optimizing_compiler.cc revision 4a34277c55279ba57ab361f7580db846a201d9b1
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 "dex_cache_array_fixups_arm.h" 24#endif 25 26#ifdef ART_ENABLE_CODEGEN_arm64 27#include "instruction_simplifier_arm64.h" 28#endif 29 30#ifdef ART_ENABLE_CODEGEN_x86 31#include "pc_relative_fixups_x86.h" 32#endif 33 34#include "art_method-inl.h" 35#include "base/arena_allocator.h" 36#include "base/arena_containers.h" 37#include "base/dumpable.h" 38#include "base/macros.h" 39#include "base/timing_logger.h" 40#include "boolean_simplifier.h" 41#include "bounds_check_elimination.h" 42#include "builder.h" 43#include "code_generator.h" 44#include "compiled_method.h" 45#include "compiler.h" 46#include "constant_folding.h" 47#include "dead_code_elimination.h" 48#include "dex/quick/dex_file_to_method_inliner_map.h" 49#include "dex/verified_method.h" 50#include "dex/verification_results.h" 51#include "driver/compiler_driver.h" 52#include "driver/compiler_driver-inl.h" 53#include "driver/compiler_options.h" 54#include "driver/dex_compilation_unit.h" 55#include "elf_writer_quick.h" 56#include "graph_checker.h" 57#include "graph_visualizer.h" 58#include "gvn.h" 59#include "induction_var_analysis.h" 60#include "inliner.h" 61#include "instruction_simplifier.h" 62#include "intrinsics.h" 63#include "jit/jit_code_cache.h" 64#include "licm.h" 65#include "jni/quick/jni_compiler.h" 66#include "load_store_elimination.h" 67#include "nodes.h" 68#include "prepare_for_register_allocation.h" 69#include "reference_type_propagation.h" 70#include "register_allocator.h" 71#include "sharpening.h" 72#include "side_effects_analysis.h" 73#include "ssa_builder.h" 74#include "ssa_phi_elimination.h" 75#include "ssa_liveness_analysis.h" 76#include "utils/assembler.h" 77#include "verifier/method_verifier.h" 78 79namespace art { 80 81/** 82 * Used by the code generator, to allocate the code in a vector. 83 */ 84class CodeVectorAllocator FINAL : public CodeAllocator { 85 public: 86 explicit CodeVectorAllocator(ArenaAllocator* arena) 87 : memory_(arena->Adapter(kArenaAllocCodeBuffer)), 88 size_(0) {} 89 90 virtual uint8_t* Allocate(size_t size) { 91 size_ = size; 92 memory_.resize(size); 93 return &memory_[0]; 94 } 95 96 size_t GetSize() const { return size_; } 97 const ArenaVector<uint8_t>& GetMemory() const { return memory_; } 98 99 private: 100 ArenaVector<uint8_t> memory_; 101 size_t size_; 102 103 DISALLOW_COPY_AND_ASSIGN(CodeVectorAllocator); 104}; 105 106/** 107 * Filter to apply to the visualizer. Methods whose name contain that filter will 108 * be dumped. 109 */ 110static constexpr const char kStringFilter[] = ""; 111 112class PassScope; 113 114class PassObserver : public ValueObject { 115 public: 116 PassObserver(HGraph* graph, 117 CodeGenerator* codegen, 118 std::ostream* visualizer_output, 119 CompilerDriver* compiler_driver) 120 : graph_(graph), 121 cached_method_name_(), 122 timing_logger_enabled_(compiler_driver->GetDumpPasses()), 123 timing_logger_(timing_logger_enabled_ ? GetMethodName() : "", true, true), 124 disasm_info_(graph->GetArena()), 125 visualizer_enabled_(!compiler_driver->GetDumpCfgFileName().empty()), 126 visualizer_(visualizer_output, graph, *codegen), 127 graph_in_bad_state_(false) { 128 if (timing_logger_enabled_ || visualizer_enabled_) { 129 if (!IsVerboseMethod(compiler_driver, GetMethodName())) { 130 timing_logger_enabled_ = visualizer_enabled_ = false; 131 } 132 if (visualizer_enabled_) { 133 visualizer_.PrintHeader(GetMethodName()); 134 codegen->SetDisassemblyInformation(&disasm_info_); 135 } 136 } 137 } 138 139 ~PassObserver() { 140 if (timing_logger_enabled_) { 141 LOG(INFO) << "TIMINGS " << GetMethodName(); 142 LOG(INFO) << Dumpable<TimingLogger>(timing_logger_); 143 } 144 } 145 146 void DumpDisassembly() const { 147 if (visualizer_enabled_) { 148 visualizer_.DumpGraphWithDisassembly(); 149 } 150 } 151 152 void SetGraphInBadState() { graph_in_bad_state_ = true; } 153 154 const char* GetMethodName() { 155 // PrettyMethod() is expensive, so we delay calling it until we actually have to. 156 if (cached_method_name_.empty()) { 157 cached_method_name_ = PrettyMethod(graph_->GetMethodIdx(), graph_->GetDexFile()); 158 } 159 return cached_method_name_.c_str(); 160 } 161 162 private: 163 void StartPass(const char* pass_name) { 164 // Dump graph first, then start timer. 165 if (visualizer_enabled_) { 166 visualizer_.DumpGraph(pass_name, /* is_after_pass */ false, graph_in_bad_state_); 167 } 168 if (timing_logger_enabled_) { 169 timing_logger_.StartTiming(pass_name); 170 } 171 } 172 173 void EndPass(const char* pass_name) { 174 // Pause timer first, then dump graph. 175 if (timing_logger_enabled_) { 176 timing_logger_.EndTiming(); 177 } 178 if (visualizer_enabled_) { 179 visualizer_.DumpGraph(pass_name, /* is_after_pass */ true, graph_in_bad_state_); 180 } 181 182 // Validate the HGraph if running in debug mode. 183 if (kIsDebugBuild) { 184 if (!graph_in_bad_state_) { 185 if (graph_->IsInSsaForm()) { 186 SSAChecker checker(graph_); 187 checker.Run(); 188 if (!checker.IsValid()) { 189 LOG(FATAL) << "Error after " << pass_name << ": " << Dumpable<SSAChecker>(checker); 190 } 191 } else { 192 GraphChecker checker(graph_); 193 checker.Run(); 194 if (!checker.IsValid()) { 195 LOG(FATAL) << "Error after " << pass_name << ": " << Dumpable<GraphChecker>(checker); 196 } 197 } 198 } 199 } 200 } 201 202 static bool IsVerboseMethod(CompilerDriver* compiler_driver, const char* method_name) { 203 // Test an exact match to --verbose-methods. If verbose-methods is set, this overrides an 204 // empty kStringFilter matching all methods. 205 if (compiler_driver->GetCompilerOptions().HasVerboseMethods()) { 206 return compiler_driver->GetCompilerOptions().IsVerboseMethod(method_name); 207 } 208 209 // Test the kStringFilter sub-string. constexpr helper variable to silence unreachable-code 210 // warning when the string is empty. 211 constexpr bool kStringFilterEmpty = arraysize(kStringFilter) <= 1; 212 if (kStringFilterEmpty || strstr(method_name, kStringFilter) != nullptr) { 213 return true; 214 } 215 216 return false; 217 } 218 219 HGraph* const graph_; 220 221 std::string cached_method_name_; 222 223 bool timing_logger_enabled_; 224 TimingLogger timing_logger_; 225 226 DisassemblyInformation disasm_info_; 227 228 bool visualizer_enabled_; 229 HGraphVisualizer visualizer_; 230 231 // Flag to be set by the compiler if the pass failed and the graph is not 232 // expected to validate. 233 bool graph_in_bad_state_; 234 235 friend PassScope; 236 237 DISALLOW_COPY_AND_ASSIGN(PassObserver); 238}; 239 240class PassScope : public ValueObject { 241 public: 242 PassScope(const char *pass_name, PassObserver* pass_observer) 243 : pass_name_(pass_name), 244 pass_observer_(pass_observer) { 245 pass_observer_->StartPass(pass_name_); 246 } 247 248 ~PassScope() { 249 pass_observer_->EndPass(pass_name_); 250 } 251 252 private: 253 const char* const pass_name_; 254 PassObserver* const pass_observer_; 255}; 256 257class OptimizingCompiler FINAL : public Compiler { 258 public: 259 explicit OptimizingCompiler(CompilerDriver* driver); 260 ~OptimizingCompiler(); 261 262 bool CanCompileMethod(uint32_t method_idx, const DexFile& dex_file, CompilationUnit* cu) const 263 OVERRIDE; 264 265 CompiledMethod* Compile(const DexFile::CodeItem* code_item, 266 uint32_t access_flags, 267 InvokeType invoke_type, 268 uint16_t class_def_idx, 269 uint32_t method_idx, 270 jobject class_loader, 271 const DexFile& dex_file, 272 Handle<mirror::DexCache> dex_cache) const OVERRIDE; 273 274 CompiledMethod* JniCompile(uint32_t access_flags, 275 uint32_t method_idx, 276 const DexFile& dex_file) const OVERRIDE { 277 return ArtQuickJniCompileMethod(GetCompilerDriver(), access_flags, method_idx, dex_file); 278 } 279 280 uintptr_t GetEntryPointOf(ArtMethod* method) const OVERRIDE 281 SHARED_REQUIRES(Locks::mutator_lock_) { 282 return reinterpret_cast<uintptr_t>(method->GetEntryPointFromQuickCompiledCodePtrSize( 283 InstructionSetPointerSize(GetCompilerDriver()->GetInstructionSet()))); 284 } 285 286 void InitCompilationUnit(CompilationUnit& cu) const OVERRIDE; 287 288 void Init() OVERRIDE; 289 290 void UnInit() const OVERRIDE; 291 292 void MaybeRecordStat(MethodCompilationStat compilation_stat) const { 293 if (compilation_stats_.get() != nullptr) { 294 compilation_stats_->RecordStat(compilation_stat); 295 } 296 } 297 298 bool JitCompile(Thread* self, jit::JitCodeCache* code_cache, ArtMethod* method) 299 OVERRIDE 300 SHARED_REQUIRES(Locks::mutator_lock_); 301 302 private: 303 // Whether we should run any optimization or register allocation. If false, will 304 // just run the code generation after the graph was built. 305 const bool run_optimizations_; 306 307 // Create a 'CompiledMethod' for an optimized graph. 308 CompiledMethod* EmitOptimized(ArenaAllocator* arena, 309 CodeVectorAllocator* code_allocator, 310 CodeGenerator* codegen, 311 CompilerDriver* driver) const; 312 313 // Create a 'CompiledMethod' for a non-optimized graph. 314 CompiledMethod* EmitBaseline(ArenaAllocator* arena, 315 CodeVectorAllocator* code_allocator, 316 CodeGenerator* codegen, 317 CompilerDriver* driver) const; 318 319 // Try compiling a method and return the code generator used for 320 // compiling it. 321 // This method: 322 // 1) Builds the graph. Returns null if it failed to build it. 323 // 2) If `run_optimizations_` is set: 324 // 2.1) Transform the graph to SSA. Returns null if it failed. 325 // 2.2) Run optimizations on the graph, including register allocator. 326 // 3) Generate code with the `code_allocator` provided. 327 CodeGenerator* TryCompile(ArenaAllocator* arena, 328 CodeVectorAllocator* code_allocator, 329 const DexFile::CodeItem* code_item, 330 uint32_t access_flags, 331 InvokeType invoke_type, 332 uint16_t class_def_idx, 333 uint32_t method_idx, 334 jobject class_loader, 335 const DexFile& dex_file, 336 Handle<mirror::DexCache> dex_cache) const; 337 338 std::unique_ptr<OptimizingCompilerStats> compilation_stats_; 339 340 std::unique_ptr<std::ostream> visualizer_output_; 341 342 DISALLOW_COPY_AND_ASSIGN(OptimizingCompiler); 343}; 344 345static const int kMaximumCompilationTimeBeforeWarning = 100; /* ms */ 346 347OptimizingCompiler::OptimizingCompiler(CompilerDriver* driver) 348 : Compiler(driver, kMaximumCompilationTimeBeforeWarning), 349 run_optimizations_( 350 driver->GetCompilerOptions().GetCompilerFilter() != CompilerOptions::kTime) {} 351 352void OptimizingCompiler::Init() { 353 // Enable C1visualizer output. Must be done in Init() because the compiler 354 // driver is not fully initialized when passed to the compiler's constructor. 355 CompilerDriver* driver = GetCompilerDriver(); 356 const std::string cfg_file_name = driver->GetDumpCfgFileName(); 357 if (!cfg_file_name.empty()) { 358 CHECK_EQ(driver->GetThreadCount(), 1U) 359 << "Graph visualizer requires the compiler to run single-threaded. " 360 << "Invoke the compiler with '-j1'."; 361 std::ios_base::openmode cfg_file_mode = 362 driver->GetDumpCfgAppend() ? std::ofstream::app : std::ofstream::out; 363 visualizer_output_.reset(new std::ofstream(cfg_file_name, cfg_file_mode)); 364 } 365 if (driver->GetDumpStats()) { 366 compilation_stats_.reset(new OptimizingCompilerStats()); 367 } 368} 369 370void OptimizingCompiler::UnInit() const { 371} 372 373OptimizingCompiler::~OptimizingCompiler() { 374 if (compilation_stats_.get() != nullptr) { 375 compilation_stats_->Log(); 376 } 377} 378 379void OptimizingCompiler::InitCompilationUnit(CompilationUnit& cu ATTRIBUTE_UNUSED) const { 380} 381 382bool OptimizingCompiler::CanCompileMethod(uint32_t method_idx ATTRIBUTE_UNUSED, 383 const DexFile& dex_file ATTRIBUTE_UNUSED, 384 CompilationUnit* cu ATTRIBUTE_UNUSED) const { 385 return true; 386} 387 388static bool IsInstructionSetSupported(InstructionSet instruction_set) { 389 return (instruction_set == kArm && !kArm32QuickCodeUseSoftFloat) 390 || instruction_set == kArm64 391 || (instruction_set == kThumb2 && !kArm32QuickCodeUseSoftFloat) 392 || instruction_set == kMips 393 || instruction_set == kMips64 394 || instruction_set == kX86 395 || instruction_set == kX86_64; 396} 397 398// Read barrier are supported on ARM, ARM64, x86 and x86-64 at the moment. 399// TODO: Add support for other architectures and remove this function 400static bool InstructionSetSupportsReadBarrier(InstructionSet instruction_set) { 401 return instruction_set == kArm64 402 || instruction_set == kThumb2 403 || instruction_set == kX86 404 || instruction_set == kX86_64; 405} 406 407static void RunOptimizations(HOptimization* optimizations[], 408 size_t length, 409 PassObserver* pass_observer) { 410 for (size_t i = 0; i < length; ++i) { 411 PassScope scope(optimizations[i]->GetPassName(), pass_observer); 412 optimizations[i]->Run(); 413 } 414} 415 416static void MaybeRunInliner(HGraph* graph, 417 CodeGenerator* codegen, 418 CompilerDriver* driver, 419 OptimizingCompilerStats* stats, 420 const DexCompilationUnit& dex_compilation_unit, 421 PassObserver* pass_observer, 422 StackHandleScopeCollection* handles) { 423 const CompilerOptions& compiler_options = driver->GetCompilerOptions(); 424 bool should_inline = (compiler_options.GetInlineDepthLimit() > 0) 425 && (compiler_options.GetInlineMaxCodeUnits() > 0); 426 if (!should_inline) { 427 return; 428 } 429 HInliner* inliner = new (graph->GetArena()) HInliner( 430 graph, codegen, dex_compilation_unit, dex_compilation_unit, driver, handles, stats); 431 HOptimization* optimizations[] = { inliner }; 432 433 RunOptimizations(optimizations, arraysize(optimizations), pass_observer); 434} 435 436static void RunArchOptimizations(InstructionSet instruction_set, 437 HGraph* graph, 438 OptimizingCompilerStats* stats, 439 PassObserver* pass_observer) { 440 ArenaAllocator* arena = graph->GetArena(); 441 switch (instruction_set) { 442#ifdef ART_ENABLE_CODEGEN_arm 443 case kThumb2: 444 case kArm: { 445 arm::DexCacheArrayFixups* fixups = new (arena) arm::DexCacheArrayFixups(graph, stats); 446 HOptimization* arm_optimizations[] = { 447 fixups 448 }; 449 RunOptimizations(arm_optimizations, arraysize(arm_optimizations), pass_observer); 450 break; 451 } 452#endif 453#ifdef ART_ENABLE_CODEGEN_arm64 454 case kArm64: { 455 arm64::InstructionSimplifierArm64* simplifier = 456 new (arena) arm64::InstructionSimplifierArm64(graph, stats); 457 SideEffectsAnalysis* side_effects = new (arena) SideEffectsAnalysis(graph); 458 GVNOptimization* gvn = new (arena) GVNOptimization(graph, *side_effects, "GVN_after_arch"); 459 HOptimization* arm64_optimizations[] = { 460 simplifier, 461 side_effects, 462 gvn 463 }; 464 RunOptimizations(arm64_optimizations, arraysize(arm64_optimizations), pass_observer); 465 break; 466 } 467#endif 468#ifdef ART_ENABLE_CODEGEN_x86 469 case kX86: { 470 x86::PcRelativeFixups* pc_relative_fixups = new (arena) x86::PcRelativeFixups(graph, stats); 471 HOptimization* x86_optimizations[] = { 472 pc_relative_fixups 473 }; 474 RunOptimizations(x86_optimizations, arraysize(x86_optimizations), pass_observer); 475 break; 476 } 477#endif 478 default: 479 break; 480 } 481} 482 483NO_INLINE // Avoid increasing caller's frame size by large stack-allocated objects. 484static void AllocateRegisters(HGraph* graph, 485 CodeGenerator* codegen, 486 PassObserver* pass_observer) { 487 PrepareForRegisterAllocation(graph).Run(); 488 SsaLivenessAnalysis liveness(graph, codegen); 489 { 490 PassScope scope(SsaLivenessAnalysis::kLivenessPassName, pass_observer); 491 liveness.Analyze(); 492 } 493 { 494 PassScope scope(RegisterAllocator::kRegisterAllocatorPassName, pass_observer); 495 RegisterAllocator(graph->GetArena(), codegen, liveness).AllocateRegisters(); 496 } 497} 498 499static void RunOptimizations(HGraph* graph, 500 CodeGenerator* codegen, 501 CompilerDriver* driver, 502 OptimizingCompilerStats* stats, 503 const DexCompilationUnit& dex_compilation_unit, 504 PassObserver* pass_observer) { 505 ScopedObjectAccess soa(Thread::Current()); 506 StackHandleScopeCollection handles(soa.Self()); 507 ScopedThreadSuspension sts(soa.Self(), kNative); 508 509 ArenaAllocator* arena = graph->GetArena(); 510 HDeadCodeElimination* dce1 = new (arena) HDeadCodeElimination( 511 graph, stats, HDeadCodeElimination::kInitialDeadCodeEliminationPassName); 512 HDeadCodeElimination* dce2 = new (arena) HDeadCodeElimination( 513 graph, stats, HDeadCodeElimination::kFinalDeadCodeEliminationPassName); 514 HConstantFolding* fold1 = new (arena) HConstantFolding(graph); 515 InstructionSimplifier* simplify1 = new (arena) InstructionSimplifier(graph, stats); 516 HBooleanSimplifier* boolean_simplify = new (arena) HBooleanSimplifier(graph); 517 HConstantFolding* fold2 = new (arena) HConstantFolding(graph, "constant_folding_after_inlining"); 518 HConstantFolding* fold3 = new (arena) HConstantFolding(graph, "constant_folding_after_bce"); 519 SideEffectsAnalysis* side_effects = new (arena) SideEffectsAnalysis(graph); 520 GVNOptimization* gvn = new (arena) GVNOptimization(graph, *side_effects); 521 LICM* licm = new (arena) LICM(graph, *side_effects); 522 LoadStoreElimination* lse = new (arena) LoadStoreElimination(graph, *side_effects); 523 HInductionVarAnalysis* induction = new (arena) HInductionVarAnalysis(graph); 524 BoundsCheckElimination* bce = new (arena) BoundsCheckElimination(graph, *side_effects, induction); 525 ReferenceTypePropagation* type_propagation = 526 new (arena) ReferenceTypePropagation(graph, &handles); 527 HSharpening* sharpening = new (arena) HSharpening(graph, codegen, dex_compilation_unit, driver); 528 InstructionSimplifier* simplify2 = new (arena) InstructionSimplifier( 529 graph, stats, "instruction_simplifier_after_types"); 530 InstructionSimplifier* simplify3 = new (arena) InstructionSimplifier( 531 graph, stats, "instruction_simplifier_after_bce"); 532 InstructionSimplifier* simplify4 = new (arena) InstructionSimplifier( 533 graph, stats, "instruction_simplifier_before_codegen"); 534 535 IntrinsicsRecognizer* intrinsics = new (arena) IntrinsicsRecognizer(graph, driver); 536 537 HOptimization* optimizations1[] = { 538 intrinsics, 539 fold1, 540 simplify1, 541 type_propagation, 542 sharpening, 543 dce1, 544 simplify2 545 }; 546 547 RunOptimizations(optimizations1, arraysize(optimizations1), pass_observer); 548 549 MaybeRunInliner(graph, codegen, driver, stats, dex_compilation_unit, pass_observer, &handles); 550 551 // TODO: Update passes incompatible with try/catch so we have the same 552 // pipeline for all methods. 553 if (graph->HasTryCatch()) { 554 HOptimization* optimizations2[] = { 555 boolean_simplify, 556 side_effects, 557 gvn, 558 dce2, 559 // The codegen has a few assumptions that only the instruction simplifier 560 // can satisfy. For example, the code generator does not expect to see a 561 // HTypeConversion from a type to the same type. 562 simplify4, 563 }; 564 565 RunOptimizations(optimizations2, arraysize(optimizations2), pass_observer); 566 } else { 567 HOptimization* optimizations2[] = { 568 // BooleanSimplifier depends on the InstructionSimplifier removing 569 // redundant suspend checks to recognize empty blocks. 570 boolean_simplify, 571 fold2, // TODO: if we don't inline we can also skip fold2. 572 side_effects, 573 gvn, 574 licm, 575 induction, 576 bce, 577 fold3, // evaluates code generated by dynamic bce 578 simplify3, 579 lse, 580 dce2, 581 // The codegen has a few assumptions that only the instruction simplifier 582 // can satisfy. For example, the code generator does not expect to see a 583 // HTypeConversion from a type to the same type. 584 simplify4, 585 }; 586 587 RunOptimizations(optimizations2, arraysize(optimizations2), pass_observer); 588 } 589 590 RunArchOptimizations(driver->GetInstructionSet(), graph, stats, pass_observer); 591 AllocateRegisters(graph, codegen, pass_observer); 592} 593 594// The stack map we generate must be 4-byte aligned on ARM. Since existing 595// maps are generated alongside these stack maps, we must also align them. 596static ArrayRef<const uint8_t> AlignVectorSize(ArenaVector<uint8_t>& vector) { 597 size_t size = vector.size(); 598 size_t aligned_size = RoundUp(size, 4); 599 for (; size < aligned_size; ++size) { 600 vector.push_back(0); 601 } 602 return ArrayRef<const uint8_t>(vector); 603} 604 605static ArenaVector<LinkerPatch> EmitAndSortLinkerPatches(CodeGenerator* codegen) { 606 ArenaVector<LinkerPatch> linker_patches(codegen->GetGraph()->GetArena()->Adapter()); 607 codegen->EmitLinkerPatches(&linker_patches); 608 609 // Sort patches by literal offset. Required for .oat_patches encoding. 610 std::sort(linker_patches.begin(), linker_patches.end(), 611 [](const LinkerPatch& lhs, const LinkerPatch& rhs) { 612 return lhs.LiteralOffset() < rhs.LiteralOffset(); 613 }); 614 615 return linker_patches; 616} 617 618CompiledMethod* OptimizingCompiler::EmitOptimized(ArenaAllocator* arena, 619 CodeVectorAllocator* code_allocator, 620 CodeGenerator* codegen, 621 CompilerDriver* compiler_driver) const { 622 ArenaVector<LinkerPatch> linker_patches = EmitAndSortLinkerPatches(codegen); 623 ArenaVector<uint8_t> stack_map(arena->Adapter(kArenaAllocStackMaps)); 624 stack_map.resize(codegen->ComputeStackMapsSize()); 625 codegen->BuildStackMaps(MemoryRegion(stack_map.data(), stack_map.size())); 626 627 MaybeRecordStat(MethodCompilationStat::kCompiledOptimized); 628 629 CompiledMethod* compiled_method = CompiledMethod::SwapAllocCompiledMethod( 630 compiler_driver, 631 codegen->GetInstructionSet(), 632 ArrayRef<const uint8_t>(code_allocator->GetMemory()), 633 // Follow Quick's behavior and set the frame size to zero if it is 634 // considered "empty" (see the definition of 635 // art::CodeGenerator::HasEmptyFrame). 636 codegen->HasEmptyFrame() ? 0 : codegen->GetFrameSize(), 637 codegen->GetCoreSpillMask(), 638 codegen->GetFpuSpillMask(), 639 ArrayRef<const SrcMapElem>(codegen->GetSrcMappingTable()), 640 ArrayRef<const uint8_t>(), // mapping_table. 641 ArrayRef<const uint8_t>(stack_map), 642 ArrayRef<const uint8_t>(), // native_gc_map. 643 ArrayRef<const uint8_t>(*codegen->GetAssembler()->cfi().data()), 644 ArrayRef<const LinkerPatch>(linker_patches)); 645 646 return compiled_method; 647} 648 649CompiledMethod* OptimizingCompiler::EmitBaseline( 650 ArenaAllocator* arena, 651 CodeVectorAllocator* code_allocator, 652 CodeGenerator* codegen, 653 CompilerDriver* compiler_driver) const { 654 ArenaVector<LinkerPatch> linker_patches = EmitAndSortLinkerPatches(codegen); 655 656 ArenaVector<uint8_t> mapping_table(arena->Adapter(kArenaAllocBaselineMaps)); 657 codegen->BuildMappingTable(&mapping_table); 658 ArenaVector<uint8_t> vmap_table(arena->Adapter(kArenaAllocBaselineMaps)); 659 codegen->BuildVMapTable(&vmap_table); 660 ArenaVector<uint8_t> gc_map(arena->Adapter(kArenaAllocBaselineMaps)); 661 codegen->BuildNativeGCMap(&gc_map, *compiler_driver); 662 663 MaybeRecordStat(MethodCompilationStat::kCompiledBaseline); 664 CompiledMethod* compiled_method = CompiledMethod::SwapAllocCompiledMethod( 665 compiler_driver, 666 codegen->GetInstructionSet(), 667 ArrayRef<const uint8_t>(code_allocator->GetMemory()), 668 // Follow Quick's behavior and set the frame size to zero if it is 669 // considered "empty" (see the definition of 670 // art::CodeGenerator::HasEmptyFrame). 671 codegen->HasEmptyFrame() ? 0 : codegen->GetFrameSize(), 672 codegen->GetCoreSpillMask(), 673 codegen->GetFpuSpillMask(), 674 ArrayRef<const SrcMapElem>(codegen->GetSrcMappingTable()), 675 AlignVectorSize(mapping_table), 676 AlignVectorSize(vmap_table), 677 AlignVectorSize(gc_map), 678 ArrayRef<const uint8_t>(*codegen->GetAssembler()->cfi().data()), 679 ArrayRef<const LinkerPatch>(linker_patches)); 680 return compiled_method; 681} 682 683CodeGenerator* OptimizingCompiler::TryCompile(ArenaAllocator* arena, 684 CodeVectorAllocator* code_allocator, 685 const DexFile::CodeItem* code_item, 686 uint32_t access_flags, 687 InvokeType invoke_type, 688 uint16_t class_def_idx, 689 uint32_t method_idx, 690 jobject class_loader, 691 const DexFile& dex_file, 692 Handle<mirror::DexCache> dex_cache) const { 693 MaybeRecordStat(MethodCompilationStat::kAttemptCompilation); 694 CompilerDriver* compiler_driver = GetCompilerDriver(); 695 InstructionSet instruction_set = compiler_driver->GetInstructionSet(); 696 697 // Always use the Thumb-2 assembler: some runtime functionality 698 // (like implicit stack overflow checks) assume Thumb-2. 699 if (instruction_set == kArm) { 700 instruction_set = kThumb2; 701 } 702 703 // Do not attempt to compile on architectures we do not support. 704 if (!IsInstructionSetSupported(instruction_set)) { 705 MaybeRecordStat(MethodCompilationStat::kNotCompiledUnsupportedIsa); 706 return nullptr; 707 } 708 709 // When read barriers are enabled, do not attempt to compile for 710 // instruction sets that have no read barrier support. 711 if (kEmitCompilerReadBarrier && !InstructionSetSupportsReadBarrier(instruction_set)) { 712 return nullptr; 713 } 714 715 if (Compiler::IsPathologicalCase(*code_item, method_idx, dex_file)) { 716 MaybeRecordStat(MethodCompilationStat::kNotCompiledPathological); 717 return nullptr; 718 } 719 720 // Implementation of the space filter: do not compile a code item whose size in 721 // code units is bigger than 128. 722 static constexpr size_t kSpaceFilterOptimizingThreshold = 128; 723 const CompilerOptions& compiler_options = compiler_driver->GetCompilerOptions(); 724 if ((compiler_options.GetCompilerFilter() == CompilerOptions::kSpace) 725 && (code_item->insns_size_in_code_units_ > kSpaceFilterOptimizingThreshold)) { 726 MaybeRecordStat(MethodCompilationStat::kNotCompiledSpaceFilter); 727 return nullptr; 728 } 729 730 DexCompilationUnit dex_compilation_unit( 731 nullptr, class_loader, Runtime::Current()->GetClassLinker(), dex_file, code_item, 732 class_def_idx, method_idx, access_flags, 733 compiler_driver->GetVerifiedMethod(&dex_file, method_idx), dex_cache); 734 735 bool requires_barrier = dex_compilation_unit.IsConstructor() 736 && compiler_driver->RequiresConstructorBarrier(Thread::Current(), 737 dex_compilation_unit.GetDexFile(), 738 dex_compilation_unit.GetClassDefIndex()); 739 HGraph* graph = new (arena) HGraph( 740 arena, dex_file, method_idx, requires_barrier, compiler_driver->GetInstructionSet(), 741 kInvalidInvokeType, compiler_driver->GetCompilerOptions().GetDebuggable()); 742 743 std::unique_ptr<CodeGenerator> codegen( 744 CodeGenerator::Create(graph, 745 instruction_set, 746 *compiler_driver->GetInstructionSetFeatures(), 747 compiler_driver->GetCompilerOptions())); 748 if (codegen.get() == nullptr) { 749 MaybeRecordStat(MethodCompilationStat::kNotCompiledNoCodegen); 750 return nullptr; 751 } 752 codegen->GetAssembler()->cfi().SetEnabled( 753 compiler_driver->GetCompilerOptions().GetGenerateDebugInfo()); 754 755 PassObserver pass_observer(graph, 756 codegen.get(), 757 visualizer_output_.get(), 758 compiler_driver); 759 760 const uint8_t* interpreter_metadata = nullptr; 761 { 762 ScopedObjectAccess soa(Thread::Current()); 763 StackHandleScope<1> hs(soa.Self()); 764 Handle<mirror::ClassLoader> loader(hs.NewHandle( 765 soa.Decode<mirror::ClassLoader*>(class_loader))); 766 ArtMethod* art_method = compiler_driver->ResolveMethod( 767 soa, dex_cache, loader, &dex_compilation_unit, method_idx, invoke_type); 768 // We may not get a method, for example if its class is erroneous. 769 // TODO: Clean this up, the compiler driver should just pass the ArtMethod to compile. 770 if (art_method != nullptr) { 771 interpreter_metadata = art_method->GetQuickenedInfo(); 772 } 773 } 774 HGraphBuilder builder(graph, 775 &dex_compilation_unit, 776 &dex_compilation_unit, 777 &dex_file, 778 compiler_driver, 779 compilation_stats_.get(), 780 interpreter_metadata, 781 dex_cache); 782 783 VLOG(compiler) << "Building " << pass_observer.GetMethodName(); 784 785 { 786 PassScope scope(HGraphBuilder::kBuilderPassName, &pass_observer); 787 if (!builder.BuildGraph(*code_item)) { 788 pass_observer.SetGraphInBadState(); 789 return nullptr; 790 } 791 } 792 793 VLOG(compiler) << "Optimizing " << pass_observer.GetMethodName(); 794 if (run_optimizations_) { 795 { 796 PassScope scope(SsaBuilder::kSsaBuilderPassName, &pass_observer); 797 if (!graph->TryBuildingSsa()) { 798 // We could not transform the graph to SSA, bailout. 799 LOG(INFO) << "Skipping compilation of " << pass_observer.GetMethodName() 800 << ": it contains a non natural loop"; 801 MaybeRecordStat(MethodCompilationStat::kNotCompiledCannotBuildSSA); 802 pass_observer.SetGraphInBadState(); 803 return nullptr; 804 } 805 } 806 807 RunOptimizations(graph, 808 codegen.get(), 809 compiler_driver, 810 compilation_stats_.get(), 811 dex_compilation_unit, 812 &pass_observer); 813 codegen->CompileOptimized(code_allocator); 814 } else { 815 codegen->CompileBaseline(code_allocator); 816 } 817 pass_observer.DumpDisassembly(); 818 819 if (kArenaAllocatorCountAllocations) { 820 if (arena->BytesAllocated() > 4 * MB) { 821 MemStats mem_stats(arena->GetMemStats()); 822 LOG(INFO) << PrettyMethod(method_idx, dex_file) << " " << Dumpable<MemStats>(mem_stats); 823 } 824 } 825 826 return codegen.release(); 827} 828 829static bool CanHandleVerificationFailure(const VerifiedMethod* verified_method) { 830 // For access errors the compiler will use the unresolved helpers (e.g. HInvokeUnresolved). 831 uint32_t unresolved_mask = verifier::VerifyError::VERIFY_ERROR_NO_CLASS 832 | verifier::VerifyError::VERIFY_ERROR_ACCESS_CLASS 833 | verifier::VerifyError::VERIFY_ERROR_ACCESS_FIELD 834 | verifier::VerifyError::VERIFY_ERROR_ACCESS_METHOD; 835 return (verified_method->GetEncounteredVerificationFailures() & (~unresolved_mask)) == 0; 836} 837 838CompiledMethod* OptimizingCompiler::Compile(const DexFile::CodeItem* code_item, 839 uint32_t access_flags, 840 InvokeType invoke_type, 841 uint16_t class_def_idx, 842 uint32_t method_idx, 843 jobject jclass_loader, 844 const DexFile& dex_file, 845 Handle<mirror::DexCache> dex_cache) const { 846 CompilerDriver* compiler_driver = GetCompilerDriver(); 847 CompiledMethod* method = nullptr; 848 DCHECK(Runtime::Current()->IsAotCompiler()); 849 const VerifiedMethod* verified_method = compiler_driver->GetVerifiedMethod(&dex_file, method_idx); 850 DCHECK(!verified_method->HasRuntimeThrow()); 851 if (compiler_driver->IsMethodVerifiedWithoutFailures(method_idx, class_def_idx, dex_file) 852 || CanHandleVerificationFailure(verified_method)) { 853 ArenaAllocator arena(Runtime::Current()->GetArenaPool()); 854 CodeVectorAllocator code_allocator(&arena); 855 std::unique_ptr<CodeGenerator> codegen( 856 TryCompile(&arena, 857 &code_allocator, 858 code_item, 859 access_flags, 860 invoke_type, 861 class_def_idx, 862 method_idx, 863 jclass_loader, 864 dex_file, 865 dex_cache)); 866 if (codegen.get() != nullptr) { 867 if (run_optimizations_) { 868 method = EmitOptimized(&arena, &code_allocator, codegen.get(), compiler_driver); 869 } else { 870 method = EmitBaseline(&arena, &code_allocator, codegen.get(), compiler_driver); 871 } 872 } 873 } else { 874 if (compiler_driver->GetCompilerOptions().VerifyAtRuntime()) { 875 MaybeRecordStat(MethodCompilationStat::kNotCompiledVerifyAtRuntime); 876 } else { 877 MaybeRecordStat(MethodCompilationStat::kNotCompiledClassNotVerified); 878 } 879 } 880 881 if (kIsDebugBuild && 882 IsCompilingWithCoreImage() && 883 IsInstructionSetSupported(compiler_driver->GetInstructionSet()) && 884 (!kEmitCompilerReadBarrier || 885 InstructionSetSupportsReadBarrier(compiler_driver->GetInstructionSet()))) { 886 // For testing purposes, we put a special marker on method names 887 // that should be compiled with this compiler (when the the 888 // instruction set is supported -- and has support for read 889 // barriers, if they are enabled). This makes sure we're not 890 // regressing. 891 std::string method_name = PrettyMethod(method_idx, dex_file); 892 bool shouldCompile = method_name.find("$opt$") != std::string::npos; 893 DCHECK((method != nullptr) || !shouldCompile) << "Didn't compile " << method_name; 894 } 895 896 return method; 897} 898 899Compiler* CreateOptimizingCompiler(CompilerDriver* driver) { 900 return new OptimizingCompiler(driver); 901} 902 903bool IsCompilingWithCoreImage() { 904 const std::string& image = Runtime::Current()->GetImageLocation(); 905 return EndsWith(image, "core.art") || EndsWith(image, "core-optimizing.art"); 906} 907 908bool OptimizingCompiler::JitCompile(Thread* self, 909 jit::JitCodeCache* code_cache, 910 ArtMethod* method) { 911 StackHandleScope<2> hs(self); 912 Handle<mirror::ClassLoader> class_loader(hs.NewHandle( 913 method->GetDeclaringClass()->GetClassLoader())); 914 Handle<mirror::DexCache> dex_cache(hs.NewHandle(method->GetDexCache())); 915 916 jobject jclass_loader = class_loader.ToJObject(); 917 const DexFile* dex_file = method->GetDexFile(); 918 const uint16_t class_def_idx = method->GetClassDefIndex(); 919 const DexFile::CodeItem* code_item = dex_file->GetCodeItem(method->GetCodeItemOffset()); 920 const uint32_t method_idx = method->GetDexMethodIndex(); 921 const uint32_t access_flags = method->GetAccessFlags(); 922 const InvokeType invoke_type = method->GetInvokeType(); 923 924 ArenaAllocator arena(Runtime::Current()->GetArenaPool()); 925 CodeVectorAllocator code_allocator(&arena); 926 std::unique_ptr<CodeGenerator> codegen; 927 { 928 // Go to native so that we don't block GC during compilation. 929 ScopedThreadSuspension sts(self, kNative); 930 931 DCHECK(run_optimizations_); 932 codegen.reset( 933 TryCompile(&arena, 934 &code_allocator, 935 code_item, 936 access_flags, 937 invoke_type, 938 class_def_idx, 939 method_idx, 940 jclass_loader, 941 *dex_file, 942 dex_cache)); 943 if (codegen.get() == nullptr) { 944 return false; 945 } 946 } 947 948 size_t stack_map_size = codegen->ComputeStackMapsSize(); 949 uint8_t* stack_map_data = code_cache->ReserveData(self, stack_map_size); 950 if (stack_map_data == nullptr) { 951 return false; 952 } 953 codegen->BuildStackMaps(MemoryRegion(stack_map_data, stack_map_size)); 954 const void* code = code_cache->CommitCode( 955 self, 956 method, 957 nullptr, 958 stack_map_data, 959 nullptr, 960 codegen->HasEmptyFrame() ? 0 : codegen->GetFrameSize(), 961 codegen->GetCoreSpillMask(), 962 codegen->GetFpuSpillMask(), 963 code_allocator.GetMemory().data(), 964 code_allocator.GetSize()); 965 966 if (code == nullptr) { 967 code_cache->ClearData(self, stack_map_data); 968 return false; 969 } 970 971 return true; 972} 973 974} // namespace art 975