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_arm
28 #include "dex_cache_array_fixups_arm.h"
29 #endif
30
31 #ifdef ART_ENABLE_CODEGEN_arm64
32 #include "instruction_simplifier_arm64.h"
33 #endif
34
35 #ifdef ART_ENABLE_CODEGEN_mips
36 #include "dex_cache_array_fixups_mips.h"
37 #include "pc_relative_fixups_mips.h"
38 #endif
39
40 #ifdef ART_ENABLE_CODEGEN_x86
41 #include "pc_relative_fixups_x86.h"
42 #endif
43
44 #if defined(ART_ENABLE_CODEGEN_x86) || defined(ART_ENABLE_CODEGEN_x86_64)
45 #include "x86_memory_gen.h"
46 #endif
47
48 #include "art_method-inl.h"
49 #include "base/arena_allocator.h"
50 #include "base/arena_containers.h"
51 #include "base/dumpable.h"
52 #include "base/macros.h"
53 #include "base/mutex.h"
54 #include "base/timing_logger.h"
55 #include "bounds_check_elimination.h"
56 #include "builder.h"
57 #include "cha_guard_optimization.h"
58 #include "code_generator.h"
59 #include "code_sinking.h"
60 #include "compiled_method.h"
61 #include "compiler.h"
62 #include "constant_folding.h"
63 #include "dead_code_elimination.h"
64 #include "debug/elf_debug_writer.h"
65 #include "debug/method_debug_info.h"
66 #include "dex/verification_results.h"
67 #include "dex/verified_method.h"
68 #include "dex_file_types.h"
69 #include "driver/compiler_driver-inl.h"
70 #include "driver/compiler_options.h"
71 #include "driver/dex_compilation_unit.h"
72 #include "elf_writer_quick.h"
73 #include "graph_checker.h"
74 #include "graph_visualizer.h"
75 #include "gvn.h"
76 #include "induction_var_analysis.h"
77 #include "inliner.h"
78 #include "instruction_simplifier.h"
79 #include "instruction_simplifier_arm.h"
80 #include "intrinsics.h"
81 #include "jit/debugger_interface.h"
82 #include "jit/jit.h"
83 #include "jit/jit_code_cache.h"
84 #include "jni/quick/jni_compiler.h"
85 #include "licm.h"
86 #include "load_store_elimination.h"
87 #include "loop_optimization.h"
88 #include "nodes.h"
89 #include "oat_quick_method_header.h"
90 #include "prepare_for_register_allocation.h"
91 #include "reference_type_propagation.h"
92 #include "register_allocator_linear_scan.h"
93 #include "select_generator.h"
94 #include "scheduler.h"
95 #include "sharpening.h"
96 #include "side_effects_analysis.h"
97 #include "ssa_builder.h"
98 #include "ssa_liveness_analysis.h"
99 #include "ssa_phi_elimination.h"
100 #include "utils/assembler.h"
101 #include "verifier/verifier_compiler_binding.h"
102
103 namespace art {
104
105 static constexpr size_t kArenaAllocatorMemoryReportThreshold = 8 * MB;
106
107 static constexpr const char* kPassNameSeparator = "$";
108
109 /**
110 * Used by the code generator, to allocate the code in a vector.
111 */
112 class CodeVectorAllocator FINAL : public CodeAllocator {
113 public:
CodeVectorAllocator(ArenaAllocator * arena)114 explicit CodeVectorAllocator(ArenaAllocator* arena)
115 : memory_(arena->Adapter(kArenaAllocCodeBuffer)),
116 size_(0) {}
117
Allocate(size_t size)118 virtual uint8_t* Allocate(size_t size) {
119 size_ = size;
120 memory_.resize(size);
121 return &memory_[0];
122 }
123
GetSize() const124 size_t GetSize() const { return size_; }
GetMemory() const125 const ArenaVector<uint8_t>& GetMemory() const { return memory_; }
GetData()126 uint8_t* GetData() { return memory_.data(); }
127
128 private:
129 ArenaVector<uint8_t> memory_;
130 size_t size_;
131
132 DISALLOW_COPY_AND_ASSIGN(CodeVectorAllocator);
133 };
134
135 /**
136 * Filter to apply to the visualizer. Methods whose name contain that filter will
137 * be dumped.
138 */
139 static constexpr const char kStringFilter[] = "";
140
141 class PassScope;
142
143 class PassObserver : public ValueObject {
144 public:
PassObserver(HGraph * graph,CodeGenerator * codegen,std::ostream * visualizer_output,CompilerDriver * compiler_driver,Mutex & dump_mutex)145 PassObserver(HGraph* graph,
146 CodeGenerator* codegen,
147 std::ostream* visualizer_output,
148 CompilerDriver* compiler_driver,
149 Mutex& dump_mutex)
150 : graph_(graph),
151 cached_method_name_(),
152 timing_logger_enabled_(compiler_driver->GetDumpPasses()),
153 timing_logger_(timing_logger_enabled_ ? GetMethodName() : "", true, true),
154 disasm_info_(graph->GetArena()),
155 visualizer_oss_(),
156 visualizer_output_(visualizer_output),
157 visualizer_enabled_(!compiler_driver->GetCompilerOptions().GetDumpCfgFileName().empty()),
158 visualizer_(&visualizer_oss_, graph, *codegen),
159 visualizer_dump_mutex_(dump_mutex),
160 graph_in_bad_state_(false) {
161 if (timing_logger_enabled_ || visualizer_enabled_) {
162 if (!IsVerboseMethod(compiler_driver, GetMethodName())) {
163 timing_logger_enabled_ = visualizer_enabled_ = false;
164 }
165 if (visualizer_enabled_) {
166 visualizer_.PrintHeader(GetMethodName());
167 codegen->SetDisassemblyInformation(&disasm_info_);
168 }
169 }
170 }
171
~PassObserver()172 ~PassObserver() {
173 if (timing_logger_enabled_) {
174 LOG(INFO) << "TIMINGS " << GetMethodName();
175 LOG(INFO) << Dumpable<TimingLogger>(timing_logger_);
176 }
177 DCHECK(visualizer_oss_.str().empty());
178 }
179
DumpDisassembly()180 void DumpDisassembly() REQUIRES(!visualizer_dump_mutex_) {
181 if (visualizer_enabled_) {
182 visualizer_.DumpGraphWithDisassembly();
183 FlushVisualizer();
184 }
185 }
186
SetGraphInBadState()187 void SetGraphInBadState() { graph_in_bad_state_ = true; }
188
GetMethodName()189 const char* GetMethodName() {
190 // PrettyMethod() is expensive, so we delay calling it until we actually have to.
191 if (cached_method_name_.empty()) {
192 cached_method_name_ = graph_->GetDexFile().PrettyMethod(graph_->GetMethodIdx());
193 }
194 return cached_method_name_.c_str();
195 }
196
197 private:
StartPass(const char * pass_name)198 void StartPass(const char* pass_name) REQUIRES(!visualizer_dump_mutex_) {
199 VLOG(compiler) << "Starting pass: " << pass_name;
200 // Dump graph first, then start timer.
201 if (visualizer_enabled_) {
202 visualizer_.DumpGraph(pass_name, /* is_after_pass */ false, graph_in_bad_state_);
203 FlushVisualizer();
204 }
205 if (timing_logger_enabled_) {
206 timing_logger_.StartTiming(pass_name);
207 }
208 }
209
FlushVisualizer()210 void FlushVisualizer() REQUIRES(!visualizer_dump_mutex_) {
211 MutexLock mu(Thread::Current(), visualizer_dump_mutex_);
212 *visualizer_output_ << visualizer_oss_.str();
213 visualizer_output_->flush();
214 visualizer_oss_.str("");
215 visualizer_oss_.clear();
216 }
217
EndPass(const char * pass_name)218 void EndPass(const char* pass_name) REQUIRES(!visualizer_dump_mutex_) {
219 // Pause timer first, then dump graph.
220 if (timing_logger_enabled_) {
221 timing_logger_.EndTiming();
222 }
223 if (visualizer_enabled_) {
224 visualizer_.DumpGraph(pass_name, /* is_after_pass */ true, graph_in_bad_state_);
225 FlushVisualizer();
226 }
227
228 // Validate the HGraph if running in debug mode.
229 if (kIsDebugBuild) {
230 if (!graph_in_bad_state_) {
231 GraphChecker checker(graph_);
232 checker.Run();
233 if (!checker.IsValid()) {
234 LOG(FATAL) << "Error after " << pass_name << ": " << Dumpable<GraphChecker>(checker);
235 }
236 }
237 }
238 }
239
IsVerboseMethod(CompilerDriver * compiler_driver,const char * method_name)240 static bool IsVerboseMethod(CompilerDriver* compiler_driver, const char* method_name) {
241 // Test an exact match to --verbose-methods. If verbose-methods is set, this overrides an
242 // empty kStringFilter matching all methods.
243 if (compiler_driver->GetCompilerOptions().HasVerboseMethods()) {
244 return compiler_driver->GetCompilerOptions().IsVerboseMethod(method_name);
245 }
246
247 // Test the kStringFilter sub-string. constexpr helper variable to silence unreachable-code
248 // warning when the string is empty.
249 constexpr bool kStringFilterEmpty = arraysize(kStringFilter) <= 1;
250 if (kStringFilterEmpty || strstr(method_name, kStringFilter) != nullptr) {
251 return true;
252 }
253
254 return false;
255 }
256
257 HGraph* const graph_;
258
259 std::string cached_method_name_;
260
261 bool timing_logger_enabled_;
262 TimingLogger timing_logger_;
263
264 DisassemblyInformation disasm_info_;
265
266 std::ostringstream visualizer_oss_;
267 std::ostream* visualizer_output_;
268 bool visualizer_enabled_;
269 HGraphVisualizer visualizer_;
270 Mutex& visualizer_dump_mutex_;
271
272 // Flag to be set by the compiler if the pass failed and the graph is not
273 // expected to validate.
274 bool graph_in_bad_state_;
275
276 friend PassScope;
277
278 DISALLOW_COPY_AND_ASSIGN(PassObserver);
279 };
280
281 class PassScope : public ValueObject {
282 public:
PassScope(const char * pass_name,PassObserver * pass_observer)283 PassScope(const char *pass_name, PassObserver* pass_observer)
284 : pass_name_(pass_name),
285 pass_observer_(pass_observer) {
286 pass_observer_->StartPass(pass_name_);
287 }
288
~PassScope()289 ~PassScope() {
290 pass_observer_->EndPass(pass_name_);
291 }
292
293 private:
294 const char* const pass_name_;
295 PassObserver* const pass_observer_;
296 };
297
298 class OptimizingCompiler FINAL : public Compiler {
299 public:
300 explicit OptimizingCompiler(CompilerDriver* driver);
301 ~OptimizingCompiler() OVERRIDE;
302
303 bool CanCompileMethod(uint32_t method_idx, const DexFile& dex_file) const OVERRIDE;
304
305 CompiledMethod* Compile(const DexFile::CodeItem* code_item,
306 uint32_t access_flags,
307 InvokeType invoke_type,
308 uint16_t class_def_idx,
309 uint32_t method_idx,
310 Handle<mirror::ClassLoader> class_loader,
311 const DexFile& dex_file,
312 Handle<mirror::DexCache> dex_cache) const OVERRIDE;
313
JniCompile(uint32_t access_flags,uint32_t method_idx,const DexFile & dex_file,JniOptimizationFlags optimization_flags) const314 CompiledMethod* JniCompile(uint32_t access_flags,
315 uint32_t method_idx,
316 const DexFile& dex_file,
317 JniOptimizationFlags optimization_flags) const OVERRIDE {
318 return ArtQuickJniCompileMethod(GetCompilerDriver(),
319 access_flags,
320 method_idx,
321 dex_file,
322 optimization_flags);
323 }
324
GetEntryPointOf(ArtMethod * method) const325 uintptr_t GetEntryPointOf(ArtMethod* method) const OVERRIDE
326 REQUIRES_SHARED(Locks::mutator_lock_) {
327 return reinterpret_cast<uintptr_t>(method->GetEntryPointFromQuickCompiledCodePtrSize(
328 InstructionSetPointerSize(GetCompilerDriver()->GetInstructionSet())));
329 }
330
331 void Init() OVERRIDE;
332
333 void UnInit() const OVERRIDE;
334
MaybeRecordStat(MethodCompilationStat compilation_stat) const335 void MaybeRecordStat(MethodCompilationStat compilation_stat) const {
336 if (compilation_stats_.get() != nullptr) {
337 compilation_stats_->RecordStat(compilation_stat);
338 }
339 }
340
341 bool JitCompile(Thread* self, jit::JitCodeCache* code_cache, ArtMethod* method, bool osr)
342 OVERRIDE
343 REQUIRES_SHARED(Locks::mutator_lock_);
344
345 private:
346 void RunOptimizations(HGraph* graph,
347 CodeGenerator* codegen,
348 CompilerDriver* driver,
349 const DexCompilationUnit& dex_compilation_unit,
350 PassObserver* pass_observer,
351 VariableSizedHandleScope* handles) const;
352
353 void RunOptimizations(HOptimization* optimizations[],
354 size_t length,
355 PassObserver* pass_observer) const;
356
357 private:
358 // Create a 'CompiledMethod' for an optimized graph.
359 CompiledMethod* Emit(ArenaAllocator* arena,
360 CodeVectorAllocator* code_allocator,
361 CodeGenerator* codegen,
362 CompilerDriver* driver,
363 const DexFile::CodeItem* item) const;
364
365 // Try compiling a method and return the code generator used for
366 // compiling it.
367 // This method:
368 // 1) Builds the graph. Returns null if it failed to build it.
369 // 2) Transforms the graph to SSA. Returns null if it failed.
370 // 3) Runs optimizations on the graph, including register allocator.
371 // 4) Generates code with the `code_allocator` provided.
372 CodeGenerator* TryCompile(ArenaAllocator* arena,
373 CodeVectorAllocator* code_allocator,
374 const DexFile::CodeItem* code_item,
375 uint32_t access_flags,
376 InvokeType invoke_type,
377 uint16_t class_def_idx,
378 uint32_t method_idx,
379 Handle<mirror::ClassLoader> class_loader,
380 const DexFile& dex_file,
381 Handle<mirror::DexCache> dex_cache,
382 ArtMethod* method,
383 bool osr,
384 VariableSizedHandleScope* handles) const;
385
386 void MaybeRunInliner(HGraph* graph,
387 CodeGenerator* codegen,
388 CompilerDriver* driver,
389 const DexCompilationUnit& dex_compilation_unit,
390 PassObserver* pass_observer,
391 VariableSizedHandleScope* handles) const;
392
393 void RunArchOptimizations(InstructionSet instruction_set,
394 HGraph* graph,
395 CodeGenerator* codegen,
396 PassObserver* pass_observer) const;
397
398 std::unique_ptr<OptimizingCompilerStats> compilation_stats_;
399
400 std::unique_ptr<std::ostream> visualizer_output_;
401
402 mutable Mutex dump_mutex_; // To synchronize visualizer writing.
403
404 DISALLOW_COPY_AND_ASSIGN(OptimizingCompiler);
405 };
406
407 static const int kMaximumCompilationTimeBeforeWarning = 100; /* ms */
408
OptimizingCompiler(CompilerDriver * driver)409 OptimizingCompiler::OptimizingCompiler(CompilerDriver* driver)
410 : Compiler(driver, kMaximumCompilationTimeBeforeWarning),
411 dump_mutex_("Visualizer dump lock") {}
412
Init()413 void OptimizingCompiler::Init() {
414 // Enable C1visualizer output. Must be done in Init() because the compiler
415 // driver is not fully initialized when passed to the compiler's constructor.
416 CompilerDriver* driver = GetCompilerDriver();
417 const std::string cfg_file_name = driver->GetCompilerOptions().GetDumpCfgFileName();
418 if (!cfg_file_name.empty()) {
419 std::ios_base::openmode cfg_file_mode =
420 driver->GetCompilerOptions().GetDumpCfgAppend() ? std::ofstream::app : std::ofstream::out;
421 visualizer_output_.reset(new std::ofstream(cfg_file_name, cfg_file_mode));
422 }
423 if (driver->GetDumpStats()) {
424 compilation_stats_.reset(new OptimizingCompilerStats());
425 }
426 }
427
UnInit() const428 void OptimizingCompiler::UnInit() const {
429 }
430
~OptimizingCompiler()431 OptimizingCompiler::~OptimizingCompiler() {
432 if (compilation_stats_.get() != nullptr) {
433 compilation_stats_->Log();
434 }
435 }
436
CanCompileMethod(uint32_t method_idx ATTRIBUTE_UNUSED,const DexFile & dex_file ATTRIBUTE_UNUSED) const437 bool OptimizingCompiler::CanCompileMethod(uint32_t method_idx ATTRIBUTE_UNUSED,
438 const DexFile& dex_file ATTRIBUTE_UNUSED) const {
439 return true;
440 }
441
IsInstructionSetSupported(InstructionSet instruction_set)442 static bool IsInstructionSetSupported(InstructionSet instruction_set) {
443 return (instruction_set == kArm && !kArm32QuickCodeUseSoftFloat)
444 || instruction_set == kArm64
445 || (instruction_set == kThumb2 && !kArm32QuickCodeUseSoftFloat)
446 || instruction_set == kMips
447 || instruction_set == kMips64
448 || instruction_set == kX86
449 || instruction_set == kX86_64;
450 }
451
452 // Strip pass name suffix to get optimization name.
ConvertPassNameToOptimizationName(const std::string & pass_name)453 static std::string ConvertPassNameToOptimizationName(const std::string& pass_name) {
454 size_t pos = pass_name.find(kPassNameSeparator);
455 return pos == std::string::npos ? pass_name : pass_name.substr(0, pos);
456 }
457
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)458 static HOptimization* BuildOptimization(
459 const std::string& pass_name,
460 ArenaAllocator* arena,
461 HGraph* graph,
462 OptimizingCompilerStats* stats,
463 CodeGenerator* codegen,
464 CompilerDriver* driver,
465 const DexCompilationUnit& dex_compilation_unit,
466 VariableSizedHandleScope* handles,
467 SideEffectsAnalysis* most_recent_side_effects,
468 HInductionVarAnalysis* most_recent_induction) {
469 std::string opt_name = ConvertPassNameToOptimizationName(pass_name);
470 if (opt_name == BoundsCheckElimination::kBoundsCheckEliminationPassName) {
471 CHECK(most_recent_side_effects != nullptr && most_recent_induction != nullptr);
472 return new (arena) BoundsCheckElimination(graph,
473 *most_recent_side_effects,
474 most_recent_induction);
475 } else if (opt_name == GVNOptimization::kGlobalValueNumberingPassName) {
476 CHECK(most_recent_side_effects != nullptr);
477 return new (arena) GVNOptimization(graph, *most_recent_side_effects, pass_name.c_str());
478 } else if (opt_name == HConstantFolding::kConstantFoldingPassName) {
479 return new (arena) HConstantFolding(graph, pass_name.c_str());
480 } else if (opt_name == HDeadCodeElimination::kDeadCodeEliminationPassName) {
481 return new (arena) HDeadCodeElimination(graph, stats, pass_name.c_str());
482 } else if (opt_name == HInliner::kInlinerPassName) {
483 size_t number_of_dex_registers = dex_compilation_unit.GetCodeItem()->registers_size_;
484 return new (arena) HInliner(graph, // outer_graph
485 graph, // outermost_graph
486 codegen,
487 dex_compilation_unit, // outer_compilation_unit
488 dex_compilation_unit, // outermost_compilation_unit
489 driver,
490 handles,
491 stats,
492 number_of_dex_registers,
493 /* total_number_of_instructions */ 0,
494 /* parent */ nullptr);
495 } else if (opt_name == HSharpening::kSharpeningPassName) {
496 return new (arena) HSharpening(graph, codegen, dex_compilation_unit, driver, handles);
497 } else if (opt_name == HSelectGenerator::kSelectGeneratorPassName) {
498 return new (arena) HSelectGenerator(graph, stats);
499 } else if (opt_name == HInductionVarAnalysis::kInductionPassName) {
500 return new (arena) HInductionVarAnalysis(graph);
501 } else if (opt_name == InstructionSimplifier::kInstructionSimplifierPassName) {
502 return new (arena) InstructionSimplifier(graph, codegen, stats, pass_name.c_str());
503 } else if (opt_name == IntrinsicsRecognizer::kIntrinsicsRecognizerPassName) {
504 return new (arena) IntrinsicsRecognizer(graph, stats);
505 } else if (opt_name == LICM::kLoopInvariantCodeMotionPassName) {
506 CHECK(most_recent_side_effects != nullptr);
507 return new (arena) LICM(graph, *most_recent_side_effects, stats);
508 } else if (opt_name == LoadStoreElimination::kLoadStoreEliminationPassName) {
509 CHECK(most_recent_side_effects != nullptr);
510 return new (arena) LoadStoreElimination(graph, *most_recent_side_effects);
511 } else if (opt_name == SideEffectsAnalysis::kSideEffectsAnalysisPassName) {
512 return new (arena) SideEffectsAnalysis(graph);
513 } else if (opt_name == HLoopOptimization::kLoopOptimizationPassName) {
514 return new (arena) HLoopOptimization(graph, driver, most_recent_induction);
515 } else if (opt_name == CHAGuardOptimization::kCHAGuardOptimizationPassName) {
516 return new (arena) CHAGuardOptimization(graph);
517 } else if (opt_name == CodeSinking::kCodeSinkingPassName) {
518 return new (arena) CodeSinking(graph, stats);
519 #ifdef ART_ENABLE_CODEGEN_arm
520 } else if (opt_name == arm::DexCacheArrayFixups::kDexCacheArrayFixupsArmPassName) {
521 return new (arena) arm::DexCacheArrayFixups(graph, codegen, stats);
522 } else if (opt_name == arm::InstructionSimplifierArm::kInstructionSimplifierArmPassName) {
523 return new (arena) arm::InstructionSimplifierArm(graph, stats);
524 #endif
525 #ifdef ART_ENABLE_CODEGEN_arm64
526 } else if (opt_name == arm64::InstructionSimplifierArm64::kInstructionSimplifierArm64PassName) {
527 return new (arena) arm64::InstructionSimplifierArm64(graph, stats);
528 #endif
529 #ifdef ART_ENABLE_CODEGEN_mips
530 } else if (opt_name == mips::DexCacheArrayFixups::kDexCacheArrayFixupsMipsPassName) {
531 return new (arena) mips::DexCacheArrayFixups(graph, codegen, stats);
532 } else if (opt_name == mips::PcRelativeFixups::kPcRelativeFixupsMipsPassName) {
533 return new (arena) mips::PcRelativeFixups(graph, codegen, stats);
534 #endif
535 #ifdef ART_ENABLE_CODEGEN_x86
536 } else if (opt_name == x86::PcRelativeFixups::kPcRelativeFixupsX86PassName) {
537 return new (arena) x86::PcRelativeFixups(graph, codegen, stats);
538 } else if (opt_name == x86::X86MemoryOperandGeneration::kX86MemoryOperandGenerationPassName) {
539 return new (arena) x86::X86MemoryOperandGeneration(graph, codegen, stats);
540 #endif
541 }
542 return nullptr;
543 }
544
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)545 static ArenaVector<HOptimization*> BuildOptimizations(
546 const std::vector<std::string>& pass_names,
547 ArenaAllocator* arena,
548 HGraph* graph,
549 OptimizingCompilerStats* stats,
550 CodeGenerator* codegen,
551 CompilerDriver* driver,
552 const DexCompilationUnit& dex_compilation_unit,
553 VariableSizedHandleScope* handles) {
554 // Few HOptimizations constructors require SideEffectsAnalysis or HInductionVarAnalysis
555 // instances. This method assumes that each of them expects the nearest instance preceeding it
556 // in the pass name list.
557 SideEffectsAnalysis* most_recent_side_effects = nullptr;
558 HInductionVarAnalysis* most_recent_induction = nullptr;
559 ArenaVector<HOptimization*> ret(arena->Adapter());
560 for (const std::string& pass_name : pass_names) {
561 HOptimization* opt = BuildOptimization(
562 pass_name,
563 arena,
564 graph,
565 stats,
566 codegen,
567 driver,
568 dex_compilation_unit,
569 handles,
570 most_recent_side_effects,
571 most_recent_induction);
572 CHECK(opt != nullptr) << "Couldn't build optimization: \"" << pass_name << "\"";
573 ret.push_back(opt);
574
575 std::string opt_name = ConvertPassNameToOptimizationName(pass_name);
576 if (opt_name == SideEffectsAnalysis::kSideEffectsAnalysisPassName) {
577 most_recent_side_effects = down_cast<SideEffectsAnalysis*>(opt);
578 } else if (opt_name == HInductionVarAnalysis::kInductionPassName) {
579 most_recent_induction = down_cast<HInductionVarAnalysis*>(opt);
580 }
581 }
582 return ret;
583 }
584
RunOptimizations(HOptimization * optimizations[],size_t length,PassObserver * pass_observer) const585 void OptimizingCompiler::RunOptimizations(HOptimization* optimizations[],
586 size_t length,
587 PassObserver* pass_observer) const {
588 for (size_t i = 0; i < length; ++i) {
589 PassScope scope(optimizations[i]->GetPassName(), pass_observer);
590 optimizations[i]->Run();
591 }
592 }
593
MaybeRunInliner(HGraph * graph,CodeGenerator * codegen,CompilerDriver * driver,const DexCompilationUnit & dex_compilation_unit,PassObserver * pass_observer,VariableSizedHandleScope * handles) const594 void OptimizingCompiler::MaybeRunInliner(HGraph* graph,
595 CodeGenerator* codegen,
596 CompilerDriver* driver,
597 const DexCompilationUnit& dex_compilation_unit,
598 PassObserver* pass_observer,
599 VariableSizedHandleScope* handles) const {
600 OptimizingCompilerStats* stats = compilation_stats_.get();
601 const CompilerOptions& compiler_options = driver->GetCompilerOptions();
602 bool should_inline = (compiler_options.GetInlineMaxCodeUnits() > 0);
603 if (!should_inline) {
604 return;
605 }
606 size_t number_of_dex_registers = dex_compilation_unit.GetCodeItem()->registers_size_;
607 HInliner* inliner = new (graph->GetArena()) HInliner(
608 graph, // outer_graph
609 graph, // outermost_graph
610 codegen,
611 dex_compilation_unit, // outer_compilation_unit
612 dex_compilation_unit, // outermost_compilation_unit
613 driver,
614 handles,
615 stats,
616 number_of_dex_registers,
617 /* total_number_of_instructions */ 0,
618 /* parent */ nullptr);
619 HOptimization* optimizations[] = { inliner };
620
621 RunOptimizations(optimizations, arraysize(optimizations), pass_observer);
622 }
623
RunArchOptimizations(InstructionSet instruction_set,HGraph * graph,CodeGenerator * codegen,PassObserver * pass_observer) const624 void OptimizingCompiler::RunArchOptimizations(InstructionSet instruction_set,
625 HGraph* graph,
626 CodeGenerator* codegen,
627 PassObserver* pass_observer) const {
628 UNUSED(codegen); // To avoid compilation error when compiling for svelte
629 OptimizingCompilerStats* stats = compilation_stats_.get();
630 ArenaAllocator* arena = graph->GetArena();
631 switch (instruction_set) {
632 #if defined(ART_ENABLE_CODEGEN_arm)
633 case kThumb2:
634 case kArm: {
635 arm::DexCacheArrayFixups* fixups =
636 new (arena) arm::DexCacheArrayFixups(graph, codegen, stats);
637 arm::InstructionSimplifierArm* simplifier =
638 new (arena) arm::InstructionSimplifierArm(graph, stats);
639 SideEffectsAnalysis* side_effects = new (arena) SideEffectsAnalysis(graph);
640 GVNOptimization* gvn = new (arena) GVNOptimization(graph, *side_effects, "GVN$after_arch");
641 HOptimization* arm_optimizations[] = {
642 simplifier,
643 side_effects,
644 gvn,
645 fixups
646 };
647 RunOptimizations(arm_optimizations, arraysize(arm_optimizations), pass_observer);
648 break;
649 }
650 #endif
651 #ifdef ART_ENABLE_CODEGEN_arm64
652 case kArm64: {
653 arm64::InstructionSimplifierArm64* simplifier =
654 new (arena) arm64::InstructionSimplifierArm64(graph, stats);
655 SideEffectsAnalysis* side_effects = new (arena) SideEffectsAnalysis(graph);
656 GVNOptimization* gvn = new (arena) GVNOptimization(graph, *side_effects, "GVN$after_arch");
657 HInstructionScheduling* scheduling =
658 new (arena) HInstructionScheduling(graph, instruction_set);
659 HOptimization* arm64_optimizations[] = {
660 simplifier,
661 side_effects,
662 gvn,
663 scheduling,
664 };
665 RunOptimizations(arm64_optimizations, arraysize(arm64_optimizations), pass_observer);
666 break;
667 }
668 #endif
669 #ifdef ART_ENABLE_CODEGEN_mips
670 case kMips: {
671 mips::PcRelativeFixups* pc_relative_fixups =
672 new (arena) mips::PcRelativeFixups(graph, codegen, stats);
673 mips::DexCacheArrayFixups* dex_cache_array_fixups =
674 new (arena) mips::DexCacheArrayFixups(graph, codegen, stats);
675 HOptimization* mips_optimizations[] = {
676 pc_relative_fixups,
677 dex_cache_array_fixups
678 };
679 RunOptimizations(mips_optimizations, arraysize(mips_optimizations), pass_observer);
680 break;
681 }
682 #endif
683 #ifdef ART_ENABLE_CODEGEN_x86
684 case kX86: {
685 x86::PcRelativeFixups* pc_relative_fixups =
686 new (arena) x86::PcRelativeFixups(graph, codegen, stats);
687 x86::X86MemoryOperandGeneration* memory_gen =
688 new (arena) x86::X86MemoryOperandGeneration(graph, codegen, stats);
689 HOptimization* x86_optimizations[] = {
690 pc_relative_fixups,
691 memory_gen
692 };
693 RunOptimizations(x86_optimizations, arraysize(x86_optimizations), pass_observer);
694 break;
695 }
696 #endif
697 #ifdef ART_ENABLE_CODEGEN_x86_64
698 case kX86_64: {
699 x86::X86MemoryOperandGeneration* memory_gen =
700 new (arena) x86::X86MemoryOperandGeneration(graph, codegen, stats);
701 HOptimization* x86_64_optimizations[] = {
702 memory_gen
703 };
704 RunOptimizations(x86_64_optimizations, arraysize(x86_64_optimizations), pass_observer);
705 break;
706 }
707 #endif
708 default:
709 break;
710 }
711 }
712
713 NO_INLINE // Avoid increasing caller's frame size by large stack-allocated objects.
AllocateRegisters(HGraph * graph,CodeGenerator * codegen,PassObserver * pass_observer,RegisterAllocator::Strategy strategy)714 static void AllocateRegisters(HGraph* graph,
715 CodeGenerator* codegen,
716 PassObserver* pass_observer,
717 RegisterAllocator::Strategy strategy) {
718 {
719 PassScope scope(PrepareForRegisterAllocation::kPrepareForRegisterAllocationPassName,
720 pass_observer);
721 PrepareForRegisterAllocation(graph).Run();
722 }
723 SsaLivenessAnalysis liveness(graph, codegen);
724 {
725 PassScope scope(SsaLivenessAnalysis::kLivenessPassName, pass_observer);
726 liveness.Analyze();
727 }
728 {
729 PassScope scope(RegisterAllocator::kRegisterAllocatorPassName, pass_observer);
730 RegisterAllocator::Create(graph->GetArena(), codegen, liveness, strategy)->AllocateRegisters();
731 }
732 }
733
RunOptimizations(HGraph * graph,CodeGenerator * codegen,CompilerDriver * driver,const DexCompilationUnit & dex_compilation_unit,PassObserver * pass_observer,VariableSizedHandleScope * handles) const734 void OptimizingCompiler::RunOptimizations(HGraph* graph,
735 CodeGenerator* codegen,
736 CompilerDriver* driver,
737 const DexCompilationUnit& dex_compilation_unit,
738 PassObserver* pass_observer,
739 VariableSizedHandleScope* handles) const {
740 OptimizingCompilerStats* stats = compilation_stats_.get();
741 ArenaAllocator* arena = graph->GetArena();
742 if (driver->GetCompilerOptions().GetPassesToRun() != nullptr) {
743 ArenaVector<HOptimization*> optimizations = BuildOptimizations(
744 *driver->GetCompilerOptions().GetPassesToRun(),
745 arena,
746 graph,
747 stats,
748 codegen,
749 driver,
750 dex_compilation_unit,
751 handles);
752 RunOptimizations(&optimizations[0], optimizations.size(), pass_observer);
753 return;
754 }
755
756 HDeadCodeElimination* dce1 = new (arena) HDeadCodeElimination(
757 graph, stats, "dead_code_elimination$initial");
758 HDeadCodeElimination* dce2 = new (arena) HDeadCodeElimination(
759 graph, stats, "dead_code_elimination$after_inlining");
760 HDeadCodeElimination* dce3 = new (arena) HDeadCodeElimination(
761 graph, stats, "dead_code_elimination$final");
762 HConstantFolding* fold1 = new (arena) HConstantFolding(graph, "constant_folding");
763 InstructionSimplifier* simplify1 = new (arena) InstructionSimplifier(graph, codegen, stats);
764 HSelectGenerator* select_generator = new (arena) HSelectGenerator(graph, stats);
765 HConstantFolding* fold2 = new (arena) HConstantFolding(
766 graph, "constant_folding$after_inlining");
767 HConstantFolding* fold3 = new (arena) HConstantFolding(graph, "constant_folding$after_bce");
768 SideEffectsAnalysis* side_effects1 = new (arena) SideEffectsAnalysis(
769 graph, "side_effects$before_gvn");
770 SideEffectsAnalysis* side_effects2 = new (arena) SideEffectsAnalysis(
771 graph, "side_effects$before_lse");
772 GVNOptimization* gvn = new (arena) GVNOptimization(graph, *side_effects1);
773 LICM* licm = new (arena) LICM(graph, *side_effects1, stats);
774 HInductionVarAnalysis* induction = new (arena) HInductionVarAnalysis(graph);
775 BoundsCheckElimination* bce = new (arena) BoundsCheckElimination(graph, *side_effects1, induction);
776 HLoopOptimization* loop = new (arena) HLoopOptimization(graph, driver, induction);
777 LoadStoreElimination* lse = new (arena) LoadStoreElimination(graph, *side_effects2);
778 HSharpening* sharpening = new (arena) HSharpening(
779 graph, codegen, dex_compilation_unit, driver, handles);
780 InstructionSimplifier* simplify2 = new (arena) InstructionSimplifier(
781 graph, codegen, stats, "instruction_simplifier$after_inlining");
782 InstructionSimplifier* simplify3 = new (arena) InstructionSimplifier(
783 graph, codegen, stats, "instruction_simplifier$after_bce");
784 InstructionSimplifier* simplify4 = new (arena) InstructionSimplifier(
785 graph, codegen, stats, "instruction_simplifier$before_codegen");
786 IntrinsicsRecognizer* intrinsics = new (arena) IntrinsicsRecognizer(graph, stats);
787 CHAGuardOptimization* cha_guard = new (arena) CHAGuardOptimization(graph);
788 CodeSinking* code_sinking = new (arena) CodeSinking(graph, stats);
789
790 HOptimization* optimizations1[] = {
791 intrinsics,
792 sharpening,
793 fold1,
794 simplify1,
795 dce1,
796 };
797 RunOptimizations(optimizations1, arraysize(optimizations1), pass_observer);
798
799 MaybeRunInliner(graph, codegen, driver, dex_compilation_unit, pass_observer, handles);
800
801 HOptimization* optimizations2[] = {
802 // SelectGenerator depends on the InstructionSimplifier removing
803 // redundant suspend checks to recognize empty blocks.
804 select_generator,
805 fold2, // TODO: if we don't inline we can also skip fold2.
806 simplify2,
807 dce2,
808 side_effects1,
809 gvn,
810 licm,
811 induction,
812 bce,
813 loop,
814 fold3, // evaluates code generated by dynamic bce
815 simplify3,
816 side_effects2,
817 lse,
818 cha_guard,
819 dce3,
820 code_sinking,
821 // The codegen has a few assumptions that only the instruction simplifier
822 // can satisfy. For example, the code generator does not expect to see a
823 // HTypeConversion from a type to the same type.
824 simplify4,
825 };
826 RunOptimizations(optimizations2, arraysize(optimizations2), pass_observer);
827
828 RunArchOptimizations(driver->GetInstructionSet(), graph, codegen, pass_observer);
829 }
830
EmitAndSortLinkerPatches(CodeGenerator * codegen)831 static ArenaVector<LinkerPatch> EmitAndSortLinkerPatches(CodeGenerator* codegen) {
832 ArenaVector<LinkerPatch> linker_patches(codegen->GetGraph()->GetArena()->Adapter());
833 codegen->EmitLinkerPatches(&linker_patches);
834
835 // Sort patches by literal offset. Required for .oat_patches encoding.
836 std::sort(linker_patches.begin(), linker_patches.end(),
837 [](const LinkerPatch& lhs, const LinkerPatch& rhs) {
838 return lhs.LiteralOffset() < rhs.LiteralOffset();
839 });
840
841 return linker_patches;
842 }
843
Emit(ArenaAllocator * arena,CodeVectorAllocator * code_allocator,CodeGenerator * codegen,CompilerDriver * compiler_driver,const DexFile::CodeItem * code_item) const844 CompiledMethod* OptimizingCompiler::Emit(ArenaAllocator* arena,
845 CodeVectorAllocator* code_allocator,
846 CodeGenerator* codegen,
847 CompilerDriver* compiler_driver,
848 const DexFile::CodeItem* code_item) const {
849 ArenaVector<LinkerPatch> linker_patches = EmitAndSortLinkerPatches(codegen);
850 ArenaVector<uint8_t> stack_map(arena->Adapter(kArenaAllocStackMaps));
851 ArenaVector<uint8_t> method_info(arena->Adapter(kArenaAllocStackMaps));
852 size_t stack_map_size = 0;
853 size_t method_info_size = 0;
854 codegen->ComputeStackMapAndMethodInfoSize(&stack_map_size, &method_info_size);
855 stack_map.resize(stack_map_size);
856 method_info.resize(method_info_size);
857 codegen->BuildStackMaps(MemoryRegion(stack_map.data(), stack_map.size()),
858 MemoryRegion(method_info.data(), method_info.size()),
859 *code_item);
860
861 CompiledMethod* compiled_method = CompiledMethod::SwapAllocCompiledMethod(
862 compiler_driver,
863 codegen->GetInstructionSet(),
864 ArrayRef<const uint8_t>(code_allocator->GetMemory()),
865 // Follow Quick's behavior and set the frame size to zero if it is
866 // considered "empty" (see the definition of
867 // art::CodeGenerator::HasEmptyFrame).
868 codegen->HasEmptyFrame() ? 0 : codegen->GetFrameSize(),
869 codegen->GetCoreSpillMask(),
870 codegen->GetFpuSpillMask(),
871 ArrayRef<const uint8_t>(method_info),
872 ArrayRef<const uint8_t>(stack_map),
873 ArrayRef<const uint8_t>(*codegen->GetAssembler()->cfi().data()),
874 ArrayRef<const LinkerPatch>(linker_patches));
875
876 return compiled_method;
877 }
878
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) const879 CodeGenerator* OptimizingCompiler::TryCompile(ArenaAllocator* arena,
880 CodeVectorAllocator* code_allocator,
881 const DexFile::CodeItem* code_item,
882 uint32_t access_flags,
883 InvokeType invoke_type,
884 uint16_t class_def_idx,
885 uint32_t method_idx,
886 Handle<mirror::ClassLoader> class_loader,
887 const DexFile& dex_file,
888 Handle<mirror::DexCache> dex_cache,
889 ArtMethod* method,
890 bool osr,
891 VariableSizedHandleScope* handles) const {
892 MaybeRecordStat(MethodCompilationStat::kAttemptCompilation);
893 CompilerDriver* compiler_driver = GetCompilerDriver();
894 InstructionSet instruction_set = compiler_driver->GetInstructionSet();
895
896 // Always use the Thumb-2 assembler: some runtime functionality
897 // (like implicit stack overflow checks) assume Thumb-2.
898 DCHECK_NE(instruction_set, kArm);
899
900 // Do not attempt to compile on architectures we do not support.
901 if (!IsInstructionSetSupported(instruction_set)) {
902 MaybeRecordStat(MethodCompilationStat::kNotCompiledUnsupportedIsa);
903 return nullptr;
904 }
905
906 if (Compiler::IsPathologicalCase(*code_item, method_idx, dex_file)) {
907 MaybeRecordStat(MethodCompilationStat::kNotCompiledPathological);
908 return nullptr;
909 }
910
911 // Implementation of the space filter: do not compile a code item whose size in
912 // code units is bigger than 128.
913 static constexpr size_t kSpaceFilterOptimizingThreshold = 128;
914 const CompilerOptions& compiler_options = compiler_driver->GetCompilerOptions();
915 if ((compiler_options.GetCompilerFilter() == CompilerFilter::kSpace)
916 && (code_item->insns_size_in_code_units_ > kSpaceFilterOptimizingThreshold)) {
917 MaybeRecordStat(MethodCompilationStat::kNotCompiledSpaceFilter);
918 return nullptr;
919 }
920
921 ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
922 DexCompilationUnit dex_compilation_unit(
923 class_loader,
924 class_linker,
925 dex_file,
926 code_item,
927 class_def_idx,
928 method_idx,
929 access_flags,
930 /* verified_method */ nullptr,
931 dex_cache);
932
933 HGraph* graph = new (arena) HGraph(
934 arena,
935 dex_file,
936 method_idx,
937 compiler_driver->GetInstructionSet(),
938 kInvalidInvokeType,
939 compiler_driver->GetCompilerOptions().GetDebuggable(),
940 osr);
941
942 const uint8_t* interpreter_metadata = nullptr;
943 if (method == nullptr) {
944 ScopedObjectAccess soa(Thread::Current());
945 method = compiler_driver->ResolveMethod(
946 soa, dex_cache, class_loader, &dex_compilation_unit, method_idx, invoke_type);
947 }
948 // For AOT compilation, we may not get a method, for example if its class is erroneous.
949 // JIT should always have a method.
950 DCHECK(Runtime::Current()->IsAotCompiler() || method != nullptr);
951 if (method != nullptr) {
952 graph->SetArtMethod(method);
953 ScopedObjectAccess soa(Thread::Current());
954 interpreter_metadata = method->GetQuickenedInfo(class_linker->GetImagePointerSize());
955 }
956
957 std::unique_ptr<CodeGenerator> codegen(
958 CodeGenerator::Create(graph,
959 instruction_set,
960 *compiler_driver->GetInstructionSetFeatures(),
961 compiler_driver->GetCompilerOptions(),
962 compilation_stats_.get()));
963 if (codegen.get() == nullptr) {
964 MaybeRecordStat(MethodCompilationStat::kNotCompiledNoCodegen);
965 return nullptr;
966 }
967 codegen->GetAssembler()->cfi().SetEnabled(
968 compiler_driver->GetCompilerOptions().GenerateAnyDebugInfo());
969
970 PassObserver pass_observer(graph,
971 codegen.get(),
972 visualizer_output_.get(),
973 compiler_driver,
974 dump_mutex_);
975
976 {
977 VLOG(compiler) << "Building " << pass_observer.GetMethodName();
978 PassScope scope(HGraphBuilder::kBuilderPassName, &pass_observer);
979 HGraphBuilder builder(graph,
980 &dex_compilation_unit,
981 &dex_compilation_unit,
982 &dex_file,
983 *code_item,
984 compiler_driver,
985 codegen.get(),
986 compilation_stats_.get(),
987 interpreter_metadata,
988 dex_cache,
989 handles);
990 GraphAnalysisResult result = builder.BuildGraph();
991 if (result != kAnalysisSuccess) {
992 switch (result) {
993 case kAnalysisSkipped:
994 MaybeRecordStat(MethodCompilationStat::kNotCompiledSkipped);
995 break;
996 case kAnalysisInvalidBytecode:
997 MaybeRecordStat(MethodCompilationStat::kNotCompiledInvalidBytecode);
998 break;
999 case kAnalysisFailThrowCatchLoop:
1000 MaybeRecordStat(MethodCompilationStat::kNotCompiledThrowCatchLoop);
1001 break;
1002 case kAnalysisFailAmbiguousArrayOp:
1003 MaybeRecordStat(MethodCompilationStat::kNotCompiledAmbiguousArrayOp);
1004 break;
1005 case kAnalysisSuccess:
1006 UNREACHABLE();
1007 }
1008 pass_observer.SetGraphInBadState();
1009 return nullptr;
1010 }
1011 }
1012
1013 RunOptimizations(graph,
1014 codegen.get(),
1015 compiler_driver,
1016 dex_compilation_unit,
1017 &pass_observer,
1018 handles);
1019
1020 RegisterAllocator::Strategy regalloc_strategy =
1021 compiler_options.GetRegisterAllocationStrategy();
1022 AllocateRegisters(graph, codegen.get(), &pass_observer, regalloc_strategy);
1023
1024 codegen->Compile(code_allocator);
1025 pass_observer.DumpDisassembly();
1026
1027 return codegen.release();
1028 }
1029
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) const1030 CompiledMethod* OptimizingCompiler::Compile(const DexFile::CodeItem* code_item,
1031 uint32_t access_flags,
1032 InvokeType invoke_type,
1033 uint16_t class_def_idx,
1034 uint32_t method_idx,
1035 Handle<mirror::ClassLoader> jclass_loader,
1036 const DexFile& dex_file,
1037 Handle<mirror::DexCache> dex_cache) const {
1038 CompilerDriver* compiler_driver = GetCompilerDriver();
1039 CompiledMethod* method = nullptr;
1040 DCHECK(Runtime::Current()->IsAotCompiler());
1041 const VerifiedMethod* verified_method = compiler_driver->GetVerifiedMethod(&dex_file, method_idx);
1042 DCHECK(!verified_method->HasRuntimeThrow());
1043 if (compiler_driver->IsMethodVerifiedWithoutFailures(method_idx, class_def_idx, dex_file)
1044 || verifier::CanCompilerHandleVerificationFailure(
1045 verified_method->GetEncounteredVerificationFailures())) {
1046 ArenaAllocator arena(Runtime::Current()->GetArenaPool());
1047 CodeVectorAllocator code_allocator(&arena);
1048 std::unique_ptr<CodeGenerator> codegen;
1049 {
1050 ScopedObjectAccess soa(Thread::Current());
1051 VariableSizedHandleScope handles(soa.Self());
1052 // Go to native so that we don't block GC during compilation.
1053 ScopedThreadSuspension sts(soa.Self(), kNative);
1054 codegen.reset(
1055 TryCompile(&arena,
1056 &code_allocator,
1057 code_item,
1058 access_flags,
1059 invoke_type,
1060 class_def_idx,
1061 method_idx,
1062 jclass_loader,
1063 dex_file,
1064 dex_cache,
1065 nullptr,
1066 /* osr */ false,
1067 &handles));
1068 }
1069 if (codegen.get() != nullptr) {
1070 MaybeRecordStat(MethodCompilationStat::kCompiled);
1071 method = Emit(&arena, &code_allocator, codegen.get(), compiler_driver, code_item);
1072
1073 if (kArenaAllocatorCountAllocations) {
1074 if (arena.BytesAllocated() > kArenaAllocatorMemoryReportThreshold) {
1075 MemStats mem_stats(arena.GetMemStats());
1076 LOG(INFO) << dex_file.PrettyMethod(method_idx) << " " << Dumpable<MemStats>(mem_stats);
1077 }
1078 }
1079 }
1080 } else {
1081 if (compiler_driver->GetCompilerOptions().VerifyAtRuntime()) {
1082 MaybeRecordStat(MethodCompilationStat::kNotCompiledVerifyAtRuntime);
1083 } else {
1084 MaybeRecordStat(MethodCompilationStat::kNotCompiledVerificationError);
1085 }
1086 }
1087
1088 if (kIsDebugBuild &&
1089 IsCompilingWithCoreImage() &&
1090 IsInstructionSetSupported(compiler_driver->GetInstructionSet())) {
1091 // For testing purposes, we put a special marker on method names
1092 // that should be compiled with this compiler (when the
1093 // instruction set is supported). This makes sure we're not
1094 // regressing.
1095 std::string method_name = dex_file.PrettyMethod(method_idx);
1096 bool shouldCompile = method_name.find("$opt$") != std::string::npos;
1097 DCHECK((method != nullptr) || !shouldCompile) << "Didn't compile " << method_name;
1098 }
1099
1100 return method;
1101 }
1102
CreateOptimizingCompiler(CompilerDriver * driver)1103 Compiler* CreateOptimizingCompiler(CompilerDriver* driver) {
1104 return new OptimizingCompiler(driver);
1105 }
1106
IsCompilingWithCoreImage()1107 bool IsCompilingWithCoreImage() {
1108 const std::string& image = Runtime::Current()->GetImageLocation();
1109 // TODO: This is under-approximating...
1110 if (android::base::EndsWith(image, "core.art") ||
1111 android::base::EndsWith(image, "core-optimizing.art")) {
1112 return true;
1113 }
1114 return false;
1115 }
1116
EncodeArtMethodInInlineInfo(ArtMethod * method ATTRIBUTE_UNUSED)1117 bool EncodeArtMethodInInlineInfo(ArtMethod* method ATTRIBUTE_UNUSED) {
1118 // Note: the runtime is null only for unit testing.
1119 return Runtime::Current() == nullptr || !Runtime::Current()->IsAotCompiler();
1120 }
1121
CanEncodeInlinedMethodInStackMap(const DexFile & caller_dex_file,ArtMethod * callee)1122 bool CanEncodeInlinedMethodInStackMap(const DexFile& caller_dex_file, ArtMethod* callee) {
1123 if (!Runtime::Current()->IsAotCompiler()) {
1124 // JIT can always encode methods in stack maps.
1125 return true;
1126 }
1127 if (IsSameDexFile(caller_dex_file, *callee->GetDexFile())) {
1128 return true;
1129 }
1130 // TODO(ngeoffray): Support more AOT cases for inlining:
1131 // - methods in multidex
1132 // - methods in boot image for on-device non-PIC compilation.
1133 return false;
1134 }
1135
JitCompile(Thread * self,jit::JitCodeCache * code_cache,ArtMethod * method,bool osr)1136 bool OptimizingCompiler::JitCompile(Thread* self,
1137 jit::JitCodeCache* code_cache,
1138 ArtMethod* method,
1139 bool osr) {
1140 StackHandleScope<3> hs(self);
1141 Handle<mirror::ClassLoader> class_loader(hs.NewHandle(
1142 method->GetDeclaringClass()->GetClassLoader()));
1143 Handle<mirror::DexCache> dex_cache(hs.NewHandle(method->GetDexCache()));
1144 DCHECK(method->IsCompilable());
1145
1146 const DexFile* dex_file = method->GetDexFile();
1147 const uint16_t class_def_idx = method->GetClassDefIndex();
1148 const DexFile::CodeItem* code_item = dex_file->GetCodeItem(method->GetCodeItemOffset());
1149 const uint32_t method_idx = method->GetDexMethodIndex();
1150 const uint32_t access_flags = method->GetAccessFlags();
1151 const InvokeType invoke_type = method->GetInvokeType();
1152
1153 ArenaAllocator arena(Runtime::Current()->GetJitArenaPool());
1154 CodeVectorAllocator code_allocator(&arena);
1155 VariableSizedHandleScope handles(self);
1156
1157 std::unique_ptr<CodeGenerator> codegen;
1158 {
1159 // Go to native so that we don't block GC during compilation.
1160 ScopedThreadSuspension sts(self, kNative);
1161 codegen.reset(
1162 TryCompile(&arena,
1163 &code_allocator,
1164 code_item,
1165 access_flags,
1166 invoke_type,
1167 class_def_idx,
1168 method_idx,
1169 class_loader,
1170 *dex_file,
1171 dex_cache,
1172 method,
1173 osr,
1174 &handles));
1175 if (codegen.get() == nullptr) {
1176 return false;
1177 }
1178
1179 if (kArenaAllocatorCountAllocations) {
1180 if (arena.BytesAllocated() > kArenaAllocatorMemoryReportThreshold) {
1181 MemStats mem_stats(arena.GetMemStats());
1182 LOG(INFO) << dex_file->PrettyMethod(method_idx) << " " << Dumpable<MemStats>(mem_stats);
1183 }
1184 }
1185 }
1186
1187 size_t stack_map_size = 0;
1188 size_t method_info_size = 0;
1189 codegen->ComputeStackMapAndMethodInfoSize(&stack_map_size, &method_info_size);
1190 size_t number_of_roots = codegen->GetNumberOfJitRoots();
1191 ClassLinker* class_linker = Runtime::Current()->GetClassLinker();
1192 // We allocate an object array to ensure the JIT roots that we will collect in EmitJitRoots
1193 // will be visible by the GC between EmitLiterals and CommitCode. Once CommitCode is
1194 // executed, this array is not needed.
1195 Handle<mirror::ObjectArray<mirror::Object>> roots(
1196 hs.NewHandle(mirror::ObjectArray<mirror::Object>::Alloc(
1197 self, class_linker->GetClassRoot(ClassLinker::kObjectArrayClass), number_of_roots)));
1198 if (roots == nullptr) {
1199 // Out of memory, just clear the exception to avoid any Java exception uncaught problems.
1200 DCHECK(self->IsExceptionPending());
1201 self->ClearException();
1202 return false;
1203 }
1204 uint8_t* stack_map_data = nullptr;
1205 uint8_t* method_info_data = nullptr;
1206 uint8_t* roots_data = nullptr;
1207 uint32_t data_size = code_cache->ReserveData(self,
1208 stack_map_size,
1209 method_info_size,
1210 number_of_roots,
1211 method,
1212 &stack_map_data,
1213 &method_info_data,
1214 &roots_data);
1215 if (stack_map_data == nullptr || roots_data == nullptr) {
1216 return false;
1217 }
1218 MaybeRecordStat(MethodCompilationStat::kCompiled);
1219 codegen->BuildStackMaps(MemoryRegion(stack_map_data, stack_map_size),
1220 MemoryRegion(method_info_data, method_info_size),
1221 *code_item);
1222 codegen->EmitJitRoots(code_allocator.GetData(), roots, roots_data);
1223
1224 const void* code = code_cache->CommitCode(
1225 self,
1226 method,
1227 stack_map_data,
1228 method_info_data,
1229 roots_data,
1230 codegen->HasEmptyFrame() ? 0 : codegen->GetFrameSize(),
1231 codegen->GetCoreSpillMask(),
1232 codegen->GetFpuSpillMask(),
1233 code_allocator.GetMemory().data(),
1234 code_allocator.GetSize(),
1235 data_size,
1236 osr,
1237 roots,
1238 codegen->GetGraph()->HasShouldDeoptimizeFlag(),
1239 codegen->GetGraph()->GetCHASingleImplementationList());
1240
1241 if (code == nullptr) {
1242 code_cache->ClearData(self, stack_map_data, roots_data);
1243 return false;
1244 }
1245
1246 const CompilerOptions& compiler_options = GetCompilerDriver()->GetCompilerOptions();
1247 if (compiler_options.GetGenerateDebugInfo()) {
1248 const auto* method_header = reinterpret_cast<const OatQuickMethodHeader*>(code);
1249 const uintptr_t code_address = reinterpret_cast<uintptr_t>(method_header->GetCode());
1250 debug::MethodDebugInfo info = debug::MethodDebugInfo();
1251 info.trampoline_name = nullptr;
1252 info.dex_file = dex_file;
1253 info.class_def_index = class_def_idx;
1254 info.dex_method_index = method_idx;
1255 info.access_flags = access_flags;
1256 info.code_item = code_item;
1257 info.isa = codegen->GetInstructionSet();
1258 info.deduped = false;
1259 info.is_native_debuggable = compiler_options.GetNativeDebuggable();
1260 info.is_optimized = true;
1261 info.is_code_address_text_relative = false;
1262 info.code_address = code_address;
1263 info.code_size = code_allocator.GetSize();
1264 info.frame_size_in_bytes = method_header->GetFrameSizeInBytes();
1265 info.code_info = stack_map_size == 0 ? nullptr : stack_map_data;
1266 info.cfi = ArrayRef<const uint8_t>(*codegen->GetAssembler()->cfi().data());
1267 std::vector<uint8_t> elf_file = debug::WriteDebugElfFileForMethods(
1268 GetCompilerDriver()->GetInstructionSet(),
1269 GetCompilerDriver()->GetInstructionSetFeatures(),
1270 ArrayRef<const debug::MethodDebugInfo>(&info, 1));
1271 CreateJITCodeEntryForAddress(code_address, std::move(elf_file));
1272 }
1273
1274 Runtime::Current()->GetJit()->AddMemoryUsage(method, arena.BytesUsed());
1275
1276 return true;
1277 }
1278
1279 } // namespace art
1280