1 /*
2 * Copyright (C) 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 "optimizing_compiler.h"
18
19 #include <fstream>
20 #include <memory>
21 #include <sstream>
22
23 #include <stdint.h>
24
25 #include "art_method-inl.h"
26 #include "base/arena_allocator.h"
27 #include "base/arena_containers.h"
28 #include "base/dumpable.h"
29 #include "base/logging.h"
30 #include "base/macros.h"
31 #include "base/mutex.h"
32 #include "base/scoped_arena_allocator.h"
33 #include "base/timing_logger.h"
34 #include "builder.h"
35 #include "class_root.h"
36 #include "code_generator.h"
37 #include "compiled_method.h"
38 #include "compiler.h"
39 #include "debug/elf_debug_writer.h"
40 #include "debug/method_debug_info.h"
41 #include "dex/dex_file_types.h"
42 #include "dex/verification_results.h"
43 #include "dex/verified_method.h"
44 #include "driver/compiled_method_storage.h"
45 #include "driver/compiler_options.h"
46 #include "driver/dex_compilation_unit.h"
47 #include "graph_checker.h"
48 #include "graph_visualizer.h"
49 #include "inliner.h"
50 #include "jit/debugger_interface.h"
51 #include "jit/jit.h"
52 #include "jit/jit_code_cache.h"
53 #include "jit/jit_logger.h"
54 #include "jni/quick/jni_compiler.h"
55 #include "linker/linker_patch.h"
56 #include "nodes.h"
57 #include "oat_quick_method_header.h"
58 #include "prepare_for_register_allocation.h"
59 #include "reference_type_propagation.h"
60 #include "register_allocator_linear_scan.h"
61 #include "select_generator.h"
62 #include "ssa_builder.h"
63 #include "ssa_liveness_analysis.h"
64 #include "ssa_phi_elimination.h"
65 #include "stack_map_stream.h"
66 #include "utils/assembler.h"
67 #include "verifier/verifier_compiler_binding.h"
68
69 namespace art {
70
71 static constexpr size_t kArenaAllocatorMemoryReportThreshold = 8 * MB;
72
73 static constexpr const char* kPassNameSeparator = "$";
74
75 /**
76 * Used by the code generator, to allocate the code in a vector.
77 */
78 class CodeVectorAllocator final : public CodeAllocator {
79 public:
CodeVectorAllocator(ArenaAllocator * allocator)80 explicit CodeVectorAllocator(ArenaAllocator* allocator)
81 : memory_(allocator->Adapter(kArenaAllocCodeBuffer)) {}
82
Allocate(size_t size)83 uint8_t* Allocate(size_t size) override {
84 memory_.resize(size);
85 return &memory_[0];
86 }
87
GetMemory() const88 ArrayRef<const uint8_t> GetMemory() const override { return ArrayRef<const uint8_t>(memory_); }
GetData()89 uint8_t* GetData() { return memory_.data(); }
90
91 private:
92 ArenaVector<uint8_t> memory_;
93
94 DISALLOW_COPY_AND_ASSIGN(CodeVectorAllocator);
95 };
96
97 /**
98 * Filter to apply to the visualizer. Methods whose name contain that filter will
99 * be dumped.
100 */
101 static constexpr const char kStringFilter[] = "";
102
103 class PassScope;
104
105 class PassObserver : public ValueObject {
106 public:
PassObserver(HGraph * graph,CodeGenerator * codegen,std::ostream * visualizer_output,const CompilerOptions & compiler_options,Mutex & dump_mutex)107 PassObserver(HGraph* graph,
108 CodeGenerator* codegen,
109 std::ostream* visualizer_output,
110 const CompilerOptions& compiler_options,
111 Mutex& dump_mutex)
112 : graph_(graph),
113 last_seen_graph_size_(0),
114 cached_method_name_(),
115 timing_logger_enabled_(compiler_options.GetDumpPassTimings()),
116 timing_logger_(timing_logger_enabled_ ? GetMethodName() : "", true, true),
117 disasm_info_(graph->GetAllocator()),
118 visualizer_oss_(),
119 visualizer_output_(visualizer_output),
120 visualizer_enabled_(!compiler_options.GetDumpCfgFileName().empty()),
121 visualizer_(&visualizer_oss_, graph, *codegen),
122 visualizer_dump_mutex_(dump_mutex),
123 graph_in_bad_state_(false) {
124 if (timing_logger_enabled_ || visualizer_enabled_) {
125 if (!IsVerboseMethod(compiler_options, GetMethodName())) {
126 timing_logger_enabled_ = visualizer_enabled_ = false;
127 }
128 if (visualizer_enabled_) {
129 visualizer_.PrintHeader(GetMethodName());
130 codegen->SetDisassemblyInformation(&disasm_info_);
131 }
132 }
133 }
134
~PassObserver()135 ~PassObserver() {
136 if (timing_logger_enabled_) {
137 LOG(INFO) << "TIMINGS " << GetMethodName();
138 LOG(INFO) << Dumpable<TimingLogger>(timing_logger_);
139 }
140 if (visualizer_enabled_) {
141 FlushVisualizer();
142 }
143 DCHECK(visualizer_oss_.str().empty());
144 }
145
DumpDisassembly()146 void DumpDisassembly() {
147 if (visualizer_enabled_) {
148 visualizer_.DumpGraphWithDisassembly();
149 }
150 }
151
SetGraphInBadState()152 void SetGraphInBadState() { graph_in_bad_state_ = true; }
153
GetMethodName()154 const char* GetMethodName() {
155 // PrettyMethod() is expensive, so we delay calling it until we actually have to.
156 if (cached_method_name_.empty()) {
157 cached_method_name_ = graph_->GetDexFile().PrettyMethod(graph_->GetMethodIdx());
158 }
159 return cached_method_name_.c_str();
160 }
161
162 private:
StartPass(const char * pass_name)163 void StartPass(const char* pass_name) {
164 VLOG(compiler) << "Starting pass: " << pass_name;
165 // Dump graph first, then start timer.
166 if (visualizer_enabled_) {
167 visualizer_.DumpGraph(pass_name, /* is_after_pass= */ false, graph_in_bad_state_);
168 }
169 if (timing_logger_enabled_) {
170 timing_logger_.StartTiming(pass_name);
171 }
172 }
173
FlushVisualizer()174 void FlushVisualizer() REQUIRES(!visualizer_dump_mutex_) {
175 MutexLock mu(Thread::Current(), visualizer_dump_mutex_);
176 *visualizer_output_ << visualizer_oss_.str();
177 visualizer_output_->flush();
178 visualizer_oss_.str("");
179 visualizer_oss_.clear();
180 }
181
EndPass(const char * pass_name,bool pass_change)182 void EndPass(const char* pass_name, bool pass_change) {
183 // Pause timer first, then dump graph.
184 if (timing_logger_enabled_) {
185 timing_logger_.EndTiming();
186 }
187 if (visualizer_enabled_) {
188 visualizer_.DumpGraph(pass_name, /* is_after_pass= */ true, graph_in_bad_state_);
189 }
190
191 // Validate the HGraph if running in debug mode.
192 if (kIsDebugBuild) {
193 if (!graph_in_bad_state_) {
194 GraphChecker checker(graph_);
195 last_seen_graph_size_ = checker.Run(pass_change, last_seen_graph_size_);
196 if (!checker.IsValid()) {
197 LOG(FATAL) << "Error after " << pass_name << ": " << Dumpable<GraphChecker>(checker);
198 }
199 }
200 }
201 }
202
IsVerboseMethod(const CompilerOptions & compiler_options,const char * method_name)203 static bool IsVerboseMethod(const CompilerOptions& compiler_options, const char* method_name) {
204 // Test an exact match to --verbose-methods. If verbose-methods is set, this overrides an
205 // empty kStringFilter matching all methods.
206 if (compiler_options.HasVerboseMethods()) {
207 return compiler_options.IsVerboseMethod(method_name);
208 }
209
210 // Test the kStringFilter sub-string. constexpr helper variable to silence unreachable-code
211 // warning when the string is empty.
212 constexpr bool kStringFilterEmpty = arraysize(kStringFilter) <= 1;
213 if (kStringFilterEmpty || strstr(method_name, kStringFilter) != nullptr) {
214 return true;
215 }
216
217 return false;
218 }
219
220 HGraph* const graph_;
221 size_t last_seen_graph_size_;
222
223 std::string cached_method_name_;
224
225 bool timing_logger_enabled_;
226 TimingLogger timing_logger_;
227
228 DisassemblyInformation disasm_info_;
229
230 std::ostringstream visualizer_oss_;
231 std::ostream* visualizer_output_;
232 bool visualizer_enabled_;
233 HGraphVisualizer visualizer_;
234 Mutex& visualizer_dump_mutex_;
235
236 // Flag to be set by the compiler if the pass failed and the graph is not
237 // expected to validate.
238 bool graph_in_bad_state_;
239
240 friend PassScope;
241
242 DISALLOW_COPY_AND_ASSIGN(PassObserver);
243 };
244
245 class PassScope : public ValueObject {
246 public:
PassScope(const char * pass_name,PassObserver * pass_observer)247 PassScope(const char *pass_name, PassObserver* pass_observer)
248 : pass_name_(pass_name),
249 pass_change_(true), // assume change
250 pass_observer_(pass_observer) {
251 pass_observer_->StartPass(pass_name_);
252 }
253
SetPassNotChanged()254 void SetPassNotChanged() {
255 pass_change_ = false;
256 }
257
~PassScope()258 ~PassScope() {
259 pass_observer_->EndPass(pass_name_, pass_change_);
260 }
261
262 private:
263 const char* const pass_name_;
264 bool pass_change_;
265 PassObserver* const pass_observer_;
266 };
267
268 class OptimizingCompiler final : public Compiler {
269 public:
270 explicit OptimizingCompiler(const CompilerOptions& compiler_options,
271 CompiledMethodStorage* storage);
272 ~OptimizingCompiler() override;
273
274 bool CanCompileMethod(uint32_t method_idx, const DexFile& dex_file) const override;
275
276 CompiledMethod* Compile(const dex::CodeItem* code_item,
277 uint32_t access_flags,
278 InvokeType invoke_type,
279 uint16_t class_def_idx,
280 uint32_t method_idx,
281 Handle<mirror::ClassLoader> class_loader,
282 const DexFile& dex_file,
283 Handle<mirror::DexCache> dex_cache) const override;
284
285 CompiledMethod* JniCompile(uint32_t access_flags,
286 uint32_t method_idx,
287 const DexFile& dex_file,
288 Handle<mirror::DexCache> dex_cache) const override;
289
GetEntryPointOf(ArtMethod * method) const290 uintptr_t GetEntryPointOf(ArtMethod* method) const override
291 REQUIRES_SHARED(Locks::mutator_lock_) {
292 return reinterpret_cast<uintptr_t>(method->GetEntryPointFromQuickCompiledCodePtrSize(
293 InstructionSetPointerSize(GetCompilerOptions().GetInstructionSet())));
294 }
295
296 bool JitCompile(Thread* self,
297 jit::JitCodeCache* code_cache,
298 jit::JitMemoryRegion* region,
299 ArtMethod* method,
300 bool baseline,
301 bool osr,
302 jit::JitLogger* jit_logger)
303 override
304 REQUIRES_SHARED(Locks::mutator_lock_);
305
306 private:
RunOptimizations(HGraph * graph,CodeGenerator * codegen,const DexCompilationUnit & dex_compilation_unit,PassObserver * pass_observer,VariableSizedHandleScope * handles,const OptimizationDef definitions[],size_t length) const307 bool RunOptimizations(HGraph* graph,
308 CodeGenerator* codegen,
309 const DexCompilationUnit& dex_compilation_unit,
310 PassObserver* pass_observer,
311 VariableSizedHandleScope* handles,
312 const OptimizationDef definitions[],
313 size_t length) const {
314 // Convert definitions to optimization passes.
315 ArenaVector<HOptimization*> optimizations = ConstructOptimizations(
316 definitions,
317 length,
318 graph->GetAllocator(),
319 graph,
320 compilation_stats_.get(),
321 codegen,
322 dex_compilation_unit,
323 handles);
324 DCHECK_EQ(length, optimizations.size());
325 // Run the optimization passes one by one. Any "depends_on" pass refers back to
326 // the most recent occurrence of that pass, skipped or executed.
327 std::bitset<static_cast<size_t>(OptimizationPass::kLast) + 1u> pass_changes;
328 pass_changes[static_cast<size_t>(OptimizationPass::kNone)] = true;
329 bool change = false;
330 for (size_t i = 0; i < length; ++i) {
331 if (pass_changes[static_cast<size_t>(definitions[i].depends_on)]) {
332 // Execute the pass and record whether it changed anything.
333 PassScope scope(optimizations[i]->GetPassName(), pass_observer);
334 bool pass_change = optimizations[i]->Run();
335 pass_changes[static_cast<size_t>(definitions[i].pass)] = pass_change;
336 if (pass_change) {
337 change = true;
338 } else {
339 scope.SetPassNotChanged();
340 }
341 } else {
342 // Skip the pass and record that nothing changed.
343 pass_changes[static_cast<size_t>(definitions[i].pass)] = false;
344 }
345 }
346 return change;
347 }
348
RunOptimizations(HGraph * graph,CodeGenerator * codegen,const DexCompilationUnit & dex_compilation_unit,PassObserver * pass_observer,VariableSizedHandleScope * handles,const OptimizationDef (& definitions)[length]) const349 template <size_t length> bool RunOptimizations(
350 HGraph* graph,
351 CodeGenerator* codegen,
352 const DexCompilationUnit& dex_compilation_unit,
353 PassObserver* pass_observer,
354 VariableSizedHandleScope* handles,
355 const OptimizationDef (&definitions)[length]) const {
356 return RunOptimizations(
357 graph, codegen, dex_compilation_unit, pass_observer, handles, definitions, length);
358 }
359
360 void RunOptimizations(HGraph* graph,
361 CodeGenerator* codegen,
362 const DexCompilationUnit& dex_compilation_unit,
363 PassObserver* pass_observer,
364 VariableSizedHandleScope* handles) const;
365
366 private:
367 // Create a 'CompiledMethod' for an optimized graph.
368 CompiledMethod* Emit(ArenaAllocator* allocator,
369 CodeVectorAllocator* code_allocator,
370 CodeGenerator* codegen,
371 const dex::CodeItem* item) const;
372
373 // Try compiling a method and return the code generator used for
374 // compiling it.
375 // This method:
376 // 1) Builds the graph. Returns null if it failed to build it.
377 // 2) Transforms the graph to SSA. Returns null if it failed.
378 // 3) Runs optimizations on the graph, including register allocator.
379 // 4) Generates code with the `code_allocator` provided.
380 CodeGenerator* TryCompile(ArenaAllocator* allocator,
381 ArenaStack* arena_stack,
382 CodeVectorAllocator* code_allocator,
383 const DexCompilationUnit& dex_compilation_unit,
384 ArtMethod* method,
385 bool baseline,
386 bool osr,
387 bool is_shared_jit_code,
388 VariableSizedHandleScope* handles) const;
389
390 CodeGenerator* TryCompileIntrinsic(ArenaAllocator* allocator,
391 ArenaStack* arena_stack,
392 CodeVectorAllocator* code_allocator,
393 const DexCompilationUnit& dex_compilation_unit,
394 ArtMethod* method,
395 VariableSizedHandleScope* handles) const;
396
397 bool RunArchOptimizations(HGraph* graph,
398 CodeGenerator* codegen,
399 const DexCompilationUnit& dex_compilation_unit,
400 PassObserver* pass_observer,
401 VariableSizedHandleScope* handles) const;
402
403 bool RunBaselineOptimizations(HGraph* graph,
404 CodeGenerator* codegen,
405 const DexCompilationUnit& dex_compilation_unit,
406 PassObserver* pass_observer,
407 VariableSizedHandleScope* handles) const;
408
409 void GenerateJitDebugInfo(const debug::MethodDebugInfo& method_debug_info);
410
411 std::unique_ptr<OptimizingCompilerStats> compilation_stats_;
412
413 std::unique_ptr<std::ostream> visualizer_output_;
414
415 mutable Mutex dump_mutex_; // To synchronize visualizer writing.
416
417 DISALLOW_COPY_AND_ASSIGN(OptimizingCompiler);
418 };
419
420 static const int kMaximumCompilationTimeBeforeWarning = 100; /* ms */
421
OptimizingCompiler(const CompilerOptions & compiler_options,CompiledMethodStorage * storage)422 OptimizingCompiler::OptimizingCompiler(const CompilerOptions& compiler_options,
423 CompiledMethodStorage* storage)
424 : Compiler(compiler_options, storage, kMaximumCompilationTimeBeforeWarning),
425 dump_mutex_("Visualizer dump lock") {
426 // Enable C1visualizer output.
427 const std::string& cfg_file_name = compiler_options.GetDumpCfgFileName();
428 if (!cfg_file_name.empty()) {
429 std::ios_base::openmode cfg_file_mode =
430 compiler_options.GetDumpCfgAppend() ? std::ofstream::app : std::ofstream::out;
431 visualizer_output_.reset(new std::ofstream(cfg_file_name, cfg_file_mode));
432 }
433 if (compiler_options.GetDumpStats()) {
434 compilation_stats_.reset(new OptimizingCompilerStats());
435 }
436 }
437
~OptimizingCompiler()438 OptimizingCompiler::~OptimizingCompiler() {
439 if (compilation_stats_.get() != nullptr) {
440 compilation_stats_->Log();
441 }
442 }
443
CanCompileMethod(uint32_t method_idx ATTRIBUTE_UNUSED,const DexFile & dex_file ATTRIBUTE_UNUSED) const444 bool OptimizingCompiler::CanCompileMethod(uint32_t method_idx ATTRIBUTE_UNUSED,
445 const DexFile& dex_file ATTRIBUTE_UNUSED) const {
446 return true;
447 }
448
IsInstructionSetSupported(InstructionSet instruction_set)449 static bool IsInstructionSetSupported(InstructionSet instruction_set) {
450 return instruction_set == InstructionSet::kArm
451 || instruction_set == InstructionSet::kArm64
452 || instruction_set == InstructionSet::kThumb2
453 || instruction_set == InstructionSet::kX86
454 || instruction_set == InstructionSet::kX86_64;
455 }
456
RunBaselineOptimizations(HGraph * graph,CodeGenerator * codegen,const DexCompilationUnit & dex_compilation_unit,PassObserver * pass_observer,VariableSizedHandleScope * handles) const457 bool OptimizingCompiler::RunBaselineOptimizations(HGraph* graph,
458 CodeGenerator* codegen,
459 const DexCompilationUnit& dex_compilation_unit,
460 PassObserver* pass_observer,
461 VariableSizedHandleScope* handles) const {
462 switch (codegen->GetCompilerOptions().GetInstructionSet()) {
463 #ifdef ART_ENABLE_CODEGEN_x86
464 case InstructionSet::kX86: {
465 OptimizationDef x86_optimizations[] = {
466 OptDef(OptimizationPass::kPcRelativeFixupsX86),
467 };
468 return RunOptimizations(graph,
469 codegen,
470 dex_compilation_unit,
471 pass_observer,
472 handles,
473 x86_optimizations);
474 }
475 #endif
476 default:
477 UNUSED(graph);
478 UNUSED(codegen);
479 UNUSED(dex_compilation_unit);
480 UNUSED(pass_observer);
481 UNUSED(handles);
482 return false;
483 }
484 }
485
RunArchOptimizations(HGraph * graph,CodeGenerator * codegen,const DexCompilationUnit & dex_compilation_unit,PassObserver * pass_observer,VariableSizedHandleScope * handles) const486 bool OptimizingCompiler::RunArchOptimizations(HGraph* graph,
487 CodeGenerator* codegen,
488 const DexCompilationUnit& dex_compilation_unit,
489 PassObserver* pass_observer,
490 VariableSizedHandleScope* handles) const {
491 switch (codegen->GetCompilerOptions().GetInstructionSet()) {
492 #if defined(ART_ENABLE_CODEGEN_arm)
493 case InstructionSet::kThumb2:
494 case InstructionSet::kArm: {
495 OptimizationDef arm_optimizations[] = {
496 OptDef(OptimizationPass::kInstructionSimplifierArm),
497 OptDef(OptimizationPass::kSideEffectsAnalysis),
498 OptDef(OptimizationPass::kGlobalValueNumbering, "GVN$after_arch"),
499 OptDef(OptimizationPass::kScheduling)
500 };
501 return RunOptimizations(graph,
502 codegen,
503 dex_compilation_unit,
504 pass_observer,
505 handles,
506 arm_optimizations);
507 }
508 #endif
509 #ifdef ART_ENABLE_CODEGEN_arm64
510 case InstructionSet::kArm64: {
511 OptimizationDef arm64_optimizations[] = {
512 OptDef(OptimizationPass::kInstructionSimplifierArm64),
513 OptDef(OptimizationPass::kSideEffectsAnalysis),
514 OptDef(OptimizationPass::kGlobalValueNumbering, "GVN$after_arch"),
515 OptDef(OptimizationPass::kScheduling)
516 };
517 return RunOptimizations(graph,
518 codegen,
519 dex_compilation_unit,
520 pass_observer,
521 handles,
522 arm64_optimizations);
523 }
524 #endif
525 #ifdef ART_ENABLE_CODEGEN_x86
526 case InstructionSet::kX86: {
527 OptimizationDef x86_optimizations[] = {
528 OptDef(OptimizationPass::kInstructionSimplifierX86),
529 OptDef(OptimizationPass::kSideEffectsAnalysis),
530 OptDef(OptimizationPass::kGlobalValueNumbering, "GVN$after_arch"),
531 OptDef(OptimizationPass::kPcRelativeFixupsX86),
532 OptDef(OptimizationPass::kX86MemoryOperandGeneration)
533 };
534 return RunOptimizations(graph,
535 codegen,
536 dex_compilation_unit,
537 pass_observer,
538 handles,
539 x86_optimizations);
540 }
541 #endif
542 #ifdef ART_ENABLE_CODEGEN_x86_64
543 case InstructionSet::kX86_64: {
544 OptimizationDef x86_64_optimizations[] = {
545 OptDef(OptimizationPass::kInstructionSimplifierX86_64),
546 OptDef(OptimizationPass::kSideEffectsAnalysis),
547 OptDef(OptimizationPass::kGlobalValueNumbering, "GVN$after_arch"),
548 OptDef(OptimizationPass::kX86MemoryOperandGeneration)
549 };
550 return RunOptimizations(graph,
551 codegen,
552 dex_compilation_unit,
553 pass_observer,
554 handles,
555 x86_64_optimizations);
556 }
557 #endif
558 default:
559 return false;
560 }
561 }
562
563 NO_INLINE // Avoid increasing caller's frame size by large stack-allocated objects.
AllocateRegisters(HGraph * graph,CodeGenerator * codegen,PassObserver * pass_observer,RegisterAllocator::Strategy strategy,OptimizingCompilerStats * stats)564 static void AllocateRegisters(HGraph* graph,
565 CodeGenerator* codegen,
566 PassObserver* pass_observer,
567 RegisterAllocator::Strategy strategy,
568 OptimizingCompilerStats* stats) {
569 {
570 PassScope scope(PrepareForRegisterAllocation::kPrepareForRegisterAllocationPassName,
571 pass_observer);
572 PrepareForRegisterAllocation(graph, codegen->GetCompilerOptions(), stats).Run();
573 }
574 // Use local allocator shared by SSA liveness analysis and register allocator.
575 // (Register allocator creates new objects in the liveness data.)
576 ScopedArenaAllocator local_allocator(graph->GetArenaStack());
577 SsaLivenessAnalysis liveness(graph, codegen, &local_allocator);
578 {
579 PassScope scope(SsaLivenessAnalysis::kLivenessPassName, pass_observer);
580 liveness.Analyze();
581 }
582 {
583 PassScope scope(RegisterAllocator::kRegisterAllocatorPassName, pass_observer);
584 std::unique_ptr<RegisterAllocator> register_allocator =
585 RegisterAllocator::Create(&local_allocator, codegen, liveness, strategy);
586 register_allocator->AllocateRegisters();
587 }
588 }
589
590 // Strip pass name suffix to get optimization name.
ConvertPassNameToOptimizationName(const std::string & pass_name)591 static std::string ConvertPassNameToOptimizationName(const std::string& pass_name) {
592 size_t pos = pass_name.find(kPassNameSeparator);
593 return pos == std::string::npos ? pass_name : pass_name.substr(0, pos);
594 }
595
RunOptimizations(HGraph * graph,CodeGenerator * codegen,const DexCompilationUnit & dex_compilation_unit,PassObserver * pass_observer,VariableSizedHandleScope * handles) const596 void OptimizingCompiler::RunOptimizations(HGraph* graph,
597 CodeGenerator* codegen,
598 const DexCompilationUnit& dex_compilation_unit,
599 PassObserver* pass_observer,
600 VariableSizedHandleScope* handles) const {
601 const std::vector<std::string>* pass_names = GetCompilerOptions().GetPassesToRun();
602 if (pass_names != nullptr) {
603 // If passes were defined on command-line, build the optimization
604 // passes and run these instead of the built-in optimizations.
605 // TODO: a way to define depends_on via command-line?
606 const size_t length = pass_names->size();
607 std::vector<OptimizationDef> optimizations;
608 for (const std::string& pass_name : *pass_names) {
609 std::string opt_name = ConvertPassNameToOptimizationName(pass_name);
610 optimizations.push_back(OptDef(OptimizationPassByName(opt_name), pass_name.c_str()));
611 }
612 RunOptimizations(graph,
613 codegen,
614 dex_compilation_unit,
615 pass_observer,
616 handles,
617 optimizations.data(),
618 length);
619 return;
620 }
621
622 OptimizationDef optimizations[] = {
623 // Initial optimizations.
624 OptDef(OptimizationPass::kConstantFolding),
625 OptDef(OptimizationPass::kInstructionSimplifier),
626 OptDef(OptimizationPass::kDeadCodeElimination,
627 "dead_code_elimination$initial"),
628 // Inlining.
629 OptDef(OptimizationPass::kInliner),
630 // Simplification (only if inlining occurred).
631 OptDef(OptimizationPass::kConstantFolding,
632 "constant_folding$after_inlining",
633 OptimizationPass::kInliner),
634 OptDef(OptimizationPass::kInstructionSimplifier,
635 "instruction_simplifier$after_inlining",
636 OptimizationPass::kInliner),
637 OptDef(OptimizationPass::kDeadCodeElimination,
638 "dead_code_elimination$after_inlining",
639 OptimizationPass::kInliner),
640 // GVN.
641 OptDef(OptimizationPass::kSideEffectsAnalysis,
642 "side_effects$before_gvn"),
643 OptDef(OptimizationPass::kGlobalValueNumbering),
644 // Simplification (TODO: only if GVN occurred).
645 OptDef(OptimizationPass::kSelectGenerator),
646 OptDef(OptimizationPass::kConstantFolding,
647 "constant_folding$after_gvn"),
648 OptDef(OptimizationPass::kInstructionSimplifier,
649 "instruction_simplifier$after_gvn"),
650 OptDef(OptimizationPass::kDeadCodeElimination,
651 "dead_code_elimination$after_gvn"),
652 // High-level optimizations.
653 OptDef(OptimizationPass::kSideEffectsAnalysis,
654 "side_effects$before_licm"),
655 OptDef(OptimizationPass::kInvariantCodeMotion),
656 OptDef(OptimizationPass::kInductionVarAnalysis),
657 OptDef(OptimizationPass::kBoundsCheckElimination),
658 OptDef(OptimizationPass::kLoopOptimization),
659 // Simplification.
660 OptDef(OptimizationPass::kConstantFolding,
661 "constant_folding$after_bce"),
662 OptDef(OptimizationPass::kInstructionSimplifier,
663 "instruction_simplifier$after_bce"),
664 // Other high-level optimizations.
665 OptDef(OptimizationPass::kSideEffectsAnalysis,
666 "side_effects$before_lse"),
667 OptDef(OptimizationPass::kLoadStoreAnalysis),
668 OptDef(OptimizationPass::kLoadStoreElimination),
669 OptDef(OptimizationPass::kCHAGuardOptimization),
670 OptDef(OptimizationPass::kDeadCodeElimination,
671 "dead_code_elimination$final"),
672 OptDef(OptimizationPass::kCodeSinking),
673 // The codegen has a few assumptions that only the instruction simplifier
674 // can satisfy. For example, the code generator does not expect to see a
675 // HTypeConversion from a type to the same type.
676 OptDef(OptimizationPass::kInstructionSimplifier,
677 "instruction_simplifier$before_codegen"),
678 // Eliminate constructor fences after code sinking to avoid
679 // complicated sinking logic to split a fence with many inputs.
680 OptDef(OptimizationPass::kConstructorFenceRedundancyElimination)
681 };
682 RunOptimizations(graph,
683 codegen,
684 dex_compilation_unit,
685 pass_observer,
686 handles,
687 optimizations);
688
689 RunArchOptimizations(graph, codegen, dex_compilation_unit, pass_observer, handles);
690 }
691
EmitAndSortLinkerPatches(CodeGenerator * codegen)692 static ArenaVector<linker::LinkerPatch> EmitAndSortLinkerPatches(CodeGenerator* codegen) {
693 ArenaVector<linker::LinkerPatch> linker_patches(codegen->GetGraph()->GetAllocator()->Adapter());
694 codegen->EmitLinkerPatches(&linker_patches);
695
696 // Sort patches by literal offset. Required for .oat_patches encoding.
697 std::sort(linker_patches.begin(), linker_patches.end(),
698 [](const linker::LinkerPatch& lhs, const linker::LinkerPatch& rhs) {
699 return lhs.LiteralOffset() < rhs.LiteralOffset();
700 });
701
702 return linker_patches;
703 }
704
Emit(ArenaAllocator * allocator,CodeVectorAllocator * code_allocator,CodeGenerator * codegen,const dex::CodeItem * code_item_for_osr_check) const705 CompiledMethod* OptimizingCompiler::Emit(ArenaAllocator* allocator,
706 CodeVectorAllocator* code_allocator,
707 CodeGenerator* codegen,
708 const dex::CodeItem* code_item_for_osr_check) const {
709 ArenaVector<linker::LinkerPatch> linker_patches = EmitAndSortLinkerPatches(codegen);
710 ScopedArenaVector<uint8_t> stack_map = codegen->BuildStackMaps(code_item_for_osr_check);
711
712 CompiledMethodStorage* storage = GetCompiledMethodStorage();
713 CompiledMethod* compiled_method = CompiledMethod::SwapAllocCompiledMethod(
714 storage,
715 codegen->GetInstructionSet(),
716 code_allocator->GetMemory(),
717 ArrayRef<const uint8_t>(stack_map),
718 ArrayRef<const uint8_t>(*codegen->GetAssembler()->cfi().data()),
719 ArrayRef<const linker::LinkerPatch>(linker_patches));
720
721 for (const linker::LinkerPatch& patch : linker_patches) {
722 if (codegen->NeedsThunkCode(patch) && storage->GetThunkCode(patch).empty()) {
723 ArenaVector<uint8_t> code(allocator->Adapter());
724 std::string debug_name;
725 codegen->EmitThunkCode(patch, &code, &debug_name);
726 storage->SetThunkCode(patch, ArrayRef<const uint8_t>(code), debug_name);
727 }
728 }
729
730 return compiled_method;
731 }
732
TryCompile(ArenaAllocator * allocator,ArenaStack * arena_stack,CodeVectorAllocator * code_allocator,const DexCompilationUnit & dex_compilation_unit,ArtMethod * method,bool baseline,bool osr,bool is_shared_jit_code,VariableSizedHandleScope * handles) const733 CodeGenerator* OptimizingCompiler::TryCompile(ArenaAllocator* allocator,
734 ArenaStack* arena_stack,
735 CodeVectorAllocator* code_allocator,
736 const DexCompilationUnit& dex_compilation_unit,
737 ArtMethod* method,
738 bool baseline,
739 bool osr,
740 bool is_shared_jit_code,
741 VariableSizedHandleScope* handles) const {
742 MaybeRecordStat(compilation_stats_.get(), MethodCompilationStat::kAttemptBytecodeCompilation);
743 const CompilerOptions& compiler_options = GetCompilerOptions();
744 InstructionSet instruction_set = compiler_options.GetInstructionSet();
745 const DexFile& dex_file = *dex_compilation_unit.GetDexFile();
746 uint32_t method_idx = dex_compilation_unit.GetDexMethodIndex();
747 const dex::CodeItem* code_item = dex_compilation_unit.GetCodeItem();
748
749 // Always use the Thumb-2 assembler: some runtime functionality
750 // (like implicit stack overflow checks) assume Thumb-2.
751 DCHECK_NE(instruction_set, InstructionSet::kArm);
752
753 // Do not attempt to compile on architectures we do not support.
754 if (!IsInstructionSetSupported(instruction_set)) {
755 MaybeRecordStat(compilation_stats_.get(),
756 MethodCompilationStat::kNotCompiledUnsupportedIsa);
757 return nullptr;
758 }
759
760 if (Compiler::IsPathologicalCase(*code_item, method_idx, dex_file)) {
761 MaybeRecordStat(compilation_stats_.get(), MethodCompilationStat::kNotCompiledPathological);
762 return nullptr;
763 }
764
765 // Implementation of the space filter: do not compile a code item whose size in
766 // code units is bigger than 128.
767 static constexpr size_t kSpaceFilterOptimizingThreshold = 128;
768 if ((compiler_options.GetCompilerFilter() == CompilerFilter::kSpace)
769 && (CodeItemInstructionAccessor(dex_file, code_item).InsnsSizeInCodeUnits() >
770 kSpaceFilterOptimizingThreshold)) {
771 MaybeRecordStat(compilation_stats_.get(), MethodCompilationStat::kNotCompiledSpaceFilter);
772 return nullptr;
773 }
774
775 CodeItemDebugInfoAccessor code_item_accessor(dex_file, code_item, method_idx);
776
777 bool dead_reference_safe;
778 ArrayRef<const uint8_t> interpreter_metadata;
779 // For AOT compilation, we may not get a method, for example if its class is erroneous,
780 // possibly due to an unavailable superclass. JIT should always have a method.
781 DCHECK(Runtime::Current()->IsAotCompiler() || method != nullptr);
782 if (method != nullptr) {
783 const dex::ClassDef* containing_class;
784 {
785 ScopedObjectAccess soa(Thread::Current());
786 containing_class = &method->GetClassDef();
787 interpreter_metadata = method->GetQuickenedInfo();
788 }
789 // MethodContainsRSensitiveAccess is currently slow, but HasDeadReferenceSafeAnnotation()
790 // is currently rarely true.
791 dead_reference_safe =
792 annotations::HasDeadReferenceSafeAnnotation(dex_file, *containing_class)
793 && !annotations::MethodContainsRSensitiveAccess(dex_file, *containing_class, method_idx);
794 } else {
795 // If we could not resolve the class, conservatively assume it's dead-reference unsafe.
796 dead_reference_safe = false;
797 }
798
799 HGraph* graph = new (allocator) HGraph(
800 allocator,
801 arena_stack,
802 dex_file,
803 method_idx,
804 compiler_options.GetInstructionSet(),
805 kInvalidInvokeType,
806 dead_reference_safe,
807 compiler_options.GetDebuggable(),
808 /* osr= */ osr,
809 /* is_shared_jit_code= */ is_shared_jit_code,
810 /* baseline= */ baseline);
811
812 if (method != nullptr) {
813 graph->SetArtMethod(method);
814 }
815
816 std::unique_ptr<CodeGenerator> codegen(
817 CodeGenerator::Create(graph,
818 compiler_options,
819 compilation_stats_.get()));
820 if (codegen.get() == nullptr) {
821 MaybeRecordStat(compilation_stats_.get(), MethodCompilationStat::kNotCompiledNoCodegen);
822 return nullptr;
823 }
824 codegen->GetAssembler()->cfi().SetEnabled(compiler_options.GenerateAnyDebugInfo());
825
826 PassObserver pass_observer(graph,
827 codegen.get(),
828 visualizer_output_.get(),
829 compiler_options,
830 dump_mutex_);
831
832 {
833 VLOG(compiler) << "Building " << pass_observer.GetMethodName();
834 PassScope scope(HGraphBuilder::kBuilderPassName, &pass_observer);
835 HGraphBuilder builder(graph,
836 code_item_accessor,
837 &dex_compilation_unit,
838 &dex_compilation_unit,
839 codegen.get(),
840 compilation_stats_.get(),
841 interpreter_metadata,
842 handles);
843 GraphAnalysisResult result = builder.BuildGraph();
844 if (result != kAnalysisSuccess) {
845 switch (result) {
846 case kAnalysisSkipped: {
847 MaybeRecordStat(compilation_stats_.get(),
848 MethodCompilationStat::kNotCompiledSkipped);
849 break;
850 }
851 case kAnalysisInvalidBytecode: {
852 MaybeRecordStat(compilation_stats_.get(),
853 MethodCompilationStat::kNotCompiledInvalidBytecode);
854 break;
855 }
856 case kAnalysisFailThrowCatchLoop: {
857 MaybeRecordStat(compilation_stats_.get(),
858 MethodCompilationStat::kNotCompiledThrowCatchLoop);
859 break;
860 }
861 case kAnalysisFailAmbiguousArrayOp: {
862 MaybeRecordStat(compilation_stats_.get(),
863 MethodCompilationStat::kNotCompiledAmbiguousArrayOp);
864 break;
865 }
866 case kAnalysisFailIrreducibleLoopAndStringInit: {
867 MaybeRecordStat(compilation_stats_.get(),
868 MethodCompilationStat::kNotCompiledIrreducibleLoopAndStringInit);
869 break;
870 }
871 case kAnalysisFailPhiEquivalentInOsr: {
872 MaybeRecordStat(compilation_stats_.get(),
873 MethodCompilationStat::kNotCompiledPhiEquivalentInOsr);
874 break;
875 }
876 case kAnalysisSuccess:
877 UNREACHABLE();
878 }
879 pass_observer.SetGraphInBadState();
880 return nullptr;
881 }
882 }
883
884 if (baseline) {
885 RunBaselineOptimizations(graph, codegen.get(), dex_compilation_unit, &pass_observer, handles);
886 } else {
887 RunOptimizations(graph, codegen.get(), dex_compilation_unit, &pass_observer, handles);
888 }
889
890 RegisterAllocator::Strategy regalloc_strategy =
891 compiler_options.GetRegisterAllocationStrategy();
892 AllocateRegisters(graph,
893 codegen.get(),
894 &pass_observer,
895 regalloc_strategy,
896 compilation_stats_.get());
897
898 codegen->Compile(code_allocator);
899 pass_observer.DumpDisassembly();
900
901 MaybeRecordStat(compilation_stats_.get(), MethodCompilationStat::kCompiledBytecode);
902 return codegen.release();
903 }
904
TryCompileIntrinsic(ArenaAllocator * allocator,ArenaStack * arena_stack,CodeVectorAllocator * code_allocator,const DexCompilationUnit & dex_compilation_unit,ArtMethod * method,VariableSizedHandleScope * handles) const905 CodeGenerator* OptimizingCompiler::TryCompileIntrinsic(
906 ArenaAllocator* allocator,
907 ArenaStack* arena_stack,
908 CodeVectorAllocator* code_allocator,
909 const DexCompilationUnit& dex_compilation_unit,
910 ArtMethod* method,
911 VariableSizedHandleScope* handles) const {
912 MaybeRecordStat(compilation_stats_.get(), MethodCompilationStat::kAttemptIntrinsicCompilation);
913 const CompilerOptions& compiler_options = GetCompilerOptions();
914 InstructionSet instruction_set = compiler_options.GetInstructionSet();
915 const DexFile& dex_file = *dex_compilation_unit.GetDexFile();
916 uint32_t method_idx = dex_compilation_unit.GetDexMethodIndex();
917
918 // Always use the Thumb-2 assembler: some runtime functionality
919 // (like implicit stack overflow checks) assume Thumb-2.
920 DCHECK_NE(instruction_set, InstructionSet::kArm);
921
922 // Do not attempt to compile on architectures we do not support.
923 if (!IsInstructionSetSupported(instruction_set)) {
924 return nullptr;
925 }
926
927 HGraph* graph = new (allocator) HGraph(
928 allocator,
929 arena_stack,
930 dex_file,
931 method_idx,
932 compiler_options.GetInstructionSet(),
933 kInvalidInvokeType,
934 /* dead_reference_safe= */ true, // Intrinsics don't affect dead reference safety.
935 compiler_options.GetDebuggable(),
936 /* osr= */ false);
937
938 DCHECK(Runtime::Current()->IsAotCompiler());
939 DCHECK(method != nullptr);
940 graph->SetArtMethod(method);
941
942 std::unique_ptr<CodeGenerator> codegen(
943 CodeGenerator::Create(graph,
944 compiler_options,
945 compilation_stats_.get()));
946 if (codegen.get() == nullptr) {
947 return nullptr;
948 }
949 codegen->GetAssembler()->cfi().SetEnabled(compiler_options.GenerateAnyDebugInfo());
950
951 PassObserver pass_observer(graph,
952 codegen.get(),
953 visualizer_output_.get(),
954 compiler_options,
955 dump_mutex_);
956
957 {
958 VLOG(compiler) << "Building intrinsic graph " << pass_observer.GetMethodName();
959 PassScope scope(HGraphBuilder::kBuilderPassName, &pass_observer);
960 HGraphBuilder builder(graph,
961 CodeItemDebugInfoAccessor(), // Null code item.
962 &dex_compilation_unit,
963 &dex_compilation_unit,
964 codegen.get(),
965 compilation_stats_.get(),
966 /* interpreter_metadata= */ ArrayRef<const uint8_t>(),
967 handles);
968 builder.BuildIntrinsicGraph(method);
969 }
970
971 OptimizationDef optimizations[] = {
972 // The codegen has a few assumptions that only the instruction simplifier
973 // can satisfy.
974 OptDef(OptimizationPass::kInstructionSimplifier),
975 };
976 RunOptimizations(graph,
977 codegen.get(),
978 dex_compilation_unit,
979 &pass_observer,
980 handles,
981 optimizations);
982
983 RunArchOptimizations(graph, codegen.get(), dex_compilation_unit, &pass_observer, handles);
984
985 AllocateRegisters(graph,
986 codegen.get(),
987 &pass_observer,
988 compiler_options.GetRegisterAllocationStrategy(),
989 compilation_stats_.get());
990 if (!codegen->IsLeafMethod()) {
991 VLOG(compiler) << "Intrinsic method is not leaf: " << method->GetIntrinsic()
992 << " " << graph->PrettyMethod();
993 return nullptr;
994 }
995
996 codegen->Compile(code_allocator);
997 pass_observer.DumpDisassembly();
998
999 VLOG(compiler) << "Compiled intrinsic: " << method->GetIntrinsic()
1000 << " " << graph->PrettyMethod();
1001 MaybeRecordStat(compilation_stats_.get(), MethodCompilationStat::kCompiledIntrinsic);
1002 return codegen.release();
1003 }
1004
Compile(const dex::CodeItem * code_item,uint32_t access_flags,InvokeType invoke_type,uint16_t class_def_idx,uint32_t method_idx,Handle<mirror::ClassLoader> jclass_loader,const DexFile & dex_file,Handle<mirror::DexCache> dex_cache) const1005 CompiledMethod* OptimizingCompiler::Compile(const dex::CodeItem* code_item,
1006 uint32_t access_flags,
1007 InvokeType invoke_type,
1008 uint16_t class_def_idx,
1009 uint32_t method_idx,
1010 Handle<mirror::ClassLoader> jclass_loader,
1011 const DexFile& dex_file,
1012 Handle<mirror::DexCache> dex_cache) const {
1013 const CompilerOptions& compiler_options = GetCompilerOptions();
1014 CompiledMethod* compiled_method = nullptr;
1015 Runtime* runtime = Runtime::Current();
1016 DCHECK(runtime->IsAotCompiler());
1017 const VerifiedMethod* verified_method = compiler_options.GetVerifiedMethod(&dex_file, method_idx);
1018 DCHECK(!verified_method->HasRuntimeThrow());
1019 if (compiler_options.IsMethodVerifiedWithoutFailures(method_idx, class_def_idx, dex_file) ||
1020 verifier::CanCompilerHandleVerificationFailure(
1021 verified_method->GetEncounteredVerificationFailures())) {
1022 ArenaAllocator allocator(runtime->GetArenaPool());
1023 ArenaStack arena_stack(runtime->GetArenaPool());
1024 CodeVectorAllocator code_allocator(&allocator);
1025 std::unique_ptr<CodeGenerator> codegen;
1026 bool compiled_intrinsic = false;
1027 {
1028 ScopedObjectAccess soa(Thread::Current());
1029 ArtMethod* method =
1030 runtime->GetClassLinker()->ResolveMethod<ClassLinker::ResolveMode::kCheckICCEAndIAE>(
1031 method_idx, dex_cache, jclass_loader, /*referrer=*/ nullptr, invoke_type);
1032 DCHECK_EQ(method == nullptr, soa.Self()->IsExceptionPending());
1033 soa.Self()->ClearException(); // Suppress exception if any.
1034 VariableSizedHandleScope handles(soa.Self());
1035 Handle<mirror::Class> compiling_class =
1036 handles.NewHandle(method != nullptr ? method->GetDeclaringClass() : nullptr);
1037 DexCompilationUnit dex_compilation_unit(
1038 jclass_loader,
1039 runtime->GetClassLinker(),
1040 dex_file,
1041 code_item,
1042 class_def_idx,
1043 method_idx,
1044 access_flags,
1045 /*verified_method=*/ nullptr, // Not needed by the Optimizing compiler.
1046 dex_cache,
1047 compiling_class);
1048 // Go to native so that we don't block GC during compilation.
1049 ScopedThreadSuspension sts(soa.Self(), kNative);
1050 if (method != nullptr && UNLIKELY(method->IsIntrinsic())) {
1051 DCHECK(compiler_options.IsBootImage());
1052 codegen.reset(
1053 TryCompileIntrinsic(&allocator,
1054 &arena_stack,
1055 &code_allocator,
1056 dex_compilation_unit,
1057 method,
1058 &handles));
1059 if (codegen != nullptr) {
1060 compiled_intrinsic = true;
1061 }
1062 }
1063 if (codegen == nullptr) {
1064 codegen.reset(
1065 TryCompile(&allocator,
1066 &arena_stack,
1067 &code_allocator,
1068 dex_compilation_unit,
1069 method,
1070 compiler_options.IsBaseline(),
1071 /* osr= */ false,
1072 /* is_shared_jit_code= */ false,
1073 &handles));
1074 }
1075 }
1076 if (codegen.get() != nullptr) {
1077 compiled_method = Emit(&allocator,
1078 &code_allocator,
1079 codegen.get(),
1080 compiled_intrinsic ? nullptr : code_item);
1081 if (compiled_intrinsic) {
1082 compiled_method->MarkAsIntrinsic();
1083 }
1084
1085 if (kArenaAllocatorCountAllocations) {
1086 codegen.reset(); // Release codegen's ScopedArenaAllocator for memory accounting.
1087 size_t total_allocated = allocator.BytesAllocated() + arena_stack.PeakBytesAllocated();
1088 if (total_allocated > kArenaAllocatorMemoryReportThreshold) {
1089 MemStats mem_stats(allocator.GetMemStats());
1090 MemStats peak_stats(arena_stack.GetPeakStats());
1091 LOG(INFO) << "Used " << total_allocated << " bytes of arena memory for compiling "
1092 << dex_file.PrettyMethod(method_idx)
1093 << "\n" << Dumpable<MemStats>(mem_stats)
1094 << "\n" << Dumpable<MemStats>(peak_stats);
1095 }
1096 }
1097 }
1098 } else {
1099 MethodCompilationStat method_stat;
1100 if (compiler_options.VerifyAtRuntime()) {
1101 method_stat = MethodCompilationStat::kNotCompiledVerifyAtRuntime;
1102 } else {
1103 method_stat = MethodCompilationStat::kNotCompiledVerificationError;
1104 }
1105 MaybeRecordStat(compilation_stats_.get(), method_stat);
1106 }
1107
1108 if (kIsDebugBuild &&
1109 compiler_options.CompilingWithCoreImage() &&
1110 IsInstructionSetSupported(compiler_options.GetInstructionSet())) {
1111 // For testing purposes, we put a special marker on method names
1112 // that should be compiled with this compiler (when the
1113 // instruction set is supported). This makes sure we're not
1114 // regressing.
1115 std::string method_name = dex_file.PrettyMethod(method_idx);
1116 bool shouldCompile = method_name.find("$opt$") != std::string::npos;
1117 DCHECK((compiled_method != nullptr) || !shouldCompile) << "Didn't compile " << method_name;
1118 }
1119
1120 return compiled_method;
1121 }
1122
CreateJniStackMap(ScopedArenaAllocator * allocator,const JniCompiledMethod & jni_compiled_method)1123 static ScopedArenaVector<uint8_t> CreateJniStackMap(ScopedArenaAllocator* allocator,
1124 const JniCompiledMethod& jni_compiled_method) {
1125 // StackMapStream is quite large, so allocate it using the ScopedArenaAllocator
1126 // to stay clear of the frame size limit.
1127 std::unique_ptr<StackMapStream> stack_map_stream(
1128 new (allocator) StackMapStream(allocator, jni_compiled_method.GetInstructionSet()));
1129 stack_map_stream->BeginMethod(
1130 jni_compiled_method.GetFrameSize(),
1131 jni_compiled_method.GetCoreSpillMask(),
1132 jni_compiled_method.GetFpSpillMask(),
1133 /* num_dex_registers= */ 0,
1134 /* baseline= */ false);
1135 stack_map_stream->EndMethod();
1136 return stack_map_stream->Encode();
1137 }
1138
JniCompile(uint32_t access_flags,uint32_t method_idx,const DexFile & dex_file,Handle<mirror::DexCache> dex_cache) const1139 CompiledMethod* OptimizingCompiler::JniCompile(uint32_t access_flags,
1140 uint32_t method_idx,
1141 const DexFile& dex_file,
1142 Handle<mirror::DexCache> dex_cache) const {
1143 Runtime* runtime = Runtime::Current();
1144 ArenaAllocator allocator(runtime->GetArenaPool());
1145 ArenaStack arena_stack(runtime->GetArenaPool());
1146
1147 const CompilerOptions& compiler_options = GetCompilerOptions();
1148 if (compiler_options.IsBootImage()) {
1149 ScopedObjectAccess soa(Thread::Current());
1150 ArtMethod* method = runtime->GetClassLinker()->LookupResolvedMethod(
1151 method_idx, dex_cache.Get(), /*class_loader=*/ nullptr);
1152 if (method != nullptr && UNLIKELY(method->IsIntrinsic())) {
1153 VariableSizedHandleScope handles(soa.Self());
1154 ScopedNullHandle<mirror::ClassLoader> class_loader; // null means boot class path loader.
1155 Handle<mirror::Class> compiling_class = handles.NewHandle(method->GetDeclaringClass());
1156 DexCompilationUnit dex_compilation_unit(
1157 class_loader,
1158 runtime->GetClassLinker(),
1159 dex_file,
1160 /*code_item=*/ nullptr,
1161 /*class_def_idx=*/ DexFile::kDexNoIndex16,
1162 method_idx,
1163 access_flags,
1164 /*verified_method=*/ nullptr,
1165 dex_cache,
1166 compiling_class);
1167 CodeVectorAllocator code_allocator(&allocator);
1168 // Go to native so that we don't block GC during compilation.
1169 ScopedThreadSuspension sts(soa.Self(), kNative);
1170 std::unique_ptr<CodeGenerator> codegen(
1171 TryCompileIntrinsic(&allocator,
1172 &arena_stack,
1173 &code_allocator,
1174 dex_compilation_unit,
1175 method,
1176 &handles));
1177 if (codegen != nullptr) {
1178 CompiledMethod* compiled_method = Emit(&allocator,
1179 &code_allocator,
1180 codegen.get(),
1181 /* item= */ nullptr);
1182 compiled_method->MarkAsIntrinsic();
1183 return compiled_method;
1184 }
1185 }
1186 }
1187
1188 JniCompiledMethod jni_compiled_method = ArtQuickJniCompileMethod(
1189 compiler_options, access_flags, method_idx, dex_file);
1190 MaybeRecordStat(compilation_stats_.get(), MethodCompilationStat::kCompiledNativeStub);
1191
1192 ScopedArenaAllocator stack_map_allocator(&arena_stack); // Will hold the stack map.
1193 ScopedArenaVector<uint8_t> stack_map = CreateJniStackMap(&stack_map_allocator,
1194 jni_compiled_method);
1195 return CompiledMethod::SwapAllocCompiledMethod(
1196 GetCompiledMethodStorage(),
1197 jni_compiled_method.GetInstructionSet(),
1198 jni_compiled_method.GetCode(),
1199 ArrayRef<const uint8_t>(stack_map),
1200 jni_compiled_method.GetCfi(),
1201 /* patches= */ ArrayRef<const linker::LinkerPatch>());
1202 }
1203
CreateOptimizingCompiler(const CompilerOptions & compiler_options,CompiledMethodStorage * storage)1204 Compiler* CreateOptimizingCompiler(const CompilerOptions& compiler_options,
1205 CompiledMethodStorage* storage) {
1206 return new OptimizingCompiler(compiler_options, storage);
1207 }
1208
EncodeArtMethodInInlineInfo(ArtMethod * method ATTRIBUTE_UNUSED)1209 bool EncodeArtMethodInInlineInfo(ArtMethod* method ATTRIBUTE_UNUSED) {
1210 // Note: the runtime is null only for unit testing.
1211 return Runtime::Current() == nullptr || !Runtime::Current()->IsAotCompiler();
1212 }
1213
JitCompile(Thread * self,jit::JitCodeCache * code_cache,jit::JitMemoryRegion * region,ArtMethod * method,bool baseline,bool osr,jit::JitLogger * jit_logger)1214 bool OptimizingCompiler::JitCompile(Thread* self,
1215 jit::JitCodeCache* code_cache,
1216 jit::JitMemoryRegion* region,
1217 ArtMethod* method,
1218 bool baseline,
1219 bool osr,
1220 jit::JitLogger* jit_logger) {
1221 StackHandleScope<3> hs(self);
1222 Handle<mirror::ClassLoader> class_loader(hs.NewHandle(
1223 method->GetDeclaringClass()->GetClassLoader()));
1224 Handle<mirror::DexCache> dex_cache(hs.NewHandle(method->GetDexCache()));
1225 DCHECK(method->IsCompilable());
1226
1227 const DexFile* dex_file = method->GetDexFile();
1228 const uint16_t class_def_idx = method->GetClassDefIndex();
1229 const dex::CodeItem* code_item = dex_file->GetCodeItem(method->GetCodeItemOffset());
1230 const uint32_t method_idx = method->GetDexMethodIndex();
1231 const uint32_t access_flags = method->GetAccessFlags();
1232
1233 Runtime* runtime = Runtime::Current();
1234 ArenaAllocator allocator(runtime->GetJitArenaPool());
1235
1236 if (UNLIKELY(method->IsNative())) {
1237 const CompilerOptions& compiler_options = GetCompilerOptions();
1238 JniCompiledMethod jni_compiled_method = ArtQuickJniCompileMethod(
1239 compiler_options, access_flags, method_idx, *dex_file);
1240 std::vector<Handle<mirror::Object>> roots;
1241 ArenaSet<ArtMethod*, std::less<ArtMethod*>> cha_single_implementation_list(
1242 allocator.Adapter(kArenaAllocCHA));
1243 ArenaStack arena_stack(runtime->GetJitArenaPool());
1244 // StackMapStream is large and it does not fit into this frame, so we need helper method.
1245 ScopedArenaAllocator stack_map_allocator(&arena_stack); // Will hold the stack map.
1246 ScopedArenaVector<uint8_t> stack_map = CreateJniStackMap(&stack_map_allocator,
1247 jni_compiled_method);
1248
1249 ArrayRef<const uint8_t> reserved_code;
1250 ArrayRef<const uint8_t> reserved_data;
1251 if (!code_cache->Reserve(self,
1252 region,
1253 jni_compiled_method.GetCode().size(),
1254 stack_map.size(),
1255 /* number_of_roots= */ 0,
1256 method,
1257 /*out*/ &reserved_code,
1258 /*out*/ &reserved_data)) {
1259 MaybeRecordStat(compilation_stats_.get(), MethodCompilationStat::kJitOutOfMemoryForCommit);
1260 return false;
1261 }
1262 const uint8_t* code = reserved_code.data() + OatQuickMethodHeader::InstructionAlignedSize();
1263
1264 // Add debug info after we know the code location but before we update entry-point.
1265 if (compiler_options.GenerateAnyDebugInfo()) {
1266 debug::MethodDebugInfo info = {};
1267 info.custom_name = "art_jni_trampoline";
1268 info.dex_file = dex_file;
1269 info.class_def_index = class_def_idx;
1270 info.dex_method_index = method_idx;
1271 info.access_flags = access_flags;
1272 info.code_item = code_item;
1273 info.isa = jni_compiled_method.GetInstructionSet();
1274 info.deduped = false;
1275 info.is_native_debuggable = compiler_options.GetNativeDebuggable();
1276 info.is_optimized = true;
1277 info.is_code_address_text_relative = false;
1278 info.code_address = reinterpret_cast<uintptr_t>(code);
1279 info.code_size = jni_compiled_method.GetCode().size();
1280 info.frame_size_in_bytes = jni_compiled_method.GetFrameSize();
1281 info.code_info = nullptr;
1282 info.cfi = jni_compiled_method.GetCfi();
1283 GenerateJitDebugInfo(info);
1284 }
1285
1286 if (!code_cache->Commit(self,
1287 region,
1288 method,
1289 reserved_code,
1290 jni_compiled_method.GetCode(),
1291 reserved_data,
1292 roots,
1293 ArrayRef<const uint8_t>(stack_map),
1294 osr,
1295 /* has_should_deoptimize_flag= */ false,
1296 cha_single_implementation_list)) {
1297 code_cache->Free(self, region, reserved_code.data(), reserved_data.data());
1298 return false;
1299 }
1300
1301 Runtime::Current()->GetJit()->AddMemoryUsage(method, allocator.BytesUsed());
1302 if (jit_logger != nullptr) {
1303 jit_logger->WriteLog(code, jni_compiled_method.GetCode().size(), method);
1304 }
1305 return true;
1306 }
1307
1308 ArenaStack arena_stack(runtime->GetJitArenaPool());
1309 CodeVectorAllocator code_allocator(&allocator);
1310 VariableSizedHandleScope handles(self);
1311
1312 std::unique_ptr<CodeGenerator> codegen;
1313 {
1314 Handle<mirror::Class> compiling_class = handles.NewHandle(method->GetDeclaringClass());
1315 DexCompilationUnit dex_compilation_unit(
1316 class_loader,
1317 runtime->GetClassLinker(),
1318 *dex_file,
1319 code_item,
1320 class_def_idx,
1321 method_idx,
1322 access_flags,
1323 /*verified_method=*/ nullptr,
1324 dex_cache,
1325 compiling_class);
1326
1327 // Go to native so that we don't block GC during compilation.
1328 ScopedThreadSuspension sts(self, kNative);
1329 codegen.reset(
1330 TryCompile(&allocator,
1331 &arena_stack,
1332 &code_allocator,
1333 dex_compilation_unit,
1334 method,
1335 baseline || GetCompilerOptions().IsBaseline(),
1336 osr,
1337 /* is_shared_jit_code= */ code_cache->IsSharedRegion(*region),
1338 &handles));
1339 if (codegen.get() == nullptr) {
1340 return false;
1341 }
1342 }
1343
1344 ScopedArenaVector<uint8_t> stack_map = codegen->BuildStackMaps(code_item);
1345
1346 ArrayRef<const uint8_t> reserved_code;
1347 ArrayRef<const uint8_t> reserved_data;
1348 if (!code_cache->Reserve(self,
1349 region,
1350 code_allocator.GetMemory().size(),
1351 stack_map.size(),
1352 /*number_of_roots=*/codegen->GetNumberOfJitRoots(),
1353 method,
1354 /*out*/ &reserved_code,
1355 /*out*/ &reserved_data)) {
1356 MaybeRecordStat(compilation_stats_.get(), MethodCompilationStat::kJitOutOfMemoryForCommit);
1357 return false;
1358 }
1359 const uint8_t* code = reserved_code.data() + OatQuickMethodHeader::InstructionAlignedSize();
1360 const uint8_t* roots_data = reserved_data.data();
1361
1362 std::vector<Handle<mirror::Object>> roots;
1363 codegen->EmitJitRoots(code_allocator.GetData(), roots_data, &roots);
1364 // The root Handle<>s filled by the codegen reference entries in the VariableSizedHandleScope.
1365 DCHECK(std::all_of(roots.begin(),
1366 roots.end(),
1367 [&handles](Handle<mirror::Object> root){
1368 return handles.Contains(root.GetReference());
1369 }));
1370
1371 // Add debug info after we know the code location but before we update entry-point.
1372 const CompilerOptions& compiler_options = GetCompilerOptions();
1373 if (compiler_options.GenerateAnyDebugInfo()) {
1374 debug::MethodDebugInfo info = {};
1375 DCHECK(info.custom_name.empty());
1376 info.dex_file = dex_file;
1377 info.class_def_index = class_def_idx;
1378 info.dex_method_index = method_idx;
1379 info.access_flags = access_flags;
1380 info.code_item = code_item;
1381 info.isa = codegen->GetInstructionSet();
1382 info.deduped = false;
1383 info.is_native_debuggable = compiler_options.GetNativeDebuggable();
1384 info.is_optimized = true;
1385 info.is_code_address_text_relative = false;
1386 info.code_address = reinterpret_cast<uintptr_t>(code);
1387 info.code_size = code_allocator.GetMemory().size();
1388 info.frame_size_in_bytes = codegen->GetFrameSize();
1389 info.code_info = stack_map.size() == 0 ? nullptr : stack_map.data();
1390 info.cfi = ArrayRef<const uint8_t>(*codegen->GetAssembler()->cfi().data());
1391 GenerateJitDebugInfo(info);
1392 }
1393
1394 if (!code_cache->Commit(self,
1395 region,
1396 method,
1397 reserved_code,
1398 code_allocator.GetMemory(),
1399 reserved_data,
1400 roots,
1401 ArrayRef<const uint8_t>(stack_map),
1402 osr,
1403 codegen->GetGraph()->HasShouldDeoptimizeFlag(),
1404 codegen->GetGraph()->GetCHASingleImplementationList())) {
1405 code_cache->Free(self, region, reserved_code.data(), reserved_data.data());
1406 return false;
1407 }
1408
1409 Runtime::Current()->GetJit()->AddMemoryUsage(method, allocator.BytesUsed());
1410 if (jit_logger != nullptr) {
1411 jit_logger->WriteLog(code, code_allocator.GetMemory().size(), method);
1412 }
1413
1414 if (kArenaAllocatorCountAllocations) {
1415 codegen.reset(); // Release codegen's ScopedArenaAllocator for memory accounting.
1416 size_t total_allocated = allocator.BytesAllocated() + arena_stack.PeakBytesAllocated();
1417 if (total_allocated > kArenaAllocatorMemoryReportThreshold) {
1418 MemStats mem_stats(allocator.GetMemStats());
1419 MemStats peak_stats(arena_stack.GetPeakStats());
1420 LOG(INFO) << "Used " << total_allocated << " bytes of arena memory for compiling "
1421 << dex_file->PrettyMethod(method_idx)
1422 << "\n" << Dumpable<MemStats>(mem_stats)
1423 << "\n" << Dumpable<MemStats>(peak_stats);
1424 }
1425 }
1426
1427 return true;
1428 }
1429
GenerateJitDebugInfo(const debug::MethodDebugInfo & info)1430 void OptimizingCompiler::GenerateJitDebugInfo(const debug::MethodDebugInfo& info) {
1431 const CompilerOptions& compiler_options = GetCompilerOptions();
1432 if (compiler_options.GenerateAnyDebugInfo()) {
1433 // If both flags are passed, generate full debug info.
1434 const bool mini_debug_info = !compiler_options.GetGenerateDebugInfo();
1435
1436 // Create entry for the single method that we just compiled.
1437 InstructionSet isa = compiler_options.GetInstructionSet();
1438 const InstructionSetFeatures* features = compiler_options.GetInstructionSetFeatures();
1439 std::vector<uint8_t> elf = debug::MakeElfFileForJIT(isa, features, mini_debug_info, info);
1440
1441 // NB: Don't allow packing of full info since it would remove non-backtrace data.
1442 MutexLock mu(Thread::Current(), *Locks::jit_lock_);
1443 const void* code_ptr = reinterpret_cast<const void*>(info.code_address);
1444 AddNativeDebugInfoForJit(code_ptr, elf, /*allow_packing=*/ mini_debug_info);
1445 }
1446 }
1447
1448 } // namespace art
1449