optimizing_compiler.cc revision 0d5a281c671444bfa75d63caf1427a8c0e6e1177
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 <stdint.h> 21 22#ifdef ART_ENABLE_CODEGEN_arm64 23#include "instruction_simplifier_arm64.h" 24#endif 25 26#ifdef ART_ENABLE_CODEGEN_x86 27#include "pc_relative_fixups_x86.h" 28#endif 29 30#include "art_method-inl.h" 31#include "base/arena_allocator.h" 32#include "base/arena_containers.h" 33#include "base/dumpable.h" 34#include "base/macros.h" 35#include "base/timing_logger.h" 36#include "boolean_simplifier.h" 37#include "bounds_check_elimination.h" 38#include "builder.h" 39#include "code_generator.h" 40#include "compiled_method.h" 41#include "compiler.h" 42#include "constant_folding.h" 43#include "dead_code_elimination.h" 44#include "dex/quick/dex_file_to_method_inliner_map.h" 45#include "dex/verified_method.h" 46#include "dex/verification_results.h" 47#include "driver/compiler_driver.h" 48#include "driver/compiler_driver-inl.h" 49#include "driver/compiler_options.h" 50#include "driver/dex_compilation_unit.h" 51#include "elf_writer_quick.h" 52#include "graph_checker.h" 53#include "graph_visualizer.h" 54#include "gvn.h" 55#include "induction_var_analysis.h" 56#include "inliner.h" 57#include "instruction_simplifier.h" 58#include "intrinsics.h" 59#include "jit/jit_code_cache.h" 60#include "licm.h" 61#include "jni/quick/jni_compiler.h" 62#include "load_store_elimination.h" 63#include "nodes.h" 64#include "prepare_for_register_allocation.h" 65#include "reference_type_propagation.h" 66#include "register_allocator.h" 67#include "sharpening.h" 68#include "side_effects_analysis.h" 69#include "ssa_builder.h" 70#include "ssa_phi_elimination.h" 71#include "ssa_liveness_analysis.h" 72#include "utils/assembler.h" 73#include "verifier/method_verifier.h" 74 75namespace art { 76 77/** 78 * Used by the code generator, to allocate the code in a vector. 79 */ 80class CodeVectorAllocator FINAL : public CodeAllocator { 81 public: 82 explicit CodeVectorAllocator(ArenaAllocator* arena) 83 : memory_(arena->Adapter(kArenaAllocCodeBuffer)), 84 size_(0) {} 85 86 virtual uint8_t* Allocate(size_t size) { 87 size_ = size; 88 memory_.resize(size); 89 return &memory_[0]; 90 } 91 92 size_t GetSize() const { return size_; } 93 const ArenaVector<uint8_t>& GetMemory() const { return memory_; } 94 95 private: 96 ArenaVector<uint8_t> memory_; 97 size_t size_; 98 99 DISALLOW_COPY_AND_ASSIGN(CodeVectorAllocator); 100}; 101 102/** 103 * Filter to apply to the visualizer. Methods whose name contain that filter will 104 * be dumped. 105 */ 106static constexpr const char kStringFilter[] = ""; 107 108class PassScope; 109 110class PassObserver : public ValueObject { 111 public: 112 PassObserver(HGraph* graph, 113 const char* method_name, 114 CodeGenerator* codegen, 115 std::ostream* visualizer_output, 116 CompilerDriver* compiler_driver) 117 : graph_(graph), 118 method_name_(method_name), 119 timing_logger_enabled_(compiler_driver->GetDumpPasses()), 120 timing_logger_(method_name, true, true), 121 disasm_info_(graph->GetArena()), 122 visualizer_enabled_(!compiler_driver->GetDumpCfgFileName().empty()), 123 visualizer_(visualizer_output, graph, *codegen), 124 graph_in_bad_state_(false) { 125 if (timing_logger_enabled_ || visualizer_enabled_) { 126 if (!IsVerboseMethod(compiler_driver, method_name)) { 127 timing_logger_enabled_ = visualizer_enabled_ = false; 128 } 129 if (visualizer_enabled_) { 130 visualizer_.PrintHeader(method_name_); 131 codegen->SetDisassemblyInformation(&disasm_info_); 132 } 133 } 134 } 135 136 ~PassObserver() { 137 if (timing_logger_enabled_) { 138 LOG(INFO) << "TIMINGS " << method_name_; 139 LOG(INFO) << Dumpable<TimingLogger>(timing_logger_); 140 } 141 } 142 143 void DumpDisassembly() const { 144 if (visualizer_enabled_) { 145 visualizer_.DumpGraphWithDisassembly(); 146 } 147 } 148 149 void SetGraphInBadState() { graph_in_bad_state_ = true; } 150 151 private: 152 void StartPass(const char* pass_name) { 153 // Dump graph first, then start timer. 154 if (visualizer_enabled_) { 155 visualizer_.DumpGraph(pass_name, /* is_after_pass */ false, graph_in_bad_state_); 156 } 157 if (timing_logger_enabled_) { 158 timing_logger_.StartTiming(pass_name); 159 } 160 } 161 162 void EndPass(const char* pass_name) { 163 // Pause timer first, then dump graph. 164 if (timing_logger_enabled_) { 165 timing_logger_.EndTiming(); 166 } 167 if (visualizer_enabled_) { 168 visualizer_.DumpGraph(pass_name, /* is_after_pass */ true, graph_in_bad_state_); 169 } 170 171 // Validate the HGraph if running in debug mode. 172 if (kIsDebugBuild) { 173 if (!graph_in_bad_state_) { 174 if (graph_->IsInSsaForm()) { 175 SSAChecker checker(graph_); 176 checker.Run(); 177 if (!checker.IsValid()) { 178 LOG(FATAL) << "Error after " << pass_name << ": " << Dumpable<SSAChecker>(checker); 179 } 180 } else { 181 GraphChecker checker(graph_); 182 checker.Run(); 183 if (!checker.IsValid()) { 184 LOG(FATAL) << "Error after " << pass_name << ": " << Dumpable<GraphChecker>(checker); 185 } 186 } 187 } 188 } 189 } 190 191 static bool IsVerboseMethod(CompilerDriver* compiler_driver, const char* method_name) { 192 // Test an exact match to --verbose-methods. If verbose-methods is set, this overrides an 193 // empty kStringFilter matching all methods. 194 if (compiler_driver->GetCompilerOptions().HasVerboseMethods()) { 195 return compiler_driver->GetCompilerOptions().IsVerboseMethod(method_name); 196 } 197 198 // Test the kStringFilter sub-string. constexpr helper variable to silence unreachable-code 199 // warning when the string is empty. 200 constexpr bool kStringFilterEmpty = arraysize(kStringFilter) <= 1; 201 if (kStringFilterEmpty || strstr(method_name, kStringFilter) != nullptr) { 202 return true; 203 } 204 205 return false; 206 } 207 208 HGraph* const graph_; 209 const char* method_name_; 210 211 bool timing_logger_enabled_; 212 TimingLogger timing_logger_; 213 214 DisassemblyInformation disasm_info_; 215 216 bool visualizer_enabled_; 217 HGraphVisualizer visualizer_; 218 219 // Flag to be set by the compiler if the pass failed and the graph is not 220 // expected to validate. 221 bool graph_in_bad_state_; 222 223 friend PassScope; 224 225 DISALLOW_COPY_AND_ASSIGN(PassObserver); 226}; 227 228class PassScope : public ValueObject { 229 public: 230 PassScope(const char *pass_name, PassObserver* pass_observer) 231 : pass_name_(pass_name), 232 pass_observer_(pass_observer) { 233 pass_observer_->StartPass(pass_name_); 234 } 235 236 ~PassScope() { 237 pass_observer_->EndPass(pass_name_); 238 } 239 240 private: 241 const char* const pass_name_; 242 PassObserver* const pass_observer_; 243}; 244 245class OptimizingCompiler FINAL : public Compiler { 246 public: 247 explicit OptimizingCompiler(CompilerDriver* driver); 248 ~OptimizingCompiler(); 249 250 bool CanCompileMethod(uint32_t method_idx, const DexFile& dex_file, CompilationUnit* cu) const 251 OVERRIDE; 252 253 CompiledMethod* Compile(const DexFile::CodeItem* code_item, 254 uint32_t access_flags, 255 InvokeType invoke_type, 256 uint16_t class_def_idx, 257 uint32_t method_idx, 258 jobject class_loader, 259 const DexFile& dex_file, 260 Handle<mirror::DexCache> dex_cache) const OVERRIDE; 261 262 CompiledMethod* JniCompile(uint32_t access_flags, 263 uint32_t method_idx, 264 const DexFile& dex_file) const OVERRIDE { 265 return ArtQuickJniCompileMethod(GetCompilerDriver(), access_flags, method_idx, dex_file); 266 } 267 268 uintptr_t GetEntryPointOf(ArtMethod* method) const OVERRIDE 269 SHARED_REQUIRES(Locks::mutator_lock_) { 270 return reinterpret_cast<uintptr_t>(method->GetEntryPointFromQuickCompiledCodePtrSize( 271 InstructionSetPointerSize(GetCompilerDriver()->GetInstructionSet()))); 272 } 273 274 void InitCompilationUnit(CompilationUnit& cu) const OVERRIDE; 275 276 void Init() OVERRIDE; 277 278 void UnInit() const OVERRIDE; 279 280 void MaybeRecordStat(MethodCompilationStat compilation_stat) const { 281 if (compilation_stats_.get() != nullptr) { 282 compilation_stats_->RecordStat(compilation_stat); 283 } 284 } 285 286 bool JitCompile(Thread* self, jit::JitCodeCache* code_cache, ArtMethod* method) 287 OVERRIDE 288 SHARED_REQUIRES(Locks::mutator_lock_); 289 290 private: 291 // Whether we should run any optimization or register allocation. If false, will 292 // just run the code generation after the graph was built. 293 const bool run_optimizations_; 294 295 // Create a 'CompiledMethod' for an optimized graph. 296 CompiledMethod* EmitOptimized(ArenaAllocator* arena, 297 CodeVectorAllocator* code_allocator, 298 CodeGenerator* codegen, 299 CompilerDriver* driver) const; 300 301 // Create a 'CompiledMethod' for a non-optimized graph. 302 CompiledMethod* EmitBaseline(ArenaAllocator* arena, 303 CodeVectorAllocator* code_allocator, 304 CodeGenerator* codegen, 305 CompilerDriver* driver) const; 306 307 // Try compiling a method and return the code generator used for 308 // compiling it. 309 // This method: 310 // 1) Builds the graph. Returns null if it failed to build it. 311 // 2) If `run_optimizations_` is set: 312 // 2.1) Transform the graph to SSA. Returns null if it failed. 313 // 2.2) Run optimizations on the graph, including register allocator. 314 // 3) Generate code with the `code_allocator` provided. 315 CodeGenerator* TryCompile(ArenaAllocator* arena, 316 CodeVectorAllocator* code_allocator, 317 const DexFile::CodeItem* code_item, 318 uint32_t access_flags, 319 InvokeType invoke_type, 320 uint16_t class_def_idx, 321 uint32_t method_idx, 322 jobject class_loader, 323 const DexFile& dex_file, 324 Handle<mirror::DexCache> dex_cache) const; 325 326 std::unique_ptr<OptimizingCompilerStats> compilation_stats_; 327 328 std::unique_ptr<std::ostream> visualizer_output_; 329 330 DISALLOW_COPY_AND_ASSIGN(OptimizingCompiler); 331}; 332 333static const int kMaximumCompilationTimeBeforeWarning = 100; /* ms */ 334 335OptimizingCompiler::OptimizingCompiler(CompilerDriver* driver) 336 : Compiler(driver, kMaximumCompilationTimeBeforeWarning), 337 run_optimizations_( 338 driver->GetCompilerOptions().GetCompilerFilter() != CompilerOptions::kTime) {} 339 340void OptimizingCompiler::Init() { 341 // Enable C1visualizer output. Must be done in Init() because the compiler 342 // driver is not fully initialized when passed to the compiler's constructor. 343 CompilerDriver* driver = GetCompilerDriver(); 344 const std::string cfg_file_name = driver->GetDumpCfgFileName(); 345 if (!cfg_file_name.empty()) { 346 CHECK_EQ(driver->GetThreadCount(), 1U) 347 << "Graph visualizer requires the compiler to run single-threaded. " 348 << "Invoke the compiler with '-j1'."; 349 std::ios_base::openmode cfg_file_mode = 350 driver->GetDumpCfgAppend() ? std::ofstream::app : std::ofstream::out; 351 visualizer_output_.reset(new std::ofstream(cfg_file_name, cfg_file_mode)); 352 } 353 if (driver->GetDumpStats()) { 354 compilation_stats_.reset(new OptimizingCompilerStats()); 355 } 356} 357 358void OptimizingCompiler::UnInit() const { 359} 360 361OptimizingCompiler::~OptimizingCompiler() { 362 if (compilation_stats_.get() != nullptr) { 363 compilation_stats_->Log(); 364 } 365} 366 367void OptimizingCompiler::InitCompilationUnit(CompilationUnit& cu ATTRIBUTE_UNUSED) const { 368} 369 370bool OptimizingCompiler::CanCompileMethod(uint32_t method_idx ATTRIBUTE_UNUSED, 371 const DexFile& dex_file ATTRIBUTE_UNUSED, 372 CompilationUnit* cu ATTRIBUTE_UNUSED) const { 373 return true; 374} 375 376static bool IsInstructionSetSupported(InstructionSet instruction_set) { 377 return (instruction_set == kArm && !kArm32QuickCodeUseSoftFloat) 378 || instruction_set == kArm64 379 || (instruction_set == kThumb2 && !kArm32QuickCodeUseSoftFloat) 380 || instruction_set == kMips 381 || instruction_set == kMips64 382 || instruction_set == kX86 383 || instruction_set == kX86_64; 384} 385 386// Read barrier are supported only on x86 and x86-64 at the moment. 387// TODO: Add support for other architectures and remove this function 388static bool InstructionSetSupportsReadBarrier(InstructionSet instruction_set) { 389 return instruction_set == kX86 390 || instruction_set == kX86_64; 391} 392 393static void RunOptimizations(HOptimization* optimizations[], 394 size_t length, 395 PassObserver* pass_observer) { 396 for (size_t i = 0; i < length; ++i) { 397 PassScope scope(optimizations[i]->GetPassName(), pass_observer); 398 optimizations[i]->Run(); 399 } 400} 401 402static void MaybeRunInliner(HGraph* graph, 403 CodeGenerator* codegen, 404 CompilerDriver* driver, 405 OptimizingCompilerStats* stats, 406 const DexCompilationUnit& dex_compilation_unit, 407 PassObserver* pass_observer, 408 StackHandleScopeCollection* handles) { 409 const CompilerOptions& compiler_options = driver->GetCompilerOptions(); 410 bool should_inline = (compiler_options.GetInlineDepthLimit() > 0) 411 && (compiler_options.GetInlineMaxCodeUnits() > 0); 412 if (!should_inline) { 413 return; 414 } 415 HInliner* inliner = new (graph->GetArena()) HInliner( 416 graph, codegen, dex_compilation_unit, dex_compilation_unit, driver, handles, stats); 417 HOptimization* optimizations[] = { inliner }; 418 419 RunOptimizations(optimizations, arraysize(optimizations), pass_observer); 420} 421 422static void RunArchOptimizations(InstructionSet instruction_set, 423 HGraph* graph, 424 OptimizingCompilerStats* stats, 425 PassObserver* pass_observer) { 426 ArenaAllocator* arena = graph->GetArena(); 427 switch (instruction_set) { 428#ifdef ART_ENABLE_CODEGEN_arm64 429 case kArm64: { 430 arm64::InstructionSimplifierArm64* simplifier = 431 new (arena) arm64::InstructionSimplifierArm64(graph, stats); 432 SideEffectsAnalysis* side_effects = new (arena) SideEffectsAnalysis(graph); 433 GVNOptimization* gvn = new (arena) GVNOptimization(graph, *side_effects, "GVN_after_arch"); 434 HOptimization* arm64_optimizations[] = { 435 simplifier, 436 side_effects, 437 gvn 438 }; 439 RunOptimizations(arm64_optimizations, arraysize(arm64_optimizations), pass_observer); 440 break; 441 } 442#endif 443#ifdef ART_ENABLE_CODEGEN_x86 444 case kX86: { 445 x86::PcRelativeFixups* pc_relative_fixups = new (arena) x86::PcRelativeFixups(graph, stats); 446 HOptimization* x86_optimizations[] = { 447 pc_relative_fixups 448 }; 449 RunOptimizations(x86_optimizations, arraysize(x86_optimizations), pass_observer); 450 break; 451 } 452#endif 453 default: 454 break; 455 } 456} 457 458NO_INLINE // Avoid increasing caller's frame size by large stack-allocated objects. 459static void AllocateRegisters(HGraph* graph, 460 CodeGenerator* codegen, 461 PassObserver* pass_observer) { 462 PrepareForRegisterAllocation(graph).Run(); 463 SsaLivenessAnalysis liveness(graph, codegen); 464 { 465 PassScope scope(SsaLivenessAnalysis::kLivenessPassName, pass_observer); 466 liveness.Analyze(); 467 } 468 { 469 PassScope scope(RegisterAllocator::kRegisterAllocatorPassName, pass_observer); 470 RegisterAllocator(graph->GetArena(), codegen, liveness).AllocateRegisters(); 471 } 472} 473 474static void RunOptimizations(HGraph* graph, 475 CodeGenerator* codegen, 476 CompilerDriver* driver, 477 OptimizingCompilerStats* stats, 478 const DexCompilationUnit& dex_compilation_unit, 479 PassObserver* pass_observer) { 480 ScopedObjectAccess soa(Thread::Current()); 481 StackHandleScopeCollection handles(soa.Self()); 482 ScopedThreadSuspension sts(soa.Self(), kNative); 483 484 ArenaAllocator* arena = graph->GetArena(); 485 HDeadCodeElimination* dce1 = new (arena) HDeadCodeElimination( 486 graph, stats, HDeadCodeElimination::kInitialDeadCodeEliminationPassName); 487 HDeadCodeElimination* dce2 = new (arena) HDeadCodeElimination( 488 graph, stats, HDeadCodeElimination::kFinalDeadCodeEliminationPassName); 489 HConstantFolding* fold1 = new (arena) HConstantFolding(graph); 490 InstructionSimplifier* simplify1 = new (arena) InstructionSimplifier(graph, stats); 491 HBooleanSimplifier* boolean_simplify = new (arena) HBooleanSimplifier(graph); 492 HConstantFolding* fold2 = new (arena) HConstantFolding(graph, "constant_folding_after_inlining"); 493 SideEffectsAnalysis* side_effects = new (arena) SideEffectsAnalysis(graph); 494 GVNOptimization* gvn = new (arena) GVNOptimization(graph, *side_effects); 495 LICM* licm = new (arena) LICM(graph, *side_effects); 496 LoadStoreElimination* lse = new (arena) LoadStoreElimination(graph, *side_effects); 497 HInductionVarAnalysis* induction = new (arena) HInductionVarAnalysis(graph); 498 BoundsCheckElimination* bce = new (arena) BoundsCheckElimination(graph, induction); 499 ReferenceTypePropagation* type_propagation = 500 new (arena) ReferenceTypePropagation(graph, &handles); 501 HSharpening* sharpening = new (arena) HSharpening(graph, codegen, dex_compilation_unit, driver); 502 InstructionSimplifier* simplify2 = new (arena) InstructionSimplifier( 503 graph, stats, "instruction_simplifier_after_types"); 504 InstructionSimplifier* simplify3 = new (arena) InstructionSimplifier( 505 graph, stats, "instruction_simplifier_after_bce"); 506 InstructionSimplifier* simplify4 = new (arena) InstructionSimplifier( 507 graph, stats, "instruction_simplifier_before_codegen"); 508 509 IntrinsicsRecognizer* intrinsics = new (arena) IntrinsicsRecognizer(graph, driver); 510 511 HOptimization* optimizations1[] = { 512 intrinsics, 513 fold1, 514 simplify1, 515 type_propagation, 516 sharpening, 517 dce1, 518 simplify2 519 }; 520 521 RunOptimizations(optimizations1, arraysize(optimizations1), pass_observer); 522 523 MaybeRunInliner(graph, codegen, driver, stats, dex_compilation_unit, pass_observer, &handles); 524 525 // TODO: Update passes incompatible with try/catch so we have the same 526 // pipeline for all methods. 527 if (graph->HasTryCatch()) { 528 HOptimization* optimizations2[] = { 529 boolean_simplify, 530 side_effects, 531 gvn, 532 dce2, 533 // The codegen has a few assumptions that only the instruction simplifier 534 // can satisfy. For example, the code generator does not expect to see a 535 // HTypeConversion from a type to the same type. 536 simplify4, 537 }; 538 539 RunOptimizations(optimizations2, arraysize(optimizations2), pass_observer); 540 } else { 541 HOptimization* optimizations2[] = { 542 // BooleanSimplifier depends on the InstructionSimplifier removing 543 // redundant suspend checks to recognize empty blocks. 544 boolean_simplify, 545 fold2, // TODO: if we don't inline we can also skip fold2. 546 side_effects, 547 gvn, 548 licm, 549 induction, 550 bce, 551 simplify3, 552 lse, 553 dce2, 554 // The codegen has a few assumptions that only the instruction simplifier 555 // can satisfy. For example, the code generator does not expect to see a 556 // HTypeConversion from a type to the same type. 557 simplify4, 558 }; 559 560 RunOptimizations(optimizations2, arraysize(optimizations2), pass_observer); 561 } 562 563 RunArchOptimizations(driver->GetInstructionSet(), graph, stats, pass_observer); 564 AllocateRegisters(graph, codegen, pass_observer); 565} 566 567// The stack map we generate must be 4-byte aligned on ARM. Since existing 568// maps are generated alongside these stack maps, we must also align them. 569static ArrayRef<const uint8_t> AlignVectorSize(ArenaVector<uint8_t>& vector) { 570 size_t size = vector.size(); 571 size_t aligned_size = RoundUp(size, 4); 572 for (; size < aligned_size; ++size) { 573 vector.push_back(0); 574 } 575 return ArrayRef<const uint8_t>(vector); 576} 577 578static ArenaVector<LinkerPatch> EmitAndSortLinkerPatches(CodeGenerator* codegen) { 579 ArenaVector<LinkerPatch> linker_patches(codegen->GetGraph()->GetArena()->Adapter()); 580 codegen->EmitLinkerPatches(&linker_patches); 581 582 // Sort patches by literal offset. Required for .oat_patches encoding. 583 std::sort(linker_patches.begin(), linker_patches.end(), 584 [](const LinkerPatch& lhs, const LinkerPatch& rhs) { 585 return lhs.LiteralOffset() < rhs.LiteralOffset(); 586 }); 587 588 return linker_patches; 589} 590 591CompiledMethod* OptimizingCompiler::EmitOptimized(ArenaAllocator* arena, 592 CodeVectorAllocator* code_allocator, 593 CodeGenerator* codegen, 594 CompilerDriver* compiler_driver) const { 595 ArenaVector<LinkerPatch> linker_patches = EmitAndSortLinkerPatches(codegen); 596 ArenaVector<uint8_t> stack_map(arena->Adapter(kArenaAllocStackMaps)); 597 stack_map.resize(codegen->ComputeStackMapsSize()); 598 codegen->BuildStackMaps(MemoryRegion(stack_map.data(), stack_map.size())); 599 600 MaybeRecordStat(MethodCompilationStat::kCompiledOptimized); 601 602 CompiledMethod* compiled_method = CompiledMethod::SwapAllocCompiledMethod( 603 compiler_driver, 604 codegen->GetInstructionSet(), 605 ArrayRef<const uint8_t>(code_allocator->GetMemory()), 606 // Follow Quick's behavior and set the frame size to zero if it is 607 // considered "empty" (see the definition of 608 // art::CodeGenerator::HasEmptyFrame). 609 codegen->HasEmptyFrame() ? 0 : codegen->GetFrameSize(), 610 codegen->GetCoreSpillMask(), 611 codegen->GetFpuSpillMask(), 612 ArrayRef<const SrcMapElem>(codegen->GetSrcMappingTable()), 613 ArrayRef<const uint8_t>(), // mapping_table. 614 ArrayRef<const uint8_t>(stack_map), 615 ArrayRef<const uint8_t>(), // native_gc_map. 616 ArrayRef<const uint8_t>(*codegen->GetAssembler()->cfi().data()), 617 ArrayRef<const LinkerPatch>(linker_patches)); 618 619 return compiled_method; 620} 621 622CompiledMethod* OptimizingCompiler::EmitBaseline( 623 ArenaAllocator* arena, 624 CodeVectorAllocator* code_allocator, 625 CodeGenerator* codegen, 626 CompilerDriver* compiler_driver) const { 627 ArenaVector<LinkerPatch> linker_patches = EmitAndSortLinkerPatches(codegen); 628 629 ArenaVector<uint8_t> mapping_table(arena->Adapter(kArenaAllocBaselineMaps)); 630 codegen->BuildMappingTable(&mapping_table); 631 ArenaVector<uint8_t> vmap_table(arena->Adapter(kArenaAllocBaselineMaps)); 632 codegen->BuildVMapTable(&vmap_table); 633 ArenaVector<uint8_t> gc_map(arena->Adapter(kArenaAllocBaselineMaps)); 634 codegen->BuildNativeGCMap(&gc_map, *compiler_driver); 635 636 MaybeRecordStat(MethodCompilationStat::kCompiledBaseline); 637 CompiledMethod* compiled_method = CompiledMethod::SwapAllocCompiledMethod( 638 compiler_driver, 639 codegen->GetInstructionSet(), 640 ArrayRef<const uint8_t>(code_allocator->GetMemory()), 641 // Follow Quick's behavior and set the frame size to zero if it is 642 // considered "empty" (see the definition of 643 // art::CodeGenerator::HasEmptyFrame). 644 codegen->HasEmptyFrame() ? 0 : codegen->GetFrameSize(), 645 codegen->GetCoreSpillMask(), 646 codegen->GetFpuSpillMask(), 647 ArrayRef<const SrcMapElem>(codegen->GetSrcMappingTable()), 648 AlignVectorSize(mapping_table), 649 AlignVectorSize(vmap_table), 650 AlignVectorSize(gc_map), 651 ArrayRef<const uint8_t>(*codegen->GetAssembler()->cfi().data()), 652 ArrayRef<const LinkerPatch>(linker_patches)); 653 return compiled_method; 654} 655 656CodeGenerator* OptimizingCompiler::TryCompile(ArenaAllocator* arena, 657 CodeVectorAllocator* code_allocator, 658 const DexFile::CodeItem* code_item, 659 uint32_t access_flags, 660 InvokeType invoke_type, 661 uint16_t class_def_idx, 662 uint32_t method_idx, 663 jobject class_loader, 664 const DexFile& dex_file, 665 Handle<mirror::DexCache> dex_cache) const { 666 std::string method_name = PrettyMethod(method_idx, dex_file); 667 MaybeRecordStat(MethodCompilationStat::kAttemptCompilation); 668 CompilerDriver* compiler_driver = GetCompilerDriver(); 669 InstructionSet instruction_set = compiler_driver->GetInstructionSet(); 670 671 // Always use the thumb2 assembler: some runtime functionality (like implicit stack 672 // overflow checks) assume thumb2. 673 if (instruction_set == kArm) { 674 instruction_set = kThumb2; 675 } 676 677 // Do not attempt to compile on architectures we do not support. 678 if (!IsInstructionSetSupported(instruction_set)) { 679 MaybeRecordStat(MethodCompilationStat::kNotCompiledUnsupportedIsa); 680 return nullptr; 681 } 682 683 // When read barriers are enabled, do not attempt to compile for 684 // instruction sets that have no read barrier support. 685 if (kEmitCompilerReadBarrier && !InstructionSetSupportsReadBarrier(instruction_set)) { 686 return nullptr; 687 } 688 689 if (Compiler::IsPathologicalCase(*code_item, method_idx, dex_file)) { 690 MaybeRecordStat(MethodCompilationStat::kNotCompiledPathological); 691 return nullptr; 692 } 693 694 // Implementation of the space filter: do not compile a code item whose size in 695 // code units is bigger than 128. 696 static constexpr size_t kSpaceFilterOptimizingThreshold = 128; 697 const CompilerOptions& compiler_options = compiler_driver->GetCompilerOptions(); 698 if ((compiler_options.GetCompilerFilter() == CompilerOptions::kSpace) 699 && (code_item->insns_size_in_code_units_ > kSpaceFilterOptimizingThreshold)) { 700 MaybeRecordStat(MethodCompilationStat::kNotCompiledSpaceFilter); 701 return nullptr; 702 } 703 704 DexCompilationUnit dex_compilation_unit( 705 nullptr, class_loader, Runtime::Current()->GetClassLinker(), dex_file, code_item, 706 class_def_idx, method_idx, access_flags, 707 compiler_driver->GetVerifiedMethod(&dex_file, method_idx), dex_cache); 708 709 bool requires_barrier = dex_compilation_unit.IsConstructor() 710 && compiler_driver->RequiresConstructorBarrier(Thread::Current(), 711 dex_compilation_unit.GetDexFile(), 712 dex_compilation_unit.GetClassDefIndex()); 713 HGraph* graph = new (arena) HGraph( 714 arena, dex_file, method_idx, requires_barrier, compiler_driver->GetInstructionSet(), 715 kInvalidInvokeType, compiler_driver->GetCompilerOptions().GetDebuggable()); 716 717 std::unique_ptr<CodeGenerator> codegen( 718 CodeGenerator::Create(graph, 719 instruction_set, 720 *compiler_driver->GetInstructionSetFeatures(), 721 compiler_driver->GetCompilerOptions())); 722 if (codegen.get() == nullptr) { 723 MaybeRecordStat(MethodCompilationStat::kNotCompiledNoCodegen); 724 return nullptr; 725 } 726 codegen->GetAssembler()->cfi().SetEnabled( 727 compiler_driver->GetCompilerOptions().GetGenerateDebugInfo()); 728 729 PassObserver pass_observer(graph, 730 method_name.c_str(), 731 codegen.get(), 732 visualizer_output_.get(), 733 compiler_driver); 734 735 const uint8_t* interpreter_metadata = nullptr; 736 { 737 ScopedObjectAccess soa(Thread::Current()); 738 StackHandleScope<1> hs(soa.Self()); 739 Handle<mirror::ClassLoader> loader(hs.NewHandle( 740 soa.Decode<mirror::ClassLoader*>(class_loader))); 741 ArtMethod* art_method = compiler_driver->ResolveMethod( 742 soa, dex_cache, loader, &dex_compilation_unit, method_idx, invoke_type); 743 // We may not get a method, for example if its class is erroneous. 744 // TODO: Clean this up, the compiler driver should just pass the ArtMethod to compile. 745 if (art_method != nullptr) { 746 interpreter_metadata = art_method->GetQuickenedInfo(); 747 } 748 } 749 HGraphBuilder builder(graph, 750 &dex_compilation_unit, 751 &dex_compilation_unit, 752 &dex_file, 753 compiler_driver, 754 compilation_stats_.get(), 755 interpreter_metadata, 756 dex_cache); 757 758 VLOG(compiler) << "Building " << method_name; 759 760 { 761 PassScope scope(HGraphBuilder::kBuilderPassName, &pass_observer); 762 if (!builder.BuildGraph(*code_item)) { 763 pass_observer.SetGraphInBadState(); 764 return nullptr; 765 } 766 } 767 768 VLOG(compiler) << "Optimizing " << method_name; 769 if (run_optimizations_) { 770 { 771 PassScope scope(SsaBuilder::kSsaBuilderPassName, &pass_observer); 772 if (!graph->TryBuildingSsa()) { 773 // We could not transform the graph to SSA, bailout. 774 LOG(INFO) << "Skipping compilation of " << method_name << ": it contains a non natural loop"; 775 MaybeRecordStat(MethodCompilationStat::kNotCompiledCannotBuildSSA); 776 pass_observer.SetGraphInBadState(); 777 return nullptr; 778 } 779 } 780 781 RunOptimizations(graph, 782 codegen.get(), 783 compiler_driver, 784 compilation_stats_.get(), 785 dex_compilation_unit, 786 &pass_observer); 787 codegen->CompileOptimized(code_allocator); 788 } else { 789 codegen->CompileBaseline(code_allocator); 790 } 791 pass_observer.DumpDisassembly(); 792 793 if (kArenaAllocatorCountAllocations) { 794 if (arena->BytesAllocated() > 4 * MB) { 795 MemStats mem_stats(arena->GetMemStats()); 796 LOG(INFO) << PrettyMethod(method_idx, dex_file) << " " << Dumpable<MemStats>(mem_stats); 797 } 798 } 799 800 return codegen.release(); 801} 802 803static bool CanHandleVerificationFailure(const VerifiedMethod* verified_method) { 804 // For access errors the compiler will use the unresolved helpers (e.g. HInvokeUnresolved). 805 uint32_t unresolved_mask = verifier::VerifyError::VERIFY_ERROR_NO_CLASS 806 | verifier::VerifyError::VERIFY_ERROR_ACCESS_CLASS 807 | verifier::VerifyError::VERIFY_ERROR_ACCESS_FIELD 808 | verifier::VerifyError::VERIFY_ERROR_ACCESS_METHOD; 809 return (verified_method->GetEncounteredVerificationFailures() & (~unresolved_mask)) == 0; 810} 811 812CompiledMethod* OptimizingCompiler::Compile(const DexFile::CodeItem* code_item, 813 uint32_t access_flags, 814 InvokeType invoke_type, 815 uint16_t class_def_idx, 816 uint32_t method_idx, 817 jobject jclass_loader, 818 const DexFile& dex_file, 819 Handle<mirror::DexCache> dex_cache) const { 820 CompilerDriver* compiler_driver = GetCompilerDriver(); 821 CompiledMethod* method = nullptr; 822 DCHECK(Runtime::Current()->IsAotCompiler()); 823 const VerifiedMethod* verified_method = compiler_driver->GetVerifiedMethod(&dex_file, method_idx); 824 DCHECK(!verified_method->HasRuntimeThrow()); 825 if (compiler_driver->IsMethodVerifiedWithoutFailures(method_idx, class_def_idx, dex_file) 826 || CanHandleVerificationFailure(verified_method)) { 827 ArenaAllocator arena(Runtime::Current()->GetArenaPool()); 828 CodeVectorAllocator code_allocator(&arena); 829 std::unique_ptr<CodeGenerator> codegen( 830 TryCompile(&arena, 831 &code_allocator, 832 code_item, 833 access_flags, 834 invoke_type, 835 class_def_idx, 836 method_idx, 837 jclass_loader, 838 dex_file, 839 dex_cache)); 840 if (codegen.get() != nullptr) { 841 if (run_optimizations_) { 842 method = EmitOptimized(&arena, &code_allocator, codegen.get(), compiler_driver); 843 } else { 844 method = EmitBaseline(&arena, &code_allocator, codegen.get(), compiler_driver); 845 } 846 } 847 } else { 848 if (compiler_driver->GetCompilerOptions().VerifyAtRuntime()) { 849 MaybeRecordStat(MethodCompilationStat::kNotCompiledVerifyAtRuntime); 850 } else { 851 MaybeRecordStat(MethodCompilationStat::kNotCompiledClassNotVerified); 852 } 853 } 854 855 if (kIsDebugBuild && 856 IsCompilingWithCoreImage() && 857 IsInstructionSetSupported(compiler_driver->GetInstructionSet()) && 858 (!kEmitCompilerReadBarrier || 859 InstructionSetSupportsReadBarrier(compiler_driver->GetInstructionSet()))) { 860 // For testing purposes, we put a special marker on method names 861 // that should be compiled with this compiler (when the the 862 // instruction set is supported -- and has support for read 863 // barriers, if they are enabled). This makes sure we're not 864 // regressing. 865 std::string method_name = PrettyMethod(method_idx, dex_file); 866 bool shouldCompile = method_name.find("$opt$") != std::string::npos; 867 DCHECK((method != nullptr) || !shouldCompile) << "Didn't compile " << method_name; 868 } 869 870 return method; 871} 872 873Compiler* CreateOptimizingCompiler(CompilerDriver* driver) { 874 return new OptimizingCompiler(driver); 875} 876 877bool IsCompilingWithCoreImage() { 878 const std::string& image = Runtime::Current()->GetImageLocation(); 879 return EndsWith(image, "core.art") || EndsWith(image, "core-optimizing.art"); 880} 881 882bool OptimizingCompiler::JitCompile(Thread* self, 883 jit::JitCodeCache* code_cache, 884 ArtMethod* method) { 885 StackHandleScope<2> hs(self); 886 Handle<mirror::ClassLoader> class_loader(hs.NewHandle( 887 method->GetDeclaringClass()->GetClassLoader())); 888 Handle<mirror::DexCache> dex_cache(hs.NewHandle(method->GetDexCache())); 889 890 jobject jclass_loader = class_loader.ToJObject(); 891 const DexFile* dex_file = method->GetDexFile(); 892 const uint16_t class_def_idx = method->GetClassDefIndex(); 893 const DexFile::CodeItem* code_item = dex_file->GetCodeItem(method->GetCodeItemOffset()); 894 const uint32_t method_idx = method->GetDexMethodIndex(); 895 const uint32_t access_flags = method->GetAccessFlags(); 896 const InvokeType invoke_type = method->GetInvokeType(); 897 898 ArenaAllocator arena(Runtime::Current()->GetArenaPool()); 899 CodeVectorAllocator code_allocator(&arena); 900 std::unique_ptr<CodeGenerator> codegen; 901 { 902 // Go to native so that we don't block GC during compilation. 903 ScopedThreadSuspension sts(self, kNative); 904 905 DCHECK(run_optimizations_); 906 codegen.reset( 907 TryCompile(&arena, 908 &code_allocator, 909 code_item, 910 access_flags, 911 invoke_type, 912 class_def_idx, 913 method_idx, 914 jclass_loader, 915 *dex_file, 916 dex_cache)); 917 if (codegen.get() == nullptr) { 918 return false; 919 } 920 } 921 922 size_t stack_map_size = codegen->ComputeStackMapsSize(); 923 uint8_t* stack_map_data = code_cache->ReserveData(self, stack_map_size); 924 if (stack_map_data == nullptr) { 925 return false; 926 } 927 codegen->BuildStackMaps(MemoryRegion(stack_map_data, stack_map_size)); 928 const void* code = code_cache->CommitCode( 929 self, 930 method, 931 nullptr, 932 stack_map_data, 933 nullptr, 934 codegen->HasEmptyFrame() ? 0 : codegen->GetFrameSize(), 935 codegen->GetCoreSpillMask(), 936 codegen->GetFpuSpillMask(), 937 code_allocator.GetMemory().data(), 938 code_allocator.GetSize()); 939 940 if (code == nullptr) { 941 code_cache->ClearData(self, stack_map_data); 942 return false; 943 } 944 945 return true; 946} 947 948} // namespace art 949