code_generator.cc revision 4acd03638fcdb4e5d1666f8eec7eb3bf6d6be035
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 + FrameEntrySpillSize(), 308 kStackAlignment)); 309 } 310} 311 312void CodeGenerator::CreateCommonInvokeLocationSummary( 313 HInvoke* invoke, InvokeDexCallingConventionVisitor* visitor) { 314 ArenaAllocator* allocator = invoke->GetBlock()->GetGraph()->GetArena(); 315 LocationSummary* locations = new (allocator) LocationSummary(invoke, 316 LocationSummary::kCallOnMainOnly); 317 318 for (size_t i = 0; i < invoke->GetNumberOfArguments(); i++) { 319 HInstruction* input = invoke->InputAt(i); 320 locations->SetInAt(i, visitor->GetNextLocation(input->GetType())); 321 } 322 323 locations->SetOut(visitor->GetReturnLocation(invoke->GetType())); 324 325 if (invoke->IsInvokeStaticOrDirect()) { 326 HInvokeStaticOrDirect* call = invoke->AsInvokeStaticOrDirect(); 327 switch (call->GetMethodLoadKind()) { 328 case HInvokeStaticOrDirect::MethodLoadKind::kRecursive: 329 locations->SetInAt(call->GetSpecialInputIndex(), visitor->GetMethodLocation()); 330 break; 331 case HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod: 332 locations->AddTemp(visitor->GetMethodLocation()); 333 locations->SetInAt(call->GetSpecialInputIndex(), Location::RequiresRegister()); 334 break; 335 default: 336 locations->AddTemp(visitor->GetMethodLocation()); 337 break; 338 } 339 } else { 340 locations->AddTemp(visitor->GetMethodLocation()); 341 } 342} 343 344void CodeGenerator::GenerateInvokeUnresolvedRuntimeCall(HInvokeUnresolved* invoke) { 345 MoveConstant(invoke->GetLocations()->GetTemp(0), invoke->GetDexMethodIndex()); 346 347 // Initialize to anything to silent compiler warnings. 348 QuickEntrypointEnum entrypoint = kQuickInvokeStaticTrampolineWithAccessCheck; 349 switch (invoke->GetInvokeType()) { 350 case kStatic: 351 entrypoint = kQuickInvokeStaticTrampolineWithAccessCheck; 352 break; 353 case kDirect: 354 entrypoint = kQuickInvokeDirectTrampolineWithAccessCheck; 355 break; 356 case kVirtual: 357 entrypoint = kQuickInvokeVirtualTrampolineWithAccessCheck; 358 break; 359 case kSuper: 360 entrypoint = kQuickInvokeSuperTrampolineWithAccessCheck; 361 break; 362 case kInterface: 363 entrypoint = kQuickInvokeInterfaceTrampolineWithAccessCheck; 364 break; 365 } 366 InvokeRuntime(entrypoint, invoke, invoke->GetDexPc(), nullptr); 367} 368 369void CodeGenerator::CreateUnresolvedFieldLocationSummary( 370 HInstruction* field_access, 371 Primitive::Type field_type, 372 const FieldAccessCallingConvention& calling_convention) { 373 bool is_instance = field_access->IsUnresolvedInstanceFieldGet() 374 || field_access->IsUnresolvedInstanceFieldSet(); 375 bool is_get = field_access->IsUnresolvedInstanceFieldGet() 376 || field_access->IsUnresolvedStaticFieldGet(); 377 378 ArenaAllocator* allocator = field_access->GetBlock()->GetGraph()->GetArena(); 379 LocationSummary* locations = 380 new (allocator) LocationSummary(field_access, LocationSummary::kCallOnMainOnly); 381 382 locations->AddTemp(calling_convention.GetFieldIndexLocation()); 383 384 if (is_instance) { 385 // Add the `this` object for instance field accesses. 386 locations->SetInAt(0, calling_convention.GetObjectLocation()); 387 } 388 389 // Note that pSetXXStatic/pGetXXStatic always takes/returns an int or int64 390 // regardless of the the type. Because of that we forced to special case 391 // the access to floating point values. 392 if (is_get) { 393 if (Primitive::IsFloatingPointType(field_type)) { 394 // The return value will be stored in regular registers while register 395 // allocator expects it in a floating point register. 396 // Note We don't need to request additional temps because the return 397 // register(s) are already blocked due the call and they may overlap with 398 // the input or field index. 399 // The transfer between the two will be done at codegen level. 400 locations->SetOut(calling_convention.GetFpuLocation(field_type)); 401 } else { 402 locations->SetOut(calling_convention.GetReturnLocation(field_type)); 403 } 404 } else { 405 size_t set_index = is_instance ? 1 : 0; 406 if (Primitive::IsFloatingPointType(field_type)) { 407 // The set value comes from a float location while the calling convention 408 // expects it in a regular register location. Allocate a temp for it and 409 // make the transfer at codegen. 410 AddLocationAsTemp(calling_convention.GetSetValueLocation(field_type, is_instance), locations); 411 locations->SetInAt(set_index, calling_convention.GetFpuLocation(field_type)); 412 } else { 413 locations->SetInAt(set_index, 414 calling_convention.GetSetValueLocation(field_type, is_instance)); 415 } 416 } 417} 418 419void CodeGenerator::GenerateUnresolvedFieldAccess( 420 HInstruction* field_access, 421 Primitive::Type field_type, 422 uint32_t field_index, 423 uint32_t dex_pc, 424 const FieldAccessCallingConvention& calling_convention) { 425 LocationSummary* locations = field_access->GetLocations(); 426 427 MoveConstant(locations->GetTemp(0), field_index); 428 429 bool is_instance = field_access->IsUnresolvedInstanceFieldGet() 430 || field_access->IsUnresolvedInstanceFieldSet(); 431 bool is_get = field_access->IsUnresolvedInstanceFieldGet() 432 || field_access->IsUnresolvedStaticFieldGet(); 433 434 if (!is_get && Primitive::IsFloatingPointType(field_type)) { 435 // Copy the float value to be set into the calling convention register. 436 // Note that using directly the temp location is problematic as we don't 437 // support temp register pairs. To avoid boilerplate conversion code, use 438 // the location from the calling convention. 439 MoveLocation(calling_convention.GetSetValueLocation(field_type, is_instance), 440 locations->InAt(is_instance ? 1 : 0), 441 (Primitive::Is64BitType(field_type) ? Primitive::kPrimLong : Primitive::kPrimInt)); 442 } 443 444 QuickEntrypointEnum entrypoint = kQuickSet8Static; // Initialize to anything to avoid warnings. 445 switch (field_type) { 446 case Primitive::kPrimBoolean: 447 entrypoint = is_instance 448 ? (is_get ? kQuickGetBooleanInstance : kQuickSet8Instance) 449 : (is_get ? kQuickGetBooleanStatic : kQuickSet8Static); 450 break; 451 case Primitive::kPrimByte: 452 entrypoint = is_instance 453 ? (is_get ? kQuickGetByteInstance : kQuickSet8Instance) 454 : (is_get ? kQuickGetByteStatic : kQuickSet8Static); 455 break; 456 case Primitive::kPrimShort: 457 entrypoint = is_instance 458 ? (is_get ? kQuickGetShortInstance : kQuickSet16Instance) 459 : (is_get ? kQuickGetShortStatic : kQuickSet16Static); 460 break; 461 case Primitive::kPrimChar: 462 entrypoint = is_instance 463 ? (is_get ? kQuickGetCharInstance : kQuickSet16Instance) 464 : (is_get ? kQuickGetCharStatic : kQuickSet16Static); 465 break; 466 case Primitive::kPrimInt: 467 case Primitive::kPrimFloat: 468 entrypoint = is_instance 469 ? (is_get ? kQuickGet32Instance : kQuickSet32Instance) 470 : (is_get ? kQuickGet32Static : kQuickSet32Static); 471 break; 472 case Primitive::kPrimNot: 473 entrypoint = is_instance 474 ? (is_get ? kQuickGetObjInstance : kQuickSetObjInstance) 475 : (is_get ? kQuickGetObjStatic : kQuickSetObjStatic); 476 break; 477 case Primitive::kPrimLong: 478 case Primitive::kPrimDouble: 479 entrypoint = is_instance 480 ? (is_get ? kQuickGet64Instance : kQuickSet64Instance) 481 : (is_get ? kQuickGet64Static : kQuickSet64Static); 482 break; 483 default: 484 LOG(FATAL) << "Invalid type " << field_type; 485 } 486 InvokeRuntime(entrypoint, field_access, dex_pc, nullptr); 487 488 if (is_get && Primitive::IsFloatingPointType(field_type)) { 489 MoveLocation(locations->Out(), calling_convention.GetReturnLocation(field_type), field_type); 490 } 491} 492 493// TODO: Remove argument `code_generator_supports_read_barrier` when 494// all code generators have read barrier support. 495void CodeGenerator::CreateLoadClassLocationSummary(HLoadClass* cls, 496 Location runtime_type_index_location, 497 Location runtime_return_location, 498 bool code_generator_supports_read_barrier) { 499 ArenaAllocator* allocator = cls->GetBlock()->GetGraph()->GetArena(); 500 LocationSummary::CallKind call_kind = cls->NeedsAccessCheck() 501 ? LocationSummary::kCallOnMainOnly 502 : (((code_generator_supports_read_barrier && kEmitCompilerReadBarrier) || 503 cls->CanCallRuntime()) 504 ? LocationSummary::kCallOnSlowPath 505 : LocationSummary::kNoCall); 506 LocationSummary* locations = new (allocator) LocationSummary(cls, call_kind); 507 if (cls->NeedsAccessCheck()) { 508 locations->SetInAt(0, Location::NoLocation()); 509 locations->AddTemp(runtime_type_index_location); 510 locations->SetOut(runtime_return_location); 511 } else { 512 locations->SetInAt(0, Location::RequiresRegister()); 513 locations->SetOut(Location::RequiresRegister()); 514 } 515} 516 517 518void CodeGenerator::BlockIfInRegister(Location location, bool is_out) const { 519 // The DCHECKS below check that a register is not specified twice in 520 // the summary. The out location can overlap with an input, so we need 521 // to special case it. 522 if (location.IsRegister()) { 523 DCHECK(is_out || !blocked_core_registers_[location.reg()]); 524 blocked_core_registers_[location.reg()] = true; 525 } else if (location.IsFpuRegister()) { 526 DCHECK(is_out || !blocked_fpu_registers_[location.reg()]); 527 blocked_fpu_registers_[location.reg()] = true; 528 } else if (location.IsFpuRegisterPair()) { 529 DCHECK(is_out || !blocked_fpu_registers_[location.AsFpuRegisterPairLow<int>()]); 530 blocked_fpu_registers_[location.AsFpuRegisterPairLow<int>()] = true; 531 DCHECK(is_out || !blocked_fpu_registers_[location.AsFpuRegisterPairHigh<int>()]); 532 blocked_fpu_registers_[location.AsFpuRegisterPairHigh<int>()] = true; 533 } else if (location.IsRegisterPair()) { 534 DCHECK(is_out || !blocked_core_registers_[location.AsRegisterPairLow<int>()]); 535 blocked_core_registers_[location.AsRegisterPairLow<int>()] = true; 536 DCHECK(is_out || !blocked_core_registers_[location.AsRegisterPairHigh<int>()]); 537 blocked_core_registers_[location.AsRegisterPairHigh<int>()] = true; 538 } 539} 540 541void CodeGenerator::AllocateLocations(HInstruction* instruction) { 542 instruction->Accept(GetLocationBuilder()); 543 DCHECK(CheckTypeConsistency(instruction)); 544 LocationSummary* locations = instruction->GetLocations(); 545 if (!instruction->IsSuspendCheckEntry()) { 546 if (locations != nullptr) { 547 if (locations->CanCall()) { 548 MarkNotLeaf(); 549 } else if (locations->Intrinsified() && 550 instruction->IsInvokeStaticOrDirect() && 551 !instruction->AsInvokeStaticOrDirect()->HasCurrentMethodInput()) { 552 // A static method call that has been fully intrinsified, and cannot call on the slow 553 // path or refer to the current method directly, no longer needs current method. 554 return; 555 } 556 } 557 if (instruction->NeedsCurrentMethod()) { 558 SetRequiresCurrentMethod(); 559 } 560 } 561} 562 563void CodeGenerator::MaybeRecordStat(MethodCompilationStat compilation_stat, size_t count) const { 564 if (stats_ != nullptr) { 565 stats_->RecordStat(compilation_stat, count); 566 } 567} 568 569std::unique_ptr<CodeGenerator> CodeGenerator::Create(HGraph* graph, 570 InstructionSet instruction_set, 571 const InstructionSetFeatures& isa_features, 572 const CompilerOptions& compiler_options, 573 OptimizingCompilerStats* stats) { 574 ArenaAllocator* arena = graph->GetArena(); 575 switch (instruction_set) { 576#ifdef ART_ENABLE_CODEGEN_arm 577 case kArm: 578 case kThumb2: { 579 if (kArmUseVIXL32) { 580 return std::unique_ptr<CodeGenerator>( 581 new (arena) arm::CodeGeneratorARMVIXL(graph, 582 *isa_features.AsArmInstructionSetFeatures(), 583 compiler_options, 584 stats)); 585 } else { 586 return std::unique_ptr<CodeGenerator>( 587 new (arena) arm::CodeGeneratorARM(graph, 588 *isa_features.AsArmInstructionSetFeatures(), 589 compiler_options, 590 stats)); 591 } 592 } 593#endif 594#ifdef ART_ENABLE_CODEGEN_arm64 595 case kArm64: { 596 return std::unique_ptr<CodeGenerator>( 597 new (arena) arm64::CodeGeneratorARM64(graph, 598 *isa_features.AsArm64InstructionSetFeatures(), 599 compiler_options, 600 stats)); 601 } 602#endif 603#ifdef ART_ENABLE_CODEGEN_mips 604 case kMips: { 605 return std::unique_ptr<CodeGenerator>( 606 new (arena) mips::CodeGeneratorMIPS(graph, 607 *isa_features.AsMipsInstructionSetFeatures(), 608 compiler_options, 609 stats)); 610 } 611#endif 612#ifdef ART_ENABLE_CODEGEN_mips64 613 case kMips64: { 614 return std::unique_ptr<CodeGenerator>( 615 new (arena) mips64::CodeGeneratorMIPS64(graph, 616 *isa_features.AsMips64InstructionSetFeatures(), 617 compiler_options, 618 stats)); 619 } 620#endif 621#ifdef ART_ENABLE_CODEGEN_x86 622 case kX86: { 623 return std::unique_ptr<CodeGenerator>( 624 new (arena) x86::CodeGeneratorX86(graph, 625 *isa_features.AsX86InstructionSetFeatures(), 626 compiler_options, 627 stats)); 628 } 629#endif 630#ifdef ART_ENABLE_CODEGEN_x86_64 631 case kX86_64: { 632 return std::unique_ptr<CodeGenerator>( 633 new (arena) x86_64::CodeGeneratorX86_64(graph, 634 *isa_features.AsX86_64InstructionSetFeatures(), 635 compiler_options, 636 stats)); 637 } 638#endif 639 default: 640 return nullptr; 641 } 642} 643 644size_t CodeGenerator::ComputeStackMapsSize() { 645 return stack_map_stream_.PrepareForFillIn(); 646} 647 648static void CheckCovers(uint32_t dex_pc, 649 const HGraph& graph, 650 const CodeInfo& code_info, 651 const ArenaVector<HSuspendCheck*>& loop_headers, 652 ArenaVector<size_t>* covered) { 653 CodeInfoEncoding encoding = code_info.ExtractEncoding(); 654 for (size_t i = 0; i < loop_headers.size(); ++i) { 655 if (loop_headers[i]->GetDexPc() == dex_pc) { 656 if (graph.IsCompilingOsr()) { 657 DCHECK(code_info.GetOsrStackMapForDexPc(dex_pc, encoding).IsValid()); 658 } 659 ++(*covered)[i]; 660 } 661 } 662} 663 664// Debug helper to ensure loop entries in compiled code are matched by 665// dex branch instructions. 666static void CheckLoopEntriesCanBeUsedForOsr(const HGraph& graph, 667 const CodeInfo& code_info, 668 const DexFile::CodeItem& code_item) { 669 if (graph.HasTryCatch()) { 670 // One can write loops through try/catch, which we do not support for OSR anyway. 671 return; 672 } 673 ArenaVector<HSuspendCheck*> loop_headers(graph.GetArena()->Adapter(kArenaAllocMisc)); 674 for (HBasicBlock* block : graph.GetReversePostOrder()) { 675 if (block->IsLoopHeader()) { 676 HSuspendCheck* suspend_check = block->GetLoopInformation()->GetSuspendCheck(); 677 if (!suspend_check->GetEnvironment()->IsFromInlinedInvoke()) { 678 loop_headers.push_back(suspend_check); 679 } 680 } 681 } 682 ArenaVector<size_t> covered(loop_headers.size(), 0, graph.GetArena()->Adapter(kArenaAllocMisc)); 683 const uint16_t* code_ptr = code_item.insns_; 684 const uint16_t* code_end = code_item.insns_ + code_item.insns_size_in_code_units_; 685 686 size_t dex_pc = 0; 687 while (code_ptr < code_end) { 688 const Instruction& instruction = *Instruction::At(code_ptr); 689 if (instruction.IsBranch()) { 690 uint32_t target = dex_pc + instruction.GetTargetOffset(); 691 CheckCovers(target, graph, code_info, loop_headers, &covered); 692 } else if (instruction.IsSwitch()) { 693 DexSwitchTable table(instruction, dex_pc); 694 uint16_t num_entries = table.GetNumEntries(); 695 size_t offset = table.GetFirstValueIndex(); 696 697 // Use a larger loop counter type to avoid overflow issues. 698 for (size_t i = 0; i < num_entries; ++i) { 699 // The target of the case. 700 uint32_t target = dex_pc + table.GetEntryAt(i + offset); 701 CheckCovers(target, graph, code_info, loop_headers, &covered); 702 } 703 } 704 dex_pc += instruction.SizeInCodeUnits(); 705 code_ptr += instruction.SizeInCodeUnits(); 706 } 707 708 for (size_t i = 0; i < covered.size(); ++i) { 709 DCHECK_NE(covered[i], 0u) << "Loop in compiled code has no dex branch equivalent"; 710 } 711} 712 713void CodeGenerator::BuildStackMaps(MemoryRegion region, const DexFile::CodeItem& code_item) { 714 stack_map_stream_.FillIn(region); 715 if (kIsDebugBuild) { 716 CheckLoopEntriesCanBeUsedForOsr(*graph_, CodeInfo(region), code_item); 717 } 718} 719 720void CodeGenerator::RecordPcInfo(HInstruction* instruction, 721 uint32_t dex_pc, 722 SlowPathCode* slow_path) { 723 if (instruction != nullptr) { 724 // The code generated for some type conversions 725 // may call the runtime, thus normally requiring a subsequent 726 // call to this method. However, the method verifier does not 727 // produce PC information for certain instructions, which are 728 // considered "atomic" (they cannot join a GC). 729 // Therefore we do not currently record PC information for such 730 // instructions. As this may change later, we added this special 731 // case so that code generators may nevertheless call 732 // CodeGenerator::RecordPcInfo without triggering an error in 733 // CodeGenerator::BuildNativeGCMap ("Missing ref for dex pc 0x") 734 // thereafter. 735 if (instruction->IsTypeConversion()) { 736 return; 737 } 738 if (instruction->IsRem()) { 739 Primitive::Type type = instruction->AsRem()->GetResultType(); 740 if ((type == Primitive::kPrimFloat) || (type == Primitive::kPrimDouble)) { 741 return; 742 } 743 } 744 } 745 746 uint32_t outer_dex_pc = dex_pc; 747 uint32_t outer_environment_size = 0; 748 uint32_t inlining_depth = 0; 749 if (instruction != nullptr) { 750 for (HEnvironment* environment = instruction->GetEnvironment(); 751 environment != nullptr; 752 environment = environment->GetParent()) { 753 outer_dex_pc = environment->GetDexPc(); 754 outer_environment_size = environment->Size(); 755 if (environment != instruction->GetEnvironment()) { 756 inlining_depth++; 757 } 758 } 759 } 760 761 // Collect PC infos for the mapping table. 762 uint32_t native_pc = GetAssembler()->CodePosition(); 763 764 if (instruction == nullptr) { 765 // For stack overflow checks and native-debug-info entries without dex register 766 // mapping (i.e. start of basic block or start of slow path). 767 stack_map_stream_.BeginStackMapEntry(outer_dex_pc, native_pc, 0, 0, 0, 0); 768 stack_map_stream_.EndStackMapEntry(); 769 return; 770 } 771 LocationSummary* locations = instruction->GetLocations(); 772 773 uint32_t register_mask = locations->GetRegisterMask(); 774 DCHECK_EQ(register_mask & ~locations->GetLiveRegisters()->GetCoreRegisters(), 0u); 775 if (locations->OnlyCallsOnSlowPath()) { 776 // In case of slow path, we currently set the location of caller-save registers 777 // to register (instead of their stack location when pushed before the slow-path 778 // call). Therefore register_mask contains both callee-save and caller-save 779 // registers that hold objects. We must remove the spilled caller-save from the 780 // mask, since they will be overwritten by the callee. 781 uint32_t spills = GetSlowPathSpills(locations, /* core_registers */ true); 782 register_mask &= ~spills; 783 } else { 784 // The register mask must be a subset of callee-save registers. 785 DCHECK_EQ(register_mask & core_callee_save_mask_, register_mask); 786 } 787 stack_map_stream_.BeginStackMapEntry(outer_dex_pc, 788 native_pc, 789 register_mask, 790 locations->GetStackMask(), 791 outer_environment_size, 792 inlining_depth); 793 794 EmitEnvironment(instruction->GetEnvironment(), slow_path); 795 stack_map_stream_.EndStackMapEntry(); 796 797 HLoopInformation* info = instruction->GetBlock()->GetLoopInformation(); 798 if (instruction->IsSuspendCheck() && 799 (info != nullptr) && 800 graph_->IsCompilingOsr() && 801 (inlining_depth == 0)) { 802 DCHECK_EQ(info->GetSuspendCheck(), instruction); 803 // We duplicate the stack map as a marker that this stack map can be an OSR entry. 804 // Duplicating it avoids having the runtime recognize and skip an OSR stack map. 805 DCHECK(info->IsIrreducible()); 806 stack_map_stream_.BeginStackMapEntry( 807 dex_pc, native_pc, register_mask, locations->GetStackMask(), outer_environment_size, 0); 808 EmitEnvironment(instruction->GetEnvironment(), slow_path); 809 stack_map_stream_.EndStackMapEntry(); 810 if (kIsDebugBuild) { 811 HEnvironment* environment = instruction->GetEnvironment(); 812 for (size_t i = 0, environment_size = environment->Size(); i < environment_size; ++i) { 813 HInstruction* in_environment = environment->GetInstructionAt(i); 814 if (in_environment != nullptr) { 815 DCHECK(in_environment->IsPhi() || in_environment->IsConstant()); 816 Location location = environment->GetLocationAt(i); 817 DCHECK(location.IsStackSlot() || 818 location.IsDoubleStackSlot() || 819 location.IsConstant() || 820 location.IsInvalid()); 821 if (location.IsStackSlot() || location.IsDoubleStackSlot()) { 822 DCHECK_LT(location.GetStackIndex(), static_cast<int32_t>(GetFrameSize())); 823 } 824 } 825 } 826 } 827 } else if (kIsDebugBuild) { 828 // Ensure stack maps are unique, by checking that the native pc in the stack map 829 // last emitted is different than the native pc of the stack map just emitted. 830 size_t number_of_stack_maps = stack_map_stream_.GetNumberOfStackMaps(); 831 if (number_of_stack_maps > 1) { 832 DCHECK_NE(stack_map_stream_.GetStackMap(number_of_stack_maps - 1).native_pc_offset, 833 stack_map_stream_.GetStackMap(number_of_stack_maps - 2).native_pc_offset); 834 } 835 } 836} 837 838bool CodeGenerator::HasStackMapAtCurrentPc() { 839 uint32_t pc = GetAssembler()->CodeSize(); 840 size_t count = stack_map_stream_.GetNumberOfStackMaps(); 841 return count > 0 && stack_map_stream_.GetStackMap(count - 1).native_pc_offset == pc; 842} 843 844void CodeGenerator::MaybeRecordNativeDebugInfo(HInstruction* instruction, 845 uint32_t dex_pc, 846 SlowPathCode* slow_path) { 847 if (GetCompilerOptions().GetNativeDebuggable() && dex_pc != kNoDexPc) { 848 if (HasStackMapAtCurrentPc()) { 849 // Ensure that we do not collide with the stack map of the previous instruction. 850 GenerateNop(); 851 } 852 RecordPcInfo(instruction, dex_pc, slow_path); 853 } 854} 855 856void CodeGenerator::RecordCatchBlockInfo() { 857 ArenaAllocator* arena = graph_->GetArena(); 858 859 for (HBasicBlock* block : *block_order_) { 860 if (!block->IsCatchBlock()) { 861 continue; 862 } 863 864 uint32_t dex_pc = block->GetDexPc(); 865 uint32_t num_vregs = graph_->GetNumberOfVRegs(); 866 uint32_t inlining_depth = 0; // Inlining of catch blocks is not supported at the moment. 867 uint32_t native_pc = GetAddressOf(block); 868 uint32_t register_mask = 0; // Not used. 869 870 // The stack mask is not used, so we leave it empty. 871 ArenaBitVector* stack_mask = 872 ArenaBitVector::Create(arena, 0, /* expandable */ true, kArenaAllocCodeGenerator); 873 874 stack_map_stream_.BeginStackMapEntry(dex_pc, 875 native_pc, 876 register_mask, 877 stack_mask, 878 num_vregs, 879 inlining_depth); 880 881 HInstruction* current_phi = block->GetFirstPhi(); 882 for (size_t vreg = 0; vreg < num_vregs; ++vreg) { 883 while (current_phi != nullptr && current_phi->AsPhi()->GetRegNumber() < vreg) { 884 HInstruction* next_phi = current_phi->GetNext(); 885 DCHECK(next_phi == nullptr || 886 current_phi->AsPhi()->GetRegNumber() <= next_phi->AsPhi()->GetRegNumber()) 887 << "Phis need to be sorted by vreg number to keep this a linear-time loop."; 888 current_phi = next_phi; 889 } 890 891 if (current_phi == nullptr || current_phi->AsPhi()->GetRegNumber() != vreg) { 892 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kNone, 0); 893 } else { 894 Location location = current_phi->GetLiveInterval()->ToLocation(); 895 switch (location.GetKind()) { 896 case Location::kStackSlot: { 897 stack_map_stream_.AddDexRegisterEntry( 898 DexRegisterLocation::Kind::kInStack, location.GetStackIndex()); 899 break; 900 } 901 case Location::kDoubleStackSlot: { 902 stack_map_stream_.AddDexRegisterEntry( 903 DexRegisterLocation::Kind::kInStack, location.GetStackIndex()); 904 stack_map_stream_.AddDexRegisterEntry( 905 DexRegisterLocation::Kind::kInStack, location.GetHighStackIndex(kVRegSize)); 906 ++vreg; 907 DCHECK_LT(vreg, num_vregs); 908 break; 909 } 910 default: { 911 // All catch phis must be allocated to a stack slot. 912 LOG(FATAL) << "Unexpected kind " << location.GetKind(); 913 UNREACHABLE(); 914 } 915 } 916 } 917 } 918 919 stack_map_stream_.EndStackMapEntry(); 920 } 921} 922 923void CodeGenerator::EmitEnvironment(HEnvironment* environment, SlowPathCode* slow_path) { 924 if (environment == nullptr) return; 925 926 if (environment->GetParent() != nullptr) { 927 // We emit the parent environment first. 928 EmitEnvironment(environment->GetParent(), slow_path); 929 stack_map_stream_.BeginInlineInfoEntry(environment->GetMethodIdx(), 930 environment->GetDexPc(), 931 environment->GetInvokeType(), 932 environment->Size()); 933 } 934 935 // Walk over the environment, and record the location of dex registers. 936 for (size_t i = 0, environment_size = environment->Size(); i < environment_size; ++i) { 937 HInstruction* current = environment->GetInstructionAt(i); 938 if (current == nullptr) { 939 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kNone, 0); 940 continue; 941 } 942 943 Location location = environment->GetLocationAt(i); 944 switch (location.GetKind()) { 945 case Location::kConstant: { 946 DCHECK_EQ(current, location.GetConstant()); 947 if (current->IsLongConstant()) { 948 int64_t value = current->AsLongConstant()->GetValue(); 949 stack_map_stream_.AddDexRegisterEntry( 950 DexRegisterLocation::Kind::kConstant, Low32Bits(value)); 951 stack_map_stream_.AddDexRegisterEntry( 952 DexRegisterLocation::Kind::kConstant, High32Bits(value)); 953 ++i; 954 DCHECK_LT(i, environment_size); 955 } else if (current->IsDoubleConstant()) { 956 int64_t value = bit_cast<int64_t, double>(current->AsDoubleConstant()->GetValue()); 957 stack_map_stream_.AddDexRegisterEntry( 958 DexRegisterLocation::Kind::kConstant, Low32Bits(value)); 959 stack_map_stream_.AddDexRegisterEntry( 960 DexRegisterLocation::Kind::kConstant, High32Bits(value)); 961 ++i; 962 DCHECK_LT(i, environment_size); 963 } else if (current->IsIntConstant()) { 964 int32_t value = current->AsIntConstant()->GetValue(); 965 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kConstant, value); 966 } else if (current->IsNullConstant()) { 967 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kConstant, 0); 968 } else { 969 DCHECK(current->IsFloatConstant()) << current->DebugName(); 970 int32_t value = bit_cast<int32_t, float>(current->AsFloatConstant()->GetValue()); 971 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kConstant, value); 972 } 973 break; 974 } 975 976 case Location::kStackSlot: { 977 stack_map_stream_.AddDexRegisterEntry( 978 DexRegisterLocation::Kind::kInStack, location.GetStackIndex()); 979 break; 980 } 981 982 case Location::kDoubleStackSlot: { 983 stack_map_stream_.AddDexRegisterEntry( 984 DexRegisterLocation::Kind::kInStack, location.GetStackIndex()); 985 stack_map_stream_.AddDexRegisterEntry( 986 DexRegisterLocation::Kind::kInStack, location.GetHighStackIndex(kVRegSize)); 987 ++i; 988 DCHECK_LT(i, environment_size); 989 break; 990 } 991 992 case Location::kRegister : { 993 int id = location.reg(); 994 if (slow_path != nullptr && slow_path->IsCoreRegisterSaved(id)) { 995 uint32_t offset = slow_path->GetStackOffsetOfCoreRegister(id); 996 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInStack, offset); 997 if (current->GetType() == Primitive::kPrimLong) { 998 stack_map_stream_.AddDexRegisterEntry( 999 DexRegisterLocation::Kind::kInStack, offset + kVRegSize); 1000 ++i; 1001 DCHECK_LT(i, environment_size); 1002 } 1003 } else { 1004 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInRegister, id); 1005 if (current->GetType() == Primitive::kPrimLong) { 1006 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInRegisterHigh, id); 1007 ++i; 1008 DCHECK_LT(i, environment_size); 1009 } 1010 } 1011 break; 1012 } 1013 1014 case Location::kFpuRegister : { 1015 int id = location.reg(); 1016 if (slow_path != nullptr && slow_path->IsFpuRegisterSaved(id)) { 1017 uint32_t offset = slow_path->GetStackOffsetOfFpuRegister(id); 1018 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInStack, offset); 1019 if (current->GetType() == Primitive::kPrimDouble) { 1020 stack_map_stream_.AddDexRegisterEntry( 1021 DexRegisterLocation::Kind::kInStack, offset + kVRegSize); 1022 ++i; 1023 DCHECK_LT(i, environment_size); 1024 } 1025 } else { 1026 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInFpuRegister, id); 1027 if (current->GetType() == Primitive::kPrimDouble) { 1028 stack_map_stream_.AddDexRegisterEntry( 1029 DexRegisterLocation::Kind::kInFpuRegisterHigh, id); 1030 ++i; 1031 DCHECK_LT(i, environment_size); 1032 } 1033 } 1034 break; 1035 } 1036 1037 case Location::kFpuRegisterPair : { 1038 int low = location.low(); 1039 int high = location.high(); 1040 if (slow_path != nullptr && slow_path->IsFpuRegisterSaved(low)) { 1041 uint32_t offset = slow_path->GetStackOffsetOfFpuRegister(low); 1042 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInStack, offset); 1043 } else { 1044 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInFpuRegister, low); 1045 } 1046 if (slow_path != nullptr && slow_path->IsFpuRegisterSaved(high)) { 1047 uint32_t offset = slow_path->GetStackOffsetOfFpuRegister(high); 1048 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInStack, offset); 1049 ++i; 1050 } else { 1051 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInFpuRegister, high); 1052 ++i; 1053 } 1054 DCHECK_LT(i, environment_size); 1055 break; 1056 } 1057 1058 case Location::kRegisterPair : { 1059 int low = location.low(); 1060 int high = location.high(); 1061 if (slow_path != nullptr && slow_path->IsCoreRegisterSaved(low)) { 1062 uint32_t offset = slow_path->GetStackOffsetOfCoreRegister(low); 1063 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInStack, offset); 1064 } else { 1065 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInRegister, low); 1066 } 1067 if (slow_path != nullptr && slow_path->IsCoreRegisterSaved(high)) { 1068 uint32_t offset = slow_path->GetStackOffsetOfCoreRegister(high); 1069 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInStack, offset); 1070 } else { 1071 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInRegister, high); 1072 } 1073 ++i; 1074 DCHECK_LT(i, environment_size); 1075 break; 1076 } 1077 1078 case Location::kInvalid: { 1079 stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kNone, 0); 1080 break; 1081 } 1082 1083 default: 1084 LOG(FATAL) << "Unexpected kind " << location.GetKind(); 1085 } 1086 } 1087 1088 if (environment->GetParent() != nullptr) { 1089 stack_map_stream_.EndInlineInfoEntry(); 1090 } 1091} 1092 1093bool CodeGenerator::CanMoveNullCheckToUser(HNullCheck* null_check) { 1094 HInstruction* first_next_not_move = null_check->GetNextDisregardingMoves(); 1095 1096 return (first_next_not_move != nullptr) 1097 && first_next_not_move->CanDoImplicitNullCheckOn(null_check->InputAt(0)); 1098} 1099 1100void CodeGenerator::MaybeRecordImplicitNullCheck(HInstruction* instr) { 1101 if (!compiler_options_.GetImplicitNullChecks()) { 1102 return; 1103 } 1104 1105 // If we are from a static path don't record the pc as we can't throw NPE. 1106 // NB: having the checks here makes the code much less verbose in the arch 1107 // specific code generators. 1108 if (instr->IsStaticFieldSet() || instr->IsStaticFieldGet()) { 1109 return; 1110 } 1111 1112 if (!instr->CanDoImplicitNullCheckOn(instr->InputAt(0))) { 1113 return; 1114 } 1115 1116 // Find the first previous instruction which is not a move. 1117 HInstruction* first_prev_not_move = instr->GetPreviousDisregardingMoves(); 1118 1119 // If the instruction is a null check it means that `instr` is the first user 1120 // and needs to record the pc. 1121 if (first_prev_not_move != nullptr && first_prev_not_move->IsNullCheck()) { 1122 HNullCheck* null_check = first_prev_not_move->AsNullCheck(); 1123 // TODO: The parallel moves modify the environment. Their changes need to be 1124 // reverted otherwise the stack maps at the throw point will not be correct. 1125 RecordPcInfo(null_check, null_check->GetDexPc()); 1126 } 1127} 1128 1129LocationSummary* CodeGenerator::CreateThrowingSlowPathLocations(HInstruction* instruction, 1130 RegisterSet caller_saves) { 1131 // Note: Using kNoCall allows the method to be treated as leaf (and eliminate the 1132 // HSuspendCheck from entry block). However, it will still get a valid stack frame 1133 // because the HNullCheck needs an environment. 1134 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 1135 // When throwing from a try block, we may need to retrieve dalvik registers from 1136 // physical registers and we also need to set up stack mask for GC. This is 1137 // implicitly achieved by passing kCallOnSlowPath to the LocationSummary. 1138 bool can_throw_into_catch_block = instruction->CanThrowIntoCatchBlock(); 1139 if (can_throw_into_catch_block) { 1140 call_kind = LocationSummary::kCallOnSlowPath; 1141 } 1142 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 1143 if (can_throw_into_catch_block && compiler_options_.GetImplicitNullChecks()) { 1144 locations->SetCustomSlowPathCallerSaves(caller_saves); // Default: no caller-save registers. 1145 } 1146 DCHECK(!instruction->HasUses()); 1147 return locations; 1148} 1149 1150void CodeGenerator::GenerateNullCheck(HNullCheck* instruction) { 1151 if (compiler_options_.GetImplicitNullChecks()) { 1152 MaybeRecordStat(kImplicitNullCheckGenerated); 1153 GenerateImplicitNullCheck(instruction); 1154 } else { 1155 MaybeRecordStat(kExplicitNullCheckGenerated); 1156 GenerateExplicitNullCheck(instruction); 1157 } 1158} 1159 1160void CodeGenerator::ClearSpillSlotsFromLoopPhisInStackMap(HSuspendCheck* suspend_check) const { 1161 LocationSummary* locations = suspend_check->GetLocations(); 1162 HBasicBlock* block = suspend_check->GetBlock(); 1163 DCHECK(block->GetLoopInformation()->GetSuspendCheck() == suspend_check); 1164 DCHECK(block->IsLoopHeader()); 1165 1166 for (HInstructionIterator it(block->GetPhis()); !it.Done(); it.Advance()) { 1167 HInstruction* current = it.Current(); 1168 LiveInterval* interval = current->GetLiveInterval(); 1169 // We only need to clear bits of loop phis containing objects and allocated in register. 1170 // Loop phis allocated on stack already have the object in the stack. 1171 if (current->GetType() == Primitive::kPrimNot 1172 && interval->HasRegister() 1173 && interval->HasSpillSlot()) { 1174 locations->ClearStackBit(interval->GetSpillSlot() / kVRegSize); 1175 } 1176 } 1177} 1178 1179void CodeGenerator::EmitParallelMoves(Location from1, 1180 Location to1, 1181 Primitive::Type type1, 1182 Location from2, 1183 Location to2, 1184 Primitive::Type type2) { 1185 HParallelMove parallel_move(GetGraph()->GetArena()); 1186 parallel_move.AddMove(from1, to1, type1, nullptr); 1187 parallel_move.AddMove(from2, to2, type2, nullptr); 1188 GetMoveResolver()->EmitNativeCode(¶llel_move); 1189} 1190 1191void CodeGenerator::ValidateInvokeRuntime(QuickEntrypointEnum entrypoint, 1192 HInstruction* instruction, 1193 SlowPathCode* slow_path) { 1194 // Ensure that the call kind indication given to the register allocator is 1195 // coherent with the runtime call generated. 1196 if (slow_path == nullptr) { 1197 DCHECK(instruction->GetLocations()->WillCall()) 1198 << "instruction->DebugName()=" << instruction->DebugName(); 1199 } else { 1200 DCHECK(instruction->GetLocations()->CallsOnSlowPath() || slow_path->IsFatal()) 1201 << "instruction->DebugName()=" << instruction->DebugName() 1202 << " slow_path->GetDescription()=" << slow_path->GetDescription(); 1203 } 1204 1205 // Check that the GC side effect is set when required. 1206 // TODO: Reverse EntrypointCanTriggerGC 1207 if (EntrypointCanTriggerGC(entrypoint)) { 1208 if (slow_path == nullptr) { 1209 DCHECK(instruction->GetSideEffects().Includes(SideEffects::CanTriggerGC())) 1210 << "instruction->DebugName()=" << instruction->DebugName() 1211 << " instruction->GetSideEffects().ToString()=" 1212 << instruction->GetSideEffects().ToString(); 1213 } else { 1214 DCHECK(instruction->GetSideEffects().Includes(SideEffects::CanTriggerGC()) || 1215 // When (non-Baker) read barriers are enabled, some instructions 1216 // use a slow path to emit a read barrier, which does not trigger 1217 // GC. 1218 (kEmitCompilerReadBarrier && 1219 !kUseBakerReadBarrier && 1220 (instruction->IsInstanceFieldGet() || 1221 instruction->IsStaticFieldGet() || 1222 instruction->IsArrayGet() || 1223 instruction->IsLoadClass() || 1224 instruction->IsLoadString() || 1225 instruction->IsInstanceOf() || 1226 instruction->IsCheckCast() || 1227 (instruction->IsInvokeVirtual() && instruction->GetLocations()->Intrinsified())))) 1228 << "instruction->DebugName()=" << instruction->DebugName() 1229 << " instruction->GetSideEffects().ToString()=" 1230 << instruction->GetSideEffects().ToString() 1231 << " slow_path->GetDescription()=" << slow_path->GetDescription(); 1232 } 1233 } else { 1234 // The GC side effect is not required for the instruction. But the instruction might still have 1235 // it, for example if it calls other entrypoints requiring it. 1236 } 1237 1238 // Check the coherency of leaf information. 1239 DCHECK(instruction->IsSuspendCheck() 1240 || ((slow_path != nullptr) && slow_path->IsFatal()) 1241 || instruction->GetLocations()->CanCall() 1242 || !IsLeafMethod()) 1243 << instruction->DebugName() << ((slow_path != nullptr) ? slow_path->GetDescription() : ""); 1244} 1245 1246void CodeGenerator::ValidateInvokeRuntimeWithoutRecordingPcInfo(HInstruction* instruction, 1247 SlowPathCode* slow_path) { 1248 DCHECK(instruction->GetLocations()->OnlyCallsOnSlowPath()) 1249 << "instruction->DebugName()=" << instruction->DebugName() 1250 << " slow_path->GetDescription()=" << slow_path->GetDescription(); 1251 // Only the Baker read barrier marking slow path used by certains 1252 // instructions is expected to invoke the runtime without recording 1253 // PC-related information. 1254 DCHECK(kUseBakerReadBarrier); 1255 DCHECK(instruction->IsInstanceFieldGet() || 1256 instruction->IsStaticFieldGet() || 1257 instruction->IsArrayGet() || 1258 instruction->IsArraySet() || 1259 instruction->IsLoadClass() || 1260 instruction->IsLoadString() || 1261 instruction->IsInstanceOf() || 1262 instruction->IsCheckCast() || 1263 (instruction->IsInvokeVirtual() && instruction->GetLocations()->Intrinsified()) || 1264 (instruction->IsInvokeStaticOrDirect() && instruction->GetLocations()->Intrinsified())) 1265 << "instruction->DebugName()=" << instruction->DebugName() 1266 << " slow_path->GetDescription()=" << slow_path->GetDescription(); 1267} 1268 1269void SlowPathCode::SaveLiveRegisters(CodeGenerator* codegen, LocationSummary* locations) { 1270 size_t stack_offset = codegen->GetFirstRegisterSlotInSlowPath(); 1271 1272 const uint32_t core_spills = codegen->GetSlowPathSpills(locations, /* core_registers */ true); 1273 for (uint32_t i : LowToHighBits(core_spills)) { 1274 // If the register holds an object, update the stack mask. 1275 if (locations->RegisterContainsObject(i)) { 1276 locations->SetStackBit(stack_offset / kVRegSize); 1277 } 1278 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 1279 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 1280 saved_core_stack_offsets_[i] = stack_offset; 1281 stack_offset += codegen->SaveCoreRegister(stack_offset, i); 1282 } 1283 1284 const uint32_t fp_spills = codegen->GetSlowPathSpills(locations, /* core_registers */ false); 1285 for (uint32_t i : LowToHighBits(fp_spills)) { 1286 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 1287 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 1288 saved_fpu_stack_offsets_[i] = stack_offset; 1289 stack_offset += codegen->SaveFloatingPointRegister(stack_offset, i); 1290 } 1291} 1292 1293void SlowPathCode::RestoreLiveRegisters(CodeGenerator* codegen, LocationSummary* locations) { 1294 size_t stack_offset = codegen->GetFirstRegisterSlotInSlowPath(); 1295 1296 const uint32_t core_spills = codegen->GetSlowPathSpills(locations, /* core_registers */ true); 1297 for (uint32_t i : LowToHighBits(core_spills)) { 1298 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 1299 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 1300 stack_offset += codegen->RestoreCoreRegister(stack_offset, i); 1301 } 1302 1303 const uint32_t fp_spills = codegen->GetSlowPathSpills(locations, /* core_registers */ false); 1304 for (uint32_t i : LowToHighBits(fp_spills)) { 1305 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 1306 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 1307 stack_offset += codegen->RestoreFloatingPointRegister(stack_offset, i); 1308 } 1309} 1310 1311void CodeGenerator::CreateSystemArrayCopyLocationSummary(HInvoke* invoke) { 1312 // Check to see if we have known failures that will cause us to have to bail out 1313 // to the runtime, and just generate the runtime call directly. 1314 HIntConstant* src_pos = invoke->InputAt(1)->AsIntConstant(); 1315 HIntConstant* dest_pos = invoke->InputAt(3)->AsIntConstant(); 1316 1317 // The positions must be non-negative. 1318 if ((src_pos != nullptr && src_pos->GetValue() < 0) || 1319 (dest_pos != nullptr && dest_pos->GetValue() < 0)) { 1320 // We will have to fail anyways. 1321 return; 1322 } 1323 1324 // The length must be >= 0. 1325 HIntConstant* length = invoke->InputAt(4)->AsIntConstant(); 1326 if (length != nullptr) { 1327 int32_t len = length->GetValue(); 1328 if (len < 0) { 1329 // Just call as normal. 1330 return; 1331 } 1332 } 1333 1334 SystemArrayCopyOptimizations optimizations(invoke); 1335 1336 if (optimizations.GetDestinationIsSource()) { 1337 if (src_pos != nullptr && dest_pos != nullptr && src_pos->GetValue() < dest_pos->GetValue()) { 1338 // We only support backward copying if source and destination are the same. 1339 return; 1340 } 1341 } 1342 1343 if (optimizations.GetDestinationIsPrimitiveArray() || optimizations.GetSourceIsPrimitiveArray()) { 1344 // We currently don't intrinsify primitive copying. 1345 return; 1346 } 1347 1348 ArenaAllocator* allocator = invoke->GetBlock()->GetGraph()->GetArena(); 1349 LocationSummary* locations = new (allocator) LocationSummary(invoke, 1350 LocationSummary::kCallOnSlowPath, 1351 kIntrinsified); 1352 // arraycopy(Object src, int src_pos, Object dest, int dest_pos, int length). 1353 locations->SetInAt(0, Location::RequiresRegister()); 1354 locations->SetInAt(1, Location::RegisterOrConstant(invoke->InputAt(1))); 1355 locations->SetInAt(2, Location::RequiresRegister()); 1356 locations->SetInAt(3, Location::RegisterOrConstant(invoke->InputAt(3))); 1357 locations->SetInAt(4, Location::RegisterOrConstant(invoke->InputAt(4))); 1358 1359 locations->AddTemp(Location::RequiresRegister()); 1360 locations->AddTemp(Location::RequiresRegister()); 1361 locations->AddTemp(Location::RequiresRegister()); 1362} 1363 1364uint32_t CodeGenerator::GetReferenceSlowFlagOffset() const { 1365 ScopedObjectAccess soa(Thread::Current()); 1366 mirror::Class* klass = mirror::Reference::GetJavaLangRefReference(); 1367 DCHECK(klass->IsInitialized()); 1368 return klass->GetSlowPathFlagOffset().Uint32Value(); 1369} 1370 1371uint32_t CodeGenerator::GetReferenceDisableFlagOffset() const { 1372 ScopedObjectAccess soa(Thread::Current()); 1373 mirror::Class* klass = mirror::Reference::GetJavaLangRefReference(); 1374 DCHECK(klass->IsInitialized()); 1375 return klass->GetDisableIntrinsicFlagOffset().Uint32Value(); 1376} 1377 1378} // namespace art 1379