&private_region_, reinterpret_cast<uint8_t*>(allocation), data);
459 }
460
FreeAllMethodHeaders(const std::unordered_set<OatQuickMethodHeader * > & method_headers)461 void JitCodeCache::FreeAllMethodHeaders(
462 const std::unordered_set<OatQuickMethodHeader*>& method_headers) {
463 // We need to remove entries in method_headers from CHA dependencies
464 // first since once we do FreeCode() below, the memory can be reused
465 // so it's possible for the same method_header to start representing
466 // different compile code.
467 {
468 MutexLock mu2(Thread::Current(), *Locks::cha_lock_);
469 Runtime::Current()->GetClassLinker()->GetClassHierarchyAnalysis()
470 ->RemoveDependentsWithMethodHeaders(method_headers);
471 }
472
473 ScopedCodeCacheWrite scc(private_region_);
474 for (const OatQuickMethodHeader* method_header : method_headers) {
475 FreeCodeAndData(method_header->GetCode());
476 }
477
478 // We have potentially removed a lot of debug info. Do maintenance pass to save space.
479 RepackNativeDebugInfoForJit();
480
481 // Check that the set of compiled methods exactly matches native debug information.
482 // Does not check zygote methods since they can change concurrently.
483 if (kIsDebugBuild && !Runtime::Current()->IsZygote()) {
484 std::map<const void*, ArtMethod*> compiled_methods;
485 VisitAllMethods([&](const void* addr, ArtMethod* method) {
486 if (!IsInZygoteExecSpace(addr)) {
487 CHECK(addr != nullptr && method != nullptr);
488 compiled_methods.emplace(addr, method);
489 }
490 });
491 std::set<const void*> debug_info;
492 ForEachNativeDebugSymbol([&](const void* addr, size_t, const char* name) {
493 addr = AlignDown(addr, GetInstructionSetInstructionAlignment(kRuntimeISA)); // Thumb-bit.
494 CHECK(debug_info.emplace(addr).second) << "Duplicate debug info: " << addr << " " << name;
495 CHECK_EQ(compiled_methods.count(addr), 1u) << "Extra debug info: " << addr << " " << name;
496 });
497 if (!debug_info.empty()) { // If debug-info generation is enabled.
498 for (auto it : compiled_methods) {
499 CHECK_EQ(debug_info.count(it.first), 1u) << "No debug info: " << it.second->PrettyMethod();
500 }
501 CHECK_EQ(compiled_methods.size(), debug_info.size());
502 }
503 }
504 }
505
RemoveMethodsIn(Thread * self,const LinearAlloc & alloc)506 void JitCodeCache::RemoveMethodsIn(Thread* self, const LinearAlloc& alloc) {
507 ScopedTrace trace(__PRETTY_FUNCTION__);
508 // We use a set to first collect all method_headers whose code need to be
509 // removed. We need to free the underlying code after we remove CHA dependencies
510 // for entries in this set. And it's more efficient to iterate through
511 // the CHA dependency map just once with an unordered_set.
512 std::unordered_set<OatQuickMethodHeader*> method_headers;
513 {
514 MutexLock mu(self, *Locks::jit_lock_);
515 // We do not check if a code cache GC is in progress, as this method comes
516 // with the classlinker_classes_lock_ held, and suspending ourselves could
517 // lead to a deadlock.
518 {
519 for (auto it = jni_stubs_map_.begin(); it != jni_stubs_map_.end();) {
520 it->second.RemoveMethodsIn(alloc);
521 if (it->second.GetMethods().empty()) {
522 method_headers.insert(OatQuickMethodHeader::FromCodePointer(it->second.GetCode()));
523 it = jni_stubs_map_.erase(it);
524 } else {
525 it->first.UpdateShorty(it->second.GetMethods().front());
526 ++it;
527 }
528 }
529 for (auto it = method_code_map_.begin(); it != method_code_map_.end();) {
530 if (alloc.ContainsUnsafe(it->second)) {
531 method_headers.insert(OatQuickMethodHeader::FromCodePointer(it->first));
532 VLOG(jit) << "JIT removed " << it->second->PrettyMethod() << ": " << it->first;
533 it = method_code_map_.erase(it);
534 } else {
535 ++it;
536 }
537 }
538 }
539 for (auto it = osr_code_map_.begin(); it != osr_code_map_.end();) {
540 if (alloc.ContainsUnsafe(it->first)) {
541 // Note that the code has already been pushed to method_headers in the loop
542 // above and is going to be removed in FreeCode() below.
543 it = osr_code_map_.erase(it);
544 } else {
545 ++it;
546 }
547 }
548 for (auto it = profiling_infos_.begin(); it != profiling_infos_.end();) {
549 ProfilingInfo* info = it->second;
550 if (alloc.ContainsUnsafe(info->GetMethod())) {
551 private_region_.FreeWritableData(reinterpret_cast<uint8_t*>(info));
552 it = profiling_infos_.erase(it);
553 } else {
554 ++it;
555 }
556 }
557 FreeAllMethodHeaders(method_headers);
558 }
559 }
560
IsWeakAccessEnabled(Thread * self) const561 bool JitCodeCache::IsWeakAccessEnabled(Thread* self) const {
562 return kUseReadBarrier
563 ? self->GetWeakRefAccessEnabled()
564 : is_weak_access_enabled_.load(std::memory_order_seq_cst);
565 }
566
WaitUntilInlineCacheAccessible(Thread * self)567 void JitCodeCache::WaitUntilInlineCacheAccessible(Thread* self) {
568 if (IsWeakAccessEnabled(self)) {
569 return;
570 }
571 ScopedThreadSuspension sts(self, kWaitingWeakGcRootRead);
572 MutexLock mu(self, *Locks::jit_lock_);
573 while (!IsWeakAccessEnabled(self)) {
574 inline_cache_cond_.Wait(self);
575 }
576 }
577
BroadcastForInlineCacheAccess()578 void JitCodeCache::BroadcastForInlineCacheAccess() {
579 Thread* self = Thread::Current();
580 MutexLock mu(self, *Locks::jit_lock_);
581 inline_cache_cond_.Broadcast(self);
582 }
583
AllowInlineCacheAccess()584 void JitCodeCache::AllowInlineCacheAccess() {
585 DCHECK(!kUseReadBarrier);
586 is_weak_access_enabled_.store(true, std::memory_order_seq_cst);
587 BroadcastForInlineCacheAccess();
588 }
589
DisallowInlineCacheAccess()590 void JitCodeCache::DisallowInlineCacheAccess() {
591 DCHECK(!kUseReadBarrier);
592 is_weak_access_enabled_.store(false, std::memory_order_seq_cst);
593 }
594
CopyInlineCacheInto(const InlineCache & ic,StackHandleScope<InlineCache::kIndividualCacheSize> * classes)595 void JitCodeCache::CopyInlineCacheInto(
596 const InlineCache& ic,
597 /*out*/StackHandleScope<InlineCache::kIndividualCacheSize>* classes) {
598 static_assert(arraysize(ic.classes_) == InlineCache::kIndividualCacheSize);
599 DCHECK_EQ(classes->NumberOfReferences(), InlineCache::kIndividualCacheSize);
600 DCHECK_EQ(classes->RemainingSlots(), InlineCache::kIndividualCacheSize);
601 WaitUntilInlineCacheAccessible(Thread::Current());
602 // Note that we don't need to lock `lock_` here, the compiler calling
603 // this method has already ensured the inline cache will not be deleted.
604 for (const GcRoot<mirror::Class>& root : ic.classes_) {
605 mirror::Class* object = root.Read();
606 if (object != nullptr) {
607 DCHECK_NE(classes->RemainingSlots(), 0u);
608 classes->NewHandle(object);
609 }
610 }
611 }
612
ClearMethodCounter(ArtMethod * method,bool was_warm)613 static void ClearMethodCounter(ArtMethod* method, bool was_warm)
614 REQUIRES_SHARED(Locks::mutator_lock_) {
615 if (was_warm) {
616 method->SetPreviouslyWarm();
617 }
618 // We reset the counter to 1 so that the profile knows that the method was executed at least once.
619 // This is required for layout purposes.
620 // We also need to make sure we'll pass the warmup threshold again, so we set to 0 if
621 // the warmup threshold is 1.
622 uint16_t jit_warmup_threshold = Runtime::Current()->GetJITOptions()->GetWarmupThreshold();
623 method->SetCounter(std::min(jit_warmup_threshold - 1, 1));
624 }
625
WaitForPotentialCollectionToCompleteRunnable(Thread * self)626 void JitCodeCache::WaitForPotentialCollectionToCompleteRunnable(Thread* self) {
627 while (collection_in_progress_) {
628 Locks::jit_lock_->Unlock(self);
629 {
630 ScopedThreadSuspension sts(self, kSuspended);
631 MutexLock mu(self, *Locks::jit_lock_);
632 WaitForPotentialCollectionToComplete(self);
633 }
634 Locks::jit_lock_->Lock(self);
635 }
636 }
637
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)638 bool JitCodeCache::Commit(Thread* self,
639 JitMemoryRegion* region,
640 ArtMethod* method,
641 ArrayRef<const uint8_t> reserved_code,
642 ArrayRef<const uint8_t> code,
643 ArrayRef<const uint8_t> reserved_data,
644 const std::vector<Handle<mirror::Object>>& roots,
645 ArrayRef<const uint8_t> stack_map,
646 const std::vector<uint8_t>& debug_info,
647 bool is_full_debug_info,
648 CompilationKind compilation_kind,
649 bool has_should_deoptimize_flag,
650 const ArenaSet<ArtMethod*>& cha_single_implementation_list) {
651 DCHECK(!method->IsNative() || (compilation_kind != CompilationKind::kOsr));
652
653 if (!method->IsNative()) {
654 // We need to do this before grabbing the lock_ because it needs to be able to see the string
655 // InternTable. Native methods do not have roots.
656 DCheckRootsAreValid(roots, IsSharedRegion(*region));
657 }
658
659 const uint8_t* roots_data = reserved_data.data();
660 size_t root_table_size = ComputeRootTableSize(roots.size());
661 const uint8_t* stack_map_data = roots_data + root_table_size;
662
663 MutexLock mu(self, *Locks::jit_lock_);
664 // We need to make sure that there will be no jit-gcs going on and wait for any ongoing one to
665 // finish.
666 WaitForPotentialCollectionToCompleteRunnable(self);
667 const uint8_t* code_ptr = region->CommitCode(
668 reserved_code, code, stack_map_data, has_should_deoptimize_flag);
669 if (code_ptr == nullptr) {
670 return false;
671 }
672 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
673
674 // Commit roots and stack maps before updating the entry point.
675 if (!region->CommitData(reserved_data, roots, stack_map)) {
676 return false;
677 }
678
679 switch (compilation_kind) {
680 case CompilationKind::kOsr:
681 number_of_osr_compilations_++;
682 break;
683 case CompilationKind::kBaseline:
684 number_of_baseline_compilations_++;
685 break;
686 case CompilationKind::kOptimized:
687 number_of_optimized_compilations_++;
688 break;
689 }
690
691 // We need to update the debug info before the entry point gets set.
692 // At the same time we want to do under JIT lock so that debug info and JIT maps are in sync.
693 if (!debug_info.empty()) {
694 // NB: Don't allow packing of full info since it would remove non-backtrace data.
695 AddNativeDebugInfoForJit(code_ptr, debug_info, /*allow_packing=*/ !is_full_debug_info);
696 }
697
698 // We need to update the entry point in the runnable state for the instrumentation.
699 {
700 // The following needs to be guarded by cha_lock_ also. Otherwise it's possible that the
701 // compiled code is considered invalidated by some class linking, but below we still make the
702 // compiled code valid for the method. Need cha_lock_ for checking all single-implementation
703 // flags and register dependencies.
704 MutexLock cha_mu(self, *Locks::cha_lock_);
705 bool single_impl_still_valid = true;
706 for (ArtMethod* single_impl : cha_single_implementation_list) {
707 if (!single_impl->HasSingleImplementation()) {
708 // Simply discard the compiled code. Clear the counter so that it may be recompiled later.
709 // Hopefully the class hierarchy will be more stable when compilation is retried.
710 single_impl_still_valid = false;
711 ClearMethodCounter(method, /*was_warm=*/ false);
712 break;
713 }
714 }
715
716 // Discard the code if any single-implementation assumptions are now invalid.
717 if (UNLIKELY(!single_impl_still_valid)) {
718 VLOG(jit) << "JIT discarded jitted code due to invalid single-implementation assumptions.";
719 return false;
720 }
721 DCHECK(cha_single_implementation_list.empty() || !Runtime::Current()->IsJavaDebuggable())
722 << "Should not be using cha on debuggable apps/runs!";
723
724 ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
725 for (ArtMethod* single_impl : cha_single_implementation_list) {
726 class_linker->GetClassHierarchyAnalysis()->AddDependency(single_impl, method, method_header);
727 }
728
729 if (UNLIKELY(method->IsNative())) {
730 auto it = jni_stubs_map_.find(JniStubKey(method));
731 DCHECK(it != jni_stubs_map_.end())
732 << "Entry inserted in NotifyCompilationOf() should be alive.";
733 JniStubData* data = &it->second;
734 DCHECK(ContainsElement(data->GetMethods(), method))
735 << "Entry inserted in NotifyCompilationOf() should contain this method.";
736 data->SetCode(code_ptr);
737 data->UpdateEntryPoints(method_header->GetEntryPoint());
738 } else {
739 if (method->IsPreCompiled() && IsSharedRegion(*region)) {
740 zygote_map_.Put(code_ptr, method);
741 } else {
742 method_code_map_.Put(code_ptr, method);
743 }
744 if (compilation_kind == CompilationKind::kOsr) {
745 osr_code_map_.Put(method, code_ptr);
746 } else if (NeedsClinitCheckBeforeCall(method) &&
747 !method->GetDeclaringClass()->IsVisiblyInitialized()) {
748 // This situation currently only occurs in the jit-zygote mode.
749 DCHECK(!garbage_collect_code_);
750 DCHECK(method->IsPreCompiled());
751 // The shared region can easily be queried. For the private region, we
752 // use a side map.
753 if (!IsSharedRegion(*region)) {
754 saved_compiled_methods_map_.Put(method, code_ptr);
755 }
756 } else {
757 Runtime::Current()->GetInstrumentation()->UpdateMethodsCode(
758 method, method_header->GetEntryPoint());
759 }
760 }
761 if (collection_in_progress_) {
762 // We need to update the live bitmap if there is a GC to ensure it sees this new
763 // code.
764 GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(code_ptr));
765 }
766 VLOG(jit)
767 << "JIT added (kind=" << compilation_kind << ") "
768 << ArtMethod::PrettyMethod(method) << "@" << method
769 << " ccache_size=" << PrettySize(CodeCacheSizeLocked()) << ": "
770 << " dcache_size=" << PrettySize(DataCacheSizeLocked()) << ": "
771 << reinterpret_cast<const void*>(method_header->GetEntryPoint()) << ","
772 << reinterpret_cast<const void*>(method_header->GetEntryPoint() +
773 method_header->GetCodeSize());
774 }
775
776 return true;
777 }
778
CodeCacheSize()779 size_t JitCodeCache::CodeCacheSize() {
780 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
781 return CodeCacheSizeLocked();
782 }
783
RemoveMethod(ArtMethod * method,bool release_memory)784 bool JitCodeCache::RemoveMethod(ArtMethod* method, bool release_memory) {
785 // This function is used only for testing and only with non-native methods.
786 CHECK(!method->IsNative());
787
788 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
789
790 bool osr = osr_code_map_.find(method) != osr_code_map_.end();
791 bool in_cache = RemoveMethodLocked(method, release_memory);
792
793 if (!in_cache) {
794 return false;
795 }
796
797 method->SetCounter(0);
798 Runtime::Current()->GetInstrumentation()->UpdateMethodsCode(
799 method, GetQuickToInterpreterBridge());
800 VLOG(jit)
801 << "JIT removed (osr=" << std::boolalpha << osr << std::noboolalpha << ") "
802 << ArtMethod::PrettyMethod(method) << "@" << method
803 << " ccache_size=" << PrettySize(CodeCacheSizeLocked()) << ": "
804 << " dcache_size=" << PrettySize(DataCacheSizeLocked());
805 return true;
806 }
807
RemoveMethodLocked(ArtMethod * method,bool release_memory)808 bool JitCodeCache::RemoveMethodLocked(ArtMethod* method, bool release_memory) {
809 if (LIKELY(!method->IsNative())) {
810 auto it = profiling_infos_.find(method);
811 if (it != profiling_infos_.end()) {
812 profiling_infos_.erase(it);
813 }
814 }
815
816 bool in_cache = false;
817 ScopedCodeCacheWrite ccw(private_region_);
818 if (UNLIKELY(method->IsNative())) {
819 auto it = jni_stubs_map_.find(JniStubKey(method));
820 if (it != jni_stubs_map_.end() && it->second.RemoveMethod(method)) {
821 in_cache = true;
822 if (it->second.GetMethods().empty()) {
823 if (release_memory) {
824 FreeCodeAndData(it->second.GetCode());
825 }
826 jni_stubs_map_.erase(it);
827 } else {
828 it->first.UpdateShorty(it->second.GetMethods().front());
829 }
830 }
831 } else {
832 for (auto it = method_code_map_.begin(); it != method_code_map_.end();) {
833 if (it->second == method) {
834 in_cache = true;
835 if (release_memory) {
836 FreeCodeAndData(it->first);
837 }
838 VLOG(jit) << "JIT removed " << it->second->PrettyMethod() << ": " << it->first;
839 it = method_code_map_.erase(it);
840 } else {
841 ++it;
842 }
843 }
844
845 auto osr_it = osr_code_map_.find(method);
846 if (osr_it != osr_code_map_.end()) {
847 osr_code_map_.erase(osr_it);
848 }
849 }
850
851 return in_cache;
852 }
853
854 // This notifies the code cache that the given method has been redefined and that it should remove
855 // any cached information it has on the method. All threads must be suspended before calling this
856 // method. The compiled code for the method (if there is any) must not be in any threads call stack.
NotifyMethodRedefined(ArtMethod * method)857 void JitCodeCache::NotifyMethodRedefined(ArtMethod* method) {
858 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
859 RemoveMethodLocked(method, /* release_memory= */ true);
860 }
861
862 // This invalidates old_method. Once this function returns one can no longer use old_method to
863 // execute code unless it is fixed up. This fixup will happen later in the process of installing a
864 // class redefinition.
865 // TODO We should add some info to ArtMethod to note that 'old_method' has been invalidated and
866 // shouldn't be used since it is no longer logically in the jit code cache.
867 // TODO We should add DCHECKS that validate that the JIT is paused when this method is entered.
MoveObsoleteMethod(ArtMethod * old_method,ArtMethod * new_method)868 void JitCodeCache::MoveObsoleteMethod(ArtMethod* old_method, ArtMethod* new_method) {
869 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
870 if (old_method->IsNative()) {
871 // Update methods in jni_stubs_map_.
872 for (auto& entry : jni_stubs_map_) {
873 JniStubData& data = entry.second;
874 data.MoveObsoleteMethod(old_method, new_method);
875 }
876 return;
877 }
878 // Update method_code_map_ to point to the new method.
879 for (auto& it : method_code_map_) {
880 if (it.second == old_method) {
881 it.second = new_method;
882 }
883 }
884 // Update osr_code_map_ to point to the new method.
885 auto code_map = osr_code_map_.find(old_method);
886 if (code_map != osr_code_map_.end()) {
887 osr_code_map_.Put(new_method, code_map->second);
888 osr_code_map_.erase(old_method);
889 }
890 }
891
TransitionToDebuggable()892 void JitCodeCache::TransitionToDebuggable() {
893 // Check that none of our methods have an entrypoint in the zygote exec
894 // space (this should be taken care of by
895 // ClassLinker::UpdateEntryPointsClassVisitor.
896 {
897 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
898 if (kIsDebugBuild) {
899 for (const auto& it : method_code_map_) {
900 ArtMethod* method = it.second;
901 DCHECK(!method->IsPreCompiled());
902 DCHECK(!IsInZygoteExecSpace(method->GetEntryPointFromQuickCompiledCode()));
903 }
904 }
905 // Not strictly necessary, but this map is useless now.
906 saved_compiled_methods_map_.clear();
907 }
908 if (kIsDebugBuild) {
909 for (const auto& entry : zygote_map_) {
910 ArtMethod* method = entry.method;
911 if (method != nullptr) {
912 DCHECK(!method->IsPreCompiled());
913 DCHECK(!IsInZygoteExecSpace(method->GetEntryPointFromQuickCompiledCode()));
914 }
915 }
916 }
917 }
918
CodeCacheSizeLocked()919 size_t JitCodeCache::CodeCacheSizeLocked() {
920 return GetCurrentRegion()->GetUsedMemoryForCode();
921 }
922
DataCacheSize()923 size_t JitCodeCache::DataCacheSize() {
924 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
925 return DataCacheSizeLocked();
926 }
927
DataCacheSizeLocked()928 size_t JitCodeCache::DataCacheSizeLocked() {
929 return GetCurrentRegion()->GetUsedMemoryForData();
930 }
931
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)932 bool JitCodeCache::Reserve(Thread* self,
933 JitMemoryRegion* region,
934 size_t code_size,
935 size_t stack_map_size,
936 size_t number_of_roots,
937 ArtMethod* method,
938 /*out*/ArrayRef<const uint8_t>* reserved_code,
939 /*out*/ArrayRef<const uint8_t>* reserved_data) {
940 code_size = OatQuickMethodHeader::InstructionAlignedSize() + code_size;
941 size_t data_size = RoundUp(ComputeRootTableSize(number_of_roots) + stack_map_size, sizeof(void*));
942
943 const uint8_t* code;
944 const uint8_t* data;
945 while (true) {
946 bool at_max_capacity = false;
947 {
948 ScopedThreadSuspension sts(self, kSuspended);
949 MutexLock mu(self, *Locks::jit_lock_);
950 WaitForPotentialCollectionToComplete(self);
951 ScopedCodeCacheWrite ccw(*region);
952 code = region->AllocateCode(code_size);
953 data = region->AllocateData(data_size);
954 at_max_capacity = IsAtMaxCapacity();
955 }
956 if (code != nullptr && data != nullptr) {
957 break;
958 }
959 Free(self, region, code, data);
960 if (at_max_capacity) {
961 VLOG(jit) << "JIT failed to allocate code of size "
962 << PrettySize(code_size)
963 << ", and data of size "
964 << PrettySize(data_size);
965 return false;
966 }
967 // Run a code cache collection and try again.
968 GarbageCollectCache(self);
969 }
970
971 *reserved_code = ArrayRef<const uint8_t>(code, code_size);
972 *reserved_data = ArrayRef<const uint8_t>(data, data_size);
973
974 MutexLock mu(self, *Locks::jit_lock_);
975 histogram_code_memory_use_.AddValue(code_size);
976 if (code_size > kCodeSizeLogThreshold) {
977 LOG(INFO) << "JIT allocated "
978 << PrettySize(code_size)
979 << " for compiled code of "
980 << ArtMethod::PrettyMethod(method);
981 }
982 histogram_stack_map_memory_use_.AddValue(data_size);
983 if (data_size > kStackMapSizeLogThreshold) {
984 LOG(INFO) << "JIT allocated "
985 << PrettySize(data_size)
986 << " for stack maps of "
987 << ArtMethod::PrettyMethod(method);
988 }
989 return true;
990 }
991
Free(Thread * self,JitMemoryRegion * region,const uint8_t * code,const uint8_t * data)992 void JitCodeCache::Free(Thread* self,
993 JitMemoryRegion* region,
994 const uint8_t* code,
995 const uint8_t* data) {
996 MutexLock mu(self, *Locks::jit_lock_);
997 ScopedCodeCacheWrite ccw(*region);
998 FreeLocked(region, code, data);
999 }
1000
FreeLocked(JitMemoryRegion * region,const uint8_t * code,const uint8_t * data)1001 void JitCodeCache::FreeLocked(JitMemoryRegion* region, const uint8_t* code, const uint8_t* data) {
1002 if (code != nullptr) {
1003 RemoveNativeDebugInfoForJit(reinterpret_cast<const void*>(FromAllocationToCode(code)));
1004 region->FreeCode(code);
1005 }
1006 if (data != nullptr) {
1007 region->FreeData(data);
1008 }
1009 }
1010
1011 class MarkCodeClosure final : public Closure {
1012 public:
MarkCodeClosure(JitCodeCache * code_cache,CodeCacheBitmap * bitmap,Barrier * barrier)1013 MarkCodeClosure(JitCodeCache* code_cache, CodeCacheBitmap* bitmap, Barrier* barrier)
1014 : code_cache_(code_cache), bitmap_(bitmap), barrier_(barrier) {}
1015
Run(Thread * thread)1016 void Run(Thread* thread) override REQUIRES_SHARED(Locks::mutator_lock_) {
1017 ScopedTrace trace(__PRETTY_FUNCTION__);
1018 DCHECK(thread == Thread::Current() || thread->IsSuspended());
1019 StackVisitor::WalkStack(
1020 [&](const art::StackVisitor* stack_visitor) {
1021 const OatQuickMethodHeader* method_header =
1022 stack_visitor->GetCurrentOatQuickMethodHeader();
1023 if (method_header == nullptr) {
1024 return true;
1025 }
1026 const void* code = method_header->GetCode();
1027 if (code_cache_->ContainsPc(code) && !code_cache_->IsInZygoteExecSpace(code)) {
1028 // Use the atomic set version, as multiple threads are executing this code.
1029 bitmap_->AtomicTestAndSet(FromCodeToAllocation(code));
1030 }
1031 return true;
1032 },
1033 thread,
1034 /* context= */ nullptr,
1035 art::StackVisitor::StackWalkKind::kSkipInlinedFrames);
1036
1037 if (kIsDebugBuild) {
1038 // The stack walking code queries the side instrumentation stack if it
1039 // sees an instrumentation exit pc, so the JIT code of methods in that stack
1040 // must have been seen. We check this below.
1041 for (const auto& it : *thread->GetInstrumentationStack()) {
1042 // The 'method_' in InstrumentationStackFrame is the one that has return_pc_ in
1043 // its stack frame, it is not the method owning return_pc_. We just pass null to
1044 // LookupMethodHeader: the method is only checked against in debug builds.
1045 OatQuickMethodHeader* method_header =
1046 code_cache_->LookupMethodHeader(it.second.return_pc_, /* method= */ nullptr);
1047 if (method_header != nullptr) {
1048 const void* code = method_header->GetCode();
1049 CHECK(bitmap_->Test(FromCodeToAllocation(code)));
1050 }
1051 }
1052 }
1053 barrier_->Pass(Thread::Current());
1054 }
1055
1056 private:
1057 JitCodeCache* const code_cache_;
1058 CodeCacheBitmap* const bitmap_;
1059 Barrier* const barrier_;
1060 };
1061
NotifyCollectionDone(Thread * self)1062 void JitCodeCache::NotifyCollectionDone(Thread* self) {
1063 collection_in_progress_ = false;
1064 lock_cond_.Broadcast(self);
1065 }
1066
MarkCompiledCodeOnThreadStacks(Thread * self)1067 void JitCodeCache::MarkCompiledCodeOnThreadStacks(Thread* self) {
1068 Barrier barrier(0);
1069 size_t threads_running_checkpoint = 0;
1070 MarkCodeClosure closure(this, GetLiveBitmap(), &barrier);
1071 threads_running_checkpoint = Runtime::Current()->GetThreadList()->RunCheckpoint(&closure);
1072 // Now that we have run our checkpoint, move to a suspended state and wait
1073 // for other threads to run the checkpoint.
1074 ScopedThreadSuspension sts(self, kSuspended);
1075 if (threads_running_checkpoint != 0) {
1076 barrier.Increment(self, threads_running_checkpoint);
1077 }
1078 }
1079
IsAtMaxCapacity() const1080 bool JitCodeCache::IsAtMaxCapacity() const {
1081 return private_region_.GetCurrentCapacity() == private_region_.GetMaxCapacity();
1082 }
1083
ShouldDoFullCollection()1084 bool JitCodeCache::ShouldDoFullCollection() {
1085 if (IsAtMaxCapacity()) {
1086 // Always do a full collection when the code cache is full.
1087 return true;
1088 } else if (private_region_.GetCurrentCapacity() < kReservedCapacity) {
1089 // Always do partial collection when the code cache size is below the reserved
1090 // capacity.
1091 return false;
1092 } else if (last_collection_increased_code_cache_) {
1093 // This time do a full collection.
1094 return true;
1095 } else {
1096 // This time do a partial collection.
1097 return false;
1098 }
1099 }
1100
GarbageCollectCache(Thread * self)1101 void JitCodeCache::GarbageCollectCache(Thread* self) {
1102 ScopedTrace trace(__FUNCTION__);
1103 // Wait for an existing collection, or let everyone know we are starting one.
1104 {
1105 ScopedThreadSuspension sts(self, kSuspended);
1106 MutexLock mu(self, *Locks::jit_lock_);
1107 if (!garbage_collect_code_) {
1108 private_region_.IncreaseCodeCacheCapacity();
1109 return;
1110 } else if (WaitForPotentialCollectionToComplete(self)) {
1111 return;
1112 } else {
1113 number_of_collections_++;
1114 live_bitmap_.reset(CodeCacheBitmap::Create(
1115 "code-cache-bitmap",
1116 reinterpret_cast<uintptr_t>(private_region_.GetExecPages()->Begin()),
1117 reinterpret_cast<uintptr_t>(
1118 private_region_.GetExecPages()->Begin() + private_region_.GetCurrentCapacity() / 2)));
1119 collection_in_progress_ = true;
1120 }
1121 }
1122
1123 TimingLogger logger("JIT code cache timing logger", true, VLOG_IS_ON(jit));
1124 {
1125 TimingLogger::ScopedTiming st("Code cache collection", &logger);
1126
1127 bool do_full_collection = false;
1128 {
1129 MutexLock mu(self, *Locks::jit_lock_);
1130 do_full_collection = ShouldDoFullCollection();
1131 }
1132
1133 VLOG(jit) << "Do "
1134 << (do_full_collection ? "full" : "partial")
1135 << " code cache collection, code="
1136 << PrettySize(CodeCacheSize())
1137 << ", data=" << PrettySize(DataCacheSize());
1138
1139 DoCollection(self, /* collect_profiling_info= */ do_full_collection);
1140
1141 VLOG(jit) << "After code cache collection, code="
1142 << PrettySize(CodeCacheSize())
1143 << ", data=" << PrettySize(DataCacheSize());
1144
1145 {
1146 MutexLock mu(self, *Locks::jit_lock_);
1147
1148 // Increase the code cache only when we do partial collections.
1149 // TODO: base this strategy on how full the code cache is?
1150 if (do_full_collection) {
1151 last_collection_increased_code_cache_ = false;
1152 } else {
1153 last_collection_increased_code_cache_ = true;
1154 private_region_.IncreaseCodeCacheCapacity();
1155 }
1156
1157 bool next_collection_will_be_full = ShouldDoFullCollection();
1158
1159 // Start polling the liveness of compiled code to prepare for the next full collection.
1160 if (next_collection_will_be_full) {
1161 for (auto it : profiling_infos_) {
1162 it.second->SetBaselineHotnessCount(0);
1163 }
1164
1165 // Change entry points of native methods back to the GenericJNI entrypoint.
1166 for (const auto& entry : jni_stubs_map_) {
1167 const JniStubData& data = entry.second;
1168 if (!data.IsCompiled() || IsInZygoteExecSpace(data.GetCode())) {
1169 continue;
1170 }
1171 // Make sure a single invocation of the GenericJNI trampoline tries to recompile.
1172 uint16_t new_counter = Runtime::Current()->GetJit()->HotMethodThreshold() - 1u;
1173 const OatQuickMethodHeader* method_header =
1174 OatQuickMethodHeader::FromCodePointer(data.GetCode());
1175 for (ArtMethod* method : data.GetMethods()) {
1176 if (method->GetEntryPointFromQuickCompiledCode() == method_header->GetEntryPoint()) {
1177 // Don't call Instrumentation::UpdateMethodsCode(), same as for normal methods above.
1178 method->SetCounter(new_counter);
1179 method->SetEntryPointFromQuickCompiledCode(GetQuickGenericJniStub());
1180 }
1181 }
1182 }
1183 }
1184 live_bitmap_.reset(nullptr);
1185 NotifyCollectionDone(self);
1186 }
1187 }
1188 Runtime::Current()->GetJit()->AddTimingLogger(logger);
1189 }
1190
RemoveUnmarkedCode(Thread * self)1191 void JitCodeCache::RemoveUnmarkedCode(Thread* self) {
1192 ScopedTrace trace(__FUNCTION__);
1193 std::unordered_set<OatQuickMethodHeader*> method_headers;
1194 {
1195 MutexLock mu(self, *Locks::jit_lock_);
1196 // Iterate over all compiled code and remove entries that are not marked.
1197 for (auto it = jni_stubs_map_.begin(); it != jni_stubs_map_.end();) {
1198 JniStubData* data = &it->second;
1199 if (IsInZygoteExecSpace(data->GetCode()) ||
1200 !data->IsCompiled() ||
1201 GetLiveBitmap()->Test(FromCodeToAllocation(data->GetCode()))) {
1202 ++it;
1203 } else {
1204 method_headers.insert(OatQuickMethodHeader::FromCodePointer(data->GetCode()));
1205 for (ArtMethod* method : data->GetMethods()) {
1206 VLOG(jit) << "JIT removed (JNI) " << method->PrettyMethod() << ": " << data->GetCode();
1207 }
1208 it = jni_stubs_map_.erase(it);
1209 }
1210 }
1211 for (auto it = method_code_map_.begin(); it != method_code_map_.end();) {
1212 const void* code_ptr = it->first;
1213 uintptr_t allocation = FromCodeToAllocation(code_ptr);
1214 if (IsInZygoteExecSpace(code_ptr) || GetLiveBitmap()->Test(allocation)) {
1215 ++it;
1216 } else {
1217 OatQuickMethodHeader* header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1218 method_headers.insert(header);
1219 VLOG(jit) << "JIT removed " << it->second->PrettyMethod() << ": " << it->first;
1220 it = method_code_map_.erase(it);
1221 }
1222 }
1223 FreeAllMethodHeaders(method_headers);
1224 }
1225 }
1226
GetGarbageCollectCode()1227 bool JitCodeCache::GetGarbageCollectCode() {
1228 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1229 return garbage_collect_code_;
1230 }
1231
SetGarbageCollectCode(bool value)1232 void JitCodeCache::SetGarbageCollectCode(bool value) {
1233 Thread* self = Thread::Current();
1234 MutexLock mu(self, *Locks::jit_lock_);
1235 // Update the flag while holding the lock to ensure no thread will try to GC.
1236 garbage_collect_code_ = value;
1237 }
1238
RemoveMethodBeingCompiled(ArtMethod * method,CompilationKind kind)1239 void JitCodeCache::RemoveMethodBeingCompiled(ArtMethod* method, CompilationKind kind) {
1240 DCHECK(IsMethodBeingCompiled(method, kind));
1241 switch (kind) {
1242 case CompilationKind::kOsr:
1243 current_osr_compilations_.erase(method);
1244 break;
1245 case CompilationKind::kBaseline:
1246 current_baseline_compilations_.erase(method);
1247 break;
1248 case CompilationKind::kOptimized:
1249 current_optimized_compilations_.erase(method);
1250 break;
1251 }
1252 }
1253
AddMethodBeingCompiled(ArtMethod * method,CompilationKind kind)1254 void JitCodeCache::AddMethodBeingCompiled(ArtMethod* method, CompilationKind kind) {
1255 DCHECK(!IsMethodBeingCompiled(method, kind));
1256 switch (kind) {
1257 case CompilationKind::kOsr:
1258 current_osr_compilations_.insert(method);
1259 break;
1260 case CompilationKind::kBaseline:
1261 current_baseline_compilations_.insert(method);
1262 break;
1263 case CompilationKind::kOptimized:
1264 current_optimized_compilations_.insert(method);
1265 break;
1266 }
1267 }
1268
IsMethodBeingCompiled(ArtMethod * method,CompilationKind kind)1269 bool JitCodeCache::IsMethodBeingCompiled(ArtMethod* method, CompilationKind kind) {
1270 switch (kind) {
1271 case CompilationKind::kOsr:
1272 return ContainsElement(current_osr_compilations_, method);
1273 case CompilationKind::kBaseline:
1274 return ContainsElement(current_baseline_compilations_, method);
1275 case CompilationKind::kOptimized:
1276 return ContainsElement(current_optimized_compilations_, method);
1277 }
1278 }
1279
IsMethodBeingCompiled(ArtMethod * method)1280 bool JitCodeCache::IsMethodBeingCompiled(ArtMethod* method) {
1281 return ContainsElement(current_optimized_compilations_, method) ||
1282 ContainsElement(current_osr_compilations_, method) ||
1283 ContainsElement(current_baseline_compilations_, method);
1284 }
1285
DoCollection(Thread * self,bool collect_profiling_info)1286 void JitCodeCache::DoCollection(Thread* self, bool collect_profiling_info) {
1287 ScopedTrace trace(__FUNCTION__);
1288 {
1289 MutexLock mu(self, *Locks::jit_lock_);
1290
1291 // Update to interpreter the methods that have baseline entrypoints and whose baseline
1292 // hotness count is zero.
1293 // Note that these methods may be in thread stack or concurrently revived
1294 // between. That's OK, as the thread executing it will mark it.
1295 for (auto it : profiling_infos_) {
1296 ProfilingInfo* info = it.second;
1297 if (info->GetBaselineHotnessCount() == 0) {
1298 const void* entry_point = info->GetMethod()->GetEntryPointFromQuickCompiledCode();
1299 if (ContainsPc(entry_point)) {
1300 OatQuickMethodHeader* method_header =
1301 OatQuickMethodHeader::FromEntryPoint(entry_point);
1302 if (CodeInfo::IsBaseline(method_header->GetOptimizedCodeInfoPtr())) {
1303 info->GetMethod()->SetEntryPointFromQuickCompiledCode(GetQuickToInterpreterBridge());
1304 }
1305 }
1306 }
1307 }
1308 // TODO: collect profiling info
1309 // TODO: collect optimized code
1310
1311 // Mark compiled code that are entrypoints of ArtMethods. Compiled code that is not
1312 // an entry point is either:
1313 // - an osr compiled code, that will be removed if not in a thread call stack.
1314 // - discarded compiled code, that will be removed if not in a thread call stack.
1315 for (const auto& entry : jni_stubs_map_) {
1316 const JniStubData& data = entry.second;
1317 const void* code_ptr = data.GetCode();
1318 if (IsInZygoteExecSpace(code_ptr)) {
1319 continue;
1320 }
1321 const OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1322 for (ArtMethod* method : data.GetMethods()) {
1323 if (method_header->GetEntryPoint() == method->GetEntryPointFromQuickCompiledCode()) {
1324 GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(code_ptr));
1325 break;
1326 }
1327 }
1328 }
1329 for (const auto& it : method_code_map_) {
1330 ArtMethod* method = it.second;
1331 const void* code_ptr = it.first;
1332 if (IsInZygoteExecSpace(code_ptr)) {
1333 continue;
1334 }
1335 const OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1336 if (method_header->GetEntryPoint() == method->GetEntryPointFromQuickCompiledCode()) {
1337 GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(code_ptr));
1338 }
1339 }
1340
1341 // Empty osr method map, as osr compiled code will be deleted (except the ones
1342 // on thread stacks).
1343 osr_code_map_.clear();
1344 }
1345
1346 // Run a checkpoint on all threads to mark the JIT compiled code they are running.
1347 MarkCompiledCodeOnThreadStacks(self);
1348
1349 // At this point, mutator threads are still running, and entrypoints of methods can
1350 // change. We do know they cannot change to a code cache entry that is not marked,
1351 // therefore we can safely remove those entries.
1352 RemoveUnmarkedCode(self);
1353
1354 if (collect_profiling_info) {
1355 // TODO: Collect unused profiling infos.
1356 }
1357 }
1358
LookupMethodHeader(uintptr_t pc,ArtMethod * method)1359 OatQuickMethodHeader* JitCodeCache::LookupMethodHeader(uintptr_t pc, ArtMethod* method) {
1360 static_assert(kRuntimeISA != InstructionSet::kThumb2, "kThumb2 cannot be a runtime ISA");
1361 if (kRuntimeISA == InstructionSet::kArm) {
1362 // On Thumb-2, the pc is offset by one.
1363 --pc;
1364 }
1365 if (!ContainsPc(reinterpret_cast<const void*>(pc))) {
1366 return nullptr;
1367 }
1368
1369 if (!kIsDebugBuild) {
1370 // Called with null `method` only from MarkCodeClosure::Run() in debug build.
1371 CHECK(method != nullptr);
1372 }
1373
1374 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1375 OatQuickMethodHeader* method_header = nullptr;
1376 ArtMethod* found_method = nullptr; // Only for DCHECK(), not for JNI stubs.
1377 if (method != nullptr && UNLIKELY(method->IsNative())) {
1378 auto it = jni_stubs_map_.find(JniStubKey(method));
1379 if (it == jni_stubs_map_.end() || !ContainsElement(it->second.GetMethods(), method)) {
1380 return nullptr;
1381 }
1382 const void* code_ptr = it->second.GetCode();
1383 method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1384 if (!method_header->Contains(pc)) {
1385 return nullptr;
1386 }
1387 } else {
1388 if (shared_region_.IsInExecSpace(reinterpret_cast<const void*>(pc))) {
1389 const void* code_ptr = zygote_map_.GetCodeFor(method, pc);
1390 if (code_ptr != nullptr) {
1391 return OatQuickMethodHeader::FromCodePointer(code_ptr);
1392 }
1393 }
1394 auto it = method_code_map_.lower_bound(reinterpret_cast<const void*>(pc));
1395 if (it != method_code_map_.begin()) {
1396 --it;
1397 const void* code_ptr = it->first;
1398 if (OatQuickMethodHeader::FromCodePointer(code_ptr)->Contains(pc)) {
1399 method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1400 found_method = it->second;
1401 }
1402 }
1403 if (method_header == nullptr && method == nullptr) {
1404 // Scan all compiled JNI stubs as well. This slow search is used only
1405 // for checks in debug build, for release builds the `method` is not null.
1406 for (auto&& entry : jni_stubs_map_) {
1407 const JniStubData& data = entry.second;
1408 if (data.IsCompiled() &&
1409 OatQuickMethodHeader::FromCodePointer(data.GetCode())->Contains(pc)) {
1410 method_header = OatQuickMethodHeader::FromCodePointer(data.GetCode());
1411 }
1412 }
1413 }
1414 if (method_header == nullptr) {
1415 return nullptr;
1416 }
1417 }
1418
1419 if (kIsDebugBuild && method != nullptr && !method->IsNative()) {
1420 DCHECK_EQ(found_method, method)
1421 << ArtMethod::PrettyMethod(method) << " "
1422 << ArtMethod::PrettyMethod(found_method) << " "
1423 << std::hex << pc;
1424 }
1425 return method_header;
1426 }
1427
LookupOsrMethodHeader(ArtMethod * method)1428 OatQuickMethodHeader* JitCodeCache::LookupOsrMethodHeader(ArtMethod* method) {
1429 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1430 auto it = osr_code_map_.find(method);
1431 if (it == osr_code_map_.end()) {
1432 return nullptr;
1433 }
1434 return OatQuickMethodHeader::FromCodePointer(it->second);
1435 }
1436
AddProfilingInfo(Thread * self,ArtMethod * method,const std::vector<uint32_t> & entries)1437 ProfilingInfo* JitCodeCache::AddProfilingInfo(Thread* self,
1438 ArtMethod* method,
1439 const std::vector<uint32_t>& entries) {
1440 DCHECK(CanAllocateProfilingInfo());
1441 ProfilingInfo* info = nullptr;
1442 {
1443 MutexLock mu(self, *Locks::jit_lock_);
1444 info = AddProfilingInfoInternal(self, method, entries);
1445 }
1446
1447 if (info == nullptr) {
1448 GarbageCollectCache(self);
1449 MutexLock mu(self, *Locks::jit_lock_);
1450 info = AddProfilingInfoInternal(self, method, entries);
1451 }
1452 return info;
1453 }
1454
AddProfilingInfoInternal(Thread * self ATTRIBUTE_UNUSED,ArtMethod * method,const std::vector<uint32_t> & entries)1455 ProfilingInfo* JitCodeCache::AddProfilingInfoInternal(Thread* self ATTRIBUTE_UNUSED,
1456 ArtMethod* method,
1457 const std::vector<uint32_t>& entries) {
1458 // Check whether some other thread has concurrently created it.
1459 auto it = profiling_infos_.find(method);
1460 if (it != profiling_infos_.end()) {
1461 return it->second;
1462 }
1463
1464 size_t profile_info_size = RoundUp(
1465 sizeof(ProfilingInfo) + sizeof(InlineCache) * entries.size(),
1466 sizeof(void*));
1467
1468 const uint8_t* data = private_region_.AllocateData(profile_info_size);
1469 if (data == nullptr) {
1470 return nullptr;
1471 }
1472 uint8_t* writable_data = private_region_.GetWritableDataAddress(data);
1473 ProfilingInfo* info = new (writable_data) ProfilingInfo(method, entries);
1474
1475 profiling_infos_.Put(method, info);
1476 histogram_profiling_info_memory_use_.AddValue(profile_info_size);
1477 return info;
1478 }
1479
MoreCore(const void * mspace,intptr_t increment)1480 void* JitCodeCache::MoreCore(const void* mspace, intptr_t increment) {
1481 return shared_region_.OwnsSpace(mspace)
1482 ? shared_region_.MoreCore(mspace, increment)
1483 : private_region_.MoreCore(mspace, increment);
1484 }
1485
GetProfiledMethods(const std::set<std::string> & dex_base_locations,std::vector<ProfileMethodInfo> & methods)1486 void JitCodeCache::GetProfiledMethods(const std::set<std::string>& dex_base_locations,
1487 std::vector<ProfileMethodInfo>& methods) {
1488 Thread* self = Thread::Current();
1489 WaitUntilInlineCacheAccessible(self);
1490 MutexLock mu(self, *Locks::jit_lock_);
1491 ScopedTrace trace(__FUNCTION__);
1492 uint16_t jit_compile_threshold = Runtime::Current()->GetJITOptions()->GetCompileThreshold();
1493 for (auto it : profiling_infos_) {
1494 ProfilingInfo* info = it.second;
1495 ArtMethod* method = info->GetMethod();
1496 const DexFile* dex_file = method->GetDexFile();
1497 const std::string base_location = DexFileLoader::GetBaseLocation(dex_file->GetLocation());
1498 if (!ContainsElement(dex_base_locations, base_location)) {
1499 // Skip dex files which are not profiled.
1500 continue;
1501 }
1502 std::vector<ProfileMethodInfo::ProfileInlineCache> inline_caches;
1503
1504 // If the method didn't reach the compilation threshold don't save the inline caches.
1505 // They might be incomplete and cause unnecessary deoptimizations.
1506 // If the inline cache is empty the compiler will generate a regular invoke virtual/interface.
1507 if (method->GetCounter() < jit_compile_threshold) {
1508 methods.emplace_back(/*ProfileMethodInfo*/
1509 MethodReference(dex_file, method->GetDexMethodIndex()), inline_caches);
1510 continue;
1511 }
1512
1513 for (size_t i = 0; i < info->number_of_inline_caches_; ++i) {
1514 std::vector<TypeReference> profile_classes;
1515 const InlineCache& cache = info->cache_[i];
1516 ArtMethod* caller = info->GetMethod();
1517 bool is_missing_types = false;
1518 for (size_t k = 0; k < InlineCache::kIndividualCacheSize; k++) {
1519 mirror::Class* cls = cache.classes_[k].Read();
1520 if (cls == nullptr) {
1521 break;
1522 }
1523
1524 // Check if the receiver is in the boot class path or if it's in the
1525 // same class loader as the caller. If not, skip it, as there is not
1526 // much we can do during AOT.
1527 if (!cls->IsBootStrapClassLoaded() &&
1528 caller->GetClassLoader() != cls->GetClassLoader()) {
1529 is_missing_types = true;
1530 continue;
1531 }
1532
1533 const DexFile* class_dex_file = nullptr;
1534 dex::TypeIndex type_index;
1535
1536 if (cls->GetDexCache() == nullptr) {
1537 DCHECK(cls->IsArrayClass()) << cls->PrettyClass();
1538 // Make a best effort to find the type index in the method's dex file.
1539 // We could search all open dex files but that might turn expensive
1540 // and probably not worth it.
1541 class_dex_file = dex_file;
1542 type_index = cls->FindTypeIndexInOtherDexFile(*dex_file);
1543 } else {
1544 class_dex_file = &(cls->GetDexFile());
1545 type_index = cls->GetDexTypeIndex();
1546 }
1547 if (!type_index.IsValid()) {
1548 // Could be a proxy class or an array for which we couldn't find the type index.
1549 is_missing_types = true;
1550 continue;
1551 }
1552 if (ContainsElement(dex_base_locations,
1553 DexFileLoader::GetBaseLocation(class_dex_file->GetLocation()))) {
1554 // Only consider classes from the same apk (including multidex).
1555 profile_classes.emplace_back(/*ProfileMethodInfo::ProfileClassReference*/
1556 class_dex_file, type_index);
1557 } else {
1558 is_missing_types = true;
1559 }
1560 }
1561 if (!profile_classes.empty()) {
1562 inline_caches.emplace_back(/*ProfileMethodInfo::ProfileInlineCache*/
1563 cache.dex_pc_, is_missing_types, profile_classes);
1564 }
1565 }
1566 methods.emplace_back(/*ProfileMethodInfo*/
1567 MethodReference(dex_file, method->GetDexMethodIndex()), inline_caches);
1568 }
1569 }
1570
IsOsrCompiled(ArtMethod * method)1571 bool JitCodeCache::IsOsrCompiled(ArtMethod* method) {
1572 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1573 return osr_code_map_.find(method) != osr_code_map_.end();
1574 }
1575
NotifyCompilationOf(ArtMethod * method,Thread * self,CompilationKind compilation_kind,bool prejit)1576 bool JitCodeCache::NotifyCompilationOf(ArtMethod* method,
1577 Thread* self,
1578 CompilationKind compilation_kind,
1579 bool prejit) {
1580 const void* existing_entry_point = method->GetEntryPointFromQuickCompiledCode();
1581 if (compilation_kind != CompilationKind::kOsr && ContainsPc(existing_entry_point)) {
1582 OatQuickMethodHeader* method_header =
1583 OatQuickMethodHeader::FromEntryPoint(existing_entry_point);
1584 bool is_baseline = (compilation_kind == CompilationKind::kBaseline);
1585 if (CodeInfo::IsBaseline(method_header->GetOptimizedCodeInfoPtr()) == is_baseline) {
1586 VLOG(jit) << "Not compiling "
1587 << method->PrettyMethod()
1588 << " because it has already been compiled"
1589 << " kind=" << compilation_kind;
1590 return false;
1591 }
1592 }
1593
1594 if (NeedsClinitCheckBeforeCall(method) && !prejit) {
1595 // We do not need a synchronization barrier for checking the visibly initialized status
1596 // or checking the initialized status just for requesting visible initialization.
1597 ClassStatus status = method->GetDeclaringClass()
1598 ->GetStatus<kDefaultVerifyFlags, /*kWithSynchronizationBarrier=*/ false>();
1599 if (status != ClassStatus::kVisiblyInitialized) {
1600 // Unless we're pre-jitting, we currently don't save the JIT compiled code if we cannot
1601 // update the entrypoint due to needing an initialization check.
1602 if (status == ClassStatus::kInitialized) {
1603 // Request visible initialization but do not block to allow compiling other methods.
1604 // Hopefully, this will complete by the time the method becomes hot again.
1605 Runtime::Current()->GetClassLinker()->MakeInitializedClassesVisiblyInitialized(
1606 self, /*wait=*/ false);
1607 }
1608 VLOG(jit) << "Not compiling "
1609 << method->PrettyMethod()
1610 << " because it has the resolution stub";
1611 // Give it a new chance to be hot.
1612 ClearMethodCounter(method, /*was_warm=*/ false);
1613 return false;
1614 }
1615 }
1616
1617 if (compilation_kind == CompilationKind::kOsr) {
1618 MutexLock mu(self, *Locks::jit_lock_);
1619 if (osr_code_map_.find(method) != osr_code_map_.end()) {
1620 return false;
1621 }
1622 }
1623
1624 if (UNLIKELY(method->IsNative())) {
1625 MutexLock mu(self, *Locks::jit_lock_);
1626 JniStubKey key(method);
1627 auto it = jni_stubs_map_.find(key);
1628 bool new_compilation = false;
1629 if (it == jni_stubs_map_.end()) {
1630 // Create a new entry to mark the stub as being compiled.
1631 it = jni_stubs_map_.Put(key, JniStubData{});
1632 new_compilation = true;
1633 }
1634 JniStubData* data = &it->second;
1635 data->AddMethod(method);
1636 if (data->IsCompiled()) {
1637 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(data->GetCode());
1638 const void* entrypoint = method_header->GetEntryPoint();
1639 // Update also entrypoints of other methods held by the JniStubData.
1640 // We could simply update the entrypoint of `method` but if the last JIT GC has
1641 // changed these entrypoints to GenericJNI in preparation for a full GC, we may
1642 // as well change them back as this stub shall not be collected anyway and this
1643 // can avoid a few expensive GenericJNI calls.
1644 data->UpdateEntryPoints(entrypoint);
1645 if (collection_in_progress_) {
1646 if (!IsInZygoteExecSpace(data->GetCode())) {
1647 GetLiveBitmap()->AtomicTestAndSet(FromCodeToAllocation(data->GetCode()));
1648 }
1649 }
1650 }
1651 return new_compilation;
1652 } else {
1653 if (CanAllocateProfilingInfo() && (compilation_kind == CompilationKind::kBaseline)) {
1654 bool has_profiling_info = false;
1655 {
1656 MutexLock mu(self, *Locks::jit_lock_);
1657 has_profiling_info = (profiling_infos_.find(method) != profiling_infos_.end());
1658 }
1659 if (!has_profiling_info) {
1660 if (ProfilingInfo::Create(self, method) == nullptr) {
1661 VLOG(jit) << method->PrettyMethod() << " needs a ProfilingInfo to be compiled baseline";
1662 ClearMethodCounter(method, /*was_warm=*/ false);
1663 return false;
1664 }
1665 }
1666 }
1667 MutexLock mu(self, *Locks::jit_lock_);
1668 if (IsMethodBeingCompiled(method, compilation_kind)) {
1669 return false;
1670 }
1671 AddMethodBeingCompiled(method, compilation_kind);
1672 return true;
1673 }
1674 }
1675
NotifyCompilerUse(ArtMethod * method,Thread * self)1676 ProfilingInfo* JitCodeCache::NotifyCompilerUse(ArtMethod* method, Thread* self) {
1677 MutexLock mu(self, *Locks::jit_lock_);
1678 auto it = profiling_infos_.find(method);
1679 if (it == profiling_infos_.end()) {
1680 return nullptr;
1681 }
1682 if (!it->second->IncrementInlineUse()) {
1683 // Overflow of inlining uses, just bail.
1684 return nullptr;
1685 }
1686 return it->second;
1687 }
1688
DoneCompilerUse(ArtMethod * method,Thread * self)1689 void JitCodeCache::DoneCompilerUse(ArtMethod* method, Thread* self) {
1690 MutexLock mu(self, *Locks::jit_lock_);
1691 auto it = profiling_infos_.find(method);
1692 DCHECK(it != profiling_infos_.end());
1693 it->second->DecrementInlineUse();
1694 }
1695
DoneCompiling(ArtMethod * method,Thread * self,CompilationKind compilation_kind)1696 void JitCodeCache::DoneCompiling(ArtMethod* method,
1697 Thread* self,
1698 CompilationKind compilation_kind) {
1699 DCHECK_EQ(Thread::Current(), self);
1700 MutexLock mu(self, *Locks::jit_lock_);
1701 if (UNLIKELY(method->IsNative())) {
1702 auto it = jni_stubs_map_.find(JniStubKey(method));
1703 DCHECK(it != jni_stubs_map_.end());
1704 JniStubData* data = &it->second;
1705 DCHECK(ContainsElement(data->GetMethods(), method));
1706 if (UNLIKELY(!data->IsCompiled())) {
1707 // Failed to compile; the JNI compiler never fails, but the cache may be full.
1708 jni_stubs_map_.erase(it); // Remove the entry added in NotifyCompilationOf().
1709 } // else Commit() updated entrypoints of all methods in the JniStubData.
1710 } else {
1711 RemoveMethodBeingCompiled(method, compilation_kind);
1712 }
1713 }
1714
InvalidateAllCompiledCode()1715 void JitCodeCache::InvalidateAllCompiledCode() {
1716 art::MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1717 VLOG(jit) << "Invalidating all compiled code";
1718 ClassLinker* linker = Runtime::Current()->GetClassLinker();
1719 for (auto it : method_code_map_) {
1720 ArtMethod* meth = it.second;
1721 // We were compiled, so we must be warm.
1722 ClearMethodCounter(meth, /*was_warm=*/true);
1723 if (meth->IsObsolete()) {
1724 linker->SetEntryPointsForObsoleteMethod(meth);
1725 } else {
1726 linker->SetEntryPointsToInterpreter(meth);
1727 }
1728 }
1729 saved_compiled_methods_map_.clear();
1730 osr_code_map_.clear();
1731 }
1732
InvalidateCompiledCodeFor(ArtMethod * method,const OatQuickMethodHeader * header)1733 void JitCodeCache::InvalidateCompiledCodeFor(ArtMethod* method,
1734 const OatQuickMethodHeader* header) {
1735 DCHECK(!method->IsNative());
1736 const void* method_entrypoint = method->GetEntryPointFromQuickCompiledCode();
1737
1738 // Clear the method counter if we are running jitted code since we might want to jit this again in
1739 // the future.
1740 if (method_entrypoint == header->GetEntryPoint()) {
1741 // The entrypoint is the one to invalidate, so we just update it to the interpreter entry point
1742 // and clear the counter to get the method Jitted again.
1743 Runtime::Current()->GetInstrumentation()->UpdateMethodsCode(
1744 method, GetQuickToInterpreterBridge());
1745 ClearMethodCounter(method, /*was_warm=*/ true);
1746 } else {
1747 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1748 auto it = osr_code_map_.find(method);
1749 if (it != osr_code_map_.end() && OatQuickMethodHeader::FromCodePointer(it->second) == header) {
1750 // Remove the OSR method, to avoid using it again.
1751 osr_code_map_.erase(it);
1752 }
1753 }
1754
1755 // In case the method was pre-compiled, clear that information so we
1756 // can recompile it ourselves.
1757 if (method->IsPreCompiled()) {
1758 method->ClearPreCompiled();
1759 }
1760 }
1761
Dump(std::ostream & os)1762 void JitCodeCache::Dump(std::ostream& os) {
1763 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1764 os << "Current JIT code cache size (used / resident): "
1765 << GetCurrentRegion()->GetUsedMemoryForCode() / KB << "KB / "
1766 << GetCurrentRegion()->GetResidentMemoryForCode() / KB << "KB\n"
1767 << "Current JIT data cache size (used / resident): "
1768 << GetCurrentRegion()->GetUsedMemoryForData() / KB << "KB / "
1769 << GetCurrentRegion()->GetResidentMemoryForData() / KB << "KB\n";
1770 if (!Runtime::Current()->IsZygote()) {
1771 os << "Zygote JIT code cache size (at point of fork): "
1772 << shared_region_.GetUsedMemoryForCode() / KB << "KB / "
1773 << shared_region_.GetResidentMemoryForCode() / KB << "KB\n"
1774 << "Zygote JIT data cache size (at point of fork): "
1775 << shared_region_.GetUsedMemoryForData() / KB << "KB / "
1776 << shared_region_.GetResidentMemoryForData() / KB << "KB\n";
1777 }
1778 os << "Current JIT mini-debug-info size: " << PrettySize(GetJitMiniDebugInfoMemUsage()) << "\n"
1779 << "Current JIT capacity: " << PrettySize(GetCurrentRegion()->GetCurrentCapacity()) << "\n"
1780 << "Current number of JIT JNI stub entries: " << jni_stubs_map_.size() << "\n"
1781 << "Current number of JIT code cache entries: " << method_code_map_.size() << "\n"
1782 << "Total number of JIT baseline compilations: " << number_of_baseline_compilations_ << "\n"
1783 << "Total number of JIT optimized compilations: " << number_of_optimized_compilations_ << "\n"
1784 << "Total number of JIT compilations for on stack replacement: "
1785 << number_of_osr_compilations_ << "\n"
1786 << "Total number of JIT code cache collections: " << number_of_collections_ << std::endl;
1787 histogram_stack_map_memory_use_.PrintMemoryUse(os);
1788 histogram_code_memory_use_.PrintMemoryUse(os);
1789 histogram_profiling_info_memory_use_.PrintMemoryUse(os);
1790 }
1791
PostForkChildAction(bool is_system_server,bool is_zygote)1792 void JitCodeCache::PostForkChildAction(bool is_system_server, bool is_zygote) {
1793 Thread* self = Thread::Current();
1794
1795 // Remove potential tasks that have been inherited from the zygote.
1796 // We do this now and not in Jit::PostForkChildAction, as system server calls
1797 // JitCodeCache::PostForkChildAction first, and then does some code loading
1798 // that may result in new JIT tasks that we want to keep.
1799 ThreadPool* pool = Runtime::Current()->GetJit()->GetThreadPool();
1800 if (pool != nullptr) {
1801 pool->RemoveAllTasks(self);
1802 }
1803
1804 MutexLock mu(self, *Locks::jit_lock_);
1805
1806 // Reset potential writable MemMaps inherited from the zygote. We never want
1807 // to write to them.
1808 shared_region_.ResetWritableMappings();
1809
1810 if (is_zygote || Runtime::Current()->IsSafeMode()) {
1811 // Don't create a private region for a child zygote. Regions are usually map shared
1812 // (to satisfy dual-view), and we don't want children of a child zygote to inherit it.
1813 return;
1814 }
1815
1816 // Reset all statistics to be specific to this process.
1817 number_of_baseline_compilations_ = 0;
1818 number_of_optimized_compilations_ = 0;
1819 number_of_osr_compilations_ = 0;
1820 number_of_collections_ = 0;
1821 histogram_stack_map_memory_use_.Reset();
1822 histogram_code_memory_use_.Reset();
1823 histogram_profiling_info_memory_use_.Reset();
1824
1825 size_t initial_capacity = Runtime::Current()->GetJITOptions()->GetCodeCacheInitialCapacity();
1826 size_t max_capacity = Runtime::Current()->GetJITOptions()->GetCodeCacheMaxCapacity();
1827 std::string error_msg;
1828 if (!private_region_.Initialize(initial_capacity,
1829 max_capacity,
1830 /* rwx_memory_allowed= */ !is_system_server,
1831 is_zygote,
1832 &error_msg)) {
1833 LOG(WARNING) << "Could not create private region after zygote fork: " << error_msg;
1834 }
1835 }
1836
GetCurrentRegion()1837 JitMemoryRegion* JitCodeCache::GetCurrentRegion() {
1838 return Runtime::Current()->IsZygote() ? &shared_region_ : &private_region_;
1839 }
1840
VisitAllMethods(const std::function<void (const void *,ArtMethod *)> & cb)1841 void JitCodeCache::VisitAllMethods(const std::function<void(const void*, ArtMethod*)>& cb) {
1842 for (const auto& it : jni_stubs_map_) {
1843 const JniStubData& data = it.second;
1844 if (data.IsCompiled()) {
1845 for (ArtMethod* method : data.GetMethods()) {
1846 cb(data.GetCode(), method);
1847 }
1848 }
1849 }
1850 for (auto it : method_code_map_) { // Includes OSR methods.
1851 cb(it.first, it.second);
1852 }
1853 for (auto it : saved_compiled_methods_map_) {
1854 cb(it.second, it.first);
1855 }
1856 for (auto it : zygote_map_) {
1857 if (it.code_ptr != nullptr && it.method != nullptr) {
1858 cb(it.code_ptr, it.method);
1859 }
1860 }
1861 }
1862
Initialize(uint32_t number_of_methods)1863 void ZygoteMap::Initialize(uint32_t number_of_methods) {
1864 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1865 // Allocate for 40-80% capacity. This will offer OK lookup times, and termination
1866 // cases.
1867 size_t capacity = RoundUpToPowerOfTwo(number_of_methods * 100 / 80);
1868 const uint8_t* memory = region_->AllocateData(
1869 capacity * sizeof(Entry) + sizeof(ZygoteCompilationState));
1870 if (memory == nullptr) {
1871 LOG(WARNING) << "Could not allocate data for the zygote map";
1872 return;
1873 }
1874 const Entry* data = reinterpret_cast<const Entry*>(memory);
1875 region_->FillData(data, capacity, Entry { nullptr, nullptr });
1876 map_ = ArrayRef(data, capacity);
1877 compilation_state_ = reinterpret_cast<const ZygoteCompilationState*>(
1878 memory + capacity * sizeof(Entry));
1879 region_->WriteData(compilation_state_, ZygoteCompilationState::kInProgress);
1880 }
1881
GetCodeFor(ArtMethod * method,uintptr_t pc) const1882 const void* ZygoteMap::GetCodeFor(ArtMethod* method, uintptr_t pc) const {
1883 if (map_.empty()) {
1884 return nullptr;
1885 }
1886
1887 if (method == nullptr) {
1888 // Do a linear search. This should only be used in debug builds.
1889 CHECK(kIsDebugBuild);
1890 for (const Entry& entry : map_) {
1891 const void* code_ptr = entry.code_ptr;
1892 if (code_ptr != nullptr) {
1893 OatQuickMethodHeader* method_header = OatQuickMethodHeader::FromCodePointer(code_ptr);
1894 if (method_header->Contains(pc)) {
1895 return code_ptr;
1896 }
1897 }
1898 }
1899 return nullptr;
1900 }
1901
1902 std::hash<ArtMethod*> hf;
1903 size_t index = hf(method) & (map_.size() - 1u);
1904 size_t original_index = index;
1905 // Loop over the array: we know this loop terminates as we will either
1906 // encounter the given method, or a null entry. Both terminate the loop.
1907 // Note that the zygote may concurrently write new entries to the map. That's OK as the
1908 // map is never resized.
1909 while (true) {
1910 const Entry& entry = map_[index];
1911 if (entry.method == nullptr) {
1912 // Not compiled yet.
1913 return nullptr;
1914 }
1915 if (entry.method == method) {
1916 if (entry.code_ptr == nullptr) {
1917 // This is a race with the zygote which wrote the method, but hasn't written the
1918 // code. Just bail and wait for the next time we need the method.
1919 return nullptr;
1920 }
1921 if (pc != 0 && !OatQuickMethodHeader::FromCodePointer(entry.code_ptr)->Contains(pc)) {
1922 return nullptr;
1923 }
1924 return entry.code_ptr;
1925 }
1926 index = (index + 1) & (map_.size() - 1);
1927 DCHECK_NE(original_index, index);
1928 }
1929 }
1930
Put(const void * code,ArtMethod * method)1931 void ZygoteMap::Put(const void* code, ArtMethod* method) {
1932 if (map_.empty()) {
1933 return;
1934 }
1935 CHECK(Runtime::Current()->IsZygote());
1936 std::hash<ArtMethod*> hf;
1937 size_t index = hf(method) & (map_.size() - 1);
1938 size_t original_index = index;
1939 // Because the size of the map is bigger than the number of methods that will
1940 // be added, we are guaranteed to find a free slot in the array, and
1941 // therefore for this loop to terminate.
1942 while (true) {
1943 const Entry* entry = &map_[index];
1944 if (entry->method == nullptr) {
1945 // Note that readers can read this memory concurrently, but that's OK as
1946 // we are writing pointers.
1947 region_->WriteData(entry, Entry { method, code });
1948 break;
1949 }
1950 index = (index + 1) & (map_.size() - 1);
1951 DCHECK_NE(original_index, index);
1952 }
1953 DCHECK_EQ(GetCodeFor(method), code);
1954 }
1955
1956 } // namespace jit
1957 } // namespace art
1958