code_generator_x86.cc revision c88ef3a10c474045a3476a02ae75d07ddd3230b7
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_x86.h" 18 19#include "art_method.h" 20#include "code_generator_utils.h" 21#include "compiled_method.h" 22#include "entrypoints/quick/quick_entrypoints.h" 23#include "entrypoints/quick/quick_entrypoints_enum.h" 24#include "gc/accounting/card_table.h" 25#include "intrinsics.h" 26#include "intrinsics_x86.h" 27#include "mirror/array-inl.h" 28#include "mirror/class-inl.h" 29#include "pc_relative_fixups_x86.h" 30#include "thread.h" 31#include "utils/assembler.h" 32#include "utils/stack_checks.h" 33#include "utils/x86/assembler_x86.h" 34#include "utils/x86/managed_register_x86.h" 35 36namespace art { 37 38template<class MirrorType> 39class GcRoot; 40 41namespace x86 { 42 43static constexpr int kCurrentMethodStackOffset = 0; 44static constexpr Register kMethodRegisterArgument = EAX; 45 46static constexpr Register kCoreCalleeSaves[] = { EBP, ESI, EDI }; 47 48static constexpr int kC2ConditionMask = 0x400; 49 50static constexpr int kFakeReturnRegister = Register(8); 51 52#define __ down_cast<X86Assembler*>(codegen->GetAssembler())-> 53#define QUICK_ENTRY_POINT(x) QUICK_ENTRYPOINT_OFFSET(kX86WordSize, x).Int32Value() 54 55class NullCheckSlowPathX86 : public SlowPathCode { 56 public: 57 explicit NullCheckSlowPathX86(HNullCheck* instruction) : instruction_(instruction) {} 58 59 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 60 CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen); 61 __ Bind(GetEntryLabel()); 62 if (instruction_->CanThrowIntoCatchBlock()) { 63 // Live registers will be restored in the catch block if caught. 64 SaveLiveRegisters(codegen, instruction_->GetLocations()); 65 } 66 x86_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pThrowNullPointer), 67 instruction_, 68 instruction_->GetDexPc(), 69 this); 70 CheckEntrypointTypes<kQuickThrowNullPointer, void, void>(); 71 } 72 73 bool IsFatal() const OVERRIDE { return true; } 74 75 const char* GetDescription() const OVERRIDE { return "NullCheckSlowPathX86"; } 76 77 private: 78 HNullCheck* const instruction_; 79 DISALLOW_COPY_AND_ASSIGN(NullCheckSlowPathX86); 80}; 81 82class DivZeroCheckSlowPathX86 : public SlowPathCode { 83 public: 84 explicit DivZeroCheckSlowPathX86(HDivZeroCheck* instruction) : instruction_(instruction) {} 85 86 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 87 CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen); 88 __ Bind(GetEntryLabel()); 89 if (instruction_->CanThrowIntoCatchBlock()) { 90 // Live registers will be restored in the catch block if caught. 91 SaveLiveRegisters(codegen, instruction_->GetLocations()); 92 } 93 x86_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pThrowDivZero), 94 instruction_, 95 instruction_->GetDexPc(), 96 this); 97 CheckEntrypointTypes<kQuickThrowDivZero, void, void>(); 98 } 99 100 bool IsFatal() const OVERRIDE { return true; } 101 102 const char* GetDescription() const OVERRIDE { return "DivZeroCheckSlowPathX86"; } 103 104 private: 105 HDivZeroCheck* const instruction_; 106 DISALLOW_COPY_AND_ASSIGN(DivZeroCheckSlowPathX86); 107}; 108 109class DivRemMinusOneSlowPathX86 : public SlowPathCode { 110 public: 111 DivRemMinusOneSlowPathX86(Register reg, bool is_div) : reg_(reg), is_div_(is_div) {} 112 113 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 114 __ Bind(GetEntryLabel()); 115 if (is_div_) { 116 __ negl(reg_); 117 } else { 118 __ movl(reg_, Immediate(0)); 119 } 120 __ jmp(GetExitLabel()); 121 } 122 123 const char* GetDescription() const OVERRIDE { return "DivRemMinusOneSlowPathX86"; } 124 125 private: 126 Register reg_; 127 bool is_div_; 128 DISALLOW_COPY_AND_ASSIGN(DivRemMinusOneSlowPathX86); 129}; 130 131class BoundsCheckSlowPathX86 : public SlowPathCode { 132 public: 133 explicit BoundsCheckSlowPathX86(HBoundsCheck* instruction) : instruction_(instruction) {} 134 135 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 136 LocationSummary* locations = instruction_->GetLocations(); 137 CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen); 138 __ Bind(GetEntryLabel()); 139 // We're moving two locations to locations that could overlap, so we need a parallel 140 // move resolver. 141 if (instruction_->CanThrowIntoCatchBlock()) { 142 // Live registers will be restored in the catch block if caught. 143 SaveLiveRegisters(codegen, instruction_->GetLocations()); 144 } 145 InvokeRuntimeCallingConvention calling_convention; 146 x86_codegen->EmitParallelMoves( 147 locations->InAt(0), 148 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 149 Primitive::kPrimInt, 150 locations->InAt(1), 151 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 152 Primitive::kPrimInt); 153 x86_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pThrowArrayBounds), 154 instruction_, 155 instruction_->GetDexPc(), 156 this); 157 CheckEntrypointTypes<kQuickThrowArrayBounds, void, int32_t, int32_t>(); 158 } 159 160 bool IsFatal() const OVERRIDE { return true; } 161 162 const char* GetDescription() const OVERRIDE { return "BoundsCheckSlowPathX86"; } 163 164 private: 165 HBoundsCheck* const instruction_; 166 167 DISALLOW_COPY_AND_ASSIGN(BoundsCheckSlowPathX86); 168}; 169 170class SuspendCheckSlowPathX86 : public SlowPathCode { 171 public: 172 SuspendCheckSlowPathX86(HSuspendCheck* instruction, HBasicBlock* successor) 173 : instruction_(instruction), successor_(successor) {} 174 175 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 176 CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen); 177 __ Bind(GetEntryLabel()); 178 SaveLiveRegisters(codegen, instruction_->GetLocations()); 179 x86_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pTestSuspend), 180 instruction_, 181 instruction_->GetDexPc(), 182 this); 183 CheckEntrypointTypes<kQuickTestSuspend, void, void>(); 184 RestoreLiveRegisters(codegen, instruction_->GetLocations()); 185 if (successor_ == nullptr) { 186 __ jmp(GetReturnLabel()); 187 } else { 188 __ jmp(x86_codegen->GetLabelOf(successor_)); 189 } 190 } 191 192 Label* GetReturnLabel() { 193 DCHECK(successor_ == nullptr); 194 return &return_label_; 195 } 196 197 HBasicBlock* GetSuccessor() const { 198 return successor_; 199 } 200 201 const char* GetDescription() const OVERRIDE { return "SuspendCheckSlowPathX86"; } 202 203 private: 204 HSuspendCheck* const instruction_; 205 HBasicBlock* const successor_; 206 Label return_label_; 207 208 DISALLOW_COPY_AND_ASSIGN(SuspendCheckSlowPathX86); 209}; 210 211class LoadStringSlowPathX86 : public SlowPathCode { 212 public: 213 explicit LoadStringSlowPathX86(HLoadString* instruction) : instruction_(instruction) {} 214 215 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 216 LocationSummary* locations = instruction_->GetLocations(); 217 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg())); 218 219 CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen); 220 __ Bind(GetEntryLabel()); 221 SaveLiveRegisters(codegen, locations); 222 223 InvokeRuntimeCallingConvention calling_convention; 224 __ movl(calling_convention.GetRegisterAt(0), Immediate(instruction_->GetStringIndex())); 225 x86_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pResolveString), 226 instruction_, 227 instruction_->GetDexPc(), 228 this); 229 CheckEntrypointTypes<kQuickResolveString, void*, uint32_t>(); 230 x86_codegen->Move32(locations->Out(), Location::RegisterLocation(EAX)); 231 RestoreLiveRegisters(codegen, locations); 232 233 __ jmp(GetExitLabel()); 234 } 235 236 const char* GetDescription() const OVERRIDE { return "LoadStringSlowPathX86"; } 237 238 private: 239 HLoadString* const instruction_; 240 241 DISALLOW_COPY_AND_ASSIGN(LoadStringSlowPathX86); 242}; 243 244class LoadClassSlowPathX86 : public SlowPathCode { 245 public: 246 LoadClassSlowPathX86(HLoadClass* cls, 247 HInstruction* at, 248 uint32_t dex_pc, 249 bool do_clinit) 250 : cls_(cls), at_(at), dex_pc_(dex_pc), do_clinit_(do_clinit) { 251 DCHECK(at->IsLoadClass() || at->IsClinitCheck()); 252 } 253 254 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 255 LocationSummary* locations = at_->GetLocations(); 256 CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen); 257 __ Bind(GetEntryLabel()); 258 SaveLiveRegisters(codegen, locations); 259 260 InvokeRuntimeCallingConvention calling_convention; 261 __ movl(calling_convention.GetRegisterAt(0), Immediate(cls_->GetTypeIndex())); 262 x86_codegen->InvokeRuntime(do_clinit_ ? QUICK_ENTRY_POINT(pInitializeStaticStorage) 263 : QUICK_ENTRY_POINT(pInitializeType), 264 at_, dex_pc_, this); 265 if (do_clinit_) { 266 CheckEntrypointTypes<kQuickInitializeStaticStorage, void*, uint32_t>(); 267 } else { 268 CheckEntrypointTypes<kQuickInitializeType, void*, uint32_t>(); 269 } 270 271 // Move the class to the desired location. 272 Location out = locations->Out(); 273 if (out.IsValid()) { 274 DCHECK(out.IsRegister() && !locations->GetLiveRegisters()->ContainsCoreRegister(out.reg())); 275 x86_codegen->Move32(out, Location::RegisterLocation(EAX)); 276 } 277 278 RestoreLiveRegisters(codegen, locations); 279 __ jmp(GetExitLabel()); 280 } 281 282 const char* GetDescription() const OVERRIDE { return "LoadClassSlowPathX86"; } 283 284 private: 285 // The class this slow path will load. 286 HLoadClass* const cls_; 287 288 // The instruction where this slow path is happening. 289 // (Might be the load class or an initialization check). 290 HInstruction* const at_; 291 292 // The dex PC of `at_`. 293 const uint32_t dex_pc_; 294 295 // Whether to initialize the class. 296 const bool do_clinit_; 297 298 DISALLOW_COPY_AND_ASSIGN(LoadClassSlowPathX86); 299}; 300 301class TypeCheckSlowPathX86 : public SlowPathCode { 302 public: 303 TypeCheckSlowPathX86(HInstruction* instruction, bool is_fatal) 304 : instruction_(instruction), is_fatal_(is_fatal) {} 305 306 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 307 LocationSummary* locations = instruction_->GetLocations(); 308 Location object_class = instruction_->IsCheckCast() ? locations->GetTemp(0) 309 : locations->Out(); 310 DCHECK(instruction_->IsCheckCast() 311 || !locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg())); 312 313 CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen); 314 __ Bind(GetEntryLabel()); 315 316 if (!is_fatal_) { 317 SaveLiveRegisters(codegen, locations); 318 } 319 320 // We're moving two locations to locations that could overlap, so we need a parallel 321 // move resolver. 322 InvokeRuntimeCallingConvention calling_convention; 323 x86_codegen->EmitParallelMoves( 324 locations->InAt(1), 325 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 326 Primitive::kPrimNot, 327 object_class, 328 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 329 Primitive::kPrimNot); 330 331 if (instruction_->IsInstanceOf()) { 332 x86_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pInstanceofNonTrivial), 333 instruction_, 334 instruction_->GetDexPc(), 335 this); 336 CheckEntrypointTypes< 337 kQuickInstanceofNonTrivial, uint32_t, const mirror::Class*, const mirror::Class*>(); 338 } else { 339 DCHECK(instruction_->IsCheckCast()); 340 x86_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pCheckCast), 341 instruction_, 342 instruction_->GetDexPc(), 343 this); 344 CheckEntrypointTypes<kQuickCheckCast, void, const mirror::Class*, const mirror::Class*>(); 345 } 346 347 if (!is_fatal_) { 348 if (instruction_->IsInstanceOf()) { 349 x86_codegen->Move32(locations->Out(), Location::RegisterLocation(EAX)); 350 } 351 RestoreLiveRegisters(codegen, locations); 352 353 __ jmp(GetExitLabel()); 354 } 355 } 356 357 const char* GetDescription() const OVERRIDE { return "TypeCheckSlowPathX86"; } 358 bool IsFatal() const OVERRIDE { return is_fatal_; } 359 360 private: 361 HInstruction* const instruction_; 362 const bool is_fatal_; 363 364 DISALLOW_COPY_AND_ASSIGN(TypeCheckSlowPathX86); 365}; 366 367class DeoptimizationSlowPathX86 : public SlowPathCode { 368 public: 369 explicit DeoptimizationSlowPathX86(HInstruction* instruction) 370 : instruction_(instruction) {} 371 372 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 373 DCHECK(instruction_->IsDeoptimize()); 374 CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen); 375 __ Bind(GetEntryLabel()); 376 SaveLiveRegisters(codegen, instruction_->GetLocations()); 377 x86_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pDeoptimize), 378 instruction_, 379 instruction_->GetDexPc(), 380 this); 381 CheckEntrypointTypes<kQuickDeoptimize, void, void>(); 382 } 383 384 const char* GetDescription() const OVERRIDE { return "DeoptimizationSlowPathX86"; } 385 386 private: 387 HInstruction* const instruction_; 388 DISALLOW_COPY_AND_ASSIGN(DeoptimizationSlowPathX86); 389}; 390 391class ArraySetSlowPathX86 : public SlowPathCode { 392 public: 393 explicit ArraySetSlowPathX86(HInstruction* instruction) : instruction_(instruction) {} 394 395 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 396 LocationSummary* locations = instruction_->GetLocations(); 397 __ Bind(GetEntryLabel()); 398 SaveLiveRegisters(codegen, locations); 399 400 InvokeRuntimeCallingConvention calling_convention; 401 HParallelMove parallel_move(codegen->GetGraph()->GetArena()); 402 parallel_move.AddMove( 403 locations->InAt(0), 404 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 405 Primitive::kPrimNot, 406 nullptr); 407 parallel_move.AddMove( 408 locations->InAt(1), 409 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 410 Primitive::kPrimInt, 411 nullptr); 412 parallel_move.AddMove( 413 locations->InAt(2), 414 Location::RegisterLocation(calling_convention.GetRegisterAt(2)), 415 Primitive::kPrimNot, 416 nullptr); 417 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 418 419 CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen); 420 x86_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pAputObject), 421 instruction_, 422 instruction_->GetDexPc(), 423 this); 424 CheckEntrypointTypes<kQuickAputObject, void, mirror::Array*, int32_t, mirror::Object*>(); 425 RestoreLiveRegisters(codegen, locations); 426 __ jmp(GetExitLabel()); 427 } 428 429 const char* GetDescription() const OVERRIDE { return "ArraySetSlowPathX86"; } 430 431 private: 432 HInstruction* const instruction_; 433 434 DISALLOW_COPY_AND_ASSIGN(ArraySetSlowPathX86); 435}; 436 437// Slow path generating a read barrier for a heap reference. 438class ReadBarrierForHeapReferenceSlowPathX86 : public SlowPathCode { 439 public: 440 ReadBarrierForHeapReferenceSlowPathX86(HInstruction* instruction, 441 Location out, 442 Location ref, 443 Location obj, 444 uint32_t offset, 445 Location index) 446 : instruction_(instruction), 447 out_(out), 448 ref_(ref), 449 obj_(obj), 450 offset_(offset), 451 index_(index) { 452 DCHECK(kEmitCompilerReadBarrier); 453 // If `obj` is equal to `out` or `ref`, it means the initial object 454 // has been overwritten by (or after) the heap object reference load 455 // to be instrumented, e.g.: 456 // 457 // __ movl(out, Address(out, offset)); 458 // codegen_->GenerateReadBarrier(instruction, out_loc, out_loc, out_loc, offset); 459 // 460 // In that case, we have lost the information about the original 461 // object, and the emitted read barrier cannot work properly. 462 DCHECK(!obj.Equals(out)) << "obj=" << obj << " out=" << out; 463 DCHECK(!obj.Equals(ref)) << "obj=" << obj << " ref=" << ref; 464 } 465 466 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 467 CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen); 468 LocationSummary* locations = instruction_->GetLocations(); 469 Register reg_out = out_.AsRegister<Register>(); 470 DCHECK(locations->CanCall()); 471 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg_out)); 472 DCHECK(!instruction_->IsInvoke() || 473 (instruction_->IsInvokeStaticOrDirect() && 474 instruction_->GetLocations()->Intrinsified())); 475 476 __ Bind(GetEntryLabel()); 477 SaveLiveRegisters(codegen, locations); 478 479 // We may have to change the index's value, but as `index_` is a 480 // constant member (like other "inputs" of this slow path), 481 // introduce a copy of it, `index`. 482 Location index = index_; 483 if (index_.IsValid()) { 484 // Handle `index_` for HArrayGet and intrinsic UnsafeGetObject. 485 if (instruction_->IsArrayGet()) { 486 // Compute the actual memory offset and store it in `index`. 487 Register index_reg = index_.AsRegister<Register>(); 488 DCHECK(locations->GetLiveRegisters()->ContainsCoreRegister(index_reg)); 489 if (codegen->IsCoreCalleeSaveRegister(index_reg)) { 490 // We are about to change the value of `index_reg` (see the 491 // calls to art::x86::X86Assembler::shll and 492 // art::x86::X86Assembler::AddImmediate below), but it has 493 // not been saved by the previous call to 494 // art::SlowPathCode::SaveLiveRegisters, as it is a 495 // callee-save register -- 496 // art::SlowPathCode::SaveLiveRegisters does not consider 497 // callee-save registers, as it has been designed with the 498 // assumption that callee-save registers are supposed to be 499 // handled by the called function. So, as a callee-save 500 // register, `index_reg` _would_ eventually be saved onto 501 // the stack, but it would be too late: we would have 502 // changed its value earlier. Therefore, we manually save 503 // it here into another freely available register, 504 // `free_reg`, chosen of course among the caller-save 505 // registers (as a callee-save `free_reg` register would 506 // exhibit the same problem). 507 // 508 // Note we could have requested a temporary register from 509 // the register allocator instead; but we prefer not to, as 510 // this is a slow path, and we know we can find a 511 // caller-save register that is available. 512 Register free_reg = FindAvailableCallerSaveRegister(codegen); 513 __ movl(free_reg, index_reg); 514 index_reg = free_reg; 515 index = Location::RegisterLocation(index_reg); 516 } else { 517 // The initial register stored in `index_` has already been 518 // saved in the call to art::SlowPathCode::SaveLiveRegisters 519 // (as it is not a callee-save register), so we can freely 520 // use it. 521 } 522 // Shifting the index value contained in `index_reg` by the scale 523 // factor (2) cannot overflow in practice, as the runtime is 524 // unable to allocate object arrays with a size larger than 525 // 2^26 - 1 (that is, 2^28 - 4 bytes). 526 __ shll(index_reg, Immediate(TIMES_4)); 527 static_assert( 528 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 529 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 530 __ AddImmediate(index_reg, Immediate(offset_)); 531 } else { 532 DCHECK(instruction_->IsInvoke()); 533 DCHECK(instruction_->GetLocations()->Intrinsified()); 534 DCHECK((instruction_->AsInvoke()->GetIntrinsic() == Intrinsics::kUnsafeGetObject) || 535 (instruction_->AsInvoke()->GetIntrinsic() == Intrinsics::kUnsafeGetObjectVolatile)) 536 << instruction_->AsInvoke()->GetIntrinsic(); 537 DCHECK_EQ(offset_, 0U); 538 DCHECK(index_.IsRegisterPair()); 539 // UnsafeGet's offset location is a register pair, the low 540 // part contains the correct offset. 541 index = index_.ToLow(); 542 } 543 } 544 545 // We're moving two or three locations to locations that could 546 // overlap, so we need a parallel move resolver. 547 InvokeRuntimeCallingConvention calling_convention; 548 HParallelMove parallel_move(codegen->GetGraph()->GetArena()); 549 parallel_move.AddMove(ref_, 550 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 551 Primitive::kPrimNot, 552 nullptr); 553 parallel_move.AddMove(obj_, 554 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 555 Primitive::kPrimNot, 556 nullptr); 557 if (index.IsValid()) { 558 parallel_move.AddMove(index, 559 Location::RegisterLocation(calling_convention.GetRegisterAt(2)), 560 Primitive::kPrimInt, 561 nullptr); 562 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 563 } else { 564 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 565 __ movl(calling_convention.GetRegisterAt(2), Immediate(offset_)); 566 } 567 x86_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierSlow), 568 instruction_, 569 instruction_->GetDexPc(), 570 this); 571 CheckEntrypointTypes< 572 kQuickReadBarrierSlow, mirror::Object*, mirror::Object*, mirror::Object*, uint32_t>(); 573 x86_codegen->Move32(out_, Location::RegisterLocation(EAX)); 574 575 RestoreLiveRegisters(codegen, locations); 576 __ jmp(GetExitLabel()); 577 } 578 579 const char* GetDescription() const OVERRIDE { return "ReadBarrierForHeapReferenceSlowPathX86"; } 580 581 private: 582 Register FindAvailableCallerSaveRegister(CodeGenerator* codegen) { 583 size_t ref = static_cast<int>(ref_.AsRegister<Register>()); 584 size_t obj = static_cast<int>(obj_.AsRegister<Register>()); 585 for (size_t i = 0, e = codegen->GetNumberOfCoreRegisters(); i < e; ++i) { 586 if (i != ref && i != obj && !codegen->IsCoreCalleeSaveRegister(i)) { 587 return static_cast<Register>(i); 588 } 589 } 590 // We shall never fail to find a free caller-save register, as 591 // there are more than two core caller-save registers on x86 592 // (meaning it is possible to find one which is different from 593 // `ref` and `obj`). 594 DCHECK_GT(codegen->GetNumberOfCoreCallerSaveRegisters(), 2u); 595 LOG(FATAL) << "Could not find a free caller-save register"; 596 UNREACHABLE(); 597 } 598 599 HInstruction* const instruction_; 600 const Location out_; 601 const Location ref_; 602 const Location obj_; 603 const uint32_t offset_; 604 // An additional location containing an index to an array. 605 // Only used for HArrayGet and the UnsafeGetObject & 606 // UnsafeGetObjectVolatile intrinsics. 607 const Location index_; 608 609 DISALLOW_COPY_AND_ASSIGN(ReadBarrierForHeapReferenceSlowPathX86); 610}; 611 612// Slow path generating a read barrier for a GC root. 613class ReadBarrierForRootSlowPathX86 : public SlowPathCode { 614 public: 615 ReadBarrierForRootSlowPathX86(HInstruction* instruction, Location out, Location root) 616 : instruction_(instruction), out_(out), root_(root) {} 617 618 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 619 LocationSummary* locations = instruction_->GetLocations(); 620 Register reg_out = out_.AsRegister<Register>(); 621 DCHECK(locations->CanCall()); 622 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg_out)); 623 DCHECK(instruction_->IsLoadClass() || instruction_->IsLoadString()); 624 625 __ Bind(GetEntryLabel()); 626 SaveLiveRegisters(codegen, locations); 627 628 InvokeRuntimeCallingConvention calling_convention; 629 CodeGeneratorX86* x86_codegen = down_cast<CodeGeneratorX86*>(codegen); 630 x86_codegen->Move32(Location::RegisterLocation(calling_convention.GetRegisterAt(0)), root_); 631 x86_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierForRootSlow), 632 instruction_, 633 instruction_->GetDexPc(), 634 this); 635 CheckEntrypointTypes<kQuickReadBarrierForRootSlow, mirror::Object*, GcRoot<mirror::Object>*>(); 636 x86_codegen->Move32(out_, Location::RegisterLocation(EAX)); 637 638 RestoreLiveRegisters(codegen, locations); 639 __ jmp(GetExitLabel()); 640 } 641 642 const char* GetDescription() const OVERRIDE { return "ReadBarrierForRootSlowPathX86"; } 643 644 private: 645 HInstruction* const instruction_; 646 const Location out_; 647 const Location root_; 648 649 DISALLOW_COPY_AND_ASSIGN(ReadBarrierForRootSlowPathX86); 650}; 651 652#undef __ 653#define __ down_cast<X86Assembler*>(GetAssembler())-> 654 655inline Condition X86Condition(IfCondition cond) { 656 switch (cond) { 657 case kCondEQ: return kEqual; 658 case kCondNE: return kNotEqual; 659 case kCondLT: return kLess; 660 case kCondLE: return kLessEqual; 661 case kCondGT: return kGreater; 662 case kCondGE: return kGreaterEqual; 663 case kCondB: return kBelow; 664 case kCondBE: return kBelowEqual; 665 case kCondA: return kAbove; 666 case kCondAE: return kAboveEqual; 667 } 668 LOG(FATAL) << "Unreachable"; 669 UNREACHABLE(); 670} 671 672// Maps signed condition to unsigned condition and FP condition to x86 name. 673inline Condition X86UnsignedOrFPCondition(IfCondition cond) { 674 switch (cond) { 675 case kCondEQ: return kEqual; 676 case kCondNE: return kNotEqual; 677 // Signed to unsigned, and FP to x86 name. 678 case kCondLT: return kBelow; 679 case kCondLE: return kBelowEqual; 680 case kCondGT: return kAbove; 681 case kCondGE: return kAboveEqual; 682 // Unsigned remain unchanged. 683 case kCondB: return kBelow; 684 case kCondBE: return kBelowEqual; 685 case kCondA: return kAbove; 686 case kCondAE: return kAboveEqual; 687 } 688 LOG(FATAL) << "Unreachable"; 689 UNREACHABLE(); 690} 691 692void CodeGeneratorX86::DumpCoreRegister(std::ostream& stream, int reg) const { 693 stream << Register(reg); 694} 695 696void CodeGeneratorX86::DumpFloatingPointRegister(std::ostream& stream, int reg) const { 697 stream << XmmRegister(reg); 698} 699 700size_t CodeGeneratorX86::SaveCoreRegister(size_t stack_index, uint32_t reg_id) { 701 __ movl(Address(ESP, stack_index), static_cast<Register>(reg_id)); 702 return kX86WordSize; 703} 704 705size_t CodeGeneratorX86::RestoreCoreRegister(size_t stack_index, uint32_t reg_id) { 706 __ movl(static_cast<Register>(reg_id), Address(ESP, stack_index)); 707 return kX86WordSize; 708} 709 710size_t CodeGeneratorX86::SaveFloatingPointRegister(size_t stack_index, uint32_t reg_id) { 711 __ movsd(Address(ESP, stack_index), XmmRegister(reg_id)); 712 return GetFloatingPointSpillSlotSize(); 713} 714 715size_t CodeGeneratorX86::RestoreFloatingPointRegister(size_t stack_index, uint32_t reg_id) { 716 __ movsd(XmmRegister(reg_id), Address(ESP, stack_index)); 717 return GetFloatingPointSpillSlotSize(); 718} 719 720void CodeGeneratorX86::InvokeRuntime(QuickEntrypointEnum entrypoint, 721 HInstruction* instruction, 722 uint32_t dex_pc, 723 SlowPathCode* slow_path) { 724 InvokeRuntime(GetThreadOffset<kX86WordSize>(entrypoint).Int32Value(), 725 instruction, 726 dex_pc, 727 slow_path); 728} 729 730void CodeGeneratorX86::InvokeRuntime(int32_t entry_point_offset, 731 HInstruction* instruction, 732 uint32_t dex_pc, 733 SlowPathCode* slow_path) { 734 ValidateInvokeRuntime(instruction, slow_path); 735 __ fs()->call(Address::Absolute(entry_point_offset)); 736 RecordPcInfo(instruction, dex_pc, slow_path); 737} 738 739CodeGeneratorX86::CodeGeneratorX86(HGraph* graph, 740 const X86InstructionSetFeatures& isa_features, 741 const CompilerOptions& compiler_options, 742 OptimizingCompilerStats* stats) 743 : CodeGenerator(graph, 744 kNumberOfCpuRegisters, 745 kNumberOfXmmRegisters, 746 kNumberOfRegisterPairs, 747 ComputeRegisterMask(reinterpret_cast<const int*>(kCoreCalleeSaves), 748 arraysize(kCoreCalleeSaves)) 749 | (1 << kFakeReturnRegister), 750 0, 751 compiler_options, 752 stats), 753 block_labels_(nullptr), 754 location_builder_(graph, this), 755 instruction_visitor_(graph, this), 756 move_resolver_(graph->GetArena(), this), 757 isa_features_(isa_features), 758 method_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 759 relative_call_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 760 pc_relative_dex_cache_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 761 fixups_to_jump_tables_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)) { 762 // Use a fake return address register to mimic Quick. 763 AddAllocatedRegister(Location::RegisterLocation(kFakeReturnRegister)); 764} 765 766Location CodeGeneratorX86::AllocateFreeRegister(Primitive::Type type) const { 767 switch (type) { 768 case Primitive::kPrimLong: { 769 size_t reg = FindFreeEntry(blocked_register_pairs_, kNumberOfRegisterPairs); 770 X86ManagedRegister pair = 771 X86ManagedRegister::FromRegisterPair(static_cast<RegisterPair>(reg)); 772 DCHECK(!blocked_core_registers_[pair.AsRegisterPairLow()]); 773 DCHECK(!blocked_core_registers_[pair.AsRegisterPairHigh()]); 774 blocked_core_registers_[pair.AsRegisterPairLow()] = true; 775 blocked_core_registers_[pair.AsRegisterPairHigh()] = true; 776 UpdateBlockedPairRegisters(); 777 return Location::RegisterPairLocation(pair.AsRegisterPairLow(), pair.AsRegisterPairHigh()); 778 } 779 780 case Primitive::kPrimByte: 781 case Primitive::kPrimBoolean: 782 case Primitive::kPrimChar: 783 case Primitive::kPrimShort: 784 case Primitive::kPrimInt: 785 case Primitive::kPrimNot: { 786 Register reg = static_cast<Register>( 787 FindFreeEntry(blocked_core_registers_, kNumberOfCpuRegisters)); 788 // Block all register pairs that contain `reg`. 789 for (int i = 0; i < kNumberOfRegisterPairs; i++) { 790 X86ManagedRegister current = 791 X86ManagedRegister::FromRegisterPair(static_cast<RegisterPair>(i)); 792 if (current.AsRegisterPairLow() == reg || current.AsRegisterPairHigh() == reg) { 793 blocked_register_pairs_[i] = true; 794 } 795 } 796 return Location::RegisterLocation(reg); 797 } 798 799 case Primitive::kPrimFloat: 800 case Primitive::kPrimDouble: { 801 return Location::FpuRegisterLocation( 802 FindFreeEntry(blocked_fpu_registers_, kNumberOfXmmRegisters)); 803 } 804 805 case Primitive::kPrimVoid: 806 LOG(FATAL) << "Unreachable type " << type; 807 } 808 809 return Location::NoLocation(); 810} 811 812void CodeGeneratorX86::SetupBlockedRegisters(bool is_baseline) const { 813 // Don't allocate the dalvik style register pair passing. 814 blocked_register_pairs_[ECX_EDX] = true; 815 816 // Stack register is always reserved. 817 blocked_core_registers_[ESP] = true; 818 819 if (is_baseline) { 820 blocked_core_registers_[EBP] = true; 821 blocked_core_registers_[ESI] = true; 822 blocked_core_registers_[EDI] = true; 823 } 824 825 UpdateBlockedPairRegisters(); 826} 827 828void CodeGeneratorX86::UpdateBlockedPairRegisters() const { 829 for (int i = 0; i < kNumberOfRegisterPairs; i++) { 830 X86ManagedRegister current = 831 X86ManagedRegister::FromRegisterPair(static_cast<RegisterPair>(i)); 832 if (blocked_core_registers_[current.AsRegisterPairLow()] 833 || blocked_core_registers_[current.AsRegisterPairHigh()]) { 834 blocked_register_pairs_[i] = true; 835 } 836 } 837} 838 839InstructionCodeGeneratorX86::InstructionCodeGeneratorX86(HGraph* graph, CodeGeneratorX86* codegen) 840 : HGraphVisitor(graph), 841 assembler_(codegen->GetAssembler()), 842 codegen_(codegen) {} 843 844static dwarf::Reg DWARFReg(Register reg) { 845 return dwarf::Reg::X86Core(static_cast<int>(reg)); 846} 847 848void CodeGeneratorX86::GenerateFrameEntry() { 849 __ cfi().SetCurrentCFAOffset(kX86WordSize); // return address 850 __ Bind(&frame_entry_label_); 851 bool skip_overflow_check = 852 IsLeafMethod() && !FrameNeedsStackCheck(GetFrameSize(), InstructionSet::kX86); 853 DCHECK(GetCompilerOptions().GetImplicitStackOverflowChecks()); 854 855 if (!skip_overflow_check) { 856 __ testl(EAX, Address(ESP, -static_cast<int32_t>(GetStackOverflowReservedBytes(kX86)))); 857 RecordPcInfo(nullptr, 0); 858 } 859 860 if (HasEmptyFrame()) { 861 return; 862 } 863 864 for (int i = arraysize(kCoreCalleeSaves) - 1; i >= 0; --i) { 865 Register reg = kCoreCalleeSaves[i]; 866 if (allocated_registers_.ContainsCoreRegister(reg)) { 867 __ pushl(reg); 868 __ cfi().AdjustCFAOffset(kX86WordSize); 869 __ cfi().RelOffset(DWARFReg(reg), 0); 870 } 871 } 872 873 int adjust = GetFrameSize() - FrameEntrySpillSize(); 874 __ subl(ESP, Immediate(adjust)); 875 __ cfi().AdjustCFAOffset(adjust); 876 __ movl(Address(ESP, kCurrentMethodStackOffset), kMethodRegisterArgument); 877} 878 879void CodeGeneratorX86::GenerateFrameExit() { 880 __ cfi().RememberState(); 881 if (!HasEmptyFrame()) { 882 int adjust = GetFrameSize() - FrameEntrySpillSize(); 883 __ addl(ESP, Immediate(adjust)); 884 __ cfi().AdjustCFAOffset(-adjust); 885 886 for (size_t i = 0; i < arraysize(kCoreCalleeSaves); ++i) { 887 Register reg = kCoreCalleeSaves[i]; 888 if (allocated_registers_.ContainsCoreRegister(reg)) { 889 __ popl(reg); 890 __ cfi().AdjustCFAOffset(-static_cast<int>(kX86WordSize)); 891 __ cfi().Restore(DWARFReg(reg)); 892 } 893 } 894 } 895 __ ret(); 896 __ cfi().RestoreState(); 897 __ cfi().DefCFAOffset(GetFrameSize()); 898} 899 900void CodeGeneratorX86::Bind(HBasicBlock* block) { 901 __ Bind(GetLabelOf(block)); 902} 903 904Location CodeGeneratorX86::GetStackLocation(HLoadLocal* load) const { 905 switch (load->GetType()) { 906 case Primitive::kPrimLong: 907 case Primitive::kPrimDouble: 908 return Location::DoubleStackSlot(GetStackSlot(load->GetLocal())); 909 910 case Primitive::kPrimInt: 911 case Primitive::kPrimNot: 912 case Primitive::kPrimFloat: 913 return Location::StackSlot(GetStackSlot(load->GetLocal())); 914 915 case Primitive::kPrimBoolean: 916 case Primitive::kPrimByte: 917 case Primitive::kPrimChar: 918 case Primitive::kPrimShort: 919 case Primitive::kPrimVoid: 920 LOG(FATAL) << "Unexpected type " << load->GetType(); 921 UNREACHABLE(); 922 } 923 924 LOG(FATAL) << "Unreachable"; 925 UNREACHABLE(); 926} 927 928Location InvokeDexCallingConventionVisitorX86::GetReturnLocation(Primitive::Type type) const { 929 switch (type) { 930 case Primitive::kPrimBoolean: 931 case Primitive::kPrimByte: 932 case Primitive::kPrimChar: 933 case Primitive::kPrimShort: 934 case Primitive::kPrimInt: 935 case Primitive::kPrimNot: 936 return Location::RegisterLocation(EAX); 937 938 case Primitive::kPrimLong: 939 return Location::RegisterPairLocation(EAX, EDX); 940 941 case Primitive::kPrimVoid: 942 return Location::NoLocation(); 943 944 case Primitive::kPrimDouble: 945 case Primitive::kPrimFloat: 946 return Location::FpuRegisterLocation(XMM0); 947 } 948 949 UNREACHABLE(); 950} 951 952Location InvokeDexCallingConventionVisitorX86::GetMethodLocation() const { 953 return Location::RegisterLocation(kMethodRegisterArgument); 954} 955 956Location InvokeDexCallingConventionVisitorX86::GetNextLocation(Primitive::Type type) { 957 switch (type) { 958 case Primitive::kPrimBoolean: 959 case Primitive::kPrimByte: 960 case Primitive::kPrimChar: 961 case Primitive::kPrimShort: 962 case Primitive::kPrimInt: 963 case Primitive::kPrimNot: { 964 uint32_t index = gp_index_++; 965 stack_index_++; 966 if (index < calling_convention.GetNumberOfRegisters()) { 967 return Location::RegisterLocation(calling_convention.GetRegisterAt(index)); 968 } else { 969 return Location::StackSlot(calling_convention.GetStackOffsetOf(stack_index_ - 1)); 970 } 971 } 972 973 case Primitive::kPrimLong: { 974 uint32_t index = gp_index_; 975 gp_index_ += 2; 976 stack_index_ += 2; 977 if (index + 1 < calling_convention.GetNumberOfRegisters()) { 978 X86ManagedRegister pair = X86ManagedRegister::FromRegisterPair( 979 calling_convention.GetRegisterPairAt(index)); 980 return Location::RegisterPairLocation(pair.AsRegisterPairLow(), pair.AsRegisterPairHigh()); 981 } else { 982 return Location::DoubleStackSlot(calling_convention.GetStackOffsetOf(stack_index_ - 2)); 983 } 984 } 985 986 case Primitive::kPrimFloat: { 987 uint32_t index = float_index_++; 988 stack_index_++; 989 if (index < calling_convention.GetNumberOfFpuRegisters()) { 990 return Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(index)); 991 } else { 992 return Location::StackSlot(calling_convention.GetStackOffsetOf(stack_index_ - 1)); 993 } 994 } 995 996 case Primitive::kPrimDouble: { 997 uint32_t index = float_index_++; 998 stack_index_ += 2; 999 if (index < calling_convention.GetNumberOfFpuRegisters()) { 1000 return Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(index)); 1001 } else { 1002 return Location::DoubleStackSlot(calling_convention.GetStackOffsetOf(stack_index_ - 2)); 1003 } 1004 } 1005 1006 case Primitive::kPrimVoid: 1007 LOG(FATAL) << "Unexpected parameter type " << type; 1008 break; 1009 } 1010 return Location::NoLocation(); 1011} 1012 1013void CodeGeneratorX86::Move32(Location destination, Location source) { 1014 if (source.Equals(destination)) { 1015 return; 1016 } 1017 if (destination.IsRegister()) { 1018 if (source.IsRegister()) { 1019 __ movl(destination.AsRegister<Register>(), source.AsRegister<Register>()); 1020 } else if (source.IsFpuRegister()) { 1021 __ movd(destination.AsRegister<Register>(), source.AsFpuRegister<XmmRegister>()); 1022 } else { 1023 DCHECK(source.IsStackSlot()); 1024 __ movl(destination.AsRegister<Register>(), Address(ESP, source.GetStackIndex())); 1025 } 1026 } else if (destination.IsFpuRegister()) { 1027 if (source.IsRegister()) { 1028 __ movd(destination.AsFpuRegister<XmmRegister>(), source.AsRegister<Register>()); 1029 } else if (source.IsFpuRegister()) { 1030 __ movaps(destination.AsFpuRegister<XmmRegister>(), source.AsFpuRegister<XmmRegister>()); 1031 } else { 1032 DCHECK(source.IsStackSlot()); 1033 __ movss(destination.AsFpuRegister<XmmRegister>(), Address(ESP, source.GetStackIndex())); 1034 } 1035 } else { 1036 DCHECK(destination.IsStackSlot()) << destination; 1037 if (source.IsRegister()) { 1038 __ movl(Address(ESP, destination.GetStackIndex()), source.AsRegister<Register>()); 1039 } else if (source.IsFpuRegister()) { 1040 __ movss(Address(ESP, destination.GetStackIndex()), source.AsFpuRegister<XmmRegister>()); 1041 } else if (source.IsConstant()) { 1042 HConstant* constant = source.GetConstant(); 1043 int32_t value = GetInt32ValueOf(constant); 1044 __ movl(Address(ESP, destination.GetStackIndex()), Immediate(value)); 1045 } else { 1046 DCHECK(source.IsStackSlot()); 1047 __ pushl(Address(ESP, source.GetStackIndex())); 1048 __ popl(Address(ESP, destination.GetStackIndex())); 1049 } 1050 } 1051} 1052 1053void CodeGeneratorX86::Move64(Location destination, Location source) { 1054 if (source.Equals(destination)) { 1055 return; 1056 } 1057 if (destination.IsRegisterPair()) { 1058 if (source.IsRegisterPair()) { 1059 EmitParallelMoves( 1060 Location::RegisterLocation(source.AsRegisterPairHigh<Register>()), 1061 Location::RegisterLocation(destination.AsRegisterPairHigh<Register>()), 1062 Primitive::kPrimInt, 1063 Location::RegisterLocation(source.AsRegisterPairLow<Register>()), 1064 Location::RegisterLocation(destination.AsRegisterPairLow<Register>()), 1065 Primitive::kPrimInt); 1066 } else if (source.IsFpuRegister()) { 1067 XmmRegister src_reg = source.AsFpuRegister<XmmRegister>(); 1068 __ movd(destination.AsRegisterPairLow<Register>(), src_reg); 1069 __ psrlq(src_reg, Immediate(32)); 1070 __ movd(destination.AsRegisterPairHigh<Register>(), src_reg); 1071 } else { 1072 // No conflict possible, so just do the moves. 1073 DCHECK(source.IsDoubleStackSlot()); 1074 __ movl(destination.AsRegisterPairLow<Register>(), Address(ESP, source.GetStackIndex())); 1075 __ movl(destination.AsRegisterPairHigh<Register>(), 1076 Address(ESP, source.GetHighStackIndex(kX86WordSize))); 1077 } 1078 } else if (destination.IsFpuRegister()) { 1079 if (source.IsFpuRegister()) { 1080 __ movaps(destination.AsFpuRegister<XmmRegister>(), source.AsFpuRegister<XmmRegister>()); 1081 } else if (source.IsDoubleStackSlot()) { 1082 __ movsd(destination.AsFpuRegister<XmmRegister>(), Address(ESP, source.GetStackIndex())); 1083 } else if (source.IsRegisterPair()) { 1084 size_t elem_size = Primitive::ComponentSize(Primitive::kPrimInt); 1085 // Create stack space for 2 elements. 1086 __ subl(ESP, Immediate(2 * elem_size)); 1087 __ movl(Address(ESP, 0), source.AsRegisterPairLow<Register>()); 1088 __ movl(Address(ESP, elem_size), source.AsRegisterPairHigh<Register>()); 1089 __ movsd(destination.AsFpuRegister<XmmRegister>(), Address(ESP, 0)); 1090 // And remove the temporary stack space we allocated. 1091 __ addl(ESP, Immediate(2 * elem_size)); 1092 } else { 1093 LOG(FATAL) << "Unimplemented"; 1094 } 1095 } else { 1096 DCHECK(destination.IsDoubleStackSlot()) << destination; 1097 if (source.IsRegisterPair()) { 1098 // No conflict possible, so just do the moves. 1099 __ movl(Address(ESP, destination.GetStackIndex()), source.AsRegisterPairLow<Register>()); 1100 __ movl(Address(ESP, destination.GetHighStackIndex(kX86WordSize)), 1101 source.AsRegisterPairHigh<Register>()); 1102 } else if (source.IsFpuRegister()) { 1103 __ movsd(Address(ESP, destination.GetStackIndex()), source.AsFpuRegister<XmmRegister>()); 1104 } else if (source.IsConstant()) { 1105 HConstant* constant = source.GetConstant(); 1106 int64_t value; 1107 if (constant->IsLongConstant()) { 1108 value = constant->AsLongConstant()->GetValue(); 1109 } else { 1110 DCHECK(constant->IsDoubleConstant()); 1111 value = bit_cast<int64_t, double>(constant->AsDoubleConstant()->GetValue()); 1112 } 1113 __ movl(Address(ESP, destination.GetStackIndex()), Immediate(Low32Bits(value))); 1114 __ movl(Address(ESP, destination.GetHighStackIndex(kX86WordSize)), Immediate(High32Bits(value))); 1115 } else { 1116 DCHECK(source.IsDoubleStackSlot()) << source; 1117 EmitParallelMoves( 1118 Location::StackSlot(source.GetStackIndex()), 1119 Location::StackSlot(destination.GetStackIndex()), 1120 Primitive::kPrimInt, 1121 Location::StackSlot(source.GetHighStackIndex(kX86WordSize)), 1122 Location::StackSlot(destination.GetHighStackIndex(kX86WordSize)), 1123 Primitive::kPrimInt); 1124 } 1125 } 1126} 1127 1128void CodeGeneratorX86::Move(HInstruction* instruction, Location location, HInstruction* move_for) { 1129 LocationSummary* locations = instruction->GetLocations(); 1130 if (instruction->IsCurrentMethod()) { 1131 Move32(location, Location::StackSlot(kCurrentMethodStackOffset)); 1132 } else if (locations != nullptr && locations->Out().Equals(location)) { 1133 return; 1134 } else if (locations != nullptr && locations->Out().IsConstant()) { 1135 HConstant* const_to_move = locations->Out().GetConstant(); 1136 if (const_to_move->IsIntConstant() || const_to_move->IsNullConstant()) { 1137 Immediate imm(GetInt32ValueOf(const_to_move)); 1138 if (location.IsRegister()) { 1139 __ movl(location.AsRegister<Register>(), imm); 1140 } else if (location.IsStackSlot()) { 1141 __ movl(Address(ESP, location.GetStackIndex()), imm); 1142 } else { 1143 DCHECK(location.IsConstant()); 1144 DCHECK_EQ(location.GetConstant(), const_to_move); 1145 } 1146 } else if (const_to_move->IsLongConstant()) { 1147 int64_t value = const_to_move->AsLongConstant()->GetValue(); 1148 if (location.IsRegisterPair()) { 1149 __ movl(location.AsRegisterPairLow<Register>(), Immediate(Low32Bits(value))); 1150 __ movl(location.AsRegisterPairHigh<Register>(), Immediate(High32Bits(value))); 1151 } else if (location.IsDoubleStackSlot()) { 1152 __ movl(Address(ESP, location.GetStackIndex()), Immediate(Low32Bits(value))); 1153 __ movl(Address(ESP, location.GetHighStackIndex(kX86WordSize)), 1154 Immediate(High32Bits(value))); 1155 } else { 1156 DCHECK(location.IsConstant()); 1157 DCHECK_EQ(location.GetConstant(), instruction); 1158 } 1159 } 1160 } else if (instruction->IsTemporary()) { 1161 Location temp_location = GetTemporaryLocation(instruction->AsTemporary()); 1162 if (temp_location.IsStackSlot()) { 1163 Move32(location, temp_location); 1164 } else { 1165 DCHECK(temp_location.IsDoubleStackSlot()); 1166 Move64(location, temp_location); 1167 } 1168 } else if (instruction->IsLoadLocal()) { 1169 int slot = GetStackSlot(instruction->AsLoadLocal()->GetLocal()); 1170 switch (instruction->GetType()) { 1171 case Primitive::kPrimBoolean: 1172 case Primitive::kPrimByte: 1173 case Primitive::kPrimChar: 1174 case Primitive::kPrimShort: 1175 case Primitive::kPrimInt: 1176 case Primitive::kPrimNot: 1177 case Primitive::kPrimFloat: 1178 Move32(location, Location::StackSlot(slot)); 1179 break; 1180 1181 case Primitive::kPrimLong: 1182 case Primitive::kPrimDouble: 1183 Move64(location, Location::DoubleStackSlot(slot)); 1184 break; 1185 1186 default: 1187 LOG(FATAL) << "Unimplemented local type " << instruction->GetType(); 1188 } 1189 } else { 1190 DCHECK((instruction->GetNext() == move_for) || instruction->GetNext()->IsTemporary()); 1191 switch (instruction->GetType()) { 1192 case Primitive::kPrimBoolean: 1193 case Primitive::kPrimByte: 1194 case Primitive::kPrimChar: 1195 case Primitive::kPrimShort: 1196 case Primitive::kPrimInt: 1197 case Primitive::kPrimNot: 1198 case Primitive::kPrimFloat: 1199 Move32(location, locations->Out()); 1200 break; 1201 1202 case Primitive::kPrimLong: 1203 case Primitive::kPrimDouble: 1204 Move64(location, locations->Out()); 1205 break; 1206 1207 default: 1208 LOG(FATAL) << "Unexpected type " << instruction->GetType(); 1209 } 1210 } 1211} 1212 1213void CodeGeneratorX86::MoveConstant(Location location, int32_t value) { 1214 DCHECK(location.IsRegister()); 1215 __ movl(location.AsRegister<Register>(), Immediate(value)); 1216} 1217 1218void CodeGeneratorX86::MoveLocation(Location dst, Location src, Primitive::Type dst_type) { 1219 if (Primitive::Is64BitType(dst_type)) { 1220 Move64(dst, src); 1221 } else { 1222 Move32(dst, src); 1223 } 1224} 1225 1226void CodeGeneratorX86::AddLocationAsTemp(Location location, LocationSummary* locations) { 1227 if (location.IsRegister()) { 1228 locations->AddTemp(location); 1229 } else if (location.IsRegisterPair()) { 1230 locations->AddTemp(Location::RegisterLocation(location.AsRegisterPairLow<Register>())); 1231 locations->AddTemp(Location::RegisterLocation(location.AsRegisterPairHigh<Register>())); 1232 } else { 1233 UNIMPLEMENTED(FATAL) << "AddLocationAsTemp not implemented for location " << location; 1234 } 1235} 1236 1237void InstructionCodeGeneratorX86::HandleGoto(HInstruction* got, HBasicBlock* successor) { 1238 DCHECK(!successor->IsExitBlock()); 1239 1240 HBasicBlock* block = got->GetBlock(); 1241 HInstruction* previous = got->GetPrevious(); 1242 1243 HLoopInformation* info = block->GetLoopInformation(); 1244 if (info != nullptr && info->IsBackEdge(*block) && info->HasSuspendCheck()) { 1245 GenerateSuspendCheck(info->GetSuspendCheck(), successor); 1246 return; 1247 } 1248 1249 if (block->IsEntryBlock() && (previous != nullptr) && previous->IsSuspendCheck()) { 1250 GenerateSuspendCheck(previous->AsSuspendCheck(), nullptr); 1251 } 1252 if (!codegen_->GoesToNextBlock(got->GetBlock(), successor)) { 1253 __ jmp(codegen_->GetLabelOf(successor)); 1254 } 1255} 1256 1257void LocationsBuilderX86::VisitGoto(HGoto* got) { 1258 got->SetLocations(nullptr); 1259} 1260 1261void InstructionCodeGeneratorX86::VisitGoto(HGoto* got) { 1262 HandleGoto(got, got->GetSuccessor()); 1263} 1264 1265void LocationsBuilderX86::VisitTryBoundary(HTryBoundary* try_boundary) { 1266 try_boundary->SetLocations(nullptr); 1267} 1268 1269void InstructionCodeGeneratorX86::VisitTryBoundary(HTryBoundary* try_boundary) { 1270 HBasicBlock* successor = try_boundary->GetNormalFlowSuccessor(); 1271 if (!successor->IsExitBlock()) { 1272 HandleGoto(try_boundary, successor); 1273 } 1274} 1275 1276void LocationsBuilderX86::VisitExit(HExit* exit) { 1277 exit->SetLocations(nullptr); 1278} 1279 1280void InstructionCodeGeneratorX86::VisitExit(HExit* exit ATTRIBUTE_UNUSED) { 1281} 1282 1283void InstructionCodeGeneratorX86::GenerateFPJumps(HCondition* cond, 1284 Label* true_label, 1285 Label* false_label) { 1286 if (cond->IsFPConditionTrueIfNaN()) { 1287 __ j(kUnordered, true_label); 1288 } else if (cond->IsFPConditionFalseIfNaN()) { 1289 __ j(kUnordered, false_label); 1290 } 1291 __ j(X86UnsignedOrFPCondition(cond->GetCondition()), true_label); 1292} 1293 1294void InstructionCodeGeneratorX86::GenerateLongComparesAndJumps(HCondition* cond, 1295 Label* true_label, 1296 Label* false_label) { 1297 LocationSummary* locations = cond->GetLocations(); 1298 Location left = locations->InAt(0); 1299 Location right = locations->InAt(1); 1300 IfCondition if_cond = cond->GetCondition(); 1301 1302 Register left_high = left.AsRegisterPairHigh<Register>(); 1303 Register left_low = left.AsRegisterPairLow<Register>(); 1304 IfCondition true_high_cond = if_cond; 1305 IfCondition false_high_cond = cond->GetOppositeCondition(); 1306 Condition final_condition = X86UnsignedOrFPCondition(if_cond); // unsigned on lower part 1307 1308 // Set the conditions for the test, remembering that == needs to be 1309 // decided using the low words. 1310 switch (if_cond) { 1311 case kCondEQ: 1312 case kCondNE: 1313 // Nothing to do. 1314 break; 1315 case kCondLT: 1316 false_high_cond = kCondGT; 1317 break; 1318 case kCondLE: 1319 true_high_cond = kCondLT; 1320 break; 1321 case kCondGT: 1322 false_high_cond = kCondLT; 1323 break; 1324 case kCondGE: 1325 true_high_cond = kCondGT; 1326 break; 1327 case kCondB: 1328 false_high_cond = kCondA; 1329 break; 1330 case kCondBE: 1331 true_high_cond = kCondB; 1332 break; 1333 case kCondA: 1334 false_high_cond = kCondB; 1335 break; 1336 case kCondAE: 1337 true_high_cond = kCondA; 1338 break; 1339 } 1340 1341 if (right.IsConstant()) { 1342 int64_t value = right.GetConstant()->AsLongConstant()->GetValue(); 1343 int32_t val_high = High32Bits(value); 1344 int32_t val_low = Low32Bits(value); 1345 1346 if (val_high == 0) { 1347 __ testl(left_high, left_high); 1348 } else { 1349 __ cmpl(left_high, Immediate(val_high)); 1350 } 1351 if (if_cond == kCondNE) { 1352 __ j(X86Condition(true_high_cond), true_label); 1353 } else if (if_cond == kCondEQ) { 1354 __ j(X86Condition(false_high_cond), false_label); 1355 } else { 1356 __ j(X86Condition(true_high_cond), true_label); 1357 __ j(X86Condition(false_high_cond), false_label); 1358 } 1359 // Must be equal high, so compare the lows. 1360 if (val_low == 0) { 1361 __ testl(left_low, left_low); 1362 } else { 1363 __ cmpl(left_low, Immediate(val_low)); 1364 } 1365 } else { 1366 Register right_high = right.AsRegisterPairHigh<Register>(); 1367 Register right_low = right.AsRegisterPairLow<Register>(); 1368 1369 __ cmpl(left_high, right_high); 1370 if (if_cond == kCondNE) { 1371 __ j(X86Condition(true_high_cond), true_label); 1372 } else if (if_cond == kCondEQ) { 1373 __ j(X86Condition(false_high_cond), false_label); 1374 } else { 1375 __ j(X86Condition(true_high_cond), true_label); 1376 __ j(X86Condition(false_high_cond), false_label); 1377 } 1378 // Must be equal high, so compare the lows. 1379 __ cmpl(left_low, right_low); 1380 } 1381 // The last comparison might be unsigned. 1382 __ j(final_condition, true_label); 1383} 1384 1385void InstructionCodeGeneratorX86::GenerateCompareTestAndBranch(HCondition* condition, 1386 Label* true_target_in, 1387 Label* false_target_in) { 1388 // Generated branching requires both targets to be explicit. If either of the 1389 // targets is nullptr (fallthrough) use and bind `fallthrough_target` instead. 1390 Label fallthrough_target; 1391 Label* true_target = true_target_in == nullptr ? &fallthrough_target : true_target_in; 1392 Label* false_target = false_target_in == nullptr ? &fallthrough_target : false_target_in; 1393 1394 LocationSummary* locations = condition->GetLocations(); 1395 Location left = locations->InAt(0); 1396 Location right = locations->InAt(1); 1397 1398 Primitive::Type type = condition->InputAt(0)->GetType(); 1399 switch (type) { 1400 case Primitive::kPrimLong: 1401 GenerateLongComparesAndJumps(condition, true_target, false_target); 1402 break; 1403 case Primitive::kPrimFloat: 1404 __ ucomiss(left.AsFpuRegister<XmmRegister>(), right.AsFpuRegister<XmmRegister>()); 1405 GenerateFPJumps(condition, true_target, false_target); 1406 break; 1407 case Primitive::kPrimDouble: 1408 __ ucomisd(left.AsFpuRegister<XmmRegister>(), right.AsFpuRegister<XmmRegister>()); 1409 GenerateFPJumps(condition, true_target, false_target); 1410 break; 1411 default: 1412 LOG(FATAL) << "Unexpected compare type " << type; 1413 } 1414 1415 if (false_target != &fallthrough_target) { 1416 __ jmp(false_target); 1417 } 1418 1419 if (fallthrough_target.IsLinked()) { 1420 __ Bind(&fallthrough_target); 1421 } 1422} 1423 1424static bool AreEflagsSetFrom(HInstruction* cond, HInstruction* branch) { 1425 // Moves may affect the eflags register (move zero uses xorl), so the EFLAGS 1426 // are set only strictly before `branch`. We can't use the eflags on long/FP 1427 // conditions if they are materialized due to the complex branching. 1428 return cond->IsCondition() && 1429 cond->GetNext() == branch && 1430 cond->InputAt(0)->GetType() != Primitive::kPrimLong && 1431 !Primitive::IsFloatingPointType(cond->InputAt(0)->GetType()); 1432} 1433 1434void InstructionCodeGeneratorX86::GenerateTestAndBranch(HInstruction* instruction, 1435 size_t condition_input_index, 1436 Label* true_target, 1437 Label* false_target) { 1438 HInstruction* cond = instruction->InputAt(condition_input_index); 1439 1440 if (true_target == nullptr && false_target == nullptr) { 1441 // Nothing to do. The code always falls through. 1442 return; 1443 } else if (cond->IsIntConstant()) { 1444 // Constant condition, statically compared against 1. 1445 if (cond->AsIntConstant()->IsOne()) { 1446 if (true_target != nullptr) { 1447 __ jmp(true_target); 1448 } 1449 } else { 1450 DCHECK(cond->AsIntConstant()->IsZero()); 1451 if (false_target != nullptr) { 1452 __ jmp(false_target); 1453 } 1454 } 1455 return; 1456 } 1457 1458 // The following code generates these patterns: 1459 // (1) true_target == nullptr && false_target != nullptr 1460 // - opposite condition true => branch to false_target 1461 // (2) true_target != nullptr && false_target == nullptr 1462 // - condition true => branch to true_target 1463 // (3) true_target != nullptr && false_target != nullptr 1464 // - condition true => branch to true_target 1465 // - branch to false_target 1466 if (IsBooleanValueOrMaterializedCondition(cond)) { 1467 if (AreEflagsSetFrom(cond, instruction)) { 1468 if (true_target == nullptr) { 1469 __ j(X86Condition(cond->AsCondition()->GetOppositeCondition()), false_target); 1470 } else { 1471 __ j(X86Condition(cond->AsCondition()->GetCondition()), true_target); 1472 } 1473 } else { 1474 // Materialized condition, compare against 0. 1475 Location lhs = instruction->GetLocations()->InAt(condition_input_index); 1476 if (lhs.IsRegister()) { 1477 __ testl(lhs.AsRegister<Register>(), lhs.AsRegister<Register>()); 1478 } else { 1479 __ cmpl(Address(ESP, lhs.GetStackIndex()), Immediate(0)); 1480 } 1481 if (true_target == nullptr) { 1482 __ j(kEqual, false_target); 1483 } else { 1484 __ j(kNotEqual, true_target); 1485 } 1486 } 1487 } else { 1488 // Condition has not been materialized, use its inputs as the comparison and 1489 // its condition as the branch condition. 1490 HCondition* condition = cond->AsCondition(); 1491 1492 // If this is a long or FP comparison that has been folded into 1493 // the HCondition, generate the comparison directly. 1494 Primitive::Type type = condition->InputAt(0)->GetType(); 1495 if (type == Primitive::kPrimLong || Primitive::IsFloatingPointType(type)) { 1496 GenerateCompareTestAndBranch(condition, true_target, false_target); 1497 return; 1498 } 1499 1500 Location lhs = condition->GetLocations()->InAt(0); 1501 Location rhs = condition->GetLocations()->InAt(1); 1502 // LHS is guaranteed to be in a register (see LocationsBuilderX86::VisitCondition). 1503 if (rhs.IsRegister()) { 1504 __ cmpl(lhs.AsRegister<Register>(), rhs.AsRegister<Register>()); 1505 } else if (rhs.IsConstant()) { 1506 int32_t constant = CodeGenerator::GetInt32ValueOf(rhs.GetConstant()); 1507 if (constant == 0) { 1508 __ testl(lhs.AsRegister<Register>(), lhs.AsRegister<Register>()); 1509 } else { 1510 __ cmpl(lhs.AsRegister<Register>(), Immediate(constant)); 1511 } 1512 } else { 1513 __ cmpl(lhs.AsRegister<Register>(), Address(ESP, rhs.GetStackIndex())); 1514 } 1515 if (true_target == nullptr) { 1516 __ j(X86Condition(condition->GetOppositeCondition()), false_target); 1517 } else { 1518 __ j(X86Condition(condition->GetCondition()), true_target); 1519 } 1520 } 1521 1522 // If neither branch falls through (case 3), the conditional branch to `true_target` 1523 // was already emitted (case 2) and we need to emit a jump to `false_target`. 1524 if (true_target != nullptr && false_target != nullptr) { 1525 __ jmp(false_target); 1526 } 1527} 1528 1529void LocationsBuilderX86::VisitIf(HIf* if_instr) { 1530 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(if_instr); 1531 if (IsBooleanValueOrMaterializedCondition(if_instr->InputAt(0))) { 1532 locations->SetInAt(0, Location::Any()); 1533 } 1534} 1535 1536void InstructionCodeGeneratorX86::VisitIf(HIf* if_instr) { 1537 HBasicBlock* true_successor = if_instr->IfTrueSuccessor(); 1538 HBasicBlock* false_successor = if_instr->IfFalseSuccessor(); 1539 Label* true_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), true_successor) ? 1540 nullptr : codegen_->GetLabelOf(true_successor); 1541 Label* false_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), false_successor) ? 1542 nullptr : codegen_->GetLabelOf(false_successor); 1543 GenerateTestAndBranch(if_instr, /* condition_input_index */ 0, true_target, false_target); 1544} 1545 1546void LocationsBuilderX86::VisitDeoptimize(HDeoptimize* deoptimize) { 1547 LocationSummary* locations = new (GetGraph()->GetArena()) 1548 LocationSummary(deoptimize, LocationSummary::kCallOnSlowPath); 1549 if (IsBooleanValueOrMaterializedCondition(deoptimize->InputAt(0))) { 1550 locations->SetInAt(0, Location::Any()); 1551 } 1552} 1553 1554void InstructionCodeGeneratorX86::VisitDeoptimize(HDeoptimize* deoptimize) { 1555 SlowPathCode* slow_path = new (GetGraph()->GetArena()) 1556 DeoptimizationSlowPathX86(deoptimize); 1557 codegen_->AddSlowPath(slow_path); 1558 GenerateTestAndBranch(deoptimize, 1559 /* condition_input_index */ 0, 1560 slow_path->GetEntryLabel(), 1561 /* false_target */ nullptr); 1562} 1563 1564void LocationsBuilderX86::VisitLocal(HLocal* local) { 1565 local->SetLocations(nullptr); 1566} 1567 1568void InstructionCodeGeneratorX86::VisitLocal(HLocal* local) { 1569 DCHECK_EQ(local->GetBlock(), GetGraph()->GetEntryBlock()); 1570} 1571 1572void LocationsBuilderX86::VisitLoadLocal(HLoadLocal* local) { 1573 local->SetLocations(nullptr); 1574} 1575 1576void InstructionCodeGeneratorX86::VisitLoadLocal(HLoadLocal* load ATTRIBUTE_UNUSED) { 1577 // Nothing to do, this is driven by the code generator. 1578} 1579 1580void LocationsBuilderX86::VisitStoreLocal(HStoreLocal* store) { 1581 LocationSummary* locations = 1582 new (GetGraph()->GetArena()) LocationSummary(store, LocationSummary::kNoCall); 1583 switch (store->InputAt(1)->GetType()) { 1584 case Primitive::kPrimBoolean: 1585 case Primitive::kPrimByte: 1586 case Primitive::kPrimChar: 1587 case Primitive::kPrimShort: 1588 case Primitive::kPrimInt: 1589 case Primitive::kPrimNot: 1590 case Primitive::kPrimFloat: 1591 locations->SetInAt(1, Location::StackSlot(codegen_->GetStackSlot(store->GetLocal()))); 1592 break; 1593 1594 case Primitive::kPrimLong: 1595 case Primitive::kPrimDouble: 1596 locations->SetInAt(1, Location::DoubleStackSlot(codegen_->GetStackSlot(store->GetLocal()))); 1597 break; 1598 1599 default: 1600 LOG(FATAL) << "Unknown local type " << store->InputAt(1)->GetType(); 1601 } 1602} 1603 1604void InstructionCodeGeneratorX86::VisitStoreLocal(HStoreLocal* store ATTRIBUTE_UNUSED) { 1605} 1606 1607void LocationsBuilderX86::VisitCondition(HCondition* cond) { 1608 LocationSummary* locations = 1609 new (GetGraph()->GetArena()) LocationSummary(cond, LocationSummary::kNoCall); 1610 // Handle the long/FP comparisons made in instruction simplification. 1611 switch (cond->InputAt(0)->GetType()) { 1612 case Primitive::kPrimLong: { 1613 locations->SetInAt(0, Location::RequiresRegister()); 1614 locations->SetInAt(1, Location::RegisterOrConstant(cond->InputAt(1))); 1615 if (cond->NeedsMaterialization()) { 1616 locations->SetOut(Location::RequiresRegister()); 1617 } 1618 break; 1619 } 1620 case Primitive::kPrimFloat: 1621 case Primitive::kPrimDouble: { 1622 locations->SetInAt(0, Location::RequiresFpuRegister()); 1623 locations->SetInAt(1, Location::RequiresFpuRegister()); 1624 if (cond->NeedsMaterialization()) { 1625 locations->SetOut(Location::RequiresRegister()); 1626 } 1627 break; 1628 } 1629 default: 1630 locations->SetInAt(0, Location::RequiresRegister()); 1631 locations->SetInAt(1, Location::Any()); 1632 if (cond->NeedsMaterialization()) { 1633 // We need a byte register. 1634 locations->SetOut(Location::RegisterLocation(ECX)); 1635 } 1636 break; 1637 } 1638} 1639 1640void InstructionCodeGeneratorX86::VisitCondition(HCondition* cond) { 1641 if (!cond->NeedsMaterialization()) { 1642 return; 1643 } 1644 1645 LocationSummary* locations = cond->GetLocations(); 1646 Location lhs = locations->InAt(0); 1647 Location rhs = locations->InAt(1); 1648 Register reg = locations->Out().AsRegister<Register>(); 1649 Label true_label, false_label; 1650 1651 switch (cond->InputAt(0)->GetType()) { 1652 default: { 1653 // Integer case. 1654 1655 // Clear output register: setb only sets the low byte. 1656 __ xorl(reg, reg); 1657 1658 if (rhs.IsRegister()) { 1659 __ cmpl(lhs.AsRegister<Register>(), rhs.AsRegister<Register>()); 1660 } else if (rhs.IsConstant()) { 1661 int32_t constant = CodeGenerator::GetInt32ValueOf(rhs.GetConstant()); 1662 if (constant == 0) { 1663 __ testl(lhs.AsRegister<Register>(), lhs.AsRegister<Register>()); 1664 } else { 1665 __ cmpl(lhs.AsRegister<Register>(), Immediate(constant)); 1666 } 1667 } else { 1668 __ cmpl(lhs.AsRegister<Register>(), Address(ESP, rhs.GetStackIndex())); 1669 } 1670 __ setb(X86Condition(cond->GetCondition()), reg); 1671 return; 1672 } 1673 case Primitive::kPrimLong: 1674 GenerateLongComparesAndJumps(cond, &true_label, &false_label); 1675 break; 1676 case Primitive::kPrimFloat: 1677 __ ucomiss(lhs.AsFpuRegister<XmmRegister>(), rhs.AsFpuRegister<XmmRegister>()); 1678 GenerateFPJumps(cond, &true_label, &false_label); 1679 break; 1680 case Primitive::kPrimDouble: 1681 __ ucomisd(lhs.AsFpuRegister<XmmRegister>(), rhs.AsFpuRegister<XmmRegister>()); 1682 GenerateFPJumps(cond, &true_label, &false_label); 1683 break; 1684 } 1685 1686 // Convert the jumps into the result. 1687 NearLabel done_label; 1688 1689 // False case: result = 0. 1690 __ Bind(&false_label); 1691 __ xorl(reg, reg); 1692 __ jmp(&done_label); 1693 1694 // True case: result = 1. 1695 __ Bind(&true_label); 1696 __ movl(reg, Immediate(1)); 1697 __ Bind(&done_label); 1698} 1699 1700void LocationsBuilderX86::VisitEqual(HEqual* comp) { 1701 VisitCondition(comp); 1702} 1703 1704void InstructionCodeGeneratorX86::VisitEqual(HEqual* comp) { 1705 VisitCondition(comp); 1706} 1707 1708void LocationsBuilderX86::VisitNotEqual(HNotEqual* comp) { 1709 VisitCondition(comp); 1710} 1711 1712void InstructionCodeGeneratorX86::VisitNotEqual(HNotEqual* comp) { 1713 VisitCondition(comp); 1714} 1715 1716void LocationsBuilderX86::VisitLessThan(HLessThan* comp) { 1717 VisitCondition(comp); 1718} 1719 1720void InstructionCodeGeneratorX86::VisitLessThan(HLessThan* comp) { 1721 VisitCondition(comp); 1722} 1723 1724void LocationsBuilderX86::VisitLessThanOrEqual(HLessThanOrEqual* comp) { 1725 VisitCondition(comp); 1726} 1727 1728void InstructionCodeGeneratorX86::VisitLessThanOrEqual(HLessThanOrEqual* comp) { 1729 VisitCondition(comp); 1730} 1731 1732void LocationsBuilderX86::VisitGreaterThan(HGreaterThan* comp) { 1733 VisitCondition(comp); 1734} 1735 1736void InstructionCodeGeneratorX86::VisitGreaterThan(HGreaterThan* comp) { 1737 VisitCondition(comp); 1738} 1739 1740void LocationsBuilderX86::VisitGreaterThanOrEqual(HGreaterThanOrEqual* comp) { 1741 VisitCondition(comp); 1742} 1743 1744void InstructionCodeGeneratorX86::VisitGreaterThanOrEqual(HGreaterThanOrEqual* comp) { 1745 VisitCondition(comp); 1746} 1747 1748void LocationsBuilderX86::VisitBelow(HBelow* comp) { 1749 VisitCondition(comp); 1750} 1751 1752void InstructionCodeGeneratorX86::VisitBelow(HBelow* comp) { 1753 VisitCondition(comp); 1754} 1755 1756void LocationsBuilderX86::VisitBelowOrEqual(HBelowOrEqual* comp) { 1757 VisitCondition(comp); 1758} 1759 1760void InstructionCodeGeneratorX86::VisitBelowOrEqual(HBelowOrEqual* comp) { 1761 VisitCondition(comp); 1762} 1763 1764void LocationsBuilderX86::VisitAbove(HAbove* comp) { 1765 VisitCondition(comp); 1766} 1767 1768void InstructionCodeGeneratorX86::VisitAbove(HAbove* comp) { 1769 VisitCondition(comp); 1770} 1771 1772void LocationsBuilderX86::VisitAboveOrEqual(HAboveOrEqual* comp) { 1773 VisitCondition(comp); 1774} 1775 1776void InstructionCodeGeneratorX86::VisitAboveOrEqual(HAboveOrEqual* comp) { 1777 VisitCondition(comp); 1778} 1779 1780void LocationsBuilderX86::VisitIntConstant(HIntConstant* constant) { 1781 LocationSummary* locations = 1782 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1783 locations->SetOut(Location::ConstantLocation(constant)); 1784} 1785 1786void InstructionCodeGeneratorX86::VisitIntConstant(HIntConstant* constant ATTRIBUTE_UNUSED) { 1787 // Will be generated at use site. 1788} 1789 1790void LocationsBuilderX86::VisitNullConstant(HNullConstant* constant) { 1791 LocationSummary* locations = 1792 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1793 locations->SetOut(Location::ConstantLocation(constant)); 1794} 1795 1796void InstructionCodeGeneratorX86::VisitNullConstant(HNullConstant* constant ATTRIBUTE_UNUSED) { 1797 // Will be generated at use site. 1798} 1799 1800void LocationsBuilderX86::VisitLongConstant(HLongConstant* constant) { 1801 LocationSummary* locations = 1802 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1803 locations->SetOut(Location::ConstantLocation(constant)); 1804} 1805 1806void InstructionCodeGeneratorX86::VisitLongConstant(HLongConstant* constant ATTRIBUTE_UNUSED) { 1807 // Will be generated at use site. 1808} 1809 1810void LocationsBuilderX86::VisitFloatConstant(HFloatConstant* constant) { 1811 LocationSummary* locations = 1812 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1813 locations->SetOut(Location::ConstantLocation(constant)); 1814} 1815 1816void InstructionCodeGeneratorX86::VisitFloatConstant(HFloatConstant* constant ATTRIBUTE_UNUSED) { 1817 // Will be generated at use site. 1818} 1819 1820void LocationsBuilderX86::VisitDoubleConstant(HDoubleConstant* constant) { 1821 LocationSummary* locations = 1822 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1823 locations->SetOut(Location::ConstantLocation(constant)); 1824} 1825 1826void InstructionCodeGeneratorX86::VisitDoubleConstant(HDoubleConstant* constant ATTRIBUTE_UNUSED) { 1827 // Will be generated at use site. 1828} 1829 1830void LocationsBuilderX86::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) { 1831 memory_barrier->SetLocations(nullptr); 1832} 1833 1834void InstructionCodeGeneratorX86::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) { 1835 GenerateMemoryBarrier(memory_barrier->GetBarrierKind()); 1836} 1837 1838void LocationsBuilderX86::VisitReturnVoid(HReturnVoid* ret) { 1839 ret->SetLocations(nullptr); 1840} 1841 1842void InstructionCodeGeneratorX86::VisitReturnVoid(HReturnVoid* ret ATTRIBUTE_UNUSED) { 1843 codegen_->GenerateFrameExit(); 1844} 1845 1846void LocationsBuilderX86::VisitReturn(HReturn* ret) { 1847 LocationSummary* locations = 1848 new (GetGraph()->GetArena()) LocationSummary(ret, LocationSummary::kNoCall); 1849 switch (ret->InputAt(0)->GetType()) { 1850 case Primitive::kPrimBoolean: 1851 case Primitive::kPrimByte: 1852 case Primitive::kPrimChar: 1853 case Primitive::kPrimShort: 1854 case Primitive::kPrimInt: 1855 case Primitive::kPrimNot: 1856 locations->SetInAt(0, Location::RegisterLocation(EAX)); 1857 break; 1858 1859 case Primitive::kPrimLong: 1860 locations->SetInAt( 1861 0, Location::RegisterPairLocation(EAX, EDX)); 1862 break; 1863 1864 case Primitive::kPrimFloat: 1865 case Primitive::kPrimDouble: 1866 locations->SetInAt( 1867 0, Location::FpuRegisterLocation(XMM0)); 1868 break; 1869 1870 default: 1871 LOG(FATAL) << "Unknown return type " << ret->InputAt(0)->GetType(); 1872 } 1873} 1874 1875void InstructionCodeGeneratorX86::VisitReturn(HReturn* ret) { 1876 if (kIsDebugBuild) { 1877 switch (ret->InputAt(0)->GetType()) { 1878 case Primitive::kPrimBoolean: 1879 case Primitive::kPrimByte: 1880 case Primitive::kPrimChar: 1881 case Primitive::kPrimShort: 1882 case Primitive::kPrimInt: 1883 case Primitive::kPrimNot: 1884 DCHECK_EQ(ret->GetLocations()->InAt(0).AsRegister<Register>(), EAX); 1885 break; 1886 1887 case Primitive::kPrimLong: 1888 DCHECK_EQ(ret->GetLocations()->InAt(0).AsRegisterPairLow<Register>(), EAX); 1889 DCHECK_EQ(ret->GetLocations()->InAt(0).AsRegisterPairHigh<Register>(), EDX); 1890 break; 1891 1892 case Primitive::kPrimFloat: 1893 case Primitive::kPrimDouble: 1894 DCHECK_EQ(ret->GetLocations()->InAt(0).AsFpuRegister<XmmRegister>(), XMM0); 1895 break; 1896 1897 default: 1898 LOG(FATAL) << "Unknown return type " << ret->InputAt(0)->GetType(); 1899 } 1900 } 1901 codegen_->GenerateFrameExit(); 1902} 1903 1904void LocationsBuilderX86::VisitInvokeUnresolved(HInvokeUnresolved* invoke) { 1905 // The trampoline uses the same calling convention as dex calling conventions, 1906 // except instead of loading arg0/r0 with the target Method*, arg0/r0 will contain 1907 // the method_idx. 1908 HandleInvoke(invoke); 1909} 1910 1911void InstructionCodeGeneratorX86::VisitInvokeUnresolved(HInvokeUnresolved* invoke) { 1912 codegen_->GenerateInvokeUnresolvedRuntimeCall(invoke); 1913} 1914 1915void LocationsBuilderX86::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) { 1916 // When we do not run baseline, explicit clinit checks triggered by static 1917 // invokes must have been pruned by art::PrepareForRegisterAllocation. 1918 DCHECK(codegen_->IsBaseline() || !invoke->IsStaticWithExplicitClinitCheck()); 1919 1920 IntrinsicLocationsBuilderX86 intrinsic(codegen_); 1921 if (intrinsic.TryDispatch(invoke)) { 1922 if (invoke->GetLocations()->CanCall() && invoke->HasPcRelativeDexCache()) { 1923 invoke->GetLocations()->SetInAt(invoke->GetSpecialInputIndex(), Location::Any()); 1924 } 1925 return; 1926 } 1927 1928 HandleInvoke(invoke); 1929 1930 // For PC-relative dex cache the invoke has an extra input, the PC-relative address base. 1931 if (invoke->HasPcRelativeDexCache()) { 1932 invoke->GetLocations()->SetInAt(invoke->GetSpecialInputIndex(), 1933 Location::RequiresRegister()); 1934 } 1935 1936 if (codegen_->IsBaseline()) { 1937 // Baseline does not have enough registers if the current method also 1938 // needs a register. We therefore do not require a register for it, and let 1939 // the code generation of the invoke handle it. 1940 LocationSummary* locations = invoke->GetLocations(); 1941 Location location = locations->InAt(invoke->GetSpecialInputIndex()); 1942 if (location.IsUnallocated() && location.GetPolicy() == Location::kRequiresRegister) { 1943 locations->SetInAt(invoke->GetSpecialInputIndex(), Location::NoLocation()); 1944 } 1945 } 1946} 1947 1948static bool TryGenerateIntrinsicCode(HInvoke* invoke, CodeGeneratorX86* codegen) { 1949 if (invoke->GetLocations()->Intrinsified()) { 1950 IntrinsicCodeGeneratorX86 intrinsic(codegen); 1951 intrinsic.Dispatch(invoke); 1952 return true; 1953 } 1954 return false; 1955} 1956 1957void InstructionCodeGeneratorX86::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) { 1958 // When we do not run baseline, explicit clinit checks triggered by static 1959 // invokes must have been pruned by art::PrepareForRegisterAllocation. 1960 DCHECK(codegen_->IsBaseline() || !invoke->IsStaticWithExplicitClinitCheck()); 1961 1962 if (TryGenerateIntrinsicCode(invoke, codegen_)) { 1963 return; 1964 } 1965 1966 LocationSummary* locations = invoke->GetLocations(); 1967 codegen_->GenerateStaticOrDirectCall( 1968 invoke, locations->HasTemps() ? locations->GetTemp(0) : Location::NoLocation()); 1969 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 1970} 1971 1972void LocationsBuilderX86::VisitInvokeVirtual(HInvokeVirtual* invoke) { 1973 HandleInvoke(invoke); 1974} 1975 1976void LocationsBuilderX86::HandleInvoke(HInvoke* invoke) { 1977 InvokeDexCallingConventionVisitorX86 calling_convention_visitor; 1978 CodeGenerator::CreateCommonInvokeLocationSummary(invoke, &calling_convention_visitor); 1979} 1980 1981void InstructionCodeGeneratorX86::VisitInvokeVirtual(HInvokeVirtual* invoke) { 1982 if (TryGenerateIntrinsicCode(invoke, codegen_)) { 1983 return; 1984 } 1985 1986 codegen_->GenerateVirtualCall(invoke, invoke->GetLocations()->GetTemp(0)); 1987 DCHECK(!codegen_->IsLeafMethod()); 1988 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 1989} 1990 1991void LocationsBuilderX86::VisitInvokeInterface(HInvokeInterface* invoke) { 1992 // This call to HandleInvoke allocates a temporary (core) register 1993 // which is also used to transfer the hidden argument from FP to 1994 // core register. 1995 HandleInvoke(invoke); 1996 // Add the hidden argument. 1997 invoke->GetLocations()->AddTemp(Location::FpuRegisterLocation(XMM7)); 1998} 1999 2000void InstructionCodeGeneratorX86::VisitInvokeInterface(HInvokeInterface* invoke) { 2001 // TODO: b/18116999, our IMTs can miss an IncompatibleClassChangeError. 2002 LocationSummary* locations = invoke->GetLocations(); 2003 Register temp = locations->GetTemp(0).AsRegister<Register>(); 2004 XmmRegister hidden_reg = locations->GetTemp(1).AsFpuRegister<XmmRegister>(); 2005 uint32_t method_offset = mirror::Class::EmbeddedImTableEntryOffset( 2006 invoke->GetImtIndex() % mirror::Class::kImtSize, kX86PointerSize).Uint32Value(); 2007 Location receiver = locations->InAt(0); 2008 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 2009 2010 // Set the hidden argument. This is safe to do this here, as XMM7 2011 // won't be modified thereafter, before the `call` instruction. 2012 DCHECK_EQ(XMM7, hidden_reg); 2013 __ movl(temp, Immediate(invoke->GetDexMethodIndex())); 2014 __ movd(hidden_reg, temp); 2015 2016 if (receiver.IsStackSlot()) { 2017 __ movl(temp, Address(ESP, receiver.GetStackIndex())); 2018 // /* HeapReference<Class> */ temp = temp->klass_ 2019 __ movl(temp, Address(temp, class_offset)); 2020 } else { 2021 // /* HeapReference<Class> */ temp = receiver->klass_ 2022 __ movl(temp, Address(receiver.AsRegister<Register>(), class_offset)); 2023 } 2024 codegen_->MaybeRecordImplicitNullCheck(invoke); 2025 // Instead of simply (possibly) unpoisoning `temp` here, we should 2026 // emit a read barrier for the previous class reference load. 2027 // However this is not required in practice, as this is an 2028 // intermediate/temporary reference and because the current 2029 // concurrent copying collector keeps the from-space memory 2030 // intact/accessible until the end of the marking phase (the 2031 // concurrent copying collector may not in the future). 2032 __ MaybeUnpoisonHeapReference(temp); 2033 // temp = temp->GetImtEntryAt(method_offset); 2034 __ movl(temp, Address(temp, method_offset)); 2035 // call temp->GetEntryPoint(); 2036 __ call(Address(temp, 2037 ArtMethod::EntryPointFromQuickCompiledCodeOffset(kX86WordSize).Int32Value())); 2038 2039 DCHECK(!codegen_->IsLeafMethod()); 2040 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 2041} 2042 2043void LocationsBuilderX86::VisitNeg(HNeg* neg) { 2044 LocationSummary* locations = 2045 new (GetGraph()->GetArena()) LocationSummary(neg, LocationSummary::kNoCall); 2046 switch (neg->GetResultType()) { 2047 case Primitive::kPrimInt: 2048 case Primitive::kPrimLong: 2049 locations->SetInAt(0, Location::RequiresRegister()); 2050 locations->SetOut(Location::SameAsFirstInput()); 2051 break; 2052 2053 case Primitive::kPrimFloat: 2054 locations->SetInAt(0, Location::RequiresFpuRegister()); 2055 locations->SetOut(Location::SameAsFirstInput()); 2056 locations->AddTemp(Location::RequiresRegister()); 2057 locations->AddTemp(Location::RequiresFpuRegister()); 2058 break; 2059 2060 case Primitive::kPrimDouble: 2061 locations->SetInAt(0, Location::RequiresFpuRegister()); 2062 locations->SetOut(Location::SameAsFirstInput()); 2063 locations->AddTemp(Location::RequiresFpuRegister()); 2064 break; 2065 2066 default: 2067 LOG(FATAL) << "Unexpected neg type " << neg->GetResultType(); 2068 } 2069} 2070 2071void InstructionCodeGeneratorX86::VisitNeg(HNeg* neg) { 2072 LocationSummary* locations = neg->GetLocations(); 2073 Location out = locations->Out(); 2074 Location in = locations->InAt(0); 2075 switch (neg->GetResultType()) { 2076 case Primitive::kPrimInt: 2077 DCHECK(in.IsRegister()); 2078 DCHECK(in.Equals(out)); 2079 __ negl(out.AsRegister<Register>()); 2080 break; 2081 2082 case Primitive::kPrimLong: 2083 DCHECK(in.IsRegisterPair()); 2084 DCHECK(in.Equals(out)); 2085 __ negl(out.AsRegisterPairLow<Register>()); 2086 // Negation is similar to subtraction from zero. The least 2087 // significant byte triggers a borrow when it is different from 2088 // zero; to take it into account, add 1 to the most significant 2089 // byte if the carry flag (CF) is set to 1 after the first NEGL 2090 // operation. 2091 __ adcl(out.AsRegisterPairHigh<Register>(), Immediate(0)); 2092 __ negl(out.AsRegisterPairHigh<Register>()); 2093 break; 2094 2095 case Primitive::kPrimFloat: { 2096 DCHECK(in.Equals(out)); 2097 Register constant = locations->GetTemp(0).AsRegister<Register>(); 2098 XmmRegister mask = locations->GetTemp(1).AsFpuRegister<XmmRegister>(); 2099 // Implement float negation with an exclusive or with value 2100 // 0x80000000 (mask for bit 31, representing the sign of a 2101 // single-precision floating-point number). 2102 __ movl(constant, Immediate(INT32_C(0x80000000))); 2103 __ movd(mask, constant); 2104 __ xorps(out.AsFpuRegister<XmmRegister>(), mask); 2105 break; 2106 } 2107 2108 case Primitive::kPrimDouble: { 2109 DCHECK(in.Equals(out)); 2110 XmmRegister mask = locations->GetTemp(0).AsFpuRegister<XmmRegister>(); 2111 // Implement double negation with an exclusive or with value 2112 // 0x8000000000000000 (mask for bit 63, representing the sign of 2113 // a double-precision floating-point number). 2114 __ LoadLongConstant(mask, INT64_C(0x8000000000000000)); 2115 __ xorpd(out.AsFpuRegister<XmmRegister>(), mask); 2116 break; 2117 } 2118 2119 default: 2120 LOG(FATAL) << "Unexpected neg type " << neg->GetResultType(); 2121 } 2122} 2123 2124void LocationsBuilderX86::VisitTypeConversion(HTypeConversion* conversion) { 2125 Primitive::Type result_type = conversion->GetResultType(); 2126 Primitive::Type input_type = conversion->GetInputType(); 2127 DCHECK_NE(result_type, input_type); 2128 2129 // The float-to-long and double-to-long type conversions rely on a 2130 // call to the runtime. 2131 LocationSummary::CallKind call_kind = 2132 ((input_type == Primitive::kPrimFloat || input_type == Primitive::kPrimDouble) 2133 && result_type == Primitive::kPrimLong) 2134 ? LocationSummary::kCall 2135 : LocationSummary::kNoCall; 2136 LocationSummary* locations = 2137 new (GetGraph()->GetArena()) LocationSummary(conversion, call_kind); 2138 2139 // The Java language does not allow treating boolean as an integral type but 2140 // our bit representation makes it safe. 2141 2142 switch (result_type) { 2143 case Primitive::kPrimByte: 2144 switch (input_type) { 2145 case Primitive::kPrimBoolean: 2146 // Boolean input is a result of code transformations. 2147 case Primitive::kPrimShort: 2148 case Primitive::kPrimInt: 2149 case Primitive::kPrimChar: 2150 // Processing a Dex `int-to-byte' instruction. 2151 locations->SetInAt(0, Location::ByteRegisterOrConstant(ECX, conversion->InputAt(0))); 2152 // Make the output overlap to please the register allocator. This greatly simplifies 2153 // the validation of the linear scan implementation 2154 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 2155 break; 2156 2157 default: 2158 LOG(FATAL) << "Unexpected type conversion from " << input_type 2159 << " to " << result_type; 2160 } 2161 break; 2162 2163 case Primitive::kPrimShort: 2164 switch (input_type) { 2165 case Primitive::kPrimBoolean: 2166 // Boolean input is a result of code transformations. 2167 case Primitive::kPrimByte: 2168 case Primitive::kPrimInt: 2169 case Primitive::kPrimChar: 2170 // Processing a Dex `int-to-short' instruction. 2171 locations->SetInAt(0, Location::Any()); 2172 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2173 break; 2174 2175 default: 2176 LOG(FATAL) << "Unexpected type conversion from " << input_type 2177 << " to " << result_type; 2178 } 2179 break; 2180 2181 case Primitive::kPrimInt: 2182 switch (input_type) { 2183 case Primitive::kPrimLong: 2184 // Processing a Dex `long-to-int' instruction. 2185 locations->SetInAt(0, Location::Any()); 2186 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2187 break; 2188 2189 case Primitive::kPrimFloat: 2190 // Processing a Dex `float-to-int' instruction. 2191 locations->SetInAt(0, Location::RequiresFpuRegister()); 2192 locations->SetOut(Location::RequiresRegister()); 2193 locations->AddTemp(Location::RequiresFpuRegister()); 2194 break; 2195 2196 case Primitive::kPrimDouble: 2197 // Processing a Dex `double-to-int' instruction. 2198 locations->SetInAt(0, Location::RequiresFpuRegister()); 2199 locations->SetOut(Location::RequiresRegister()); 2200 locations->AddTemp(Location::RequiresFpuRegister()); 2201 break; 2202 2203 default: 2204 LOG(FATAL) << "Unexpected type conversion from " << input_type 2205 << " to " << result_type; 2206 } 2207 break; 2208 2209 case Primitive::kPrimLong: 2210 switch (input_type) { 2211 case Primitive::kPrimBoolean: 2212 // Boolean input is a result of code transformations. 2213 case Primitive::kPrimByte: 2214 case Primitive::kPrimShort: 2215 case Primitive::kPrimInt: 2216 case Primitive::kPrimChar: 2217 // Processing a Dex `int-to-long' instruction. 2218 locations->SetInAt(0, Location::RegisterLocation(EAX)); 2219 locations->SetOut(Location::RegisterPairLocation(EAX, EDX)); 2220 break; 2221 2222 case Primitive::kPrimFloat: 2223 case Primitive::kPrimDouble: { 2224 // Processing a Dex `float-to-long' or 'double-to-long' instruction. 2225 InvokeRuntimeCallingConvention calling_convention; 2226 XmmRegister parameter = calling_convention.GetFpuRegisterAt(0); 2227 locations->SetInAt(0, Location::FpuRegisterLocation(parameter)); 2228 2229 // The runtime helper puts the result in EAX, EDX. 2230 locations->SetOut(Location::RegisterPairLocation(EAX, EDX)); 2231 } 2232 break; 2233 2234 default: 2235 LOG(FATAL) << "Unexpected type conversion from " << input_type 2236 << " to " << result_type; 2237 } 2238 break; 2239 2240 case Primitive::kPrimChar: 2241 switch (input_type) { 2242 case Primitive::kPrimBoolean: 2243 // Boolean input is a result of code transformations. 2244 case Primitive::kPrimByte: 2245 case Primitive::kPrimShort: 2246 case Primitive::kPrimInt: 2247 // Processing a Dex `int-to-char' instruction. 2248 locations->SetInAt(0, Location::Any()); 2249 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2250 break; 2251 2252 default: 2253 LOG(FATAL) << "Unexpected type conversion from " << input_type 2254 << " to " << result_type; 2255 } 2256 break; 2257 2258 case Primitive::kPrimFloat: 2259 switch (input_type) { 2260 case Primitive::kPrimBoolean: 2261 // Boolean input is a result of code transformations. 2262 case Primitive::kPrimByte: 2263 case Primitive::kPrimShort: 2264 case Primitive::kPrimInt: 2265 case Primitive::kPrimChar: 2266 // Processing a Dex `int-to-float' instruction. 2267 locations->SetInAt(0, Location::RequiresRegister()); 2268 locations->SetOut(Location::RequiresFpuRegister()); 2269 break; 2270 2271 case Primitive::kPrimLong: 2272 // Processing a Dex `long-to-float' instruction. 2273 locations->SetInAt(0, Location::Any()); 2274 locations->SetOut(Location::Any()); 2275 break; 2276 2277 case Primitive::kPrimDouble: 2278 // Processing a Dex `double-to-float' instruction. 2279 locations->SetInAt(0, Location::RequiresFpuRegister()); 2280 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2281 break; 2282 2283 default: 2284 LOG(FATAL) << "Unexpected type conversion from " << input_type 2285 << " to " << result_type; 2286 }; 2287 break; 2288 2289 case Primitive::kPrimDouble: 2290 switch (input_type) { 2291 case Primitive::kPrimBoolean: 2292 // Boolean input is a result of code transformations. 2293 case Primitive::kPrimByte: 2294 case Primitive::kPrimShort: 2295 case Primitive::kPrimInt: 2296 case Primitive::kPrimChar: 2297 // Processing a Dex `int-to-double' instruction. 2298 locations->SetInAt(0, Location::RequiresRegister()); 2299 locations->SetOut(Location::RequiresFpuRegister()); 2300 break; 2301 2302 case Primitive::kPrimLong: 2303 // Processing a Dex `long-to-double' instruction. 2304 locations->SetInAt(0, Location::Any()); 2305 locations->SetOut(Location::Any()); 2306 break; 2307 2308 case Primitive::kPrimFloat: 2309 // Processing a Dex `float-to-double' instruction. 2310 locations->SetInAt(0, Location::RequiresFpuRegister()); 2311 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2312 break; 2313 2314 default: 2315 LOG(FATAL) << "Unexpected type conversion from " << input_type 2316 << " to " << result_type; 2317 } 2318 break; 2319 2320 default: 2321 LOG(FATAL) << "Unexpected type conversion from " << input_type 2322 << " to " << result_type; 2323 } 2324} 2325 2326void InstructionCodeGeneratorX86::VisitTypeConversion(HTypeConversion* conversion) { 2327 LocationSummary* locations = conversion->GetLocations(); 2328 Location out = locations->Out(); 2329 Location in = locations->InAt(0); 2330 Primitive::Type result_type = conversion->GetResultType(); 2331 Primitive::Type input_type = conversion->GetInputType(); 2332 DCHECK_NE(result_type, input_type); 2333 switch (result_type) { 2334 case Primitive::kPrimByte: 2335 switch (input_type) { 2336 case Primitive::kPrimBoolean: 2337 // Boolean input is a result of code transformations. 2338 case Primitive::kPrimShort: 2339 case Primitive::kPrimInt: 2340 case Primitive::kPrimChar: 2341 // Processing a Dex `int-to-byte' instruction. 2342 if (in.IsRegister()) { 2343 __ movsxb(out.AsRegister<Register>(), in.AsRegister<ByteRegister>()); 2344 } else { 2345 DCHECK(in.GetConstant()->IsIntConstant()); 2346 int32_t value = in.GetConstant()->AsIntConstant()->GetValue(); 2347 __ movl(out.AsRegister<Register>(), Immediate(static_cast<int8_t>(value))); 2348 } 2349 break; 2350 2351 default: 2352 LOG(FATAL) << "Unexpected type conversion from " << input_type 2353 << " to " << result_type; 2354 } 2355 break; 2356 2357 case Primitive::kPrimShort: 2358 switch (input_type) { 2359 case Primitive::kPrimBoolean: 2360 // Boolean input is a result of code transformations. 2361 case Primitive::kPrimByte: 2362 case Primitive::kPrimInt: 2363 case Primitive::kPrimChar: 2364 // Processing a Dex `int-to-short' instruction. 2365 if (in.IsRegister()) { 2366 __ movsxw(out.AsRegister<Register>(), in.AsRegister<Register>()); 2367 } else if (in.IsStackSlot()) { 2368 __ movsxw(out.AsRegister<Register>(), Address(ESP, in.GetStackIndex())); 2369 } else { 2370 DCHECK(in.GetConstant()->IsIntConstant()); 2371 int32_t value = in.GetConstant()->AsIntConstant()->GetValue(); 2372 __ movl(out.AsRegister<Register>(), Immediate(static_cast<int16_t>(value))); 2373 } 2374 break; 2375 2376 default: 2377 LOG(FATAL) << "Unexpected type conversion from " << input_type 2378 << " to " << result_type; 2379 } 2380 break; 2381 2382 case Primitive::kPrimInt: 2383 switch (input_type) { 2384 case Primitive::kPrimLong: 2385 // Processing a Dex `long-to-int' instruction. 2386 if (in.IsRegisterPair()) { 2387 __ movl(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>()); 2388 } else if (in.IsDoubleStackSlot()) { 2389 __ movl(out.AsRegister<Register>(), Address(ESP, in.GetStackIndex())); 2390 } else { 2391 DCHECK(in.IsConstant()); 2392 DCHECK(in.GetConstant()->IsLongConstant()); 2393 int64_t value = in.GetConstant()->AsLongConstant()->GetValue(); 2394 __ movl(out.AsRegister<Register>(), Immediate(static_cast<int32_t>(value))); 2395 } 2396 break; 2397 2398 case Primitive::kPrimFloat: { 2399 // Processing a Dex `float-to-int' instruction. 2400 XmmRegister input = in.AsFpuRegister<XmmRegister>(); 2401 Register output = out.AsRegister<Register>(); 2402 XmmRegister temp = locations->GetTemp(0).AsFpuRegister<XmmRegister>(); 2403 NearLabel done, nan; 2404 2405 __ movl(output, Immediate(kPrimIntMax)); 2406 // temp = int-to-float(output) 2407 __ cvtsi2ss(temp, output); 2408 // if input >= temp goto done 2409 __ comiss(input, temp); 2410 __ j(kAboveEqual, &done); 2411 // if input == NaN goto nan 2412 __ j(kUnordered, &nan); 2413 // output = float-to-int-truncate(input) 2414 __ cvttss2si(output, input); 2415 __ jmp(&done); 2416 __ Bind(&nan); 2417 // output = 0 2418 __ xorl(output, output); 2419 __ Bind(&done); 2420 break; 2421 } 2422 2423 case Primitive::kPrimDouble: { 2424 // Processing a Dex `double-to-int' instruction. 2425 XmmRegister input = in.AsFpuRegister<XmmRegister>(); 2426 Register output = out.AsRegister<Register>(); 2427 XmmRegister temp = locations->GetTemp(0).AsFpuRegister<XmmRegister>(); 2428 NearLabel done, nan; 2429 2430 __ movl(output, Immediate(kPrimIntMax)); 2431 // temp = int-to-double(output) 2432 __ cvtsi2sd(temp, output); 2433 // if input >= temp goto done 2434 __ comisd(input, temp); 2435 __ j(kAboveEqual, &done); 2436 // if input == NaN goto nan 2437 __ j(kUnordered, &nan); 2438 // output = double-to-int-truncate(input) 2439 __ cvttsd2si(output, input); 2440 __ jmp(&done); 2441 __ Bind(&nan); 2442 // output = 0 2443 __ xorl(output, output); 2444 __ Bind(&done); 2445 break; 2446 } 2447 2448 default: 2449 LOG(FATAL) << "Unexpected type conversion from " << input_type 2450 << " to " << result_type; 2451 } 2452 break; 2453 2454 case Primitive::kPrimLong: 2455 switch (input_type) { 2456 case Primitive::kPrimBoolean: 2457 // Boolean input is a result of code transformations. 2458 case Primitive::kPrimByte: 2459 case Primitive::kPrimShort: 2460 case Primitive::kPrimInt: 2461 case Primitive::kPrimChar: 2462 // Processing a Dex `int-to-long' instruction. 2463 DCHECK_EQ(out.AsRegisterPairLow<Register>(), EAX); 2464 DCHECK_EQ(out.AsRegisterPairHigh<Register>(), EDX); 2465 DCHECK_EQ(in.AsRegister<Register>(), EAX); 2466 __ cdq(); 2467 break; 2468 2469 case Primitive::kPrimFloat: 2470 // Processing a Dex `float-to-long' instruction. 2471 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pF2l), 2472 conversion, 2473 conversion->GetDexPc(), 2474 nullptr); 2475 CheckEntrypointTypes<kQuickF2l, int64_t, float>(); 2476 break; 2477 2478 case Primitive::kPrimDouble: 2479 // Processing a Dex `double-to-long' instruction. 2480 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pD2l), 2481 conversion, 2482 conversion->GetDexPc(), 2483 nullptr); 2484 CheckEntrypointTypes<kQuickD2l, int64_t, double>(); 2485 break; 2486 2487 default: 2488 LOG(FATAL) << "Unexpected type conversion from " << input_type 2489 << " to " << result_type; 2490 } 2491 break; 2492 2493 case Primitive::kPrimChar: 2494 switch (input_type) { 2495 case Primitive::kPrimBoolean: 2496 // Boolean input is a result of code transformations. 2497 case Primitive::kPrimByte: 2498 case Primitive::kPrimShort: 2499 case Primitive::kPrimInt: 2500 // Processing a Dex `Process a Dex `int-to-char'' instruction. 2501 if (in.IsRegister()) { 2502 __ movzxw(out.AsRegister<Register>(), in.AsRegister<Register>()); 2503 } else if (in.IsStackSlot()) { 2504 __ movzxw(out.AsRegister<Register>(), Address(ESP, in.GetStackIndex())); 2505 } else { 2506 DCHECK(in.GetConstant()->IsIntConstant()); 2507 int32_t value = in.GetConstant()->AsIntConstant()->GetValue(); 2508 __ movl(out.AsRegister<Register>(), Immediate(static_cast<uint16_t>(value))); 2509 } 2510 break; 2511 2512 default: 2513 LOG(FATAL) << "Unexpected type conversion from " << input_type 2514 << " to " << result_type; 2515 } 2516 break; 2517 2518 case Primitive::kPrimFloat: 2519 switch (input_type) { 2520 case Primitive::kPrimBoolean: 2521 // Boolean input is a result of code transformations. 2522 case Primitive::kPrimByte: 2523 case Primitive::kPrimShort: 2524 case Primitive::kPrimInt: 2525 case Primitive::kPrimChar: 2526 // Processing a Dex `int-to-float' instruction. 2527 __ cvtsi2ss(out.AsFpuRegister<XmmRegister>(), in.AsRegister<Register>()); 2528 break; 2529 2530 case Primitive::kPrimLong: { 2531 // Processing a Dex `long-to-float' instruction. 2532 size_t adjustment = 0; 2533 2534 // Create stack space for the call to 2535 // InstructionCodeGeneratorX86::PushOntoFPStack and/or X86Assembler::fstps below. 2536 // TODO: enhance register allocator to ask for stack temporaries. 2537 if (!in.IsDoubleStackSlot() || !out.IsStackSlot()) { 2538 adjustment = Primitive::ComponentSize(Primitive::kPrimLong); 2539 __ subl(ESP, Immediate(adjustment)); 2540 } 2541 2542 // Load the value to the FP stack, using temporaries if needed. 2543 PushOntoFPStack(in, 0, adjustment, false, true); 2544 2545 if (out.IsStackSlot()) { 2546 __ fstps(Address(ESP, out.GetStackIndex() + adjustment)); 2547 } else { 2548 __ fstps(Address(ESP, 0)); 2549 Location stack_temp = Location::StackSlot(0); 2550 codegen_->Move32(out, stack_temp); 2551 } 2552 2553 // Remove the temporary stack space we allocated. 2554 if (adjustment != 0) { 2555 __ addl(ESP, Immediate(adjustment)); 2556 } 2557 break; 2558 } 2559 2560 case Primitive::kPrimDouble: 2561 // Processing a Dex `double-to-float' instruction. 2562 __ cvtsd2ss(out.AsFpuRegister<XmmRegister>(), in.AsFpuRegister<XmmRegister>()); 2563 break; 2564 2565 default: 2566 LOG(FATAL) << "Unexpected type conversion from " << input_type 2567 << " to " << result_type; 2568 }; 2569 break; 2570 2571 case Primitive::kPrimDouble: 2572 switch (input_type) { 2573 case Primitive::kPrimBoolean: 2574 // Boolean input is a result of code transformations. 2575 case Primitive::kPrimByte: 2576 case Primitive::kPrimShort: 2577 case Primitive::kPrimInt: 2578 case Primitive::kPrimChar: 2579 // Processing a Dex `int-to-double' instruction. 2580 __ cvtsi2sd(out.AsFpuRegister<XmmRegister>(), in.AsRegister<Register>()); 2581 break; 2582 2583 case Primitive::kPrimLong: { 2584 // Processing a Dex `long-to-double' instruction. 2585 size_t adjustment = 0; 2586 2587 // Create stack space for the call to 2588 // InstructionCodeGeneratorX86::PushOntoFPStack and/or X86Assembler::fstpl below. 2589 // TODO: enhance register allocator to ask for stack temporaries. 2590 if (!in.IsDoubleStackSlot() || !out.IsDoubleStackSlot()) { 2591 adjustment = Primitive::ComponentSize(Primitive::kPrimLong); 2592 __ subl(ESP, Immediate(adjustment)); 2593 } 2594 2595 // Load the value to the FP stack, using temporaries if needed. 2596 PushOntoFPStack(in, 0, adjustment, false, true); 2597 2598 if (out.IsDoubleStackSlot()) { 2599 __ fstpl(Address(ESP, out.GetStackIndex() + adjustment)); 2600 } else { 2601 __ fstpl(Address(ESP, 0)); 2602 Location stack_temp = Location::DoubleStackSlot(0); 2603 codegen_->Move64(out, stack_temp); 2604 } 2605 2606 // Remove the temporary stack space we allocated. 2607 if (adjustment != 0) { 2608 __ addl(ESP, Immediate(adjustment)); 2609 } 2610 break; 2611 } 2612 2613 case Primitive::kPrimFloat: 2614 // Processing a Dex `float-to-double' instruction. 2615 __ cvtss2sd(out.AsFpuRegister<XmmRegister>(), in.AsFpuRegister<XmmRegister>()); 2616 break; 2617 2618 default: 2619 LOG(FATAL) << "Unexpected type conversion from " << input_type 2620 << " to " << result_type; 2621 }; 2622 break; 2623 2624 default: 2625 LOG(FATAL) << "Unexpected type conversion from " << input_type 2626 << " to " << result_type; 2627 } 2628} 2629 2630void LocationsBuilderX86::VisitAdd(HAdd* add) { 2631 LocationSummary* locations = 2632 new (GetGraph()->GetArena()) LocationSummary(add, LocationSummary::kNoCall); 2633 switch (add->GetResultType()) { 2634 case Primitive::kPrimInt: { 2635 locations->SetInAt(0, Location::RequiresRegister()); 2636 locations->SetInAt(1, Location::RegisterOrConstant(add->InputAt(1))); 2637 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2638 break; 2639 } 2640 2641 case Primitive::kPrimLong: { 2642 locations->SetInAt(0, Location::RequiresRegister()); 2643 locations->SetInAt(1, Location::Any()); 2644 locations->SetOut(Location::SameAsFirstInput()); 2645 break; 2646 } 2647 2648 case Primitive::kPrimFloat: 2649 case Primitive::kPrimDouble: { 2650 locations->SetInAt(0, Location::RequiresFpuRegister()); 2651 locations->SetInAt(1, Location::Any()); 2652 locations->SetOut(Location::SameAsFirstInput()); 2653 break; 2654 } 2655 2656 default: 2657 LOG(FATAL) << "Unexpected add type " << add->GetResultType(); 2658 break; 2659 } 2660} 2661 2662void InstructionCodeGeneratorX86::VisitAdd(HAdd* add) { 2663 LocationSummary* locations = add->GetLocations(); 2664 Location first = locations->InAt(0); 2665 Location second = locations->InAt(1); 2666 Location out = locations->Out(); 2667 2668 switch (add->GetResultType()) { 2669 case Primitive::kPrimInt: { 2670 if (second.IsRegister()) { 2671 if (out.AsRegister<Register>() == first.AsRegister<Register>()) { 2672 __ addl(out.AsRegister<Register>(), second.AsRegister<Register>()); 2673 } else if (out.AsRegister<Register>() == second.AsRegister<Register>()) { 2674 __ addl(out.AsRegister<Register>(), first.AsRegister<Register>()); 2675 } else { 2676 __ leal(out.AsRegister<Register>(), Address( 2677 first.AsRegister<Register>(), second.AsRegister<Register>(), TIMES_1, 0)); 2678 } 2679 } else if (second.IsConstant()) { 2680 int32_t value = second.GetConstant()->AsIntConstant()->GetValue(); 2681 if (out.AsRegister<Register>() == first.AsRegister<Register>()) { 2682 __ addl(out.AsRegister<Register>(), Immediate(value)); 2683 } else { 2684 __ leal(out.AsRegister<Register>(), Address(first.AsRegister<Register>(), value)); 2685 } 2686 } else { 2687 DCHECK(first.Equals(locations->Out())); 2688 __ addl(first.AsRegister<Register>(), Address(ESP, second.GetStackIndex())); 2689 } 2690 break; 2691 } 2692 2693 case Primitive::kPrimLong: { 2694 if (second.IsRegisterPair()) { 2695 __ addl(first.AsRegisterPairLow<Register>(), second.AsRegisterPairLow<Register>()); 2696 __ adcl(first.AsRegisterPairHigh<Register>(), second.AsRegisterPairHigh<Register>()); 2697 } else if (second.IsDoubleStackSlot()) { 2698 __ addl(first.AsRegisterPairLow<Register>(), Address(ESP, second.GetStackIndex())); 2699 __ adcl(first.AsRegisterPairHigh<Register>(), 2700 Address(ESP, second.GetHighStackIndex(kX86WordSize))); 2701 } else { 2702 DCHECK(second.IsConstant()) << second; 2703 int64_t value = second.GetConstant()->AsLongConstant()->GetValue(); 2704 __ addl(first.AsRegisterPairLow<Register>(), Immediate(Low32Bits(value))); 2705 __ adcl(first.AsRegisterPairHigh<Register>(), Immediate(High32Bits(value))); 2706 } 2707 break; 2708 } 2709 2710 case Primitive::kPrimFloat: { 2711 if (second.IsFpuRegister()) { 2712 __ addss(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>()); 2713 } else if (add->InputAt(1)->IsX86LoadFromConstantTable()) { 2714 HX86LoadFromConstantTable* const_area = add->InputAt(1)->AsX86LoadFromConstantTable(); 2715 DCHECK(!const_area->NeedsMaterialization()); 2716 __ addss(first.AsFpuRegister<XmmRegister>(), 2717 codegen_->LiteralFloatAddress( 2718 const_area->GetConstant()->AsFloatConstant()->GetValue(), 2719 const_area->GetLocations()->InAt(0).AsRegister<Register>())); 2720 } else { 2721 DCHECK(second.IsStackSlot()); 2722 __ addss(first.AsFpuRegister<XmmRegister>(), Address(ESP, second.GetStackIndex())); 2723 } 2724 break; 2725 } 2726 2727 case Primitive::kPrimDouble: { 2728 if (second.IsFpuRegister()) { 2729 __ addsd(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>()); 2730 } else if (add->InputAt(1)->IsX86LoadFromConstantTable()) { 2731 HX86LoadFromConstantTable* const_area = add->InputAt(1)->AsX86LoadFromConstantTable(); 2732 DCHECK(!const_area->NeedsMaterialization()); 2733 __ addsd(first.AsFpuRegister<XmmRegister>(), 2734 codegen_->LiteralDoubleAddress( 2735 const_area->GetConstant()->AsDoubleConstant()->GetValue(), 2736 const_area->GetLocations()->InAt(0).AsRegister<Register>())); 2737 } else { 2738 DCHECK(second.IsDoubleStackSlot()); 2739 __ addsd(first.AsFpuRegister<XmmRegister>(), Address(ESP, second.GetStackIndex())); 2740 } 2741 break; 2742 } 2743 2744 default: 2745 LOG(FATAL) << "Unexpected add type " << add->GetResultType(); 2746 } 2747} 2748 2749void LocationsBuilderX86::VisitSub(HSub* sub) { 2750 LocationSummary* locations = 2751 new (GetGraph()->GetArena()) LocationSummary(sub, LocationSummary::kNoCall); 2752 switch (sub->GetResultType()) { 2753 case Primitive::kPrimInt: 2754 case Primitive::kPrimLong: { 2755 locations->SetInAt(0, Location::RequiresRegister()); 2756 locations->SetInAt(1, Location::Any()); 2757 locations->SetOut(Location::SameAsFirstInput()); 2758 break; 2759 } 2760 case Primitive::kPrimFloat: 2761 case Primitive::kPrimDouble: { 2762 locations->SetInAt(0, Location::RequiresFpuRegister()); 2763 locations->SetInAt(1, Location::Any()); 2764 locations->SetOut(Location::SameAsFirstInput()); 2765 break; 2766 } 2767 2768 default: 2769 LOG(FATAL) << "Unexpected sub type " << sub->GetResultType(); 2770 } 2771} 2772 2773void InstructionCodeGeneratorX86::VisitSub(HSub* sub) { 2774 LocationSummary* locations = sub->GetLocations(); 2775 Location first = locations->InAt(0); 2776 Location second = locations->InAt(1); 2777 DCHECK(first.Equals(locations->Out())); 2778 switch (sub->GetResultType()) { 2779 case Primitive::kPrimInt: { 2780 if (second.IsRegister()) { 2781 __ subl(first.AsRegister<Register>(), second.AsRegister<Register>()); 2782 } else if (second.IsConstant()) { 2783 __ subl(first.AsRegister<Register>(), 2784 Immediate(second.GetConstant()->AsIntConstant()->GetValue())); 2785 } else { 2786 __ subl(first.AsRegister<Register>(), Address(ESP, second.GetStackIndex())); 2787 } 2788 break; 2789 } 2790 2791 case Primitive::kPrimLong: { 2792 if (second.IsRegisterPair()) { 2793 __ subl(first.AsRegisterPairLow<Register>(), second.AsRegisterPairLow<Register>()); 2794 __ sbbl(first.AsRegisterPairHigh<Register>(), second.AsRegisterPairHigh<Register>()); 2795 } else if (second.IsDoubleStackSlot()) { 2796 __ subl(first.AsRegisterPairLow<Register>(), Address(ESP, second.GetStackIndex())); 2797 __ sbbl(first.AsRegisterPairHigh<Register>(), 2798 Address(ESP, second.GetHighStackIndex(kX86WordSize))); 2799 } else { 2800 DCHECK(second.IsConstant()) << second; 2801 int64_t value = second.GetConstant()->AsLongConstant()->GetValue(); 2802 __ subl(first.AsRegisterPairLow<Register>(), Immediate(Low32Bits(value))); 2803 __ sbbl(first.AsRegisterPairHigh<Register>(), Immediate(High32Bits(value))); 2804 } 2805 break; 2806 } 2807 2808 case Primitive::kPrimFloat: { 2809 if (second.IsFpuRegister()) { 2810 __ subss(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>()); 2811 } else if (sub->InputAt(1)->IsX86LoadFromConstantTable()) { 2812 HX86LoadFromConstantTable* const_area = sub->InputAt(1)->AsX86LoadFromConstantTable(); 2813 DCHECK(!const_area->NeedsMaterialization()); 2814 __ subss(first.AsFpuRegister<XmmRegister>(), 2815 codegen_->LiteralFloatAddress( 2816 const_area->GetConstant()->AsFloatConstant()->GetValue(), 2817 const_area->GetLocations()->InAt(0).AsRegister<Register>())); 2818 } else { 2819 DCHECK(second.IsStackSlot()); 2820 __ subss(first.AsFpuRegister<XmmRegister>(), Address(ESP, second.GetStackIndex())); 2821 } 2822 break; 2823 } 2824 2825 case Primitive::kPrimDouble: { 2826 if (second.IsFpuRegister()) { 2827 __ subsd(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>()); 2828 } else if (sub->InputAt(1)->IsX86LoadFromConstantTable()) { 2829 HX86LoadFromConstantTable* const_area = sub->InputAt(1)->AsX86LoadFromConstantTable(); 2830 DCHECK(!const_area->NeedsMaterialization()); 2831 __ subsd(first.AsFpuRegister<XmmRegister>(), 2832 codegen_->LiteralDoubleAddress( 2833 const_area->GetConstant()->AsDoubleConstant()->GetValue(), 2834 const_area->GetLocations()->InAt(0).AsRegister<Register>())); 2835 } else { 2836 DCHECK(second.IsDoubleStackSlot()); 2837 __ subsd(first.AsFpuRegister<XmmRegister>(), Address(ESP, second.GetStackIndex())); 2838 } 2839 break; 2840 } 2841 2842 default: 2843 LOG(FATAL) << "Unexpected sub type " << sub->GetResultType(); 2844 } 2845} 2846 2847void LocationsBuilderX86::VisitMul(HMul* mul) { 2848 LocationSummary* locations = 2849 new (GetGraph()->GetArena()) LocationSummary(mul, LocationSummary::kNoCall); 2850 switch (mul->GetResultType()) { 2851 case Primitive::kPrimInt: 2852 locations->SetInAt(0, Location::RequiresRegister()); 2853 locations->SetInAt(1, Location::Any()); 2854 if (mul->InputAt(1)->IsIntConstant()) { 2855 // Can use 3 operand multiply. 2856 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2857 } else { 2858 locations->SetOut(Location::SameAsFirstInput()); 2859 } 2860 break; 2861 case Primitive::kPrimLong: { 2862 locations->SetInAt(0, Location::RequiresRegister()); 2863 locations->SetInAt(1, Location::Any()); 2864 locations->SetOut(Location::SameAsFirstInput()); 2865 // Needed for imul on 32bits with 64bits output. 2866 locations->AddTemp(Location::RegisterLocation(EAX)); 2867 locations->AddTemp(Location::RegisterLocation(EDX)); 2868 break; 2869 } 2870 case Primitive::kPrimFloat: 2871 case Primitive::kPrimDouble: { 2872 locations->SetInAt(0, Location::RequiresFpuRegister()); 2873 locations->SetInAt(1, Location::Any()); 2874 locations->SetOut(Location::SameAsFirstInput()); 2875 break; 2876 } 2877 2878 default: 2879 LOG(FATAL) << "Unexpected mul type " << mul->GetResultType(); 2880 } 2881} 2882 2883void InstructionCodeGeneratorX86::VisitMul(HMul* mul) { 2884 LocationSummary* locations = mul->GetLocations(); 2885 Location first = locations->InAt(0); 2886 Location second = locations->InAt(1); 2887 Location out = locations->Out(); 2888 2889 switch (mul->GetResultType()) { 2890 case Primitive::kPrimInt: 2891 // The constant may have ended up in a register, so test explicitly to avoid 2892 // problems where the output may not be the same as the first operand. 2893 if (mul->InputAt(1)->IsIntConstant()) { 2894 Immediate imm(mul->InputAt(1)->AsIntConstant()->GetValue()); 2895 __ imull(out.AsRegister<Register>(), first.AsRegister<Register>(), imm); 2896 } else if (second.IsRegister()) { 2897 DCHECK(first.Equals(out)); 2898 __ imull(first.AsRegister<Register>(), second.AsRegister<Register>()); 2899 } else { 2900 DCHECK(second.IsStackSlot()); 2901 DCHECK(first.Equals(out)); 2902 __ imull(first.AsRegister<Register>(), Address(ESP, second.GetStackIndex())); 2903 } 2904 break; 2905 2906 case Primitive::kPrimLong: { 2907 Register in1_hi = first.AsRegisterPairHigh<Register>(); 2908 Register in1_lo = first.AsRegisterPairLow<Register>(); 2909 Register eax = locations->GetTemp(0).AsRegister<Register>(); 2910 Register edx = locations->GetTemp(1).AsRegister<Register>(); 2911 2912 DCHECK_EQ(EAX, eax); 2913 DCHECK_EQ(EDX, edx); 2914 2915 // input: in1 - 64 bits, in2 - 64 bits. 2916 // output: in1 2917 // formula: in1.hi : in1.lo = (in1.lo * in2.hi + in1.hi * in2.lo)* 2^32 + in1.lo * in2.lo 2918 // parts: in1.hi = in1.lo * in2.hi + in1.hi * in2.lo + (in1.lo * in2.lo)[63:32] 2919 // parts: in1.lo = (in1.lo * in2.lo)[31:0] 2920 if (second.IsConstant()) { 2921 DCHECK(second.GetConstant()->IsLongConstant()); 2922 2923 int64_t value = second.GetConstant()->AsLongConstant()->GetValue(); 2924 int32_t low_value = Low32Bits(value); 2925 int32_t high_value = High32Bits(value); 2926 Immediate low(low_value); 2927 Immediate high(high_value); 2928 2929 __ movl(eax, high); 2930 // eax <- in1.lo * in2.hi 2931 __ imull(eax, in1_lo); 2932 // in1.hi <- in1.hi * in2.lo 2933 __ imull(in1_hi, low); 2934 // in1.hi <- in1.lo * in2.hi + in1.hi * in2.lo 2935 __ addl(in1_hi, eax); 2936 // move in2_lo to eax to prepare for double precision 2937 __ movl(eax, low); 2938 // edx:eax <- in1.lo * in2.lo 2939 __ mull(in1_lo); 2940 // in1.hi <- in2.hi * in1.lo + in2.lo * in1.hi + (in1.lo * in2.lo)[63:32] 2941 __ addl(in1_hi, edx); 2942 // in1.lo <- (in1.lo * in2.lo)[31:0]; 2943 __ movl(in1_lo, eax); 2944 } else if (second.IsRegisterPair()) { 2945 Register in2_hi = second.AsRegisterPairHigh<Register>(); 2946 Register in2_lo = second.AsRegisterPairLow<Register>(); 2947 2948 __ movl(eax, in2_hi); 2949 // eax <- in1.lo * in2.hi 2950 __ imull(eax, in1_lo); 2951 // in1.hi <- in1.hi * in2.lo 2952 __ imull(in1_hi, in2_lo); 2953 // in1.hi <- in1.lo * in2.hi + in1.hi * in2.lo 2954 __ addl(in1_hi, eax); 2955 // move in1_lo to eax to prepare for double precision 2956 __ movl(eax, in1_lo); 2957 // edx:eax <- in1.lo * in2.lo 2958 __ mull(in2_lo); 2959 // in1.hi <- in2.hi * in1.lo + in2.lo * in1.hi + (in1.lo * in2.lo)[63:32] 2960 __ addl(in1_hi, edx); 2961 // in1.lo <- (in1.lo * in2.lo)[31:0]; 2962 __ movl(in1_lo, eax); 2963 } else { 2964 DCHECK(second.IsDoubleStackSlot()) << second; 2965 Address in2_hi(ESP, second.GetHighStackIndex(kX86WordSize)); 2966 Address in2_lo(ESP, second.GetStackIndex()); 2967 2968 __ movl(eax, in2_hi); 2969 // eax <- in1.lo * in2.hi 2970 __ imull(eax, in1_lo); 2971 // in1.hi <- in1.hi * in2.lo 2972 __ imull(in1_hi, in2_lo); 2973 // in1.hi <- in1.lo * in2.hi + in1.hi * in2.lo 2974 __ addl(in1_hi, eax); 2975 // move in1_lo to eax to prepare for double precision 2976 __ movl(eax, in1_lo); 2977 // edx:eax <- in1.lo * in2.lo 2978 __ mull(in2_lo); 2979 // in1.hi <- in2.hi * in1.lo + in2.lo * in1.hi + (in1.lo * in2.lo)[63:32] 2980 __ addl(in1_hi, edx); 2981 // in1.lo <- (in1.lo * in2.lo)[31:0]; 2982 __ movl(in1_lo, eax); 2983 } 2984 2985 break; 2986 } 2987 2988 case Primitive::kPrimFloat: { 2989 DCHECK(first.Equals(locations->Out())); 2990 if (second.IsFpuRegister()) { 2991 __ mulss(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>()); 2992 } else if (mul->InputAt(1)->IsX86LoadFromConstantTable()) { 2993 HX86LoadFromConstantTable* const_area = mul->InputAt(1)->AsX86LoadFromConstantTable(); 2994 DCHECK(!const_area->NeedsMaterialization()); 2995 __ mulss(first.AsFpuRegister<XmmRegister>(), 2996 codegen_->LiteralFloatAddress( 2997 const_area->GetConstant()->AsFloatConstant()->GetValue(), 2998 const_area->GetLocations()->InAt(0).AsRegister<Register>())); 2999 } else { 3000 DCHECK(second.IsStackSlot()); 3001 __ mulss(first.AsFpuRegister<XmmRegister>(), Address(ESP, second.GetStackIndex())); 3002 } 3003 break; 3004 } 3005 3006 case Primitive::kPrimDouble: { 3007 DCHECK(first.Equals(locations->Out())); 3008 if (second.IsFpuRegister()) { 3009 __ mulsd(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>()); 3010 } else if (mul->InputAt(1)->IsX86LoadFromConstantTable()) { 3011 HX86LoadFromConstantTable* const_area = mul->InputAt(1)->AsX86LoadFromConstantTable(); 3012 DCHECK(!const_area->NeedsMaterialization()); 3013 __ mulsd(first.AsFpuRegister<XmmRegister>(), 3014 codegen_->LiteralDoubleAddress( 3015 const_area->GetConstant()->AsDoubleConstant()->GetValue(), 3016 const_area->GetLocations()->InAt(0).AsRegister<Register>())); 3017 } else { 3018 DCHECK(second.IsDoubleStackSlot()); 3019 __ mulsd(first.AsFpuRegister<XmmRegister>(), Address(ESP, second.GetStackIndex())); 3020 } 3021 break; 3022 } 3023 3024 default: 3025 LOG(FATAL) << "Unexpected mul type " << mul->GetResultType(); 3026 } 3027} 3028 3029void InstructionCodeGeneratorX86::PushOntoFPStack(Location source, 3030 uint32_t temp_offset, 3031 uint32_t stack_adjustment, 3032 bool is_fp, 3033 bool is_wide) { 3034 if (source.IsStackSlot()) { 3035 DCHECK(!is_wide); 3036 if (is_fp) { 3037 __ flds(Address(ESP, source.GetStackIndex() + stack_adjustment)); 3038 } else { 3039 __ filds(Address(ESP, source.GetStackIndex() + stack_adjustment)); 3040 } 3041 } else if (source.IsDoubleStackSlot()) { 3042 DCHECK(is_wide); 3043 if (is_fp) { 3044 __ fldl(Address(ESP, source.GetStackIndex() + stack_adjustment)); 3045 } else { 3046 __ fildl(Address(ESP, source.GetStackIndex() + stack_adjustment)); 3047 } 3048 } else { 3049 // Write the value to the temporary location on the stack and load to FP stack. 3050 if (!is_wide) { 3051 Location stack_temp = Location::StackSlot(temp_offset); 3052 codegen_->Move32(stack_temp, source); 3053 if (is_fp) { 3054 __ flds(Address(ESP, temp_offset)); 3055 } else { 3056 __ filds(Address(ESP, temp_offset)); 3057 } 3058 } else { 3059 Location stack_temp = Location::DoubleStackSlot(temp_offset); 3060 codegen_->Move64(stack_temp, source); 3061 if (is_fp) { 3062 __ fldl(Address(ESP, temp_offset)); 3063 } else { 3064 __ fildl(Address(ESP, temp_offset)); 3065 } 3066 } 3067 } 3068} 3069 3070void InstructionCodeGeneratorX86::GenerateRemFP(HRem *rem) { 3071 Primitive::Type type = rem->GetResultType(); 3072 bool is_float = type == Primitive::kPrimFloat; 3073 size_t elem_size = Primitive::ComponentSize(type); 3074 LocationSummary* locations = rem->GetLocations(); 3075 Location first = locations->InAt(0); 3076 Location second = locations->InAt(1); 3077 Location out = locations->Out(); 3078 3079 // Create stack space for 2 elements. 3080 // TODO: enhance register allocator to ask for stack temporaries. 3081 __ subl(ESP, Immediate(2 * elem_size)); 3082 3083 // Load the values to the FP stack in reverse order, using temporaries if needed. 3084 const bool is_wide = !is_float; 3085 PushOntoFPStack(second, elem_size, 2 * elem_size, /* is_fp */ true, is_wide); 3086 PushOntoFPStack(first, 0, 2 * elem_size, /* is_fp */ true, is_wide); 3087 3088 // Loop doing FPREM until we stabilize. 3089 NearLabel retry; 3090 __ Bind(&retry); 3091 __ fprem(); 3092 3093 // Move FP status to AX. 3094 __ fstsw(); 3095 3096 // And see if the argument reduction is complete. This is signaled by the 3097 // C2 FPU flag bit set to 0. 3098 __ andl(EAX, Immediate(kC2ConditionMask)); 3099 __ j(kNotEqual, &retry); 3100 3101 // We have settled on the final value. Retrieve it into an XMM register. 3102 // Store FP top of stack to real stack. 3103 if (is_float) { 3104 __ fsts(Address(ESP, 0)); 3105 } else { 3106 __ fstl(Address(ESP, 0)); 3107 } 3108 3109 // Pop the 2 items from the FP stack. 3110 __ fucompp(); 3111 3112 // Load the value from the stack into an XMM register. 3113 DCHECK(out.IsFpuRegister()) << out; 3114 if (is_float) { 3115 __ movss(out.AsFpuRegister<XmmRegister>(), Address(ESP, 0)); 3116 } else { 3117 __ movsd(out.AsFpuRegister<XmmRegister>(), Address(ESP, 0)); 3118 } 3119 3120 // And remove the temporary stack space we allocated. 3121 __ addl(ESP, Immediate(2 * elem_size)); 3122} 3123 3124 3125void InstructionCodeGeneratorX86::DivRemOneOrMinusOne(HBinaryOperation* instruction) { 3126 DCHECK(instruction->IsDiv() || instruction->IsRem()); 3127 3128 LocationSummary* locations = instruction->GetLocations(); 3129 DCHECK(locations->InAt(1).IsConstant()); 3130 DCHECK(locations->InAt(1).GetConstant()->IsIntConstant()); 3131 3132 Register out_register = locations->Out().AsRegister<Register>(); 3133 Register input_register = locations->InAt(0).AsRegister<Register>(); 3134 int32_t imm = locations->InAt(1).GetConstant()->AsIntConstant()->GetValue(); 3135 3136 DCHECK(imm == 1 || imm == -1); 3137 3138 if (instruction->IsRem()) { 3139 __ xorl(out_register, out_register); 3140 } else { 3141 __ movl(out_register, input_register); 3142 if (imm == -1) { 3143 __ negl(out_register); 3144 } 3145 } 3146} 3147 3148 3149void InstructionCodeGeneratorX86::DivByPowerOfTwo(HDiv* instruction) { 3150 LocationSummary* locations = instruction->GetLocations(); 3151 3152 Register out_register = locations->Out().AsRegister<Register>(); 3153 Register input_register = locations->InAt(0).AsRegister<Register>(); 3154 int32_t imm = locations->InAt(1).GetConstant()->AsIntConstant()->GetValue(); 3155 3156 DCHECK(IsPowerOfTwo(std::abs(imm))); 3157 Register num = locations->GetTemp(0).AsRegister<Register>(); 3158 3159 __ leal(num, Address(input_register, std::abs(imm) - 1)); 3160 __ testl(input_register, input_register); 3161 __ cmovl(kGreaterEqual, num, input_register); 3162 int shift = CTZ(imm); 3163 __ sarl(num, Immediate(shift)); 3164 3165 if (imm < 0) { 3166 __ negl(num); 3167 } 3168 3169 __ movl(out_register, num); 3170} 3171 3172void InstructionCodeGeneratorX86::GenerateDivRemWithAnyConstant(HBinaryOperation* instruction) { 3173 DCHECK(instruction->IsDiv() || instruction->IsRem()); 3174 3175 LocationSummary* locations = instruction->GetLocations(); 3176 int imm = locations->InAt(1).GetConstant()->AsIntConstant()->GetValue(); 3177 3178 Register eax = locations->InAt(0).AsRegister<Register>(); 3179 Register out = locations->Out().AsRegister<Register>(); 3180 Register num; 3181 Register edx; 3182 3183 if (instruction->IsDiv()) { 3184 edx = locations->GetTemp(0).AsRegister<Register>(); 3185 num = locations->GetTemp(1).AsRegister<Register>(); 3186 } else { 3187 edx = locations->Out().AsRegister<Register>(); 3188 num = locations->GetTemp(0).AsRegister<Register>(); 3189 } 3190 3191 DCHECK_EQ(EAX, eax); 3192 DCHECK_EQ(EDX, edx); 3193 if (instruction->IsDiv()) { 3194 DCHECK_EQ(EAX, out); 3195 } else { 3196 DCHECK_EQ(EDX, out); 3197 } 3198 3199 int64_t magic; 3200 int shift; 3201 CalculateMagicAndShiftForDivRem(imm, false /* is_long */, &magic, &shift); 3202 3203 NearLabel ndiv; 3204 NearLabel end; 3205 // If numerator is 0, the result is 0, no computation needed. 3206 __ testl(eax, eax); 3207 __ j(kNotEqual, &ndiv); 3208 3209 __ xorl(out, out); 3210 __ jmp(&end); 3211 3212 __ Bind(&ndiv); 3213 3214 // Save the numerator. 3215 __ movl(num, eax); 3216 3217 // EAX = magic 3218 __ movl(eax, Immediate(magic)); 3219 3220 // EDX:EAX = magic * numerator 3221 __ imull(num); 3222 3223 if (imm > 0 && magic < 0) { 3224 // EDX += num 3225 __ addl(edx, num); 3226 } else if (imm < 0 && magic > 0) { 3227 __ subl(edx, num); 3228 } 3229 3230 // Shift if needed. 3231 if (shift != 0) { 3232 __ sarl(edx, Immediate(shift)); 3233 } 3234 3235 // EDX += 1 if EDX < 0 3236 __ movl(eax, edx); 3237 __ shrl(edx, Immediate(31)); 3238 __ addl(edx, eax); 3239 3240 if (instruction->IsRem()) { 3241 __ movl(eax, num); 3242 __ imull(edx, Immediate(imm)); 3243 __ subl(eax, edx); 3244 __ movl(edx, eax); 3245 } else { 3246 __ movl(eax, edx); 3247 } 3248 __ Bind(&end); 3249} 3250 3251void InstructionCodeGeneratorX86::GenerateDivRemIntegral(HBinaryOperation* instruction) { 3252 DCHECK(instruction->IsDiv() || instruction->IsRem()); 3253 3254 LocationSummary* locations = instruction->GetLocations(); 3255 Location out = locations->Out(); 3256 Location first = locations->InAt(0); 3257 Location second = locations->InAt(1); 3258 bool is_div = instruction->IsDiv(); 3259 3260 switch (instruction->GetResultType()) { 3261 case Primitive::kPrimInt: { 3262 DCHECK_EQ(EAX, first.AsRegister<Register>()); 3263 DCHECK_EQ(is_div ? EAX : EDX, out.AsRegister<Register>()); 3264 3265 if (second.IsConstant()) { 3266 int32_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 3267 3268 if (imm == 0) { 3269 // Do not generate anything for 0. DivZeroCheck would forbid any generated code. 3270 } else if (imm == 1 || imm == -1) { 3271 DivRemOneOrMinusOne(instruction); 3272 } else if (is_div && IsPowerOfTwo(std::abs(imm))) { 3273 DivByPowerOfTwo(instruction->AsDiv()); 3274 } else { 3275 DCHECK(imm <= -2 || imm >= 2); 3276 GenerateDivRemWithAnyConstant(instruction); 3277 } 3278 } else { 3279 SlowPathCode* slow_path = 3280 new (GetGraph()->GetArena()) DivRemMinusOneSlowPathX86(out.AsRegister<Register>(), 3281 is_div); 3282 codegen_->AddSlowPath(slow_path); 3283 3284 Register second_reg = second.AsRegister<Register>(); 3285 // 0x80000000/-1 triggers an arithmetic exception! 3286 // Dividing by -1 is actually negation and -0x800000000 = 0x80000000 so 3287 // it's safe to just use negl instead of more complex comparisons. 3288 3289 __ cmpl(second_reg, Immediate(-1)); 3290 __ j(kEqual, slow_path->GetEntryLabel()); 3291 3292 // edx:eax <- sign-extended of eax 3293 __ cdq(); 3294 // eax = quotient, edx = remainder 3295 __ idivl(second_reg); 3296 __ Bind(slow_path->GetExitLabel()); 3297 } 3298 break; 3299 } 3300 3301 case Primitive::kPrimLong: { 3302 InvokeRuntimeCallingConvention calling_convention; 3303 DCHECK_EQ(calling_convention.GetRegisterAt(0), first.AsRegisterPairLow<Register>()); 3304 DCHECK_EQ(calling_convention.GetRegisterAt(1), first.AsRegisterPairHigh<Register>()); 3305 DCHECK_EQ(calling_convention.GetRegisterAt(2), second.AsRegisterPairLow<Register>()); 3306 DCHECK_EQ(calling_convention.GetRegisterAt(3), second.AsRegisterPairHigh<Register>()); 3307 DCHECK_EQ(EAX, out.AsRegisterPairLow<Register>()); 3308 DCHECK_EQ(EDX, out.AsRegisterPairHigh<Register>()); 3309 3310 if (is_div) { 3311 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pLdiv), 3312 instruction, 3313 instruction->GetDexPc(), 3314 nullptr); 3315 CheckEntrypointTypes<kQuickLdiv, int64_t, int64_t, int64_t>(); 3316 } else { 3317 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pLmod), 3318 instruction, 3319 instruction->GetDexPc(), 3320 nullptr); 3321 CheckEntrypointTypes<kQuickLmod, int64_t, int64_t, int64_t>(); 3322 } 3323 break; 3324 } 3325 3326 default: 3327 LOG(FATAL) << "Unexpected type for GenerateDivRemIntegral " << instruction->GetResultType(); 3328 } 3329} 3330 3331void LocationsBuilderX86::VisitDiv(HDiv* div) { 3332 LocationSummary::CallKind call_kind = (div->GetResultType() == Primitive::kPrimLong) 3333 ? LocationSummary::kCall 3334 : LocationSummary::kNoCall; 3335 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(div, call_kind); 3336 3337 switch (div->GetResultType()) { 3338 case Primitive::kPrimInt: { 3339 locations->SetInAt(0, Location::RegisterLocation(EAX)); 3340 locations->SetInAt(1, Location::RegisterOrConstant(div->InputAt(1))); 3341 locations->SetOut(Location::SameAsFirstInput()); 3342 // Intel uses edx:eax as the dividend. 3343 locations->AddTemp(Location::RegisterLocation(EDX)); 3344 // We need to save the numerator while we tweak eax and edx. As we are using imul in a way 3345 // which enforces results to be in EAX and EDX, things are simpler if we use EAX also as 3346 // output and request another temp. 3347 if (div->InputAt(1)->IsIntConstant()) { 3348 locations->AddTemp(Location::RequiresRegister()); 3349 } 3350 break; 3351 } 3352 case Primitive::kPrimLong: { 3353 InvokeRuntimeCallingConvention calling_convention; 3354 locations->SetInAt(0, Location::RegisterPairLocation( 3355 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1))); 3356 locations->SetInAt(1, Location::RegisterPairLocation( 3357 calling_convention.GetRegisterAt(2), calling_convention.GetRegisterAt(3))); 3358 // Runtime helper puts the result in EAX, EDX. 3359 locations->SetOut(Location::RegisterPairLocation(EAX, EDX)); 3360 break; 3361 } 3362 case Primitive::kPrimFloat: 3363 case Primitive::kPrimDouble: { 3364 locations->SetInAt(0, Location::RequiresFpuRegister()); 3365 locations->SetInAt(1, Location::Any()); 3366 locations->SetOut(Location::SameAsFirstInput()); 3367 break; 3368 } 3369 3370 default: 3371 LOG(FATAL) << "Unexpected div type " << div->GetResultType(); 3372 } 3373} 3374 3375void InstructionCodeGeneratorX86::VisitDiv(HDiv* div) { 3376 LocationSummary* locations = div->GetLocations(); 3377 Location first = locations->InAt(0); 3378 Location second = locations->InAt(1); 3379 3380 switch (div->GetResultType()) { 3381 case Primitive::kPrimInt: 3382 case Primitive::kPrimLong: { 3383 GenerateDivRemIntegral(div); 3384 break; 3385 } 3386 3387 case Primitive::kPrimFloat: { 3388 if (second.IsFpuRegister()) { 3389 __ divss(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>()); 3390 } else if (div->InputAt(1)->IsX86LoadFromConstantTable()) { 3391 HX86LoadFromConstantTable* const_area = div->InputAt(1)->AsX86LoadFromConstantTable(); 3392 DCHECK(!const_area->NeedsMaterialization()); 3393 __ divss(first.AsFpuRegister<XmmRegister>(), 3394 codegen_->LiteralFloatAddress( 3395 const_area->GetConstant()->AsFloatConstant()->GetValue(), 3396 const_area->GetLocations()->InAt(0).AsRegister<Register>())); 3397 } else { 3398 DCHECK(second.IsStackSlot()); 3399 __ divss(first.AsFpuRegister<XmmRegister>(), Address(ESP, second.GetStackIndex())); 3400 } 3401 break; 3402 } 3403 3404 case Primitive::kPrimDouble: { 3405 if (second.IsFpuRegister()) { 3406 __ divsd(first.AsFpuRegister<XmmRegister>(), second.AsFpuRegister<XmmRegister>()); 3407 } else if (div->InputAt(1)->IsX86LoadFromConstantTable()) { 3408 HX86LoadFromConstantTable* const_area = div->InputAt(1)->AsX86LoadFromConstantTable(); 3409 DCHECK(!const_area->NeedsMaterialization()); 3410 __ divsd(first.AsFpuRegister<XmmRegister>(), 3411 codegen_->LiteralDoubleAddress( 3412 const_area->GetConstant()->AsDoubleConstant()->GetValue(), 3413 const_area->GetLocations()->InAt(0).AsRegister<Register>())); 3414 } else { 3415 DCHECK(second.IsDoubleStackSlot()); 3416 __ divsd(first.AsFpuRegister<XmmRegister>(), Address(ESP, second.GetStackIndex())); 3417 } 3418 break; 3419 } 3420 3421 default: 3422 LOG(FATAL) << "Unexpected div type " << div->GetResultType(); 3423 } 3424} 3425 3426void LocationsBuilderX86::VisitRem(HRem* rem) { 3427 Primitive::Type type = rem->GetResultType(); 3428 3429 LocationSummary::CallKind call_kind = (rem->GetResultType() == Primitive::kPrimLong) 3430 ? LocationSummary::kCall 3431 : LocationSummary::kNoCall; 3432 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(rem, call_kind); 3433 3434 switch (type) { 3435 case Primitive::kPrimInt: { 3436 locations->SetInAt(0, Location::RegisterLocation(EAX)); 3437 locations->SetInAt(1, Location::RegisterOrConstant(rem->InputAt(1))); 3438 locations->SetOut(Location::RegisterLocation(EDX)); 3439 // We need to save the numerator while we tweak eax and edx. As we are using imul in a way 3440 // which enforces results to be in EAX and EDX, things are simpler if we use EDX also as 3441 // output and request another temp. 3442 if (rem->InputAt(1)->IsIntConstant()) { 3443 locations->AddTemp(Location::RequiresRegister()); 3444 } 3445 break; 3446 } 3447 case Primitive::kPrimLong: { 3448 InvokeRuntimeCallingConvention calling_convention; 3449 locations->SetInAt(0, Location::RegisterPairLocation( 3450 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1))); 3451 locations->SetInAt(1, Location::RegisterPairLocation( 3452 calling_convention.GetRegisterAt(2), calling_convention.GetRegisterAt(3))); 3453 // Runtime helper puts the result in EAX, EDX. 3454 locations->SetOut(Location::RegisterPairLocation(EAX, EDX)); 3455 break; 3456 } 3457 case Primitive::kPrimDouble: 3458 case Primitive::kPrimFloat: { 3459 locations->SetInAt(0, Location::Any()); 3460 locations->SetInAt(1, Location::Any()); 3461 locations->SetOut(Location::RequiresFpuRegister()); 3462 locations->AddTemp(Location::RegisterLocation(EAX)); 3463 break; 3464 } 3465 3466 default: 3467 LOG(FATAL) << "Unexpected rem type " << type; 3468 } 3469} 3470 3471void InstructionCodeGeneratorX86::VisitRem(HRem* rem) { 3472 Primitive::Type type = rem->GetResultType(); 3473 switch (type) { 3474 case Primitive::kPrimInt: 3475 case Primitive::kPrimLong: { 3476 GenerateDivRemIntegral(rem); 3477 break; 3478 } 3479 case Primitive::kPrimFloat: 3480 case Primitive::kPrimDouble: { 3481 GenerateRemFP(rem); 3482 break; 3483 } 3484 default: 3485 LOG(FATAL) << "Unexpected rem type " << type; 3486 } 3487} 3488 3489void LocationsBuilderX86::VisitDivZeroCheck(HDivZeroCheck* instruction) { 3490 LocationSummary::CallKind call_kind = instruction->CanThrowIntoCatchBlock() 3491 ? LocationSummary::kCallOnSlowPath 3492 : LocationSummary::kNoCall; 3493 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 3494 switch (instruction->GetType()) { 3495 case Primitive::kPrimByte: 3496 case Primitive::kPrimChar: 3497 case Primitive::kPrimShort: 3498 case Primitive::kPrimInt: { 3499 locations->SetInAt(0, Location::Any()); 3500 break; 3501 } 3502 case Primitive::kPrimLong: { 3503 locations->SetInAt(0, Location::RegisterOrConstant(instruction->InputAt(0))); 3504 if (!instruction->IsConstant()) { 3505 locations->AddTemp(Location::RequiresRegister()); 3506 } 3507 break; 3508 } 3509 default: 3510 LOG(FATAL) << "Unexpected type for HDivZeroCheck " << instruction->GetType(); 3511 } 3512 if (instruction->HasUses()) { 3513 locations->SetOut(Location::SameAsFirstInput()); 3514 } 3515} 3516 3517void InstructionCodeGeneratorX86::VisitDivZeroCheck(HDivZeroCheck* instruction) { 3518 SlowPathCode* slow_path = new (GetGraph()->GetArena()) DivZeroCheckSlowPathX86(instruction); 3519 codegen_->AddSlowPath(slow_path); 3520 3521 LocationSummary* locations = instruction->GetLocations(); 3522 Location value = locations->InAt(0); 3523 3524 switch (instruction->GetType()) { 3525 case Primitive::kPrimByte: 3526 case Primitive::kPrimChar: 3527 case Primitive::kPrimShort: 3528 case Primitive::kPrimInt: { 3529 if (value.IsRegister()) { 3530 __ testl(value.AsRegister<Register>(), value.AsRegister<Register>()); 3531 __ j(kEqual, slow_path->GetEntryLabel()); 3532 } else if (value.IsStackSlot()) { 3533 __ cmpl(Address(ESP, value.GetStackIndex()), Immediate(0)); 3534 __ j(kEqual, slow_path->GetEntryLabel()); 3535 } else { 3536 DCHECK(value.IsConstant()) << value; 3537 if (value.GetConstant()->AsIntConstant()->GetValue() == 0) { 3538 __ jmp(slow_path->GetEntryLabel()); 3539 } 3540 } 3541 break; 3542 } 3543 case Primitive::kPrimLong: { 3544 if (value.IsRegisterPair()) { 3545 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3546 __ movl(temp, value.AsRegisterPairLow<Register>()); 3547 __ orl(temp, value.AsRegisterPairHigh<Register>()); 3548 __ j(kEqual, slow_path->GetEntryLabel()); 3549 } else { 3550 DCHECK(value.IsConstant()) << value; 3551 if (value.GetConstant()->AsLongConstant()->GetValue() == 0) { 3552 __ jmp(slow_path->GetEntryLabel()); 3553 } 3554 } 3555 break; 3556 } 3557 default: 3558 LOG(FATAL) << "Unexpected type for HDivZeroCheck" << instruction->GetType(); 3559 } 3560} 3561 3562void LocationsBuilderX86::HandleShift(HBinaryOperation* op) { 3563 DCHECK(op->IsShl() || op->IsShr() || op->IsUShr()); 3564 3565 LocationSummary* locations = 3566 new (GetGraph()->GetArena()) LocationSummary(op, LocationSummary::kNoCall); 3567 3568 switch (op->GetResultType()) { 3569 case Primitive::kPrimInt: 3570 case Primitive::kPrimLong: { 3571 // Can't have Location::Any() and output SameAsFirstInput() 3572 locations->SetInAt(0, Location::RequiresRegister()); 3573 // The shift count needs to be in CL or a constant. 3574 locations->SetInAt(1, Location::ByteRegisterOrConstant(ECX, op->InputAt(1))); 3575 locations->SetOut(Location::SameAsFirstInput()); 3576 break; 3577 } 3578 default: 3579 LOG(FATAL) << "Unexpected op type " << op->GetResultType(); 3580 } 3581} 3582 3583void InstructionCodeGeneratorX86::HandleShift(HBinaryOperation* op) { 3584 DCHECK(op->IsShl() || op->IsShr() || op->IsUShr()); 3585 3586 LocationSummary* locations = op->GetLocations(); 3587 Location first = locations->InAt(0); 3588 Location second = locations->InAt(1); 3589 DCHECK(first.Equals(locations->Out())); 3590 3591 switch (op->GetResultType()) { 3592 case Primitive::kPrimInt: { 3593 DCHECK(first.IsRegister()); 3594 Register first_reg = first.AsRegister<Register>(); 3595 if (second.IsRegister()) { 3596 Register second_reg = second.AsRegister<Register>(); 3597 DCHECK_EQ(ECX, second_reg); 3598 if (op->IsShl()) { 3599 __ shll(first_reg, second_reg); 3600 } else if (op->IsShr()) { 3601 __ sarl(first_reg, second_reg); 3602 } else { 3603 __ shrl(first_reg, second_reg); 3604 } 3605 } else { 3606 int32_t shift = second.GetConstant()->AsIntConstant()->GetValue() & kMaxIntShiftValue; 3607 if (shift == 0) { 3608 return; 3609 } 3610 Immediate imm(shift); 3611 if (op->IsShl()) { 3612 __ shll(first_reg, imm); 3613 } else if (op->IsShr()) { 3614 __ sarl(first_reg, imm); 3615 } else { 3616 __ shrl(first_reg, imm); 3617 } 3618 } 3619 break; 3620 } 3621 case Primitive::kPrimLong: { 3622 if (second.IsRegister()) { 3623 Register second_reg = second.AsRegister<Register>(); 3624 DCHECK_EQ(ECX, second_reg); 3625 if (op->IsShl()) { 3626 GenerateShlLong(first, second_reg); 3627 } else if (op->IsShr()) { 3628 GenerateShrLong(first, second_reg); 3629 } else { 3630 GenerateUShrLong(first, second_reg); 3631 } 3632 } else { 3633 // Shift by a constant. 3634 int shift = second.GetConstant()->AsIntConstant()->GetValue() & kMaxLongShiftValue; 3635 // Nothing to do if the shift is 0, as the input is already the output. 3636 if (shift != 0) { 3637 if (op->IsShl()) { 3638 GenerateShlLong(first, shift); 3639 } else if (op->IsShr()) { 3640 GenerateShrLong(first, shift); 3641 } else { 3642 GenerateUShrLong(first, shift); 3643 } 3644 } 3645 } 3646 break; 3647 } 3648 default: 3649 LOG(FATAL) << "Unexpected op type " << op->GetResultType(); 3650 } 3651} 3652 3653void InstructionCodeGeneratorX86::GenerateShlLong(const Location& loc, int shift) { 3654 Register low = loc.AsRegisterPairLow<Register>(); 3655 Register high = loc.AsRegisterPairHigh<Register>(); 3656 if (shift == 1) { 3657 // This is just an addition. 3658 __ addl(low, low); 3659 __ adcl(high, high); 3660 } else if (shift == 32) { 3661 // Shift by 32 is easy. High gets low, and low gets 0. 3662 codegen_->EmitParallelMoves( 3663 loc.ToLow(), 3664 loc.ToHigh(), 3665 Primitive::kPrimInt, 3666 Location::ConstantLocation(GetGraph()->GetIntConstant(0)), 3667 loc.ToLow(), 3668 Primitive::kPrimInt); 3669 } else if (shift > 32) { 3670 // Low part becomes 0. High part is low part << (shift-32). 3671 __ movl(high, low); 3672 __ shll(high, Immediate(shift - 32)); 3673 __ xorl(low, low); 3674 } else { 3675 // Between 1 and 31. 3676 __ shld(high, low, Immediate(shift)); 3677 __ shll(low, Immediate(shift)); 3678 } 3679} 3680 3681void InstructionCodeGeneratorX86::GenerateShlLong(const Location& loc, Register shifter) { 3682 NearLabel done; 3683 __ shld(loc.AsRegisterPairHigh<Register>(), loc.AsRegisterPairLow<Register>(), shifter); 3684 __ shll(loc.AsRegisterPairLow<Register>(), shifter); 3685 __ testl(shifter, Immediate(32)); 3686 __ j(kEqual, &done); 3687 __ movl(loc.AsRegisterPairHigh<Register>(), loc.AsRegisterPairLow<Register>()); 3688 __ movl(loc.AsRegisterPairLow<Register>(), Immediate(0)); 3689 __ Bind(&done); 3690} 3691 3692void InstructionCodeGeneratorX86::GenerateShrLong(const Location& loc, int shift) { 3693 Register low = loc.AsRegisterPairLow<Register>(); 3694 Register high = loc.AsRegisterPairHigh<Register>(); 3695 if (shift == 32) { 3696 // Need to copy the sign. 3697 DCHECK_NE(low, high); 3698 __ movl(low, high); 3699 __ sarl(high, Immediate(31)); 3700 } else if (shift > 32) { 3701 DCHECK_NE(low, high); 3702 // High part becomes sign. Low part is shifted by shift - 32. 3703 __ movl(low, high); 3704 __ sarl(high, Immediate(31)); 3705 __ sarl(low, Immediate(shift - 32)); 3706 } else { 3707 // Between 1 and 31. 3708 __ shrd(low, high, Immediate(shift)); 3709 __ sarl(high, Immediate(shift)); 3710 } 3711} 3712 3713void InstructionCodeGeneratorX86::GenerateShrLong(const Location& loc, Register shifter) { 3714 NearLabel done; 3715 __ shrd(loc.AsRegisterPairLow<Register>(), loc.AsRegisterPairHigh<Register>(), shifter); 3716 __ sarl(loc.AsRegisterPairHigh<Register>(), shifter); 3717 __ testl(shifter, Immediate(32)); 3718 __ j(kEqual, &done); 3719 __ movl(loc.AsRegisterPairLow<Register>(), loc.AsRegisterPairHigh<Register>()); 3720 __ sarl(loc.AsRegisterPairHigh<Register>(), Immediate(31)); 3721 __ Bind(&done); 3722} 3723 3724void InstructionCodeGeneratorX86::GenerateUShrLong(const Location& loc, int shift) { 3725 Register low = loc.AsRegisterPairLow<Register>(); 3726 Register high = loc.AsRegisterPairHigh<Register>(); 3727 if (shift == 32) { 3728 // Shift by 32 is easy. Low gets high, and high gets 0. 3729 codegen_->EmitParallelMoves( 3730 loc.ToHigh(), 3731 loc.ToLow(), 3732 Primitive::kPrimInt, 3733 Location::ConstantLocation(GetGraph()->GetIntConstant(0)), 3734 loc.ToHigh(), 3735 Primitive::kPrimInt); 3736 } else if (shift > 32) { 3737 // Low part is high >> (shift - 32). High part becomes 0. 3738 __ movl(low, high); 3739 __ shrl(low, Immediate(shift - 32)); 3740 __ xorl(high, high); 3741 } else { 3742 // Between 1 and 31. 3743 __ shrd(low, high, Immediate(shift)); 3744 __ shrl(high, Immediate(shift)); 3745 } 3746} 3747 3748void InstructionCodeGeneratorX86::GenerateUShrLong(const Location& loc, Register shifter) { 3749 NearLabel done; 3750 __ shrd(loc.AsRegisterPairLow<Register>(), loc.AsRegisterPairHigh<Register>(), shifter); 3751 __ shrl(loc.AsRegisterPairHigh<Register>(), shifter); 3752 __ testl(shifter, Immediate(32)); 3753 __ j(kEqual, &done); 3754 __ movl(loc.AsRegisterPairLow<Register>(), loc.AsRegisterPairHigh<Register>()); 3755 __ movl(loc.AsRegisterPairHigh<Register>(), Immediate(0)); 3756 __ Bind(&done); 3757} 3758 3759void LocationsBuilderX86::VisitShl(HShl* shl) { 3760 HandleShift(shl); 3761} 3762 3763void InstructionCodeGeneratorX86::VisitShl(HShl* shl) { 3764 HandleShift(shl); 3765} 3766 3767void LocationsBuilderX86::VisitShr(HShr* shr) { 3768 HandleShift(shr); 3769} 3770 3771void InstructionCodeGeneratorX86::VisitShr(HShr* shr) { 3772 HandleShift(shr); 3773} 3774 3775void LocationsBuilderX86::VisitUShr(HUShr* ushr) { 3776 HandleShift(ushr); 3777} 3778 3779void InstructionCodeGeneratorX86::VisitUShr(HUShr* ushr) { 3780 HandleShift(ushr); 3781} 3782 3783void LocationsBuilderX86::VisitNewInstance(HNewInstance* instruction) { 3784 LocationSummary* locations = 3785 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall); 3786 locations->SetOut(Location::RegisterLocation(EAX)); 3787 InvokeRuntimeCallingConvention calling_convention; 3788 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 3789 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 3790} 3791 3792void InstructionCodeGeneratorX86::VisitNewInstance(HNewInstance* instruction) { 3793 // Note: if heap poisoning is enabled, the entry point takes cares 3794 // of poisoning the reference. 3795 codegen_->InvokeRuntime(instruction->GetEntrypoint(), 3796 instruction, 3797 instruction->GetDexPc(), 3798 nullptr); 3799 CheckEntrypointTypes<kQuickAllocObjectWithAccessCheck, void*, uint32_t, ArtMethod*>(); 3800 DCHECK(!codegen_->IsLeafMethod()); 3801} 3802 3803void LocationsBuilderX86::VisitNewArray(HNewArray* instruction) { 3804 LocationSummary* locations = 3805 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall); 3806 locations->SetOut(Location::RegisterLocation(EAX)); 3807 InvokeRuntimeCallingConvention calling_convention; 3808 locations->AddTemp(Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 3809 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 3810 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(2))); 3811} 3812 3813void InstructionCodeGeneratorX86::VisitNewArray(HNewArray* instruction) { 3814 InvokeRuntimeCallingConvention calling_convention; 3815 __ movl(calling_convention.GetRegisterAt(0), Immediate(instruction->GetTypeIndex())); 3816 // Note: if heap poisoning is enabled, the entry point takes cares 3817 // of poisoning the reference. 3818 codegen_->InvokeRuntime(instruction->GetEntrypoint(), 3819 instruction, 3820 instruction->GetDexPc(), 3821 nullptr); 3822 CheckEntrypointTypes<kQuickAllocArrayWithAccessCheck, void*, uint32_t, int32_t, ArtMethod*>(); 3823 DCHECK(!codegen_->IsLeafMethod()); 3824} 3825 3826void LocationsBuilderX86::VisitParameterValue(HParameterValue* instruction) { 3827 LocationSummary* locations = 3828 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 3829 Location location = parameter_visitor_.GetNextLocation(instruction->GetType()); 3830 if (location.IsStackSlot()) { 3831 location = Location::StackSlot(location.GetStackIndex() + codegen_->GetFrameSize()); 3832 } else if (location.IsDoubleStackSlot()) { 3833 location = Location::DoubleStackSlot(location.GetStackIndex() + codegen_->GetFrameSize()); 3834 } 3835 locations->SetOut(location); 3836} 3837 3838void InstructionCodeGeneratorX86::VisitParameterValue( 3839 HParameterValue* instruction ATTRIBUTE_UNUSED) { 3840} 3841 3842void LocationsBuilderX86::VisitCurrentMethod(HCurrentMethod* instruction) { 3843 LocationSummary* locations = 3844 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 3845 locations->SetOut(Location::RegisterLocation(kMethodRegisterArgument)); 3846} 3847 3848void InstructionCodeGeneratorX86::VisitCurrentMethod(HCurrentMethod* instruction ATTRIBUTE_UNUSED) { 3849} 3850 3851void LocationsBuilderX86::VisitNot(HNot* not_) { 3852 LocationSummary* locations = 3853 new (GetGraph()->GetArena()) LocationSummary(not_, LocationSummary::kNoCall); 3854 locations->SetInAt(0, Location::RequiresRegister()); 3855 locations->SetOut(Location::SameAsFirstInput()); 3856} 3857 3858void InstructionCodeGeneratorX86::VisitNot(HNot* not_) { 3859 LocationSummary* locations = not_->GetLocations(); 3860 Location in = locations->InAt(0); 3861 Location out = locations->Out(); 3862 DCHECK(in.Equals(out)); 3863 switch (not_->GetResultType()) { 3864 case Primitive::kPrimInt: 3865 __ notl(out.AsRegister<Register>()); 3866 break; 3867 3868 case Primitive::kPrimLong: 3869 __ notl(out.AsRegisterPairLow<Register>()); 3870 __ notl(out.AsRegisterPairHigh<Register>()); 3871 break; 3872 3873 default: 3874 LOG(FATAL) << "Unimplemented type for not operation " << not_->GetResultType(); 3875 } 3876} 3877 3878void LocationsBuilderX86::VisitBooleanNot(HBooleanNot* bool_not) { 3879 LocationSummary* locations = 3880 new (GetGraph()->GetArena()) LocationSummary(bool_not, LocationSummary::kNoCall); 3881 locations->SetInAt(0, Location::RequiresRegister()); 3882 locations->SetOut(Location::SameAsFirstInput()); 3883} 3884 3885void InstructionCodeGeneratorX86::VisitBooleanNot(HBooleanNot* bool_not) { 3886 LocationSummary* locations = bool_not->GetLocations(); 3887 Location in = locations->InAt(0); 3888 Location out = locations->Out(); 3889 DCHECK(in.Equals(out)); 3890 __ xorl(out.AsRegister<Register>(), Immediate(1)); 3891} 3892 3893void LocationsBuilderX86::VisitCompare(HCompare* compare) { 3894 LocationSummary* locations = 3895 new (GetGraph()->GetArena()) LocationSummary(compare, LocationSummary::kNoCall); 3896 switch (compare->InputAt(0)->GetType()) { 3897 case Primitive::kPrimLong: { 3898 locations->SetInAt(0, Location::RequiresRegister()); 3899 locations->SetInAt(1, Location::Any()); 3900 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3901 break; 3902 } 3903 case Primitive::kPrimFloat: 3904 case Primitive::kPrimDouble: { 3905 locations->SetInAt(0, Location::RequiresFpuRegister()); 3906 locations->SetInAt(1, Location::RequiresFpuRegister()); 3907 locations->SetOut(Location::RequiresRegister()); 3908 break; 3909 } 3910 default: 3911 LOG(FATAL) << "Unexpected type for compare operation " << compare->InputAt(0)->GetType(); 3912 } 3913} 3914 3915void InstructionCodeGeneratorX86::VisitCompare(HCompare* compare) { 3916 LocationSummary* locations = compare->GetLocations(); 3917 Register out = locations->Out().AsRegister<Register>(); 3918 Location left = locations->InAt(0); 3919 Location right = locations->InAt(1); 3920 3921 NearLabel less, greater, done; 3922 switch (compare->InputAt(0)->GetType()) { 3923 case Primitive::kPrimLong: { 3924 Register left_low = left.AsRegisterPairLow<Register>(); 3925 Register left_high = left.AsRegisterPairHigh<Register>(); 3926 int32_t val_low = 0; 3927 int32_t val_high = 0; 3928 bool right_is_const = false; 3929 3930 if (right.IsConstant()) { 3931 DCHECK(right.GetConstant()->IsLongConstant()); 3932 right_is_const = true; 3933 int64_t val = right.GetConstant()->AsLongConstant()->GetValue(); 3934 val_low = Low32Bits(val); 3935 val_high = High32Bits(val); 3936 } 3937 3938 if (right.IsRegisterPair()) { 3939 __ cmpl(left_high, right.AsRegisterPairHigh<Register>()); 3940 } else if (right.IsDoubleStackSlot()) { 3941 __ cmpl(left_high, Address(ESP, right.GetHighStackIndex(kX86WordSize))); 3942 } else { 3943 DCHECK(right_is_const) << right; 3944 if (val_high == 0) { 3945 __ testl(left_high, left_high); 3946 } else { 3947 __ cmpl(left_high, Immediate(val_high)); 3948 } 3949 } 3950 __ j(kLess, &less); // Signed compare. 3951 __ j(kGreater, &greater); // Signed compare. 3952 if (right.IsRegisterPair()) { 3953 __ cmpl(left_low, right.AsRegisterPairLow<Register>()); 3954 } else if (right.IsDoubleStackSlot()) { 3955 __ cmpl(left_low, Address(ESP, right.GetStackIndex())); 3956 } else { 3957 DCHECK(right_is_const) << right; 3958 if (val_low == 0) { 3959 __ testl(left_low, left_low); 3960 } else { 3961 __ cmpl(left_low, Immediate(val_low)); 3962 } 3963 } 3964 break; 3965 } 3966 case Primitive::kPrimFloat: { 3967 __ ucomiss(left.AsFpuRegister<XmmRegister>(), right.AsFpuRegister<XmmRegister>()); 3968 __ j(kUnordered, compare->IsGtBias() ? &greater : &less); 3969 break; 3970 } 3971 case Primitive::kPrimDouble: { 3972 __ ucomisd(left.AsFpuRegister<XmmRegister>(), right.AsFpuRegister<XmmRegister>()); 3973 __ j(kUnordered, compare->IsGtBias() ? &greater : &less); 3974 break; 3975 } 3976 default: 3977 LOG(FATAL) << "Unexpected type for compare operation " << compare->InputAt(0)->GetType(); 3978 } 3979 __ movl(out, Immediate(0)); 3980 __ j(kEqual, &done); 3981 __ j(kBelow, &less); // kBelow is for CF (unsigned & floats). 3982 3983 __ Bind(&greater); 3984 __ movl(out, Immediate(1)); 3985 __ jmp(&done); 3986 3987 __ Bind(&less); 3988 __ movl(out, Immediate(-1)); 3989 3990 __ Bind(&done); 3991} 3992 3993void LocationsBuilderX86::VisitPhi(HPhi* instruction) { 3994 LocationSummary* locations = 3995 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 3996 for (size_t i = 0, e = instruction->InputCount(); i < e; ++i) { 3997 locations->SetInAt(i, Location::Any()); 3998 } 3999 locations->SetOut(Location::Any()); 4000} 4001 4002void InstructionCodeGeneratorX86::VisitPhi(HPhi* instruction ATTRIBUTE_UNUSED) { 4003 LOG(FATAL) << "Unreachable"; 4004} 4005 4006void InstructionCodeGeneratorX86::GenerateMemoryBarrier(MemBarrierKind kind) { 4007 /* 4008 * According to the JSR-133 Cookbook, for x86 only StoreLoad/AnyAny barriers need memory fence. 4009 * All other barriers (LoadAny, AnyStore, StoreStore) are nops due to the x86 memory model. 4010 * For those cases, all we need to ensure is that there is a scheduling barrier in place. 4011 */ 4012 switch (kind) { 4013 case MemBarrierKind::kAnyAny: { 4014 __ mfence(); 4015 break; 4016 } 4017 case MemBarrierKind::kAnyStore: 4018 case MemBarrierKind::kLoadAny: 4019 case MemBarrierKind::kStoreStore: { 4020 // nop 4021 break; 4022 } 4023 default: 4024 LOG(FATAL) << "Unexpected memory barrier " << kind; 4025 } 4026} 4027 4028HInvokeStaticOrDirect::DispatchInfo CodeGeneratorX86::GetSupportedInvokeStaticOrDirectDispatch( 4029 const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info, 4030 MethodReference target_method ATTRIBUTE_UNUSED) { 4031 switch (desired_dispatch_info.code_ptr_location) { 4032 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup: 4033 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect: 4034 // For direct code, we actually prefer to call via the code pointer from ArtMethod*. 4035 // (Though the direct CALL ptr16:32 is available for consideration). 4036 return HInvokeStaticOrDirect::DispatchInfo { 4037 desired_dispatch_info.method_load_kind, 4038 HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod, 4039 desired_dispatch_info.method_load_data, 4040 0u 4041 }; 4042 default: 4043 return desired_dispatch_info; 4044 } 4045} 4046 4047Register CodeGeneratorX86::GetInvokeStaticOrDirectExtraParameter(HInvokeStaticOrDirect* invoke, 4048 Register temp) { 4049 DCHECK_EQ(invoke->InputCount(), invoke->GetNumberOfArguments() + 1u); 4050 Location location = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 4051 if (!invoke->GetLocations()->Intrinsified()) { 4052 return location.AsRegister<Register>(); 4053 } 4054 // For intrinsics we allow any location, so it may be on the stack. 4055 if (!location.IsRegister()) { 4056 __ movl(temp, Address(ESP, location.GetStackIndex())); 4057 return temp; 4058 } 4059 // For register locations, check if the register was saved. If so, get it from the stack. 4060 // Note: There is a chance that the register was saved but not overwritten, so we could 4061 // save one load. However, since this is just an intrinsic slow path we prefer this 4062 // simple and more robust approach rather that trying to determine if that's the case. 4063 SlowPathCode* slow_path = GetCurrentSlowPath(); 4064 DCHECK(slow_path != nullptr); // For intrinsified invokes the call is emitted on the slow path. 4065 if (slow_path->IsCoreRegisterSaved(location.AsRegister<Register>())) { 4066 int stack_offset = slow_path->GetStackOffsetOfCoreRegister(location.AsRegister<Register>()); 4067 __ movl(temp, Address(ESP, stack_offset)); 4068 return temp; 4069 } 4070 return location.AsRegister<Register>(); 4071} 4072 4073void CodeGeneratorX86::GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) { 4074 Location callee_method = temp; // For all kinds except kRecursive, callee will be in temp. 4075 switch (invoke->GetMethodLoadKind()) { 4076 case HInvokeStaticOrDirect::MethodLoadKind::kStringInit: 4077 // temp = thread->string_init_entrypoint 4078 __ fs()->movl(temp.AsRegister<Register>(), Address::Absolute(invoke->GetStringInitOffset())); 4079 break; 4080 case HInvokeStaticOrDirect::MethodLoadKind::kRecursive: 4081 callee_method = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 4082 break; 4083 case HInvokeStaticOrDirect::MethodLoadKind::kDirectAddress: 4084 __ movl(temp.AsRegister<Register>(), Immediate(invoke->GetMethodAddress())); 4085 break; 4086 case HInvokeStaticOrDirect::MethodLoadKind::kDirectAddressWithFixup: 4087 __ movl(temp.AsRegister<Register>(), Immediate(0)); // Placeholder. 4088 method_patches_.emplace_back(invoke->GetTargetMethod()); 4089 __ Bind(&method_patches_.back().label); // Bind the label at the end of the "movl" insn. 4090 break; 4091 case HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative: { 4092 Register base_reg = GetInvokeStaticOrDirectExtraParameter(invoke, 4093 temp.AsRegister<Register>()); 4094 uint32_t offset = invoke->GetDexCacheArrayOffset(); 4095 __ movl(temp.AsRegister<Register>(), Address(base_reg, kDummy32BitOffset)); 4096 // Add the patch entry and bind its label at the end of the instruction. 4097 pc_relative_dex_cache_patches_.emplace_back(*invoke->GetTargetMethod().dex_file, offset); 4098 __ Bind(&pc_relative_dex_cache_patches_.back().label); 4099 break; 4100 } 4101 case HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod: { 4102 Location current_method = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 4103 Register method_reg; 4104 Register reg = temp.AsRegister<Register>(); 4105 if (current_method.IsRegister()) { 4106 method_reg = current_method.AsRegister<Register>(); 4107 } else { 4108 DCHECK(IsBaseline() || invoke->GetLocations()->Intrinsified()); 4109 DCHECK(!current_method.IsValid()); 4110 method_reg = reg; 4111 __ movl(reg, Address(ESP, kCurrentMethodStackOffset)); 4112 } 4113 // /* ArtMethod*[] */ temp = temp.ptr_sized_fields_->dex_cache_resolved_methods_; 4114 __ movl(reg, Address(method_reg, 4115 ArtMethod::DexCacheResolvedMethodsOffset(kX86PointerSize).Int32Value())); 4116 // temp = temp[index_in_cache] 4117 uint32_t index_in_cache = invoke->GetTargetMethod().dex_method_index; 4118 __ movl(reg, Address(reg, CodeGenerator::GetCachePointerOffset(index_in_cache))); 4119 break; 4120 } 4121 } 4122 4123 switch (invoke->GetCodePtrLocation()) { 4124 case HInvokeStaticOrDirect::CodePtrLocation::kCallSelf: 4125 __ call(GetFrameEntryLabel()); 4126 break; 4127 case HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative: { 4128 relative_call_patches_.emplace_back(invoke->GetTargetMethod()); 4129 Label* label = &relative_call_patches_.back().label; 4130 __ call(label); // Bind to the patch label, override at link time. 4131 __ Bind(label); // Bind the label at the end of the "call" insn. 4132 break; 4133 } 4134 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup: 4135 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect: 4136 // Filtered out by GetSupportedInvokeStaticOrDirectDispatch(). 4137 LOG(FATAL) << "Unsupported"; 4138 UNREACHABLE(); 4139 case HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod: 4140 // (callee_method + offset_of_quick_compiled_code)() 4141 __ call(Address(callee_method.AsRegister<Register>(), 4142 ArtMethod::EntryPointFromQuickCompiledCodeOffset( 4143 kX86WordSize).Int32Value())); 4144 break; 4145 } 4146 4147 DCHECK(!IsLeafMethod()); 4148} 4149 4150void CodeGeneratorX86::GenerateVirtualCall(HInvokeVirtual* invoke, Location temp_in) { 4151 Register temp = temp_in.AsRegister<Register>(); 4152 uint32_t method_offset = mirror::Class::EmbeddedVTableEntryOffset( 4153 invoke->GetVTableIndex(), kX86PointerSize).Uint32Value(); 4154 LocationSummary* locations = invoke->GetLocations(); 4155 Location receiver = locations->InAt(0); 4156 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 4157 DCHECK(receiver.IsRegister()); 4158 // /* HeapReference<Class> */ temp = receiver->klass_ 4159 __ movl(temp, Address(receiver.AsRegister<Register>(), class_offset)); 4160 MaybeRecordImplicitNullCheck(invoke); 4161 // Instead of simply (possibly) unpoisoning `temp` here, we should 4162 // emit a read barrier for the previous class reference load. 4163 // However this is not required in practice, as this is an 4164 // intermediate/temporary reference and because the current 4165 // concurrent copying collector keeps the from-space memory 4166 // intact/accessible until the end of the marking phase (the 4167 // concurrent copying collector may not in the future). 4168 __ MaybeUnpoisonHeapReference(temp); 4169 // temp = temp->GetMethodAt(method_offset); 4170 __ movl(temp, Address(temp, method_offset)); 4171 // call temp->GetEntryPoint(); 4172 __ call(Address( 4173 temp, ArtMethod::EntryPointFromQuickCompiledCodeOffset(kX86WordSize).Int32Value())); 4174} 4175 4176void CodeGeneratorX86::EmitLinkerPatches(ArenaVector<LinkerPatch>* linker_patches) { 4177 DCHECK(linker_patches->empty()); 4178 size_t size = 4179 method_patches_.size() + 4180 relative_call_patches_.size() + 4181 pc_relative_dex_cache_patches_.size(); 4182 linker_patches->reserve(size); 4183 // The label points to the end of the "movl" insn but the literal offset for method 4184 // patch needs to point to the embedded constant which occupies the last 4 bytes. 4185 constexpr uint32_t kLabelPositionToLiteralOffsetAdjustment = 4u; 4186 for (const MethodPatchInfo<Label>& info : method_patches_) { 4187 uint32_t literal_offset = info.label.Position() - kLabelPositionToLiteralOffsetAdjustment; 4188 linker_patches->push_back(LinkerPatch::MethodPatch(literal_offset, 4189 info.target_method.dex_file, 4190 info.target_method.dex_method_index)); 4191 } 4192 for (const MethodPatchInfo<Label>& info : relative_call_patches_) { 4193 uint32_t literal_offset = info.label.Position() - kLabelPositionToLiteralOffsetAdjustment; 4194 linker_patches->push_back(LinkerPatch::RelativeCodePatch(literal_offset, 4195 info.target_method.dex_file, 4196 info.target_method.dex_method_index)); 4197 } 4198 for (const PcRelativeDexCacheAccessInfo& info : pc_relative_dex_cache_patches_) { 4199 uint32_t literal_offset = info.label.Position() - kLabelPositionToLiteralOffsetAdjustment; 4200 linker_patches->push_back(LinkerPatch::DexCacheArrayPatch(literal_offset, 4201 &info.target_dex_file, 4202 GetMethodAddressOffset(), 4203 info.element_offset)); 4204 } 4205} 4206 4207void CodeGeneratorX86::MarkGCCard(Register temp, 4208 Register card, 4209 Register object, 4210 Register value, 4211 bool value_can_be_null) { 4212 NearLabel is_null; 4213 if (value_can_be_null) { 4214 __ testl(value, value); 4215 __ j(kEqual, &is_null); 4216 } 4217 __ fs()->movl(card, Address::Absolute(Thread::CardTableOffset<kX86WordSize>().Int32Value())); 4218 __ movl(temp, object); 4219 __ shrl(temp, Immediate(gc::accounting::CardTable::kCardShift)); 4220 __ movb(Address(temp, card, TIMES_1, 0), 4221 X86ManagedRegister::FromCpuRegister(card).AsByteRegister()); 4222 if (value_can_be_null) { 4223 __ Bind(&is_null); 4224 } 4225} 4226 4227void LocationsBuilderX86::HandleFieldGet(HInstruction* instruction, const FieldInfo& field_info) { 4228 DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet()); 4229 4230 bool object_field_get_with_read_barrier = 4231 kEmitCompilerReadBarrier && (instruction->GetType() == Primitive::kPrimNot); 4232 LocationSummary* locations = 4233 new (GetGraph()->GetArena()) LocationSummary(instruction, 4234 kEmitCompilerReadBarrier ? 4235 LocationSummary::kCallOnSlowPath : 4236 LocationSummary::kNoCall); 4237 locations->SetInAt(0, Location::RequiresRegister()); 4238 4239 if (Primitive::IsFloatingPointType(instruction->GetType())) { 4240 locations->SetOut(Location::RequiresFpuRegister()); 4241 } else { 4242 // The output overlaps in case of long: we don't want the low move 4243 // to overwrite the object's location. Likewise, in the case of 4244 // an object field get with read barriers enabled, we do not want 4245 // the move to overwrite the object's location, as we need it to emit 4246 // the read barrier. 4247 locations->SetOut( 4248 Location::RequiresRegister(), 4249 (object_field_get_with_read_barrier || instruction->GetType() == Primitive::kPrimLong) ? 4250 Location::kOutputOverlap : 4251 Location::kNoOutputOverlap); 4252 } 4253 4254 if (field_info.IsVolatile() && (field_info.GetFieldType() == Primitive::kPrimLong)) { 4255 // Long values can be loaded atomically into an XMM using movsd. 4256 // So we use an XMM register as a temp to achieve atomicity (first load the temp into the XMM 4257 // and then copy the XMM into the output 32bits at a time). 4258 locations->AddTemp(Location::RequiresFpuRegister()); 4259 } 4260} 4261 4262void InstructionCodeGeneratorX86::HandleFieldGet(HInstruction* instruction, 4263 const FieldInfo& field_info) { 4264 DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet()); 4265 4266 LocationSummary* locations = instruction->GetLocations(); 4267 Location base_loc = locations->InAt(0); 4268 Register base = base_loc.AsRegister<Register>(); 4269 Location out = locations->Out(); 4270 bool is_volatile = field_info.IsVolatile(); 4271 Primitive::Type field_type = field_info.GetFieldType(); 4272 uint32_t offset = field_info.GetFieldOffset().Uint32Value(); 4273 4274 switch (field_type) { 4275 case Primitive::kPrimBoolean: { 4276 __ movzxb(out.AsRegister<Register>(), Address(base, offset)); 4277 break; 4278 } 4279 4280 case Primitive::kPrimByte: { 4281 __ movsxb(out.AsRegister<Register>(), Address(base, offset)); 4282 break; 4283 } 4284 4285 case Primitive::kPrimShort: { 4286 __ movsxw(out.AsRegister<Register>(), Address(base, offset)); 4287 break; 4288 } 4289 4290 case Primitive::kPrimChar: { 4291 __ movzxw(out.AsRegister<Register>(), Address(base, offset)); 4292 break; 4293 } 4294 4295 case Primitive::kPrimInt: 4296 case Primitive::kPrimNot: { 4297 __ movl(out.AsRegister<Register>(), Address(base, offset)); 4298 break; 4299 } 4300 4301 case Primitive::kPrimLong: { 4302 if (is_volatile) { 4303 XmmRegister temp = locations->GetTemp(0).AsFpuRegister<XmmRegister>(); 4304 __ movsd(temp, Address(base, offset)); 4305 codegen_->MaybeRecordImplicitNullCheck(instruction); 4306 __ movd(out.AsRegisterPairLow<Register>(), temp); 4307 __ psrlq(temp, Immediate(32)); 4308 __ movd(out.AsRegisterPairHigh<Register>(), temp); 4309 } else { 4310 DCHECK_NE(base, out.AsRegisterPairLow<Register>()); 4311 __ movl(out.AsRegisterPairLow<Register>(), Address(base, offset)); 4312 codegen_->MaybeRecordImplicitNullCheck(instruction); 4313 __ movl(out.AsRegisterPairHigh<Register>(), Address(base, kX86WordSize + offset)); 4314 } 4315 break; 4316 } 4317 4318 case Primitive::kPrimFloat: { 4319 __ movss(out.AsFpuRegister<XmmRegister>(), Address(base, offset)); 4320 break; 4321 } 4322 4323 case Primitive::kPrimDouble: { 4324 __ movsd(out.AsFpuRegister<XmmRegister>(), Address(base, offset)); 4325 break; 4326 } 4327 4328 case Primitive::kPrimVoid: 4329 LOG(FATAL) << "Unreachable type " << field_type; 4330 UNREACHABLE(); 4331 } 4332 4333 // Longs are handled in the switch. 4334 if (field_type != Primitive::kPrimLong) { 4335 codegen_->MaybeRecordImplicitNullCheck(instruction); 4336 } 4337 4338 if (is_volatile) { 4339 GenerateMemoryBarrier(MemBarrierKind::kLoadAny); 4340 } 4341 4342 if (field_type == Primitive::kPrimNot) { 4343 codegen_->MaybeGenerateReadBarrier(instruction, out, out, base_loc, offset); 4344 } 4345} 4346 4347void LocationsBuilderX86::HandleFieldSet(HInstruction* instruction, const FieldInfo& field_info) { 4348 DCHECK(instruction->IsInstanceFieldSet() || instruction->IsStaticFieldSet()); 4349 4350 LocationSummary* locations = 4351 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 4352 locations->SetInAt(0, Location::RequiresRegister()); 4353 bool is_volatile = field_info.IsVolatile(); 4354 Primitive::Type field_type = field_info.GetFieldType(); 4355 bool is_byte_type = (field_type == Primitive::kPrimBoolean) 4356 || (field_type == Primitive::kPrimByte); 4357 4358 // The register allocator does not support multiple 4359 // inputs that die at entry with one in a specific register. 4360 if (is_byte_type) { 4361 // Ensure the value is in a byte register. 4362 locations->SetInAt(1, Location::RegisterLocation(EAX)); 4363 } else if (Primitive::IsFloatingPointType(field_type)) { 4364 if (is_volatile && field_type == Primitive::kPrimDouble) { 4365 // In order to satisfy the semantics of volatile, this must be a single instruction store. 4366 locations->SetInAt(1, Location::RequiresFpuRegister()); 4367 } else { 4368 locations->SetInAt(1, Location::FpuRegisterOrConstant(instruction->InputAt(1))); 4369 } 4370 } else if (is_volatile && field_type == Primitive::kPrimLong) { 4371 // In order to satisfy the semantics of volatile, this must be a single instruction store. 4372 locations->SetInAt(1, Location::RequiresRegister()); 4373 4374 // 64bits value can be atomically written to an address with movsd and an XMM register. 4375 // We need two XMM registers because there's no easier way to (bit) copy a register pair 4376 // into a single XMM register (we copy each pair part into the XMMs and then interleave them). 4377 // NB: We could make the register allocator understand fp_reg <-> core_reg moves but given the 4378 // isolated cases when we need this it isn't worth adding the extra complexity. 4379 locations->AddTemp(Location::RequiresFpuRegister()); 4380 locations->AddTemp(Location::RequiresFpuRegister()); 4381 } else { 4382 locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1))); 4383 4384 if (CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1))) { 4385 // Temporary registers for the write barrier. 4386 locations->AddTemp(Location::RequiresRegister()); // May be used for reference poisoning too. 4387 // Ensure the card is in a byte register. 4388 locations->AddTemp(Location::RegisterLocation(ECX)); 4389 } 4390 } 4391} 4392 4393void InstructionCodeGeneratorX86::HandleFieldSet(HInstruction* instruction, 4394 const FieldInfo& field_info, 4395 bool value_can_be_null) { 4396 DCHECK(instruction->IsInstanceFieldSet() || instruction->IsStaticFieldSet()); 4397 4398 LocationSummary* locations = instruction->GetLocations(); 4399 Register base = locations->InAt(0).AsRegister<Register>(); 4400 Location value = locations->InAt(1); 4401 bool is_volatile = field_info.IsVolatile(); 4402 Primitive::Type field_type = field_info.GetFieldType(); 4403 uint32_t offset = field_info.GetFieldOffset().Uint32Value(); 4404 bool needs_write_barrier = 4405 CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1)); 4406 4407 if (is_volatile) { 4408 GenerateMemoryBarrier(MemBarrierKind::kAnyStore); 4409 } 4410 4411 bool maybe_record_implicit_null_check_done = false; 4412 4413 switch (field_type) { 4414 case Primitive::kPrimBoolean: 4415 case Primitive::kPrimByte: { 4416 __ movb(Address(base, offset), value.AsRegister<ByteRegister>()); 4417 break; 4418 } 4419 4420 case Primitive::kPrimShort: 4421 case Primitive::kPrimChar: { 4422 if (value.IsConstant()) { 4423 int16_t v = CodeGenerator::GetInt32ValueOf(value.GetConstant()); 4424 __ movw(Address(base, offset), Immediate(v)); 4425 } else { 4426 __ movw(Address(base, offset), value.AsRegister<Register>()); 4427 } 4428 break; 4429 } 4430 4431 case Primitive::kPrimInt: 4432 case Primitive::kPrimNot: { 4433 if (kPoisonHeapReferences && needs_write_barrier) { 4434 // Note that in the case where `value` is a null reference, 4435 // we do not enter this block, as the reference does not 4436 // need poisoning. 4437 DCHECK_EQ(field_type, Primitive::kPrimNot); 4438 Register temp = locations->GetTemp(0).AsRegister<Register>(); 4439 __ movl(temp, value.AsRegister<Register>()); 4440 __ PoisonHeapReference(temp); 4441 __ movl(Address(base, offset), temp); 4442 } else if (value.IsConstant()) { 4443 int32_t v = CodeGenerator::GetInt32ValueOf(value.GetConstant()); 4444 __ movl(Address(base, offset), Immediate(v)); 4445 } else { 4446 __ movl(Address(base, offset), value.AsRegister<Register>()); 4447 } 4448 break; 4449 } 4450 4451 case Primitive::kPrimLong: { 4452 if (is_volatile) { 4453 XmmRegister temp1 = locations->GetTemp(0).AsFpuRegister<XmmRegister>(); 4454 XmmRegister temp2 = locations->GetTemp(1).AsFpuRegister<XmmRegister>(); 4455 __ movd(temp1, value.AsRegisterPairLow<Register>()); 4456 __ movd(temp2, value.AsRegisterPairHigh<Register>()); 4457 __ punpckldq(temp1, temp2); 4458 __ movsd(Address(base, offset), temp1); 4459 codegen_->MaybeRecordImplicitNullCheck(instruction); 4460 } else if (value.IsConstant()) { 4461 int64_t v = CodeGenerator::GetInt64ValueOf(value.GetConstant()); 4462 __ movl(Address(base, offset), Immediate(Low32Bits(v))); 4463 codegen_->MaybeRecordImplicitNullCheck(instruction); 4464 __ movl(Address(base, kX86WordSize + offset), Immediate(High32Bits(v))); 4465 } else { 4466 __ movl(Address(base, offset), value.AsRegisterPairLow<Register>()); 4467 codegen_->MaybeRecordImplicitNullCheck(instruction); 4468 __ movl(Address(base, kX86WordSize + offset), value.AsRegisterPairHigh<Register>()); 4469 } 4470 maybe_record_implicit_null_check_done = true; 4471 break; 4472 } 4473 4474 case Primitive::kPrimFloat: { 4475 if (value.IsConstant()) { 4476 int32_t v = CodeGenerator::GetInt32ValueOf(value.GetConstant()); 4477 __ movl(Address(base, offset), Immediate(v)); 4478 } else { 4479 __ movss(Address(base, offset), value.AsFpuRegister<XmmRegister>()); 4480 } 4481 break; 4482 } 4483 4484 case Primitive::kPrimDouble: { 4485 if (value.IsConstant()) { 4486 int64_t v = CodeGenerator::GetInt64ValueOf(value.GetConstant()); 4487 __ movl(Address(base, offset), Immediate(Low32Bits(v))); 4488 codegen_->MaybeRecordImplicitNullCheck(instruction); 4489 __ movl(Address(base, kX86WordSize + offset), Immediate(High32Bits(v))); 4490 maybe_record_implicit_null_check_done = true; 4491 } else { 4492 __ movsd(Address(base, offset), value.AsFpuRegister<XmmRegister>()); 4493 } 4494 break; 4495 } 4496 4497 case Primitive::kPrimVoid: 4498 LOG(FATAL) << "Unreachable type " << field_type; 4499 UNREACHABLE(); 4500 } 4501 4502 if (!maybe_record_implicit_null_check_done) { 4503 codegen_->MaybeRecordImplicitNullCheck(instruction); 4504 } 4505 4506 if (needs_write_barrier) { 4507 Register temp = locations->GetTemp(0).AsRegister<Register>(); 4508 Register card = locations->GetTemp(1).AsRegister<Register>(); 4509 codegen_->MarkGCCard(temp, card, base, value.AsRegister<Register>(), value_can_be_null); 4510 } 4511 4512 if (is_volatile) { 4513 GenerateMemoryBarrier(MemBarrierKind::kAnyAny); 4514 } 4515} 4516 4517void LocationsBuilderX86::VisitStaticFieldGet(HStaticFieldGet* instruction) { 4518 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4519} 4520 4521void InstructionCodeGeneratorX86::VisitStaticFieldGet(HStaticFieldGet* instruction) { 4522 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4523} 4524 4525void LocationsBuilderX86::VisitStaticFieldSet(HStaticFieldSet* instruction) { 4526 HandleFieldSet(instruction, instruction->GetFieldInfo()); 4527} 4528 4529void InstructionCodeGeneratorX86::VisitStaticFieldSet(HStaticFieldSet* instruction) { 4530 HandleFieldSet(instruction, instruction->GetFieldInfo(), instruction->GetValueCanBeNull()); 4531} 4532 4533void LocationsBuilderX86::VisitInstanceFieldSet(HInstanceFieldSet* instruction) { 4534 HandleFieldSet(instruction, instruction->GetFieldInfo()); 4535} 4536 4537void InstructionCodeGeneratorX86::VisitInstanceFieldSet(HInstanceFieldSet* instruction) { 4538 HandleFieldSet(instruction, instruction->GetFieldInfo(), instruction->GetValueCanBeNull()); 4539} 4540 4541void LocationsBuilderX86::VisitInstanceFieldGet(HInstanceFieldGet* instruction) { 4542 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4543} 4544 4545void InstructionCodeGeneratorX86::VisitInstanceFieldGet(HInstanceFieldGet* instruction) { 4546 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4547} 4548 4549void LocationsBuilderX86::VisitUnresolvedInstanceFieldGet( 4550 HUnresolvedInstanceFieldGet* instruction) { 4551 FieldAccessCallingConventionX86 calling_convention; 4552 codegen_->CreateUnresolvedFieldLocationSummary( 4553 instruction, instruction->GetFieldType(), calling_convention); 4554} 4555 4556void InstructionCodeGeneratorX86::VisitUnresolvedInstanceFieldGet( 4557 HUnresolvedInstanceFieldGet* instruction) { 4558 FieldAccessCallingConventionX86 calling_convention; 4559 codegen_->GenerateUnresolvedFieldAccess(instruction, 4560 instruction->GetFieldType(), 4561 instruction->GetFieldIndex(), 4562 instruction->GetDexPc(), 4563 calling_convention); 4564} 4565 4566void LocationsBuilderX86::VisitUnresolvedInstanceFieldSet( 4567 HUnresolvedInstanceFieldSet* instruction) { 4568 FieldAccessCallingConventionX86 calling_convention; 4569 codegen_->CreateUnresolvedFieldLocationSummary( 4570 instruction, instruction->GetFieldType(), calling_convention); 4571} 4572 4573void InstructionCodeGeneratorX86::VisitUnresolvedInstanceFieldSet( 4574 HUnresolvedInstanceFieldSet* instruction) { 4575 FieldAccessCallingConventionX86 calling_convention; 4576 codegen_->GenerateUnresolvedFieldAccess(instruction, 4577 instruction->GetFieldType(), 4578 instruction->GetFieldIndex(), 4579 instruction->GetDexPc(), 4580 calling_convention); 4581} 4582 4583void LocationsBuilderX86::VisitUnresolvedStaticFieldGet( 4584 HUnresolvedStaticFieldGet* instruction) { 4585 FieldAccessCallingConventionX86 calling_convention; 4586 codegen_->CreateUnresolvedFieldLocationSummary( 4587 instruction, instruction->GetFieldType(), calling_convention); 4588} 4589 4590void InstructionCodeGeneratorX86::VisitUnresolvedStaticFieldGet( 4591 HUnresolvedStaticFieldGet* instruction) { 4592 FieldAccessCallingConventionX86 calling_convention; 4593 codegen_->GenerateUnresolvedFieldAccess(instruction, 4594 instruction->GetFieldType(), 4595 instruction->GetFieldIndex(), 4596 instruction->GetDexPc(), 4597 calling_convention); 4598} 4599 4600void LocationsBuilderX86::VisitUnresolvedStaticFieldSet( 4601 HUnresolvedStaticFieldSet* instruction) { 4602 FieldAccessCallingConventionX86 calling_convention; 4603 codegen_->CreateUnresolvedFieldLocationSummary( 4604 instruction, instruction->GetFieldType(), calling_convention); 4605} 4606 4607void InstructionCodeGeneratorX86::VisitUnresolvedStaticFieldSet( 4608 HUnresolvedStaticFieldSet* instruction) { 4609 FieldAccessCallingConventionX86 calling_convention; 4610 codegen_->GenerateUnresolvedFieldAccess(instruction, 4611 instruction->GetFieldType(), 4612 instruction->GetFieldIndex(), 4613 instruction->GetDexPc(), 4614 calling_convention); 4615} 4616 4617void LocationsBuilderX86::VisitNullCheck(HNullCheck* instruction) { 4618 LocationSummary::CallKind call_kind = instruction->CanThrowIntoCatchBlock() 4619 ? LocationSummary::kCallOnSlowPath 4620 : LocationSummary::kNoCall; 4621 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 4622 Location loc = codegen_->IsImplicitNullCheckAllowed(instruction) 4623 ? Location::RequiresRegister() 4624 : Location::Any(); 4625 locations->SetInAt(0, loc); 4626 if (instruction->HasUses()) { 4627 locations->SetOut(Location::SameAsFirstInput()); 4628 } 4629} 4630 4631void InstructionCodeGeneratorX86::GenerateImplicitNullCheck(HNullCheck* instruction) { 4632 if (codegen_->CanMoveNullCheckToUser(instruction)) { 4633 return; 4634 } 4635 LocationSummary* locations = instruction->GetLocations(); 4636 Location obj = locations->InAt(0); 4637 4638 __ testl(EAX, Address(obj.AsRegister<Register>(), 0)); 4639 codegen_->RecordPcInfo(instruction, instruction->GetDexPc()); 4640} 4641 4642void InstructionCodeGeneratorX86::GenerateExplicitNullCheck(HNullCheck* instruction) { 4643 SlowPathCode* slow_path = new (GetGraph()->GetArena()) NullCheckSlowPathX86(instruction); 4644 codegen_->AddSlowPath(slow_path); 4645 4646 LocationSummary* locations = instruction->GetLocations(); 4647 Location obj = locations->InAt(0); 4648 4649 if (obj.IsRegister()) { 4650 __ testl(obj.AsRegister<Register>(), obj.AsRegister<Register>()); 4651 } else if (obj.IsStackSlot()) { 4652 __ cmpl(Address(ESP, obj.GetStackIndex()), Immediate(0)); 4653 } else { 4654 DCHECK(obj.IsConstant()) << obj; 4655 DCHECK(obj.GetConstant()->IsNullConstant()); 4656 __ jmp(slow_path->GetEntryLabel()); 4657 return; 4658 } 4659 __ j(kEqual, slow_path->GetEntryLabel()); 4660} 4661 4662void InstructionCodeGeneratorX86::VisitNullCheck(HNullCheck* instruction) { 4663 if (codegen_->IsImplicitNullCheckAllowed(instruction)) { 4664 GenerateImplicitNullCheck(instruction); 4665 } else { 4666 GenerateExplicitNullCheck(instruction); 4667 } 4668} 4669 4670void LocationsBuilderX86::VisitArrayGet(HArrayGet* instruction) { 4671 bool object_array_get_with_read_barrier = 4672 kEmitCompilerReadBarrier && (instruction->GetType() == Primitive::kPrimNot); 4673 LocationSummary* locations = 4674 new (GetGraph()->GetArena()) LocationSummary(instruction, 4675 object_array_get_with_read_barrier ? 4676 LocationSummary::kCallOnSlowPath : 4677 LocationSummary::kNoCall); 4678 locations->SetInAt(0, Location::RequiresRegister()); 4679 locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1))); 4680 if (Primitive::IsFloatingPointType(instruction->GetType())) { 4681 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 4682 } else { 4683 // The output overlaps in case of long: we don't want the low move 4684 // to overwrite the array's location. Likewise, in the case of an 4685 // object array get with read barriers enabled, we do not want the 4686 // move to overwrite the array's location, as we need it to emit 4687 // the read barrier. 4688 locations->SetOut( 4689 Location::RequiresRegister(), 4690 (instruction->GetType() == Primitive::kPrimLong || object_array_get_with_read_barrier) ? 4691 Location::kOutputOverlap : 4692 Location::kNoOutputOverlap); 4693 } 4694} 4695 4696void InstructionCodeGeneratorX86::VisitArrayGet(HArrayGet* instruction) { 4697 LocationSummary* locations = instruction->GetLocations(); 4698 Location obj_loc = locations->InAt(0); 4699 Register obj = obj_loc.AsRegister<Register>(); 4700 Location index = locations->InAt(1); 4701 4702 Primitive::Type type = instruction->GetType(); 4703 switch (type) { 4704 case Primitive::kPrimBoolean: { 4705 uint32_t data_offset = mirror::Array::DataOffset(sizeof(uint8_t)).Uint32Value(); 4706 Register out = locations->Out().AsRegister<Register>(); 4707 if (index.IsConstant()) { 4708 __ movzxb(out, Address(obj, 4709 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_1) + data_offset)); 4710 } else { 4711 __ movzxb(out, Address(obj, index.AsRegister<Register>(), TIMES_1, data_offset)); 4712 } 4713 break; 4714 } 4715 4716 case Primitive::kPrimByte: { 4717 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int8_t)).Uint32Value(); 4718 Register out = locations->Out().AsRegister<Register>(); 4719 if (index.IsConstant()) { 4720 __ movsxb(out, Address(obj, 4721 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_1) + data_offset)); 4722 } else { 4723 __ movsxb(out, Address(obj, index.AsRegister<Register>(), TIMES_1, data_offset)); 4724 } 4725 break; 4726 } 4727 4728 case Primitive::kPrimShort: { 4729 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int16_t)).Uint32Value(); 4730 Register out = locations->Out().AsRegister<Register>(); 4731 if (index.IsConstant()) { 4732 __ movsxw(out, Address(obj, 4733 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_2) + data_offset)); 4734 } else { 4735 __ movsxw(out, Address(obj, index.AsRegister<Register>(), TIMES_2, data_offset)); 4736 } 4737 break; 4738 } 4739 4740 case Primitive::kPrimChar: { 4741 uint32_t data_offset = mirror::Array::DataOffset(sizeof(uint16_t)).Uint32Value(); 4742 Register out = locations->Out().AsRegister<Register>(); 4743 if (index.IsConstant()) { 4744 __ movzxw(out, Address(obj, 4745 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_2) + data_offset)); 4746 } else { 4747 __ movzxw(out, Address(obj, index.AsRegister<Register>(), TIMES_2, data_offset)); 4748 } 4749 break; 4750 } 4751 4752 case Primitive::kPrimInt: 4753 case Primitive::kPrimNot: { 4754 static_assert( 4755 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 4756 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 4757 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int32_t)).Uint32Value(); 4758 Register out = locations->Out().AsRegister<Register>(); 4759 if (index.IsConstant()) { 4760 __ movl(out, Address(obj, 4761 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset)); 4762 } else { 4763 __ movl(out, Address(obj, index.AsRegister<Register>(), TIMES_4, data_offset)); 4764 } 4765 break; 4766 } 4767 4768 case Primitive::kPrimLong: { 4769 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int64_t)).Uint32Value(); 4770 Location out = locations->Out(); 4771 DCHECK_NE(obj, out.AsRegisterPairLow<Register>()); 4772 if (index.IsConstant()) { 4773 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 4774 __ movl(out.AsRegisterPairLow<Register>(), Address(obj, offset)); 4775 codegen_->MaybeRecordImplicitNullCheck(instruction); 4776 __ movl(out.AsRegisterPairHigh<Register>(), Address(obj, offset + kX86WordSize)); 4777 } else { 4778 __ movl(out.AsRegisterPairLow<Register>(), 4779 Address(obj, index.AsRegister<Register>(), TIMES_8, data_offset)); 4780 codegen_->MaybeRecordImplicitNullCheck(instruction); 4781 __ movl(out.AsRegisterPairHigh<Register>(), 4782 Address(obj, index.AsRegister<Register>(), TIMES_8, data_offset + kX86WordSize)); 4783 } 4784 break; 4785 } 4786 4787 case Primitive::kPrimFloat: { 4788 uint32_t data_offset = mirror::Array::DataOffset(sizeof(float)).Uint32Value(); 4789 XmmRegister out = locations->Out().AsFpuRegister<XmmRegister>(); 4790 if (index.IsConstant()) { 4791 __ movss(out, Address(obj, 4792 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset)); 4793 } else { 4794 __ movss(out, Address(obj, index.AsRegister<Register>(), TIMES_4, data_offset)); 4795 } 4796 break; 4797 } 4798 4799 case Primitive::kPrimDouble: { 4800 uint32_t data_offset = mirror::Array::DataOffset(sizeof(double)).Uint32Value(); 4801 XmmRegister out = locations->Out().AsFpuRegister<XmmRegister>(); 4802 if (index.IsConstant()) { 4803 __ movsd(out, Address(obj, 4804 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset)); 4805 } else { 4806 __ movsd(out, Address(obj, index.AsRegister<Register>(), TIMES_8, data_offset)); 4807 } 4808 break; 4809 } 4810 4811 case Primitive::kPrimVoid: 4812 LOG(FATAL) << "Unreachable type " << type; 4813 UNREACHABLE(); 4814 } 4815 4816 if (type != Primitive::kPrimLong) { 4817 codegen_->MaybeRecordImplicitNullCheck(instruction); 4818 } 4819 4820 if (type == Primitive::kPrimNot) { 4821 static_assert( 4822 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 4823 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 4824 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int32_t)).Uint32Value(); 4825 Location out = locations->Out(); 4826 if (index.IsConstant()) { 4827 uint32_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4828 codegen_->MaybeGenerateReadBarrier(instruction, out, out, obj_loc, offset); 4829 } else { 4830 codegen_->MaybeGenerateReadBarrier(instruction, out, out, obj_loc, data_offset, index); 4831 } 4832 } 4833} 4834 4835void LocationsBuilderX86::VisitArraySet(HArraySet* instruction) { 4836 // This location builder might end up asking to up to four registers, which is 4837 // not currently possible for baseline. The situation in which we need four 4838 // registers cannot be met by baseline though, because it has not run any 4839 // optimization. 4840 4841 Primitive::Type value_type = instruction->GetComponentType(); 4842 4843 bool needs_write_barrier = 4844 CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue()); 4845 bool may_need_runtime_call_for_type_check = instruction->NeedsTypeCheck(); 4846 bool object_array_set_with_read_barrier = 4847 kEmitCompilerReadBarrier && (value_type == Primitive::kPrimNot); 4848 4849 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary( 4850 instruction, 4851 (may_need_runtime_call_for_type_check || object_array_set_with_read_barrier) ? 4852 LocationSummary::kCallOnSlowPath : 4853 LocationSummary::kNoCall); 4854 4855 bool is_byte_type = (value_type == Primitive::kPrimBoolean) 4856 || (value_type == Primitive::kPrimByte); 4857 // We need the inputs to be different than the output in case of long operation. 4858 // In case of a byte operation, the register allocator does not support multiple 4859 // inputs that die at entry with one in a specific register. 4860 locations->SetInAt(0, Location::RequiresRegister()); 4861 locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1))); 4862 if (is_byte_type) { 4863 // Ensure the value is in a byte register. 4864 locations->SetInAt(2, Location::ByteRegisterOrConstant(EAX, instruction->InputAt(2))); 4865 } else if (Primitive::IsFloatingPointType(value_type)) { 4866 locations->SetInAt(2, Location::FpuRegisterOrConstant(instruction->InputAt(2))); 4867 } else { 4868 locations->SetInAt(2, Location::RegisterOrConstant(instruction->InputAt(2))); 4869 } 4870 if (needs_write_barrier) { 4871 // Temporary registers for the write barrier. 4872 locations->AddTemp(Location::RequiresRegister()); // Possibly used for ref. poisoning too. 4873 // Ensure the card is in a byte register. 4874 locations->AddTemp(Location::RegisterLocation(ECX)); 4875 } 4876} 4877 4878void InstructionCodeGeneratorX86::VisitArraySet(HArraySet* instruction) { 4879 LocationSummary* locations = instruction->GetLocations(); 4880 Location array_loc = locations->InAt(0); 4881 Register array = array_loc.AsRegister<Register>(); 4882 Location index = locations->InAt(1); 4883 Location value = locations->InAt(2); 4884 Primitive::Type value_type = instruction->GetComponentType(); 4885 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 4886 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 4887 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 4888 bool may_need_runtime_call_for_type_check = instruction->NeedsTypeCheck(); 4889 bool needs_write_barrier = 4890 CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue()); 4891 4892 switch (value_type) { 4893 case Primitive::kPrimBoolean: 4894 case Primitive::kPrimByte: { 4895 uint32_t offset = mirror::Array::DataOffset(sizeof(uint8_t)).Uint32Value(); 4896 Address address = index.IsConstant() 4897 ? Address(array, (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_1) + offset) 4898 : Address(array, index.AsRegister<Register>(), TIMES_1, offset); 4899 if (value.IsRegister()) { 4900 __ movb(address, value.AsRegister<ByteRegister>()); 4901 } else { 4902 __ movb(address, Immediate(value.GetConstant()->AsIntConstant()->GetValue())); 4903 } 4904 codegen_->MaybeRecordImplicitNullCheck(instruction); 4905 break; 4906 } 4907 4908 case Primitive::kPrimShort: 4909 case Primitive::kPrimChar: { 4910 uint32_t offset = mirror::Array::DataOffset(sizeof(uint16_t)).Uint32Value(); 4911 Address address = index.IsConstant() 4912 ? Address(array, (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_2) + offset) 4913 : Address(array, index.AsRegister<Register>(), TIMES_2, offset); 4914 if (value.IsRegister()) { 4915 __ movw(address, value.AsRegister<Register>()); 4916 } else { 4917 __ movw(address, Immediate(value.GetConstant()->AsIntConstant()->GetValue())); 4918 } 4919 codegen_->MaybeRecordImplicitNullCheck(instruction); 4920 break; 4921 } 4922 4923 case Primitive::kPrimNot: { 4924 uint32_t offset = mirror::Array::DataOffset(sizeof(int32_t)).Uint32Value(); 4925 Address address = index.IsConstant() 4926 ? Address(array, (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + offset) 4927 : Address(array, index.AsRegister<Register>(), TIMES_4, offset); 4928 4929 if (!value.IsRegister()) { 4930 // Just setting null. 4931 DCHECK(instruction->InputAt(2)->IsNullConstant()); 4932 DCHECK(value.IsConstant()) << value; 4933 __ movl(address, Immediate(0)); 4934 codegen_->MaybeRecordImplicitNullCheck(instruction); 4935 DCHECK(!needs_write_barrier); 4936 DCHECK(!may_need_runtime_call_for_type_check); 4937 break; 4938 } 4939 4940 DCHECK(needs_write_barrier); 4941 Register register_value = value.AsRegister<Register>(); 4942 NearLabel done, not_null, do_put; 4943 SlowPathCode* slow_path = nullptr; 4944 Register temp = locations->GetTemp(0).AsRegister<Register>(); 4945 if (may_need_runtime_call_for_type_check) { 4946 slow_path = new (GetGraph()->GetArena()) ArraySetSlowPathX86(instruction); 4947 codegen_->AddSlowPath(slow_path); 4948 if (instruction->GetValueCanBeNull()) { 4949 __ testl(register_value, register_value); 4950 __ j(kNotEqual, ¬_null); 4951 __ movl(address, Immediate(0)); 4952 codegen_->MaybeRecordImplicitNullCheck(instruction); 4953 __ jmp(&done); 4954 __ Bind(¬_null); 4955 } 4956 4957 if (kEmitCompilerReadBarrier) { 4958 // When read barriers are enabled, the type checking 4959 // instrumentation requires two read barriers: 4960 // 4961 // __ movl(temp2, temp); 4962 // // /* HeapReference<Class> */ temp = temp->component_type_ 4963 // __ movl(temp, Address(temp, component_offset)); 4964 // codegen_->GenerateReadBarrier( 4965 // instruction, temp_loc, temp_loc, temp2_loc, component_offset); 4966 // 4967 // // /* HeapReference<Class> */ temp2 = register_value->klass_ 4968 // __ movl(temp2, Address(register_value, class_offset)); 4969 // codegen_->GenerateReadBarrier( 4970 // instruction, temp2_loc, temp2_loc, value, class_offset, temp_loc); 4971 // 4972 // __ cmpl(temp, temp2); 4973 // 4974 // However, the second read barrier may trash `temp`, as it 4975 // is a temporary register, and as such would not be saved 4976 // along with live registers before calling the runtime (nor 4977 // restored afterwards). So in this case, we bail out and 4978 // delegate the work to the array set slow path. 4979 // 4980 // TODO: Extend the register allocator to support a new 4981 // "(locally) live temp" location so as to avoid always 4982 // going into the slow path when read barriers are enabled. 4983 __ jmp(slow_path->GetEntryLabel()); 4984 } else { 4985 // /* HeapReference<Class> */ temp = array->klass_ 4986 __ movl(temp, Address(array, class_offset)); 4987 codegen_->MaybeRecordImplicitNullCheck(instruction); 4988 __ MaybeUnpoisonHeapReference(temp); 4989 4990 // /* HeapReference<Class> */ temp = temp->component_type_ 4991 __ movl(temp, Address(temp, component_offset)); 4992 // If heap poisoning is enabled, no need to unpoison `temp` 4993 // nor the object reference in `register_value->klass`, as 4994 // we are comparing two poisoned references. 4995 __ cmpl(temp, Address(register_value, class_offset)); 4996 4997 if (instruction->StaticTypeOfArrayIsObjectArray()) { 4998 __ j(kEqual, &do_put); 4999 // If heap poisoning is enabled, the `temp` reference has 5000 // not been unpoisoned yet; unpoison it now. 5001 __ MaybeUnpoisonHeapReference(temp); 5002 5003 // /* HeapReference<Class> */ temp = temp->super_class_ 5004 __ movl(temp, Address(temp, super_offset)); 5005 // If heap poisoning is enabled, no need to unpoison 5006 // `temp`, as we are comparing against null below. 5007 __ testl(temp, temp); 5008 __ j(kNotEqual, slow_path->GetEntryLabel()); 5009 __ Bind(&do_put); 5010 } else { 5011 __ j(kNotEqual, slow_path->GetEntryLabel()); 5012 } 5013 } 5014 } 5015 5016 if (kPoisonHeapReferences) { 5017 __ movl(temp, register_value); 5018 __ PoisonHeapReference(temp); 5019 __ movl(address, temp); 5020 } else { 5021 __ movl(address, register_value); 5022 } 5023 if (!may_need_runtime_call_for_type_check) { 5024 codegen_->MaybeRecordImplicitNullCheck(instruction); 5025 } 5026 5027 Register card = locations->GetTemp(1).AsRegister<Register>(); 5028 codegen_->MarkGCCard( 5029 temp, card, array, value.AsRegister<Register>(), instruction->GetValueCanBeNull()); 5030 __ Bind(&done); 5031 5032 if (slow_path != nullptr) { 5033 __ Bind(slow_path->GetExitLabel()); 5034 } 5035 5036 break; 5037 } 5038 5039 case Primitive::kPrimInt: { 5040 uint32_t offset = mirror::Array::DataOffset(sizeof(int32_t)).Uint32Value(); 5041 Address address = index.IsConstant() 5042 ? Address(array, (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + offset) 5043 : Address(array, index.AsRegister<Register>(), TIMES_4, offset); 5044 if (value.IsRegister()) { 5045 __ movl(address, value.AsRegister<Register>()); 5046 } else { 5047 DCHECK(value.IsConstant()) << value; 5048 int32_t v = CodeGenerator::GetInt32ValueOf(value.GetConstant()); 5049 __ movl(address, Immediate(v)); 5050 } 5051 codegen_->MaybeRecordImplicitNullCheck(instruction); 5052 break; 5053 } 5054 5055 case Primitive::kPrimLong: { 5056 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int64_t)).Uint32Value(); 5057 if (index.IsConstant()) { 5058 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 5059 if (value.IsRegisterPair()) { 5060 __ movl(Address(array, offset), value.AsRegisterPairLow<Register>()); 5061 codegen_->MaybeRecordImplicitNullCheck(instruction); 5062 __ movl(Address(array, offset + kX86WordSize), value.AsRegisterPairHigh<Register>()); 5063 } else { 5064 DCHECK(value.IsConstant()); 5065 int64_t val = value.GetConstant()->AsLongConstant()->GetValue(); 5066 __ movl(Address(array, offset), Immediate(Low32Bits(val))); 5067 codegen_->MaybeRecordImplicitNullCheck(instruction); 5068 __ movl(Address(array, offset + kX86WordSize), Immediate(High32Bits(val))); 5069 } 5070 } else { 5071 if (value.IsRegisterPair()) { 5072 __ movl(Address(array, index.AsRegister<Register>(), TIMES_8, data_offset), 5073 value.AsRegisterPairLow<Register>()); 5074 codegen_->MaybeRecordImplicitNullCheck(instruction); 5075 __ movl(Address(array, index.AsRegister<Register>(), TIMES_8, data_offset + kX86WordSize), 5076 value.AsRegisterPairHigh<Register>()); 5077 } else { 5078 DCHECK(value.IsConstant()); 5079 int64_t val = value.GetConstant()->AsLongConstant()->GetValue(); 5080 __ movl(Address(array, index.AsRegister<Register>(), TIMES_8, data_offset), 5081 Immediate(Low32Bits(val))); 5082 codegen_->MaybeRecordImplicitNullCheck(instruction); 5083 __ movl(Address(array, index.AsRegister<Register>(), TIMES_8, data_offset + kX86WordSize), 5084 Immediate(High32Bits(val))); 5085 } 5086 } 5087 break; 5088 } 5089 5090 case Primitive::kPrimFloat: { 5091 uint32_t offset = mirror::Array::DataOffset(sizeof(float)).Uint32Value(); 5092 Address address = index.IsConstant() 5093 ? Address(array, (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + offset) 5094 : Address(array, index.AsRegister<Register>(), TIMES_4, offset); 5095 if (value.IsFpuRegister()) { 5096 __ movss(address, value.AsFpuRegister<XmmRegister>()); 5097 } else { 5098 DCHECK(value.IsConstant()); 5099 int32_t v = bit_cast<int32_t, float>(value.GetConstant()->AsFloatConstant()->GetValue()); 5100 __ movl(address, Immediate(v)); 5101 } 5102 codegen_->MaybeRecordImplicitNullCheck(instruction); 5103 break; 5104 } 5105 5106 case Primitive::kPrimDouble: { 5107 uint32_t offset = mirror::Array::DataOffset(sizeof(double)).Uint32Value(); 5108 Address address = index.IsConstant() 5109 ? Address(array, (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + offset) 5110 : Address(array, index.AsRegister<Register>(), TIMES_8, offset); 5111 if (value.IsFpuRegister()) { 5112 __ movsd(address, value.AsFpuRegister<XmmRegister>()); 5113 } else { 5114 DCHECK(value.IsConstant()); 5115 Address address_hi = index.IsConstant() ? 5116 Address(array, (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + 5117 offset + kX86WordSize) : 5118 Address(array, index.AsRegister<Register>(), TIMES_8, offset + kX86WordSize); 5119 int64_t v = bit_cast<int64_t, double>(value.GetConstant()->AsDoubleConstant()->GetValue()); 5120 __ movl(address, Immediate(Low32Bits(v))); 5121 codegen_->MaybeRecordImplicitNullCheck(instruction); 5122 __ movl(address_hi, Immediate(High32Bits(v))); 5123 } 5124 break; 5125 } 5126 5127 case Primitive::kPrimVoid: 5128 LOG(FATAL) << "Unreachable type " << instruction->GetType(); 5129 UNREACHABLE(); 5130 } 5131} 5132 5133void LocationsBuilderX86::VisitArrayLength(HArrayLength* instruction) { 5134 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction); 5135 locations->SetInAt(0, Location::RequiresRegister()); 5136 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 5137} 5138 5139void InstructionCodeGeneratorX86::VisitArrayLength(HArrayLength* instruction) { 5140 LocationSummary* locations = instruction->GetLocations(); 5141 uint32_t offset = mirror::Array::LengthOffset().Uint32Value(); 5142 Register obj = locations->InAt(0).AsRegister<Register>(); 5143 Register out = locations->Out().AsRegister<Register>(); 5144 __ movl(out, Address(obj, offset)); 5145 codegen_->MaybeRecordImplicitNullCheck(instruction); 5146} 5147 5148void LocationsBuilderX86::VisitBoundsCheck(HBoundsCheck* instruction) { 5149 LocationSummary::CallKind call_kind = instruction->CanThrowIntoCatchBlock() 5150 ? LocationSummary::kCallOnSlowPath 5151 : LocationSummary::kNoCall; 5152 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 5153 locations->SetInAt(0, Location::RegisterOrConstant(instruction->InputAt(0))); 5154 locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1))); 5155 if (instruction->HasUses()) { 5156 locations->SetOut(Location::SameAsFirstInput()); 5157 } 5158} 5159 5160void InstructionCodeGeneratorX86::VisitBoundsCheck(HBoundsCheck* instruction) { 5161 LocationSummary* locations = instruction->GetLocations(); 5162 Location index_loc = locations->InAt(0); 5163 Location length_loc = locations->InAt(1); 5164 SlowPathCode* slow_path = 5165 new (GetGraph()->GetArena()) BoundsCheckSlowPathX86(instruction); 5166 5167 if (length_loc.IsConstant()) { 5168 int32_t length = CodeGenerator::GetInt32ValueOf(length_loc.GetConstant()); 5169 if (index_loc.IsConstant()) { 5170 // BCE will remove the bounds check if we are guarenteed to pass. 5171 int32_t index = CodeGenerator::GetInt32ValueOf(index_loc.GetConstant()); 5172 if (index < 0 || index >= length) { 5173 codegen_->AddSlowPath(slow_path); 5174 __ jmp(slow_path->GetEntryLabel()); 5175 } else { 5176 // Some optimization after BCE may have generated this, and we should not 5177 // generate a bounds check if it is a valid range. 5178 } 5179 return; 5180 } 5181 5182 // We have to reverse the jump condition because the length is the constant. 5183 Register index_reg = index_loc.AsRegister<Register>(); 5184 __ cmpl(index_reg, Immediate(length)); 5185 codegen_->AddSlowPath(slow_path); 5186 __ j(kAboveEqual, slow_path->GetEntryLabel()); 5187 } else { 5188 Register length = length_loc.AsRegister<Register>(); 5189 if (index_loc.IsConstant()) { 5190 int32_t value = CodeGenerator::GetInt32ValueOf(index_loc.GetConstant()); 5191 __ cmpl(length, Immediate(value)); 5192 } else { 5193 __ cmpl(length, index_loc.AsRegister<Register>()); 5194 } 5195 codegen_->AddSlowPath(slow_path); 5196 __ j(kBelowEqual, slow_path->GetEntryLabel()); 5197 } 5198} 5199 5200void LocationsBuilderX86::VisitTemporary(HTemporary* temp) { 5201 temp->SetLocations(nullptr); 5202} 5203 5204void InstructionCodeGeneratorX86::VisitTemporary(HTemporary* temp ATTRIBUTE_UNUSED) { 5205 // Nothing to do, this is driven by the code generator. 5206} 5207 5208void LocationsBuilderX86::VisitParallelMove(HParallelMove* instruction ATTRIBUTE_UNUSED) { 5209 LOG(FATAL) << "Unreachable"; 5210} 5211 5212void InstructionCodeGeneratorX86::VisitParallelMove(HParallelMove* instruction) { 5213 codegen_->GetMoveResolver()->EmitNativeCode(instruction); 5214} 5215 5216void LocationsBuilderX86::VisitSuspendCheck(HSuspendCheck* instruction) { 5217 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnSlowPath); 5218} 5219 5220void InstructionCodeGeneratorX86::VisitSuspendCheck(HSuspendCheck* instruction) { 5221 HBasicBlock* block = instruction->GetBlock(); 5222 if (block->GetLoopInformation() != nullptr) { 5223 DCHECK(block->GetLoopInformation()->GetSuspendCheck() == instruction); 5224 // The back edge will generate the suspend check. 5225 return; 5226 } 5227 if (block->IsEntryBlock() && instruction->GetNext()->IsGoto()) { 5228 // The goto will generate the suspend check. 5229 return; 5230 } 5231 GenerateSuspendCheck(instruction, nullptr); 5232} 5233 5234void InstructionCodeGeneratorX86::GenerateSuspendCheck(HSuspendCheck* instruction, 5235 HBasicBlock* successor) { 5236 SuspendCheckSlowPathX86* slow_path = 5237 down_cast<SuspendCheckSlowPathX86*>(instruction->GetSlowPath()); 5238 if (slow_path == nullptr) { 5239 slow_path = new (GetGraph()->GetArena()) SuspendCheckSlowPathX86(instruction, successor); 5240 instruction->SetSlowPath(slow_path); 5241 codegen_->AddSlowPath(slow_path); 5242 if (successor != nullptr) { 5243 DCHECK(successor->IsLoopHeader()); 5244 codegen_->ClearSpillSlotsFromLoopPhisInStackMap(instruction); 5245 } 5246 } else { 5247 DCHECK_EQ(slow_path->GetSuccessor(), successor); 5248 } 5249 5250 __ fs()->cmpw(Address::Absolute( 5251 Thread::ThreadFlagsOffset<kX86WordSize>().Int32Value()), Immediate(0)); 5252 if (successor == nullptr) { 5253 __ j(kNotEqual, slow_path->GetEntryLabel()); 5254 __ Bind(slow_path->GetReturnLabel()); 5255 } else { 5256 __ j(kEqual, codegen_->GetLabelOf(successor)); 5257 __ jmp(slow_path->GetEntryLabel()); 5258 } 5259} 5260 5261X86Assembler* ParallelMoveResolverX86::GetAssembler() const { 5262 return codegen_->GetAssembler(); 5263} 5264 5265void ParallelMoveResolverX86::MoveMemoryToMemory32(int dst, int src) { 5266 ScratchRegisterScope ensure_scratch( 5267 this, kNoRegister, EAX, codegen_->GetNumberOfCoreRegisters()); 5268 Register temp_reg = static_cast<Register>(ensure_scratch.GetRegister()); 5269 int stack_offset = ensure_scratch.IsSpilled() ? kX86WordSize : 0; 5270 __ movl(temp_reg, Address(ESP, src + stack_offset)); 5271 __ movl(Address(ESP, dst + stack_offset), temp_reg); 5272} 5273 5274void ParallelMoveResolverX86::MoveMemoryToMemory64(int dst, int src) { 5275 ScratchRegisterScope ensure_scratch( 5276 this, kNoRegister, EAX, codegen_->GetNumberOfCoreRegisters()); 5277 Register temp_reg = static_cast<Register>(ensure_scratch.GetRegister()); 5278 int stack_offset = ensure_scratch.IsSpilled() ? kX86WordSize : 0; 5279 __ movl(temp_reg, Address(ESP, src + stack_offset)); 5280 __ movl(Address(ESP, dst + stack_offset), temp_reg); 5281 __ movl(temp_reg, Address(ESP, src + stack_offset + kX86WordSize)); 5282 __ movl(Address(ESP, dst + stack_offset + kX86WordSize), temp_reg); 5283} 5284 5285void ParallelMoveResolverX86::EmitMove(size_t index) { 5286 MoveOperands* move = moves_[index]; 5287 Location source = move->GetSource(); 5288 Location destination = move->GetDestination(); 5289 5290 if (source.IsRegister()) { 5291 if (destination.IsRegister()) { 5292 __ movl(destination.AsRegister<Register>(), source.AsRegister<Register>()); 5293 } else { 5294 DCHECK(destination.IsStackSlot()); 5295 __ movl(Address(ESP, destination.GetStackIndex()), source.AsRegister<Register>()); 5296 } 5297 } else if (source.IsFpuRegister()) { 5298 if (destination.IsFpuRegister()) { 5299 __ movaps(destination.AsFpuRegister<XmmRegister>(), source.AsFpuRegister<XmmRegister>()); 5300 } else if (destination.IsStackSlot()) { 5301 __ movss(Address(ESP, destination.GetStackIndex()), source.AsFpuRegister<XmmRegister>()); 5302 } else { 5303 DCHECK(destination.IsDoubleStackSlot()); 5304 __ movsd(Address(ESP, destination.GetStackIndex()), source.AsFpuRegister<XmmRegister>()); 5305 } 5306 } else if (source.IsStackSlot()) { 5307 if (destination.IsRegister()) { 5308 __ movl(destination.AsRegister<Register>(), Address(ESP, source.GetStackIndex())); 5309 } else if (destination.IsFpuRegister()) { 5310 __ movss(destination.AsFpuRegister<XmmRegister>(), Address(ESP, source.GetStackIndex())); 5311 } else { 5312 DCHECK(destination.IsStackSlot()); 5313 MoveMemoryToMemory32(destination.GetStackIndex(), source.GetStackIndex()); 5314 } 5315 } else if (source.IsDoubleStackSlot()) { 5316 if (destination.IsFpuRegister()) { 5317 __ movsd(destination.AsFpuRegister<XmmRegister>(), Address(ESP, source.GetStackIndex())); 5318 } else { 5319 DCHECK(destination.IsDoubleStackSlot()) << destination; 5320 MoveMemoryToMemory64(destination.GetStackIndex(), source.GetStackIndex()); 5321 } 5322 } else if (source.IsConstant()) { 5323 HConstant* constant = source.GetConstant(); 5324 if (constant->IsIntConstant() || constant->IsNullConstant()) { 5325 int32_t value = CodeGenerator::GetInt32ValueOf(constant); 5326 if (destination.IsRegister()) { 5327 if (value == 0) { 5328 __ xorl(destination.AsRegister<Register>(), destination.AsRegister<Register>()); 5329 } else { 5330 __ movl(destination.AsRegister<Register>(), Immediate(value)); 5331 } 5332 } else { 5333 DCHECK(destination.IsStackSlot()) << destination; 5334 __ movl(Address(ESP, destination.GetStackIndex()), Immediate(value)); 5335 } 5336 } else if (constant->IsFloatConstant()) { 5337 float fp_value = constant->AsFloatConstant()->GetValue(); 5338 int32_t value = bit_cast<int32_t, float>(fp_value); 5339 Immediate imm(value); 5340 if (destination.IsFpuRegister()) { 5341 XmmRegister dest = destination.AsFpuRegister<XmmRegister>(); 5342 if (value == 0) { 5343 // Easy handling of 0.0. 5344 __ xorps(dest, dest); 5345 } else { 5346 ScratchRegisterScope ensure_scratch( 5347 this, kNoRegister, EAX, codegen_->GetNumberOfCoreRegisters()); 5348 Register temp = static_cast<Register>(ensure_scratch.GetRegister()); 5349 __ movl(temp, Immediate(value)); 5350 __ movd(dest, temp); 5351 } 5352 } else { 5353 DCHECK(destination.IsStackSlot()) << destination; 5354 __ movl(Address(ESP, destination.GetStackIndex()), imm); 5355 } 5356 } else if (constant->IsLongConstant()) { 5357 int64_t value = constant->AsLongConstant()->GetValue(); 5358 int32_t low_value = Low32Bits(value); 5359 int32_t high_value = High32Bits(value); 5360 Immediate low(low_value); 5361 Immediate high(high_value); 5362 if (destination.IsDoubleStackSlot()) { 5363 __ movl(Address(ESP, destination.GetStackIndex()), low); 5364 __ movl(Address(ESP, destination.GetHighStackIndex(kX86WordSize)), high); 5365 } else { 5366 __ movl(destination.AsRegisterPairLow<Register>(), low); 5367 __ movl(destination.AsRegisterPairHigh<Register>(), high); 5368 } 5369 } else { 5370 DCHECK(constant->IsDoubleConstant()); 5371 double dbl_value = constant->AsDoubleConstant()->GetValue(); 5372 int64_t value = bit_cast<int64_t, double>(dbl_value); 5373 int32_t low_value = Low32Bits(value); 5374 int32_t high_value = High32Bits(value); 5375 Immediate low(low_value); 5376 Immediate high(high_value); 5377 if (destination.IsFpuRegister()) { 5378 XmmRegister dest = destination.AsFpuRegister<XmmRegister>(); 5379 if (value == 0) { 5380 // Easy handling of 0.0. 5381 __ xorpd(dest, dest); 5382 } else { 5383 __ pushl(high); 5384 __ pushl(low); 5385 __ movsd(dest, Address(ESP, 0)); 5386 __ addl(ESP, Immediate(8)); 5387 } 5388 } else { 5389 DCHECK(destination.IsDoubleStackSlot()) << destination; 5390 __ movl(Address(ESP, destination.GetStackIndex()), low); 5391 __ movl(Address(ESP, destination.GetHighStackIndex(kX86WordSize)), high); 5392 } 5393 } 5394 } else { 5395 LOG(FATAL) << "Unimplemented move: " << destination << " <- " << source; 5396 } 5397} 5398 5399void ParallelMoveResolverX86::Exchange(Register reg, int mem) { 5400 Register suggested_scratch = reg == EAX ? EBX : EAX; 5401 ScratchRegisterScope ensure_scratch( 5402 this, reg, suggested_scratch, codegen_->GetNumberOfCoreRegisters()); 5403 5404 int stack_offset = ensure_scratch.IsSpilled() ? kX86WordSize : 0; 5405 __ movl(static_cast<Register>(ensure_scratch.GetRegister()), Address(ESP, mem + stack_offset)); 5406 __ movl(Address(ESP, mem + stack_offset), reg); 5407 __ movl(reg, static_cast<Register>(ensure_scratch.GetRegister())); 5408} 5409 5410void ParallelMoveResolverX86::Exchange32(XmmRegister reg, int mem) { 5411 ScratchRegisterScope ensure_scratch( 5412 this, kNoRegister, EAX, codegen_->GetNumberOfCoreRegisters()); 5413 5414 Register temp_reg = static_cast<Register>(ensure_scratch.GetRegister()); 5415 int stack_offset = ensure_scratch.IsSpilled() ? kX86WordSize : 0; 5416 __ movl(temp_reg, Address(ESP, mem + stack_offset)); 5417 __ movss(Address(ESP, mem + stack_offset), reg); 5418 __ movd(reg, temp_reg); 5419} 5420 5421void ParallelMoveResolverX86::Exchange(int mem1, int mem2) { 5422 ScratchRegisterScope ensure_scratch1( 5423 this, kNoRegister, EAX, codegen_->GetNumberOfCoreRegisters()); 5424 5425 Register suggested_scratch = ensure_scratch1.GetRegister() == EAX ? EBX : EAX; 5426 ScratchRegisterScope ensure_scratch2( 5427 this, ensure_scratch1.GetRegister(), suggested_scratch, codegen_->GetNumberOfCoreRegisters()); 5428 5429 int stack_offset = ensure_scratch1.IsSpilled() ? kX86WordSize : 0; 5430 stack_offset += ensure_scratch2.IsSpilled() ? kX86WordSize : 0; 5431 __ movl(static_cast<Register>(ensure_scratch1.GetRegister()), Address(ESP, mem1 + stack_offset)); 5432 __ movl(static_cast<Register>(ensure_scratch2.GetRegister()), Address(ESP, mem2 + stack_offset)); 5433 __ movl(Address(ESP, mem2 + stack_offset), static_cast<Register>(ensure_scratch1.GetRegister())); 5434 __ movl(Address(ESP, mem1 + stack_offset), static_cast<Register>(ensure_scratch2.GetRegister())); 5435} 5436 5437void ParallelMoveResolverX86::EmitSwap(size_t index) { 5438 MoveOperands* move = moves_[index]; 5439 Location source = move->GetSource(); 5440 Location destination = move->GetDestination(); 5441 5442 if (source.IsRegister() && destination.IsRegister()) { 5443 // Use XOR swap algorithm to avoid serializing XCHG instruction or using a temporary. 5444 DCHECK_NE(destination.AsRegister<Register>(), source.AsRegister<Register>()); 5445 __ xorl(destination.AsRegister<Register>(), source.AsRegister<Register>()); 5446 __ xorl(source.AsRegister<Register>(), destination.AsRegister<Register>()); 5447 __ xorl(destination.AsRegister<Register>(), source.AsRegister<Register>()); 5448 } else if (source.IsRegister() && destination.IsStackSlot()) { 5449 Exchange(source.AsRegister<Register>(), destination.GetStackIndex()); 5450 } else if (source.IsStackSlot() && destination.IsRegister()) { 5451 Exchange(destination.AsRegister<Register>(), source.GetStackIndex()); 5452 } else if (source.IsStackSlot() && destination.IsStackSlot()) { 5453 Exchange(destination.GetStackIndex(), source.GetStackIndex()); 5454 } else if (source.IsFpuRegister() && destination.IsFpuRegister()) { 5455 // Use XOR Swap algorithm to avoid a temporary. 5456 DCHECK_NE(source.reg(), destination.reg()); 5457 __ xorpd(destination.AsFpuRegister<XmmRegister>(), source.AsFpuRegister<XmmRegister>()); 5458 __ xorpd(source.AsFpuRegister<XmmRegister>(), destination.AsFpuRegister<XmmRegister>()); 5459 __ xorpd(destination.AsFpuRegister<XmmRegister>(), source.AsFpuRegister<XmmRegister>()); 5460 } else if (source.IsFpuRegister() && destination.IsStackSlot()) { 5461 Exchange32(source.AsFpuRegister<XmmRegister>(), destination.GetStackIndex()); 5462 } else if (destination.IsFpuRegister() && source.IsStackSlot()) { 5463 Exchange32(destination.AsFpuRegister<XmmRegister>(), source.GetStackIndex()); 5464 } else if (source.IsFpuRegister() && destination.IsDoubleStackSlot()) { 5465 // Take advantage of the 16 bytes in the XMM register. 5466 XmmRegister reg = source.AsFpuRegister<XmmRegister>(); 5467 Address stack(ESP, destination.GetStackIndex()); 5468 // Load the double into the high doubleword. 5469 __ movhpd(reg, stack); 5470 5471 // Store the low double into the destination. 5472 __ movsd(stack, reg); 5473 5474 // Move the high double to the low double. 5475 __ psrldq(reg, Immediate(8)); 5476 } else if (destination.IsFpuRegister() && source.IsDoubleStackSlot()) { 5477 // Take advantage of the 16 bytes in the XMM register. 5478 XmmRegister reg = destination.AsFpuRegister<XmmRegister>(); 5479 Address stack(ESP, source.GetStackIndex()); 5480 // Load the double into the high doubleword. 5481 __ movhpd(reg, stack); 5482 5483 // Store the low double into the destination. 5484 __ movsd(stack, reg); 5485 5486 // Move the high double to the low double. 5487 __ psrldq(reg, Immediate(8)); 5488 } else if (destination.IsDoubleStackSlot() && source.IsDoubleStackSlot()) { 5489 Exchange(destination.GetStackIndex(), source.GetStackIndex()); 5490 Exchange(destination.GetHighStackIndex(kX86WordSize), source.GetHighStackIndex(kX86WordSize)); 5491 } else { 5492 LOG(FATAL) << "Unimplemented: source: " << source << ", destination: " << destination; 5493 } 5494} 5495 5496void ParallelMoveResolverX86::SpillScratch(int reg) { 5497 __ pushl(static_cast<Register>(reg)); 5498} 5499 5500void ParallelMoveResolverX86::RestoreScratch(int reg) { 5501 __ popl(static_cast<Register>(reg)); 5502} 5503 5504void LocationsBuilderX86::VisitLoadClass(HLoadClass* cls) { 5505 InvokeRuntimeCallingConvention calling_convention; 5506 CodeGenerator::CreateLoadClassLocationSummary( 5507 cls, 5508 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 5509 Location::RegisterLocation(EAX), 5510 /* code_generator_supports_read_barrier */ true); 5511} 5512 5513void InstructionCodeGeneratorX86::VisitLoadClass(HLoadClass* cls) { 5514 LocationSummary* locations = cls->GetLocations(); 5515 if (cls->NeedsAccessCheck()) { 5516 codegen_->MoveConstant(locations->GetTemp(0), cls->GetTypeIndex()); 5517 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pInitializeTypeAndVerifyAccess), 5518 cls, 5519 cls->GetDexPc(), 5520 nullptr); 5521 CheckEntrypointTypes<kQuickInitializeTypeAndVerifyAccess, void*, uint32_t>(); 5522 return; 5523 } 5524 5525 Location out_loc = locations->Out(); 5526 Register out = out_loc.AsRegister<Register>(); 5527 Register current_method = locations->InAt(0).AsRegister<Register>(); 5528 5529 if (cls->IsReferrersClass()) { 5530 DCHECK(!cls->CanCallRuntime()); 5531 DCHECK(!cls->MustGenerateClinitCheck()); 5532 uint32_t declaring_class_offset = ArtMethod::DeclaringClassOffset().Int32Value(); 5533 if (kEmitCompilerReadBarrier) { 5534 // /* GcRoot<mirror::Class>* */ out = &(current_method->declaring_class_) 5535 __ leal(out, Address(current_method, declaring_class_offset)); 5536 // /* mirror::Class* */ out = out->Read() 5537 codegen_->GenerateReadBarrierForRoot(cls, out_loc, out_loc); 5538 } else { 5539 // /* GcRoot<mirror::Class> */ out = current_method->declaring_class_ 5540 __ movl(out, Address(current_method, declaring_class_offset)); 5541 } 5542 } else { 5543 // /* GcRoot<mirror::Class>[] */ out = 5544 // current_method.ptr_sized_fields_->dex_cache_resolved_types_ 5545 __ movl(out, Address(current_method, 5546 ArtMethod::DexCacheResolvedTypesOffset(kX86PointerSize).Int32Value())); 5547 5548 size_t cache_offset = CodeGenerator::GetCacheOffset(cls->GetTypeIndex()); 5549 if (kEmitCompilerReadBarrier) { 5550 // /* GcRoot<mirror::Class>* */ out = &out[type_index] 5551 __ leal(out, Address(out, cache_offset)); 5552 // /* mirror::Class* */ out = out->Read() 5553 codegen_->GenerateReadBarrierForRoot(cls, out_loc, out_loc); 5554 } else { 5555 // /* GcRoot<mirror::Class> */ out = out[type_index] 5556 __ movl(out, Address(out, cache_offset)); 5557 } 5558 5559 if (!cls->IsInDexCache() || cls->MustGenerateClinitCheck()) { 5560 DCHECK(cls->CanCallRuntime()); 5561 SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathX86( 5562 cls, cls, cls->GetDexPc(), cls->MustGenerateClinitCheck()); 5563 codegen_->AddSlowPath(slow_path); 5564 5565 if (!cls->IsInDexCache()) { 5566 __ testl(out, out); 5567 __ j(kEqual, slow_path->GetEntryLabel()); 5568 } 5569 5570 if (cls->MustGenerateClinitCheck()) { 5571 GenerateClassInitializationCheck(slow_path, out); 5572 } else { 5573 __ Bind(slow_path->GetExitLabel()); 5574 } 5575 } 5576 } 5577} 5578 5579void LocationsBuilderX86::VisitClinitCheck(HClinitCheck* check) { 5580 LocationSummary* locations = 5581 new (GetGraph()->GetArena()) LocationSummary(check, LocationSummary::kCallOnSlowPath); 5582 locations->SetInAt(0, Location::RequiresRegister()); 5583 if (check->HasUses()) { 5584 locations->SetOut(Location::SameAsFirstInput()); 5585 } 5586} 5587 5588void InstructionCodeGeneratorX86::VisitClinitCheck(HClinitCheck* check) { 5589 // We assume the class to not be null. 5590 SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathX86( 5591 check->GetLoadClass(), check, check->GetDexPc(), true); 5592 codegen_->AddSlowPath(slow_path); 5593 GenerateClassInitializationCheck(slow_path, 5594 check->GetLocations()->InAt(0).AsRegister<Register>()); 5595} 5596 5597void InstructionCodeGeneratorX86::GenerateClassInitializationCheck( 5598 SlowPathCode* slow_path, Register class_reg) { 5599 __ cmpl(Address(class_reg, mirror::Class::StatusOffset().Int32Value()), 5600 Immediate(mirror::Class::kStatusInitialized)); 5601 __ j(kLess, slow_path->GetEntryLabel()); 5602 __ Bind(slow_path->GetExitLabel()); 5603 // No need for memory fence, thanks to the X86 memory model. 5604} 5605 5606void LocationsBuilderX86::VisitLoadString(HLoadString* load) { 5607 LocationSummary* locations = 5608 new (GetGraph()->GetArena()) LocationSummary(load, LocationSummary::kCallOnSlowPath); 5609 locations->SetInAt(0, Location::RequiresRegister()); 5610 locations->SetOut(Location::RequiresRegister()); 5611} 5612 5613void InstructionCodeGeneratorX86::VisitLoadString(HLoadString* load) { 5614 SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadStringSlowPathX86(load); 5615 codegen_->AddSlowPath(slow_path); 5616 5617 LocationSummary* locations = load->GetLocations(); 5618 Location out_loc = locations->Out(); 5619 Register out = out_loc.AsRegister<Register>(); 5620 Register current_method = locations->InAt(0).AsRegister<Register>(); 5621 5622 uint32_t declaring_class_offset = ArtMethod::DeclaringClassOffset().Int32Value(); 5623 if (kEmitCompilerReadBarrier) { 5624 // /* GcRoot<mirror::Class>* */ out = &(current_method->declaring_class_) 5625 __ leal(out, Address(current_method, declaring_class_offset)); 5626 // /* mirror::Class* */ out = out->Read() 5627 codegen_->GenerateReadBarrierForRoot(load, out_loc, out_loc); 5628 } else { 5629 // /* GcRoot<mirror::Class> */ out = current_method->declaring_class_ 5630 __ movl(out, Address(current_method, declaring_class_offset)); 5631 } 5632 5633 // /* GcRoot<mirror::String>[] */ out = out->dex_cache_strings_ 5634 __ movl(out, Address(out, mirror::Class::DexCacheStringsOffset().Int32Value())); 5635 5636 size_t cache_offset = CodeGenerator::GetCacheOffset(load->GetStringIndex()); 5637 if (kEmitCompilerReadBarrier) { 5638 // /* GcRoot<mirror::String>* */ out = &out[string_index] 5639 __ leal(out, Address(out, cache_offset)); 5640 // /* mirror::String* */ out = out->Read() 5641 codegen_->GenerateReadBarrierForRoot(load, out_loc, out_loc); 5642 } else { 5643 // /* GcRoot<mirror::String> */ out = out[string_index] 5644 __ movl(out, Address(out, cache_offset)); 5645 } 5646 5647 __ testl(out, out); 5648 __ j(kEqual, slow_path->GetEntryLabel()); 5649 __ Bind(slow_path->GetExitLabel()); 5650} 5651 5652static Address GetExceptionTlsAddress() { 5653 return Address::Absolute(Thread::ExceptionOffset<kX86WordSize>().Int32Value()); 5654} 5655 5656void LocationsBuilderX86::VisitLoadException(HLoadException* load) { 5657 LocationSummary* locations = 5658 new (GetGraph()->GetArena()) LocationSummary(load, LocationSummary::kNoCall); 5659 locations->SetOut(Location::RequiresRegister()); 5660} 5661 5662void InstructionCodeGeneratorX86::VisitLoadException(HLoadException* load) { 5663 __ fs()->movl(load->GetLocations()->Out().AsRegister<Register>(), GetExceptionTlsAddress()); 5664} 5665 5666void LocationsBuilderX86::VisitClearException(HClearException* clear) { 5667 new (GetGraph()->GetArena()) LocationSummary(clear, LocationSummary::kNoCall); 5668} 5669 5670void InstructionCodeGeneratorX86::VisitClearException(HClearException* clear ATTRIBUTE_UNUSED) { 5671 __ fs()->movl(GetExceptionTlsAddress(), Immediate(0)); 5672} 5673 5674void LocationsBuilderX86::VisitThrow(HThrow* instruction) { 5675 LocationSummary* locations = 5676 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall); 5677 InvokeRuntimeCallingConvention calling_convention; 5678 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 5679} 5680 5681void InstructionCodeGeneratorX86::VisitThrow(HThrow* instruction) { 5682 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pDeliverException), 5683 instruction, 5684 instruction->GetDexPc(), 5685 nullptr); 5686 CheckEntrypointTypes<kQuickDeliverException, void, mirror::Object*>(); 5687} 5688 5689void LocationsBuilderX86::VisitInstanceOf(HInstanceOf* instruction) { 5690 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 5691 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 5692 switch (type_check_kind) { 5693 case TypeCheckKind::kExactCheck: 5694 case TypeCheckKind::kAbstractClassCheck: 5695 case TypeCheckKind::kClassHierarchyCheck: 5696 case TypeCheckKind::kArrayObjectCheck: 5697 call_kind = 5698 kEmitCompilerReadBarrier ? LocationSummary::kCallOnSlowPath : LocationSummary::kNoCall; 5699 break; 5700 case TypeCheckKind::kArrayCheck: 5701 case TypeCheckKind::kUnresolvedCheck: 5702 case TypeCheckKind::kInterfaceCheck: 5703 call_kind = LocationSummary::kCallOnSlowPath; 5704 break; 5705 } 5706 5707 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 5708 locations->SetInAt(0, Location::RequiresRegister()); 5709 locations->SetInAt(1, Location::Any()); 5710 // Note that TypeCheckSlowPathX86 uses this "out" register too. 5711 locations->SetOut(Location::RequiresRegister()); 5712 // When read barriers are enabled, we need a temporary register for 5713 // some cases. 5714 if (kEmitCompilerReadBarrier && 5715 (type_check_kind == TypeCheckKind::kAbstractClassCheck || 5716 type_check_kind == TypeCheckKind::kClassHierarchyCheck || 5717 type_check_kind == TypeCheckKind::kArrayObjectCheck)) { 5718 locations->AddTemp(Location::RequiresRegister()); 5719 } 5720} 5721 5722void InstructionCodeGeneratorX86::VisitInstanceOf(HInstanceOf* instruction) { 5723 LocationSummary* locations = instruction->GetLocations(); 5724 Location obj_loc = locations->InAt(0); 5725 Register obj = obj_loc.AsRegister<Register>(); 5726 Location cls = locations->InAt(1); 5727 Location out_loc = locations->Out(); 5728 Register out = out_loc.AsRegister<Register>(); 5729 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 5730 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 5731 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 5732 uint32_t primitive_offset = mirror::Class::PrimitiveTypeOffset().Int32Value(); 5733 SlowPathCode* slow_path = nullptr; 5734 NearLabel done, zero; 5735 5736 // Return 0 if `obj` is null. 5737 // Avoid null check if we know obj is not null. 5738 if (instruction->MustDoNullCheck()) { 5739 __ testl(obj, obj); 5740 __ j(kEqual, &zero); 5741 } 5742 5743 // /* HeapReference<Class> */ out = obj->klass_ 5744 __ movl(out, Address(obj, class_offset)); 5745 codegen_->MaybeGenerateReadBarrier(instruction, out_loc, out_loc, obj_loc, class_offset); 5746 5747 switch (instruction->GetTypeCheckKind()) { 5748 case TypeCheckKind::kExactCheck: { 5749 if (cls.IsRegister()) { 5750 __ cmpl(out, cls.AsRegister<Register>()); 5751 } else { 5752 DCHECK(cls.IsStackSlot()) << cls; 5753 __ cmpl(out, Address(ESP, cls.GetStackIndex())); 5754 } 5755 5756 // Classes must be equal for the instanceof to succeed. 5757 __ j(kNotEqual, &zero); 5758 __ movl(out, Immediate(1)); 5759 __ jmp(&done); 5760 break; 5761 } 5762 5763 case TypeCheckKind::kAbstractClassCheck: { 5764 // If the class is abstract, we eagerly fetch the super class of the 5765 // object to avoid doing a comparison we know will fail. 5766 NearLabel loop; 5767 __ Bind(&loop); 5768 Location temp_loc = kEmitCompilerReadBarrier ? locations->GetTemp(0) : Location::NoLocation(); 5769 if (kEmitCompilerReadBarrier) { 5770 // Save the value of `out` into `temp` before overwriting it 5771 // in the following move operation, as we will need it for the 5772 // read barrier below. 5773 Register temp = temp_loc.AsRegister<Register>(); 5774 __ movl(temp, out); 5775 } 5776 // /* HeapReference<Class> */ out = out->super_class_ 5777 __ movl(out, Address(out, super_offset)); 5778 codegen_->MaybeGenerateReadBarrier(instruction, out_loc, out_loc, temp_loc, super_offset); 5779 __ testl(out, out); 5780 // If `out` is null, we use it for the result, and jump to `done`. 5781 __ j(kEqual, &done); 5782 if (cls.IsRegister()) { 5783 __ cmpl(out, cls.AsRegister<Register>()); 5784 } else { 5785 DCHECK(cls.IsStackSlot()) << cls; 5786 __ cmpl(out, Address(ESP, cls.GetStackIndex())); 5787 } 5788 __ j(kNotEqual, &loop); 5789 __ movl(out, Immediate(1)); 5790 if (zero.IsLinked()) { 5791 __ jmp(&done); 5792 } 5793 break; 5794 } 5795 5796 case TypeCheckKind::kClassHierarchyCheck: { 5797 // Walk over the class hierarchy to find a match. 5798 NearLabel loop, success; 5799 __ Bind(&loop); 5800 if (cls.IsRegister()) { 5801 __ cmpl(out, cls.AsRegister<Register>()); 5802 } else { 5803 DCHECK(cls.IsStackSlot()) << cls; 5804 __ cmpl(out, Address(ESP, cls.GetStackIndex())); 5805 } 5806 __ j(kEqual, &success); 5807 Location temp_loc = kEmitCompilerReadBarrier ? locations->GetTemp(0) : Location::NoLocation(); 5808 if (kEmitCompilerReadBarrier) { 5809 // Save the value of `out` into `temp` before overwriting it 5810 // in the following move operation, as we will need it for the 5811 // read barrier below. 5812 Register temp = temp_loc.AsRegister<Register>(); 5813 __ movl(temp, out); 5814 } 5815 // /* HeapReference<Class> */ out = out->super_class_ 5816 __ movl(out, Address(out, super_offset)); 5817 codegen_->MaybeGenerateReadBarrier(instruction, out_loc, out_loc, temp_loc, super_offset); 5818 __ testl(out, out); 5819 __ j(kNotEqual, &loop); 5820 // If `out` is null, we use it for the result, and jump to `done`. 5821 __ jmp(&done); 5822 __ Bind(&success); 5823 __ movl(out, Immediate(1)); 5824 if (zero.IsLinked()) { 5825 __ jmp(&done); 5826 } 5827 break; 5828 } 5829 5830 case TypeCheckKind::kArrayObjectCheck: { 5831 // Do an exact check. 5832 NearLabel exact_check; 5833 if (cls.IsRegister()) { 5834 __ cmpl(out, cls.AsRegister<Register>()); 5835 } else { 5836 DCHECK(cls.IsStackSlot()) << cls; 5837 __ cmpl(out, Address(ESP, cls.GetStackIndex())); 5838 } 5839 __ j(kEqual, &exact_check); 5840 // Otherwise, we need to check that the object's class is a non-primitive array. 5841 Location temp_loc = kEmitCompilerReadBarrier ? locations->GetTemp(0) : Location::NoLocation(); 5842 if (kEmitCompilerReadBarrier) { 5843 // Save the value of `out` into `temp` before overwriting it 5844 // in the following move operation, as we will need it for the 5845 // read barrier below. 5846 Register temp = temp_loc.AsRegister<Register>(); 5847 __ movl(temp, out); 5848 } 5849 // /* HeapReference<Class> */ out = out->component_type_ 5850 __ movl(out, Address(out, component_offset)); 5851 codegen_->MaybeGenerateReadBarrier(instruction, out_loc, out_loc, temp_loc, component_offset); 5852 __ testl(out, out); 5853 // If `out` is null, we use it for the result, and jump to `done`. 5854 __ j(kEqual, &done); 5855 __ cmpw(Address(out, primitive_offset), Immediate(Primitive::kPrimNot)); 5856 __ j(kNotEqual, &zero); 5857 __ Bind(&exact_check); 5858 __ movl(out, Immediate(1)); 5859 __ jmp(&done); 5860 break; 5861 } 5862 5863 case TypeCheckKind::kArrayCheck: { 5864 if (cls.IsRegister()) { 5865 __ cmpl(out, cls.AsRegister<Register>()); 5866 } else { 5867 DCHECK(cls.IsStackSlot()) << cls; 5868 __ cmpl(out, Address(ESP, cls.GetStackIndex())); 5869 } 5870 DCHECK(locations->OnlyCallsOnSlowPath()); 5871 slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathX86(instruction, 5872 /* is_fatal */ false); 5873 codegen_->AddSlowPath(slow_path); 5874 __ j(kNotEqual, slow_path->GetEntryLabel()); 5875 __ movl(out, Immediate(1)); 5876 if (zero.IsLinked()) { 5877 __ jmp(&done); 5878 } 5879 break; 5880 } 5881 5882 case TypeCheckKind::kUnresolvedCheck: 5883 case TypeCheckKind::kInterfaceCheck: { 5884 // Note that we indeed only call on slow path, but we always go 5885 // into the slow path for the unresolved & interface check 5886 // cases. 5887 // 5888 // We cannot directly call the InstanceofNonTrivial runtime 5889 // entry point without resorting to a type checking slow path 5890 // here (i.e. by calling InvokeRuntime directly), as it would 5891 // require to assign fixed registers for the inputs of this 5892 // HInstanceOf instruction (following the runtime calling 5893 // convention), which might be cluttered by the potential first 5894 // read barrier emission at the beginning of this method. 5895 DCHECK(locations->OnlyCallsOnSlowPath()); 5896 slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathX86(instruction, 5897 /* is_fatal */ false); 5898 codegen_->AddSlowPath(slow_path); 5899 __ jmp(slow_path->GetEntryLabel()); 5900 if (zero.IsLinked()) { 5901 __ jmp(&done); 5902 } 5903 break; 5904 } 5905 } 5906 5907 if (zero.IsLinked()) { 5908 __ Bind(&zero); 5909 __ xorl(out, out); 5910 } 5911 5912 if (done.IsLinked()) { 5913 __ Bind(&done); 5914 } 5915 5916 if (slow_path != nullptr) { 5917 __ Bind(slow_path->GetExitLabel()); 5918 } 5919} 5920 5921void LocationsBuilderX86::VisitCheckCast(HCheckCast* instruction) { 5922 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 5923 bool throws_into_catch = instruction->CanThrowIntoCatchBlock(); 5924 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 5925 switch (type_check_kind) { 5926 case TypeCheckKind::kExactCheck: 5927 case TypeCheckKind::kAbstractClassCheck: 5928 case TypeCheckKind::kClassHierarchyCheck: 5929 case TypeCheckKind::kArrayObjectCheck: 5930 call_kind = (throws_into_catch || kEmitCompilerReadBarrier) ? 5931 LocationSummary::kCallOnSlowPath : 5932 LocationSummary::kNoCall; // In fact, call on a fatal (non-returning) slow path. 5933 break; 5934 case TypeCheckKind::kArrayCheck: 5935 case TypeCheckKind::kUnresolvedCheck: 5936 case TypeCheckKind::kInterfaceCheck: 5937 call_kind = LocationSummary::kCallOnSlowPath; 5938 break; 5939 } 5940 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 5941 locations->SetInAt(0, Location::RequiresRegister()); 5942 locations->SetInAt(1, Location::Any()); 5943 // Note that TypeCheckSlowPathX86 uses this "temp" register too. 5944 locations->AddTemp(Location::RequiresRegister()); 5945 // When read barriers are enabled, we need an additional temporary 5946 // register for some cases. 5947 if (kEmitCompilerReadBarrier && 5948 (type_check_kind == TypeCheckKind::kAbstractClassCheck || 5949 type_check_kind == TypeCheckKind::kClassHierarchyCheck || 5950 type_check_kind == TypeCheckKind::kArrayObjectCheck)) { 5951 locations->AddTemp(Location::RequiresRegister()); 5952 } 5953} 5954 5955void InstructionCodeGeneratorX86::VisitCheckCast(HCheckCast* instruction) { 5956 LocationSummary* locations = instruction->GetLocations(); 5957 Location obj_loc = locations->InAt(0); 5958 Register obj = obj_loc.AsRegister<Register>(); 5959 Location cls = locations->InAt(1); 5960 Location temp_loc = locations->GetTemp(0); 5961 Register temp = temp_loc.AsRegister<Register>(); 5962 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 5963 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 5964 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 5965 uint32_t primitive_offset = mirror::Class::PrimitiveTypeOffset().Int32Value(); 5966 5967 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 5968 bool is_type_check_slow_path_fatal = 5969 (type_check_kind == TypeCheckKind::kExactCheck || 5970 type_check_kind == TypeCheckKind::kAbstractClassCheck || 5971 type_check_kind == TypeCheckKind::kClassHierarchyCheck || 5972 type_check_kind == TypeCheckKind::kArrayObjectCheck) && 5973 !instruction->CanThrowIntoCatchBlock(); 5974 SlowPathCode* type_check_slow_path = 5975 new (GetGraph()->GetArena()) TypeCheckSlowPathX86(instruction, 5976 is_type_check_slow_path_fatal); 5977 codegen_->AddSlowPath(type_check_slow_path); 5978 5979 NearLabel done; 5980 // Avoid null check if we know obj is not null. 5981 if (instruction->MustDoNullCheck()) { 5982 __ testl(obj, obj); 5983 __ j(kEqual, &done); 5984 } 5985 5986 // /* HeapReference<Class> */ temp = obj->klass_ 5987 __ movl(temp, Address(obj, class_offset)); 5988 codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset); 5989 5990 switch (type_check_kind) { 5991 case TypeCheckKind::kExactCheck: 5992 case TypeCheckKind::kArrayCheck: { 5993 if (cls.IsRegister()) { 5994 __ cmpl(temp, cls.AsRegister<Register>()); 5995 } else { 5996 DCHECK(cls.IsStackSlot()) << cls; 5997 __ cmpl(temp, Address(ESP, cls.GetStackIndex())); 5998 } 5999 // Jump to slow path for throwing the exception or doing a 6000 // more involved array check. 6001 __ j(kNotEqual, type_check_slow_path->GetEntryLabel()); 6002 break; 6003 } 6004 6005 case TypeCheckKind::kAbstractClassCheck: { 6006 // If the class is abstract, we eagerly fetch the super class of the 6007 // object to avoid doing a comparison we know will fail. 6008 NearLabel loop, compare_classes; 6009 __ Bind(&loop); 6010 Location temp2_loc = 6011 kEmitCompilerReadBarrier ? locations->GetTemp(1) : Location::NoLocation(); 6012 if (kEmitCompilerReadBarrier) { 6013 // Save the value of `temp` into `temp2` before overwriting it 6014 // in the following move operation, as we will need it for the 6015 // read barrier below. 6016 Register temp2 = temp2_loc.AsRegister<Register>(); 6017 __ movl(temp2, temp); 6018 } 6019 // /* HeapReference<Class> */ temp = temp->super_class_ 6020 __ movl(temp, Address(temp, super_offset)); 6021 codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, temp2_loc, super_offset); 6022 6023 // If the class reference currently in `temp` is not null, jump 6024 // to the `compare_classes` label to compare it with the checked 6025 // class. 6026 __ testl(temp, temp); 6027 __ j(kNotEqual, &compare_classes); 6028 // Otherwise, jump to the slow path to throw the exception. 6029 // 6030 // But before, move back the object's class into `temp` before 6031 // going into the slow path, as it has been overwritten in the 6032 // meantime. 6033 // /* HeapReference<Class> */ temp = obj->klass_ 6034 __ movl(temp, Address(obj, class_offset)); 6035 codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset); 6036 __ jmp(type_check_slow_path->GetEntryLabel()); 6037 6038 __ Bind(&compare_classes); 6039 if (cls.IsRegister()) { 6040 __ cmpl(temp, cls.AsRegister<Register>()); 6041 } else { 6042 DCHECK(cls.IsStackSlot()) << cls; 6043 __ cmpl(temp, Address(ESP, cls.GetStackIndex())); 6044 } 6045 __ j(kNotEqual, &loop); 6046 break; 6047 } 6048 6049 case TypeCheckKind::kClassHierarchyCheck: { 6050 // Walk over the class hierarchy to find a match. 6051 NearLabel loop; 6052 __ Bind(&loop); 6053 if (cls.IsRegister()) { 6054 __ cmpl(temp, cls.AsRegister<Register>()); 6055 } else { 6056 DCHECK(cls.IsStackSlot()) << cls; 6057 __ cmpl(temp, Address(ESP, cls.GetStackIndex())); 6058 } 6059 __ j(kEqual, &done); 6060 6061 Location temp2_loc = 6062 kEmitCompilerReadBarrier ? locations->GetTemp(1) : Location::NoLocation(); 6063 if (kEmitCompilerReadBarrier) { 6064 // Save the value of `temp` into `temp2` before overwriting it 6065 // in the following move operation, as we will need it for the 6066 // read barrier below. 6067 Register temp2 = temp2_loc.AsRegister<Register>(); 6068 __ movl(temp2, temp); 6069 } 6070 // /* HeapReference<Class> */ temp = temp->super_class_ 6071 __ movl(temp, Address(temp, super_offset)); 6072 codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, temp2_loc, super_offset); 6073 6074 // If the class reference currently in `temp` is not null, jump 6075 // back at the beginning of the loop. 6076 __ testl(temp, temp); 6077 __ j(kNotEqual, &loop); 6078 // Otherwise, jump to the slow path to throw the exception. 6079 // 6080 // But before, move back the object's class into `temp` before 6081 // going into the slow path, as it has been overwritten in the 6082 // meantime. 6083 // /* HeapReference<Class> */ temp = obj->klass_ 6084 __ movl(temp, Address(obj, class_offset)); 6085 codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset); 6086 __ jmp(type_check_slow_path->GetEntryLabel()); 6087 break; 6088 } 6089 6090 case TypeCheckKind::kArrayObjectCheck: { 6091 // Do an exact check. 6092 NearLabel check_non_primitive_component_type; 6093 if (cls.IsRegister()) { 6094 __ cmpl(temp, cls.AsRegister<Register>()); 6095 } else { 6096 DCHECK(cls.IsStackSlot()) << cls; 6097 __ cmpl(temp, Address(ESP, cls.GetStackIndex())); 6098 } 6099 __ j(kEqual, &done); 6100 6101 // Otherwise, we need to check that the object's class is a non-primitive array. 6102 Location temp2_loc = 6103 kEmitCompilerReadBarrier ? locations->GetTemp(1) : Location::NoLocation(); 6104 if (kEmitCompilerReadBarrier) { 6105 // Save the value of `temp` into `temp2` before overwriting it 6106 // in the following move operation, as we will need it for the 6107 // read barrier below. 6108 Register temp2 = temp2_loc.AsRegister<Register>(); 6109 __ movl(temp2, temp); 6110 } 6111 // /* HeapReference<Class> */ temp = temp->component_type_ 6112 __ movl(temp, Address(temp, component_offset)); 6113 codegen_->MaybeGenerateReadBarrier( 6114 instruction, temp_loc, temp_loc, temp2_loc, component_offset); 6115 6116 // If the component type is not null (i.e. the object is indeed 6117 // an array), jump to label `check_non_primitive_component_type` 6118 // to further check that this component type is not a primitive 6119 // type. 6120 __ testl(temp, temp); 6121 __ j(kNotEqual, &check_non_primitive_component_type); 6122 // Otherwise, jump to the slow path to throw the exception. 6123 // 6124 // But before, move back the object's class into `temp` before 6125 // going into the slow path, as it has been overwritten in the 6126 // meantime. 6127 // /* HeapReference<Class> */ temp = obj->klass_ 6128 __ movl(temp, Address(obj, class_offset)); 6129 codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset); 6130 __ jmp(type_check_slow_path->GetEntryLabel()); 6131 6132 __ Bind(&check_non_primitive_component_type); 6133 __ cmpw(Address(temp, primitive_offset), Immediate(Primitive::kPrimNot)); 6134 __ j(kEqual, &done); 6135 // Same comment as above regarding `temp` and the slow path. 6136 // /* HeapReference<Class> */ temp = obj->klass_ 6137 __ movl(temp, Address(obj, class_offset)); 6138 codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset); 6139 __ jmp(type_check_slow_path->GetEntryLabel()); 6140 break; 6141 } 6142 6143 case TypeCheckKind::kUnresolvedCheck: 6144 case TypeCheckKind::kInterfaceCheck: 6145 // We always go into the type check slow path for the unresolved & 6146 // interface check cases. 6147 // 6148 // We cannot directly call the CheckCast runtime entry point 6149 // without resorting to a type checking slow path here (i.e. by 6150 // calling InvokeRuntime directly), as it would require to 6151 // assign fixed registers for the inputs of this HInstanceOf 6152 // instruction (following the runtime calling convention), which 6153 // might be cluttered by the potential first read barrier 6154 // emission at the beginning of this method. 6155 __ jmp(type_check_slow_path->GetEntryLabel()); 6156 break; 6157 } 6158 __ Bind(&done); 6159 6160 __ Bind(type_check_slow_path->GetExitLabel()); 6161} 6162 6163void LocationsBuilderX86::VisitMonitorOperation(HMonitorOperation* instruction) { 6164 LocationSummary* locations = 6165 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall); 6166 InvokeRuntimeCallingConvention calling_convention; 6167 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 6168} 6169 6170void InstructionCodeGeneratorX86::VisitMonitorOperation(HMonitorOperation* instruction) { 6171 codegen_->InvokeRuntime(instruction->IsEnter() ? QUICK_ENTRY_POINT(pLockObject) 6172 : QUICK_ENTRY_POINT(pUnlockObject), 6173 instruction, 6174 instruction->GetDexPc(), 6175 nullptr); 6176 if (instruction->IsEnter()) { 6177 CheckEntrypointTypes<kQuickLockObject, void, mirror::Object*>(); 6178 } else { 6179 CheckEntrypointTypes<kQuickUnlockObject, void, mirror::Object*>(); 6180 } 6181} 6182 6183void LocationsBuilderX86::VisitAnd(HAnd* instruction) { HandleBitwiseOperation(instruction); } 6184void LocationsBuilderX86::VisitOr(HOr* instruction) { HandleBitwiseOperation(instruction); } 6185void LocationsBuilderX86::VisitXor(HXor* instruction) { HandleBitwiseOperation(instruction); } 6186 6187void LocationsBuilderX86::HandleBitwiseOperation(HBinaryOperation* instruction) { 6188 LocationSummary* locations = 6189 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 6190 DCHECK(instruction->GetResultType() == Primitive::kPrimInt 6191 || instruction->GetResultType() == Primitive::kPrimLong); 6192 locations->SetInAt(0, Location::RequiresRegister()); 6193 locations->SetInAt(1, Location::Any()); 6194 locations->SetOut(Location::SameAsFirstInput()); 6195} 6196 6197void InstructionCodeGeneratorX86::VisitAnd(HAnd* instruction) { 6198 HandleBitwiseOperation(instruction); 6199} 6200 6201void InstructionCodeGeneratorX86::VisitOr(HOr* instruction) { 6202 HandleBitwiseOperation(instruction); 6203} 6204 6205void InstructionCodeGeneratorX86::VisitXor(HXor* instruction) { 6206 HandleBitwiseOperation(instruction); 6207} 6208 6209void InstructionCodeGeneratorX86::HandleBitwiseOperation(HBinaryOperation* instruction) { 6210 LocationSummary* locations = instruction->GetLocations(); 6211 Location first = locations->InAt(0); 6212 Location second = locations->InAt(1); 6213 DCHECK(first.Equals(locations->Out())); 6214 6215 if (instruction->GetResultType() == Primitive::kPrimInt) { 6216 if (second.IsRegister()) { 6217 if (instruction->IsAnd()) { 6218 __ andl(first.AsRegister<Register>(), second.AsRegister<Register>()); 6219 } else if (instruction->IsOr()) { 6220 __ orl(first.AsRegister<Register>(), second.AsRegister<Register>()); 6221 } else { 6222 DCHECK(instruction->IsXor()); 6223 __ xorl(first.AsRegister<Register>(), second.AsRegister<Register>()); 6224 } 6225 } else if (second.IsConstant()) { 6226 if (instruction->IsAnd()) { 6227 __ andl(first.AsRegister<Register>(), 6228 Immediate(second.GetConstant()->AsIntConstant()->GetValue())); 6229 } else if (instruction->IsOr()) { 6230 __ orl(first.AsRegister<Register>(), 6231 Immediate(second.GetConstant()->AsIntConstant()->GetValue())); 6232 } else { 6233 DCHECK(instruction->IsXor()); 6234 __ xorl(first.AsRegister<Register>(), 6235 Immediate(second.GetConstant()->AsIntConstant()->GetValue())); 6236 } 6237 } else { 6238 if (instruction->IsAnd()) { 6239 __ andl(first.AsRegister<Register>(), Address(ESP, second.GetStackIndex())); 6240 } else if (instruction->IsOr()) { 6241 __ orl(first.AsRegister<Register>(), Address(ESP, second.GetStackIndex())); 6242 } else { 6243 DCHECK(instruction->IsXor()); 6244 __ xorl(first.AsRegister<Register>(), Address(ESP, second.GetStackIndex())); 6245 } 6246 } 6247 } else { 6248 DCHECK_EQ(instruction->GetResultType(), Primitive::kPrimLong); 6249 if (second.IsRegisterPair()) { 6250 if (instruction->IsAnd()) { 6251 __ andl(first.AsRegisterPairLow<Register>(), second.AsRegisterPairLow<Register>()); 6252 __ andl(first.AsRegisterPairHigh<Register>(), second.AsRegisterPairHigh<Register>()); 6253 } else if (instruction->IsOr()) { 6254 __ orl(first.AsRegisterPairLow<Register>(), second.AsRegisterPairLow<Register>()); 6255 __ orl(first.AsRegisterPairHigh<Register>(), second.AsRegisterPairHigh<Register>()); 6256 } else { 6257 DCHECK(instruction->IsXor()); 6258 __ xorl(first.AsRegisterPairLow<Register>(), second.AsRegisterPairLow<Register>()); 6259 __ xorl(first.AsRegisterPairHigh<Register>(), second.AsRegisterPairHigh<Register>()); 6260 } 6261 } else if (second.IsDoubleStackSlot()) { 6262 if (instruction->IsAnd()) { 6263 __ andl(first.AsRegisterPairLow<Register>(), Address(ESP, second.GetStackIndex())); 6264 __ andl(first.AsRegisterPairHigh<Register>(), 6265 Address(ESP, second.GetHighStackIndex(kX86WordSize))); 6266 } else if (instruction->IsOr()) { 6267 __ orl(first.AsRegisterPairLow<Register>(), Address(ESP, second.GetStackIndex())); 6268 __ orl(first.AsRegisterPairHigh<Register>(), 6269 Address(ESP, second.GetHighStackIndex(kX86WordSize))); 6270 } else { 6271 DCHECK(instruction->IsXor()); 6272 __ xorl(first.AsRegisterPairLow<Register>(), Address(ESP, second.GetStackIndex())); 6273 __ xorl(first.AsRegisterPairHigh<Register>(), 6274 Address(ESP, second.GetHighStackIndex(kX86WordSize))); 6275 } 6276 } else { 6277 DCHECK(second.IsConstant()) << second; 6278 int64_t value = second.GetConstant()->AsLongConstant()->GetValue(); 6279 int32_t low_value = Low32Bits(value); 6280 int32_t high_value = High32Bits(value); 6281 Immediate low(low_value); 6282 Immediate high(high_value); 6283 Register first_low = first.AsRegisterPairLow<Register>(); 6284 Register first_high = first.AsRegisterPairHigh<Register>(); 6285 if (instruction->IsAnd()) { 6286 if (low_value == 0) { 6287 __ xorl(first_low, first_low); 6288 } else if (low_value != -1) { 6289 __ andl(first_low, low); 6290 } 6291 if (high_value == 0) { 6292 __ xorl(first_high, first_high); 6293 } else if (high_value != -1) { 6294 __ andl(first_high, high); 6295 } 6296 } else if (instruction->IsOr()) { 6297 if (low_value != 0) { 6298 __ orl(first_low, low); 6299 } 6300 if (high_value != 0) { 6301 __ orl(first_high, high); 6302 } 6303 } else { 6304 DCHECK(instruction->IsXor()); 6305 if (low_value != 0) { 6306 __ xorl(first_low, low); 6307 } 6308 if (high_value != 0) { 6309 __ xorl(first_high, high); 6310 } 6311 } 6312 } 6313 } 6314} 6315 6316void CodeGeneratorX86::GenerateReadBarrier(HInstruction* instruction, 6317 Location out, 6318 Location ref, 6319 Location obj, 6320 uint32_t offset, 6321 Location index) { 6322 DCHECK(kEmitCompilerReadBarrier); 6323 6324 // If heap poisoning is enabled, the unpoisoning of the loaded 6325 // reference will be carried out by the runtime within the slow 6326 // path. 6327 // 6328 // Note that `ref` currently does not get unpoisoned (when heap 6329 // poisoning is enabled), which is alright as the `ref` argument is 6330 // not used by the artReadBarrierSlow entry point. 6331 // 6332 // TODO: Unpoison `ref` when it is used by artReadBarrierSlow. 6333 SlowPathCode* slow_path = new (GetGraph()->GetArena()) 6334 ReadBarrierForHeapReferenceSlowPathX86(instruction, out, ref, obj, offset, index); 6335 AddSlowPath(slow_path); 6336 6337 // TODO: When read barrier has a fast path, add it here. 6338 /* Currently the read barrier call is inserted after the original load. 6339 * However, if we have a fast path, we need to perform the load of obj.LockWord *before* the 6340 * original load. This load-load ordering is required by the read barrier. 6341 * The fast path/slow path (for Baker's algorithm) should look like: 6342 * 6343 * bool isGray = obj.LockWord & kReadBarrierMask; 6344 * lfence; // load fence or artificial data dependence to prevent load-load reordering 6345 * ref = obj.field; // this is the original load 6346 * if (isGray) { 6347 * ref = Mark(ref); // ideally the slow path just does Mark(ref) 6348 * } 6349 */ 6350 6351 __ jmp(slow_path->GetEntryLabel()); 6352 __ Bind(slow_path->GetExitLabel()); 6353} 6354 6355void CodeGeneratorX86::MaybeGenerateReadBarrier(HInstruction* instruction, 6356 Location out, 6357 Location ref, 6358 Location obj, 6359 uint32_t offset, 6360 Location index) { 6361 if (kEmitCompilerReadBarrier) { 6362 // If heap poisoning is enabled, unpoisoning will be taken care of 6363 // by the runtime within the slow path. 6364 GenerateReadBarrier(instruction, out, ref, obj, offset, index); 6365 } else if (kPoisonHeapReferences) { 6366 __ UnpoisonHeapReference(out.AsRegister<Register>()); 6367 } 6368} 6369 6370void CodeGeneratorX86::GenerateReadBarrierForRoot(HInstruction* instruction, 6371 Location out, 6372 Location root) { 6373 DCHECK(kEmitCompilerReadBarrier); 6374 6375 // Note that GC roots are not affected by heap poisoning, so we do 6376 // not need to do anything special for this here. 6377 SlowPathCode* slow_path = 6378 new (GetGraph()->GetArena()) ReadBarrierForRootSlowPathX86(instruction, out, root); 6379 AddSlowPath(slow_path); 6380 6381 // TODO: Implement a fast path for ReadBarrierForRoot, performing 6382 // the following operation (for Baker's algorithm): 6383 // 6384 // if (thread.tls32_.is_gc_marking) { 6385 // root = Mark(root); 6386 // } 6387 6388 __ jmp(slow_path->GetEntryLabel()); 6389 __ Bind(slow_path->GetExitLabel()); 6390} 6391 6392void LocationsBuilderX86::VisitBoundType(HBoundType* instruction ATTRIBUTE_UNUSED) { 6393 // Nothing to do, this should be removed during prepare for register allocator. 6394 LOG(FATAL) << "Unreachable"; 6395} 6396 6397void InstructionCodeGeneratorX86::VisitBoundType(HBoundType* instruction ATTRIBUTE_UNUSED) { 6398 // Nothing to do, this should be removed during prepare for register allocator. 6399 LOG(FATAL) << "Unreachable"; 6400} 6401 6402void LocationsBuilderX86::VisitFakeString(HFakeString* instruction) { 6403 DCHECK(codegen_->IsBaseline()); 6404 LocationSummary* locations = 6405 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 6406 locations->SetOut(Location::ConstantLocation(GetGraph()->GetNullConstant())); 6407} 6408 6409void InstructionCodeGeneratorX86::VisitFakeString(HFakeString* instruction ATTRIBUTE_UNUSED) { 6410 DCHECK(codegen_->IsBaseline()); 6411 // Will be generated at use site. 6412} 6413 6414// Simple implementation of packed switch - generate cascaded compare/jumps. 6415void LocationsBuilderX86::VisitPackedSwitch(HPackedSwitch* switch_instr) { 6416 LocationSummary* locations = 6417 new (GetGraph()->GetArena()) LocationSummary(switch_instr, LocationSummary::kNoCall); 6418 locations->SetInAt(0, Location::RequiresRegister()); 6419} 6420 6421void InstructionCodeGeneratorX86::VisitPackedSwitch(HPackedSwitch* switch_instr) { 6422 int32_t lower_bound = switch_instr->GetStartValue(); 6423 int32_t num_entries = switch_instr->GetNumEntries(); 6424 LocationSummary* locations = switch_instr->GetLocations(); 6425 Register value_reg = locations->InAt(0).AsRegister<Register>(); 6426 HBasicBlock* default_block = switch_instr->GetDefaultBlock(); 6427 6428 // Create a series of compare/jumps. 6429 const ArenaVector<HBasicBlock*>& successors = switch_instr->GetBlock()->GetSuccessors(); 6430 for (int i = 0; i < num_entries; i++) { 6431 int32_t case_value = lower_bound + i; 6432 if (case_value == 0) { 6433 __ testl(value_reg, value_reg); 6434 } else { 6435 __ cmpl(value_reg, Immediate(case_value)); 6436 } 6437 __ j(kEqual, codegen_->GetLabelOf(successors[i])); 6438 } 6439 6440 // And the default for any other value. 6441 if (!codegen_->GoesToNextBlock(switch_instr->GetBlock(), default_block)) { 6442 __ jmp(codegen_->GetLabelOf(default_block)); 6443 } 6444} 6445 6446void LocationsBuilderX86::VisitX86PackedSwitch(HX86PackedSwitch* switch_instr) { 6447 LocationSummary* locations = 6448 new (GetGraph()->GetArena()) LocationSummary(switch_instr, LocationSummary::kNoCall); 6449 locations->SetInAt(0, Location::RequiresRegister()); 6450 6451 // Constant area pointer. 6452 locations->SetInAt(1, Location::RequiresRegister()); 6453 6454 // And the temporary we need. 6455 locations->AddTemp(Location::RequiresRegister()); 6456} 6457 6458void InstructionCodeGeneratorX86::VisitX86PackedSwitch(HX86PackedSwitch* switch_instr) { 6459 int32_t lower_bound = switch_instr->GetStartValue(); 6460 int32_t num_entries = switch_instr->GetNumEntries(); 6461 LocationSummary* locations = switch_instr->GetLocations(); 6462 Register value_reg = locations->InAt(0).AsRegister<Register>(); 6463 HBasicBlock* default_block = switch_instr->GetDefaultBlock(); 6464 6465 // Optimizing has a jump area. 6466 Register temp_reg = locations->GetTemp(0).AsRegister<Register>(); 6467 Register constant_area = locations->InAt(1).AsRegister<Register>(); 6468 6469 // Remove the bias, if needed. 6470 if (lower_bound != 0) { 6471 __ leal(temp_reg, Address(value_reg, -lower_bound)); 6472 value_reg = temp_reg; 6473 } 6474 6475 // Is the value in range? 6476 DCHECK_GE(num_entries, 1); 6477 __ cmpl(value_reg, Immediate(num_entries - 1)); 6478 __ j(kAbove, codegen_->GetLabelOf(default_block)); 6479 6480 // We are in the range of the table. 6481 // Load (target-constant_area) from the jump table, indexing by the value. 6482 __ movl(temp_reg, codegen_->LiteralCaseTable(switch_instr, constant_area, value_reg)); 6483 6484 // Compute the actual target address by adding in constant_area. 6485 __ addl(temp_reg, constant_area); 6486 6487 // And jump. 6488 __ jmp(temp_reg); 6489} 6490 6491void LocationsBuilderX86::VisitX86ComputeBaseMethodAddress( 6492 HX86ComputeBaseMethodAddress* insn) { 6493 LocationSummary* locations = 6494 new (GetGraph()->GetArena()) LocationSummary(insn, LocationSummary::kNoCall); 6495 locations->SetOut(Location::RequiresRegister()); 6496} 6497 6498void InstructionCodeGeneratorX86::VisitX86ComputeBaseMethodAddress( 6499 HX86ComputeBaseMethodAddress* insn) { 6500 LocationSummary* locations = insn->GetLocations(); 6501 Register reg = locations->Out().AsRegister<Register>(); 6502 6503 // Generate call to next instruction. 6504 Label next_instruction; 6505 __ call(&next_instruction); 6506 __ Bind(&next_instruction); 6507 6508 // Remember this offset for later use with constant area. 6509 codegen_->SetMethodAddressOffset(GetAssembler()->CodeSize()); 6510 6511 // Grab the return address off the stack. 6512 __ popl(reg); 6513} 6514 6515void LocationsBuilderX86::VisitX86LoadFromConstantTable( 6516 HX86LoadFromConstantTable* insn) { 6517 LocationSummary* locations = 6518 new (GetGraph()->GetArena()) LocationSummary(insn, LocationSummary::kNoCall); 6519 6520 locations->SetInAt(0, Location::RequiresRegister()); 6521 locations->SetInAt(1, Location::ConstantLocation(insn->GetConstant())); 6522 6523 // If we don't need to be materialized, we only need the inputs to be set. 6524 if (!insn->NeedsMaterialization()) { 6525 return; 6526 } 6527 6528 switch (insn->GetType()) { 6529 case Primitive::kPrimFloat: 6530 case Primitive::kPrimDouble: 6531 locations->SetOut(Location::RequiresFpuRegister()); 6532 break; 6533 6534 case Primitive::kPrimInt: 6535 locations->SetOut(Location::RequiresRegister()); 6536 break; 6537 6538 default: 6539 LOG(FATAL) << "Unsupported x86 constant area type " << insn->GetType(); 6540 } 6541} 6542 6543void InstructionCodeGeneratorX86::VisitX86LoadFromConstantTable(HX86LoadFromConstantTable* insn) { 6544 if (!insn->NeedsMaterialization()) { 6545 return; 6546 } 6547 6548 LocationSummary* locations = insn->GetLocations(); 6549 Location out = locations->Out(); 6550 Register const_area = locations->InAt(0).AsRegister<Register>(); 6551 HConstant *value = insn->GetConstant(); 6552 6553 switch (insn->GetType()) { 6554 case Primitive::kPrimFloat: 6555 __ movss(out.AsFpuRegister<XmmRegister>(), 6556 codegen_->LiteralFloatAddress(value->AsFloatConstant()->GetValue(), const_area)); 6557 break; 6558 6559 case Primitive::kPrimDouble: 6560 __ movsd(out.AsFpuRegister<XmmRegister>(), 6561 codegen_->LiteralDoubleAddress(value->AsDoubleConstant()->GetValue(), const_area)); 6562 break; 6563 6564 case Primitive::kPrimInt: 6565 __ movl(out.AsRegister<Register>(), 6566 codegen_->LiteralInt32Address(value->AsIntConstant()->GetValue(), const_area)); 6567 break; 6568 6569 default: 6570 LOG(FATAL) << "Unsupported x86 constant area type " << insn->GetType(); 6571 } 6572} 6573 6574/** 6575 * Class to handle late fixup of offsets into constant area. 6576 */ 6577class RIPFixup : public AssemblerFixup, public ArenaObject<kArenaAllocCodeGenerator> { 6578 public: 6579 RIPFixup(CodeGeneratorX86& codegen, size_t offset) 6580 : codegen_(&codegen), offset_into_constant_area_(offset) {} 6581 6582 protected: 6583 void SetOffset(size_t offset) { offset_into_constant_area_ = offset; } 6584 6585 CodeGeneratorX86* codegen_; 6586 6587 private: 6588 void Process(const MemoryRegion& region, int pos) OVERRIDE { 6589 // Patch the correct offset for the instruction. The place to patch is the 6590 // last 4 bytes of the instruction. 6591 // The value to patch is the distance from the offset in the constant area 6592 // from the address computed by the HX86ComputeBaseMethodAddress instruction. 6593 int32_t constant_offset = codegen_->ConstantAreaStart() + offset_into_constant_area_; 6594 int32_t relative_position = constant_offset - codegen_->GetMethodAddressOffset();; 6595 6596 // Patch in the right value. 6597 region.StoreUnaligned<int32_t>(pos - 4, relative_position); 6598 } 6599 6600 // Location in constant area that the fixup refers to. 6601 int32_t offset_into_constant_area_; 6602}; 6603 6604/** 6605 * Class to handle late fixup of offsets to a jump table that will be created in the 6606 * constant area. 6607 */ 6608class JumpTableRIPFixup : public RIPFixup { 6609 public: 6610 JumpTableRIPFixup(CodeGeneratorX86& codegen, HX86PackedSwitch* switch_instr) 6611 : RIPFixup(codegen, static_cast<size_t>(-1)), switch_instr_(switch_instr) {} 6612 6613 void CreateJumpTable() { 6614 X86Assembler* assembler = codegen_->GetAssembler(); 6615 6616 // Ensure that the reference to the jump table has the correct offset. 6617 const int32_t offset_in_constant_table = assembler->ConstantAreaSize(); 6618 SetOffset(offset_in_constant_table); 6619 6620 // The label values in the jump table are computed relative to the 6621 // instruction addressing the constant area. 6622 const int32_t relative_offset = codegen_->GetMethodAddressOffset(); 6623 6624 // Populate the jump table with the correct values for the jump table. 6625 int32_t num_entries = switch_instr_->GetNumEntries(); 6626 HBasicBlock* block = switch_instr_->GetBlock(); 6627 const ArenaVector<HBasicBlock*>& successors = block->GetSuccessors(); 6628 // The value that we want is the target offset - the position of the table. 6629 for (int32_t i = 0; i < num_entries; i++) { 6630 HBasicBlock* b = successors[i]; 6631 Label* l = codegen_->GetLabelOf(b); 6632 DCHECK(l->IsBound()); 6633 int32_t offset_to_block = l->Position() - relative_offset; 6634 assembler->AppendInt32(offset_to_block); 6635 } 6636 } 6637 6638 private: 6639 const HX86PackedSwitch* switch_instr_; 6640}; 6641 6642void CodeGeneratorX86::Finalize(CodeAllocator* allocator) { 6643 // Generate the constant area if needed. 6644 X86Assembler* assembler = GetAssembler(); 6645 if (!assembler->IsConstantAreaEmpty() || !fixups_to_jump_tables_.empty()) { 6646 // Align to 4 byte boundary to reduce cache misses, as the data is 4 and 8 6647 // byte values. 6648 assembler->Align(4, 0); 6649 constant_area_start_ = assembler->CodeSize(); 6650 6651 // Populate any jump tables. 6652 for (auto jump_table : fixups_to_jump_tables_) { 6653 jump_table->CreateJumpTable(); 6654 } 6655 6656 // And now add the constant area to the generated code. 6657 assembler->AddConstantArea(); 6658 } 6659 6660 // And finish up. 6661 CodeGenerator::Finalize(allocator); 6662} 6663 6664Address CodeGeneratorX86::LiteralDoubleAddress(double v, Register reg) { 6665 AssemblerFixup* fixup = new (GetGraph()->GetArena()) RIPFixup(*this, __ AddDouble(v)); 6666 return Address(reg, kDummy32BitOffset, fixup); 6667} 6668 6669Address CodeGeneratorX86::LiteralFloatAddress(float v, Register reg) { 6670 AssemblerFixup* fixup = new (GetGraph()->GetArena()) RIPFixup(*this, __ AddFloat(v)); 6671 return Address(reg, kDummy32BitOffset, fixup); 6672} 6673 6674Address CodeGeneratorX86::LiteralInt32Address(int32_t v, Register reg) { 6675 AssemblerFixup* fixup = new (GetGraph()->GetArena()) RIPFixup(*this, __ AddInt32(v)); 6676 return Address(reg, kDummy32BitOffset, fixup); 6677} 6678 6679Address CodeGeneratorX86::LiteralInt64Address(int64_t v, Register reg) { 6680 AssemblerFixup* fixup = new (GetGraph()->GetArena()) RIPFixup(*this, __ AddInt64(v)); 6681 return Address(reg, kDummy32BitOffset, fixup); 6682} 6683 6684Address CodeGeneratorX86::LiteralCaseTable(HX86PackedSwitch* switch_instr, 6685 Register reg, 6686 Register value) { 6687 // Create a fixup to be used to create and address the jump table. 6688 JumpTableRIPFixup* table_fixup = 6689 new (GetGraph()->GetArena()) JumpTableRIPFixup(*this, switch_instr); 6690 6691 // We have to populate the jump tables. 6692 fixups_to_jump_tables_.push_back(table_fixup); 6693 6694 // We want a scaled address, as we are extracting the correct offset from the table. 6695 return Address(reg, value, TIMES_4, kDummy32BitOffset, table_fixup); 6696} 6697 6698// TODO: target as memory. 6699void CodeGeneratorX86::MoveFromReturnRegister(Location target, Primitive::Type type) { 6700 if (!target.IsValid()) { 6701 DCHECK(type == Primitive::kPrimVoid); 6702 return; 6703 } 6704 6705 DCHECK_NE(type, Primitive::kPrimVoid); 6706 6707 Location return_loc = InvokeDexCallingConventionVisitorX86().GetReturnLocation(type); 6708 if (target.Equals(return_loc)) { 6709 return; 6710 } 6711 6712 // TODO: Consider pairs in the parallel move resolver, then this could be nicely merged 6713 // with the else branch. 6714 if (type == Primitive::kPrimLong) { 6715 HParallelMove parallel_move(GetGraph()->GetArena()); 6716 parallel_move.AddMove(return_loc.ToLow(), target.ToLow(), Primitive::kPrimInt, nullptr); 6717 parallel_move.AddMove(return_loc.ToHigh(), target.ToHigh(), Primitive::kPrimInt, nullptr); 6718 GetMoveResolver()->EmitNativeCode(¶llel_move); 6719 } else { 6720 // Let the parallel move resolver take care of all of this. 6721 HParallelMove parallel_move(GetGraph()->GetArena()); 6722 parallel_move.AddMove(return_loc, target, type, nullptr); 6723 GetMoveResolver()->EmitNativeCode(¶llel_move); 6724 } 6725} 6726 6727#undef __ 6728 6729} // namespace x86 6730} // namespace art 6731