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 "android-base/strings.h"
26
27 #ifdef ART_ENABLE_CODEGEN_arm64
28 #include "instruction_simplifier_arm64.h"
29 #endif
30
31 #ifdef ART_ENABLE_CODEGEN_mips
32 #include "pc_relative_fixups_mips.h"
33 #endif
34
35 #ifdef ART_ENABLE_CODEGEN_x86
36 #include "pc_relative_fixups_x86.h"
37 #endif
38
39 #if defined(ART_ENABLE_CODEGEN_x86) || defined(ART_ENABLE_CODEGEN_x86_64)
40 #include "x86_memory_gen.h"
41 #endif
42
43 #include "art_method-inl.h"
44 #include "base/arena_allocator.h"
45 #include "base/arena_containers.h"
46 #include "base/dumpable.h"
47 #include "base/macros.h"
48 #include "base/mutex.h"
49 #include "base/timing_logger.h"
50 #include "bounds_check_elimination.h"
51 #include "builder.h"
52 #include "cha_guard_optimization.h"
53 #include "code_generator.h"
54 #include "code_sinking.h"
55 #include "compiled_method.h"
56 #include "compiler.h"
57 #include "constant_folding.h"
58 #include "dead_code_elimination.h"
59 #include "debug/elf_debug_writer.h"
60 #include "debug/method_debug_info.h"
61 #include "dex/verification_results.h"
62 #include "dex/verified_method.h"
63 #include "dex_file_types.h"
64 #include "driver/compiler_driver-inl.h"
65 #include "driver/compiler_options.h"
66 #include "driver/dex_compilation_unit.h"
67 #include "elf_writer_quick.h"
68 #include "graph_checker.h"
69 #include "graph_visualizer.h"
70 #include "gvn.h"
71 #include "induction_var_analysis.h"
72 #include "inliner.h"
73 #include "instruction_simplifier.h"
74 #include "instruction_simplifier_arm.h"
75 #include "intrinsics.h"
76 #include "jit/debugger_interface.h"
77 #include "jit/jit.h"
78 #include "jit/jit_code_cache.h"
79 #include "jit/jit_logger.h"
80 #include "jni/quick/jni_compiler.h"
81 #include "licm.h"
82 #include "load_store_analysis.h"
83 #include "load_store_elimination.h"
84 #include "loop_optimization.h"
85 #include "nodes.h"
86 #include "oat_quick_method_header.h"
87 #include "prepare_for_register_allocation.h"
88 #include "reference_type_propagation.h"
89 #include "register_allocator_linear_scan.h"
90 #include "select_generator.h"
91 #include "scheduler.h"
92 #include "sharpening.h"
93 #include "side_effects_analysis.h"
94 #include "ssa_builder.h"
95 #include "ssa_liveness_analysis.h"
96 #include "ssa_phi_elimination.h"
97 #include "utils/assembler.h"
98 #include "verifier/verifier_compiler_binding.h"
99
100 namespace art {
101
102 static constexpr size_t kArenaAllocatorMemoryReportThreshold = 8 * MB;
103
104 static constexpr const char* kPassNameSeparator = "$";
105
106 /**
107 * Used by the code generator, to allocate the code in a vector.
108 */
109 class CodeVectorAllocator FINAL : public CodeAllocator {
110 public:
CodeVectorAllocator(ArenaAllocator * arena)111 explicit CodeVectorAllocator(ArenaAllocator* arena)
112 : memory_(arena->Adapter(kArenaAllocCodeBuffer)),
113 size_(0) {}
114
Allocate(size_t size)115 virtual uint8_t* Allocate(size_t size) {
116 size_ = size;
117 memory_.resize(size);
118 return &memory_[0];
119 }
120
GetSize() const121 size_t GetSize() const { return size_; }
GetMemory() const122 const ArenaVector<uint8_t>& GetMemory() const { return memory_; }
GetData()123 uint8_t* GetData() { return memory_.data(); }
124
125 private:
126 ArenaVector<uint8_t> memory_;
127 size_t size_;
128
129 DISALLOW_COPY_AND_ASSIGN(CodeVectorAllocator);
130 };
131
132 /**
133 * Filter to apply to the visualizer. Methods whose name contain that filter will
134 * be dumped.
135 */
136 static constexpr const char kStringFilter[] = "";
137
138 class PassScope;
139
140 class PassObserver : public ValueObject {
141 public:
PassObserver(HGraph * graph,CodeGenerator * codegen,std::ostream * visualizer_output,CompilerDriver * compiler_driver,Mutex & dump_mutex)142 PassObserver(HGraph* graph,
143 CodeGenerator* codegen,
144 std::ostream* visualizer_output,
145 CompilerDriver* compiler_driver,
146 Mutex& dump_mutex)
147 : graph_(graph),
148 cached_method_name_(),
149 timing_logger_enabled_(compiler_driver->GetDumpPasses()),
150 timing_logger_(timing_logger_enabled_ ? GetMethodName() : "", true, true),
151 disasm_info_(graph->GetArena()),
152 visualizer_oss_(),
153 visualizer_output_(visualizer_output),
154 visualizer_enabled_(!compiler_driver->GetCompilerOptions().GetDumpCfgFileName().empty()),
155 visualizer_(&visualizer_oss_, graph, *codegen),
156 visualizer_dump_mutex_(dump_mutex),
157 graph_in_bad_state_(false) {
158 if (timing_logger_enabled_ || visualizer_enabled_) {
159 if (!IsVerboseMethod(compiler_driver, GetMethodName())) {
160 timing_logger_enabled_ = visualizer_enabled_ = false;
161 }
162 if (visualizer_enabled_) {
163 visualizer_.PrintHeader(GetMethodName());
164 codegen->SetDisassemblyInformation(&disasm_info_);
165 }
166 }
167 }
168
~PassObserver()169 ~PassObserver() {
170 if (timing_logger_enabled_) {
171 LOG(INFO) << "TIMINGS " << GetMethodName();
172 LOG(INFO) << Dumpable<TimingLogger>(timing_logger_);
173 }
174 DCHECK(visualizer_oss_.str().empty());
175 }
176
DumpDisassembly()177 void DumpDisassembly() REQUIRES(!visualizer_dump_mutex_) {
178 if (visualizer_enabled_) {
179 visualizer_.DumpGraphWithDisassembly();
180 FlushVisualizer();
181 }
182 }
183
SetGraphInBadState()184 void SetGraphInBadState() { graph_in_bad_state_ = true; }
185
GetMethodName()186 const char* GetMethodName() {
187 // PrettyMethod() is expensive, so we delay calling it until we actually have to.
188 if (cached_method_name_.empty()) {
189 cached_method_name_ = graph_->GetDexFile().PrettyMethod(graph_->GetMethodIdx());
190 }
191 return cached_method_name_.c_str();
192 }
193
194 private:
StartPass(const char * pass_name)195 void StartPass(const char* pass_name) REQUIRES(!visualizer_dump_mutex_) {
196 VLOG(compiler) << "Starting pass: " << pass_name;
197 // Dump graph first, then start timer.
198 if (visualizer_enabled_) {
199 visualizer_.DumpGraph(pass_name, /* is_after_pass */ false, graph_in_bad_state_);
200 FlushVisualizer();
201 }
202 if (timing_logger_enabled_) {
203 timing_logger_.StartTiming(pass_name);
204 }
205 }
206
FlushVisualizer()207 void FlushVisualizer() REQUIRES(!visualizer_dump_mutex_) {
208 MutexLock mu(Thread::Current(), visualizer_dump_mutex_);
209 *visualizer_output_ << visualizer_oss_.str();
210 visualizer_output_->flush();
211 visualizer_oss_.str("");
212 visualizer_oss_.clear();
213 }
214
EndPass(const char * pass_name)215 void EndPass(const char* pass_name) REQUIRES(!visualizer_dump_mutex_) {
216 // Pause timer first, then dump graph.
217 if (timing_logger_enabled_) {
218 timing_logger_.EndTiming();
219 }
220 if (visualizer_enabled_) {
221 visualizer_.DumpGraph(pass_name, /* is_after_pass */ true, graph_in_bad_state_);
222 FlushVisualizer();
223 }
224
225 // Validate the HGraph if running in debug mode.
226 if (kIsDebugBuild) {
227 if (!graph_in_bad_state_) {
228 GraphChecker checker(graph_);
229 checker.Run();
230 if (!checker.IsValid()) {
231 LOG(FATAL) << "Error after " << pass_name << ": " << Dumpable<GraphChecker>(checker);
232 }
233 }
234 }
235 }
236
IsVerboseMethod(CompilerDriver * compiler_driver,const char * method_name)237 static bool IsVerboseMethod(CompilerDriver* compiler_driver, const char* method_name) {
238 // Test an exact match to --verbose-methods. If verbose-methods is set, this overrides an
239 // empty kStringFilter matching all methods.
240 if (compiler_driver->GetCompilerOptions().HasVerboseMethods()) {
241 return compiler_driver->GetCompilerOptions().IsVerboseMethod(method_name);
242 }
243
244 // Test the kStringFilter sub-string. constexpr helper variable to silence unreachable-code
245 // warning when the string is empty.
246 constexpr bool kStringFilterEmpty = arraysize(kStringFilter) <= 1;
247 if (kStringFilterEmpty || strstr(method_name, kStringFilter) != nullptr) {
248 return true;
249 }
250
251 return false;
252 }
253
254 HGraph* const graph_;
255
256 std::string cached_method_name_;
257
258 bool timing_logger_enabled_;
259 TimingLogger timing_logger_;
260
261 DisassemblyInformation disasm_info_;
262
263 std::ostringstream visualizer_oss_;
264 std::ostream* visualizer_output_;
265 bool visualizer_enabled_;
266 HGraphVisualizer visualizer_;
267 Mutex& visualizer_dump_mutex_;
268
269 // Flag to be set by the compiler if the pass failed and the graph is not
270 // expected to validate.
271 bool graph_in_bad_state_;
272
273 friend PassScope;
274
275 DISALLOW_COPY_AND_ASSIGN(PassObserver);
276 };
277
278 class PassScope : public ValueObject {
279 public:
PassScope(const char * pass_name,PassObserver * pass_observer)280 PassScope(const char *pass_name, PassObserver* pass_observer)
281 : pass_name_(pass_name),
282 pass_observer_(pass_observer) {
283 pass_observer_->StartPass(pass_name_);
284 }
285
~PassScope()286 ~PassScope() {
287 pass_observer_->EndPass(pass_name_);
288 }
289
290 private:
291 const char* const pass_name_;
292 PassObserver* const pass_observer_;
293 };
294
295 class OptimizingCompiler FINAL : public Compiler {
296 public:
297 explicit OptimizingCompiler(CompilerDriver* driver);
298 ~OptimizingCompiler() OVERRIDE;
299
300 bool CanCompileMethod(uint32_t method_idx, const DexFile& dex_file) const OVERRIDE;
301
302 CompiledMethod* Compile(const DexFile::CodeItem* code_item,
303 uint32_t access_flags,
304 InvokeType invoke_type,
305 uint16_t class_def_idx,
306 uint32_t method_idx,
307 Handle<mirror::ClassLoader> class_loader,
308 const DexFile& dex_file,
309 Handle<mirror::DexCache> dex_cache) const OVERRIDE;
310
JniCompile(uint32_t access_flags,uint32_t method_idx,const DexFile & dex_file,JniOptimizationFlags optimization_flags) const311 CompiledMethod* JniCompile(uint32_t access_flags,
312 uint32_t method_idx,
313 const DexFile& dex_file,
314 JniOptimizationFlags optimization_flags) const OVERRIDE {
315 return ArtQuickJniCompileMethod(GetCompilerDriver(),
316 access_flags,
317 method_idx,
318 dex_file,
319 optimization_flags);
320 }
321
GetEntryPointOf(ArtMethod * method) const322 uintptr_t GetEntryPointOf(ArtMethod* method) const OVERRIDE
323 REQUIRES_SHARED(Locks::mutator_lock_) {
324 return reinterpret_cast<uintptr_t>(method->GetEntryPointFromQuickCompiledCodePtrSize(
325 InstructionSetPointerSize(GetCompilerDriver()->GetInstructionSet())));
326 }
327
328 void Init() OVERRIDE;
329
330 void UnInit() const OVERRIDE;
331
MaybeRecordStat(MethodCompilationStat compilation_stat) const332 void MaybeRecordStat(MethodCompilationStat compilation_stat) const {
333 if (compilation_stats_.get() != nullptr) {
334 compilation_stats_->RecordStat(compilation_stat);
335 }
336 }
337
338 bool JitCompile(Thread* self,
339 jit::JitCodeCache* code_cache,
340 ArtMethod* method,
341 bool osr,
342 jit::JitLogger* jit_logger)
343 OVERRIDE
344 REQUIRES_SHARED(Locks::mutator_lock_);
345
346 private:
347 void RunOptimizations(HGraph* graph,
348 CodeGenerator* codegen,
349 CompilerDriver* driver,
350 const DexCompilationUnit& dex_compilation_unit,
351 PassObserver* pass_observer,
352 VariableSizedHandleScope* handles) const;
353
354 void RunOptimizations(HOptimization* optimizations[],
355 size_t length,
356 PassObserver* pass_observer) const;
357
358 private:
359 // Create a 'CompiledMethod' for an optimized graph.
360 CompiledMethod* Emit(ArenaAllocator* arena,
361 CodeVectorAllocator* code_allocator,
362 CodeGenerator* codegen,
363 CompilerDriver* driver,
364 const DexFile::CodeItem* item) const;
365
366 // Try compiling a method and return the code generator used for
367 // compiling it.
368 // This method:
369 // 1) Builds the graph. Returns null if it failed to build it.
370 // 2) Transforms the graph to SSA. Returns null if it failed.
371 // 3) Runs optimizations on the graph, including register allocator.
372 // 4) Generates code with the `code_allocator` provided.
373 CodeGenerator* TryCompile(ArenaAllocator* arena,
374 CodeVectorAllocator* code_allocator,
375 const DexFile::CodeItem* code_item,
376 uint32_t access_flags,
377 InvokeType invoke_type,
378 uint16_t class_def_idx,
379 uint32_t method_idx,
380 Handle<mirror::ClassLoader> class_loader,
381 const DexFile& dex_file,
382 Handle<mirror::DexCache> dex_cache,
383 ArtMethod* method,
384 bool osr,
385 VariableSizedHandleScope* handles) const;
386
387 void MaybeRunInliner(HGraph* graph,
388 CodeGenerator* codegen,
389 CompilerDriver* driver,
390 const DexCompilationUnit& dex_compilation_unit,
391 PassObserver* pass_observer,
392 VariableSizedHandleScope* handles) const;
393
394 void RunArchOptimizations(InstructionSet instruction_set,
395 HGraph* graph,
396 CodeGenerator* codegen,
397 PassObserver* pass_observer) const;
398
399 std::unique_ptr<OptimizingCompilerStats> compilation_stats_;
400
401 std::unique_ptr<std::ostream> visualizer_output_;
402
403 mutable Mutex dump_mutex_; // To synchronize visualizer writing.
404
405 DISALLOW_COPY_AND_ASSIGN(OptimizingCompiler);
406 };
407
408 static const int kMaximumCompilationTimeBeforeWarning = 100; /* ms */
409
OptimizingCompiler(CompilerDriver * driver)410 OptimizingCompiler::OptimizingCompiler(CompilerDriver* driver)
411 : Compiler(driver, kMaximumCompilationTimeBeforeWarning),
412 dump_mutex_("Visualizer dump lock") {}
413
Init()414 void OptimizingCompiler::Init() {
415 // Enable C1visualizer output. Must be done in Init() because the compiler
416 // driver is not fully initialized when passed to the compiler's constructor.
417 CompilerDriver* driver = GetCompilerDriver();
418 const std::string cfg_file_name = driver->GetCompilerOptions().GetDumpCfgFileName();
419 if (!cfg_file_name.empty()) {
420 std::ios_base::openmode cfg_file_mode =
421 driver->GetCompilerOptions().GetDumpCfgAppend() ? std::ofstream::app : std::ofstream::out;
422 visualizer_output_.reset(new std::ofstream(cfg_file_name, cfg_file_mode));
423 }
424 if (driver->GetDumpStats()) {
425 compilation_stats_.reset(new OptimizingCompilerStats());
426 }
427 }
428
UnInit() const429 void OptimizingCompiler::UnInit() const {
430 }
431
~OptimizingCompiler()432 OptimizingCompiler::~OptimizingCompiler() {
433 if (compilation_stats_.get() != nullptr) {
434 compilation_stats_->Log();
435 }
436 }
437
CanCompileMethod(uint32_t method_idx ATTRIBUTE_UNUSED,const DexFile & dex_file ATTRIBUTE_UNUSED) const438 bool OptimizingCompiler::CanCompileMethod(uint32_t method_idx ATTRIBUTE_UNUSED,
439 const DexFile& dex_file ATTRIBUTE_UNUSED) const {
440 return true;
441 }
442
IsInstructionSetSupported(InstructionSet instruction_set)443 static bool IsInstructionSetSupported(InstructionSet instruction_set) {
444 return (instruction_set == kArm && !kArm32QuickCodeUseSoftFloat)
445 || instruction_set == kArm64
446 || (instruction_set == kThumb2 && !kArm32QuickCodeUseSoftFloat)
447 || instruction_set == kMips
448 || instruction_set == kMips64
449 || instruction_set == kX86
450 || instruction_set == kX86_64;
451 }
452
453 // Strip pass name suffix to get optimization name.
ConvertPassNameToOptimizationName(const std::string & pass_name)454 static std::string ConvertPassNameToOptimizationName(const std::string& pass_name) {
455 size_t pos = pass_name.find(kPassNameSeparator);
456 return pos == std::string::npos ? pass_name : pass_name.substr(0, pos);
457 }
458
BuildOptimization(const std::string & pass_name,ArenaAllocator * arena,HGraph * graph,OptimizingCompilerStats * stats,CodeGenerator * codegen,CompilerDriver * driver,const DexCompilationUnit & dex_compilation_unit,VariableSizedHandleScope * handles,SideEffectsAnalysis * most_recent_side_effects,HInductionVarAnalysis * most_recent_induction,LoadStoreAnalysis * most_recent_lsa)459 static HOptimization* BuildOptimization(
460 const std::string& pass_name,
461 ArenaAllocator* arena,
462 HGraph* graph,
463 OptimizingCompilerStats* stats,
464 CodeGenerator* codegen,
465 CompilerDriver* driver,
466 const DexCompilationUnit& dex_compilation_unit,
467 VariableSizedHandleScope* handles,
468 SideEffectsAnalysis* most_recent_side_effects,
469 HInductionVarAnalysis* most_recent_induction,
470 LoadStoreAnalysis* most_recent_lsa) {
471 std::string opt_name = ConvertPassNameToOptimizationName(pass_name);
472 if (opt_name == BoundsCheckElimination::kBoundsCheckEliminationPassName) {
473 CHECK(most_recent_side_effects != nullptr && most_recent_induction != nullptr);
474 return new (arena) BoundsCheckElimination(graph,
475 *most_recent_side_effects,
476 most_recent_induction);
477 } else if (opt_name == GVNOptimization::kGlobalValueNumberingPassName) {
478 CHECK(most_recent_side_effects != nullptr);
479 return new (arena) GVNOptimization(graph, *most_recent_side_effects, pass_name.c_str());
480 } else if (opt_name == HConstantFolding::kConstantFoldingPassName) {
481 return new (arena) HConstantFolding(graph, pass_name.c_str());
482 } else if (opt_name == HDeadCodeElimination::kDeadCodeEliminationPassName) {
483 return new (arena) HDeadCodeElimination(graph, stats, pass_name.c_str());
484 } else if (opt_name == HInliner::kInlinerPassName) {
485 size_t number_of_dex_registers = dex_compilation_unit.GetCodeItem()->registers_size_;
486 return new (arena) HInliner(graph, // outer_graph
487 graph, // outermost_graph
488 codegen,
489 dex_compilation_unit, // outer_compilation_unit
490 dex_compilation_unit, // outermost_compilation_unit
491 driver,
492 handles,
493 stats,
494 number_of_dex_registers,
495 /* total_number_of_instructions */ 0,
496 /* parent */ nullptr);
497 } else if (opt_name == HSharpening::kSharpeningPassName) {
498 return new (arena) HSharpening(graph, codegen, dex_compilation_unit, driver, handles);
499 } else if (opt_name == HSelectGenerator::kSelectGeneratorPassName) {
500 return new (arena) HSelectGenerator(graph, stats);
501 } else if (opt_name == HInductionVarAnalysis::kInductionPassName) {
502 return new (arena) HInductionVarAnalysis(graph);
503 } else if (opt_name == InstructionSimplifier::kInstructionSimplifierPassName) {
504 return new (arena) InstructionSimplifier(graph, codegen, driver, stats, pass_name.c_str());
505 } else if (opt_name == IntrinsicsRecognizer::kIntrinsicsRecognizerPassName) {
506 return new (arena) IntrinsicsRecognizer(graph, stats);
507 } else if (opt_name == LICM::kLoopInvariantCodeMotionPassName) {
508 CHECK(most_recent_side_effects != nullptr);
509 return new (arena) LICM(graph, *most_recent_side_effects, stats);
510 } else if (opt_name == LoadStoreAnalysis::kLoadStoreAnalysisPassName) {
511 return new (arena) LoadStoreAnalysis(graph);
512 } else if (opt_name == LoadStoreElimination::kLoadStoreEliminationPassName) {
513 CHECK(most_recent_side_effects != nullptr);
514 CHECK(most_recent_lsa != nullptr);
515 return new (arena) LoadStoreElimination(graph, *most_recent_side_effects, *most_recent_lsa);
516 } else if (opt_name == SideEffectsAnalysis::kSideEffectsAnalysisPassName) {
517 return new (arena) SideEffectsAnalysis(graph);
518 } else if (opt_name == HLoopOptimization::kLoopOptimizationPassName) {
519 return new (arena) HLoopOptimization(graph, driver, most_recent_induction);
520 } else if (opt_name == CHAGuardOptimization::kCHAGuardOptimizationPassName) {
521 return new (arena) CHAGuardOptimization(graph);
522 } else if (opt_name == CodeSinking::kCodeSinkingPassName) {
523 return new (arena) CodeSinking(graph, stats);
524 #ifdef ART_ENABLE_CODEGEN_arm
525 } else if (opt_name == arm::InstructionSimplifierArm::kInstructionSimplifierArmPassName) {
526 return new (arena) arm::InstructionSimplifierArm(graph, stats);
527 #endif
528 #ifdef ART_ENABLE_CODEGEN_arm64
529 } else if (opt_name == arm64::InstructionSimplifierArm64::kInstructionSimplifierArm64PassName) {
530 return new (arena) arm64::InstructionSimplifierArm64(graph, stats);
531 #endif
532 #ifdef ART_ENABLE_CODEGEN_mips
533 } else if (opt_name == mips::PcRelativeFixups::kPcRelativeFixupsMipsPassName) {
534 return new (arena) mips::PcRelativeFixups(graph, codegen, stats);
535 #endif
536 #ifdef ART_ENABLE_CODEGEN_x86
537 } else if (opt_name == x86::PcRelativeFixups::kPcRelativeFixupsX86PassName) {
538 return new (arena) x86::PcRelativeFixups(graph, codegen, stats);
539 } else if (opt_name == x86::X86MemoryOperandGeneration::kX86MemoryOperandGenerationPassName) {
540 return new (arena) x86::X86MemoryOperandGeneration(graph, codegen, stats);
541 #endif
542 }
543 return nullptr;
544 }
545
BuildOptimizations(const std::vector<std::string> & pass_names,ArenaAllocator * arena,HGraph * graph,OptimizingCompilerStats * stats,CodeGenerator * codegen,CompilerDriver * driver,const DexCompilationUnit & dex_compilation_unit,VariableSizedHandleScope * handles)546 static ArenaVector<HOptimization*> BuildOptimizations(
547 const std::vector<std::string>& pass_names,
548 ArenaAllocator* arena,
549 HGraph* graph,
550 OptimizingCompilerStats* stats,
551 CodeGenerator* codegen,
552 CompilerDriver* driver,
553 const DexCompilationUnit& dex_compilation_unit,
554 VariableSizedHandleScope* handles) {
555 // Few HOptimizations constructors require SideEffectsAnalysis or HInductionVarAnalysis
556 // instances. This method assumes that each of them expects the nearest instance preceeding it
557 // in the pass name list.
558 SideEffectsAnalysis* most_recent_side_effects = nullptr;
559 HInductionVarAnalysis* most_recent_induction = nullptr;
560 LoadStoreAnalysis* most_recent_lsa = nullptr;
561 ArenaVector<HOptimization*> ret(arena->Adapter());
562 for (const std::string& pass_name : pass_names) {
563 HOptimization* opt = BuildOptimization(
564 pass_name,
565 arena,
566 graph,
567 stats,
568 codegen,
569 driver,
570 dex_compilation_unit,
571 handles,
572 most_recent_side_effects,
573 most_recent_induction,
574 most_recent_lsa);
575 CHECK(opt != nullptr) << "Couldn't build optimization: \"" << pass_name << "\"";
576 ret.push_back(opt);
577
578 std::string opt_name = ConvertPassNameToOptimizationName(pass_name);
579 if (opt_name == SideEffectsAnalysis::kSideEffectsAnalysisPassName) {
580 most_recent_side_effects = down_cast<SideEffectsAnalysis*>(opt);
581 } else if (opt_name == HInductionVarAnalysis::kInductionPassName) {
582 most_recent_induction = down_cast<HInductionVarAnalysis*>(opt);
583 } else if (opt_name == LoadStoreAnalysis::kLoadStoreAnalysisPassName) {
584 most_recent_lsa = down_cast<LoadStoreAnalysis*>(opt);
585 }
586 }
587 return ret;
588 }
589
RunOptimizations(HOptimization * optimizations[],size_t length,PassObserver * pass_observer) const590 void OptimizingCompiler::RunOptimizations(HOptimization* optimizations[],
591 size_t length,
592 PassObserver* pass_observer) const {
593 for (size_t i = 0; i < length; ++i) {
594 PassScope scope(optimizations[i]->GetPassName(), pass_observer);
595 optimizations[i]->Run();
596 }
597 }
598
MaybeRunInliner(HGraph * graph,CodeGenerator * codegen,CompilerDriver * driver,const DexCompilationUnit & dex_compilation_unit,PassObserver * pass_observer,VariableSizedHandleScope * handles) const599 void OptimizingCompiler::MaybeRunInliner(HGraph* graph,
600 CodeGenerator* codegen,
601 CompilerDriver* driver,
602 const DexCompilationUnit& dex_compilation_unit,
603 PassObserver* pass_observer,
604 VariableSizedHandleScope* handles) const {
605 OptimizingCompilerStats* stats = compilation_stats_.get();
606 const CompilerOptions& compiler_options = driver->GetCompilerOptions();
607 bool should_inline = (compiler_options.GetInlineMaxCodeUnits() > 0);
608 if (!should_inline) {
609 return;
610 }
611 size_t number_of_dex_registers = dex_compilation_unit.GetCodeItem()->registers_size_;
612 HInliner* inliner = new (graph->GetArena()) HInliner(
613 graph, // outer_graph
614 graph, // outermost_graph
615 codegen,
616 dex_compilation_unit, // outer_compilation_unit
617 dex_compilation_unit, // outermost_compilation_unit
618 driver,
619 handles,
620 stats,
621 number_of_dex_registers,
622 /* total_number_of_instructions */ 0,
623 /* parent */ nullptr);
624 HOptimization* optimizations[] = { inliner };
625
626 RunOptimizations(optimizations, arraysize(optimizations), pass_observer);
627 }
628
RunArchOptimizations(InstructionSet instruction_set,HGraph * graph,CodeGenerator * codegen,PassObserver * pass_observer) const629 void OptimizingCompiler::RunArchOptimizations(InstructionSet instruction_set,
630 HGraph* graph,
631 CodeGenerator* codegen,
632 PassObserver* pass_observer) const {
633 UNUSED(codegen); // To avoid compilation error when compiling for svelte
634 OptimizingCompilerStats* stats = compilation_stats_.get();
635 ArenaAllocator* arena = graph->GetArena();
636 switch (instruction_set) {
637 #if defined(ART_ENABLE_CODEGEN_arm)
638 case kThumb2:
639 case kArm: {
640 arm::InstructionSimplifierArm* simplifier =
641 new (arena) arm::InstructionSimplifierArm(graph, stats);
642 SideEffectsAnalysis* side_effects = new (arena) SideEffectsAnalysis(graph);
643 GVNOptimization* gvn = new (arena) GVNOptimization(graph, *side_effects, "GVN$after_arch");
644 HInstructionScheduling* scheduling =
645 new (arena) HInstructionScheduling(graph, instruction_set, codegen);
646 HOptimization* arm_optimizations[] = {
647 simplifier,
648 side_effects,
649 gvn,
650 scheduling,
651 };
652 RunOptimizations(arm_optimizations, arraysize(arm_optimizations), pass_observer);
653 break;
654 }
655 #endif
656 #ifdef ART_ENABLE_CODEGEN_arm64
657 case kArm64: {
658 arm64::InstructionSimplifierArm64* simplifier =
659 new (arena) arm64::InstructionSimplifierArm64(graph, stats);
660 SideEffectsAnalysis* side_effects = new (arena) SideEffectsAnalysis(graph);
661 GVNOptimization* gvn = new (arena) GVNOptimization(graph, *side_effects, "GVN$after_arch");
662 HInstructionScheduling* scheduling =
663 new (arena) HInstructionScheduling(graph, instruction_set);
664 HOptimization* arm64_optimizations[] = {
665 simplifier,
666 side_effects,
667 gvn,
668 scheduling,
669 };
670 RunOptimizations(arm64_optimizations, arraysize(arm64_optimizations), pass_observer);
671 break;
672 }
673 #endif
674 #ifdef ART_ENABLE_CODEGEN_mips
675 case kMips: {
676 mips::PcRelativeFixups* pc_relative_fixups =
677 new (arena) mips::PcRelativeFixups(graph, codegen, stats);
678 HOptimization* mips_optimizations[] = {
679 pc_relative_fixups,
680 };
681 RunOptimizations(mips_optimizations, arraysize(mips_optimizations), pass_observer);
682 break;
683 }
684 #endif
685 #ifdef ART_ENABLE_CODEGEN_x86
686 case kX86: {
687 x86::PcRelativeFixups* pc_relative_fixups =
688 new (arena) x86::PcRelativeFixups(graph, codegen, stats);
689 x86::X86MemoryOperandGeneration* memory_gen =
690 new (arena) x86::X86MemoryOperandGeneration(graph, codegen, stats);
691 HOptimization* x86_optimizations[] = {
692 pc_relative_fixups,
693 memory_gen
694 };
695 RunOptimizations(x86_optimizations, arraysize(x86_optimizations), pass_observer);
696 break;
697 }
698 #endif
699 #ifdef ART_ENABLE_CODEGEN_x86_64
700 case kX86_64: {
701 x86::X86MemoryOperandGeneration* memory_gen =
702 new (arena) x86::X86MemoryOperandGeneration(graph, codegen, stats);
703 HOptimization* x86_64_optimizations[] = {
704 memory_gen
705 };
706 RunOptimizations(x86_64_optimizations, arraysize(x86_64_optimizations), pass_observer);
707 break;
708 }
709 #endif
710 default:
711 break;
712 }
713 }
714
715 NO_INLINE // Avoid increasing caller's frame size by large stack-allocated objects.
AllocateRegisters(HGraph * graph,CodeGenerator * codegen,PassObserver * pass_observer,RegisterAllocator::Strategy strategy)716 static void AllocateRegisters(HGraph* graph,
717 CodeGenerator* codegen,
718 PassObserver* pass_observer,
719 RegisterAllocator::Strategy strategy) {
720 {
721 PassScope scope(PrepareForRegisterAllocation::kPrepareForRegisterAllocationPassName,
722 pass_observer);
723 PrepareForRegisterAllocation(graph).Run();
724 }
725 SsaLivenessAnalysis liveness(graph, codegen);
726 {
727 PassScope scope(SsaLivenessAnalysis::kLivenessPassName, pass_observer);
728 liveness.Analyze();
729 }
730 {
731 PassScope scope(RegisterAllocator::kRegisterAllocatorPassName, pass_observer);
732 RegisterAllocator::Create(graph->GetArena(), codegen, liveness, strategy)->AllocateRegisters();
733 }
734 }
735
RunOptimizations(HGraph * graph,CodeGenerator * codegen,CompilerDriver * driver,const DexCompilationUnit & dex_compilation_unit,PassObserver * pass_observer,VariableSizedHandleScope * handles) const736 void OptimizingCompiler::RunOptimizations(HGraph* graph,
737 CodeGenerator* codegen,
738 CompilerDriver* driver,
739 const DexCompilationUnit& dex_compilation_unit,
740 PassObserver* pass_observer,
741 VariableSizedHandleScope* handles) const {
742 OptimizingCompilerStats* stats = compilation_stats_.get();
743 ArenaAllocator* arena = graph->GetArena();
744 if (driver->GetCompilerOptions().GetPassesToRun() != nullptr) {
745 ArenaVector<HOptimization*> optimizations = BuildOptimizations(
746 *driver->GetCompilerOptions().GetPassesToRun(),
747 arena,
748 graph,
749 stats,
750 codegen,
751 driver,
752 dex_compilation_unit,
753 handles);
754 RunOptimizations(&optimizations[0], optimizations.size(), pass_observer);
755 return;
756 }
757
758 HDeadCodeElimination* dce1 = new (arena) HDeadCodeElimination(
759 graph, stats, "dead_code_elimination$initial");
760 HDeadCodeElimination* dce2 = new (arena) HDeadCodeElimination(
761 graph, stats, "dead_code_elimination$after_inlining");
762 HDeadCodeElimination* dce3 = new (arena) HDeadCodeElimination(
763 graph, stats, "dead_code_elimination$final");
764 HConstantFolding* fold1 = new (arena) HConstantFolding(graph, "constant_folding");
765 InstructionSimplifier* simplify1 = new (arena) InstructionSimplifier(
766 graph, codegen, driver, stats);
767 HSelectGenerator* select_generator = new (arena) HSelectGenerator(graph, stats);
768 HConstantFolding* fold2 = new (arena) HConstantFolding(
769 graph, "constant_folding$after_inlining");
770 HConstantFolding* fold3 = new (arena) HConstantFolding(graph, "constant_folding$after_bce");
771 SideEffectsAnalysis* side_effects1 = new (arena) SideEffectsAnalysis(
772 graph, "side_effects$before_gvn");
773 SideEffectsAnalysis* side_effects2 = new (arena) SideEffectsAnalysis(
774 graph, "side_effects$before_lse");
775 GVNOptimization* gvn = new (arena) GVNOptimization(graph, *side_effects1);
776 LICM* licm = new (arena) LICM(graph, *side_effects1, stats);
777 HInductionVarAnalysis* induction = new (arena) HInductionVarAnalysis(graph);
778 BoundsCheckElimination* bce = new (arena) BoundsCheckElimination(graph, *side_effects1, induction);
779 HLoopOptimization* loop = new (arena) HLoopOptimization(graph, driver, induction);
780 LoadStoreAnalysis* lsa = new (arena) LoadStoreAnalysis(graph);
781 LoadStoreElimination* lse = new (arena) LoadStoreElimination(graph, *side_effects2, *lsa);
782 HSharpening* sharpening = new (arena) HSharpening(
783 graph, codegen, dex_compilation_unit, driver, handles);
784 InstructionSimplifier* simplify2 = new (arena) InstructionSimplifier(
785 graph, codegen, driver, stats, "instruction_simplifier$after_inlining");
786 InstructionSimplifier* simplify3 = new (arena) InstructionSimplifier(
787 graph, codegen, driver, stats, "instruction_simplifier$after_bce");
788 InstructionSimplifier* simplify4 = new (arena) InstructionSimplifier(
789 graph, codegen, driver, stats, "instruction_simplifier$before_codegen");
790 IntrinsicsRecognizer* intrinsics = new (arena) IntrinsicsRecognizer(graph, stats);
791 CHAGuardOptimization* cha_guard = new (arena) CHAGuardOptimization(graph);
792 CodeSinking* code_sinking = new (arena) CodeSinking(graph, stats);
793
794 HOptimization* optimizations1[] = {
795 intrinsics,
796 sharpening,
797 fold1,
798 simplify1,
799 dce1,
800 };
801 RunOptimizations(optimizations1, arraysize(optimizations1), pass_observer);
802
803 MaybeRunInliner(graph, codegen, driver, dex_compilation_unit, pass_observer, handles);
804
805 HOptimization* optimizations2[] = {
806 // SelectGenerator depends on the InstructionSimplifier removing
807 // redundant suspend checks to recognize empty blocks.
808 select_generator,
809 fold2, // TODO: if we don't inline we can also skip fold2.
810 simplify2,
811 dce2,
812 side_effects1,
813 gvn,
814 licm,
815 induction,
816 bce,
817 loop,
818 fold3, // evaluates code generated by dynamic bce
819 simplify3,
820 side_effects2,
821 lsa,
822 lse,
823 cha_guard,
824 dce3,
825 code_sinking,
826 // The codegen has a few assumptions that only the instruction simplifier
827 // can satisfy. For example, the code generator does not expect to see a
828 // HTypeConversion from a type to the same type.
829 simplify4,
830 };
831 RunOptimizations(optimizations2, arraysize(optimizations2), pass_observer);
832
833 RunArchOptimizations(driver->GetInstructionSet(), graph, codegen, pass_observer);
834 }
835
EmitAndSortLinkerPatches(CodeGenerator * codegen)836 static ArenaVector<LinkerPatch> EmitAndSortLinkerPatches(CodeGenerator* codegen) {
837 ArenaVector<LinkerPatch> linker_patches(codegen->GetGraph()->GetArena()->Adapter());
838 codegen->EmitLinkerPatches(&linker_patches);
839
840 // Sort patches by literal offset. Required for .oat_patches encoding.
841 std::sort(linker_patches.begin(), linker_patches.end(),
842 [](const LinkerPatch& lhs, const LinkerPatch& rhs) {
843 return lhs.LiteralOffset() < rhs.LiteralOffset();
844 });
845
846 return linker_patches;
847 }
848
Emit(ArenaAllocator * arena,CodeVectorAllocator * code_allocator,CodeGenerator * codegen,CompilerDriver * compiler_driver,const DexFile::CodeItem * code_item) const849 CompiledMethod* OptimizingCompiler::Emit(ArenaAllocator* arena,
850 CodeVectorAllocator* code_allocator,
851 CodeGenerator* codegen,
852 CompilerDriver* compiler_driver,
853 const DexFile::CodeItem* code_item) const {
854 ArenaVector<LinkerPatch> linker_patches = EmitAndSortLinkerPatches(codegen);
855 ArenaVector<uint8_t> stack_map(arena->Adapter(kArenaAllocStackMaps));
856 ArenaVector<uint8_t> method_info(arena->Adapter(kArenaAllocStackMaps));
857 size_t stack_map_size = 0;
858 size_t method_info_size = 0;
859 codegen->ComputeStackMapAndMethodInfoSize(&stack_map_size, &method_info_size);
860 stack_map.resize(stack_map_size);
861 method_info.resize(method_info_size);
862 codegen->BuildStackMaps(MemoryRegion(stack_map.data(), stack_map.size()),
863 MemoryRegion(method_info.data(), method_info.size()),
864 *code_item);
865
866 CompiledMethod* compiled_method = CompiledMethod::SwapAllocCompiledMethod(
867 compiler_driver,
868 codegen->GetInstructionSet(),
869 ArrayRef<const uint8_t>(code_allocator->GetMemory()),
870 // Follow Quick's behavior and set the frame size to zero if it is
871 // considered "empty" (see the definition of
872 // art::CodeGenerator::HasEmptyFrame).
873 codegen->HasEmptyFrame() ? 0 : codegen->GetFrameSize(),
874 codegen->GetCoreSpillMask(),
875 codegen->GetFpuSpillMask(),
876 ArrayRef<const uint8_t>(method_info),
877 ArrayRef<const uint8_t>(stack_map),
878 ArrayRef<const uint8_t>(*codegen->GetAssembler()->cfi().data()),
879 ArrayRef<const LinkerPatch>(linker_patches));
880
881 return compiled_method;
882 }
883
TryCompile(ArenaAllocator * arena,CodeVectorAllocator * code_allocator,const DexFile::CodeItem * code_item,uint32_t access_flags,InvokeType invoke_type,uint16_t class_def_idx,uint32_t method_idx,Handle<mirror::ClassLoader> class_loader,const DexFile & dex_file,Handle<mirror::DexCache> dex_cache,ArtMethod * method,bool osr,VariableSizedHandleScope * handles) const884 CodeGenerator* OptimizingCompiler::TryCompile(ArenaAllocator* arena,
885 CodeVectorAllocator* code_allocator,
886 const DexFile::CodeItem* code_item,
887 uint32_t access_flags,
888 InvokeType invoke_type,
889 uint16_t class_def_idx,
890 uint32_t method_idx,
891 Handle<mirror::ClassLoader> class_loader,
892 const DexFile& dex_file,
893 Handle<mirror::DexCache> dex_cache,
894 ArtMethod* method,
895 bool osr,
896 VariableSizedHandleScope* handles) const {
897 MaybeRecordStat(MethodCompilationStat::kAttemptCompilation);
898 CompilerDriver* compiler_driver = GetCompilerDriver();
899 InstructionSet instruction_set = compiler_driver->GetInstructionSet();
900
901 // Always use the Thumb-2 assembler: some runtime functionality
902 // (like implicit stack overflow checks) assume Thumb-2.
903 DCHECK_NE(instruction_set, kArm);
904
905 // Do not attempt to compile on architectures we do not support.
906 if (!IsInstructionSetSupported(instruction_set)) {
907 MaybeRecordStat(MethodCompilationStat::kNotCompiledUnsupportedIsa);
908 return nullptr;
909 }
910
911 if (Compiler::IsPathologicalCase(*code_item, method_idx, dex_file)) {
912 MaybeRecordStat(MethodCompilationStat::kNotCompiledPathological);
913 return nullptr;
914 }
915
916 // Implementation of the space filter: do not compile a code item whose size in
917 // code units is bigger than 128.
918 static constexpr size_t kSpaceFilterOptimizingThreshold = 128;
919 const CompilerOptions& compiler_options = compiler_driver->GetCompilerOptions();
920 if ((compiler_options.GetCompilerFilter() == CompilerFilter::kSpace)
921 && (code_item->insns_size_in_code_units_ > kSpaceFilterOptimizingThreshold)) {
922 MaybeRecordStat(MethodCompilationStat::kNotCompiledSpaceFilter);
923 return nullptr;
924 }
925
926 ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
927 DexCompilationUnit dex_compilation_unit(
928 class_loader,
929 class_linker,
930 dex_file,
931 code_item,
932 class_def_idx,
933 method_idx,
934 access_flags,
935 /* verified_method */ nullptr,
936 dex_cache);
937
938 HGraph* graph = new (arena) HGraph(
939 arena,
940 dex_file,
941 method_idx,
942 compiler_driver->GetInstructionSet(),
943 kInvalidInvokeType,
944 compiler_driver->GetCompilerOptions().GetDebuggable(),
945 osr);
946
947 const uint8_t* interpreter_metadata = nullptr;
948 if (method == nullptr) {
949 ScopedObjectAccess soa(Thread::Current());
950 method = compiler_driver->ResolveMethod(
951 soa, dex_cache, class_loader, &dex_compilation_unit, method_idx, invoke_type);
952 }
953 // For AOT compilation, we may not get a method, for example if its class is erroneous.
954 // JIT should always have a method.
955 DCHECK(Runtime::Current()->IsAotCompiler() || method != nullptr);
956 if (method != nullptr) {
957 graph->SetArtMethod(method);
958 ScopedObjectAccess soa(Thread::Current());
959 interpreter_metadata = method->GetQuickenedInfo(class_linker->GetImagePointerSize());
960 }
961
962 std::unique_ptr<CodeGenerator> codegen(
963 CodeGenerator::Create(graph,
964 instruction_set,
965 *compiler_driver->GetInstructionSetFeatures(),
966 compiler_driver->GetCompilerOptions(),
967 compilation_stats_.get()));
968 if (codegen.get() == nullptr) {
969 MaybeRecordStat(MethodCompilationStat::kNotCompiledNoCodegen);
970 return nullptr;
971 }
972 codegen->GetAssembler()->cfi().SetEnabled(
973 compiler_driver->GetCompilerOptions().GenerateAnyDebugInfo());
974
975 PassObserver pass_observer(graph,
976 codegen.get(),
977 visualizer_output_.get(),
978 compiler_driver,
979 dump_mutex_);
980
981 {
982 VLOG(compiler) << "Building " << pass_observer.GetMethodName();
983 PassScope scope(HGraphBuilder::kBuilderPassName, &pass_observer);
984 HGraphBuilder builder(graph,
985 &dex_compilation_unit,
986 &dex_compilation_unit,
987 &dex_file,
988 *code_item,
989 compiler_driver,
990 codegen.get(),
991 compilation_stats_.get(),
992 interpreter_metadata,
993 dex_cache,
994 handles);
995 GraphAnalysisResult result = builder.BuildGraph();
996 if (result != kAnalysisSuccess) {
997 switch (result) {
998 case kAnalysisSkipped:
999 MaybeRecordStat(MethodCompilationStat::kNotCompiledSkipped);
1000 break;
1001 case kAnalysisInvalidBytecode:
1002 MaybeRecordStat(MethodCompilationStat::kNotCompiledInvalidBytecode);
1003 break;
1004 case kAnalysisFailThrowCatchLoop:
1005 MaybeRecordStat(MethodCompilationStat::kNotCompiledThrowCatchLoop);
1006 break;
1007 case kAnalysisFailAmbiguousArrayOp:
1008 MaybeRecordStat(MethodCompilationStat::kNotCompiledAmbiguousArrayOp);
1009 break;
1010 case kAnalysisSuccess:
1011 UNREACHABLE();
1012 }
1013 pass_observer.SetGraphInBadState();
1014 return nullptr;
1015 }
1016 }
1017
1018 RunOptimizations(graph,
1019 codegen.get(),
1020 compiler_driver,
1021 dex_compilation_unit,
1022 &pass_observer,
1023 handles);
1024
1025 RegisterAllocator::Strategy regalloc_strategy =
1026 compiler_options.GetRegisterAllocationStrategy();
1027 AllocateRegisters(graph, codegen.get(), &pass_observer, regalloc_strategy);
1028
1029 codegen->Compile(code_allocator);
1030 pass_observer.DumpDisassembly();
1031
1032 return codegen.release();
1033 }
1034
Compile(const DexFile::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) const1035 CompiledMethod* OptimizingCompiler::Compile(const DexFile::CodeItem* code_item,
1036 uint32_t access_flags,
1037 InvokeType invoke_type,
1038 uint16_t class_def_idx,
1039 uint32_t method_idx,
1040 Handle<mirror::ClassLoader> jclass_loader,
1041 const DexFile& dex_file,
1042 Handle<mirror::DexCache> dex_cache) const {
1043 CompilerDriver* compiler_driver = GetCompilerDriver();
1044 CompiledMethod* method = nullptr;
1045 DCHECK(Runtime::Current()->IsAotCompiler());
1046 const VerifiedMethod* verified_method = compiler_driver->GetVerifiedMethod(&dex_file, method_idx);
1047 DCHECK(!verified_method->HasRuntimeThrow());
1048 if (compiler_driver->IsMethodVerifiedWithoutFailures(method_idx, class_def_idx, dex_file)
1049 || verifier::CanCompilerHandleVerificationFailure(
1050 verified_method->GetEncounteredVerificationFailures())) {
1051 ArenaAllocator arena(Runtime::Current()->GetArenaPool());
1052 CodeVectorAllocator code_allocator(&arena);
1053 std::unique_ptr<CodeGenerator> codegen;
1054 {
1055 ScopedObjectAccess soa(Thread::Current());
1056 VariableSizedHandleScope handles(soa.Self());
1057 // Go to native so that we don't block GC during compilation.
1058 ScopedThreadSuspension sts(soa.Self(), kNative);
1059 codegen.reset(
1060 TryCompile(&arena,
1061 &code_allocator,
1062 code_item,
1063 access_flags,
1064 invoke_type,
1065 class_def_idx,
1066 method_idx,
1067 jclass_loader,
1068 dex_file,
1069 dex_cache,
1070 nullptr,
1071 /* osr */ false,
1072 &handles));
1073 }
1074 if (codegen.get() != nullptr) {
1075 MaybeRecordStat(MethodCompilationStat::kCompiled);
1076 method = Emit(&arena, &code_allocator, codegen.get(), compiler_driver, code_item);
1077
1078 if (kArenaAllocatorCountAllocations) {
1079 if (arena.BytesAllocated() > kArenaAllocatorMemoryReportThreshold) {
1080 MemStats mem_stats(arena.GetMemStats());
1081 LOG(INFO) << dex_file.PrettyMethod(method_idx) << " " << Dumpable<MemStats>(mem_stats);
1082 }
1083 }
1084 }
1085 } else {
1086 if (compiler_driver->GetCompilerOptions().VerifyAtRuntime()) {
1087 MaybeRecordStat(MethodCompilationStat::kNotCompiledVerifyAtRuntime);
1088 } else {
1089 MaybeRecordStat(MethodCompilationStat::kNotCompiledVerificationError);
1090 }
1091 }
1092
1093 if (kIsDebugBuild &&
1094 IsCompilingWithCoreImage() &&
1095 IsInstructionSetSupported(compiler_driver->GetInstructionSet())) {
1096 // For testing purposes, we put a special marker on method names
1097 // that should be compiled with this compiler (when the
1098 // instruction set is supported). This makes sure we're not
1099 // regressing.
1100 std::string method_name = dex_file.PrettyMethod(method_idx);
1101 bool shouldCompile = method_name.find("$opt$") != std::string::npos;
1102 DCHECK((method != nullptr) || !shouldCompile) << "Didn't compile " << method_name;
1103 }
1104
1105 return method;
1106 }
1107
CreateOptimizingCompiler(CompilerDriver * driver)1108 Compiler* CreateOptimizingCompiler(CompilerDriver* driver) {
1109 return new OptimizingCompiler(driver);
1110 }
1111
IsCompilingWithCoreImage()1112 bool IsCompilingWithCoreImage() {
1113 const std::string& image = Runtime::Current()->GetImageLocation();
1114 // TODO: This is under-approximating...
1115 if (android::base::EndsWith(image, "core.art") ||
1116 android::base::EndsWith(image, "core-optimizing.art")) {
1117 return true;
1118 }
1119 return false;
1120 }
1121
EncodeArtMethodInInlineInfo(ArtMethod * method ATTRIBUTE_UNUSED)1122 bool EncodeArtMethodInInlineInfo(ArtMethod* method ATTRIBUTE_UNUSED) {
1123 // Note: the runtime is null only for unit testing.
1124 return Runtime::Current() == nullptr || !Runtime::Current()->IsAotCompiler();
1125 }
1126
CanEncodeInlinedMethodInStackMap(const DexFile & caller_dex_file,ArtMethod * callee)1127 bool CanEncodeInlinedMethodInStackMap(const DexFile& caller_dex_file, ArtMethod* callee) {
1128 if (!Runtime::Current()->IsAotCompiler()) {
1129 // JIT can always encode methods in stack maps.
1130 return true;
1131 }
1132 if (IsSameDexFile(caller_dex_file, *callee->GetDexFile())) {
1133 return true;
1134 }
1135 // TODO(ngeoffray): Support more AOT cases for inlining:
1136 // - methods in multidex
1137 // - methods in boot image for on-device non-PIC compilation.
1138 return false;
1139 }
1140
JitCompile(Thread * self,jit::JitCodeCache * code_cache,ArtMethod * method,bool osr,jit::JitLogger * jit_logger)1141 bool OptimizingCompiler::JitCompile(Thread* self,
1142 jit::JitCodeCache* code_cache,
1143 ArtMethod* method,
1144 bool osr,
1145 jit::JitLogger* jit_logger) {
1146 StackHandleScope<3> hs(self);
1147 Handle<mirror::ClassLoader> class_loader(hs.NewHandle(
1148 method->GetDeclaringClass()->GetClassLoader()));
1149 Handle<mirror::DexCache> dex_cache(hs.NewHandle(method->GetDexCache()));
1150 DCHECK(method->IsCompilable());
1151
1152 const DexFile* dex_file = method->GetDexFile();
1153 const uint16_t class_def_idx = method->GetClassDefIndex();
1154 const DexFile::CodeItem* code_item = dex_file->GetCodeItem(method->GetCodeItemOffset());
1155 const uint32_t method_idx = method->GetDexMethodIndex();
1156 const uint32_t access_flags = method->GetAccessFlags();
1157 const InvokeType invoke_type = method->GetInvokeType();
1158
1159 ArenaAllocator arena(Runtime::Current()->GetJitArenaPool());
1160 CodeVectorAllocator code_allocator(&arena);
1161 VariableSizedHandleScope handles(self);
1162
1163 std::unique_ptr<CodeGenerator> codegen;
1164 {
1165 // Go to native so that we don't block GC during compilation.
1166 ScopedThreadSuspension sts(self, kNative);
1167 codegen.reset(
1168 TryCompile(&arena,
1169 &code_allocator,
1170 code_item,
1171 access_flags,
1172 invoke_type,
1173 class_def_idx,
1174 method_idx,
1175 class_loader,
1176 *dex_file,
1177 dex_cache,
1178 method,
1179 osr,
1180 &handles));
1181 if (codegen.get() == nullptr) {
1182 return false;
1183 }
1184
1185 if (kArenaAllocatorCountAllocations) {
1186 if (arena.BytesAllocated() > kArenaAllocatorMemoryReportThreshold) {
1187 MemStats mem_stats(arena.GetMemStats());
1188 LOG(INFO) << dex_file->PrettyMethod(method_idx) << " " << Dumpable<MemStats>(mem_stats);
1189 }
1190 }
1191 }
1192
1193 size_t stack_map_size = 0;
1194 size_t method_info_size = 0;
1195 codegen->ComputeStackMapAndMethodInfoSize(&stack_map_size, &method_info_size);
1196 size_t number_of_roots = codegen->GetNumberOfJitRoots();
1197 ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
1198 // We allocate an object array to ensure the JIT roots that we will collect in EmitJitRoots
1199 // will be visible by the GC between EmitLiterals and CommitCode. Once CommitCode is
1200 // executed, this array is not needed.
1201 Handle<mirror::ObjectArray<mirror::Object>> roots(
1202 hs.NewHandle(mirror::ObjectArray<mirror::Object>::Alloc(
1203 self, class_linker->GetClassRoot(ClassLinker::kObjectArrayClass), number_of_roots)));
1204 if (roots == nullptr) {
1205 // Out of memory, just clear the exception to avoid any Java exception uncaught problems.
1206 DCHECK(self->IsExceptionPending());
1207 self->ClearException();
1208 return false;
1209 }
1210 uint8_t* stack_map_data = nullptr;
1211 uint8_t* method_info_data = nullptr;
1212 uint8_t* roots_data = nullptr;
1213 uint32_t data_size = code_cache->ReserveData(self,
1214 stack_map_size,
1215 method_info_size,
1216 number_of_roots,
1217 method,
1218 &stack_map_data,
1219 &method_info_data,
1220 &roots_data);
1221 if (stack_map_data == nullptr || roots_data == nullptr) {
1222 return false;
1223 }
1224 MaybeRecordStat(MethodCompilationStat::kCompiled);
1225 codegen->BuildStackMaps(MemoryRegion(stack_map_data, stack_map_size),
1226 MemoryRegion(method_info_data, method_info_size),
1227 *code_item);
1228 codegen->EmitJitRoots(code_allocator.GetData(), roots, roots_data);
1229
1230 const void* code = code_cache->CommitCode(
1231 self,
1232 method,
1233 stack_map_data,
1234 method_info_data,
1235 roots_data,
1236 codegen->HasEmptyFrame() ? 0 : codegen->GetFrameSize(),
1237 codegen->GetCoreSpillMask(),
1238 codegen->GetFpuSpillMask(),
1239 code_allocator.GetMemory().data(),
1240 code_allocator.GetSize(),
1241 data_size,
1242 osr,
1243 roots,
1244 codegen->GetGraph()->HasShouldDeoptimizeFlag(),
1245 codegen->GetGraph()->GetCHASingleImplementationList());
1246
1247 if (code == nullptr) {
1248 code_cache->ClearData(self, stack_map_data, roots_data);
1249 return false;
1250 }
1251
1252 const CompilerOptions& compiler_options = GetCompilerDriver()->GetCompilerOptions();
1253 if (compiler_options.GetGenerateDebugInfo()) {
1254 const auto* method_header = reinterpret_cast<const OatQuickMethodHeader*>(code);
1255 const uintptr_t code_address = reinterpret_cast<uintptr_t>(method_header->GetCode());
1256 debug::MethodDebugInfo info = debug::MethodDebugInfo();
1257 info.trampoline_name = nullptr;
1258 info.dex_file = dex_file;
1259 info.class_def_index = class_def_idx;
1260 info.dex_method_index = method_idx;
1261 info.access_flags = access_flags;
1262 info.code_item = code_item;
1263 info.isa = codegen->GetInstructionSet();
1264 info.deduped = false;
1265 info.is_native_debuggable = compiler_options.GetNativeDebuggable();
1266 info.is_optimized = true;
1267 info.is_code_address_text_relative = false;
1268 info.code_address = code_address;
1269 info.code_size = code_allocator.GetSize();
1270 info.frame_size_in_bytes = method_header->GetFrameSizeInBytes();
1271 info.code_info = stack_map_size == 0 ? nullptr : stack_map_data;
1272 info.cfi = ArrayRef<const uint8_t>(*codegen->GetAssembler()->cfi().data());
1273 std::vector<uint8_t> elf_file = debug::WriteDebugElfFileForMethods(
1274 GetCompilerDriver()->GetInstructionSet(),
1275 GetCompilerDriver()->GetInstructionSetFeatures(),
1276 ArrayRef<const debug::MethodDebugInfo>(&info, 1));
1277 CreateJITCodeEntryForAddress(code_address, std::move(elf_file));
1278 }
1279
1280 Runtime::Current()->GetJit()->AddMemoryUsage(method, arena.BytesUsed());
1281 if (jit_logger != nullptr) {
1282 jit_logger->WriteLog(code, code_allocator.GetSize(), method);
1283 }
1284
1285 return true;
1286 }
1287
1288 } // namespace art
1289