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 "code_generator.h" 18 19 #ifdef ART_ENABLE_CODEGEN_arm 20 #include "code_generator_arm_vixl.h" 21 #endif 22 23 #ifdef ART_ENABLE_CODEGEN_arm64 24 #include "code_generator_arm64.h" 25 #endif 26 27 #ifdef ART_ENABLE_CODEGEN_x86 28 #include "code_generator_x86.h" 29 #endif 30 31 #ifdef ART_ENABLE_CODEGEN_x86_64 32 #include "code_generator_x86_64.h" 33 #endif 34 35 #ifdef ART_ENABLE_CODEGEN_mips 36 #include "code_generator_mips.h" 37 #endif 38 39 #ifdef ART_ENABLE_CODEGEN_mips64 40 #include "code_generator_mips64.h" 41 #endif 42 43 #include "base/bit_utils.h" 44 #include "base/bit_utils_iterator.h" 45 #include "base/casts.h" 46 #include "base/leb128.h" 47 #include "class_linker.h" 48 #include "compiled_method.h" 49 #include "dex/bytecode_utils.h" 50 #include "dex/code_item_accessors-inl.h" 51 #include "dex/verified_method.h" 52 #include "graph_visualizer.h" 53 #include "image.h" 54 #include "gc/space/image_space.h" 55 #include "intern_table.h" 56 #include "intrinsics.h" 57 #include "mirror/array-inl.h" 58 #include "mirror/object_array-inl.h" 59 #include "mirror/object_reference.h" 60 #include "mirror/reference.h" 61 #include "mirror/string.h" 62 #include "parallel_move_resolver.h" 63 #include "scoped_thread_state_change-inl.h" 64 #include "ssa_liveness_analysis.h" 65 #include "stack_map.h" 66 #include "stack_map_stream.h" 67 #include "thread-current-inl.h" 68 #include "utils/assembler.h" 69 70 namespace art { 71 72 // Return whether a location is consistent with a type. 73 static bool CheckType(DataType::Type type, Location location) { 74 if (location.IsFpuRegister() 75 || (location.IsUnallocated() && (location.GetPolicy() == Location::kRequiresFpuRegister))) { 76 return (type == DataType::Type::kFloat32) || (type == DataType::Type::kFloat64); 77 } else if (location.IsRegister() || 78 (location.IsUnallocated() && (location.GetPolicy() == Location::kRequiresRegister))) { 79 return DataType::IsIntegralType(type) || (type == DataType::Type::kReference); 80 } else if (location.IsRegisterPair()) { 81 return type == DataType::Type::kInt64; 82 } else if (location.IsFpuRegisterPair()) { 83 return type == DataType::Type::kFloat64; 84 } else if (location.IsStackSlot()) { 85 return (DataType::IsIntegralType(type) && type != DataType::Type::kInt64) 86 || (type == DataType::Type::kFloat32) 87 || (type == DataType::Type::kReference); 88 } else if (location.IsDoubleStackSlot()) { 89 return (type == DataType::Type::kInt64) || (type == DataType::Type::kFloat64); 90 } else if (location.IsConstant()) { 91 if (location.GetConstant()->IsIntConstant()) { 92 return DataType::IsIntegralType(type) && (type != DataType::Type::kInt64); 93 } else if (location.GetConstant()->IsNullConstant()) { 94 return type == DataType::Type::kReference; 95 } else if (location.GetConstant()->IsLongConstant()) { 96 return type == DataType::Type::kInt64; 97 } else if (location.GetConstant()->IsFloatConstant()) { 98 return type == DataType::Type::kFloat32; 99 } else { 100 return location.GetConstant()->IsDoubleConstant() 101 && (type == DataType::Type::kFloat64); 102 } 103 } else { 104 return location.IsInvalid() || (location.GetPolicy() == Location::kAny); 105 } 106 } 107 108 // Check that a location summary is consistent with an instruction. 109 static bool CheckTypeConsistency(HInstruction* instruction) { 110 LocationSummary* locations = instruction->GetLocations(); 111 if (locations == nullptr) { 112 return true; 113 } 114 115 if (locations->Out().IsUnallocated() 116 && (locations->Out().GetPolicy() == Location::kSameAsFirstInput)) { 117 DCHECK(CheckType(instruction->GetType(), locations->InAt(0))) 118 << instruction->GetType() 119 << " " << locations->InAt(0); 120 } else { 121 DCHECK(CheckType(instruction->GetType(), locations->Out())) 122 << instruction->GetType() 123 << " " << locations->Out(); 124 } 125 126 HConstInputsRef inputs = instruction->GetInputs(); 127 for (size_t i = 0; i < inputs.size(); ++i) { 128 DCHECK(CheckType(inputs[i]->GetType(), locations->InAt(i))) 129 << inputs[i]->GetType() << " " << locations->InAt(i); 130 } 131 132 HEnvironment* environment = instruction->GetEnvironment(); 133 for (size_t i = 0; i < instruction->EnvironmentSize(); ++i) { 134 if (environment->GetInstructionAt(i) != nullptr) { 135 DataType::Type type = environment->GetInstructionAt(i)->GetType(); 136 DCHECK(CheckType(type, environment->GetLocationAt(i))) 137 << type << " " << environment->GetLocationAt(i); 138 } else { 139 DCHECK(environment->GetLocationAt(i).IsInvalid()) 140 << environment->GetLocationAt(i); 141 } 142 } 143 return true; 144 } 145 146 class CodeGenerator::CodeGenerationData : public DeletableArenaObject<kArenaAllocCodeGenerator> { 147 public: 148 static std::unique_ptr<CodeGenerationData> Create(ArenaStack* arena_stack, 149 InstructionSet instruction_set) { 150 ScopedArenaAllocator allocator(arena_stack); 151 void* memory = allocator.Alloc<CodeGenerationData>(kArenaAllocCodeGenerator); 152 return std::unique_ptr<CodeGenerationData>( 153 ::new (memory) CodeGenerationData(std::move(allocator), instruction_set)); 154 } 155 156 ScopedArenaAllocator* GetScopedAllocator() { 157 return &allocator_; 158 } 159 160 void AddSlowPath(SlowPathCode* slow_path) { 161 slow_paths_.emplace_back(std::unique_ptr<SlowPathCode>(slow_path)); 162 } 163 164 ArrayRef<const std::unique_ptr<SlowPathCode>> GetSlowPaths() const { 165 return ArrayRef<const std::unique_ptr<SlowPathCode>>(slow_paths_); 166 } 167 168 StackMapStream* GetStackMapStream() { return &stack_map_stream_; } 169 170 void ReserveJitStringRoot(StringReference string_reference, Handle<mirror::String> string) { 171 jit_string_roots_.Overwrite(string_reference, 172 reinterpret_cast64<uint64_t>(string.GetReference())); 173 } 174 175 uint64_t GetJitStringRootIndex(StringReference string_reference) const { 176 return jit_string_roots_.Get(string_reference); 177 } 178 179 size_t GetNumberOfJitStringRoots() const { 180 return jit_string_roots_.size(); 181 } 182 183 void ReserveJitClassRoot(TypeReference type_reference, Handle<mirror::Class> klass) { 184 jit_class_roots_.Overwrite(type_reference, reinterpret_cast64<uint64_t>(klass.GetReference())); 185 } 186 187 uint64_t GetJitClassRootIndex(TypeReference type_reference) const { 188 return jit_class_roots_.Get(type_reference); 189 } 190 191 size_t GetNumberOfJitClassRoots() const { 192 return jit_class_roots_.size(); 193 } 194 195 size_t GetNumberOfJitRoots() const { 196 return GetNumberOfJitStringRoots() + GetNumberOfJitClassRoots(); 197 } 198 199 void EmitJitRoots(/*out*/std::vector<Handle<mirror::Object>>* roots) 200 REQUIRES_SHARED(Locks::mutator_lock_); 201 202 private: 203 CodeGenerationData(ScopedArenaAllocator&& allocator, InstructionSet instruction_set) 204 : allocator_(std::move(allocator)), 205 stack_map_stream_(&allocator_, instruction_set), 206 slow_paths_(allocator_.Adapter(kArenaAllocCodeGenerator)), 207 jit_string_roots_(StringReferenceValueComparator(), 208 allocator_.Adapter(kArenaAllocCodeGenerator)), 209 jit_class_roots_(TypeReferenceValueComparator(), 210 allocator_.Adapter(kArenaAllocCodeGenerator)) { 211 slow_paths_.reserve(kDefaultSlowPathsCapacity); 212 } 213 214 static constexpr size_t kDefaultSlowPathsCapacity = 8; 215 216 ScopedArenaAllocator allocator_; 217 StackMapStream stack_map_stream_; 218 ScopedArenaVector<std::unique_ptr<SlowPathCode>> slow_paths_; 219 220 // Maps a StringReference (dex_file, string_index) to the index in the literal table. 221 // Entries are intially added with a pointer in the handle zone, and `EmitJitRoots` 222 // will compute all the indices. 223 ScopedArenaSafeMap<StringReference, uint64_t, StringReferenceValueComparator> jit_string_roots_; 224 225 // Maps a ClassReference (dex_file, type_index) to the index in the literal table. 226 // Entries are intially added with a pointer in the handle zone, and `EmitJitRoots` 227 // will compute all the indices. 228 ScopedArenaSafeMap<TypeReference, uint64_t, TypeReferenceValueComparator> jit_class_roots_; 229 }; 230 231 void CodeGenerator::CodeGenerationData::EmitJitRoots( 232 /*out*/std::vector<Handle<mirror::Object>>* roots) { 233 DCHECK(roots->empty()); 234 roots->reserve(GetNumberOfJitRoots()); 235 ClassLinker* class_linker = Runtime::Current()->GetClassLinker(); 236 size_t index = 0; 237 for (auto& entry : jit_string_roots_) { 238 // Update the `roots` with the string, and replace the address temporarily 239 // stored to the index in the table. 240 uint64_t address = entry.second; 241 roots->emplace_back(reinterpret_cast<StackReference<mirror::Object>*>(address)); 242 DCHECK(roots->back() != nullptr); 243 DCHECK(roots->back()->IsString()); 244 entry.second = index; 245 // Ensure the string is strongly interned. This is a requirement on how the JIT 246 // handles strings. b/32995596 247 class_linker->GetInternTable()->InternStrong(roots->back()->AsString()); 248 ++index; 249 } 250 for (auto& entry : jit_class_roots_) { 251 // Update the `roots` with the class, and replace the address temporarily 252 // stored to the index in the table. 253 uint64_t address = entry.second; 254 roots->emplace_back(reinterpret_cast<StackReference<mirror::Object>*>(address)); 255 DCHECK(roots->back() != nullptr); 256 DCHECK(roots->back()->IsClass()); 257 entry.second = index; 258 ++index; 259 } 260 } 261 262 ScopedArenaAllocator* CodeGenerator::GetScopedAllocator() { 263 DCHECK(code_generation_data_ != nullptr); 264 return code_generation_data_->GetScopedAllocator(); 265 } 266 267 StackMapStream* CodeGenerator::GetStackMapStream() { 268 DCHECK(code_generation_data_ != nullptr); 269 return code_generation_data_->GetStackMapStream(); 270 } 271 272 void CodeGenerator::ReserveJitStringRoot(StringReference string_reference, 273 Handle<mirror::String> string) { 274 DCHECK(code_generation_data_ != nullptr); 275 code_generation_data_->ReserveJitStringRoot(string_reference, string); 276 } 277 278 uint64_t CodeGenerator::GetJitStringRootIndex(StringReference string_reference) { 279 DCHECK(code_generation_data_ != nullptr); 280 return code_generation_data_->GetJitStringRootIndex(string_reference); 281 } 282 283 void CodeGenerator::ReserveJitClassRoot(TypeReference type_reference, Handle<mirror::Class> klass) { 284 DCHECK(code_generation_data_ != nullptr); 285 code_generation_data_->ReserveJitClassRoot(type_reference, klass); 286 } 287 288 uint64_t CodeGenerator::GetJitClassRootIndex(TypeReference type_reference) { 289 DCHECK(code_generation_data_ != nullptr); 290 return code_generation_data_->GetJitClassRootIndex(type_reference); 291 } 292 293 void CodeGenerator::EmitJitRootPatches(uint8_t* code ATTRIBUTE_UNUSED, 294 const uint8_t* roots_data ATTRIBUTE_UNUSED) { 295 DCHECK(code_generation_data_ != nullptr); 296 DCHECK_EQ(code_generation_data_->GetNumberOfJitStringRoots(), 0u); 297 DCHECK_EQ(code_generation_data_->GetNumberOfJitClassRoots(), 0u); 298 } 299 300 uint32_t CodeGenerator::GetArrayLengthOffset(HArrayLength* array_length) { 301 return array_length->IsStringLength() 302 ? mirror::String::CountOffset().Uint32Value() 303 : mirror::Array::LengthOffset().Uint32Value(); 304 } 305 306 uint32_t CodeGenerator::GetArrayDataOffset(HArrayGet* array_get) { 307 DCHECK(array_get->GetType() == DataType::Type::kUint16 || !array_get->IsStringCharAt()); 308 return array_get->IsStringCharAt() 309 ? mirror::String::ValueOffset().Uint32Value() 310 : mirror::Array::DataOffset(DataType::Size(array_get->GetType())).Uint32Value(); 311 } 312 313 bool CodeGenerator::GoesToNextBlock(HBasicBlock* current, HBasicBlock* next) const { 314 DCHECK_EQ((*block_order_)[current_block_index_], current); 315 return GetNextBlockToEmit() == FirstNonEmptyBlock(next); 316 } 317 318 HBasicBlock* CodeGenerator::GetNextBlockToEmit() const { 319 for (size_t i = current_block_index_ + 1; i < block_order_->size(); ++i) { 320 HBasicBlock* block = (*block_order_)[i]; 321 if (!block->IsSingleJump()) { 322 return block; 323 } 324 } 325 return nullptr; 326 } 327 328 HBasicBlock* CodeGenerator::FirstNonEmptyBlock(HBasicBlock* block) const { 329 while (block->IsSingleJump()) { 330 block = block->GetSuccessors()[0]; 331 } 332 return block; 333 } 334 335 class DisassemblyScope { 336 public: 337 DisassemblyScope(HInstruction* instruction, const CodeGenerator& codegen) 338 : codegen_(codegen), instruction_(instruction), start_offset_(static_cast<size_t>(-1)) { 339 if (codegen_.GetDisassemblyInformation() != nullptr) { 340 start_offset_ = codegen_.GetAssembler().CodeSize(); 341 } 342 } 343 344 ~DisassemblyScope() { 345 // We avoid building this data when we know it will not be used. 346 if (codegen_.GetDisassemblyInformation() != nullptr) { 347 codegen_.GetDisassemblyInformation()->AddInstructionInterval( 348 instruction_, start_offset_, codegen_.GetAssembler().CodeSize()); 349 } 350 } 351 352 private: 353 const CodeGenerator& codegen_; 354 HInstruction* instruction_; 355 size_t start_offset_; 356 }; 357 358 359 void CodeGenerator::GenerateSlowPaths() { 360 DCHECK(code_generation_data_ != nullptr); 361 size_t code_start = 0; 362 for (const std::unique_ptr<SlowPathCode>& slow_path_ptr : code_generation_data_->GetSlowPaths()) { 363 SlowPathCode* slow_path = slow_path_ptr.get(); 364 current_slow_path_ = slow_path; 365 if (disasm_info_ != nullptr) { 366 code_start = GetAssembler()->CodeSize(); 367 } 368 // Record the dex pc at start of slow path (required for java line number mapping). 369 MaybeRecordNativeDebugInfo(slow_path->GetInstruction(), slow_path->GetDexPc(), slow_path); 370 slow_path->EmitNativeCode(this); 371 if (disasm_info_ != nullptr) { 372 disasm_info_->AddSlowPathInterval(slow_path, code_start, GetAssembler()->CodeSize()); 373 } 374 } 375 current_slow_path_ = nullptr; 376 } 377 378 void CodeGenerator::InitializeCodeGenerationData() { 379 DCHECK(code_generation_data_ == nullptr); 380 code_generation_data_ = CodeGenerationData::Create(graph_->GetArenaStack(), GetInstructionSet()); 381 } 382 383 void CodeGenerator::Compile(CodeAllocator* allocator) { 384 InitializeCodeGenerationData(); 385 386 // The register allocator already called `InitializeCodeGeneration`, 387 // where the frame size has been computed. 388 DCHECK(block_order_ != nullptr); 389 Initialize(); 390 391 HGraphVisitor* instruction_visitor = GetInstructionVisitor(); 392 DCHECK_EQ(current_block_index_, 0u); 393 394 GetStackMapStream()->BeginMethod(HasEmptyFrame() ? 0 : frame_size_, 395 core_spill_mask_, 396 fpu_spill_mask_, 397 GetGraph()->GetNumberOfVRegs()); 398 399 size_t frame_start = GetAssembler()->CodeSize(); 400 GenerateFrameEntry(); 401 DCHECK_EQ(GetAssembler()->cfi().GetCurrentCFAOffset(), static_cast<int>(frame_size_)); 402 if (disasm_info_ != nullptr) { 403 disasm_info_->SetFrameEntryInterval(frame_start, GetAssembler()->CodeSize()); 404 } 405 406 for (size_t e = block_order_->size(); current_block_index_ < e; ++current_block_index_) { 407 HBasicBlock* block = (*block_order_)[current_block_index_]; 408 // Don't generate code for an empty block. Its predecessors will branch to its successor 409 // directly. Also, the label of that block will not be emitted, so this helps catch 410 // errors where we reference that label. 411 if (block->IsSingleJump()) continue; 412 Bind(block); 413 // This ensures that we have correct native line mapping for all native instructions. 414 // It is necessary to make stepping over a statement work. Otherwise, any initial 415 // instructions (e.g. moves) would be assumed to be the start of next statement. 416 MaybeRecordNativeDebugInfo(/* instruction= */ nullptr, block->GetDexPc()); 417 for (HInstructionIterator it(block->GetInstructions()); !it.Done(); it.Advance()) { 418 HInstruction* current = it.Current(); 419 if (current->HasEnvironment()) { 420 // Create stackmap for HNativeDebugInfo or any instruction which calls native code. 421 // Note that we need correct mapping for the native PC of the call instruction, 422 // so the runtime's stackmap is not sufficient since it is at PC after the call. 423 MaybeRecordNativeDebugInfo(current, block->GetDexPc()); 424 } 425 DisassemblyScope disassembly_scope(current, *this); 426 DCHECK(CheckTypeConsistency(current)); 427 current->Accept(instruction_visitor); 428 } 429 } 430 431 GenerateSlowPaths(); 432 433 // Emit catch stack maps at the end of the stack map stream as expected by the 434 // runtime exception handler. 435 if (graph_->HasTryCatch()) { 436 RecordCatchBlockInfo(); 437 } 438 439 // Finalize instructions in assember; 440 Finalize(allocator); 441 442 GetStackMapStream()->EndMethod(); 443 } 444 445 void CodeGenerator::Finalize(CodeAllocator* allocator) { 446 size_t code_size = GetAssembler()->CodeSize(); 447 uint8_t* buffer = allocator->Allocate(code_size); 448 449 MemoryRegion code(buffer, code_size); 450 GetAssembler()->FinalizeInstructions(code); 451 } 452 453 void CodeGenerator::EmitLinkerPatches( 454 ArenaVector<linker::LinkerPatch>* linker_patches ATTRIBUTE_UNUSED) { 455 // No linker patches by default. 456 } 457 458 bool CodeGenerator::NeedsThunkCode(const linker::LinkerPatch& patch ATTRIBUTE_UNUSED) const { 459 // Code generators that create patches requiring thunk compilation should override this function. 460 return false; 461 } 462 463 void CodeGenerator::EmitThunkCode(const linker::LinkerPatch& patch ATTRIBUTE_UNUSED, 464 /*out*/ ArenaVector<uint8_t>* code ATTRIBUTE_UNUSED, 465 /*out*/ std::string* debug_name ATTRIBUTE_UNUSED) { 466 // Code generators that create patches requiring thunk compilation should override this function. 467 LOG(FATAL) << "Unexpected call to EmitThunkCode()."; 468 } 469 470 void CodeGenerator::InitializeCodeGeneration(size_t number_of_spill_slots, 471 size_t maximum_safepoint_spill_size, 472 size_t number_of_out_slots, 473 const ArenaVector<HBasicBlock*>& block_order) { 474 block_order_ = &block_order; 475 DCHECK(!block_order.empty()); 476 DCHECK(block_order[0] == GetGraph()->GetEntryBlock()); 477 ComputeSpillMask(); 478 first_register_slot_in_slow_path_ = RoundUp( 479 (number_of_out_slots + number_of_spill_slots) * kVRegSize, GetPreferredSlotsAlignment()); 480 481 if (number_of_spill_slots == 0 482 && !HasAllocatedCalleeSaveRegisters() 483 && IsLeafMethod() 484 && !RequiresCurrentMethod()) { 485 DCHECK_EQ(maximum_safepoint_spill_size, 0u); 486 SetFrameSize(CallPushesPC() ? GetWordSize() : 0); 487 } else { 488 SetFrameSize(RoundUp( 489 first_register_slot_in_slow_path_ 490 + maximum_safepoint_spill_size 491 + (GetGraph()->HasShouldDeoptimizeFlag() ? kShouldDeoptimizeFlagSize : 0) 492 + FrameEntrySpillSize(), 493 kStackAlignment)); 494 } 495 } 496 497 void CodeGenerator::CreateCommonInvokeLocationSummary( 498 HInvoke* invoke, InvokeDexCallingConventionVisitor* visitor) { 499 ArenaAllocator* allocator = invoke->GetBlock()->GetGraph()->GetAllocator(); 500 LocationSummary* locations = new (allocator) LocationSummary(invoke, 501 LocationSummary::kCallOnMainOnly); 502 503 for (size_t i = 0; i < invoke->GetNumberOfArguments(); i++) { 504 HInstruction* input = invoke->InputAt(i); 505 locations->SetInAt(i, visitor->GetNextLocation(input->GetType())); 506 } 507 508 locations->SetOut(visitor->GetReturnLocation(invoke->GetType())); 509 510 if (invoke->IsInvokeStaticOrDirect()) { 511 HInvokeStaticOrDirect* call = invoke->AsInvokeStaticOrDirect(); 512 switch (call->GetMethodLoadKind()) { 513 case HInvokeStaticOrDirect::MethodLoadKind::kRecursive: 514 locations->SetInAt(call->GetSpecialInputIndex(), visitor->GetMethodLocation()); 515 break; 516 case HInvokeStaticOrDirect::MethodLoadKind::kRuntimeCall: 517 locations->AddTemp(visitor->GetMethodLocation()); 518 locations->SetInAt(call->GetSpecialInputIndex(), Location::RequiresRegister()); 519 break; 520 default: 521 locations->AddTemp(visitor->GetMethodLocation()); 522 break; 523 } 524 } else if (!invoke->IsInvokePolymorphic()) { 525 locations->AddTemp(visitor->GetMethodLocation()); 526 } 527 } 528 529 void CodeGenerator::GenerateInvokeStaticOrDirectRuntimeCall( 530 HInvokeStaticOrDirect* invoke, Location temp, SlowPathCode* slow_path) { 531 MoveConstant(temp, invoke->GetDexMethodIndex()); 532 533 // The access check is unnecessary but we do not want to introduce 534 // extra entrypoints for the codegens that do not support some 535 // invoke type and fall back to the runtime call. 536 537 // Initialize to anything to silent compiler warnings. 538 QuickEntrypointEnum entrypoint = kQuickInvokeStaticTrampolineWithAccessCheck; 539 switch (invoke->GetInvokeType()) { 540 case kStatic: 541 entrypoint = kQuickInvokeStaticTrampolineWithAccessCheck; 542 break; 543 case kDirect: 544 entrypoint = kQuickInvokeDirectTrampolineWithAccessCheck; 545 break; 546 case kSuper: 547 entrypoint = kQuickInvokeSuperTrampolineWithAccessCheck; 548 break; 549 case kVirtual: 550 case kInterface: 551 case kPolymorphic: 552 case kCustom: 553 LOG(FATAL) << "Unexpected invoke type: " << invoke->GetInvokeType(); 554 UNREACHABLE(); 555 } 556 557 InvokeRuntime(entrypoint, invoke, invoke->GetDexPc(), slow_path); 558 } 559 void CodeGenerator::GenerateInvokeUnresolvedRuntimeCall(HInvokeUnresolved* invoke) { 560 MoveConstant(invoke->GetLocations()->GetTemp(0), invoke->GetDexMethodIndex()); 561 562 // Initialize to anything to silent compiler warnings. 563 QuickEntrypointEnum entrypoint = kQuickInvokeStaticTrampolineWithAccessCheck; 564 switch (invoke->GetInvokeType()) { 565 case kStatic: 566 entrypoint = kQuickInvokeStaticTrampolineWithAccessCheck; 567 break; 568 case kDirect: 569 entrypoint = kQuickInvokeDirectTrampolineWithAccessCheck; 570 break; 571 case kVirtual: 572 entrypoint = kQuickInvokeVirtualTrampolineWithAccessCheck; 573 break; 574 case kSuper: 575 entrypoint = kQuickInvokeSuperTrampolineWithAccessCheck; 576 break; 577 case kInterface: 578 entrypoint = kQuickInvokeInterfaceTrampolineWithAccessCheck; 579 break; 580 case kPolymorphic: 581 case kCustom: 582 LOG(FATAL) << "Unexpected invoke type: " << invoke->GetInvokeType(); 583 UNREACHABLE(); 584 } 585 InvokeRuntime(entrypoint, invoke, invoke->GetDexPc(), nullptr); 586 } 587 588 void CodeGenerator::GenerateInvokePolymorphicCall(HInvokePolymorphic* invoke) { 589 // invoke-polymorphic does not use a temporary to convey any additional information (e.g. a 590 // method index) since it requires multiple info from the instruction (registers A, B, H). Not 591 // using the reservation has no effect on the registers used in the runtime call. 592 QuickEntrypointEnum entrypoint = kQuickInvokePolymorphic; 593 InvokeRuntime(entrypoint, invoke, invoke->GetDexPc(), nullptr); 594 } 595 596 void CodeGenerator::GenerateInvokeCustomCall(HInvokeCustom* invoke) { 597 MoveConstant(invoke->GetLocations()->GetTemp(0), invoke->GetCallSiteIndex()); 598 QuickEntrypointEnum entrypoint = kQuickInvokeCustom; 599 InvokeRuntime(entrypoint, invoke, invoke->GetDexPc(), nullptr); 600 } 601 602 void CodeGenerator::CreateUnresolvedFieldLocationSummary( 603 HInstruction* field_access, 604 DataType::Type field_type, 605 const FieldAccessCallingConvention& calling_convention) { 606 bool is_instance = field_access->IsUnresolvedInstanceFieldGet() 607 || field_access->IsUnresolvedInstanceFieldSet(); 608 bool is_get = field_access->IsUnresolvedInstanceFieldGet() 609 || field_access->IsUnresolvedStaticFieldGet(); 610 611 ArenaAllocator* allocator = field_access->GetBlock()->GetGraph()->GetAllocator(); 612 LocationSummary* locations = 613 new (allocator) LocationSummary(field_access, LocationSummary::kCallOnMainOnly); 614 615 locations->AddTemp(calling_convention.GetFieldIndexLocation()); 616 617 if (is_instance) { 618 // Add the `this` object for instance field accesses. 619 locations->SetInAt(0, calling_convention.GetObjectLocation()); 620 } 621 622 // Note that pSetXXStatic/pGetXXStatic always takes/returns an int or int64 623 // regardless of the the type. Because of that we forced to special case 624 // the access to floating point values. 625 if (is_get) { 626 if (DataType::IsFloatingPointType(field_type)) { 627 // The return value will be stored in regular registers while register 628 // allocator expects it in a floating point register. 629 // Note We don't need to request additional temps because the return 630 // register(s) are already blocked due the call and they may overlap with 631 // the input or field index. 632 // The transfer between the two will be done at codegen level. 633 locations->SetOut(calling_convention.GetFpuLocation(field_type)); 634 } else { 635 locations->SetOut(calling_convention.GetReturnLocation(field_type)); 636 } 637 } else { 638 size_t set_index = is_instance ? 1 : 0; 639 if (DataType::IsFloatingPointType(field_type)) { 640 // The set value comes from a float location while the calling convention 641 // expects it in a regular register location. Allocate a temp for it and 642 // make the transfer at codegen. 643 AddLocationAsTemp(calling_convention.GetSetValueLocation(field_type, is_instance), locations); 644 locations->SetInAt(set_index, calling_convention.GetFpuLocation(field_type)); 645 } else { 646 locations->SetInAt(set_index, 647 calling_convention.GetSetValueLocation(field_type, is_instance)); 648 } 649 } 650 } 651 652 void CodeGenerator::GenerateUnresolvedFieldAccess( 653 HInstruction* field_access, 654 DataType::Type field_type, 655 uint32_t field_index, 656 uint32_t dex_pc, 657 const FieldAccessCallingConvention& calling_convention) { 658 LocationSummary* locations = field_access->GetLocations(); 659 660 MoveConstant(locations->GetTemp(0), field_index); 661 662 bool is_instance = field_access->IsUnresolvedInstanceFieldGet() 663 || field_access->IsUnresolvedInstanceFieldSet(); 664 bool is_get = field_access->IsUnresolvedInstanceFieldGet() 665 || field_access->IsUnresolvedStaticFieldGet(); 666 667 if (!is_get && DataType::IsFloatingPointType(field_type)) { 668 // Copy the float value to be set into the calling convention register. 669 // Note that using directly the temp location is problematic as we don't 670 // support temp register pairs. To avoid boilerplate conversion code, use 671 // the location from the calling convention. 672 MoveLocation(calling_convention.GetSetValueLocation(field_type, is_instance), 673 locations->InAt(is_instance ? 1 : 0), 674 (DataType::Is64BitType(field_type) ? DataType::Type::kInt64 675 : DataType::Type::kInt32)); 676 } 677 678 QuickEntrypointEnum entrypoint = kQuickSet8Static; // Initialize to anything to avoid warnings. 679 switch (field_type) { 680 case DataType::Type::kBool: 681 entrypoint = is_instance 682 ? (is_get ? kQuickGetBooleanInstance : kQuickSet8Instance) 683 : (is_get ? kQuickGetBooleanStatic : kQuickSet8Static); 684 break; 685 case DataType::Type::kInt8: 686 entrypoint = is_instance 687 ? (is_get ? kQuickGetByteInstance : kQuickSet8Instance) 688 : (is_get ? kQuickGetByteStatic : kQuickSet8Static); 689 break; 690 case DataType::Type::kInt16: 691 entrypoint = is_instance 692 ? (is_get ? kQuickGetShortInstance : kQuickSet16Instance) 693 : (is_get ? kQuickGetShortStatic : kQuickSet16Static); 694 break; 695 case DataType::Type::kUint16: 696 entrypoint = is_instance 697 ? (is_get ? kQuickGetCharInstance : kQuickSet16Instance) 698 : (is_get ? kQuickGetCharStatic : kQuickSet16Static); 699 break; 700 case DataType::Type::kInt32: 701 case DataType::Type::kFloat32: 702 entrypoint = is_instance 703 ? (is_get ? kQuickGet32Instance : kQuickSet32Instance) 704 : (is_get ? kQuickGet32Static : kQuickSet32Static); 705 break; 706 case DataType::Type::kReference: 707 entrypoint = is_instance 708 ? (is_get ? kQuickGetObjInstance : kQuickSetObjInstance) 709 : (is_get ? kQuickGetObjStatic : kQuickSetObjStatic); 710 break; 711 case DataType::Type::kInt64: 712 case DataType::Type::kFloat64: 713 entrypoint = is_instance 714 ? (is_get ? kQuickGet64Instance : kQuickSet64Instance) 715 : (is_get ? kQuickGet64Static : kQuickSet64Static); 716 break; 717 default: 718 LOG(FATAL) << "Invalid type " << field_type; 719 } 720 InvokeRuntime(entrypoint, field_access, dex_pc, nullptr); 721 722 if (is_get && DataType::IsFloatingPointType(field_type)) { 723 MoveLocation(locations->Out(), calling_convention.GetReturnLocation(field_type), field_type); 724 } 725 } 726 727 void CodeGenerator::CreateLoadClassRuntimeCallLocationSummary(HLoadClass* cls, 728 Location runtime_type_index_location, 729 Location runtime_return_location) { 730 DCHECK_EQ(cls->GetLoadKind(), HLoadClass::LoadKind::kRuntimeCall); 731 DCHECK_EQ(cls->InputCount(), 1u); 732 LocationSummary* locations = new (cls->GetBlock()->GetGraph()->GetAllocator()) LocationSummary( 733 cls, LocationSummary::kCallOnMainOnly); 734 locations->SetInAt(0, Location::NoLocation()); 735 locations->AddTemp(runtime_type_index_location); 736 locations->SetOut(runtime_return_location); 737 } 738 739 void CodeGenerator::GenerateLoadClassRuntimeCall(HLoadClass* cls) { 740 DCHECK_EQ(cls->GetLoadKind(), HLoadClass::LoadKind::kRuntimeCall); 741 DCHECK(!cls->MustGenerateClinitCheck()); 742 LocationSummary* locations = cls->GetLocations(); 743 MoveConstant(locations->GetTemp(0), cls->GetTypeIndex().index_); 744 if (cls->NeedsAccessCheck()) { 745 CheckEntrypointTypes<kQuickResolveTypeAndVerifyAccess, void*, uint32_t>(); 746 InvokeRuntime(kQuickResolveTypeAndVerifyAccess, cls, cls->GetDexPc()); 747 } else { 748 CheckEntrypointTypes<kQuickResolveType, void*, uint32_t>(); 749 InvokeRuntime(kQuickResolveType, cls, cls->GetDexPc()); 750 } 751 } 752 753 void CodeGenerator::CreateLoadMethodHandleRuntimeCallLocationSummary( 754 HLoadMethodHandle* method_handle, 755 Location runtime_proto_index_location, 756 Location runtime_return_location) { 757 DCHECK_EQ(method_handle->InputCount(), 1u); 758 LocationSummary* locations = 759 new (method_handle->GetBlock()->GetGraph()->GetAllocator()) LocationSummary( 760 method_handle, LocationSummary::kCallOnMainOnly); 761 locations->SetInAt(0, Location::NoLocation()); 762 locations->AddTemp(runtime_proto_index_location); 763 locations->SetOut(runtime_return_location); 764 } 765 766 void CodeGenerator::GenerateLoadMethodHandleRuntimeCall(HLoadMethodHandle* method_handle) { 767 LocationSummary* locations = method_handle->GetLocations(); 768 MoveConstant(locations->GetTemp(0), method_handle->GetMethodHandleIndex()); 769 CheckEntrypointTypes<kQuickResolveMethodHandle, void*, uint32_t>(); 770 InvokeRuntime(kQuickResolveMethodHandle, method_handle, method_handle->GetDexPc()); 771 } 772 773 void CodeGenerator::CreateLoadMethodTypeRuntimeCallLocationSummary( 774 HLoadMethodType* method_type, 775 Location runtime_proto_index_location, 776 Location runtime_return_location) { 777 DCHECK_EQ(method_type->InputCount(), 1u); 778 LocationSummary* locations = 779 new (method_type->GetBlock()->GetGraph()->GetAllocator()) LocationSummary( 780 method_type, LocationSummary::kCallOnMainOnly); 781 locations->SetInAt(0, Location::NoLocation()); 782 locations->AddTemp(runtime_proto_index_location); 783 locations->SetOut(runtime_return_location); 784 } 785 786 void CodeGenerator::GenerateLoadMethodTypeRuntimeCall(HLoadMethodType* method_type) { 787 LocationSummary* locations = method_type->GetLocations(); 788 MoveConstant(locations->GetTemp(0), method_type->GetProtoIndex().index_); 789 CheckEntrypointTypes<kQuickResolveMethodType, void*, uint32_t>(); 790 InvokeRuntime(kQuickResolveMethodType, method_type, method_type->GetDexPc()); 791 } 792 793 static uint32_t GetBootImageOffsetImpl(const void* object, ImageHeader::ImageSections section) { 794 Runtime* runtime = Runtime::Current(); 795 DCHECK(runtime->IsAotCompiler()); 796 const std::vector<gc::space::ImageSpace*>& boot_image_spaces = 797 runtime->GetHeap()->GetBootImageSpaces(); 798 // Check that the `object` is in the expected section of one of the boot image files. 799 DCHECK(std::any_of(boot_image_spaces.begin(), 800 boot_image_spaces.end(), 801 [object, section](gc::space::ImageSpace* space) { 802 uintptr_t begin = reinterpret_cast<uintptr_t>(space->Begin()); 803 uintptr_t offset = reinterpret_cast<uintptr_t>(object) - begin; 804 return space->GetImageHeader().GetImageSection(section).Contains(offset); 805 })); 806 uintptr_t begin = reinterpret_cast<uintptr_t>(boot_image_spaces.front()->Begin()); 807 uintptr_t offset = reinterpret_cast<uintptr_t>(object) - begin; 808 return dchecked_integral_cast<uint32_t>(offset); 809 } 810 811 // NO_THREAD_SAFETY_ANALYSIS: Avoid taking the mutator lock, boot image classes are non-moveable. 812 uint32_t CodeGenerator::GetBootImageOffset(HLoadClass* load_class) NO_THREAD_SAFETY_ANALYSIS { 813 DCHECK_EQ(load_class->GetLoadKind(), HLoadClass::LoadKind::kBootImageRelRo); 814 ObjPtr<mirror::Class> klass = load_class->GetClass().Get(); 815 DCHECK(klass != nullptr); 816 return GetBootImageOffsetImpl(klass.Ptr(), ImageHeader::kSectionObjects); 817 } 818 819 // NO_THREAD_SAFETY_ANALYSIS: Avoid taking the mutator lock, boot image strings are non-moveable. 820 uint32_t CodeGenerator::GetBootImageOffset(HLoadString* load_string) NO_THREAD_SAFETY_ANALYSIS { 821 DCHECK_EQ(load_string->GetLoadKind(), HLoadString::LoadKind::kBootImageRelRo); 822 ObjPtr<mirror::String> string = load_string->GetString().Get(); 823 DCHECK(string != nullptr); 824 return GetBootImageOffsetImpl(string.Ptr(), ImageHeader::kSectionObjects); 825 } 826 827 uint32_t CodeGenerator::GetBootImageOffset(HInvokeStaticOrDirect* invoke) { 828 DCHECK_EQ(invoke->GetMethodLoadKind(), HInvokeStaticOrDirect::MethodLoadKind::kBootImageRelRo); 829 ArtMethod* method = invoke->GetResolvedMethod(); 830 DCHECK(method != nullptr); 831 return GetBootImageOffsetImpl(method, ImageHeader::kSectionArtMethods); 832 } 833 834 void CodeGenerator::BlockIfInRegister(Location location, bool is_out) const { 835 // The DCHECKS below check that a register is not specified twice in 836 // the summary. The out location can overlap with an input, so we need 837 // to special case it. 838 if (location.IsRegister()) { 839 DCHECK(is_out || !blocked_core_registers_[location.reg()]); 840 blocked_core_registers_[location.reg()] = true; 841 } else if (location.IsFpuRegister()) { 842 DCHECK(is_out || !blocked_fpu_registers_[location.reg()]); 843 blocked_fpu_registers_[location.reg()] = true; 844 } else if (location.IsFpuRegisterPair()) { 845 DCHECK(is_out || !blocked_fpu_registers_[location.AsFpuRegisterPairLow<int>()]); 846 blocked_fpu_registers_[location.AsFpuRegisterPairLow<int>()] = true; 847 DCHECK(is_out || !blocked_fpu_registers_[location.AsFpuRegisterPairHigh<int>()]); 848 blocked_fpu_registers_[location.AsFpuRegisterPairHigh<int>()] = true; 849 } else if (location.IsRegisterPair()) { 850 DCHECK(is_out || !blocked_core_registers_[location.AsRegisterPairLow<int>()]); 851 blocked_core_registers_[location.AsRegisterPairLow<int>()] = true; 852 DCHECK(is_out || !blocked_core_registers_[location.AsRegisterPairHigh<int>()]); 853 blocked_core_registers_[location.AsRegisterPairHigh<int>()] = true; 854 } 855 } 856 857 void CodeGenerator::AllocateLocations(HInstruction* instruction) { 858 for (HEnvironment* env = instruction->GetEnvironment(); env != nullptr; env = env->GetParent()) { 859 env->AllocateLocations(); 860 } 861 instruction->Accept(GetLocationBuilder()); 862 DCHECK(CheckTypeConsistency(instruction)); 863 LocationSummary* locations = instruction->GetLocations(); 864 if (!instruction->IsSuspendCheckEntry()) { 865 if (locations != nullptr) { 866 if (locations->CanCall()) { 867 MarkNotLeaf(); 868 } else if (locations->Intrinsified() && 869 instruction->IsInvokeStaticOrDirect() && 870 !instruction->AsInvokeStaticOrDirect()->HasCurrentMethodInput()) { 871 // A static method call that has been fully intrinsified, and cannot call on the slow 872 // path or refer to the current method directly, no longer needs current method. 873 return; 874 } 875 } 876 if (instruction->NeedsCurrentMethod()) { 877 SetRequiresCurrentMethod(); 878 } 879 } 880 } 881 882 std::unique_ptr<CodeGenerator> CodeGenerator::Create(HGraph* graph, 883 const CompilerOptions& compiler_options, 884 OptimizingCompilerStats* stats) { 885 ArenaAllocator* allocator = graph->GetAllocator(); 886 switch (compiler_options.GetInstructionSet()) { 887 #ifdef ART_ENABLE_CODEGEN_arm 888 case InstructionSet::kArm: 889 case InstructionSet::kThumb2: { 890 return std::unique_ptr<CodeGenerator>( 891 new (allocator) arm::CodeGeneratorARMVIXL(graph, compiler_options, stats)); 892 } 893 #endif 894 #ifdef ART_ENABLE_CODEGEN_arm64 895 case InstructionSet::kArm64: { 896 return std::unique_ptr<CodeGenerator>( 897 new (allocator) arm64::CodeGeneratorARM64(graph, compiler_options, stats)); 898 } 899 #endif 900 #ifdef ART_ENABLE_CODEGEN_mips 901 case InstructionSet::kMips: { 902 return std::unique_ptr<CodeGenerator>( 903 new (allocator) mips::CodeGeneratorMIPS(graph, compiler_options, stats)); 904 } 905 #endif 906 #ifdef ART_ENABLE_CODEGEN_mips64 907 case InstructionSet::kMips64: { 908 return std::unique_ptr<CodeGenerator>( 909 new (allocator) mips64::CodeGeneratorMIPS64(graph, compiler_options, stats)); 910 } 911 #endif 912 #ifdef ART_ENABLE_CODEGEN_x86 913 case InstructionSet::kX86: { 914 return std::unique_ptr<CodeGenerator>( 915 new (allocator) x86::CodeGeneratorX86(graph, compiler_options, stats)); 916 } 917 #endif 918 #ifdef ART_ENABLE_CODEGEN_x86_64 919 case InstructionSet::kX86_64: { 920 return std::unique_ptr<CodeGenerator>( 921 new (allocator) x86_64::CodeGeneratorX86_64(graph, compiler_options, stats)); 922 } 923 #endif 924 default: 925 return nullptr; 926 } 927 } 928 929 CodeGenerator::CodeGenerator(HGraph* graph, 930 size_t number_of_core_registers, 931 size_t number_of_fpu_registers, 932 size_t number_of_register_pairs, 933 uint32_t core_callee_save_mask, 934 uint32_t fpu_callee_save_mask, 935 const CompilerOptions& compiler_options, 936 OptimizingCompilerStats* stats) 937 : frame_size_(0), 938 core_spill_mask_(0), 939 fpu_spill_mask_(0), 940 first_register_slot_in_slow_path_(0), 941 allocated_registers_(RegisterSet::Empty()), 942 blocked_core_registers_(graph->GetAllocator()->AllocArray<bool>(number_of_core_registers, 943 kArenaAllocCodeGenerator)), 944 blocked_fpu_registers_(graph->GetAllocator()->AllocArray<bool>(number_of_fpu_registers, 945 kArenaAllocCodeGenerator)), 946 number_of_core_registers_(number_of_core_registers), 947 number_of_fpu_registers_(number_of_fpu_registers), 948 number_of_register_pairs_(number_of_register_pairs), 949 core_callee_save_mask_(core_callee_save_mask), 950 fpu_callee_save_mask_(fpu_callee_save_mask), 951 block_order_(nullptr), 952 disasm_info_(nullptr), 953 stats_(stats), 954 graph_(graph), 955 compiler_options_(compiler_options), 956 current_slow_path_(nullptr), 957 current_block_index_(0), 958 is_leaf_(true), 959 requires_current_method_(false), 960 code_generation_data_() { 961 } 962 963 CodeGenerator::~CodeGenerator() {} 964 965 size_t CodeGenerator::GetNumberOfJitRoots() const { 966 DCHECK(code_generation_data_ != nullptr); 967 return code_generation_data_->GetNumberOfJitRoots(); 968 } 969 970 static void CheckCovers(uint32_t dex_pc, 971 const HGraph& graph, 972 const CodeInfo& code_info, 973 const ArenaVector<HSuspendCheck*>& loop_headers, 974 ArenaVector<size_t>* covered) { 975 for (size_t i = 0; i < loop_headers.size(); ++i) { 976 if (loop_headers[i]->GetDexPc() == dex_pc) { 977 if (graph.IsCompilingOsr()) { 978 DCHECK(code_info.GetOsrStackMapForDexPc(dex_pc).IsValid()); 979 } 980 ++(*covered)[i]; 981 } 982 } 983 } 984 985 // Debug helper to ensure loop entries in compiled code are matched by 986 // dex branch instructions. 987 static void CheckLoopEntriesCanBeUsedForOsr(const HGraph& graph, 988 const CodeInfo& code_info, 989 const dex::CodeItem& code_item) { 990 if (graph.HasTryCatch()) { 991 // One can write loops through try/catch, which we do not support for OSR anyway. 992 return; 993 } 994 ArenaVector<HSuspendCheck*> loop_headers(graph.GetAllocator()->Adapter(kArenaAllocMisc)); 995 for (HBasicBlock* block : graph.GetReversePostOrder()) { 996 if (block->IsLoopHeader()) { 997 HSuspendCheck* suspend_check = block->GetLoopInformation()->GetSuspendCheck(); 998 if (!suspend_check->GetEnvironment()->IsFromInlinedInvoke()) { 999 loop_headers.push_back(suspend_check); 1000 } 1001 } 1002 } 1003 ArenaVector<size_t> covered( 1004 loop_headers.size(), 0, graph.GetAllocator()->Adapter(kArenaAllocMisc)); 1005 for (const DexInstructionPcPair& pair : CodeItemInstructionAccessor(graph.GetDexFile(), 1006 &code_item)) { 1007 const uint32_t dex_pc = pair.DexPc(); 1008 const Instruction& instruction = pair.Inst(); 1009 if (instruction.IsBranch()) { 1010 uint32_t target = dex_pc + instruction.GetTargetOffset(); 1011 CheckCovers(target, graph, code_info, loop_headers, &covered); 1012 } else if (instruction.IsSwitch()) { 1013 DexSwitchTable table(instruction, dex_pc); 1014 uint16_t num_entries = table.GetNumEntries(); 1015 size_t offset = table.GetFirstValueIndex(); 1016 1017 // Use a larger loop counter type to avoid overflow issues. 1018 for (size_t i = 0; i < num_entries; ++i) { 1019 // The target of the case. 1020 uint32_t target = dex_pc + table.GetEntryAt(i + offset); 1021 CheckCovers(target, graph, code_info, loop_headers, &covered); 1022 } 1023 } 1024 } 1025 1026 for (size_t i = 0; i < covered.size(); ++i) { 1027 DCHECK_NE(covered[i], 0u) << "Loop in compiled code has no dex branch equivalent"; 1028 } 1029 } 1030 1031 ScopedArenaVector<uint8_t> CodeGenerator::BuildStackMaps(const dex::CodeItem* code_item) { 1032 ScopedArenaVector<uint8_t> stack_map = GetStackMapStream()->Encode(); 1033 if (kIsDebugBuild && code_item != nullptr) { 1034 CheckLoopEntriesCanBeUsedForOsr(*graph_, CodeInfo(stack_map.data()), *code_item); 1035 } 1036 return stack_map; 1037 } 1038 1039 void CodeGenerator::RecordPcInfo(HInstruction* instruction, 1040 uint32_t dex_pc, 1041 SlowPathCode* slow_path, 1042 bool native_debug_info) { 1043 if (instruction != nullptr) { 1044 // The code generated for some type conversions 1045 // may call the runtime, thus normally requiring a subsequent 1046 // call to this method. However, the method verifier does not 1047 // produce PC information for certain instructions, which are 1048 // considered "atomic" (they cannot join a GC). 1049 // Therefore we do not currently record PC information for such 1050 // instructions. As this may change later, we added this special 1051 // case so that code generators may nevertheless call 1052 // CodeGenerator::RecordPcInfo without triggering an error in 1053 // CodeGenerator::BuildNativeGCMap ("Missing ref for dex pc 0x") 1054 // thereafter. 1055 if (instruction->IsTypeConversion()) { 1056 return; 1057 } 1058 if (instruction->IsRem()) { 1059 DataType::Type type = instruction->AsRem()->GetResultType(); 1060 if ((type == DataType::Type::kFloat32) || (type == DataType::Type::kFloat64)) { 1061 return; 1062 } 1063 } 1064 } 1065 1066 // Collect PC infos for the mapping table. 1067 uint32_t native_pc = GetAssembler()->CodePosition(); 1068 1069 StackMapStream* stack_map_stream = GetStackMapStream(); 1070 if (instruction == nullptr) { 1071 // For stack overflow checks and native-debug-info entries without dex register 1072 // mapping (i.e. start of basic block or start of slow path). 1073 stack_map_stream->BeginStackMapEntry(dex_pc, native_pc); 1074 stack_map_stream->EndStackMapEntry(); 1075 return; 1076 } 1077 1078 LocationSummary* locations = instruction->GetLocations(); 1079 uint32_t register_mask = locations->GetRegisterMask(); 1080 DCHECK_EQ(register_mask & ~locations->GetLiveRegisters()->GetCoreRegisters(), 0u); 1081 if (locations->OnlyCallsOnSlowPath()) { 1082 // In case of slow path, we currently set the location of caller-save registers 1083 // to register (instead of their stack location when pushed before the slow-path 1084 // call). Therefore register_mask contains both callee-save and caller-save 1085 // registers that hold objects. We must remove the spilled caller-save from the 1086 // mask, since they will be overwritten by the callee. 1087 uint32_t spills = GetSlowPathSpills(locations, /* core_registers= */ true); 1088 register_mask &= ~spills; 1089 } else { 1090 // The register mask must be a subset of callee-save registers. 1091 DCHECK_EQ(register_mask & core_callee_save_mask_, register_mask); 1092 } 1093 1094 uint32_t outer_dex_pc = dex_pc; 1095 uint32_t outer_environment_size = 0u; 1096 uint32_t inlining_depth = 0; 1097 HEnvironment* const environment = instruction->GetEnvironment(); 1098 if (environment != nullptr) { 1099 HEnvironment* outer_environment = environment; 1100 while (outer_environment->GetParent() != nullptr) { 1101 outer_environment = outer_environment->GetParent(); 1102 ++inlining_depth; 1103 } 1104 outer_dex_pc = outer_environment->GetDexPc(); 1105 outer_environment_size = outer_environment->Size(); 1106 } 1107 1108 HLoopInformation* info = instruction->GetBlock()->GetLoopInformation(); 1109 bool osr = 1110 instruction->IsSuspendCheck() && 1111 (info != nullptr) && 1112 graph_->IsCompilingOsr() && 1113 (inlining_depth == 0); 1114 StackMap::Kind kind = native_debug_info 1115 ? StackMap::Kind::Debug 1116 : (osr ? StackMap::Kind::OSR : StackMap::Kind::Default); 1117 stack_map_stream->BeginStackMapEntry(outer_dex_pc, 1118 native_pc, 1119 register_mask, 1120 locations->GetStackMask(), 1121 kind); 1122 EmitEnvironment(environment, slow_path); 1123 stack_map_stream->EndStackMapEntry(); 1124 1125 if (osr) { 1126 DCHECK_EQ(info->GetSuspendCheck(), instruction); 1127 DCHECK(info->IsIrreducible()); 1128 DCHECK(environment != nullptr); 1129 if (kIsDebugBuild) { 1130 for (size_t i = 0, environment_size = environment->Size(); i < environment_size; ++i) { 1131 HInstruction* in_environment = environment->GetInstructionAt(i); 1132 if (in_environment != nullptr) { 1133 DCHECK(in_environment->IsPhi() || in_environment->IsConstant()); 1134 Location location = environment->GetLocationAt(i); 1135 DCHECK(location.IsStackSlot() || 1136 location.IsDoubleStackSlot() || 1137 location.IsConstant() || 1138 location.IsInvalid()); 1139 if (location.IsStackSlot() || location.IsDoubleStackSlot()) { 1140 DCHECK_LT(location.GetStackIndex(), static_cast<int32_t>(GetFrameSize())); 1141 } 1142 } 1143 } 1144 } 1145 } 1146 } 1147 1148 bool CodeGenerator::HasStackMapAtCurrentPc() { 1149 uint32_t pc = GetAssembler()->CodeSize(); 1150 StackMapStream* stack_map_stream = GetStackMapStream(); 1151 size_t count = stack_map_stream->GetNumberOfStackMaps(); 1152 if (count == 0) { 1153 return false; 1154 } 1155 return stack_map_stream->GetStackMapNativePcOffset(count - 1) == pc; 1156 } 1157 1158 void CodeGenerator::MaybeRecordNativeDebugInfo(HInstruction* instruction, 1159 uint32_t dex_pc, 1160 SlowPathCode* slow_path) { 1161 if (GetCompilerOptions().GetNativeDebuggable() && dex_pc != kNoDexPc) { 1162 if (HasStackMapAtCurrentPc()) { 1163 // Ensure that we do not collide with the stack map of the previous instruction. 1164 GenerateNop(); 1165 } 1166 RecordPcInfo(instruction, dex_pc, slow_path, /* native_debug_info= */ true); 1167 } 1168 } 1169 1170 void CodeGenerator::RecordCatchBlockInfo() { 1171 StackMapStream* stack_map_stream = GetStackMapStream(); 1172 1173 for (HBasicBlock* block : *block_order_) { 1174 if (!block->IsCatchBlock()) { 1175 continue; 1176 } 1177 1178 uint32_t dex_pc = block->GetDexPc(); 1179 uint32_t num_vregs = graph_->GetNumberOfVRegs(); 1180 uint32_t native_pc = GetAddressOf(block); 1181 1182 stack_map_stream->BeginStackMapEntry(dex_pc, 1183 native_pc, 1184 /* register_mask= */ 0, 1185 /* sp_mask= */ nullptr, 1186 StackMap::Kind::Catch); 1187 1188 HInstruction* current_phi = block->GetFirstPhi(); 1189 for (size_t vreg = 0; vreg < num_vregs; ++vreg) { 1190 while (current_phi != nullptr && current_phi->AsPhi()->GetRegNumber() < vreg) { 1191 HInstruction* next_phi = current_phi->GetNext(); 1192 DCHECK(next_phi == nullptr || 1193 current_phi->AsPhi()->GetRegNumber() <= next_phi->AsPhi()->GetRegNumber()) 1194 << "Phis need to be sorted by vreg number to keep this a linear-time loop."; 1195 current_phi = next_phi; 1196 } 1197 1198 if (current_phi == nullptr || current_phi->AsPhi()->GetRegNumber() != vreg) { 1199 stack_map_stream->AddDexRegisterEntry(DexRegisterLocation::Kind::kNone, 0); 1200 } else { 1201 Location location = current_phi->GetLocations()->Out(); 1202 switch (location.GetKind()) { 1203 case Location::kStackSlot: { 1204 stack_map_stream->AddDexRegisterEntry( 1205 DexRegisterLocation::Kind::kInStack, location.GetStackIndex()); 1206 break; 1207 } 1208 case Location::kDoubleStackSlot: { 1209 stack_map_stream->AddDexRegisterEntry( 1210 DexRegisterLocation::Kind::kInStack, location.GetStackIndex()); 1211 stack_map_stream->AddDexRegisterEntry( 1212 DexRegisterLocation::Kind::kInStack, location.GetHighStackIndex(kVRegSize)); 1213 ++vreg; 1214 DCHECK_LT(vreg, num_vregs); 1215 break; 1216 } 1217 default: { 1218 // All catch phis must be allocated to a stack slot. 1219 LOG(FATAL) << "Unexpected kind " << location.GetKind(); 1220 UNREACHABLE(); 1221 } 1222 } 1223 } 1224 } 1225 1226 stack_map_stream->EndStackMapEntry(); 1227 } 1228 } 1229 1230 void CodeGenerator::AddSlowPath(SlowPathCode* slow_path) { 1231 DCHECK(code_generation_data_ != nullptr); 1232 code_generation_data_->AddSlowPath(slow_path); 1233 } 1234 1235 void CodeGenerator::EmitEnvironment(HEnvironment* environment, SlowPathCode* slow_path) { 1236 if (environment == nullptr) return; 1237 1238 StackMapStream* stack_map_stream = GetStackMapStream(); 1239 if (environment->GetParent() != nullptr) { 1240 // We emit the parent environment first. 1241 EmitEnvironment(environment->GetParent(), slow_path); 1242 stack_map_stream->BeginInlineInfoEntry(environment->GetMethod(), 1243 environment->GetDexPc(), 1244 environment->Size(), 1245 &graph_->GetDexFile()); 1246 } 1247 1248 // Walk over the environment, and record the location of dex registers. 1249 for (size_t i = 0, environment_size = environment->Size(); i < environment_size; ++i) { 1250 HInstruction* current = environment->GetInstructionAt(i); 1251 if (current == nullptr) { 1252 stack_map_stream->AddDexRegisterEntry(DexRegisterLocation::Kind::kNone, 0); 1253 continue; 1254 } 1255 1256 using Kind = DexRegisterLocation::Kind; 1257 Location location = environment->GetLocationAt(i); 1258 switch (location.GetKind()) { 1259 case Location::kConstant: { 1260 DCHECK_EQ(current, location.GetConstant()); 1261 if (current->IsLongConstant()) { 1262 int64_t value = current->AsLongConstant()->GetValue(); 1263 stack_map_stream->AddDexRegisterEntry(Kind::kConstant, Low32Bits(value)); 1264 stack_map_stream->AddDexRegisterEntry(Kind::kConstant, High32Bits(value)); 1265 ++i; 1266 DCHECK_LT(i, environment_size); 1267 } else if (current->IsDoubleConstant()) { 1268 int64_t value = bit_cast<int64_t, double>(current->AsDoubleConstant()->GetValue()); 1269 stack_map_stream->AddDexRegisterEntry(Kind::kConstant, Low32Bits(value)); 1270 stack_map_stream->AddDexRegisterEntry(Kind::kConstant, High32Bits(value)); 1271 ++i; 1272 DCHECK_LT(i, environment_size); 1273 } else if (current->IsIntConstant()) { 1274 int32_t value = current->AsIntConstant()->GetValue(); 1275 stack_map_stream->AddDexRegisterEntry(Kind::kConstant, value); 1276 } else if (current->IsNullConstant()) { 1277 stack_map_stream->AddDexRegisterEntry(Kind::kConstant, 0); 1278 } else { 1279 DCHECK(current->IsFloatConstant()) << current->DebugName(); 1280 int32_t value = bit_cast<int32_t, float>(current->AsFloatConstant()->GetValue()); 1281 stack_map_stream->AddDexRegisterEntry(Kind::kConstant, value); 1282 } 1283 break; 1284 } 1285 1286 case Location::kStackSlot: { 1287 stack_map_stream->AddDexRegisterEntry(Kind::kInStack, location.GetStackIndex()); 1288 break; 1289 } 1290 1291 case Location::kDoubleStackSlot: { 1292 stack_map_stream->AddDexRegisterEntry(Kind::kInStack, location.GetStackIndex()); 1293 stack_map_stream->AddDexRegisterEntry( 1294 Kind::kInStack, location.GetHighStackIndex(kVRegSize)); 1295 ++i; 1296 DCHECK_LT(i, environment_size); 1297 break; 1298 } 1299 1300 case Location::kRegister : { 1301 int id = location.reg(); 1302 if (slow_path != nullptr && slow_path->IsCoreRegisterSaved(id)) { 1303 uint32_t offset = slow_path->GetStackOffsetOfCoreRegister(id); 1304 stack_map_stream->AddDexRegisterEntry(Kind::kInStack, offset); 1305 if (current->GetType() == DataType::Type::kInt64) { 1306 stack_map_stream->AddDexRegisterEntry(Kind::kInStack, offset + kVRegSize); 1307 ++i; 1308 DCHECK_LT(i, environment_size); 1309 } 1310 } else { 1311 stack_map_stream->AddDexRegisterEntry(Kind::kInRegister, id); 1312 if (current->GetType() == DataType::Type::kInt64) { 1313 stack_map_stream->AddDexRegisterEntry(Kind::kInRegisterHigh, id); 1314 ++i; 1315 DCHECK_LT(i, environment_size); 1316 } 1317 } 1318 break; 1319 } 1320 1321 case Location::kFpuRegister : { 1322 int id = location.reg(); 1323 if (slow_path != nullptr && slow_path->IsFpuRegisterSaved(id)) { 1324 uint32_t offset = slow_path->GetStackOffsetOfFpuRegister(id); 1325 stack_map_stream->AddDexRegisterEntry(Kind::kInStack, offset); 1326 if (current->GetType() == DataType::Type::kFloat64) { 1327 stack_map_stream->AddDexRegisterEntry(Kind::kInStack, offset + kVRegSize); 1328 ++i; 1329 DCHECK_LT(i, environment_size); 1330 } 1331 } else { 1332 stack_map_stream->AddDexRegisterEntry(Kind::kInFpuRegister, id); 1333 if (current->GetType() == DataType::Type::kFloat64) { 1334 stack_map_stream->AddDexRegisterEntry(Kind::kInFpuRegisterHigh, id); 1335 ++i; 1336 DCHECK_LT(i, environment_size); 1337 } 1338 } 1339 break; 1340 } 1341 1342 case Location::kFpuRegisterPair : { 1343 int low = location.low(); 1344 int high = location.high(); 1345 if (slow_path != nullptr && slow_path->IsFpuRegisterSaved(low)) { 1346 uint32_t offset = slow_path->GetStackOffsetOfFpuRegister(low); 1347 stack_map_stream->AddDexRegisterEntry(Kind::kInStack, offset); 1348 } else { 1349 stack_map_stream->AddDexRegisterEntry(Kind::kInFpuRegister, low); 1350 } 1351 if (slow_path != nullptr && slow_path->IsFpuRegisterSaved(high)) { 1352 uint32_t offset = slow_path->GetStackOffsetOfFpuRegister(high); 1353 stack_map_stream->AddDexRegisterEntry(Kind::kInStack, offset); 1354 ++i; 1355 } else { 1356 stack_map_stream->AddDexRegisterEntry(Kind::kInFpuRegister, high); 1357 ++i; 1358 } 1359 DCHECK_LT(i, environment_size); 1360 break; 1361 } 1362 1363 case Location::kRegisterPair : { 1364 int low = location.low(); 1365 int high = location.high(); 1366 if (slow_path != nullptr && slow_path->IsCoreRegisterSaved(low)) { 1367 uint32_t offset = slow_path->GetStackOffsetOfCoreRegister(low); 1368 stack_map_stream->AddDexRegisterEntry(Kind::kInStack, offset); 1369 } else { 1370 stack_map_stream->AddDexRegisterEntry(Kind::kInRegister, low); 1371 } 1372 if (slow_path != nullptr && slow_path->IsCoreRegisterSaved(high)) { 1373 uint32_t offset = slow_path->GetStackOffsetOfCoreRegister(high); 1374 stack_map_stream->AddDexRegisterEntry(Kind::kInStack, offset); 1375 } else { 1376 stack_map_stream->AddDexRegisterEntry(Kind::kInRegister, high); 1377 } 1378 ++i; 1379 DCHECK_LT(i, environment_size); 1380 break; 1381 } 1382 1383 case Location::kInvalid: { 1384 stack_map_stream->AddDexRegisterEntry(Kind::kNone, 0); 1385 break; 1386 } 1387 1388 default: 1389 LOG(FATAL) << "Unexpected kind " << location.GetKind(); 1390 } 1391 } 1392 1393 if (environment->GetParent() != nullptr) { 1394 stack_map_stream->EndInlineInfoEntry(); 1395 } 1396 } 1397 1398 bool CodeGenerator::CanMoveNullCheckToUser(HNullCheck* null_check) { 1399 return null_check->IsEmittedAtUseSite(); 1400 } 1401 1402 void CodeGenerator::MaybeRecordImplicitNullCheck(HInstruction* instr) { 1403 HNullCheck* null_check = instr->GetImplicitNullCheck(); 1404 if (null_check != nullptr) { 1405 RecordPcInfo(null_check, null_check->GetDexPc()); 1406 } 1407 } 1408 1409 LocationSummary* CodeGenerator::CreateThrowingSlowPathLocations(HInstruction* instruction, 1410 RegisterSet caller_saves) { 1411 // Note: Using kNoCall allows the method to be treated as leaf (and eliminate the 1412 // HSuspendCheck from entry block). However, it will still get a valid stack frame 1413 // because the HNullCheck needs an environment. 1414 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 1415 // When throwing from a try block, we may need to retrieve dalvik registers from 1416 // physical registers and we also need to set up stack mask for GC. This is 1417 // implicitly achieved by passing kCallOnSlowPath to the LocationSummary. 1418 bool can_throw_into_catch_block = instruction->CanThrowIntoCatchBlock(); 1419 if (can_throw_into_catch_block) { 1420 call_kind = LocationSummary::kCallOnSlowPath; 1421 } 1422 LocationSummary* locations = 1423 new (GetGraph()->GetAllocator()) LocationSummary(instruction, call_kind); 1424 if (can_throw_into_catch_block && compiler_options_.GetImplicitNullChecks()) { 1425 locations->SetCustomSlowPathCallerSaves(caller_saves); // Default: no caller-save registers. 1426 } 1427 DCHECK(!instruction->HasUses()); 1428 return locations; 1429 } 1430 1431 void CodeGenerator::GenerateNullCheck(HNullCheck* instruction) { 1432 if (compiler_options_.GetImplicitNullChecks()) { 1433 MaybeRecordStat(stats_, MethodCompilationStat::kImplicitNullCheckGenerated); 1434 GenerateImplicitNullCheck(instruction); 1435 } else { 1436 MaybeRecordStat(stats_, MethodCompilationStat::kExplicitNullCheckGenerated); 1437 GenerateExplicitNullCheck(instruction); 1438 } 1439 } 1440 1441 void CodeGenerator::ClearSpillSlotsFromLoopPhisInStackMap(HSuspendCheck* suspend_check, 1442 HParallelMove* spills) const { 1443 LocationSummary* locations = suspend_check->GetLocations(); 1444 HBasicBlock* block = suspend_check->GetBlock(); 1445 DCHECK(block->GetLoopInformation()->GetSuspendCheck() == suspend_check); 1446 DCHECK(block->IsLoopHeader()); 1447 DCHECK(block->GetFirstInstruction() == spills); 1448 1449 for (size_t i = 0, num_moves = spills->NumMoves(); i != num_moves; ++i) { 1450 Location dest = spills->MoveOperandsAt(i)->GetDestination(); 1451 // All parallel moves in loop headers are spills. 1452 DCHECK(dest.IsStackSlot() || dest.IsDoubleStackSlot() || dest.IsSIMDStackSlot()) << dest; 1453 // Clear the stack bit marking a reference. Do not bother to check if the spill is 1454 // actually a reference spill, clearing bits that are already zero is harmless. 1455 locations->ClearStackBit(dest.GetStackIndex() / kVRegSize); 1456 } 1457 } 1458 1459 void CodeGenerator::EmitParallelMoves(Location from1, 1460 Location to1, 1461 DataType::Type type1, 1462 Location from2, 1463 Location to2, 1464 DataType::Type type2) { 1465 HParallelMove parallel_move(GetGraph()->GetAllocator()); 1466 parallel_move.AddMove(from1, to1, type1, nullptr); 1467 parallel_move.AddMove(from2, to2, type2, nullptr); 1468 GetMoveResolver()->EmitNativeCode(¶llel_move); 1469 } 1470 1471 void CodeGenerator::ValidateInvokeRuntime(QuickEntrypointEnum entrypoint, 1472 HInstruction* instruction, 1473 SlowPathCode* slow_path) { 1474 // Ensure that the call kind indication given to the register allocator is 1475 // coherent with the runtime call generated. 1476 if (slow_path == nullptr) { 1477 DCHECK(instruction->GetLocations()->WillCall()) 1478 << "instruction->DebugName()=" << instruction->DebugName(); 1479 } else { 1480 DCHECK(instruction->GetLocations()->CallsOnSlowPath() || slow_path->IsFatal()) 1481 << "instruction->DebugName()=" << instruction->DebugName() 1482 << " slow_path->GetDescription()=" << slow_path->GetDescription(); 1483 } 1484 1485 // Check that the GC side effect is set when required. 1486 // TODO: Reverse EntrypointCanTriggerGC 1487 if (EntrypointCanTriggerGC(entrypoint)) { 1488 if (slow_path == nullptr) { 1489 DCHECK(instruction->GetSideEffects().Includes(SideEffects::CanTriggerGC())) 1490 << "instruction->DebugName()=" << instruction->DebugName() 1491 << " instruction->GetSideEffects().ToString()=" 1492 << instruction->GetSideEffects().ToString(); 1493 } else { 1494 // 'CanTriggerGC' side effect is used to restrict optimization of instructions which depend 1495 // on GC (e.g. IntermediateAddress) - to ensure they are not alive across GC points. However 1496 // if execution never returns to the compiled code from a GC point this restriction is 1497 // unnecessary - in particular for fatal slow paths which might trigger GC. 1498 DCHECK((slow_path->IsFatal() && !instruction->GetLocations()->WillCall()) || 1499 instruction->GetSideEffects().Includes(SideEffects::CanTriggerGC()) || 1500 // When (non-Baker) read barriers are enabled, some instructions 1501 // use a slow path to emit a read barrier, which does not trigger 1502 // GC. 1503 (kEmitCompilerReadBarrier && 1504 !kUseBakerReadBarrier && 1505 (instruction->IsInstanceFieldGet() || 1506 instruction->IsStaticFieldGet() || 1507 instruction->IsArrayGet() || 1508 instruction->IsLoadClass() || 1509 instruction->IsLoadString() || 1510 instruction->IsInstanceOf() || 1511 instruction->IsCheckCast() || 1512 (instruction->IsInvokeVirtual() && instruction->GetLocations()->Intrinsified())))) 1513 << "instruction->DebugName()=" << instruction->DebugName() 1514 << " instruction->GetSideEffects().ToString()=" 1515 << instruction->GetSideEffects().ToString() 1516 << " slow_path->GetDescription()=" << slow_path->GetDescription(); 1517 } 1518 } else { 1519 // The GC side effect is not required for the instruction. But the instruction might still have 1520 // it, for example if it calls other entrypoints requiring it. 1521 } 1522 1523 // Check the coherency of leaf information. 1524 DCHECK(instruction->IsSuspendCheck() 1525 || ((slow_path != nullptr) && slow_path->IsFatal()) 1526 || instruction->GetLocations()->CanCall() 1527 || !IsLeafMethod()) 1528 << instruction->DebugName() << ((slow_path != nullptr) ? slow_path->GetDescription() : ""); 1529 } 1530 1531 void CodeGenerator::ValidateInvokeRuntimeWithoutRecordingPcInfo(HInstruction* instruction, 1532 SlowPathCode* slow_path) { 1533 DCHECK(instruction->GetLocations()->OnlyCallsOnSlowPath()) 1534 << "instruction->DebugName()=" << instruction->DebugName() 1535 << " slow_path->GetDescription()=" << slow_path->GetDescription(); 1536 // Only the Baker read barrier marking slow path used by certains 1537 // instructions is expected to invoke the runtime without recording 1538 // PC-related information. 1539 DCHECK(kUseBakerReadBarrier); 1540 DCHECK(instruction->IsInstanceFieldGet() || 1541 instruction->IsStaticFieldGet() || 1542 instruction->IsArrayGet() || 1543 instruction->IsArraySet() || 1544 instruction->IsLoadClass() || 1545 instruction->IsLoadString() || 1546 instruction->IsInstanceOf() || 1547 instruction->IsCheckCast() || 1548 (instruction->IsInvokeVirtual() && instruction->GetLocations()->Intrinsified()) || 1549 (instruction->IsInvokeStaticOrDirect() && instruction->GetLocations()->Intrinsified())) 1550 << "instruction->DebugName()=" << instruction->DebugName() 1551 << " slow_path->GetDescription()=" << slow_path->GetDescription(); 1552 } 1553 1554 void SlowPathCode::SaveLiveRegisters(CodeGenerator* codegen, LocationSummary* locations) { 1555 size_t stack_offset = codegen->GetFirstRegisterSlotInSlowPath(); 1556 1557 const uint32_t core_spills = codegen->GetSlowPathSpills(locations, /* core_registers= */ true); 1558 for (uint32_t i : LowToHighBits(core_spills)) { 1559 // If the register holds an object, update the stack mask. 1560 if (locations->RegisterContainsObject(i)) { 1561 locations->SetStackBit(stack_offset / kVRegSize); 1562 } 1563 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 1564 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 1565 saved_core_stack_offsets_[i] = stack_offset; 1566 stack_offset += codegen->SaveCoreRegister(stack_offset, i); 1567 } 1568 1569 const uint32_t fp_spills = codegen->GetSlowPathSpills(locations, /* core_registers= */ false); 1570 for (uint32_t i : LowToHighBits(fp_spills)) { 1571 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 1572 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 1573 saved_fpu_stack_offsets_[i] = stack_offset; 1574 stack_offset += codegen->SaveFloatingPointRegister(stack_offset, i); 1575 } 1576 } 1577 1578 void SlowPathCode::RestoreLiveRegisters(CodeGenerator* codegen, LocationSummary* locations) { 1579 size_t stack_offset = codegen->GetFirstRegisterSlotInSlowPath(); 1580 1581 const uint32_t core_spills = codegen->GetSlowPathSpills(locations, /* core_registers= */ true); 1582 for (uint32_t i : LowToHighBits(core_spills)) { 1583 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 1584 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 1585 stack_offset += codegen->RestoreCoreRegister(stack_offset, i); 1586 } 1587 1588 const uint32_t fp_spills = codegen->GetSlowPathSpills(locations, /* core_registers= */ false); 1589 for (uint32_t i : LowToHighBits(fp_spills)) { 1590 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 1591 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 1592 stack_offset += codegen->RestoreFloatingPointRegister(stack_offset, i); 1593 } 1594 } 1595 1596 void CodeGenerator::CreateSystemArrayCopyLocationSummary(HInvoke* invoke) { 1597 // Check to see if we have known failures that will cause us to have to bail out 1598 // to the runtime, and just generate the runtime call directly. 1599 HIntConstant* src_pos = invoke->InputAt(1)->AsIntConstant(); 1600 HIntConstant* dest_pos = invoke->InputAt(3)->AsIntConstant(); 1601 1602 // The positions must be non-negative. 1603 if ((src_pos != nullptr && src_pos->GetValue() < 0) || 1604 (dest_pos != nullptr && dest_pos->GetValue() < 0)) { 1605 // We will have to fail anyways. 1606 return; 1607 } 1608 1609 // The length must be >= 0. 1610 HIntConstant* length = invoke->InputAt(4)->AsIntConstant(); 1611 if (length != nullptr) { 1612 int32_t len = length->GetValue(); 1613 if (len < 0) { 1614 // Just call as normal. 1615 return; 1616 } 1617 } 1618 1619 SystemArrayCopyOptimizations optimizations(invoke); 1620 1621 if (optimizations.GetDestinationIsSource()) { 1622 if (src_pos != nullptr && dest_pos != nullptr && src_pos->GetValue() < dest_pos->GetValue()) { 1623 // We only support backward copying if source and destination are the same. 1624 return; 1625 } 1626 } 1627 1628 if (optimizations.GetDestinationIsPrimitiveArray() || optimizations.GetSourceIsPrimitiveArray()) { 1629 // We currently don't intrinsify primitive copying. 1630 return; 1631 } 1632 1633 ArenaAllocator* allocator = invoke->GetBlock()->GetGraph()->GetAllocator(); 1634 LocationSummary* locations = new (allocator) LocationSummary(invoke, 1635 LocationSummary::kCallOnSlowPath, 1636 kIntrinsified); 1637 // arraycopy(Object src, int src_pos, Object dest, int dest_pos, int length). 1638 locations->SetInAt(0, Location::RequiresRegister()); 1639 locations->SetInAt(1, Location::RegisterOrConstant(invoke->InputAt(1))); 1640 locations->SetInAt(2, Location::RequiresRegister()); 1641 locations->SetInAt(3, Location::RegisterOrConstant(invoke->InputAt(3))); 1642 locations->SetInAt(4, Location::RegisterOrConstant(invoke->InputAt(4))); 1643 1644 locations->AddTemp(Location::RequiresRegister()); 1645 locations->AddTemp(Location::RequiresRegister()); 1646 locations->AddTemp(Location::RequiresRegister()); 1647 } 1648 1649 void CodeGenerator::EmitJitRoots(uint8_t* code, 1650 const uint8_t* roots_data, 1651 /*out*/std::vector<Handle<mirror::Object>>* roots) { 1652 code_generation_data_->EmitJitRoots(roots); 1653 EmitJitRootPatches(code, roots_data); 1654 } 1655 1656 QuickEntrypointEnum CodeGenerator::GetArrayAllocationEntrypoint(HNewArray* new_array) { 1657 switch (new_array->GetComponentSizeShift()) { 1658 case 0: return kQuickAllocArrayResolved8; 1659 case 1: return kQuickAllocArrayResolved16; 1660 case 2: return kQuickAllocArrayResolved32; 1661 case 3: return kQuickAllocArrayResolved64; 1662 } 1663 LOG(FATAL) << "Unreachable"; 1664 UNREACHABLE(); 1665 } 1666 1667 } // namespace art 1668