1 /*
2 * Copyright (C) 2011 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 "dex_to_dex_compiler.h"
18
19 #include <android-base/logging.h>
20 #include <android-base/stringprintf.h>
21
22 #include "art_field-inl.h"
23 #include "art_method-inl.h"
24 #include "base/logging.h" // For VLOG
25 #include "base/macros.h"
26 #include "base/mutex.h"
27 #include "compiled_method.h"
28 #include "dex/bytecode_utils.h"
29 #include "dex/class_accessor-inl.h"
30 #include "dex/dex_file-inl.h"
31 #include "dex/dex_instruction-inl.h"
32 #include "dex_to_dex_decompiler.h"
33 #include "driver/compiler_driver.h"
34 #include "driver/compiler_options.h"
35 #include "driver/dex_compilation_unit.h"
36 #include "mirror/dex_cache.h"
37 #include "quicken_info.h"
38 #include "thread-current-inl.h"
39
40 namespace art {
41 namespace optimizer {
42
43 using android::base::StringPrintf;
44
45 // Controls quickening activation.
46 const bool kEnableQuickening = true;
47 // Control check-cast elision.
48 const bool kEnableCheckCastEllision = true;
49
50 // Holds the state for compiling a single method.
51 struct DexToDexCompiler::CompilationState {
52 struct QuickenedInfo {
QuickenedInfoart::optimizer::DexToDexCompiler::CompilationState::QuickenedInfo53 QuickenedInfo(uint32_t pc, uint16_t index) : dex_pc(pc), dex_member_index(index) {}
54
55 uint32_t dex_pc;
56 uint16_t dex_member_index;
57 };
58
59 CompilationState(DexToDexCompiler* compiler,
60 const DexCompilationUnit& unit,
61 const CompilationLevel compilation_level,
62 const std::vector<uint8_t>* quicken_data);
63
GetQuickenedInfoart::optimizer::DexToDexCompiler::CompilationState64 const std::vector<QuickenedInfo>& GetQuickenedInfo() const {
65 return quickened_info_;
66 }
67
68 // Returns the quickening info, or an empty array if it was not quickened.
69 // If already_quickened is true, then don't change anything but still return what the quicken
70 // data would have been.
71 std::vector<uint8_t> Compile();
72
73 const DexFile& GetDexFile() const;
74
75 // Compiles a RETURN-VOID into a RETURN-VOID-BARRIER within a constructor where
76 // a barrier is required.
77 void CompileReturnVoid(Instruction* inst, uint32_t dex_pc);
78
79 // Compiles a CHECK-CAST into 2 NOP instructions if it is known to be safe. In
80 // this case, returns the second NOP instruction pointer. Otherwise, returns
81 // the given "inst".
82 Instruction* CompileCheckCast(Instruction* inst, uint32_t dex_pc);
83
84 // Compiles a field access into a quick field access.
85 // The field index is replaced by an offset within an Object where we can read
86 // from / write to this field. Therefore, this does not involve any resolution
87 // at runtime.
88 // Since the field index is encoded with 16 bits, we can replace it only if the
89 // field offset can be encoded with 16 bits too.
90 void CompileInstanceFieldAccess(Instruction* inst, uint32_t dex_pc,
91 Instruction::Code new_opcode, bool is_put);
92
93 // Compiles a virtual method invocation into a quick virtual method invocation.
94 // The method index is replaced by the vtable index where the corresponding
95 // executable can be found. Therefore, this does not involve any resolution
96 // at runtime.
97 // Since the method index is encoded with 16 bits, we can replace it only if the
98 // vtable index can be encoded with 16 bits too.
99 void CompileInvokeVirtual(Instruction* inst, uint32_t dex_pc,
100 Instruction::Code new_opcode, bool is_range);
101
102 // Return the next index.
103 uint16_t NextIndex();
104
105 // Returns the dequickened index if an instruction is quickened, otherwise return index.
106 uint16_t GetIndexForInstruction(const Instruction* inst, uint32_t index);
107
108 DexToDexCompiler* const compiler_;
109 CompilerDriver& driver_;
110 const DexCompilationUnit& unit_;
111 const CompilationLevel compilation_level_;
112
113 // Filled by the compiler when quickening, in order to encode that information
114 // in the .oat file. The runtime will use that information to get to the original
115 // opcodes.
116 std::vector<QuickenedInfo> quickened_info_;
117
118 // True if we optimized a return void to a return void no barrier.
119 bool optimized_return_void_ = false;
120
121 // If the code item was already quickened previously.
122 const bool already_quickened_;
123 const QuickenInfoTable existing_quicken_info_;
124 uint32_t quicken_index_ = 0u;
125
126 DISALLOW_COPY_AND_ASSIGN(CompilationState);
127 };
128
DexToDexCompiler(CompilerDriver * driver)129 DexToDexCompiler::DexToDexCompiler(CompilerDriver* driver)
130 : driver_(driver),
131 lock_("Quicken lock", kDexToDexCompilerLock) {
132 DCHECK(driver != nullptr);
133 }
134
ClearState()135 void DexToDexCompiler::ClearState() {
136 MutexLock lock(Thread::Current(), lock_);
137 active_dex_file_ = nullptr;
138 active_bit_vector_ = nullptr;
139 should_quicken_.clear();
140 shared_code_item_quicken_info_.clear();
141 }
142
NumCodeItemsToQuicken(Thread * self) const143 size_t DexToDexCompiler::NumCodeItemsToQuicken(Thread* self) const {
144 MutexLock lock(self, lock_);
145 return num_code_items_;
146 }
147
GetOrAddBitVectorForDex(const DexFile * dex_file)148 BitVector* DexToDexCompiler::GetOrAddBitVectorForDex(const DexFile* dex_file) {
149 if (active_dex_file_ != dex_file) {
150 active_dex_file_ = dex_file;
151 auto inserted = should_quicken_.emplace(dex_file,
152 BitVector(dex_file->NumMethodIds(),
153 /*expandable*/ false,
154 Allocator::GetMallocAllocator()));
155 active_bit_vector_ = &inserted.first->second;
156 }
157 return active_bit_vector_;
158 }
159
MarkForCompilation(Thread * self,const MethodReference & method_ref)160 void DexToDexCompiler::MarkForCompilation(Thread* self,
161 const MethodReference& method_ref) {
162 MutexLock lock(self, lock_);
163 BitVector* const bitmap = GetOrAddBitVectorForDex(method_ref.dex_file);
164 DCHECK(bitmap != nullptr);
165 DCHECK(!bitmap->IsBitSet(method_ref.index));
166 bitmap->SetBit(method_ref.index);
167 ++num_code_items_;
168 }
169
CompilationState(DexToDexCompiler * compiler,const DexCompilationUnit & unit,const CompilationLevel compilation_level,const std::vector<uint8_t> * quicken_data)170 DexToDexCompiler::CompilationState::CompilationState(DexToDexCompiler* compiler,
171 const DexCompilationUnit& unit,
172 const CompilationLevel compilation_level,
173 const std::vector<uint8_t>* quicken_data)
174 : compiler_(compiler),
175 driver_(*compiler->GetDriver()),
176 unit_(unit),
177 compilation_level_(compilation_level),
178 already_quickened_(quicken_data != nullptr),
179 existing_quicken_info_(already_quickened_
180 ? ArrayRef<const uint8_t>(*quicken_data) : ArrayRef<const uint8_t>()) {}
181
NextIndex()182 uint16_t DexToDexCompiler::CompilationState::NextIndex() {
183 DCHECK(already_quickened_);
184 if (kIsDebugBuild && quicken_index_ >= existing_quicken_info_.NumIndices()) {
185 for (const DexInstructionPcPair& pair : unit_.GetCodeItemAccessor()) {
186 LOG(ERROR) << pair->DumpString(nullptr);
187 }
188 LOG(FATAL) << "Mismatched number of quicken slots.";
189 }
190 const uint16_t ret = existing_quicken_info_.GetData(quicken_index_);
191 quicken_index_++;
192 return ret;
193 }
194
GetIndexForInstruction(const Instruction * inst,uint32_t index)195 uint16_t DexToDexCompiler::CompilationState::GetIndexForInstruction(const Instruction* inst,
196 uint32_t index) {
197 if (UNLIKELY(already_quickened_)) {
198 return inst->IsQuickened() ? NextIndex() : index;
199 }
200 DCHECK(!inst->IsQuickened());
201 return index;
202 }
203
ShouldCompileMethod(const MethodReference & ref)204 bool DexToDexCompiler::ShouldCompileMethod(const MethodReference& ref) {
205 // TODO: It's probably safe to avoid the lock here if the active_dex_file_ matches since we only
206 // only call ShouldCompileMethod on one dex at a time.
207 MutexLock lock(Thread::Current(), lock_);
208 return GetOrAddBitVectorForDex(ref.dex_file)->IsBitSet(ref.index);
209 }
210
Compile()211 std::vector<uint8_t> DexToDexCompiler::CompilationState::Compile() {
212 DCHECK_EQ(compilation_level_, CompilationLevel::kOptimize);
213 const CodeItemDataAccessor& instructions = unit_.GetCodeItemAccessor();
214 for (DexInstructionIterator it = instructions.begin(); it != instructions.end(); ++it) {
215 const uint32_t dex_pc = it.DexPc();
216 Instruction* inst = const_cast<Instruction*>(&it.Inst());
217
218 if (!already_quickened_) {
219 DCHECK(!inst->IsQuickened());
220 }
221
222 switch (inst->Opcode()) {
223 case Instruction::RETURN_VOID:
224 CompileReturnVoid(inst, dex_pc);
225 break;
226
227 case Instruction::CHECK_CAST:
228 inst = CompileCheckCast(inst, dex_pc);
229 if (inst->Opcode() == Instruction::NOP) {
230 // We turned the CHECK_CAST into two NOPs, avoid visiting the second NOP twice since this
231 // would add 2 quickening info entries.
232 ++it;
233 }
234 break;
235
236 case Instruction::IGET:
237 case Instruction::IGET_QUICK:
238 CompileInstanceFieldAccess(inst, dex_pc, Instruction::IGET_QUICK, false);
239 break;
240
241 case Instruction::IGET_WIDE:
242 case Instruction::IGET_WIDE_QUICK:
243 CompileInstanceFieldAccess(inst, dex_pc, Instruction::IGET_WIDE_QUICK, false);
244 break;
245
246 case Instruction::IGET_OBJECT:
247 case Instruction::IGET_OBJECT_QUICK:
248 CompileInstanceFieldAccess(inst, dex_pc, Instruction::IGET_OBJECT_QUICK, false);
249 break;
250
251 case Instruction::IGET_BOOLEAN:
252 case Instruction::IGET_BOOLEAN_QUICK:
253 CompileInstanceFieldAccess(inst, dex_pc, Instruction::IGET_BOOLEAN_QUICK, false);
254 break;
255
256 case Instruction::IGET_BYTE:
257 case Instruction::IGET_BYTE_QUICK:
258 CompileInstanceFieldAccess(inst, dex_pc, Instruction::IGET_BYTE_QUICK, false);
259 break;
260
261 case Instruction::IGET_CHAR:
262 case Instruction::IGET_CHAR_QUICK:
263 CompileInstanceFieldAccess(inst, dex_pc, Instruction::IGET_CHAR_QUICK, false);
264 break;
265
266 case Instruction::IGET_SHORT:
267 case Instruction::IGET_SHORT_QUICK:
268 CompileInstanceFieldAccess(inst, dex_pc, Instruction::IGET_SHORT_QUICK, false);
269 break;
270
271 case Instruction::IPUT:
272 case Instruction::IPUT_QUICK:
273 CompileInstanceFieldAccess(inst, dex_pc, Instruction::IPUT_QUICK, true);
274 break;
275
276 case Instruction::IPUT_BOOLEAN:
277 case Instruction::IPUT_BOOLEAN_QUICK:
278 CompileInstanceFieldAccess(inst, dex_pc, Instruction::IPUT_BOOLEAN_QUICK, true);
279 break;
280
281 case Instruction::IPUT_BYTE:
282 case Instruction::IPUT_BYTE_QUICK:
283 CompileInstanceFieldAccess(inst, dex_pc, Instruction::IPUT_BYTE_QUICK, true);
284 break;
285
286 case Instruction::IPUT_CHAR:
287 case Instruction::IPUT_CHAR_QUICK:
288 CompileInstanceFieldAccess(inst, dex_pc, Instruction::IPUT_CHAR_QUICK, true);
289 break;
290
291 case Instruction::IPUT_SHORT:
292 case Instruction::IPUT_SHORT_QUICK:
293 CompileInstanceFieldAccess(inst, dex_pc, Instruction::IPUT_SHORT_QUICK, true);
294 break;
295
296 case Instruction::IPUT_WIDE:
297 case Instruction::IPUT_WIDE_QUICK:
298 CompileInstanceFieldAccess(inst, dex_pc, Instruction::IPUT_WIDE_QUICK, true);
299 break;
300
301 case Instruction::IPUT_OBJECT:
302 case Instruction::IPUT_OBJECT_QUICK:
303 CompileInstanceFieldAccess(inst, dex_pc, Instruction::IPUT_OBJECT_QUICK, true);
304 break;
305
306 case Instruction::INVOKE_VIRTUAL:
307 case Instruction::INVOKE_VIRTUAL_QUICK:
308 CompileInvokeVirtual(inst, dex_pc, Instruction::INVOKE_VIRTUAL_QUICK, false);
309 break;
310
311 case Instruction::INVOKE_VIRTUAL_RANGE:
312 case Instruction::INVOKE_VIRTUAL_RANGE_QUICK:
313 CompileInvokeVirtual(inst, dex_pc, Instruction::INVOKE_VIRTUAL_RANGE_QUICK, true);
314 break;
315
316 case Instruction::NOP:
317 if (already_quickened_) {
318 const uint16_t reference_index = NextIndex();
319 quickened_info_.push_back(QuickenedInfo(dex_pc, reference_index));
320 if (reference_index == DexFile::kDexNoIndex16) {
321 // This means it was a normal nop and not a check-cast.
322 break;
323 }
324 const uint16_t type_index = NextIndex();
325 if (driver_.IsSafeCast(&unit_, dex_pc)) {
326 quickened_info_.push_back(QuickenedInfo(dex_pc, type_index));
327 }
328 ++it;
329 } else {
330 // We need to differentiate between check cast inserted NOP and normal NOP, put an invalid
331 // index in the map for normal nops. This should be rare in real code.
332 quickened_info_.push_back(QuickenedInfo(dex_pc, DexFile::kDexNoIndex16));
333 }
334 break;
335
336 default:
337 // Nothing to do.
338 break;
339 }
340 }
341
342 if (already_quickened_) {
343 DCHECK_EQ(quicken_index_, existing_quicken_info_.NumIndices());
344 }
345
346 // Even if there are no indices, generate an empty quicken info so that we know the method was
347 // quickened.
348
349 std::vector<uint8_t> quicken_data;
350 if (kIsDebugBuild) {
351 // Double check that the counts line up with the size of the quicken info.
352 size_t quicken_count = 0;
353 for (const DexInstructionPcPair& pair : instructions) {
354 if (QuickenInfoTable::NeedsIndexForInstruction(&pair.Inst())) {
355 ++quicken_count;
356 }
357 }
358 CHECK_EQ(quicken_count, GetQuickenedInfo().size());
359 }
360
361 QuickenInfoTable::Builder builder(&quicken_data, GetQuickenedInfo().size());
362 // Length is encoded by the constructor.
363 for (const CompilationState::QuickenedInfo& info : GetQuickenedInfo()) {
364 // Dex pc is not serialized, only used for checking the instructions. Since we access the
365 // array based on the index of the quickened instruction, the indexes must line up perfectly.
366 // The reader side uses the NeedsIndexForInstruction function too.
367 const Instruction& inst = instructions.InstructionAt(info.dex_pc);
368 CHECK(QuickenInfoTable::NeedsIndexForInstruction(&inst)) << inst.Opcode();
369 builder.AddIndex(info.dex_member_index);
370 }
371 DCHECK(!quicken_data.empty());
372 return quicken_data;
373 }
374
CompileReturnVoid(Instruction * inst,uint32_t dex_pc)375 void DexToDexCompiler::CompilationState::CompileReturnVoid(Instruction* inst, uint32_t dex_pc) {
376 DCHECK_EQ(inst->Opcode(), Instruction::RETURN_VOID);
377 if (unit_.IsConstructor()) {
378 // Are we compiling a non clinit constructor which needs a barrier ?
379 if (!unit_.IsStatic() && unit_.RequiresConstructorBarrier()) {
380 return;
381 }
382 }
383 // Replace RETURN_VOID by RETURN_VOID_NO_BARRIER.
384 VLOG(compiler) << "Replacing " << Instruction::Name(inst->Opcode())
385 << " by " << Instruction::Name(Instruction::RETURN_VOID_NO_BARRIER)
386 << " at dex pc " << StringPrintf("0x%x", dex_pc) << " in method "
387 << GetDexFile().PrettyMethod(unit_.GetDexMethodIndex(), true);
388 inst->SetOpcode(Instruction::RETURN_VOID_NO_BARRIER);
389 optimized_return_void_ = true;
390 }
391
CompileCheckCast(Instruction * inst,uint32_t dex_pc)392 Instruction* DexToDexCompiler::CompilationState::CompileCheckCast(Instruction* inst,
393 uint32_t dex_pc) {
394 if (!kEnableCheckCastEllision) {
395 return inst;
396 }
397 if (!driver_.IsSafeCast(&unit_, dex_pc)) {
398 return inst;
399 }
400 // Ok, this is a safe cast. Since the "check-cast" instruction size is 2 code
401 // units and a "nop" instruction size is 1 code unit, we need to replace it by
402 // 2 consecutive NOP instructions.
403 // Because the caller loops over instructions by calling Instruction::Next onto
404 // the current instruction, we need to return the 2nd NOP instruction. Indeed,
405 // its next instruction is the former check-cast's next instruction.
406 VLOG(compiler) << "Removing " << Instruction::Name(inst->Opcode())
407 << " by replacing it with 2 NOPs at dex pc "
408 << StringPrintf("0x%x", dex_pc) << " in method "
409 << GetDexFile().PrettyMethod(unit_.GetDexMethodIndex(), true);
410 if (!already_quickened_) {
411 quickened_info_.push_back(QuickenedInfo(dex_pc, inst->VRegA_21c()));
412 quickened_info_.push_back(QuickenedInfo(dex_pc, inst->VRegB_21c()));
413
414 // We are modifying 4 consecutive bytes.
415 inst->SetOpcode(Instruction::NOP);
416 inst->SetVRegA_10x(0u); // keep compliant with verifier.
417 // Get to next instruction which is the second half of check-cast and replace
418 // it by a NOP.
419 inst = const_cast<Instruction*>(inst->Next());
420 inst->SetOpcode(Instruction::NOP);
421 inst->SetVRegA_10x(0u); // keep compliant with verifier.
422 }
423 return inst;
424 }
425
CompileInstanceFieldAccess(Instruction * inst,uint32_t dex_pc,Instruction::Code new_opcode,bool is_put)426 void DexToDexCompiler::CompilationState::CompileInstanceFieldAccess(Instruction* inst,
427 uint32_t dex_pc,
428 Instruction::Code new_opcode,
429 bool is_put) {
430 if (!kEnableQuickening) {
431 return;
432 }
433 uint32_t field_idx = GetIndexForInstruction(inst, inst->VRegC_22c());
434 MemberOffset field_offset(0u);
435 bool is_volatile;
436 bool fast_path = driver_.ComputeInstanceFieldInfo(field_idx, &unit_, is_put,
437 &field_offset, &is_volatile);
438 if (fast_path && !is_volatile && IsUint<16>(field_offset.Int32Value())) {
439 VLOG(compiler) << "Quickening " << Instruction::Name(inst->Opcode())
440 << " to " << Instruction::Name(new_opcode)
441 << " by replacing field index " << field_idx
442 << " by field offset " << field_offset.Int32Value()
443 << " at dex pc " << StringPrintf("0x%x", dex_pc) << " in method "
444 << GetDexFile().PrettyMethod(unit_.GetDexMethodIndex(), true);
445 if (!already_quickened_) {
446 // We are modifying 4 consecutive bytes.
447 inst->SetOpcode(new_opcode);
448 // Replace field index by field offset.
449 inst->SetVRegC_22c(static_cast<uint16_t>(field_offset.Int32Value()));
450 }
451 quickened_info_.push_back(QuickenedInfo(dex_pc, field_idx));
452 }
453 }
454
GetDexFile() const455 const DexFile& DexToDexCompiler::CompilationState::GetDexFile() const {
456 return *unit_.GetDexFile();
457 }
458
CompileInvokeVirtual(Instruction * inst,uint32_t dex_pc,Instruction::Code new_opcode,bool is_range)459 void DexToDexCompiler::CompilationState::CompileInvokeVirtual(Instruction* inst,
460 uint32_t dex_pc,
461 Instruction::Code new_opcode,
462 bool is_range) {
463 if (!kEnableQuickening) {
464 return;
465 }
466 uint32_t method_idx = GetIndexForInstruction(inst,
467 is_range ? inst->VRegB_3rc() : inst->VRegB_35c());
468 ScopedObjectAccess soa(Thread::Current());
469
470 ClassLinker* class_linker = unit_.GetClassLinker();
471 ArtMethod* resolved_method =
472 class_linker->ResolveMethod<ClassLinker::ResolveMode::kCheckICCEAndIAE>(
473 method_idx,
474 unit_.GetDexCache(),
475 unit_.GetClassLoader(),
476 /* referrer= */ nullptr,
477 kVirtual);
478
479 if (UNLIKELY(resolved_method == nullptr)) {
480 // Clean up any exception left by type resolution.
481 soa.Self()->ClearException();
482 return;
483 }
484
485 uint32_t vtable_idx = resolved_method->GetMethodIndex();
486 DCHECK(IsUint<16>(vtable_idx));
487 VLOG(compiler) << "Quickening " << Instruction::Name(inst->Opcode())
488 << "(" << GetDexFile().PrettyMethod(method_idx, true) << ")"
489 << " to " << Instruction::Name(new_opcode)
490 << " by replacing method index " << method_idx
491 << " by vtable index " << vtable_idx
492 << " at dex pc " << StringPrintf("0x%x", dex_pc) << " in method "
493 << GetDexFile().PrettyMethod(unit_.GetDexMethodIndex(), true);
494 if (!already_quickened_) {
495 // We are modifying 4 consecutive bytes.
496 inst->SetOpcode(new_opcode);
497 // Replace method index by vtable index.
498 if (is_range) {
499 inst->SetVRegB_3rc(static_cast<uint16_t>(vtable_idx));
500 } else {
501 inst->SetVRegB_35c(static_cast<uint16_t>(vtable_idx));
502 }
503 }
504 quickened_info_.push_back(QuickenedInfo(dex_pc, method_idx));
505 }
506
CompileMethod(const dex::CodeItem * code_item,uint32_t access_flags,InvokeType invoke_type ATTRIBUTE_UNUSED,uint16_t class_def_idx,uint32_t method_idx,Handle<mirror::ClassLoader> class_loader,const DexFile & dex_file,CompilationLevel compilation_level)507 CompiledMethod* DexToDexCompiler::CompileMethod(
508 const dex::CodeItem* code_item,
509 uint32_t access_flags,
510 InvokeType invoke_type ATTRIBUTE_UNUSED,
511 uint16_t class_def_idx,
512 uint32_t method_idx,
513 Handle<mirror::ClassLoader> class_loader,
514 const DexFile& dex_file,
515 CompilationLevel compilation_level) {
516 if (compilation_level == CompilationLevel::kDontDexToDexCompile) {
517 return nullptr;
518 }
519
520 ScopedObjectAccess soa(Thread::Current());
521 StackHandleScope<1> hs(soa.Self());
522 ClassLinker* const class_linker = Runtime::Current()->GetClassLinker();
523 art::DexCompilationUnit unit(
524 class_loader,
525 class_linker,
526 dex_file,
527 code_item,
528 class_def_idx,
529 method_idx,
530 access_flags,
531 driver_->GetCompilerOptions().GetVerifiedMethod(&dex_file, method_idx),
532 hs.NewHandle(class_linker->FindDexCache(soa.Self(), dex_file)));
533
534 std::vector<uint8_t> quicken_data;
535 // If the code item is shared with multiple different method ids, make sure that we quicken only
536 // once and verify that all the dequicken maps match.
537 if (UNLIKELY(shared_code_items_.find(code_item) != shared_code_items_.end())) {
538 // Avoid quickening the shared code items for now because the existing conflict detection logic
539 // does not currently handle cases where the code item is quickened in one place but
540 // compiled in another.
541 static constexpr bool kAvoidQuickeningSharedCodeItems = true;
542 if (kAvoidQuickeningSharedCodeItems) {
543 return nullptr;
544 }
545 // For shared code items, use a lock to prevent races.
546 MutexLock mu(soa.Self(), lock_);
547 auto existing = shared_code_item_quicken_info_.find(code_item);
548 QuickenState* existing_data = nullptr;
549 std::vector<uint8_t>* existing_quicken_data = nullptr;
550 if (existing != shared_code_item_quicken_info_.end()) {
551 existing_data = &existing->second;
552 if (existing_data->conflict_) {
553 return nullptr;
554 }
555 existing_quicken_data = &existing_data->quicken_data_;
556 }
557 bool optimized_return_void;
558 {
559 CompilationState state(this, unit, compilation_level, existing_quicken_data);
560 quicken_data = state.Compile();
561 optimized_return_void = state.optimized_return_void_;
562 }
563
564 // Already quickened, check that the data matches what was previously seen.
565 MethodReference method_ref(&dex_file, method_idx);
566 if (existing_data != nullptr) {
567 if (*existing_quicken_data != quicken_data ||
568 existing_data->optimized_return_void_ != optimized_return_void) {
569 VLOG(compiler) << "Quicken data mismatch, for method "
570 << dex_file.PrettyMethod(method_idx);
571 // Mark the method as a conflict to never attempt to quicken it in the future.
572 existing_data->conflict_ = true;
573 }
574 existing_data->methods_.push_back(method_ref);
575 } else {
576 QuickenState new_state;
577 new_state.methods_.push_back(method_ref);
578 new_state.quicken_data_ = quicken_data;
579 new_state.optimized_return_void_ = optimized_return_void;
580 bool inserted = shared_code_item_quicken_info_.emplace(code_item, new_state).second;
581 CHECK(inserted) << "Failed to insert " << dex_file.PrettyMethod(method_idx);
582 }
583
584 // Easy sanity check is to check that the existing stuff matches by re-quickening using the
585 // newly produced quicken data.
586 // Note that this needs to be behind the lock for this case since we may unquicken in another
587 // thread.
588 if (kIsDebugBuild) {
589 CompilationState state2(this, unit, compilation_level, &quicken_data);
590 std::vector<uint8_t> new_data = state2.Compile();
591 CHECK(new_data == quicken_data) << "Mismatch producing new quicken data";
592 }
593 } else {
594 CompilationState state(this, unit, compilation_level, /*quicken_data*/ nullptr);
595 quicken_data = state.Compile();
596
597 // Easy sanity check is to check that the existing stuff matches by re-quickening using the
598 // newly produced quicken data.
599 if (kIsDebugBuild) {
600 CompilationState state2(this, unit, compilation_level, &quicken_data);
601 std::vector<uint8_t> new_data = state2.Compile();
602 CHECK(new_data == quicken_data) << "Mismatch producing new quicken data";
603 }
604 }
605
606 if (quicken_data.empty()) {
607 return nullptr;
608 }
609
610 // Create a `CompiledMethod`, with the quickened information in the vmap table.
611 InstructionSet instruction_set = driver_->GetCompilerOptions().GetInstructionSet();
612 if (instruction_set == InstructionSet::kThumb2) {
613 // Don't use the thumb2 instruction set to avoid the one off code delta.
614 instruction_set = InstructionSet::kArm;
615 }
616 CompiledMethod* ret = CompiledMethod::SwapAllocCompiledMethod(
617 driver_->GetCompiledMethodStorage(),
618 instruction_set,
619 ArrayRef<const uint8_t>(), // no code
620 ArrayRef<const uint8_t>(quicken_data), // vmap_table
621 ArrayRef<const uint8_t>(), // cfi data
622 ArrayRef<const linker::LinkerPatch>());
623 DCHECK(ret != nullptr);
624 return ret;
625 }
626
SetDexFiles(const std::vector<const DexFile * > & dex_files)627 void DexToDexCompiler::SetDexFiles(const std::vector<const DexFile*>& dex_files) {
628 // Record what code items are already seen to detect when multiple methods have the same code
629 // item.
630 std::unordered_set<const dex::CodeItem*> seen_code_items;
631 for (const DexFile* dex_file : dex_files) {
632 for (ClassAccessor accessor : dex_file->GetClasses()) {
633 for (const ClassAccessor::Method& method : accessor.GetMethods()) {
634 const dex::CodeItem* code_item = method.GetCodeItem();
635 // Detect the shared code items.
636 if (!seen_code_items.insert(code_item).second) {
637 shared_code_items_.insert(code_item);
638 }
639 }
640 }
641 }
642 VLOG(compiler) << "Shared code items " << shared_code_items_.size();
643 }
644
UnquickenConflictingMethods()645 void DexToDexCompiler::UnquickenConflictingMethods() {
646 MutexLock mu(Thread::Current(), lock_);
647 size_t unquicken_count = 0;
648 for (const auto& pair : shared_code_item_quicken_info_) {
649 const dex::CodeItem* code_item = pair.first;
650 const QuickenState& state = pair.second;
651 CHECK_GE(state.methods_.size(), 1u);
652 if (state.conflict_) {
653 // Unquicken using the existing quicken data.
654 // TODO: Do we really need to pass a dex file in?
655 optimizer::ArtDecompileDEX(*state.methods_[0].dex_file,
656 *code_item,
657 ArrayRef<const uint8_t>(state.quicken_data_),
658 /* decompile_return_instruction*/ true);
659 ++unquicken_count;
660 // Go clear the vmaps for all the methods that were already quickened to avoid writing them
661 // out during oat writing.
662 for (const MethodReference& ref : state.methods_) {
663 CompiledMethod* method = driver_->RemoveCompiledMethod(ref);
664 if (method != nullptr) {
665 // There is up to one compiled method for each method ref. Releasing it leaves the
666 // deduped data intact, this means its safe to do even when other threads might be
667 // compiling.
668 CompiledMethod::ReleaseSwapAllocatedCompiledMethod(driver_->GetCompiledMethodStorage(),
669 method);
670 }
671 }
672 }
673 }
674 }
675
676 } // namespace optimizer
677
678 } // namespace art
679