code_generator_x86_64.h revision 5f7b58ea1adfc0639dd605b65f59198d3763f801
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#ifndef ART_COMPILER_OPTIMIZING_CODE_GENERATOR_X86_64_H_ 18#define ART_COMPILER_OPTIMIZING_CODE_GENERATOR_X86_64_H_ 19 20#include "arch/x86_64/instruction_set_features_x86_64.h" 21#include "code_generator.h" 22#include "dex/compiler_enums.h" 23#include "driver/compiler_options.h" 24#include "nodes.h" 25#include "parallel_move_resolver.h" 26#include "utils/x86_64/assembler_x86_64.h" 27 28namespace art { 29namespace x86_64 { 30 31// Use a local definition to prevent copying mistakes. 32static constexpr size_t kX86_64WordSize = kX86_64PointerSize; 33 34// Some x86_64 instructions require a register to be available as temp. 35static constexpr Register TMP = R11; 36 37static constexpr Register kParameterCoreRegisters[] = { RSI, RDX, RCX, R8, R9 }; 38static constexpr FloatRegister kParameterFloatRegisters[] = 39 { XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7 }; 40 41static constexpr size_t kParameterCoreRegistersLength = arraysize(kParameterCoreRegisters); 42static constexpr size_t kParameterFloatRegistersLength = arraysize(kParameterFloatRegisters); 43 44static constexpr Register kRuntimeParameterCoreRegisters[] = { RDI, RSI, RDX, RCX }; 45static constexpr size_t kRuntimeParameterCoreRegistersLength = 46 arraysize(kRuntimeParameterCoreRegisters); 47static constexpr FloatRegister kRuntimeParameterFpuRegisters[] = { XMM0, XMM1 }; 48static constexpr size_t kRuntimeParameterFpuRegistersLength = 49 arraysize(kRuntimeParameterFpuRegisters); 50 51// These XMM registers are non-volatile in ART ABI, but volatile in native ABI. 52// If the ART ABI changes, this list must be updated. It is used to ensure that 53// these are not clobbered by any direct call to native code (such as math intrinsics). 54static constexpr FloatRegister non_volatile_xmm_regs[] = { XMM12, XMM13, XMM14, XMM15 }; 55 56 57class InvokeRuntimeCallingConvention : public CallingConvention<Register, FloatRegister> { 58 public: 59 InvokeRuntimeCallingConvention() 60 : CallingConvention(kRuntimeParameterCoreRegisters, 61 kRuntimeParameterCoreRegistersLength, 62 kRuntimeParameterFpuRegisters, 63 kRuntimeParameterFpuRegistersLength, 64 kX86_64PointerSize) {} 65 66 private: 67 DISALLOW_COPY_AND_ASSIGN(InvokeRuntimeCallingConvention); 68}; 69 70class InvokeDexCallingConvention : public CallingConvention<Register, FloatRegister> { 71 public: 72 InvokeDexCallingConvention() : CallingConvention( 73 kParameterCoreRegisters, 74 kParameterCoreRegistersLength, 75 kParameterFloatRegisters, 76 kParameterFloatRegistersLength, 77 kX86_64PointerSize) {} 78 79 private: 80 DISALLOW_COPY_AND_ASSIGN(InvokeDexCallingConvention); 81}; 82 83class FieldAccessCallingConventionX86_64 : public FieldAccessCallingConvention { 84 public: 85 FieldAccessCallingConventionX86_64() {} 86 87 Location GetObjectLocation() const OVERRIDE { 88 return Location::RegisterLocation(RSI); 89 } 90 Location GetFieldIndexLocation() const OVERRIDE { 91 return Location::RegisterLocation(RDI); 92 } 93 Location GetReturnLocation(Primitive::Type type ATTRIBUTE_UNUSED) const OVERRIDE { 94 return Location::RegisterLocation(RAX); 95 } 96 Location GetSetValueLocation(Primitive::Type type, bool is_instance) const OVERRIDE { 97 return Primitive::Is64BitType(type) 98 ? Location::RegisterLocation(RDX) 99 : (is_instance 100 ? Location::RegisterLocation(RDX) 101 : Location::RegisterLocation(RSI)); 102 } 103 Location GetFpuLocation(Primitive::Type type ATTRIBUTE_UNUSED) const OVERRIDE { 104 return Location::FpuRegisterLocation(XMM0); 105 } 106 107 private: 108 DISALLOW_COPY_AND_ASSIGN(FieldAccessCallingConventionX86_64); 109}; 110 111 112class InvokeDexCallingConventionVisitorX86_64 : public InvokeDexCallingConventionVisitor { 113 public: 114 InvokeDexCallingConventionVisitorX86_64() {} 115 virtual ~InvokeDexCallingConventionVisitorX86_64() {} 116 117 Location GetNextLocation(Primitive::Type type) OVERRIDE; 118 Location GetReturnLocation(Primitive::Type type) const OVERRIDE; 119 Location GetMethodLocation() const OVERRIDE; 120 121 private: 122 InvokeDexCallingConvention calling_convention; 123 124 DISALLOW_COPY_AND_ASSIGN(InvokeDexCallingConventionVisitorX86_64); 125}; 126 127class CodeGeneratorX86_64; 128 129class ParallelMoveResolverX86_64 : public ParallelMoveResolverWithSwap { 130 public: 131 ParallelMoveResolverX86_64(ArenaAllocator* allocator, CodeGeneratorX86_64* codegen) 132 : ParallelMoveResolverWithSwap(allocator), codegen_(codegen) {} 133 134 void EmitMove(size_t index) OVERRIDE; 135 void EmitSwap(size_t index) OVERRIDE; 136 void SpillScratch(int reg) OVERRIDE; 137 void RestoreScratch(int reg) OVERRIDE; 138 139 X86_64Assembler* GetAssembler() const; 140 141 private: 142 void Exchange32(CpuRegister reg, int mem); 143 void Exchange32(XmmRegister reg, int mem); 144 void Exchange32(int mem1, int mem2); 145 void Exchange64(CpuRegister reg, int mem); 146 void Exchange64(XmmRegister reg, int mem); 147 void Exchange64(int mem1, int mem2); 148 149 CodeGeneratorX86_64* const codegen_; 150 151 DISALLOW_COPY_AND_ASSIGN(ParallelMoveResolverX86_64); 152}; 153 154class LocationsBuilderX86_64 : public HGraphVisitor { 155 public: 156 LocationsBuilderX86_64(HGraph* graph, CodeGeneratorX86_64* codegen) 157 : HGraphVisitor(graph), codegen_(codegen) {} 158 159#define DECLARE_VISIT_INSTRUCTION(name, super) \ 160 void Visit##name(H##name* instr) OVERRIDE; 161 162 FOR_EACH_CONCRETE_INSTRUCTION_COMMON(DECLARE_VISIT_INSTRUCTION) 163 FOR_EACH_CONCRETE_INSTRUCTION_X86_64(DECLARE_VISIT_INSTRUCTION) 164 165#undef DECLARE_VISIT_INSTRUCTION 166 167 void VisitInstruction(HInstruction* instruction) OVERRIDE { 168 LOG(FATAL) << "Unreachable instruction " << instruction->DebugName() 169 << " (id " << instruction->GetId() << ")"; 170 } 171 172 private: 173 void HandleInvoke(HInvoke* invoke); 174 void HandleBitwiseOperation(HBinaryOperation* operation); 175 void HandleCondition(HCondition* condition); 176 void HandleShift(HBinaryOperation* operation); 177 void HandleFieldSet(HInstruction* instruction, const FieldInfo& field_info); 178 void HandleFieldGet(HInstruction* instruction); 179 180 CodeGeneratorX86_64* const codegen_; 181 InvokeDexCallingConventionVisitorX86_64 parameter_visitor_; 182 183 DISALLOW_COPY_AND_ASSIGN(LocationsBuilderX86_64); 184}; 185 186class InstructionCodeGeneratorX86_64 : public HGraphVisitor { 187 public: 188 InstructionCodeGeneratorX86_64(HGraph* graph, CodeGeneratorX86_64* codegen); 189 190#define DECLARE_VISIT_INSTRUCTION(name, super) \ 191 void Visit##name(H##name* instr) OVERRIDE; 192 193 FOR_EACH_CONCRETE_INSTRUCTION_COMMON(DECLARE_VISIT_INSTRUCTION) 194 FOR_EACH_CONCRETE_INSTRUCTION_X86_64(DECLARE_VISIT_INSTRUCTION) 195 196#undef DECLARE_VISIT_INSTRUCTION 197 198 void VisitInstruction(HInstruction* instruction) OVERRIDE { 199 LOG(FATAL) << "Unreachable instruction " << instruction->DebugName() 200 << " (id " << instruction->GetId() << ")"; 201 } 202 203 X86_64Assembler* GetAssembler() const { return assembler_; } 204 205 private: 206 // Generate code for the given suspend check. If not null, `successor` 207 // is the block to branch to if the suspend check is not needed, and after 208 // the suspend call. 209 void GenerateSuspendCheck(HSuspendCheck* instruction, HBasicBlock* successor); 210 void GenerateClassInitializationCheck(SlowPathCode* slow_path, CpuRegister class_reg); 211 void HandleBitwiseOperation(HBinaryOperation* operation); 212 void GenerateRemFP(HRem* rem); 213 void DivRemOneOrMinusOne(HBinaryOperation* instruction); 214 void DivByPowerOfTwo(HDiv* instruction); 215 void GenerateDivRemWithAnyConstant(HBinaryOperation* instruction); 216 void GenerateDivRemIntegral(HBinaryOperation* instruction); 217 void HandleCondition(HCondition* condition); 218 void HandleShift(HBinaryOperation* operation); 219 220 void HandleFieldSet(HInstruction* instruction, 221 const FieldInfo& field_info, 222 bool value_can_be_null); 223 void HandleFieldGet(HInstruction* instruction, const FieldInfo& field_info); 224 225 // Generate a heap reference load using one register `out`: 226 // 227 // out <- *(out + offset) 228 // 229 // while honoring heap poisoning and/or read barriers (if any). 230 // Register `temp` is used when generating a read barrier. 231 void GenerateReferenceLoadOneRegister(HInstruction* instruction, 232 Location out, 233 uint32_t offset, 234 Location temp); 235 // Generate a heap reference load using two different registers 236 // `out` and `obj`: 237 // 238 // out <- *(obj + offset) 239 // 240 // while honoring heap poisoning and/or read barriers (if any). 241 // Register `temp` is used when generating a Baker's read barrier. 242 void GenerateReferenceLoadTwoRegisters(HInstruction* instruction, 243 Location out, 244 Location obj, 245 uint32_t offset, 246 Location temp); 247 // Generate a GC root reference load: 248 // 249 // root <- *(obj + offset) 250 // 251 // while honoring read barriers (if any). 252 void GenerateGcRootFieldLoad(HInstruction* instruction, 253 Location root, 254 CpuRegister obj, 255 uint32_t offset); 256 257 void GenerateImplicitNullCheck(HNullCheck* instruction); 258 void GenerateExplicitNullCheck(HNullCheck* instruction); 259 void PushOntoFPStack(Location source, uint32_t temp_offset, 260 uint32_t stack_adjustment, bool is_float); 261 void GenerateTestAndBranch(HInstruction* instruction, 262 size_t condition_input_index, 263 Label* true_target, 264 Label* false_target); 265 void GenerateCompareTestAndBranch(HCondition* condition, 266 Label* true_target, 267 Label* false_target); 268 void GenerateFPJumps(HCondition* cond, Label* true_label, Label* false_label); 269 void HandleGoto(HInstruction* got, HBasicBlock* successor); 270 271 X86_64Assembler* const assembler_; 272 CodeGeneratorX86_64* const codegen_; 273 274 DISALLOW_COPY_AND_ASSIGN(InstructionCodeGeneratorX86_64); 275}; 276 277// Class for fixups to jump tables. 278class JumpTableRIPFixup; 279 280class CodeGeneratorX86_64 : public CodeGenerator { 281 public: 282 CodeGeneratorX86_64(HGraph* graph, 283 const X86_64InstructionSetFeatures& isa_features, 284 const CompilerOptions& compiler_options, 285 OptimizingCompilerStats* stats = nullptr); 286 virtual ~CodeGeneratorX86_64() {} 287 288 void GenerateFrameEntry() OVERRIDE; 289 void GenerateFrameExit() OVERRIDE; 290 void Bind(HBasicBlock* block) OVERRIDE; 291 void Move(HInstruction* instruction, Location location, HInstruction* move_for) OVERRIDE; 292 void MoveConstant(Location destination, int32_t value) OVERRIDE; 293 void MoveLocation(Location dst, Location src, Primitive::Type dst_type) OVERRIDE; 294 void AddLocationAsTemp(Location location, LocationSummary* locations) OVERRIDE; 295 296 size_t SaveCoreRegister(size_t stack_index, uint32_t reg_id) OVERRIDE; 297 size_t RestoreCoreRegister(size_t stack_index, uint32_t reg_id) OVERRIDE; 298 size_t SaveFloatingPointRegister(size_t stack_index, uint32_t reg_id) OVERRIDE; 299 size_t RestoreFloatingPointRegister(size_t stack_index, uint32_t reg_id) OVERRIDE; 300 301 // Generate code to invoke a runtime entry point. 302 void InvokeRuntime(QuickEntrypointEnum entrypoint, 303 HInstruction* instruction, 304 uint32_t dex_pc, 305 SlowPathCode* slow_path) OVERRIDE; 306 307 void InvokeRuntime(int32_t entry_point_offset, 308 HInstruction* instruction, 309 uint32_t dex_pc, 310 SlowPathCode* slow_path); 311 312 size_t GetWordSize() const OVERRIDE { 313 return kX86_64WordSize; 314 } 315 316 size_t GetFloatingPointSpillSlotSize() const OVERRIDE { 317 return kX86_64WordSize; 318 } 319 320 HGraphVisitor* GetLocationBuilder() OVERRIDE { 321 return &location_builder_; 322 } 323 324 HGraphVisitor* GetInstructionVisitor() OVERRIDE { 325 return &instruction_visitor_; 326 } 327 328 X86_64Assembler* GetAssembler() OVERRIDE { 329 return &assembler_; 330 } 331 332 const X86_64Assembler& GetAssembler() const OVERRIDE { 333 return assembler_; 334 } 335 336 ParallelMoveResolverX86_64* GetMoveResolver() OVERRIDE { 337 return &move_resolver_; 338 } 339 340 uintptr_t GetAddressOf(HBasicBlock* block) const OVERRIDE { 341 return GetLabelOf(block)->Position(); 342 } 343 344 Location GetStackLocation(HLoadLocal* load) const OVERRIDE; 345 346 void SetupBlockedRegisters(bool is_baseline) const OVERRIDE; 347 Location AllocateFreeRegister(Primitive::Type type) const OVERRIDE; 348 void DumpCoreRegister(std::ostream& stream, int reg) const OVERRIDE; 349 void DumpFloatingPointRegister(std::ostream& stream, int reg) const OVERRIDE; 350 void Finalize(CodeAllocator* allocator) OVERRIDE; 351 352 InstructionSet GetInstructionSet() const OVERRIDE { 353 return InstructionSet::kX86_64; 354 } 355 356 // Emit a write barrier. 357 void MarkGCCard(CpuRegister temp, 358 CpuRegister card, 359 CpuRegister object, 360 CpuRegister value, 361 bool value_can_be_null); 362 363 void GenerateMemoryBarrier(MemBarrierKind kind); 364 365 // Helper method to move a value between two locations. 366 void Move(Location destination, Location source); 367 368 Label* GetLabelOf(HBasicBlock* block) const { 369 return CommonGetLabelOf<Label>(block_labels_, block); 370 } 371 372 void Initialize() OVERRIDE { 373 block_labels_ = CommonInitializeLabels<Label>(); 374 } 375 376 bool NeedsTwoRegisters(Primitive::Type type ATTRIBUTE_UNUSED) const OVERRIDE { 377 return false; 378 } 379 380 // Check if the desired_dispatch_info is supported. If it is, return it, 381 // otherwise return a fall-back info that should be used instead. 382 HInvokeStaticOrDirect::DispatchInfo GetSupportedInvokeStaticOrDirectDispatch( 383 const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info, 384 MethodReference target_method) OVERRIDE; 385 386 void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) OVERRIDE; 387 void GenerateVirtualCall(HInvokeVirtual* invoke, Location temp) OVERRIDE; 388 389 void MoveFromReturnRegister(Location trg, Primitive::Type type) OVERRIDE; 390 391 void EmitLinkerPatches(ArenaVector<LinkerPatch>* linker_patches) OVERRIDE; 392 393 const X86_64InstructionSetFeatures& GetInstructionSetFeatures() const { 394 return isa_features_; 395 } 396 397 // Fast path implementation of ReadBarrier::Barrier for a heap 398 // reference field load when Baker's read barriers are used. 399 void GenerateFieldLoadWithBakerReadBarrier(HInstruction* instruction, 400 Location out, 401 CpuRegister obj, 402 uint32_t offset, 403 Location temp, 404 bool needs_null_check); 405 // Fast path implementation of ReadBarrier::Barrier for a heap 406 // reference array load when Baker's read barriers are used. 407 void GenerateArrayLoadWithBakerReadBarrier(HInstruction* instruction, 408 Location out, 409 CpuRegister obj, 410 uint32_t data_offset, 411 Location index, 412 Location temp, 413 bool needs_null_check); 414 415 // Generate a read barrier for a heap reference within `instruction` 416 // using a slow path. 417 // 418 // A read barrier for an object reference read from the heap is 419 // implemented as a call to the artReadBarrierSlow runtime entry 420 // point, which is passed the values in locations `ref`, `obj`, and 421 // `offset`: 422 // 423 // mirror::Object* artReadBarrierSlow(mirror::Object* ref, 424 // mirror::Object* obj, 425 // uint32_t offset); 426 // 427 // The `out` location contains the value returned by 428 // artReadBarrierSlow. 429 // 430 // When `index` provided (i.e., when it is different from 431 // Location::NoLocation()), the offset value passed to 432 // artReadBarrierSlow is adjusted to take `index` into account. 433 void GenerateReadBarrierSlow(HInstruction* instruction, 434 Location out, 435 Location ref, 436 Location obj, 437 uint32_t offset, 438 Location index = Location::NoLocation()); 439 440 // If read barriers are enabled, generate a read barrier for a heap 441 // reference using a slow path. If heap poisoning is enabled, also 442 // unpoison the reference in `out`. 443 void MaybeGenerateReadBarrierSlow(HInstruction* instruction, 444 Location out, 445 Location ref, 446 Location obj, 447 uint32_t offset, 448 Location index = Location::NoLocation()); 449 450 // Generate a read barrier for a GC root within `instruction` using 451 // a slow path. 452 // 453 // A read barrier for an object reference GC root is implemented as 454 // a call to the artReadBarrierForRootSlow runtime entry point, 455 // which is passed the value in location `root`: 456 // 457 // mirror::Object* artReadBarrierForRootSlow(GcRoot<mirror::Object>* root); 458 // 459 // The `out` location contains the value returned by 460 // artReadBarrierForRootSlow. 461 void GenerateReadBarrierForRootSlow(HInstruction* instruction, Location out, Location root); 462 463 int ConstantAreaStart() const { 464 return constant_area_start_; 465 } 466 467 Address LiteralDoubleAddress(double v); 468 Address LiteralFloatAddress(float v); 469 Address LiteralInt32Address(int32_t v); 470 Address LiteralInt64Address(int64_t v); 471 472 // Load a 64 bit value into a register in the most efficient manner. 473 void Load64BitValue(CpuRegister dest, int64_t value); 474 Address LiteralCaseTable(HPackedSwitch* switch_instr); 475 476 // Store a 64 bit value into a DoubleStackSlot in the most efficient manner. 477 void Store64BitValueToStack(Location dest, int64_t value); 478 479 // Assign a 64 bit constant to an address. 480 void MoveInt64ToAddress(const Address& addr_low, 481 const Address& addr_high, 482 int64_t v, 483 HInstruction* instruction); 484 485 // Ensure that prior stores complete to memory before subsequent loads. 486 // The locked add implementation will avoid serializing device memory, but will 487 // touch (but not change) the top of the stack. The locked add should not be used for 488 // ordering non-temporal stores. 489 void MemoryFence(bool force_mfence = false) { 490 if (!force_mfence && isa_features_.PrefersLockedAddSynchronization()) { 491 assembler_.lock()->addl(Address(CpuRegister(RSP), 0), Immediate(0)); 492 } else { 493 assembler_.mfence(); 494 } 495 } 496 497 private: 498 // Factored implementation of GenerateFieldLoadWithBakerReadBarrier 499 // and GenerateArrayLoadWithBakerReadBarrier. 500 void GenerateReferenceLoadWithBakerReadBarrier(HInstruction* instruction, 501 Location ref, 502 CpuRegister obj, 503 const Address& src, 504 Location temp, 505 bool needs_null_check); 506 507 struct PcRelativeDexCacheAccessInfo { 508 PcRelativeDexCacheAccessInfo(const DexFile& dex_file, uint32_t element_off) 509 : target_dex_file(dex_file), element_offset(element_off), label() { } 510 511 const DexFile& target_dex_file; 512 uint32_t element_offset; 513 Label label; 514 }; 515 516 // Labels for each block that will be compiled. 517 Label* block_labels_; // Indexed by block id. 518 Label frame_entry_label_; 519 LocationsBuilderX86_64 location_builder_; 520 InstructionCodeGeneratorX86_64 instruction_visitor_; 521 ParallelMoveResolverX86_64 move_resolver_; 522 X86_64Assembler assembler_; 523 const X86_64InstructionSetFeatures& isa_features_; 524 525 // Offset to the start of the constant area in the assembled code. 526 // Used for fixups to the constant area. 527 int constant_area_start_; 528 529 // Method patch info. Using ArenaDeque<> which retains element addresses on push/emplace_back(). 530 ArenaDeque<MethodPatchInfo<Label>> method_patches_; 531 ArenaDeque<MethodPatchInfo<Label>> relative_call_patches_; 532 // PC-relative DexCache access info. 533 ArenaDeque<PcRelativeDexCacheAccessInfo> pc_relative_dex_cache_patches_; 534 535 // When we don't know the proper offset for the value, we use kDummy32BitOffset. 536 // We will fix this up in the linker later to have the right value. 537 static constexpr int32_t kDummy32BitOffset = 256; 538 539 // Fixups for jump tables need to be handled specially. 540 ArenaVector<JumpTableRIPFixup*> fixups_to_jump_tables_; 541 542 DISALLOW_COPY_AND_ASSIGN(CodeGeneratorX86_64); 543}; 544 545} // namespace x86_64 546} // namespace art 547 548#endif // ART_COMPILER_OPTIMIZING_CODE_GENERATOR_X86_64_H_ 549