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 #include "code_generator_arm.h" 20 #include "code_generator_arm64.h" 21 #include "code_generator_x86.h" 22 #include "code_generator_x86_64.h" 23 #include "code_generator_mips64.h" 24 #include "compiled_method.h" 25 #include "dex/verified_method.h" 26 #include "driver/dex_compilation_unit.h" 27 #include "gc_map_builder.h" 28 #include "leb128.h" 29 #include "mapping_table.h" 30 #include "mirror/array-inl.h" 31 #include "mirror/object_array-inl.h" 32 #include "mirror/object_reference.h" 33 #include "ssa_liveness_analysis.h" 34 #include "utils/assembler.h" 35 #include "verifier/dex_gc_map.h" 36 #include "vmap_table.h" 37 38 namespace art { 39 40 // Return whether a location is consistent with a type. 41 static bool CheckType(Primitive::Type type, Location location) { 42 if (location.IsFpuRegister() 43 || (location.IsUnallocated() && (location.GetPolicy() == Location::kRequiresFpuRegister))) { 44 return (type == Primitive::kPrimFloat) || (type == Primitive::kPrimDouble); 45 } else if (location.IsRegister() || 46 (location.IsUnallocated() && (location.GetPolicy() == Location::kRequiresRegister))) { 47 return Primitive::IsIntegralType(type) || (type == Primitive::kPrimNot); 48 } else if (location.IsRegisterPair()) { 49 return type == Primitive::kPrimLong; 50 } else if (location.IsFpuRegisterPair()) { 51 return type == Primitive::kPrimDouble; 52 } else if (location.IsStackSlot()) { 53 return (Primitive::IsIntegralType(type) && type != Primitive::kPrimLong) 54 || (type == Primitive::kPrimFloat) 55 || (type == Primitive::kPrimNot); 56 } else if (location.IsDoubleStackSlot()) { 57 return (type == Primitive::kPrimLong) || (type == Primitive::kPrimDouble); 58 } else if (location.IsConstant()) { 59 if (location.GetConstant()->IsIntConstant()) { 60 return Primitive::IsIntegralType(type) && (type != Primitive::kPrimLong); 61 } else if (location.GetConstant()->IsNullConstant()) { 62 return type == Primitive::kPrimNot; 63 } else if (location.GetConstant()->IsLongConstant()) { 64 return type == Primitive::kPrimLong; 65 } else if (location.GetConstant()->IsFloatConstant()) { 66 return type == Primitive::kPrimFloat; 67 } else { 68 return location.GetConstant()->IsDoubleConstant() 69 && (type == Primitive::kPrimDouble); 70 } 71 } else { 72 return location.IsInvalid() || (location.GetPolicy() == Location::kAny); 73 } 74 } 75 76 // Check that a location summary is consistent with an instruction. 77 static bool CheckTypeConsistency(HInstruction* instruction) { 78 LocationSummary* locations = instruction->GetLocations(); 79 if (locations == nullptr) { 80 return true; 81 } 82 83 if (locations->Out().IsUnallocated() 84 && (locations->Out().GetPolicy() == Location::kSameAsFirstInput)) { 85 DCHECK(CheckType(instruction->GetType(), locations->InAt(0))) 86 << instruction->GetType() 87 << " " << locations->InAt(0); 88 } else { 89 DCHECK(CheckType(instruction->GetType(), locations->Out())) 90 << instruction->GetType() 91 << " " << locations->Out(); 92 } 93 94 for (size_t i = 0, e = instruction->InputCount(); i < e; ++i) { 95 DCHECK(CheckType(instruction->InputAt(i)->GetType(), locations->InAt(i))) 96 << instruction->InputAt(i)->GetType() 97 << " " << locations->InAt(i); 98 } 99 100 HEnvironment* environment = instruction->GetEnvironment(); 101 for (size_t i = 0; i < instruction->EnvironmentSize(); ++i) { 102 if (environment->GetInstructionAt(i) != nullptr) { 103 Primitive::Type type = environment->GetInstructionAt(i)->GetType(); 104 DCHECK(CheckType(type, environment->GetLocationAt(i))) 105 << type << " " << environment->GetLocationAt(i); 106 } else { 107 DCHECK(environment->GetLocationAt(i).IsInvalid()) 108 << environment->GetLocationAt(i); 109 } 110 } 111 return true; 112 } 113 114 size_t CodeGenerator::GetCacheOffset(uint32_t index) { 115 return mirror::ObjectArray<mirror::Object>::OffsetOfElement(index).SizeValue(); 116 } 117 118 size_t CodeGenerator::GetCachePointerOffset(uint32_t index) { 119 auto pointer_size = InstructionSetPointerSize(GetInstructionSet()); 120 return mirror::Array::DataOffset(pointer_size).Uint32Value() + pointer_size * index; 121 } 122 123 void CodeGenerator::CompileBaseline(CodeAllocator* allocator, bool is_leaf) { 124 Initialize(); 125 if (!is_leaf) { 126 MarkNotLeaf(); 127 } 128 const bool is_64_bit = Is64BitInstructionSet(GetInstructionSet()); 129 InitializeCodeGeneration(GetGraph()->GetNumberOfLocalVRegs() 130 + GetGraph()->GetTemporariesVRegSlots() 131 + 1 /* filler */, 132 0, /* the baseline compiler does not have live registers at slow path */ 133 0, /* the baseline compiler does not have live registers at slow path */ 134 GetGraph()->GetMaximumNumberOfOutVRegs() 135 + (is_64_bit ? 2 : 1) /* current method */, 136 GetGraph()->GetBlocks()); 137 CompileInternal(allocator, /* is_baseline */ true); 138 } 139 140 bool CodeGenerator::GoesToNextBlock(HBasicBlock* current, HBasicBlock* next) const { 141 DCHECK_EQ(block_order_->Get(current_block_index_), current); 142 return GetNextBlockToEmit() == FirstNonEmptyBlock(next); 143 } 144 145 HBasicBlock* CodeGenerator::GetNextBlockToEmit() const { 146 for (size_t i = current_block_index_ + 1; i < block_order_->Size(); ++i) { 147 HBasicBlock* block = block_order_->Get(i); 148 if (!block->IsSingleGoto()) { 149 return block; 150 } 151 } 152 return nullptr; 153 } 154 155 HBasicBlock* CodeGenerator::FirstNonEmptyBlock(HBasicBlock* block) const { 156 while (block->IsSingleGoto()) { 157 block = block->GetSuccessors().Get(0); 158 } 159 return block; 160 } 161 162 void CodeGenerator::CompileInternal(CodeAllocator* allocator, bool is_baseline) { 163 is_baseline_ = is_baseline; 164 HGraphVisitor* instruction_visitor = GetInstructionVisitor(); 165 DCHECK_EQ(current_block_index_, 0u); 166 GenerateFrameEntry(); 167 DCHECK_EQ(GetAssembler()->cfi().GetCurrentCFAOffset(), static_cast<int>(frame_size_)); 168 for (size_t e = block_order_->Size(); current_block_index_ < e; ++current_block_index_) { 169 HBasicBlock* block = block_order_->Get(current_block_index_); 170 // Don't generate code for an empty block. Its predecessors will branch to its successor 171 // directly. Also, the label of that block will not be emitted, so this helps catch 172 // errors where we reference that label. 173 if (block->IsSingleGoto()) continue; 174 Bind(block); 175 for (HInstructionIterator it(block->GetInstructions()); !it.Done(); it.Advance()) { 176 HInstruction* current = it.Current(); 177 if (is_baseline) { 178 InitLocationsBaseline(current); 179 } 180 DCHECK(CheckTypeConsistency(current)); 181 current->Accept(instruction_visitor); 182 } 183 } 184 185 // Generate the slow paths. 186 for (size_t i = 0, e = slow_paths_.Size(); i < e; ++i) { 187 slow_paths_.Get(i)->EmitNativeCode(this); 188 } 189 190 // Finalize instructions in assember; 191 Finalize(allocator); 192 } 193 194 void CodeGenerator::CompileOptimized(CodeAllocator* allocator) { 195 // The register allocator already called `InitializeCodeGeneration`, 196 // where the frame size has been computed. 197 DCHECK(block_order_ != nullptr); 198 Initialize(); 199 CompileInternal(allocator, /* is_baseline */ false); 200 } 201 202 void CodeGenerator::Finalize(CodeAllocator* allocator) { 203 size_t code_size = GetAssembler()->CodeSize(); 204 uint8_t* buffer = allocator->Allocate(code_size); 205 206 MemoryRegion code(buffer, code_size); 207 GetAssembler()->FinalizeInstructions(code); 208 } 209 210 size_t CodeGenerator::FindFreeEntry(bool* array, size_t length) { 211 for (size_t i = 0; i < length; ++i) { 212 if (!array[i]) { 213 array[i] = true; 214 return i; 215 } 216 } 217 LOG(FATAL) << "Could not find a register in baseline register allocator"; 218 UNREACHABLE(); 219 } 220 221 size_t CodeGenerator::FindTwoFreeConsecutiveAlignedEntries(bool* array, size_t length) { 222 for (size_t i = 0; i < length - 1; i += 2) { 223 if (!array[i] && !array[i + 1]) { 224 array[i] = true; 225 array[i + 1] = true; 226 return i; 227 } 228 } 229 LOG(FATAL) << "Could not find a register in baseline register allocator"; 230 UNREACHABLE(); 231 } 232 233 void CodeGenerator::InitializeCodeGeneration(size_t number_of_spill_slots, 234 size_t maximum_number_of_live_core_registers, 235 size_t maximum_number_of_live_fp_registers, 236 size_t number_of_out_slots, 237 const GrowableArray<HBasicBlock*>& block_order) { 238 block_order_ = &block_order; 239 DCHECK(block_order_->Get(0) == GetGraph()->GetEntryBlock()); 240 ComputeSpillMask(); 241 first_register_slot_in_slow_path_ = (number_of_out_slots + number_of_spill_slots) * kVRegSize; 242 243 if (number_of_spill_slots == 0 244 && !HasAllocatedCalleeSaveRegisters() 245 && IsLeafMethod() 246 && !RequiresCurrentMethod()) { 247 DCHECK_EQ(maximum_number_of_live_core_registers, 0u); 248 DCHECK_EQ(maximum_number_of_live_fp_registers, 0u); 249 SetFrameSize(CallPushesPC() ? GetWordSize() : 0); 250 } else { 251 SetFrameSize(RoundUp( 252 number_of_spill_slots * kVRegSize 253 + number_of_out_slots * kVRegSize 254 + maximum_number_of_live_core_registers * GetWordSize() 255 + maximum_number_of_live_fp_registers * GetFloatingPointSpillSlotSize() 256 + FrameEntrySpillSize(), 257 kStackAlignment)); 258 } 259 } 260 261 Location CodeGenerator::GetTemporaryLocation(HTemporary* temp) const { 262 uint16_t number_of_locals = GetGraph()->GetNumberOfLocalVRegs(); 263 // The type of the previous instruction tells us if we need a single or double stack slot. 264 Primitive::Type type = temp->GetType(); 265 int32_t temp_size = (type == Primitive::kPrimLong) || (type == Primitive::kPrimDouble) ? 2 : 1; 266 // Use the temporary region (right below the dex registers). 267 int32_t slot = GetFrameSize() - FrameEntrySpillSize() 268 - kVRegSize // filler 269 - (number_of_locals * kVRegSize) 270 - ((temp_size + temp->GetIndex()) * kVRegSize); 271 return temp_size == 2 ? Location::DoubleStackSlot(slot) : Location::StackSlot(slot); 272 } 273 274 int32_t CodeGenerator::GetStackSlot(HLocal* local) const { 275 uint16_t reg_number = local->GetRegNumber(); 276 uint16_t number_of_locals = GetGraph()->GetNumberOfLocalVRegs(); 277 if (reg_number >= number_of_locals) { 278 // Local is a parameter of the method. It is stored in the caller's frame. 279 // TODO: Share this logic with StackVisitor::GetVRegOffsetFromQuickCode. 280 return GetFrameSize() + InstructionSetPointerSize(GetInstructionSet()) // ART method 281 + (reg_number - number_of_locals) * kVRegSize; 282 } else { 283 // Local is a temporary in this method. It is stored in this method's frame. 284 return GetFrameSize() - FrameEntrySpillSize() 285 - kVRegSize // filler. 286 - (number_of_locals * kVRegSize) 287 + (reg_number * kVRegSize); 288 } 289 } 290 291 void CodeGenerator::BlockIfInRegister(Location location, bool is_out) const { 292 // The DCHECKS below check that a register is not specified twice in 293 // the summary. The out location can overlap with an input, so we need 294 // to special case it. 295 if (location.IsRegister()) { 296 DCHECK(is_out || !blocked_core_registers_[location.reg()]); 297 blocked_core_registers_[location.reg()] = true; 298 } else if (location.IsFpuRegister()) { 299 DCHECK(is_out || !blocked_fpu_registers_[location.reg()]); 300 blocked_fpu_registers_[location.reg()] = true; 301 } else if (location.IsFpuRegisterPair()) { 302 DCHECK(is_out || !blocked_fpu_registers_[location.AsFpuRegisterPairLow<int>()]); 303 blocked_fpu_registers_[location.AsFpuRegisterPairLow<int>()] = true; 304 DCHECK(is_out || !blocked_fpu_registers_[location.AsFpuRegisterPairHigh<int>()]); 305 blocked_fpu_registers_[location.AsFpuRegisterPairHigh<int>()] = true; 306 } else if (location.IsRegisterPair()) { 307 DCHECK(is_out || !blocked_core_registers_[location.AsRegisterPairLow<int>()]); 308 blocked_core_registers_[location.AsRegisterPairLow<int>()] = true; 309 DCHECK(is_out || !blocked_core_registers_[location.AsRegisterPairHigh<int>()]); 310 blocked_core_registers_[location.AsRegisterPairHigh<int>()] = true; 311 } 312 } 313 314 void CodeGenerator::AllocateRegistersLocally(HInstruction* instruction) const { 315 LocationSummary* locations = instruction->GetLocations(); 316 if (locations == nullptr) return; 317 318 for (size_t i = 0, e = GetNumberOfCoreRegisters(); i < e; ++i) { 319 blocked_core_registers_[i] = false; 320 } 321 322 for (size_t i = 0, e = GetNumberOfFloatingPointRegisters(); i < e; ++i) { 323 blocked_fpu_registers_[i] = false; 324 } 325 326 for (size_t i = 0, e = number_of_register_pairs_; i < e; ++i) { 327 blocked_register_pairs_[i] = false; 328 } 329 330 // Mark all fixed input, temp and output registers as used. 331 for (size_t i = 0, e = locations->GetInputCount(); i < e; ++i) { 332 BlockIfInRegister(locations->InAt(i)); 333 } 334 335 for (size_t i = 0, e = locations->GetTempCount(); i < e; ++i) { 336 Location loc = locations->GetTemp(i); 337 BlockIfInRegister(loc); 338 } 339 Location result_location = locations->Out(); 340 if (locations->OutputCanOverlapWithInputs()) { 341 BlockIfInRegister(result_location, /* is_out */ true); 342 } 343 344 SetupBlockedRegisters(/* is_baseline */ true); 345 346 // Allocate all unallocated input locations. 347 for (size_t i = 0, e = locations->GetInputCount(); i < e; ++i) { 348 Location loc = locations->InAt(i); 349 HInstruction* input = instruction->InputAt(i); 350 if (loc.IsUnallocated()) { 351 if ((loc.GetPolicy() == Location::kRequiresRegister) 352 || (loc.GetPolicy() == Location::kRequiresFpuRegister)) { 353 loc = AllocateFreeRegister(input->GetType()); 354 } else { 355 DCHECK_EQ(loc.GetPolicy(), Location::kAny); 356 HLoadLocal* load = input->AsLoadLocal(); 357 if (load != nullptr) { 358 loc = GetStackLocation(load); 359 } else { 360 loc = AllocateFreeRegister(input->GetType()); 361 } 362 } 363 locations->SetInAt(i, loc); 364 } 365 } 366 367 // Allocate all unallocated temp locations. 368 for (size_t i = 0, e = locations->GetTempCount(); i < e; ++i) { 369 Location loc = locations->GetTemp(i); 370 if (loc.IsUnallocated()) { 371 switch (loc.GetPolicy()) { 372 case Location::kRequiresRegister: 373 // Allocate a core register (large enough to fit a 32-bit integer). 374 loc = AllocateFreeRegister(Primitive::kPrimInt); 375 break; 376 377 case Location::kRequiresFpuRegister: 378 // Allocate a core register (large enough to fit a 64-bit double). 379 loc = AllocateFreeRegister(Primitive::kPrimDouble); 380 break; 381 382 default: 383 LOG(FATAL) << "Unexpected policy for temporary location " 384 << loc.GetPolicy(); 385 } 386 locations->SetTempAt(i, loc); 387 } 388 } 389 if (result_location.IsUnallocated()) { 390 switch (result_location.GetPolicy()) { 391 case Location::kAny: 392 case Location::kRequiresRegister: 393 case Location::kRequiresFpuRegister: 394 result_location = AllocateFreeRegister(instruction->GetType()); 395 break; 396 case Location::kSameAsFirstInput: 397 result_location = locations->InAt(0); 398 break; 399 } 400 locations->UpdateOut(result_location); 401 } 402 } 403 404 void CodeGenerator::InitLocationsBaseline(HInstruction* instruction) { 405 AllocateLocations(instruction); 406 if (instruction->GetLocations() == nullptr) { 407 if (instruction->IsTemporary()) { 408 HInstruction* previous = instruction->GetPrevious(); 409 Location temp_location = GetTemporaryLocation(instruction->AsTemporary()); 410 Move(previous, temp_location, instruction); 411 } 412 return; 413 } 414 AllocateRegistersLocally(instruction); 415 for (size_t i = 0, e = instruction->InputCount(); i < e; ++i) { 416 Location location = instruction->GetLocations()->InAt(i); 417 HInstruction* input = instruction->InputAt(i); 418 if (location.IsValid()) { 419 // Move the input to the desired location. 420 if (input->GetNext()->IsTemporary()) { 421 // If the input was stored in a temporary, use that temporary to 422 // perform the move. 423 Move(input->GetNext(), location, instruction); 424 } else { 425 Move(input, location, instruction); 426 } 427 } 428 } 429 } 430 431 void CodeGenerator::AllocateLocations(HInstruction* instruction) { 432 instruction->Accept(GetLocationBuilder()); 433 DCHECK(CheckTypeConsistency(instruction)); 434 LocationSummary* locations = instruction->GetLocations(); 435 if (!instruction->IsSuspendCheckEntry()) { 436 if (locations != nullptr && locations->CanCall()) { 437 MarkNotLeaf(); 438 } 439 if (instruction->NeedsCurrentMethod()) { 440 SetRequiresCurrentMethod(); 441 } 442 } 443 } 444 445 CodeGenerator* CodeGenerator::Create(HGraph* graph, 446 InstructionSet instruction_set, 447 const InstructionSetFeatures& isa_features, 448 const CompilerOptions& compiler_options) { 449 switch (instruction_set) { 450 case kArm: 451 case kThumb2: { 452 return new arm::CodeGeneratorARM(graph, 453 *isa_features.AsArmInstructionSetFeatures(), 454 compiler_options); 455 } 456 case kArm64: { 457 return new arm64::CodeGeneratorARM64(graph, 458 *isa_features.AsArm64InstructionSetFeatures(), 459 compiler_options); 460 } 461 case kMips: 462 return nullptr; 463 case kMips64: { 464 return new mips64::CodeGeneratorMIPS64(graph, 465 *isa_features.AsMips64InstructionSetFeatures(), 466 compiler_options); 467 } 468 case kX86: { 469 return new x86::CodeGeneratorX86(graph, 470 *isa_features.AsX86InstructionSetFeatures(), 471 compiler_options); 472 } 473 case kX86_64: { 474 return new x86_64::CodeGeneratorX86_64(graph, 475 *isa_features.AsX86_64InstructionSetFeatures(), 476 compiler_options); 477 } 478 default: 479 return nullptr; 480 } 481 } 482 483 void CodeGenerator::BuildNativeGCMap( 484 std::vector<uint8_t>* data, const DexCompilationUnit& dex_compilation_unit) const { 485 const std::vector<uint8_t>& gc_map_raw = 486 dex_compilation_unit.GetVerifiedMethod()->GetDexGcMap(); 487 verifier::DexPcToReferenceMap dex_gc_map(&(gc_map_raw)[0]); 488 489 uint32_t max_native_offset = 0; 490 for (size_t i = 0; i < pc_infos_.Size(); i++) { 491 uint32_t native_offset = pc_infos_.Get(i).native_pc; 492 if (native_offset > max_native_offset) { 493 max_native_offset = native_offset; 494 } 495 } 496 497 GcMapBuilder builder(data, pc_infos_.Size(), max_native_offset, dex_gc_map.RegWidth()); 498 for (size_t i = 0; i < pc_infos_.Size(); i++) { 499 struct PcInfo pc_info = pc_infos_.Get(i); 500 uint32_t native_offset = pc_info.native_pc; 501 uint32_t dex_pc = pc_info.dex_pc; 502 const uint8_t* references = dex_gc_map.FindBitMap(dex_pc, false); 503 CHECK(references != nullptr) << "Missing ref for dex pc 0x" << std::hex << dex_pc; 504 builder.AddEntry(native_offset, references); 505 } 506 } 507 508 void CodeGenerator::BuildSourceMap(DefaultSrcMap* src_map) const { 509 for (size_t i = 0; i < pc_infos_.Size(); i++) { 510 struct PcInfo pc_info = pc_infos_.Get(i); 511 uint32_t pc2dex_offset = pc_info.native_pc; 512 int32_t pc2dex_dalvik_offset = pc_info.dex_pc; 513 src_map->push_back(SrcMapElem({pc2dex_offset, pc2dex_dalvik_offset})); 514 } 515 } 516 517 void CodeGenerator::BuildMappingTable(std::vector<uint8_t>* data) const { 518 uint32_t pc2dex_data_size = 0u; 519 uint32_t pc2dex_entries = pc_infos_.Size(); 520 uint32_t pc2dex_offset = 0u; 521 int32_t pc2dex_dalvik_offset = 0; 522 uint32_t dex2pc_data_size = 0u; 523 uint32_t dex2pc_entries = 0u; 524 uint32_t dex2pc_offset = 0u; 525 int32_t dex2pc_dalvik_offset = 0; 526 527 for (size_t i = 0; i < pc2dex_entries; i++) { 528 struct PcInfo pc_info = pc_infos_.Get(i); 529 pc2dex_data_size += UnsignedLeb128Size(pc_info.native_pc - pc2dex_offset); 530 pc2dex_data_size += SignedLeb128Size(pc_info.dex_pc - pc2dex_dalvik_offset); 531 pc2dex_offset = pc_info.native_pc; 532 pc2dex_dalvik_offset = pc_info.dex_pc; 533 } 534 535 // Walk over the blocks and find which ones correspond to catch block entries. 536 for (size_t i = 0; i < graph_->GetBlocks().Size(); ++i) { 537 HBasicBlock* block = graph_->GetBlocks().Get(i); 538 if (block->IsCatchBlock()) { 539 intptr_t native_pc = GetAddressOf(block); 540 ++dex2pc_entries; 541 dex2pc_data_size += UnsignedLeb128Size(native_pc - dex2pc_offset); 542 dex2pc_data_size += SignedLeb128Size(block->GetDexPc() - dex2pc_dalvik_offset); 543 dex2pc_offset = native_pc; 544 dex2pc_dalvik_offset = block->GetDexPc(); 545 } 546 } 547 548 uint32_t total_entries = pc2dex_entries + dex2pc_entries; 549 uint32_t hdr_data_size = UnsignedLeb128Size(total_entries) + UnsignedLeb128Size(pc2dex_entries); 550 uint32_t data_size = hdr_data_size + pc2dex_data_size + dex2pc_data_size; 551 data->resize(data_size); 552 553 uint8_t* data_ptr = &(*data)[0]; 554 uint8_t* write_pos = data_ptr; 555 556 write_pos = EncodeUnsignedLeb128(write_pos, total_entries); 557 write_pos = EncodeUnsignedLeb128(write_pos, pc2dex_entries); 558 DCHECK_EQ(static_cast<size_t>(write_pos - data_ptr), hdr_data_size); 559 uint8_t* write_pos2 = write_pos + pc2dex_data_size; 560 561 pc2dex_offset = 0u; 562 pc2dex_dalvik_offset = 0u; 563 dex2pc_offset = 0u; 564 dex2pc_dalvik_offset = 0u; 565 566 for (size_t i = 0; i < pc2dex_entries; i++) { 567 struct PcInfo pc_info = pc_infos_.Get(i); 568 DCHECK(pc2dex_offset <= pc_info.native_pc); 569 write_pos = EncodeUnsignedLeb128(write_pos, pc_info.native_pc - pc2dex_offset); 570 write_pos = EncodeSignedLeb128(write_pos, pc_info.dex_pc - pc2dex_dalvik_offset); 571 pc2dex_offset = pc_info.native_pc; 572 pc2dex_dalvik_offset = pc_info.dex_pc; 573 } 574 575 for (size_t i = 0; i < graph_->GetBlocks().Size(); ++i) { 576 HBasicBlock* block = graph_->GetBlocks().Get(i); 577 if (block->IsCatchBlock()) { 578 intptr_t native_pc = GetAddressOf(block); 579 write_pos2 = EncodeUnsignedLeb128(write_pos2, native_pc - dex2pc_offset); 580 write_pos2 = EncodeSignedLeb128(write_pos2, block->GetDexPc() - dex2pc_dalvik_offset); 581 dex2pc_offset = native_pc; 582 dex2pc_dalvik_offset = block->GetDexPc(); 583 } 584 } 585 586 587 DCHECK_EQ(static_cast<size_t>(write_pos - data_ptr), hdr_data_size + pc2dex_data_size); 588 DCHECK_EQ(static_cast<size_t>(write_pos2 - data_ptr), data_size); 589 590 if (kIsDebugBuild) { 591 // Verify the encoded table holds the expected data. 592 MappingTable table(data_ptr); 593 CHECK_EQ(table.TotalSize(), total_entries); 594 CHECK_EQ(table.PcToDexSize(), pc2dex_entries); 595 auto it = table.PcToDexBegin(); 596 auto it2 = table.DexToPcBegin(); 597 for (size_t i = 0; i < pc2dex_entries; i++) { 598 struct PcInfo pc_info = pc_infos_.Get(i); 599 CHECK_EQ(pc_info.native_pc, it.NativePcOffset()); 600 CHECK_EQ(pc_info.dex_pc, it.DexPc()); 601 ++it; 602 } 603 for (size_t i = 0; i < graph_->GetBlocks().Size(); ++i) { 604 HBasicBlock* block = graph_->GetBlocks().Get(i); 605 if (block->IsCatchBlock()) { 606 CHECK_EQ(GetAddressOf(block), it2.NativePcOffset()); 607 CHECK_EQ(block->GetDexPc(), it2.DexPc()); 608 ++it2; 609 } 610 } 611 CHECK(it == table.PcToDexEnd()); 612 CHECK(it2 == table.DexToPcEnd()); 613 } 614 } 615 616 void CodeGenerator::BuildVMapTable(std::vector<uint8_t>* data) const { 617 Leb128EncodingVector vmap_encoder; 618 // We currently don't use callee-saved registers. 619 size_t size = 0 + 1 /* marker */ + 0; 620 vmap_encoder.Reserve(size + 1u); // All values are likely to be one byte in ULEB128 (<128). 621 vmap_encoder.PushBackUnsigned(size); 622 vmap_encoder.PushBackUnsigned(VmapTable::kAdjustedFpMarker); 623 624 *data = vmap_encoder.GetData(); 625 } 626 627 void CodeGenerator::BuildStackMaps(std::vector<uint8_t>* data) { 628 uint32_t size = stack_map_stream_.PrepareForFillIn(); 629 data->resize(size); 630 MemoryRegion region(data->data(), size); 631 stack_map_stream_.FillIn(region); 632 } 633 634 void CodeGenerator::RecordPcInfo(HInstruction* instruction, 635 uint32_t dex_pc, 636 SlowPathCode* slow_path) { 637 if (instruction != nullptr) { 638 // The code generated for some type conversions and comparisons 639 // may call the runtime, thus normally requiring a subsequent 640 // call to this method. However, the method verifier does not 641 // produce PC information for certain instructions, which are 642 // considered "atomic" (they cannot join a GC). 643 // Therefore we do not currently record PC information for such 644 // instructions. As this may change later, we added this special 645 // case so that code generators may nevertheless call 646 // CodeGenerator::RecordPcInfo without triggering an error in 647 // CodeGenerator::BuildNativeGCMap ("Missing ref for dex pc 0x") 648 // thereafter. 649 if (instruction->IsTypeConversion() || instruction->IsCompare()) { 650 return; 651 } 652 if (instruction->IsRem()) { 653 Primitive::Type type = instruction->AsRem()->GetResultType(); 654 if ((type == Primitive::kPrimFloat) || (type == Primitive::kPrimDouble)) { 655 return; 656 } 657 } 658 } 659 660 // Collect PC infos for the mapping table. 661 struct PcInfo pc_info; 662 pc_info.dex_pc = dex_pc; 663 pc_info.native_pc = GetAssembler()->CodeSize(); 664 pc_infos_.Add(pc_info); 665 666 uint32_t inlining_depth = 0; 667 668 if (instruction == nullptr) { 669 // For stack overflow checks. 670 stack_map_stream_.BeginStackMapEntry(dex_pc, pc_info.native_pc, 0, 0, 0, inlining_depth); 671 stack_map_stream_.EndStackMapEntry(); 672 return; 673 } 674 LocationSummary* locations = instruction->GetLocations(); 675 HEnvironment* environment = instruction->GetEnvironment(); 676 size_t environment_size = instruction->EnvironmentSize(); 677 678 uint32_t register_mask = locations->GetRegisterMask(); 679 if (locations->OnlyCallsOnSlowPath()) { 680 // In case of slow path, we currently set the location of caller-save registers 681 // to register (instead of their stack location when pushed before the slow-path 682 // call). Therefore register_mask contains both callee-save and caller-save 683 // registers that hold objects. We must remove the caller-save from the mask, since 684 // they will be overwritten by the callee. 685 register_mask &= core_callee_save_mask_; 686 } 687 // The register mask must be a subset of callee-save registers. 688 DCHECK_EQ(register_mask & core_callee_save_mask_, register_mask); 689 stack_map_stream_.BeginStackMapEntry(dex_pc, 690 pc_info.native_pc, 691 register_mask, 692 locations->GetStackMask(), 693 environment_size, 694 inlining_depth); 695 if (environment != nullptr) { 696 // TODO: Handle parent environment. 697 DCHECK(environment->GetParent() == nullptr); 698 DCHECK_EQ(environment->GetDexPc(), dex_pc); 699 } 700 701 // Walk over the environment, and record the location of dex registers. 702 for (size_t i = 0; i < environment_size; ++i) { 703 HInstruction* current = environment->GetInstructionAt(i); 704 if (current == nullptr) { 705 stack_map_stream_.AddDexRegisterEntry(i, DexRegisterLocation::Kind::kNone, 0); 706 continue; 707 } 708 709 Location location = environment->GetLocationAt(i); 710 switch (location.GetKind()) { 711 case Location::kConstant: { 712 DCHECK_EQ(current, location.GetConstant()); 713 if (current->IsLongConstant()) { 714 int64_t value = current->AsLongConstant()->GetValue(); 715 stack_map_stream_.AddDexRegisterEntry( 716 i, DexRegisterLocation::Kind::kConstant, Low32Bits(value)); 717 stack_map_stream_.AddDexRegisterEntry( 718 ++i, DexRegisterLocation::Kind::kConstant, High32Bits(value)); 719 DCHECK_LT(i, environment_size); 720 } else if (current->IsDoubleConstant()) { 721 int64_t value = bit_cast<int64_t, double>(current->AsDoubleConstant()->GetValue()); 722 stack_map_stream_.AddDexRegisterEntry( 723 i, DexRegisterLocation::Kind::kConstant, Low32Bits(value)); 724 stack_map_stream_.AddDexRegisterEntry( 725 ++i, DexRegisterLocation::Kind::kConstant, High32Bits(value)); 726 DCHECK_LT(i, environment_size); 727 } else if (current->IsIntConstant()) { 728 int32_t value = current->AsIntConstant()->GetValue(); 729 stack_map_stream_.AddDexRegisterEntry(i, DexRegisterLocation::Kind::kConstant, value); 730 } else if (current->IsNullConstant()) { 731 stack_map_stream_.AddDexRegisterEntry(i, DexRegisterLocation::Kind::kConstant, 0); 732 } else { 733 DCHECK(current->IsFloatConstant()) << current->DebugName(); 734 int32_t value = bit_cast<int32_t, float>(current->AsFloatConstant()->GetValue()); 735 stack_map_stream_.AddDexRegisterEntry(i, DexRegisterLocation::Kind::kConstant, value); 736 } 737 break; 738 } 739 740 case Location::kStackSlot: { 741 stack_map_stream_.AddDexRegisterEntry( 742 i, DexRegisterLocation::Kind::kInStack, location.GetStackIndex()); 743 break; 744 } 745 746 case Location::kDoubleStackSlot: { 747 stack_map_stream_.AddDexRegisterEntry( 748 i, DexRegisterLocation::Kind::kInStack, location.GetStackIndex()); 749 stack_map_stream_.AddDexRegisterEntry( 750 ++i, DexRegisterLocation::Kind::kInStack, location.GetHighStackIndex(kVRegSize)); 751 DCHECK_LT(i, environment_size); 752 break; 753 } 754 755 case Location::kRegister : { 756 int id = location.reg(); 757 if (slow_path != nullptr && slow_path->IsCoreRegisterSaved(id)) { 758 uint32_t offset = slow_path->GetStackOffsetOfCoreRegister(id); 759 stack_map_stream_.AddDexRegisterEntry(i, DexRegisterLocation::Kind::kInStack, offset); 760 if (current->GetType() == Primitive::kPrimLong) { 761 stack_map_stream_.AddDexRegisterEntry( 762 ++i, DexRegisterLocation::Kind::kInStack, offset + kVRegSize); 763 DCHECK_LT(i, environment_size); 764 } 765 } else { 766 stack_map_stream_.AddDexRegisterEntry(i, DexRegisterLocation::Kind::kInRegister, id); 767 if (current->GetType() == Primitive::kPrimLong) { 768 stack_map_stream_.AddDexRegisterEntry(++i, DexRegisterLocation::Kind::kInRegister, id); 769 DCHECK_LT(i, environment_size); 770 } 771 } 772 break; 773 } 774 775 case Location::kFpuRegister : { 776 int id = location.reg(); 777 if (slow_path != nullptr && slow_path->IsFpuRegisterSaved(id)) { 778 uint32_t offset = slow_path->GetStackOffsetOfFpuRegister(id); 779 stack_map_stream_.AddDexRegisterEntry(i, DexRegisterLocation::Kind::kInStack, offset); 780 if (current->GetType() == Primitive::kPrimDouble) { 781 stack_map_stream_.AddDexRegisterEntry( 782 ++i, DexRegisterLocation::Kind::kInStack, offset + kVRegSize); 783 DCHECK_LT(i, environment_size); 784 } 785 } else { 786 stack_map_stream_.AddDexRegisterEntry(i, DexRegisterLocation::Kind::kInFpuRegister, id); 787 if (current->GetType() == Primitive::kPrimDouble) { 788 stack_map_stream_.AddDexRegisterEntry( 789 ++i, DexRegisterLocation::Kind::kInFpuRegister, id); 790 DCHECK_LT(i, environment_size); 791 } 792 } 793 break; 794 } 795 796 case Location::kFpuRegisterPair : { 797 int low = location.low(); 798 int high = location.high(); 799 if (slow_path != nullptr && slow_path->IsFpuRegisterSaved(low)) { 800 uint32_t offset = slow_path->GetStackOffsetOfFpuRegister(low); 801 stack_map_stream_.AddDexRegisterEntry(i, DexRegisterLocation::Kind::kInStack, offset); 802 } else { 803 stack_map_stream_.AddDexRegisterEntry(i, DexRegisterLocation::Kind::kInFpuRegister, low); 804 } 805 if (slow_path != nullptr && slow_path->IsFpuRegisterSaved(high)) { 806 uint32_t offset = slow_path->GetStackOffsetOfFpuRegister(high); 807 stack_map_stream_.AddDexRegisterEntry(++i, DexRegisterLocation::Kind::kInStack, offset); 808 } else { 809 stack_map_stream_.AddDexRegisterEntry( 810 ++i, DexRegisterLocation::Kind::kInFpuRegister, high); 811 } 812 DCHECK_LT(i, environment_size); 813 break; 814 } 815 816 case Location::kRegisterPair : { 817 int low = location.low(); 818 int high = location.high(); 819 if (slow_path != nullptr && slow_path->IsCoreRegisterSaved(low)) { 820 uint32_t offset = slow_path->GetStackOffsetOfCoreRegister(low); 821 stack_map_stream_.AddDexRegisterEntry(i, DexRegisterLocation::Kind::kInStack, offset); 822 } else { 823 stack_map_stream_.AddDexRegisterEntry(i, DexRegisterLocation::Kind::kInRegister, low); 824 } 825 if (slow_path != nullptr && slow_path->IsCoreRegisterSaved(high)) { 826 uint32_t offset = slow_path->GetStackOffsetOfCoreRegister(high); 827 stack_map_stream_.AddDexRegisterEntry(++i, DexRegisterLocation::Kind::kInStack, offset); 828 } else { 829 stack_map_stream_.AddDexRegisterEntry( 830 ++i, DexRegisterLocation::Kind::kInRegister, high); 831 } 832 DCHECK_LT(i, environment_size); 833 break; 834 } 835 836 case Location::kInvalid: { 837 stack_map_stream_.AddDexRegisterEntry(i, DexRegisterLocation::Kind::kNone, 0); 838 break; 839 } 840 841 default: 842 LOG(FATAL) << "Unexpected kind " << location.GetKind(); 843 } 844 } 845 stack_map_stream_.EndStackMapEntry(); 846 } 847 848 bool CodeGenerator::CanMoveNullCheckToUser(HNullCheck* null_check) { 849 HInstruction* first_next_not_move = null_check->GetNextDisregardingMoves(); 850 851 return (first_next_not_move != nullptr) 852 && first_next_not_move->CanDoImplicitNullCheckOn(null_check->InputAt(0)); 853 } 854 855 void CodeGenerator::MaybeRecordImplicitNullCheck(HInstruction* instr) { 856 // If we are from a static path don't record the pc as we can't throw NPE. 857 // NB: having the checks here makes the code much less verbose in the arch 858 // specific code generators. 859 if (instr->IsStaticFieldSet() || instr->IsStaticFieldGet()) { 860 return; 861 } 862 863 if (!compiler_options_.GetImplicitNullChecks()) { 864 return; 865 } 866 867 if (!instr->CanDoImplicitNullCheckOn(instr->InputAt(0))) { 868 return; 869 } 870 871 // Find the first previous instruction which is not a move. 872 HInstruction* first_prev_not_move = instr->GetPreviousDisregardingMoves(); 873 874 // If the instruction is a null check it means that `instr` is the first user 875 // and needs to record the pc. 876 if (first_prev_not_move != nullptr && first_prev_not_move->IsNullCheck()) { 877 HNullCheck* null_check = first_prev_not_move->AsNullCheck(); 878 // TODO: The parallel moves modify the environment. Their changes need to be reverted 879 // otherwise the stack maps at the throw point will not be correct. 880 RecordPcInfo(null_check, null_check->GetDexPc()); 881 } 882 } 883 884 void CodeGenerator::ClearSpillSlotsFromLoopPhisInStackMap(HSuspendCheck* suspend_check) const { 885 LocationSummary* locations = suspend_check->GetLocations(); 886 HBasicBlock* block = suspend_check->GetBlock(); 887 DCHECK(block->GetLoopInformation()->GetSuspendCheck() == suspend_check); 888 DCHECK(block->IsLoopHeader()); 889 890 for (HInstructionIterator it(block->GetPhis()); !it.Done(); it.Advance()) { 891 HInstruction* current = it.Current(); 892 LiveInterval* interval = current->GetLiveInterval(); 893 // We only need to clear bits of loop phis containing objects and allocated in register. 894 // Loop phis allocated on stack already have the object in the stack. 895 if (current->GetType() == Primitive::kPrimNot 896 && interval->HasRegister() 897 && interval->HasSpillSlot()) { 898 locations->ClearStackBit(interval->GetSpillSlot() / kVRegSize); 899 } 900 } 901 } 902 903 void CodeGenerator::EmitParallelMoves(Location from1, 904 Location to1, 905 Primitive::Type type1, 906 Location from2, 907 Location to2, 908 Primitive::Type type2) { 909 HParallelMove parallel_move(GetGraph()->GetArena()); 910 parallel_move.AddMove(from1, to1, type1, nullptr); 911 parallel_move.AddMove(from2, to2, type2, nullptr); 912 GetMoveResolver()->EmitNativeCode(¶llel_move); 913 } 914 915 void SlowPathCode::RecordPcInfo(CodeGenerator* codegen, HInstruction* instruction, uint32_t dex_pc) { 916 codegen->RecordPcInfo(instruction, dex_pc, this); 917 } 918 919 void SlowPathCode::SaveLiveRegisters(CodeGenerator* codegen, LocationSummary* locations) { 920 RegisterSet* register_set = locations->GetLiveRegisters(); 921 size_t stack_offset = codegen->GetFirstRegisterSlotInSlowPath(); 922 for (size_t i = 0, e = codegen->GetNumberOfCoreRegisters(); i < e; ++i) { 923 if (!codegen->IsCoreCalleeSaveRegister(i)) { 924 if (register_set->ContainsCoreRegister(i)) { 925 // If the register holds an object, update the stack mask. 926 if (locations->RegisterContainsObject(i)) { 927 locations->SetStackBit(stack_offset / kVRegSize); 928 } 929 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 930 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 931 saved_core_stack_offsets_[i] = stack_offset; 932 stack_offset += codegen->SaveCoreRegister(stack_offset, i); 933 } 934 } 935 } 936 937 for (size_t i = 0, e = codegen->GetNumberOfFloatingPointRegisters(); i < e; ++i) { 938 if (!codegen->IsFloatingPointCalleeSaveRegister(i)) { 939 if (register_set->ContainsFloatingPointRegister(i)) { 940 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 941 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 942 saved_fpu_stack_offsets_[i] = stack_offset; 943 stack_offset += codegen->SaveFloatingPointRegister(stack_offset, i); 944 } 945 } 946 } 947 } 948 949 void SlowPathCode::RestoreLiveRegisters(CodeGenerator* codegen, LocationSummary* locations) { 950 RegisterSet* register_set = locations->GetLiveRegisters(); 951 size_t stack_offset = codegen->GetFirstRegisterSlotInSlowPath(); 952 for (size_t i = 0, e = codegen->GetNumberOfCoreRegisters(); i < e; ++i) { 953 if (!codegen->IsCoreCalleeSaveRegister(i)) { 954 if (register_set->ContainsCoreRegister(i)) { 955 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 956 stack_offset += codegen->RestoreCoreRegister(stack_offset, i); 957 } 958 } 959 } 960 961 for (size_t i = 0, e = codegen->GetNumberOfFloatingPointRegisters(); i < e; ++i) { 962 if (!codegen->IsFloatingPointCalleeSaveRegister(i)) { 963 if (register_set->ContainsFloatingPointRegister(i)) { 964 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 965 stack_offset += codegen->RestoreFloatingPointRegister(stack_offset, i); 966 } 967 } 968 } 969 } 970 971 } // namespace art 972