code_generator.cc revision fa4333dcb481e564f54726b4e6f8153612df835e
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.h" 21#include "code_generator_arm_vixl.h" 22#endif 23 24#ifdef ART_ENABLE_CODEGEN_arm64 25#include "code_generator_arm64.h" 26#endif 27 28#ifdef ART_ENABLE_CODEGEN_x86 29#include "code_generator_x86.h" 30#endif 31 32#ifdef ART_ENABLE_CODEGEN_x86_64 33#include "code_generator_x86_64.h" 34#endif 35 36#ifdef ART_ENABLE_CODEGEN_mips 37#include "code_generator_mips.h" 38#endif 39 40#ifdef ART_ENABLE_CODEGEN_mips64 41#include "code_generator_mips64.h" 42#endif 43 44#include "bytecode_utils.h" 45#include "compiled_method.h" 46#include "dex/verified_method.h" 47#include "driver/compiler_driver.h" 48#include "graph_visualizer.h" 49#include "intrinsics.h" 50#include "leb128.h" 51#include "mirror/array-inl.h" 52#include "mirror/object_array-inl.h" 53#include "mirror/object_reference.h" 54#include "mirror/string.h" 55#include "parallel_move_resolver.h" 56#include "ssa_liveness_analysis.h" 57#include "utils/assembler.h" 58 59namespace art { 60 61// Return whether a location is consistent with a type. 62static bool CheckType(Primitive::Type type, Location location) { 63 if (location.IsFpuRegister() 64 || (location.IsUnallocated() && (location.GetPolicy() == Location::kRequiresFpuRegister))) { 65 return (type == Primitive::kPrimFloat) || (type == Primitive::kPrimDouble); 66 } else if (location.IsRegister() || 67 (location.IsUnallocated() && (location.GetPolicy() == Location::kRequiresRegister))) { 68 return Primitive::IsIntegralType(type) || (type == Primitive::kPrimNot); 69 } else if (location.IsRegisterPair()) { 70 return type == Primitive::kPrimLong; 71 } else if (location.IsFpuRegisterPair()) { 72 return type == Primitive::kPrimDouble; 73 } else if (location.IsStackSlot()) { 74 return (Primitive::IsIntegralType(type) && type != Primitive::kPrimLong) 75 || (type == Primitive::kPrimFloat) 76 || (type == Primitive::kPrimNot); 77 } else if (location.IsDoubleStackSlot()) { 78 return (type == Primitive::kPrimLong) || (type == Primitive::kPrimDouble); 79 } else if (location.IsConstant()) { 80 if (location.GetConstant()->IsIntConstant()) { 81 return Primitive::IsIntegralType(type) && (type != Primitive::kPrimLong); 82 } else if (location.GetConstant()->IsNullConstant()) { 83 return type == Primitive::kPrimNot; 84 } else if (location.GetConstant()->IsLongConstant()) { 85 return type == Primitive::kPrimLong; 86 } else if (location.GetConstant()->IsFloatConstant()) { 87 return type == Primitive::kPrimFloat; 88 } else { 89 return location.GetConstant()->IsDoubleConstant() 90 && (type == Primitive::kPrimDouble); 91 } 92 } else { 93 return location.IsInvalid() || (location.GetPolicy() == Location::kAny); 94 } 95} 96 97// Check that a location summary is consistent with an instruction. 98static bool CheckTypeConsistency(HInstruction* instruction) { 99 LocationSummary* locations = instruction->GetLocations(); 100 if (locations == nullptr) { 101 return true; 102 } 103 104 if (locations->Out().IsUnallocated() 105 && (locations->Out().GetPolicy() == Location::kSameAsFirstInput)) { 106 DCHECK(CheckType(instruction->GetType(), locations->InAt(0))) 107 << instruction->GetType() 108 << " " << locations->InAt(0); 109 } else { 110 DCHECK(CheckType(instruction->GetType(), locations->Out())) 111 << instruction->GetType() 112 << " " << locations->Out(); 113 } 114 115 HConstInputsRef inputs = instruction->GetInputs(); 116 for (size_t i = 0; i < inputs.size(); ++i) { 117 DCHECK(CheckType(inputs[i]->GetType(), locations->InAt(i))) 118 << inputs[i]->GetType() << " " << locations->InAt(i); 119 } 120 121 HEnvironment* environment = instruction->GetEnvironment(); 122 for (size_t i = 0; i < instruction->EnvironmentSize(); ++i) { 123 if (environment->GetInstructionAt(i) != nullptr) { 124 Primitive::Type type = environment->GetInstructionAt(i)->GetType(); 125 DCHECK(CheckType(type, environment->GetLocationAt(i))) 126 << type << " " << environment->GetLocationAt(i); 127 } else { 128 DCHECK(environment->GetLocationAt(i).IsInvalid()) 129 << environment->GetLocationAt(i); 130 } 131 } 132 return true; 133} 134 135size_t CodeGenerator::GetCacheOffset(uint32_t index) { 136 return sizeof(GcRoot<mirror::Object>) * index; 137} 138 139size_t CodeGenerator::GetCachePointerOffset(uint32_t index) { 140 auto pointer_size = InstructionSetPointerSize(GetInstructionSet()); 141 return static_cast<size_t>(pointer_size) * index; 142} 143 144uint32_t CodeGenerator::GetArrayLengthOffset(HArrayLength* array_length) { 145 return array_length->IsStringLength() 146 ? mirror::String::CountOffset().Uint32Value() 147 : mirror::Array::LengthOffset().Uint32Value(); 148} 149 150uint32_t CodeGenerator::GetArrayDataOffset(HArrayGet* array_get) { 151 DCHECK(array_get->GetType() == Primitive::kPrimChar || !array_get->IsStringCharAt()); 152 return array_get->IsStringCharAt() 153 ? mirror::String::ValueOffset().Uint32Value() 154 : mirror::Array::DataOffset(Primitive::ComponentSize(array_get->GetType())).Uint32Value(); 155} 156 157bool CodeGenerator::GoesToNextBlock(HBasicBlock* current, HBasicBlock* next) const { 158 DCHECK_EQ((*block_order_)[current_block_index_], current); 159 return GetNextBlockToEmit() == FirstNonEmptyBlock(next); 160} 161 162HBasicBlock* CodeGenerator::GetNextBlockToEmit() const { 163 for (size_t i = current_block_index_ + 1; i < block_order_->size(); ++i) { 164 HBasicBlock* block = (*block_order_)[i]; 165 if (!block->IsSingleJump()) { 166 return block; 167 } 168 } 169 return nullptr; 170} 171 172HBasicBlock* CodeGenerator::FirstNonEmptyBlock(HBasicBlock* block) const { 173 while (block->IsSingleJump()) { 174 block = block->GetSuccessors()[0]; 175 } 176 return block; 177} 178 179class DisassemblyScope { 180 public: 181 DisassemblyScope(HInstruction* instruction, const CodeGenerator& codegen) 182 : codegen_(codegen), instruction_(instruction), start_offset_(static_cast<size_t>(-1)) { 183 if (codegen_.GetDisassemblyInformation() != nullptr) { 184 start_offset_ = codegen_.GetAssembler().CodeSize(); 185 } 186 } 187 188 ~DisassemblyScope() { 189 // We avoid building this data when we know it will not be used. 190 if (codegen_.GetDisassemblyInformation() != nullptr) { 191 codegen_.GetDisassemblyInformation()->AddInstructionInterval( 192 instruction_, start_offset_, codegen_.GetAssembler().CodeSize()); 193 } 194 } 195 196 private: 197 const CodeGenerator& codegen_; 198 HInstruction* instruction_; 199 size_t start_offset_; 200}; 201 202 203void CodeGenerator::GenerateSlowPaths() { 204 size_t code_start = 0; 205 for (const std::unique_ptr<SlowPathCode>& slow_path_unique_ptr : slow_paths_) { 206 SlowPathCode* slow_path = slow_path_unique_ptr.get(); 207 current_slow_path_ = slow_path; 208 if (disasm_info_ != nullptr) { 209 code_start = GetAssembler()->CodeSize(); 210 } 211 // Record the dex pc at start of slow path (required for java line number mapping). 212 MaybeRecordNativeDebugInfo(slow_path->GetInstruction(), slow_path->GetDexPc(), slow_path); 213 slow_path->EmitNativeCode(this); 214 if (disasm_info_ != nullptr) { 215 disasm_info_->AddSlowPathInterval(slow_path, code_start, GetAssembler()->CodeSize()); 216 } 217 } 218 current_slow_path_ = nullptr; 219} 220 221void CodeGenerator::Compile(CodeAllocator* allocator) { 222 // The register allocator already called `InitializeCodeGeneration`, 223 // where the frame size has been computed. 224 DCHECK(block_order_ != nullptr); 225 Initialize(); 226 227 HGraphVisitor* instruction_visitor = GetInstructionVisitor(); 228 DCHECK_EQ(current_block_index_, 0u); 229 230 size_t frame_start = GetAssembler()->CodeSize(); 231 GenerateFrameEntry(); 232 DCHECK_EQ(GetAssembler()->cfi().GetCurrentCFAOffset(), static_cast<int>(frame_size_)); 233 if (disasm_info_ != nullptr) { 234 disasm_info_->SetFrameEntryInterval(frame_start, GetAssembler()->CodeSize()); 235 } 236 237 for (size_t e = block_order_->size(); current_block_index_ < e; ++current_block_index_) { 238 HBasicBlock* block = (*block_order_)[current_block_index_]; 239 // Don't generate code for an empty block. Its predecessors will branch to its successor 240 // directly. Also, the label of that block will not be emitted, so this helps catch 241 // errors where we reference that label. 242 if (block->IsSingleJump()) continue; 243 Bind(block); 244 // This ensures that we have correct native line mapping for all native instructions. 245 // It is necessary to make stepping over a statement work. Otherwise, any initial 246 // instructions (e.g. moves) would be assumed to be the start of next statement. 247 MaybeRecordNativeDebugInfo(nullptr /* instruction */, block->GetDexPc()); 248 for (HInstructionIterator it(block->GetInstructions()); !it.Done(); it.Advance()) { 249 HInstruction* current = it.Current(); 250 if (current->HasEnvironment()) { 251 // Create stackmap for HNativeDebugInfo or any instruction which calls native code. 252 // Note that we need correct mapping for the native PC of the call instruction, 253 // so the runtime's stackmap is not sufficient since it is at PC after the call. 254 MaybeRecordNativeDebugInfo(current, block->GetDexPc()); 255 } 256 DisassemblyScope disassembly_scope(current, *this); 257 DCHECK(CheckTypeConsistency(current)); 258 current->Accept(instruction_visitor); 259 } 260 } 261 262 GenerateSlowPaths(); 263 264 // Emit catch stack maps at the end of the stack map stream as expected by the 265 // runtime exception handler. 266 if (graph_->HasTryCatch()) { 267 RecordCatchBlockInfo(); 268 } 269 270 // Finalize instructions in assember; 271 Finalize(allocator); 272} 273 274void CodeGenerator::Finalize(CodeAllocator* allocator) { 275 size_t code_size = GetAssembler()->CodeSize(); 276 uint8_t* buffer = allocator->Allocate(code_size); 277 278 MemoryRegion code(buffer, code_size); 279 GetAssembler()->FinalizeInstructions(code); 280} 281 282void CodeGenerator::EmitLinkerPatches(ArenaVector<LinkerPatch>* linker_patches ATTRIBUTE_UNUSED) { 283 // No linker patches by default. 284} 285 286void CodeGenerator::InitializeCodeGeneration(size_t number_of_spill_slots, 287 size_t maximum_safepoint_spill_size, 288 size_t number_of_out_slots, 289 const ArenaVector<HBasicBlock*>& block_order) { 290 block_order_ = &block_order; 291 DCHECK(!block_order.empty()); 292 DCHECK(block_order[0] == GetGraph()->GetEntryBlock()); 293 ComputeSpillMask(); 294 first_register_slot_in_slow_path_ = RoundUp( 295 (number_of_out_slots + number_of_spill_slots) * kVRegSize, GetPreferredSlotsAlignment()); 296 297 if (number_of_spill_slots == 0 298 && !HasAllocatedCalleeSaveRegisters() 299 && IsLeafMethod() 300 && !RequiresCurrentMethod()) { 301 DCHECK_EQ(maximum_safepoint_spill_size, 0u); 302 SetFrameSize(CallPushesPC() ? GetWordSize() : 0); 303 } else { 304 SetFrameSize(RoundUp( 305 first_register_slot_in_slow_path_ 306 + maximum_safepoint_spill_size 307 + (GetGraph()->HasShouldDeoptimizeFlag() ? kShouldDeoptimizeFlagSize : 0) 308 + FrameEntrySpillSize(), 309 kStackAlignment)); 310 } 311} 312 313void CodeGenerator::CreateCommonInvokeLocationSummary( 314 HInvoke* invoke, InvokeDexCallingConventionVisitor* visitor) { 315 ArenaAllocator* allocator = invoke->GetBlock()->GetGraph()->GetArena(); 316 LocationSummary* locations = new (allocator) LocationSummary(invoke, 317 LocationSummary::kCallOnMainOnly); 318 319 for (size_t i = 0; i < invoke->GetNumberOfArguments(); i++) { 320 HInstruction* input = invoke->InputAt(i); 321 locations->SetInAt(i, visitor->GetNextLocation(input->GetType())); 322 } 323 324 locations->SetOut(visitor->GetReturnLocation(invoke->GetType())); 325 326 if (invoke->IsInvokeStaticOrDirect()) { 327 HInvokeStaticOrDirect* call = invoke->AsInvokeStaticOrDirect(); 328 switch (call->GetMethodLoadKind()) { 329 case HInvokeStaticOrDirect::MethodLoadKind::kRecursive: 330 locations->SetInAt(call->GetSpecialInputIndex(), visitor->GetMethodLocation()); 331 break; 332 case HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod: 333 locations->AddTemp(visitor->GetMethodLocation()); 334 locations->SetInAt(call->GetSpecialInputIndex(), Location::RequiresRegister()); 335 break; 336 default: 337 locations->AddTemp(visitor->GetMethodLocation()); 338 break; 339 } 340 } else { 341 locations->AddTemp(visitor->GetMethodLocation()); 342 } 343} 344 345void CodeGenerator::GenerateInvokeUnresolvedRuntimeCall(HInvokeUnresolved* invoke) { 346 MoveConstant(invoke->GetLocations()->GetTemp(0), invoke->GetDexMethodIndex()); 347 348 // Initialize to anything to silent compiler warnings. 349 QuickEntrypointEnum entrypoint = kQuickInvokeStaticTrampolineWithAccessCheck; 350 switch (invoke->GetInvokeType()) { 351 case kStatic: 352 entrypoint = kQuickInvokeStaticTrampolineWithAccessCheck; 353 break; 354 case kDirect: 355 entrypoint = kQuickInvokeDirectTrampolineWithAccessCheck; 356 break; 357 case kVirtual: 358 entrypoint = kQuickInvokeVirtualTrampolineWithAccessCheck; 359 break; 360 case kSuper: 361 entrypoint = kQuickInvokeSuperTrampolineWithAccessCheck; 362 break; 363 case kInterface: 364 entrypoint = kQuickInvokeInterfaceTrampolineWithAccessCheck; 365 break; 366 } 367 InvokeRuntime(entrypoint, invoke, invoke->GetDexPc(), nullptr); 368} 369 370void CodeGenerator::GenerateInvokePolymorphicCall(HInvokePolymorphic* invoke) { 371 MoveConstant(invoke->GetLocations()->GetTemp(0), static_cast<int32_t>(invoke->GetType())); 372 QuickEntrypointEnum entrypoint = kQuickInvokePolymorphic; 373 InvokeRuntime(entrypoint, invoke, invoke->GetDexPc(), nullptr); 374} 375 376void CodeGenerator::CreateUnresolvedFieldLocationSummary( 377 HInstruction* field_access, 378 Primitive::Type field_type, 379 const FieldAccessCallingConvention& calling_convention) { 380 bool is_instance = field_access->IsUnresolvedInstanceFieldGet() 381 || field_access->IsUnresolvedInstanceFieldSet(); 382 bool is_get = field_access->IsUnresolvedInstanceFieldGet() 383 || field_access->IsUnresolvedStaticFieldGet(); 384 385 ArenaAllocator* allocator = field_access->GetBlock()->GetGraph()->GetArena(); 386 LocationSummary* locations = 387 new (allocator) LocationSummary(field_access, LocationSummary::kCallOnMainOnly); 388 389 locations->AddTemp(calling_convention.GetFieldIndexLocation()); 390 391 if (is_instance) { 392 // Add the `this` object for instance field accesses. 393 locations->SetInAt(0, calling_convention.GetObjectLocation()); 394 } 395 396 // Note that pSetXXStatic/pGetXXStatic always takes/returns an int or int64 397 // regardless of the the type. Because of that we forced to special case 398 // the access to floating point values. 399 if (is_get) { 400 if (Primitive::IsFloatingPointType(field_type)) { 401 // The return value will be stored in regular registers while register 402 // allocator expects it in a floating point register. 403 // Note We don't need to request additional temps because the return 404 // register(s) are already blocked due the call and they may overlap with 405 // the input or field index. 406 // The transfer between the two will be done at codegen level. 407 locations->SetOut(calling_convention.GetFpuLocation(field_type)); 408 } else { 409 locations->SetOut(calling_convention.GetReturnLocation(field_type)); 410 } 411 } else { 412 size_t set_index = is_instance ? 1 : 0; 413 if (Primitive::IsFloatingPointType(field_type)) { 414 // The set value comes from a float location while the calling convention 415 // expects it in a regular register location. Allocate a temp for it and 416 // make the transfer at codegen. 417 AddLocationAsTemp(calling_convention.GetSetValueLocation(field_type, is_instance), locations); 418 locations->SetInAt(set_index, calling_convention.GetFpuLocation(field_type)); 419 } else { 420 locations->SetInAt(set_index, 421 calling_convention.GetSetValueLocation(field_type, is_instance)); 422 } 423 } 424} 425 426void CodeGenerator::GenerateUnresolvedFieldAccess( 427 HInstruction* field_access, 428 Primitive::Type field_type, 429 uint32_t field_index, 430 uint32_t dex_pc, 431 const FieldAccessCallingConvention& calling_convention) { 432 LocationSummary* locations = field_access->GetLocations(); 433 434 MoveConstant(locations->GetTemp(0), field_index); 435 436 bool is_instance = field_access->IsUnresolvedInstanceFieldGet() 437 || field_access->IsUnresolvedInstanceFieldSet(); 438 bool is_get = field_access->IsUnresolvedInstanceFieldGet() 439 || field_access->IsUnresolvedStaticFieldGet(); 440 441 if (!is_get && Primitive::IsFloatingPointType(field_type)) { 442 // Copy the float value to be set into the calling convention register. 443 // Note that using directly the temp location is problematic as we don't 444 // support temp register pairs. To avoid boilerplate conversion code, use 445 // the location from the calling convention. 446 MoveLocation(calling_convention.GetSetValueLocation(field_type, is_instance), 447 locations->InAt(is_instance ? 1 : 0), 448 (Primitive::Is64BitType(field_type) ? Primitive::kPrimLong : Primitive::kPrimInt)); 449 } 450 451 QuickEntrypointEnum entrypoint = kQuickSet8Static; // Initialize to anything to avoid warnings. 452 switch (field_type) { 453 case Primitive::kPrimBoolean: 454 entrypoint = is_instance 455 ? (is_get ? kQuickGetBooleanInstance : kQuickSet8Instance) 456 : (is_get ? kQuickGetBooleanStatic : kQuickSet8Static); 457 break; 458 case Primitive::kPrimByte: 459 entrypoint = is_instance 460 ? (is_get ? kQuickGetByteInstance : kQuickSet8Instance) 461 : (is_get ? kQuickGetByteStatic : kQuickSet8Static); 462 break; 463 case Primitive::kPrimShort: 464 entrypoint = is_instance 465 ? (is_get ? kQuickGetShortInstance : kQuickSet16Instance) 466 : (is_get ? kQuickGetShortStatic : kQuickSet16Static); 467 break; 468 case Primitive::kPrimChar: 469 entrypoint = is_instance 470 ? (is_get ? kQuickGetCharInstance : kQuickSet16Instance) 471 : (is_get ? kQuickGetCharStatic : kQuickSet16Static); 472 break; 473 case Primitive::kPrimInt: 474 case Primitive::kPrimFloat: 475 entrypoint = is_instance 476 ? (is_get ? kQuickGet32Instance : kQuickSet32Instance) 477 : (is_get ? kQuickGet32Static : kQuickSet32Static); 478 break; 479 case Primitive::kPrimNot: 480 entrypoint = is_instance 481 ? (is_get ? kQuickGetObjInstance : kQuickSetObjInstance) 482 : (is_get ? kQuickGetObjStatic : kQuickSetObjStatic); 483 break; 484 case Primitive::kPrimLong: 485 case Primitive::kPrimDouble: 486 entrypoint = is_instance 487 ? (is_get ? kQuickGet64Instance : kQuickSet64Instance) 488 : (is_get ? kQuickGet64Static : kQuickSet64Static); 489 break; 490 default: 491 LOG(FATAL) << "Invalid type " << field_type; 492 } 493 InvokeRuntime(entrypoint, field_access, dex_pc, nullptr); 494 495 if (is_get && Primitive::IsFloatingPointType(field_type)) { 496 MoveLocation(locations->Out(), calling_convention.GetReturnLocation(field_type), field_type); 497 } 498} 499 500void CodeGenerator::CreateLoadClassRuntimeCallLocationSummary(HLoadClass* cls, 501 Location runtime_type_index_location, 502 Location runtime_return_location) { 503 DCHECK_EQ(cls->GetLoadKind(), HLoadClass::LoadKind::kDexCacheViaMethod); 504 DCHECK_EQ(cls->InputCount(), 1u); 505 LocationSummary* locations = new (cls->GetBlock()->GetGraph()->GetArena()) LocationSummary( 506 cls, LocationSummary::kCallOnMainOnly); 507 locations->SetInAt(0, Location::NoLocation()); 508 locations->AddTemp(runtime_type_index_location); 509 locations->SetOut(runtime_return_location); 510} 511 512void CodeGenerator::GenerateLoadClassRuntimeCall(HLoadClass* cls) { 513 DCHECK_EQ(cls->GetLoadKind(), HLoadClass::LoadKind::kDexCacheViaMethod); 514 LocationSummary* locations = cls->GetLocations(); 515 MoveConstant(locations->GetTemp(0), cls->GetTypeIndex().index_); 516 if (cls->NeedsAccessCheck()) { 517 CheckEntrypointTypes<kQuickInitializeTypeAndVerifyAccess, void*, uint32_t>(); 518 InvokeRuntime(kQuickInitializeTypeAndVerifyAccess, cls, cls->GetDexPc()); 519 } else if (cls->MustGenerateClinitCheck()) { 520 CheckEntrypointTypes<kQuickInitializeStaticStorage, void*, uint32_t>(); 521 InvokeRuntime(kQuickInitializeStaticStorage, cls, cls->GetDexPc()); 522 } else { 523 CheckEntrypointTypes<kQuickInitializeType, void*, uint32_t>(); 524 InvokeRuntime(kQuickInitializeType, cls, cls->GetDexPc()); 525 } 526} 527 528void CodeGenerator::BlockIfInRegister(Location location, bool is_out) const { 529 // The DCHECKS below check that a register is not specified twice in 530 // the summary. The out location can overlap with an input, so we need 531 // to special case it. 532 if (location.IsRegister()) { 533 DCHECK(is_out || !blocked_core_registers_[location.reg()]); 534 blocked_core_registers_[location.reg()] = true; 535 } else if (location.IsFpuRegister()) { 536 DCHECK(is_out || !blocked_fpu_registers_[location.reg()]); 537 blocked_fpu_registers_[location.reg()] = true; 538 } else if (location.IsFpuRegisterPair()) { 539 DCHECK(is_out || !blocked_fpu_registers_[location.AsFpuRegisterPairLow<int>()]); 540 blocked_fpu_registers_[location.AsFpuRegisterPairLow<int>()] = true; 541 DCHECK(is_out || !blocked_fpu_registers_[location.AsFpuRegisterPairHigh<int>()]); 542 blocked_fpu_registers_[location.AsFpuRegisterPairHigh<int>()] = true; 543 } else if (location.IsRegisterPair()) { 544 DCHECK(is_out || !blocked_core_registers_[location.AsRegisterPairLow<int>()]); 545 blocked_core_registers_[location.AsRegisterPairLow<int>()] = true; 546 DCHECK(is_out || !blocked_core_registers_[location.AsRegisterPairHigh<int>()]); 547 blocked_core_registers_[location.AsRegisterPairHigh<int>()] = true; 548 } 549} 550 551void CodeGenerator::AllocateLocations(HInstruction* instruction) { 552 instruction->Accept(GetLocationBuilder()); 553 DCHECK(CheckTypeConsistency(instruction)); 554 LocationSummary* locations = instruction->GetLocations(); 555 if (!instruction->IsSuspendCheckEntry()) { 556 if (locations != nullptr) { 557 if (locations->CanCall()) { 558 MarkNotLeaf(); 559 } else if (locations->Intrinsified() && 560 instruction->IsInvokeStaticOrDirect() && 561 !instruction->AsInvokeStaticOrDirect()->HasCurrentMethodInput()) { 562 // A static method call that has been fully intrinsified, and cannot call on the slow 563 // path or refer to the current method directly, no longer needs current method. 564 return; 565 } 566 } 567 if (instruction->NeedsCurrentMethod()) { 568 SetRequiresCurrentMethod(); 569 } 570 } 571} 572 573void CodeGenerator::MaybeRecordStat(MethodCompilationStat compilation_stat, size_t count) const { 574 if (stats_ != nullptr) { 575 stats_->RecordStat(compilation_stat, count); 576 } 577} 578 579std::unique_ptr<CodeGenerator> CodeGenerator::Create(HGraph* graph, 580 InstructionSet instruction_set, 581 const InstructionSetFeatures& isa_features, 582 const CompilerOptions& compiler_options, 583 OptimizingCompilerStats* stats) { 584 ArenaAllocator* arena = graph->GetArena(); 585 switch (instruction_set) { 586#ifdef ART_ENABLE_CODEGEN_arm 587 case kArm: 588 case kThumb2: { 589 if (kArmUseVIXL32) { 590 return std::unique_ptr<CodeGenerator>( 591 new (arena) arm::CodeGeneratorARMVIXL(graph, 592 *isa_features.AsArmInstructionSetFeatures(), 593 compiler_options, 594 stats)); 595 } else { 596 return std::unique_ptr<CodeGenerator>( 597 new (arena) arm::CodeGeneratorARM(graph, 598 *isa_features.AsArmInstructionSetFeatures(), 599 compiler_options, 600 stats)); 601 } 602 } 603#endif 604#ifdef ART_ENABLE_CODEGEN_arm64 605 case kArm64: { 606 return std::unique_ptr<CodeGenerator>( 607 new (arena) arm64::CodeGeneratorARM64(graph, 608 *isa_features.AsArm64InstructionSetFeatures(), 609 compiler_options, 610 stats)); 611 } 612#endif 613#ifdef ART_ENABLE_CODEGEN_mips 614 case kMips: { 615 return std::unique_ptr<CodeGenerator>( 616 new (arena) mips::CodeGeneratorMIPS(graph, 617 *isa_features.AsMipsInstructionSetFeatures(), 618 compiler_options, 619 stats)); 620 } 621#endif 622#ifdef ART_ENABLE_CODEGEN_mips64 623 case kMips64: { 624 return std::unique_ptr<CodeGenerator>( 625 new (arena) mips64::CodeGeneratorMIPS64(graph, 626 *isa_features.AsMips64InstructionSetFeatures(), 627 compiler_options, 628 stats)); 629 } 630#endif 631#ifdef ART_ENABLE_CODEGEN_x86 632 case kX86: { 633 return std::unique_ptr<CodeGenerator>( 634 new (arena) x86::CodeGeneratorX86(graph, 635 *isa_features.AsX86InstructionSetFeatures(), 636 compiler_options, 637 stats)); 638 } 639#endif 640#ifdef ART_ENABLE_CODEGEN_x86_64 641 case kX86_64: { 642 return std::unique_ptr<CodeGenerator>( 643 new (arena) x86_64::CodeGeneratorX86_64(graph, 644 *isa_features.AsX86_64InstructionSetFeatures(), 645 compiler_options, 646 stats)); 647 } 648#endif 649 default: 650 return nullptr; 651 } 652} 653 654size_t CodeGenerator::ComputeStackMapsSize() { 655 return stack_map_stream_.PrepareForFillIn(); 656} 657 658static void CheckCovers(uint32_t dex_pc, 659 const HGraph& graph, 660 const CodeInfo& code_info, 661 const ArenaVector<HSuspendCheck*>& loop_headers, 662 ArenaVector<size_t>* covered) { 663 CodeInfoEncoding encoding = code_info.ExtractEncoding(); 664 for (size_t i = 0; i < loop_headers.size(); ++i) { 665 if (loop_headers[i]->GetDexPc() == dex_pc) { 666 if (graph.IsCompilingOsr()) { 667 DCHECK(code_info.GetOsrStackMapForDexPc(dex_pc, encoding).IsValid()); 668 } 669 ++(*covered)[i]; 670 } 671 } 672} 673 674// Debug helper to ensure loop entries in compiled code are matched by 675// dex branch instructions. 676static void CheckLoopEntriesCanBeUsedForOsr(const HGraph& graph, 677 const CodeInfo& code_info, 678 const DexFile::CodeItem& code_item) { 679 if (graph.HasTryCatch()) { 680 // One can write loops through try/catch, which we do not support for OSR anyway. 681 return; 682 } 683 ArenaVector<HSuspendCheck*> loop_headers(graph.GetArena()->Adapter(kArenaAllocMisc)); 684 for (HBasicBlock* block : graph.GetReversePostOrder()) { 685 if (block->IsLoopHeader()) { 686 HSuspendCheck* suspend_check = block->GetLoopInformation()->GetSuspendCheck(); 687 if (!suspend_check->GetEnvironment()->IsFromInlinedInvoke()) { 688 loop_headers.push_back(suspend_check); 689 } 690 } 691 } 692 ArenaVector<size_t> covered(loop_headers.size(), 0, graph.GetArena()->Adapter(kArenaAllocMisc)); 693 const uint16_t* code_ptr = code_item.insns_; 694 const uint16_t* code_end = code_item.insns_ + code_item.insns_size_in_code_units_; 695 696 size_t dex_pc = 0; 697 while (code_ptr < code_end) { 698 const Instruction& instruction = *Instruction::At(code_ptr); 699 if (instruction.IsBranch()) { 700 uint32_t target = dex_pc + instruction.GetTargetOffset(); 701 CheckCovers(target, graph, code_info, loop_headers, &covered); 702 } else if (instruction.IsSwitch()) { 703 DexSwitchTable table(instruction, dex_pc); 704 uint16_t num_entries = table.GetNumEntries(); 705 size_t offset = table.GetFirstValueIndex(); 706 707 // Use a larger loop counter type to avoid overflow issues. 708 for (size_t i = 0; i < num_entries; ++i) { 709 // The target of the case. 710 uint32_t target = dex_pc + table.GetEntryAt(i + offset); 711 CheckCovers(target, graph, code_info, loop_headers, &covered); 712 } 713 } 714 dex_pc += instruction.SizeInCodeUnits(); 715 code_ptr += instruction.SizeInCodeUnits(); 716 } 717 718 for (size_t i = 0; i < covered.size(); ++i) { 719 DCHECK_NE(covered[i], 0u) << "Loop in compiled code has no dex branch equivalent"; 720 } 721} 722 723void CodeGenerator::BuildStackMaps(MemoryRegion region, const DexFile::CodeItem& code_item) { 724 stack_map_stream_.FillIn(region); 725 if (kIsDebugBuild) { 726 CheckLoopEntriesCanBeUsedForOsr(*graph_, CodeInfo(region), code_item); 727 } 728} 729 730void CodeGenerator::RecordPcInfo(HInstruction* instruction, 731 uint32_t dex_pc, 732 SlowPathCode* slow_path) { 733 if (instruction != nullptr) { 734 // The code generated for some type conversions 735 // may call the runtime, thus normally requiring a subsequent 736 // call to this method. However, the method verifier does not 737 // produce PC information for certain instructions, which are 738 // considered "atomic" (they cannot join a GC). 739 // Therefore we do not currently record PC information for such 740 // instructions. As this may change later, we added this special 741 // case so that code generators may nevertheless call 742 // CodeGenerator::RecordPcInfo without triggering an error in 743 // CodeGenerator::BuildNativeGCMap ("Missing ref for dex pc 0x") 744 // thereafter. 745 if (instruction->IsTypeConversion()) { 746 return; 747 } 748 if (instruction->IsRem()) { 749 Primitive::Type type = instruction->AsRem()->GetResultType(); 750 if ((type == Primitive::kPrimFloat) || (type == Primitive::kPrimDouble)) { 751 return; 752 } 753 } 754 } 755 756 uint32_t outer_dex_pc = dex_pc; 757 uint32_t outer_environment_size = 0; 758 uint32_t inlining_depth = 0; 759 if (instruction != nullptr) { 760 for (HEnvironment* environment = instruction->GetEnvironment(); 761 environment != nullptr; 762 environment = environment->GetParent()) { 763 outer_dex_pc = environment->GetDexPc(); 764 outer_environment_size = environment->Size(); 765 if (environment != instruction->GetEnvironment()) { 766 inlining_depth++; 767 } 768 } 769 } 770 771 // Collect PC infos for the mapping table. 772 uint32_t native_pc = GetAssembler()->CodePosition(); 773 774 if (instruction == nullptr) { 775 // For stack overflow checks and native-debug-info entries without dex register 776 // mapping (i.e. start of basic block or start of slow path). 777 stack_map_stream_.BeginStackMapEntry(outer_dex_pc, native_pc, 0, 0, 0, 0); 778 stack_map_stream_.EndStackMapEntry(); 779 return; 780 } 781 LocationSummary* locations = instruction->GetLocations(); 782 783 uint32_t register_mask = locations->GetRegisterMask(); 784 DCHECK_EQ(register_mask & ~locations->GetLiveRegisters()->GetCoreRegisters(), 0u); 785 if (locations->OnlyCallsOnSlowPath()) { 786 // In case of slow path, we currently set the location of caller-save registers 787 // to register (instead of their stack location when pushed before the slow-path 788 // call). Therefore register_mask contains both callee-save and caller-save 789 // registers that hold objects. We must remove the spilled caller-save from the 790 // mask, since they will be overwritten by the callee. 791 uint32_t spills = GetSlowPathSpills(locations, /* core_registers */ true); 792 register_mask &= ~spills; 793 } else { 794 // The register mask must be a subset of callee-save registers. 795 DCHECK_EQ(register_mask & core_callee_save_mask_, register_mask); 796 } 797 stack_map_stream_.BeginStackMapEntry(outer_dex_pc, 798 native_pc, 799 register_mask, 800 locations->GetStackMask(), 801 outer_environment_size, 802 inlining_depth); 803 804 EmitEnvironment(instruction->GetEnvironment(), slow_path); 805 stack_map_stream_.EndStackMapEntry(); 806 807 HLoopInformation* info = instruction->GetBlock()->GetLoopInformation(); 808 if (instruction->IsSuspendCheck() && 809 (info != nullptr) && 810 graph_->IsCompilingOsr() && 811 (inlining_depth == 0)) { 812 DCHECK_EQ(info->GetSuspendCheck(), instruction); 813 // We duplicate the stack map as a marker that this stack map can be an OSR entry. 814 // Duplicating it avoids having the runtime recognize and skip an OSR stack map. 815 DCHECK(info->IsIrreducible()); 816 stack_map_stream_.BeginStackMapEntry( 817 dex_pc, native_pc, register_mask, locations->GetStackMask(), outer_environment_size, 0); 818 EmitEnvironment(instruction->GetEnvironment(), slow_path); 819 stack_map_stream_.EndStackMapEntry(); 820 if (kIsDebugBuild) { 821 HEnvironment* environment = instruction->GetEnvironment(); 822 for (size_t i = 0, environment_size = environment->Size(); i < environment_size; ++i) { 823 HInstruction* in_environment = environment->GetInstructionAt(i); 824 if (in_environment != nullptr) { 825 DCHECK(in_environment->IsPhi() || in_environment->IsConstant()); 826 Location location = environment->GetLocationAt(i); 827 DCHECK(location.IsStackSlot() || 828 location.IsDoubleStackSlot() || 829 location.IsConstant() || 830 location.IsInvalid()); 831 if (location.IsStackSlot() || location.IsDoubleStackSlot()) { 832 DCHECK_LT(location.GetStackIndex(), static_cast<int32_t>(GetFrameSize())); 833 } 834 } 835 } 836 } 837 } else if (kIsDebugBuild) { 838 // Ensure stack maps are unique, by checking that the native pc in the stack map 839 // last emitted is different than the native pc of the stack map just emitted. 840 size_t number_of_stack_maps = stack_map_stream_.GetNumberOfStackMaps(); 841 if (number_of_stack_maps > 1) { 842 DCHECK_NE(stack_map_stream_.GetStackMap(number_of_stack_maps - 1).native_pc_code_offset, 843 stack_map_stream_.GetStackMap(number_of_stack_maps - 2).native_pc_code_offset); 844 } 845 } 846} 847 848bool CodeGenerator::HasStackMapAtCurrentPc() { 849 uint32_t pc = GetAssembler()->CodeSize(); 850 size_t count = stack_map_stream_.GetNumberOfStackMaps(); 851 CodeOffset native_pc_offset = stack_map_stream_.GetStackMap(count - 1).native_pc_code_offset; 852 return (count > 0) && (native_pc_offset.Uint32Value(GetInstructionSet()) == pc); 853} 854 855void CodeGenerator::MaybeRecordNativeDebugInfo(HInstruction* instruction, 856 uint32_t dex_pc, 857 SlowPathCode* slow_path) { 858 if (GetCompilerOptions().GetNativeDebuggable() && dex_pc != kNoDexPc) { 859 if (HasStackMapAtCurrentPc()) { 860 // Ensure that we do not collide with the stack map of the previous instruction. 861 GenerateNop(); 862 } 863 RecordPcInfo(instruction, dex_pc, slow_path); 864 } 865} 866 867void CodeGenerator::RecordCatchBlockInfo() { 868 ArenaAllocator* arena = graph_->GetArena(); 869 870 for (HBasicBlock* block : *block_order_) { 871 if (!block->IsCatchBlock()) { 872 continue; 873 } 874 875 uint32_t dex_pc = block->GetDexPc(); 876 uint32_t num_vregs = graph_->GetNumberOfVRegs(); 877 uint32_t inlining_depth = 0; // Inlining of catch blocks is not supported at the moment. 878 uint32_t native_pc = GetAddressOf(block); 879 uint32_t register_mask = 0; // Not used. 880 881 // The stack mask is not used, so we leave it empty. 882 ArenaBitVector* stack_mask = 883 ArenaBitVector::Create(arena, 0, /* expandable */ true, kArenaAllocCodeGenerator); 884 885 stack_map_stream_.BeginStackMapEntry(dex_pc, 886 native_pc, 887 register_mask, 888 stack_mask, 889 num_vregs, 890 inlining_depth); 891 892 HInstruction* current_phi = block->GetFirstPhi(); 893 for (size_t vreg = 0; vreg < num_vregs; ++vreg) { 894 while (current_phi != nullptr && current_phi->AsPhi()->GetRegNumber() < vreg) { 895 HInstruction* next_phi = current_phi->GetNext(); 896 DCHECK(next_phi == nullptr || 897 current_phi->AsPhi()->GetRegNumber() <= next_phi->AsPhi()->GetRegNumber()) 898 << "Phis need to be sorted by vreg number to keep this a linear-time loop."; 899 current_phi = next_phi; 900 } 901 902 if (current_phi == nullptr || current_phi->AsPhi()->GetRegNumber() != vreg) { 903 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kNone, 0); 904 } else { 905 Location location = current_phi->GetLiveInterval()->ToLocation(); 906 switch (location.GetKind()) { 907 case Location::kStackSlot: { 908 stack_map_stream_.AddDexRegisterEntry( 909 DexRegisterLocation::Kind::kInStack, location.GetStackIndex()); 910 break; 911 } 912 case Location::kDoubleStackSlot: { 913 stack_map_stream_.AddDexRegisterEntry( 914 DexRegisterLocation::Kind::kInStack, location.GetStackIndex()); 915 stack_map_stream_.AddDexRegisterEntry( 916 DexRegisterLocation::Kind::kInStack, location.GetHighStackIndex(kVRegSize)); 917 ++vreg; 918 DCHECK_LT(vreg, num_vregs); 919 break; 920 } 921 default: { 922 // All catch phis must be allocated to a stack slot. 923 LOG(FATAL) << "Unexpected kind " << location.GetKind(); 924 UNREACHABLE(); 925 } 926 } 927 } 928 } 929 930 stack_map_stream_.EndStackMapEntry(); 931 } 932} 933 934void CodeGenerator::EmitEnvironment(HEnvironment* environment, SlowPathCode* slow_path) { 935 if (environment == nullptr) return; 936 937 if (environment->GetParent() != nullptr) { 938 // We emit the parent environment first. 939 EmitEnvironment(environment->GetParent(), slow_path); 940 stack_map_stream_.BeginInlineInfoEntry(environment->GetMethod(), 941 environment->GetDexPc(), 942 environment->Size(), 943 &graph_->GetDexFile()); 944 } 945 946 // Walk over the environment, and record the location of dex registers. 947 for (size_t i = 0, environment_size = environment->Size(); i < environment_size; ++i) { 948 HInstruction* current = environment->GetInstructionAt(i); 949 if (current == nullptr) { 950 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kNone, 0); 951 continue; 952 } 953 954 Location location = environment->GetLocationAt(i); 955 switch (location.GetKind()) { 956 case Location::kConstant: { 957 DCHECK_EQ(current, location.GetConstant()); 958 if (current->IsLongConstant()) { 959 int64_t value = current->AsLongConstant()->GetValue(); 960 stack_map_stream_.AddDexRegisterEntry( 961 DexRegisterLocation::Kind::kConstant, Low32Bits(value)); 962 stack_map_stream_.AddDexRegisterEntry( 963 DexRegisterLocation::Kind::kConstant, High32Bits(value)); 964 ++i; 965 DCHECK_LT(i, environment_size); 966 } else if (current->IsDoubleConstant()) { 967 int64_t value = bit_cast<int64_t, double>(current->AsDoubleConstant()->GetValue()); 968 stack_map_stream_.AddDexRegisterEntry( 969 DexRegisterLocation::Kind::kConstant, Low32Bits(value)); 970 stack_map_stream_.AddDexRegisterEntry( 971 DexRegisterLocation::Kind::kConstant, High32Bits(value)); 972 ++i; 973 DCHECK_LT(i, environment_size); 974 } else if (current->IsIntConstant()) { 975 int32_t value = current->AsIntConstant()->GetValue(); 976 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kConstant, value); 977 } else if (current->IsNullConstant()) { 978 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kConstant, 0); 979 } else { 980 DCHECK(current->IsFloatConstant()) << current->DebugName(); 981 int32_t value = bit_cast<int32_t, float>(current->AsFloatConstant()->GetValue()); 982 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kConstant, value); 983 } 984 break; 985 } 986 987 case Location::kStackSlot: { 988 stack_map_stream_.AddDexRegisterEntry( 989 DexRegisterLocation::Kind::kInStack, location.GetStackIndex()); 990 break; 991 } 992 993 case Location::kDoubleStackSlot: { 994 stack_map_stream_.AddDexRegisterEntry( 995 DexRegisterLocation::Kind::kInStack, location.GetStackIndex()); 996 stack_map_stream_.AddDexRegisterEntry( 997 DexRegisterLocation::Kind::kInStack, location.GetHighStackIndex(kVRegSize)); 998 ++i; 999 DCHECK_LT(i, environment_size); 1000 break; 1001 } 1002 1003 case Location::kRegister : { 1004 int id = location.reg(); 1005 if (slow_path != nullptr && slow_path->IsCoreRegisterSaved(id)) { 1006 uint32_t offset = slow_path->GetStackOffsetOfCoreRegister(id); 1007 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInStack, offset); 1008 if (current->GetType() == Primitive::kPrimLong) { 1009 stack_map_stream_.AddDexRegisterEntry( 1010 DexRegisterLocation::Kind::kInStack, offset + kVRegSize); 1011 ++i; 1012 DCHECK_LT(i, environment_size); 1013 } 1014 } else { 1015 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInRegister, id); 1016 if (current->GetType() == Primitive::kPrimLong) { 1017 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInRegisterHigh, id); 1018 ++i; 1019 DCHECK_LT(i, environment_size); 1020 } 1021 } 1022 break; 1023 } 1024 1025 case Location::kFpuRegister : { 1026 int id = location.reg(); 1027 if (slow_path != nullptr && slow_path->IsFpuRegisterSaved(id)) { 1028 uint32_t offset = slow_path->GetStackOffsetOfFpuRegister(id); 1029 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInStack, offset); 1030 if (current->GetType() == Primitive::kPrimDouble) { 1031 stack_map_stream_.AddDexRegisterEntry( 1032 DexRegisterLocation::Kind::kInStack, offset + kVRegSize); 1033 ++i; 1034 DCHECK_LT(i, environment_size); 1035 } 1036 } else { 1037 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInFpuRegister, id); 1038 if (current->GetType() == Primitive::kPrimDouble) { 1039 stack_map_stream_.AddDexRegisterEntry( 1040 DexRegisterLocation::Kind::kInFpuRegisterHigh, id); 1041 ++i; 1042 DCHECK_LT(i, environment_size); 1043 } 1044 } 1045 break; 1046 } 1047 1048 case Location::kFpuRegisterPair : { 1049 int low = location.low(); 1050 int high = location.high(); 1051 if (slow_path != nullptr && slow_path->IsFpuRegisterSaved(low)) { 1052 uint32_t offset = slow_path->GetStackOffsetOfFpuRegister(low); 1053 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInStack, offset); 1054 } else { 1055 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInFpuRegister, low); 1056 } 1057 if (slow_path != nullptr && slow_path->IsFpuRegisterSaved(high)) { 1058 uint32_t offset = slow_path->GetStackOffsetOfFpuRegister(high); 1059 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInStack, offset); 1060 ++i; 1061 } else { 1062 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInFpuRegister, high); 1063 ++i; 1064 } 1065 DCHECK_LT(i, environment_size); 1066 break; 1067 } 1068 1069 case Location::kRegisterPair : { 1070 int low = location.low(); 1071 int high = location.high(); 1072 if (slow_path != nullptr && slow_path->IsCoreRegisterSaved(low)) { 1073 uint32_t offset = slow_path->GetStackOffsetOfCoreRegister(low); 1074 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInStack, offset); 1075 } else { 1076 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInRegister, low); 1077 } 1078 if (slow_path != nullptr && slow_path->IsCoreRegisterSaved(high)) { 1079 uint32_t offset = slow_path->GetStackOffsetOfCoreRegister(high); 1080 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInStack, offset); 1081 } else { 1082 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInRegister, high); 1083 } 1084 ++i; 1085 DCHECK_LT(i, environment_size); 1086 break; 1087 } 1088 1089 case Location::kInvalid: { 1090 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kNone, 0); 1091 break; 1092 } 1093 1094 default: 1095 LOG(FATAL) << "Unexpected kind " << location.GetKind(); 1096 } 1097 } 1098 1099 if (environment->GetParent() != nullptr) { 1100 stack_map_stream_.EndInlineInfoEntry(); 1101 } 1102} 1103 1104bool CodeGenerator::CanMoveNullCheckToUser(HNullCheck* null_check) { 1105 HInstruction* first_next_not_move = null_check->GetNextDisregardingMoves(); 1106 1107 return (first_next_not_move != nullptr) 1108 && first_next_not_move->CanDoImplicitNullCheckOn(null_check->InputAt(0)); 1109} 1110 1111void CodeGenerator::MaybeRecordImplicitNullCheck(HInstruction* instr) { 1112 if (!compiler_options_.GetImplicitNullChecks()) { 1113 return; 1114 } 1115 1116 // If we are from a static path don't record the pc as we can't throw NPE. 1117 // NB: having the checks here makes the code much less verbose in the arch 1118 // specific code generators. 1119 if (instr->IsStaticFieldSet() || instr->IsStaticFieldGet()) { 1120 return; 1121 } 1122 1123 if (!instr->CanDoImplicitNullCheckOn(instr->InputAt(0))) { 1124 return; 1125 } 1126 1127 // Find the first previous instruction which is not a move. 1128 HInstruction* first_prev_not_move = instr->GetPreviousDisregardingMoves(); 1129 1130 // If the instruction is a null check it means that `instr` is the first user 1131 // and needs to record the pc. 1132 if (first_prev_not_move != nullptr && first_prev_not_move->IsNullCheck()) { 1133 HNullCheck* null_check = first_prev_not_move->AsNullCheck(); 1134 // TODO: The parallel moves modify the environment. Their changes need to be 1135 // reverted otherwise the stack maps at the throw point will not be correct. 1136 RecordPcInfo(null_check, null_check->GetDexPc()); 1137 } 1138} 1139 1140LocationSummary* CodeGenerator::CreateThrowingSlowPathLocations(HInstruction* instruction, 1141 RegisterSet caller_saves) { 1142 // Note: Using kNoCall allows the method to be treated as leaf (and eliminate the 1143 // HSuspendCheck from entry block). However, it will still get a valid stack frame 1144 // because the HNullCheck needs an environment. 1145 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 1146 // When throwing from a try block, we may need to retrieve dalvik registers from 1147 // physical registers and we also need to set up stack mask for GC. This is 1148 // implicitly achieved by passing kCallOnSlowPath to the LocationSummary. 1149 bool can_throw_into_catch_block = instruction->CanThrowIntoCatchBlock(); 1150 if (can_throw_into_catch_block) { 1151 call_kind = LocationSummary::kCallOnSlowPath; 1152 } 1153 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 1154 if (can_throw_into_catch_block && compiler_options_.GetImplicitNullChecks()) { 1155 locations->SetCustomSlowPathCallerSaves(caller_saves); // Default: no caller-save registers. 1156 } 1157 DCHECK(!instruction->HasUses()); 1158 return locations; 1159} 1160 1161void CodeGenerator::GenerateNullCheck(HNullCheck* instruction) { 1162 if (compiler_options_.GetImplicitNullChecks()) { 1163 MaybeRecordStat(kImplicitNullCheckGenerated); 1164 GenerateImplicitNullCheck(instruction); 1165 } else { 1166 MaybeRecordStat(kExplicitNullCheckGenerated); 1167 GenerateExplicitNullCheck(instruction); 1168 } 1169} 1170 1171void CodeGenerator::ClearSpillSlotsFromLoopPhisInStackMap(HSuspendCheck* suspend_check) const { 1172 LocationSummary* locations = suspend_check->GetLocations(); 1173 HBasicBlock* block = suspend_check->GetBlock(); 1174 DCHECK(block->GetLoopInformation()->GetSuspendCheck() == suspend_check); 1175 DCHECK(block->IsLoopHeader()); 1176 1177 for (HInstructionIterator it(block->GetPhis()); !it.Done(); it.Advance()) { 1178 HInstruction* current = it.Current(); 1179 LiveInterval* interval = current->GetLiveInterval(); 1180 // We only need to clear bits of loop phis containing objects and allocated in register. 1181 // Loop phis allocated on stack already have the object in the stack. 1182 if (current->GetType() == Primitive::kPrimNot 1183 && interval->HasRegister() 1184 && interval->HasSpillSlot()) { 1185 locations->ClearStackBit(interval->GetSpillSlot() / kVRegSize); 1186 } 1187 } 1188} 1189 1190void CodeGenerator::EmitParallelMoves(Location from1, 1191 Location to1, 1192 Primitive::Type type1, 1193 Location from2, 1194 Location to2, 1195 Primitive::Type type2) { 1196 HParallelMove parallel_move(GetGraph()->GetArena()); 1197 parallel_move.AddMove(from1, to1, type1, nullptr); 1198 parallel_move.AddMove(from2, to2, type2, nullptr); 1199 GetMoveResolver()->EmitNativeCode(¶llel_move); 1200} 1201 1202void CodeGenerator::ValidateInvokeRuntime(QuickEntrypointEnum entrypoint, 1203 HInstruction* instruction, 1204 SlowPathCode* slow_path) { 1205 // Ensure that the call kind indication given to the register allocator is 1206 // coherent with the runtime call generated. 1207 if (slow_path == nullptr) { 1208 DCHECK(instruction->GetLocations()->WillCall()) 1209 << "instruction->DebugName()=" << instruction->DebugName(); 1210 } else { 1211 DCHECK(instruction->GetLocations()->CallsOnSlowPath() || slow_path->IsFatal()) 1212 << "instruction->DebugName()=" << instruction->DebugName() 1213 << " slow_path->GetDescription()=" << slow_path->GetDescription(); 1214 } 1215 1216 // Check that the GC side effect is set when required. 1217 // TODO: Reverse EntrypointCanTriggerGC 1218 if (EntrypointCanTriggerGC(entrypoint)) { 1219 if (slow_path == nullptr) { 1220 DCHECK(instruction->GetSideEffects().Includes(SideEffects::CanTriggerGC())) 1221 << "instruction->DebugName()=" << instruction->DebugName() 1222 << " instruction->GetSideEffects().ToString()=" 1223 << instruction->GetSideEffects().ToString(); 1224 } else { 1225 DCHECK(instruction->GetSideEffects().Includes(SideEffects::CanTriggerGC()) || 1226 // When (non-Baker) read barriers are enabled, some instructions 1227 // use a slow path to emit a read barrier, which does not trigger 1228 // GC. 1229 (kEmitCompilerReadBarrier && 1230 !kUseBakerReadBarrier && 1231 (instruction->IsInstanceFieldGet() || 1232 instruction->IsStaticFieldGet() || 1233 instruction->IsArrayGet() || 1234 instruction->IsLoadClass() || 1235 instruction->IsLoadString() || 1236 instruction->IsInstanceOf() || 1237 instruction->IsCheckCast() || 1238 (instruction->IsInvokeVirtual() && instruction->GetLocations()->Intrinsified())))) 1239 << "instruction->DebugName()=" << instruction->DebugName() 1240 << " instruction->GetSideEffects().ToString()=" 1241 << instruction->GetSideEffects().ToString() 1242 << " slow_path->GetDescription()=" << slow_path->GetDescription(); 1243 } 1244 } else { 1245 // The GC side effect is not required for the instruction. But the instruction might still have 1246 // it, for example if it calls other entrypoints requiring it. 1247 } 1248 1249 // Check the coherency of leaf information. 1250 DCHECK(instruction->IsSuspendCheck() 1251 || ((slow_path != nullptr) && slow_path->IsFatal()) 1252 || instruction->GetLocations()->CanCall() 1253 || !IsLeafMethod()) 1254 << instruction->DebugName() << ((slow_path != nullptr) ? slow_path->GetDescription() : ""); 1255} 1256 1257void CodeGenerator::ValidateInvokeRuntimeWithoutRecordingPcInfo(HInstruction* instruction, 1258 SlowPathCode* slow_path) { 1259 DCHECK(instruction->GetLocations()->OnlyCallsOnSlowPath()) 1260 << "instruction->DebugName()=" << instruction->DebugName() 1261 << " slow_path->GetDescription()=" << slow_path->GetDescription(); 1262 // Only the Baker read barrier marking slow path used by certains 1263 // instructions is expected to invoke the runtime without recording 1264 // PC-related information. 1265 DCHECK(kUseBakerReadBarrier); 1266 DCHECK(instruction->IsInstanceFieldGet() || 1267 instruction->IsStaticFieldGet() || 1268 instruction->IsArrayGet() || 1269 instruction->IsArraySet() || 1270 instruction->IsLoadClass() || 1271 instruction->IsLoadString() || 1272 instruction->IsInstanceOf() || 1273 instruction->IsCheckCast() || 1274 (instruction->IsInvokeVirtual() && instruction->GetLocations()->Intrinsified()) || 1275 (instruction->IsInvokeStaticOrDirect() && instruction->GetLocations()->Intrinsified())) 1276 << "instruction->DebugName()=" << instruction->DebugName() 1277 << " slow_path->GetDescription()=" << slow_path->GetDescription(); 1278} 1279 1280void SlowPathCode::SaveLiveRegisters(CodeGenerator* codegen, LocationSummary* locations) { 1281 size_t stack_offset = codegen->GetFirstRegisterSlotInSlowPath(); 1282 1283 const uint32_t core_spills = codegen->GetSlowPathSpills(locations, /* core_registers */ true); 1284 for (uint32_t i : LowToHighBits(core_spills)) { 1285 // If the register holds an object, update the stack mask. 1286 if (locations->RegisterContainsObject(i)) { 1287 locations->SetStackBit(stack_offset / kVRegSize); 1288 } 1289 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 1290 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 1291 saved_core_stack_offsets_[i] = stack_offset; 1292 stack_offset += codegen->SaveCoreRegister(stack_offset, i); 1293 } 1294 1295 const uint32_t fp_spills = codegen->GetSlowPathSpills(locations, /* core_registers */ false); 1296 for (uint32_t i : LowToHighBits(fp_spills)) { 1297 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 1298 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 1299 saved_fpu_stack_offsets_[i] = stack_offset; 1300 stack_offset += codegen->SaveFloatingPointRegister(stack_offset, i); 1301 } 1302} 1303 1304void SlowPathCode::RestoreLiveRegisters(CodeGenerator* codegen, LocationSummary* locations) { 1305 size_t stack_offset = codegen->GetFirstRegisterSlotInSlowPath(); 1306 1307 const uint32_t core_spills = codegen->GetSlowPathSpills(locations, /* core_registers */ true); 1308 for (uint32_t i : LowToHighBits(core_spills)) { 1309 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 1310 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 1311 stack_offset += codegen->RestoreCoreRegister(stack_offset, i); 1312 } 1313 1314 const uint32_t fp_spills = codegen->GetSlowPathSpills(locations, /* core_registers */ false); 1315 for (uint32_t i : LowToHighBits(fp_spills)) { 1316 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 1317 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 1318 stack_offset += codegen->RestoreFloatingPointRegister(stack_offset, i); 1319 } 1320} 1321 1322void CodeGenerator::CreateSystemArrayCopyLocationSummary(HInvoke* invoke) { 1323 // Check to see if we have known failures that will cause us to have to bail out 1324 // to the runtime, and just generate the runtime call directly. 1325 HIntConstant* src_pos = invoke->InputAt(1)->AsIntConstant(); 1326 HIntConstant* dest_pos = invoke->InputAt(3)->AsIntConstant(); 1327 1328 // The positions must be non-negative. 1329 if ((src_pos != nullptr && src_pos->GetValue() < 0) || 1330 (dest_pos != nullptr && dest_pos->GetValue() < 0)) { 1331 // We will have to fail anyways. 1332 return; 1333 } 1334 1335 // The length must be >= 0. 1336 HIntConstant* length = invoke->InputAt(4)->AsIntConstant(); 1337 if (length != nullptr) { 1338 int32_t len = length->GetValue(); 1339 if (len < 0) { 1340 // Just call as normal. 1341 return; 1342 } 1343 } 1344 1345 SystemArrayCopyOptimizations optimizations(invoke); 1346 1347 if (optimizations.GetDestinationIsSource()) { 1348 if (src_pos != nullptr && dest_pos != nullptr && src_pos->GetValue() < dest_pos->GetValue()) { 1349 // We only support backward copying if source and destination are the same. 1350 return; 1351 } 1352 } 1353 1354 if (optimizations.GetDestinationIsPrimitiveArray() || optimizations.GetSourceIsPrimitiveArray()) { 1355 // We currently don't intrinsify primitive copying. 1356 return; 1357 } 1358 1359 ArenaAllocator* allocator = invoke->GetBlock()->GetGraph()->GetArena(); 1360 LocationSummary* locations = new (allocator) LocationSummary(invoke, 1361 LocationSummary::kCallOnSlowPath, 1362 kIntrinsified); 1363 // arraycopy(Object src, int src_pos, Object dest, int dest_pos, int length). 1364 locations->SetInAt(0, Location::RequiresRegister()); 1365 locations->SetInAt(1, Location::RegisterOrConstant(invoke->InputAt(1))); 1366 locations->SetInAt(2, Location::RequiresRegister()); 1367 locations->SetInAt(3, Location::RegisterOrConstant(invoke->InputAt(3))); 1368 locations->SetInAt(4, Location::RegisterOrConstant(invoke->InputAt(4))); 1369 1370 locations->AddTemp(Location::RequiresRegister()); 1371 locations->AddTemp(Location::RequiresRegister()); 1372 locations->AddTemp(Location::RequiresRegister()); 1373} 1374 1375uint32_t CodeGenerator::GetReferenceSlowFlagOffset() const { 1376 ScopedObjectAccess soa(Thread::Current()); 1377 mirror::Class* klass = mirror::Reference::GetJavaLangRefReference(); 1378 DCHECK(klass->IsInitialized()); 1379 return klass->GetSlowPathFlagOffset().Uint32Value(); 1380} 1381 1382uint32_t CodeGenerator::GetReferenceDisableFlagOffset() const { 1383 ScopedObjectAccess soa(Thread::Current()); 1384 mirror::Class* klass = mirror::Reference::GetJavaLangRefReference(); 1385 DCHECK(klass->IsInitialized()); 1386 return klass->GetDisableIntrinsicFlagOffset().Uint32Value(); 1387} 1388 1389void CodeGenerator::EmitJitRoots(uint8_t* code, 1390 Handle<mirror::ObjectArray<mirror::Object>> roots, 1391 const uint8_t* roots_data) { 1392 DCHECK_EQ(static_cast<size_t>(roots->GetLength()), GetNumberOfJitRoots()); 1393 ClassLinker* class_linker = Runtime::Current()->GetClassLinker(); 1394 size_t index = 0; 1395 for (auto& entry : jit_string_roots_) { 1396 // Update the `roots` with the string, and replace the address temporarily 1397 // stored to the index in the table. 1398 uint64_t address = entry.second; 1399 roots->Set(index, reinterpret_cast<StackReference<mirror::String>*>(address)->AsMirrorPtr()); 1400 DCHECK(roots->Get(index) != nullptr); 1401 entry.second = index; 1402 // Ensure the string is strongly interned. This is a requirement on how the JIT 1403 // handles strings. b/32995596 1404 class_linker->GetInternTable()->InternStrong( 1405 reinterpret_cast<mirror::String*>(roots->Get(index))); 1406 ++index; 1407 } 1408 for (auto& entry : jit_class_roots_) { 1409 // Update the `roots` with the class, and replace the address temporarily 1410 // stored to the index in the table. 1411 uint64_t address = entry.second; 1412 roots->Set(index, reinterpret_cast<StackReference<mirror::Class>*>(address)->AsMirrorPtr()); 1413 DCHECK(roots->Get(index) != nullptr); 1414 entry.second = index; 1415 ++index; 1416 } 1417 EmitJitRootPatches(code, roots_data); 1418} 1419 1420QuickEntrypointEnum CodeGenerator::GetArrayAllocationEntrypoint(Handle<mirror::Class> array_klass) { 1421 ScopedObjectAccess soa(Thread::Current()); 1422 if (array_klass == nullptr) { 1423 // This can only happen for non-primitive arrays, as primitive arrays can always 1424 // be resolved. 1425 return kQuickAllocArrayResolved32; 1426 } 1427 1428 switch (array_klass->GetComponentSize()) { 1429 case 1: return kQuickAllocArrayResolved8; 1430 case 2: return kQuickAllocArrayResolved16; 1431 case 4: return kQuickAllocArrayResolved32; 1432 case 8: return kQuickAllocArrayResolved64; 1433 } 1434 LOG(FATAL) << "Unreachable"; 1435 return kQuickAllocArrayResolved; 1436} 1437 1438} // namespace art 1439