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_arm.h" 18 19 #include "arch/arm/instruction_set_features_arm.h" 20 #include "art_method.h" 21 #include "code_generator_utils.h" 22 #include "compiled_method.h" 23 #include "entrypoints/quick/quick_entrypoints.h" 24 #include "gc/accounting/card_table.h" 25 #include "intrinsics.h" 26 #include "intrinsics_arm.h" 27 #include "mirror/array-inl.h" 28 #include "mirror/class-inl.h" 29 #include "thread.h" 30 #include "utils/arm/assembler_arm.h" 31 #include "utils/arm/managed_register_arm.h" 32 #include "utils/assembler.h" 33 #include "utils/stack_checks.h" 34 35 namespace art { 36 37 template<class MirrorType> 38 class GcRoot; 39 40 namespace arm { 41 42 static bool ExpectedPairLayout(Location location) { 43 // We expected this for both core and fpu register pairs. 44 return ((location.low() & 1) == 0) && (location.low() + 1 == location.high()); 45 } 46 47 static constexpr int kCurrentMethodStackOffset = 0; 48 static constexpr Register kMethodRegisterArgument = R0; 49 50 static constexpr Register kCoreAlwaysSpillRegister = R5; 51 static constexpr Register kCoreCalleeSaves[] = 52 { R5, R6, R7, R8, R10, R11, LR }; 53 static constexpr SRegister kFpuCalleeSaves[] = 54 { S16, S17, S18, S19, S20, S21, S22, S23, S24, S25, S26, S27, S28, S29, S30, S31 }; 55 56 // D31 cannot be split into two S registers, and the register allocator only works on 57 // S registers. Therefore there is no need to block it. 58 static constexpr DRegister DTMP = D31; 59 60 static constexpr uint32_t kPackedSwitchCompareJumpThreshold = 7; 61 62 #define __ down_cast<ArmAssembler*>(codegen->GetAssembler())-> 63 #define QUICK_ENTRY_POINT(x) QUICK_ENTRYPOINT_OFFSET(kArmWordSize, x).Int32Value() 64 65 class NullCheckSlowPathARM : public SlowPathCode { 66 public: 67 explicit NullCheckSlowPathARM(HNullCheck* instruction) : SlowPathCode(instruction) {} 68 69 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 70 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 71 __ Bind(GetEntryLabel()); 72 if (instruction_->CanThrowIntoCatchBlock()) { 73 // Live registers will be restored in the catch block if caught. 74 SaveLiveRegisters(codegen, instruction_->GetLocations()); 75 } 76 arm_codegen->InvokeRuntime( 77 QUICK_ENTRY_POINT(pThrowNullPointer), instruction_, instruction_->GetDexPc(), this); 78 CheckEntrypointTypes<kQuickThrowNullPointer, void, void>(); 79 } 80 81 bool IsFatal() const OVERRIDE { return true; } 82 83 const char* GetDescription() const OVERRIDE { return "NullCheckSlowPathARM"; } 84 85 private: 86 DISALLOW_COPY_AND_ASSIGN(NullCheckSlowPathARM); 87 }; 88 89 class DivZeroCheckSlowPathARM : public SlowPathCode { 90 public: 91 explicit DivZeroCheckSlowPathARM(HDivZeroCheck* instruction) : SlowPathCode(instruction) {} 92 93 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 94 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 95 __ Bind(GetEntryLabel()); 96 if (instruction_->CanThrowIntoCatchBlock()) { 97 // Live registers will be restored in the catch block if caught. 98 SaveLiveRegisters(codegen, instruction_->GetLocations()); 99 } 100 arm_codegen->InvokeRuntime( 101 QUICK_ENTRY_POINT(pThrowDivZero), instruction_, instruction_->GetDexPc(), this); 102 CheckEntrypointTypes<kQuickThrowDivZero, void, void>(); 103 } 104 105 bool IsFatal() const OVERRIDE { return true; } 106 107 const char* GetDescription() const OVERRIDE { return "DivZeroCheckSlowPathARM"; } 108 109 private: 110 DISALLOW_COPY_AND_ASSIGN(DivZeroCheckSlowPathARM); 111 }; 112 113 class SuspendCheckSlowPathARM : public SlowPathCode { 114 public: 115 SuspendCheckSlowPathARM(HSuspendCheck* instruction, HBasicBlock* successor) 116 : SlowPathCode(instruction), successor_(successor) {} 117 118 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 119 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 120 __ Bind(GetEntryLabel()); 121 SaveLiveRegisters(codegen, instruction_->GetLocations()); 122 arm_codegen->InvokeRuntime( 123 QUICK_ENTRY_POINT(pTestSuspend), instruction_, instruction_->GetDexPc(), this); 124 CheckEntrypointTypes<kQuickTestSuspend, void, void>(); 125 RestoreLiveRegisters(codegen, instruction_->GetLocations()); 126 if (successor_ == nullptr) { 127 __ b(GetReturnLabel()); 128 } else { 129 __ b(arm_codegen->GetLabelOf(successor_)); 130 } 131 } 132 133 Label* GetReturnLabel() { 134 DCHECK(successor_ == nullptr); 135 return &return_label_; 136 } 137 138 HBasicBlock* GetSuccessor() const { 139 return successor_; 140 } 141 142 const char* GetDescription() const OVERRIDE { return "SuspendCheckSlowPathARM"; } 143 144 private: 145 // If not null, the block to branch to after the suspend check. 146 HBasicBlock* const successor_; 147 148 // If `successor_` is null, the label to branch to after the suspend check. 149 Label return_label_; 150 151 DISALLOW_COPY_AND_ASSIGN(SuspendCheckSlowPathARM); 152 }; 153 154 class BoundsCheckSlowPathARM : public SlowPathCode { 155 public: 156 explicit BoundsCheckSlowPathARM(HBoundsCheck* instruction) 157 : SlowPathCode(instruction) {} 158 159 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 160 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 161 LocationSummary* locations = instruction_->GetLocations(); 162 163 __ Bind(GetEntryLabel()); 164 if (instruction_->CanThrowIntoCatchBlock()) { 165 // Live registers will be restored in the catch block if caught. 166 SaveLiveRegisters(codegen, instruction_->GetLocations()); 167 } 168 // We're moving two locations to locations that could overlap, so we need a parallel 169 // move resolver. 170 InvokeRuntimeCallingConvention calling_convention; 171 codegen->EmitParallelMoves( 172 locations->InAt(0), 173 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 174 Primitive::kPrimInt, 175 locations->InAt(1), 176 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 177 Primitive::kPrimInt); 178 arm_codegen->InvokeRuntime( 179 QUICK_ENTRY_POINT(pThrowArrayBounds), instruction_, instruction_->GetDexPc(), this); 180 CheckEntrypointTypes<kQuickThrowArrayBounds, void, int32_t, int32_t>(); 181 } 182 183 bool IsFatal() const OVERRIDE { return true; } 184 185 const char* GetDescription() const OVERRIDE { return "BoundsCheckSlowPathARM"; } 186 187 private: 188 DISALLOW_COPY_AND_ASSIGN(BoundsCheckSlowPathARM); 189 }; 190 191 class LoadClassSlowPathARM : public SlowPathCode { 192 public: 193 LoadClassSlowPathARM(HLoadClass* cls, 194 HInstruction* at, 195 uint32_t dex_pc, 196 bool do_clinit) 197 : SlowPathCode(at), cls_(cls), at_(at), dex_pc_(dex_pc), do_clinit_(do_clinit) { 198 DCHECK(at->IsLoadClass() || at->IsClinitCheck()); 199 } 200 201 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 202 LocationSummary* locations = at_->GetLocations(); 203 204 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 205 __ Bind(GetEntryLabel()); 206 SaveLiveRegisters(codegen, locations); 207 208 InvokeRuntimeCallingConvention calling_convention; 209 __ LoadImmediate(calling_convention.GetRegisterAt(0), cls_->GetTypeIndex()); 210 int32_t entry_point_offset = do_clinit_ 211 ? QUICK_ENTRY_POINT(pInitializeStaticStorage) 212 : QUICK_ENTRY_POINT(pInitializeType); 213 arm_codegen->InvokeRuntime(entry_point_offset, at_, dex_pc_, this); 214 if (do_clinit_) { 215 CheckEntrypointTypes<kQuickInitializeStaticStorage, void*, uint32_t>(); 216 } else { 217 CheckEntrypointTypes<kQuickInitializeType, void*, uint32_t>(); 218 } 219 220 // Move the class to the desired location. 221 Location out = locations->Out(); 222 if (out.IsValid()) { 223 DCHECK(out.IsRegister() && !locations->GetLiveRegisters()->ContainsCoreRegister(out.reg())); 224 arm_codegen->Move32(locations->Out(), Location::RegisterLocation(R0)); 225 } 226 RestoreLiveRegisters(codegen, locations); 227 __ b(GetExitLabel()); 228 } 229 230 const char* GetDescription() const OVERRIDE { return "LoadClassSlowPathARM"; } 231 232 private: 233 // The class this slow path will load. 234 HLoadClass* const cls_; 235 236 // The instruction where this slow path is happening. 237 // (Might be the load class or an initialization check). 238 HInstruction* const at_; 239 240 // The dex PC of `at_`. 241 const uint32_t dex_pc_; 242 243 // Whether to initialize the class. 244 const bool do_clinit_; 245 246 DISALLOW_COPY_AND_ASSIGN(LoadClassSlowPathARM); 247 }; 248 249 class LoadStringSlowPathARM : public SlowPathCode { 250 public: 251 explicit LoadStringSlowPathARM(HLoadString* instruction) : SlowPathCode(instruction) {} 252 253 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 254 LocationSummary* locations = instruction_->GetLocations(); 255 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg())); 256 257 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 258 __ Bind(GetEntryLabel()); 259 SaveLiveRegisters(codegen, locations); 260 261 InvokeRuntimeCallingConvention calling_convention; 262 const uint32_t string_index = instruction_->AsLoadString()->GetStringIndex(); 263 __ LoadImmediate(calling_convention.GetRegisterAt(0), string_index); 264 arm_codegen->InvokeRuntime( 265 QUICK_ENTRY_POINT(pResolveString), instruction_, instruction_->GetDexPc(), this); 266 CheckEntrypointTypes<kQuickResolveString, void*, uint32_t>(); 267 arm_codegen->Move32(locations->Out(), Location::RegisterLocation(R0)); 268 269 RestoreLiveRegisters(codegen, locations); 270 __ b(GetExitLabel()); 271 } 272 273 const char* GetDescription() const OVERRIDE { return "LoadStringSlowPathARM"; } 274 275 private: 276 DISALLOW_COPY_AND_ASSIGN(LoadStringSlowPathARM); 277 }; 278 279 class TypeCheckSlowPathARM : public SlowPathCode { 280 public: 281 TypeCheckSlowPathARM(HInstruction* instruction, bool is_fatal) 282 : SlowPathCode(instruction), is_fatal_(is_fatal) {} 283 284 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 285 LocationSummary* locations = instruction_->GetLocations(); 286 Location object_class = instruction_->IsCheckCast() ? locations->GetTemp(0) 287 : locations->Out(); 288 DCHECK(instruction_->IsCheckCast() 289 || !locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg())); 290 291 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 292 __ Bind(GetEntryLabel()); 293 294 if (!is_fatal_) { 295 SaveLiveRegisters(codegen, locations); 296 } 297 298 // We're moving two locations to locations that could overlap, so we need a parallel 299 // move resolver. 300 InvokeRuntimeCallingConvention calling_convention; 301 codegen->EmitParallelMoves( 302 locations->InAt(1), 303 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 304 Primitive::kPrimNot, 305 object_class, 306 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 307 Primitive::kPrimNot); 308 309 if (instruction_->IsInstanceOf()) { 310 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pInstanceofNonTrivial), 311 instruction_, 312 instruction_->GetDexPc(), 313 this); 314 CheckEntrypointTypes< 315 kQuickInstanceofNonTrivial, uint32_t, const mirror::Class*, const mirror::Class*>(); 316 arm_codegen->Move32(locations->Out(), Location::RegisterLocation(R0)); 317 } else { 318 DCHECK(instruction_->IsCheckCast()); 319 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pCheckCast), 320 instruction_, 321 instruction_->GetDexPc(), 322 this); 323 CheckEntrypointTypes<kQuickCheckCast, void, const mirror::Class*, const mirror::Class*>(); 324 } 325 326 if (!is_fatal_) { 327 RestoreLiveRegisters(codegen, locations); 328 __ b(GetExitLabel()); 329 } 330 } 331 332 const char* GetDescription() const OVERRIDE { return "TypeCheckSlowPathARM"; } 333 334 bool IsFatal() const OVERRIDE { return is_fatal_; } 335 336 private: 337 const bool is_fatal_; 338 339 DISALLOW_COPY_AND_ASSIGN(TypeCheckSlowPathARM); 340 }; 341 342 class DeoptimizationSlowPathARM : public SlowPathCode { 343 public: 344 explicit DeoptimizationSlowPathARM(HDeoptimize* instruction) 345 : SlowPathCode(instruction) {} 346 347 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 348 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 349 __ Bind(GetEntryLabel()); 350 SaveLiveRegisters(codegen, instruction_->GetLocations()); 351 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pDeoptimize), 352 instruction_, 353 instruction_->GetDexPc(), 354 this); 355 CheckEntrypointTypes<kQuickDeoptimize, void, void>(); 356 } 357 358 const char* GetDescription() const OVERRIDE { return "DeoptimizationSlowPathARM"; } 359 360 private: 361 DISALLOW_COPY_AND_ASSIGN(DeoptimizationSlowPathARM); 362 }; 363 364 class ArraySetSlowPathARM : public SlowPathCode { 365 public: 366 explicit ArraySetSlowPathARM(HInstruction* instruction) : SlowPathCode(instruction) {} 367 368 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 369 LocationSummary* locations = instruction_->GetLocations(); 370 __ Bind(GetEntryLabel()); 371 SaveLiveRegisters(codegen, locations); 372 373 InvokeRuntimeCallingConvention calling_convention; 374 HParallelMove parallel_move(codegen->GetGraph()->GetArena()); 375 parallel_move.AddMove( 376 locations->InAt(0), 377 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 378 Primitive::kPrimNot, 379 nullptr); 380 parallel_move.AddMove( 381 locations->InAt(1), 382 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 383 Primitive::kPrimInt, 384 nullptr); 385 parallel_move.AddMove( 386 locations->InAt(2), 387 Location::RegisterLocation(calling_convention.GetRegisterAt(2)), 388 Primitive::kPrimNot, 389 nullptr); 390 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 391 392 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 393 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pAputObject), 394 instruction_, 395 instruction_->GetDexPc(), 396 this); 397 CheckEntrypointTypes<kQuickAputObject, void, mirror::Array*, int32_t, mirror::Object*>(); 398 RestoreLiveRegisters(codegen, locations); 399 __ b(GetExitLabel()); 400 } 401 402 const char* GetDescription() const OVERRIDE { return "ArraySetSlowPathARM"; } 403 404 private: 405 DISALLOW_COPY_AND_ASSIGN(ArraySetSlowPathARM); 406 }; 407 408 // Slow path marking an object during a read barrier. 409 class ReadBarrierMarkSlowPathARM : public SlowPathCode { 410 public: 411 ReadBarrierMarkSlowPathARM(HInstruction* instruction, Location out, Location obj) 412 : SlowPathCode(instruction), out_(out), obj_(obj) { 413 DCHECK(kEmitCompilerReadBarrier); 414 } 415 416 const char* GetDescription() const OVERRIDE { return "ReadBarrierMarkSlowPathARM"; } 417 418 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 419 LocationSummary* locations = instruction_->GetLocations(); 420 Register reg_out = out_.AsRegister<Register>(); 421 DCHECK(locations->CanCall()); 422 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg_out)); 423 DCHECK(instruction_->IsInstanceFieldGet() || 424 instruction_->IsStaticFieldGet() || 425 instruction_->IsArrayGet() || 426 instruction_->IsLoadClass() || 427 instruction_->IsLoadString() || 428 instruction_->IsInstanceOf() || 429 instruction_->IsCheckCast()) 430 << "Unexpected instruction in read barrier marking slow path: " 431 << instruction_->DebugName(); 432 433 __ Bind(GetEntryLabel()); 434 SaveLiveRegisters(codegen, locations); 435 436 InvokeRuntimeCallingConvention calling_convention; 437 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 438 arm_codegen->Move32(Location::RegisterLocation(calling_convention.GetRegisterAt(0)), obj_); 439 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierMark), 440 instruction_, 441 instruction_->GetDexPc(), 442 this); 443 CheckEntrypointTypes<kQuickReadBarrierMark, mirror::Object*, mirror::Object*>(); 444 arm_codegen->Move32(out_, Location::RegisterLocation(R0)); 445 446 RestoreLiveRegisters(codegen, locations); 447 __ b(GetExitLabel()); 448 } 449 450 private: 451 const Location out_; 452 const Location obj_; 453 454 DISALLOW_COPY_AND_ASSIGN(ReadBarrierMarkSlowPathARM); 455 }; 456 457 // Slow path generating a read barrier for a heap reference. 458 class ReadBarrierForHeapReferenceSlowPathARM : public SlowPathCode { 459 public: 460 ReadBarrierForHeapReferenceSlowPathARM(HInstruction* instruction, 461 Location out, 462 Location ref, 463 Location obj, 464 uint32_t offset, 465 Location index) 466 : SlowPathCode(instruction), 467 out_(out), 468 ref_(ref), 469 obj_(obj), 470 offset_(offset), 471 index_(index) { 472 DCHECK(kEmitCompilerReadBarrier); 473 // If `obj` is equal to `out` or `ref`, it means the initial object 474 // has been overwritten by (or after) the heap object reference load 475 // to be instrumented, e.g.: 476 // 477 // __ LoadFromOffset(kLoadWord, out, out, offset); 478 // codegen_->GenerateReadBarrierSlow(instruction, out_loc, out_loc, out_loc, offset); 479 // 480 // In that case, we have lost the information about the original 481 // object, and the emitted read barrier cannot work properly. 482 DCHECK(!obj.Equals(out)) << "obj=" << obj << " out=" << out; 483 DCHECK(!obj.Equals(ref)) << "obj=" << obj << " ref=" << ref; 484 } 485 486 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 487 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 488 LocationSummary* locations = instruction_->GetLocations(); 489 Register reg_out = out_.AsRegister<Register>(); 490 DCHECK(locations->CanCall()); 491 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg_out)); 492 DCHECK(!instruction_->IsInvoke() || 493 (instruction_->IsInvokeStaticOrDirect() && 494 instruction_->GetLocations()->Intrinsified())) 495 << "Unexpected instruction in read barrier for heap reference slow path: " 496 << instruction_->DebugName(); 497 498 __ Bind(GetEntryLabel()); 499 SaveLiveRegisters(codegen, locations); 500 501 // We may have to change the index's value, but as `index_` is a 502 // constant member (like other "inputs" of this slow path), 503 // introduce a copy of it, `index`. 504 Location index = index_; 505 if (index_.IsValid()) { 506 // Handle `index_` for HArrayGet and intrinsic UnsafeGetObject. 507 if (instruction_->IsArrayGet()) { 508 // Compute the actual memory offset and store it in `index`. 509 Register index_reg = index_.AsRegister<Register>(); 510 DCHECK(locations->GetLiveRegisters()->ContainsCoreRegister(index_reg)); 511 if (codegen->IsCoreCalleeSaveRegister(index_reg)) { 512 // We are about to change the value of `index_reg` (see the 513 // calls to art::arm::Thumb2Assembler::Lsl and 514 // art::arm::Thumb2Assembler::AddConstant below), but it has 515 // not been saved by the previous call to 516 // art::SlowPathCode::SaveLiveRegisters, as it is a 517 // callee-save register -- 518 // art::SlowPathCode::SaveLiveRegisters does not consider 519 // callee-save registers, as it has been designed with the 520 // assumption that callee-save registers are supposed to be 521 // handled by the called function. So, as a callee-save 522 // register, `index_reg` _would_ eventually be saved onto 523 // the stack, but it would be too late: we would have 524 // changed its value earlier. Therefore, we manually save 525 // it here into another freely available register, 526 // `free_reg`, chosen of course among the caller-save 527 // registers (as a callee-save `free_reg` register would 528 // exhibit the same problem). 529 // 530 // Note we could have requested a temporary register from 531 // the register allocator instead; but we prefer not to, as 532 // this is a slow path, and we know we can find a 533 // caller-save register that is available. 534 Register free_reg = FindAvailableCallerSaveRegister(codegen); 535 __ Mov(free_reg, index_reg); 536 index_reg = free_reg; 537 index = Location::RegisterLocation(index_reg); 538 } else { 539 // The initial register stored in `index_` has already been 540 // saved in the call to art::SlowPathCode::SaveLiveRegisters 541 // (as it is not a callee-save register), so we can freely 542 // use it. 543 } 544 // Shifting the index value contained in `index_reg` by the scale 545 // factor (2) cannot overflow in practice, as the runtime is 546 // unable to allocate object arrays with a size larger than 547 // 2^26 - 1 (that is, 2^28 - 4 bytes). 548 __ Lsl(index_reg, index_reg, TIMES_4); 549 static_assert( 550 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 551 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 552 __ AddConstant(index_reg, index_reg, offset_); 553 } else { 554 DCHECK(instruction_->IsInvoke()); 555 DCHECK(instruction_->GetLocations()->Intrinsified()); 556 DCHECK((instruction_->AsInvoke()->GetIntrinsic() == Intrinsics::kUnsafeGetObject) || 557 (instruction_->AsInvoke()->GetIntrinsic() == Intrinsics::kUnsafeGetObjectVolatile)) 558 << instruction_->AsInvoke()->GetIntrinsic(); 559 DCHECK_EQ(offset_, 0U); 560 DCHECK(index_.IsRegisterPair()); 561 // UnsafeGet's offset location is a register pair, the low 562 // part contains the correct offset. 563 index = index_.ToLow(); 564 } 565 } 566 567 // We're moving two or three locations to locations that could 568 // overlap, so we need a parallel move resolver. 569 InvokeRuntimeCallingConvention calling_convention; 570 HParallelMove parallel_move(codegen->GetGraph()->GetArena()); 571 parallel_move.AddMove(ref_, 572 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 573 Primitive::kPrimNot, 574 nullptr); 575 parallel_move.AddMove(obj_, 576 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 577 Primitive::kPrimNot, 578 nullptr); 579 if (index.IsValid()) { 580 parallel_move.AddMove(index, 581 Location::RegisterLocation(calling_convention.GetRegisterAt(2)), 582 Primitive::kPrimInt, 583 nullptr); 584 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 585 } else { 586 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 587 __ LoadImmediate(calling_convention.GetRegisterAt(2), offset_); 588 } 589 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierSlow), 590 instruction_, 591 instruction_->GetDexPc(), 592 this); 593 CheckEntrypointTypes< 594 kQuickReadBarrierSlow, mirror::Object*, mirror::Object*, mirror::Object*, uint32_t>(); 595 arm_codegen->Move32(out_, Location::RegisterLocation(R0)); 596 597 RestoreLiveRegisters(codegen, locations); 598 __ b(GetExitLabel()); 599 } 600 601 const char* GetDescription() const OVERRIDE { return "ReadBarrierForHeapReferenceSlowPathARM"; } 602 603 private: 604 Register FindAvailableCallerSaveRegister(CodeGenerator* codegen) { 605 size_t ref = static_cast<int>(ref_.AsRegister<Register>()); 606 size_t obj = static_cast<int>(obj_.AsRegister<Register>()); 607 for (size_t i = 0, e = codegen->GetNumberOfCoreRegisters(); i < e; ++i) { 608 if (i != ref && i != obj && !codegen->IsCoreCalleeSaveRegister(i)) { 609 return static_cast<Register>(i); 610 } 611 } 612 // We shall never fail to find a free caller-save register, as 613 // there are more than two core caller-save registers on ARM 614 // (meaning it is possible to find one which is different from 615 // `ref` and `obj`). 616 DCHECK_GT(codegen->GetNumberOfCoreCallerSaveRegisters(), 2u); 617 LOG(FATAL) << "Could not find a free caller-save register"; 618 UNREACHABLE(); 619 } 620 621 const Location out_; 622 const Location ref_; 623 const Location obj_; 624 const uint32_t offset_; 625 // An additional location containing an index to an array. 626 // Only used for HArrayGet and the UnsafeGetObject & 627 // UnsafeGetObjectVolatile intrinsics. 628 const Location index_; 629 630 DISALLOW_COPY_AND_ASSIGN(ReadBarrierForHeapReferenceSlowPathARM); 631 }; 632 633 // Slow path generating a read barrier for a GC root. 634 class ReadBarrierForRootSlowPathARM : public SlowPathCode { 635 public: 636 ReadBarrierForRootSlowPathARM(HInstruction* instruction, Location out, Location root) 637 : SlowPathCode(instruction), out_(out), root_(root) { 638 DCHECK(kEmitCompilerReadBarrier); 639 } 640 641 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 642 LocationSummary* locations = instruction_->GetLocations(); 643 Register reg_out = out_.AsRegister<Register>(); 644 DCHECK(locations->CanCall()); 645 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg_out)); 646 DCHECK(instruction_->IsLoadClass() || instruction_->IsLoadString()) 647 << "Unexpected instruction in read barrier for GC root slow path: " 648 << instruction_->DebugName(); 649 650 __ Bind(GetEntryLabel()); 651 SaveLiveRegisters(codegen, locations); 652 653 InvokeRuntimeCallingConvention calling_convention; 654 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 655 arm_codegen->Move32(Location::RegisterLocation(calling_convention.GetRegisterAt(0)), root_); 656 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierForRootSlow), 657 instruction_, 658 instruction_->GetDexPc(), 659 this); 660 CheckEntrypointTypes<kQuickReadBarrierForRootSlow, mirror::Object*, GcRoot<mirror::Object>*>(); 661 arm_codegen->Move32(out_, Location::RegisterLocation(R0)); 662 663 RestoreLiveRegisters(codegen, locations); 664 __ b(GetExitLabel()); 665 } 666 667 const char* GetDescription() const OVERRIDE { return "ReadBarrierForRootSlowPathARM"; } 668 669 private: 670 const Location out_; 671 const Location root_; 672 673 DISALLOW_COPY_AND_ASSIGN(ReadBarrierForRootSlowPathARM); 674 }; 675 676 #undef __ 677 #define __ down_cast<ArmAssembler*>(GetAssembler())-> 678 679 inline Condition ARMCondition(IfCondition cond) { 680 switch (cond) { 681 case kCondEQ: return EQ; 682 case kCondNE: return NE; 683 case kCondLT: return LT; 684 case kCondLE: return LE; 685 case kCondGT: return GT; 686 case kCondGE: return GE; 687 case kCondB: return LO; 688 case kCondBE: return LS; 689 case kCondA: return HI; 690 case kCondAE: return HS; 691 } 692 LOG(FATAL) << "Unreachable"; 693 UNREACHABLE(); 694 } 695 696 // Maps signed condition to unsigned condition. 697 inline Condition ARMUnsignedCondition(IfCondition cond) { 698 switch (cond) { 699 case kCondEQ: return EQ; 700 case kCondNE: return NE; 701 // Signed to unsigned. 702 case kCondLT: return LO; 703 case kCondLE: return LS; 704 case kCondGT: return HI; 705 case kCondGE: return HS; 706 // Unsigned remain unchanged. 707 case kCondB: return LO; 708 case kCondBE: return LS; 709 case kCondA: return HI; 710 case kCondAE: return HS; 711 } 712 LOG(FATAL) << "Unreachable"; 713 UNREACHABLE(); 714 } 715 716 inline Condition ARMFPCondition(IfCondition cond, bool gt_bias) { 717 // The ARM condition codes can express all the necessary branches, see the 718 // "Meaning (floating-point)" column in the table A8-1 of the ARMv7 reference manual. 719 // There is no dex instruction or HIR that would need the missing conditions 720 // "equal or unordered" or "not equal". 721 switch (cond) { 722 case kCondEQ: return EQ; 723 case kCondNE: return NE /* unordered */; 724 case kCondLT: return gt_bias ? CC : LT /* unordered */; 725 case kCondLE: return gt_bias ? LS : LE /* unordered */; 726 case kCondGT: return gt_bias ? HI /* unordered */ : GT; 727 case kCondGE: return gt_bias ? CS /* unordered */ : GE; 728 default: 729 LOG(FATAL) << "UNREACHABLE"; 730 UNREACHABLE(); 731 } 732 } 733 734 void CodeGeneratorARM::DumpCoreRegister(std::ostream& stream, int reg) const { 735 stream << Register(reg); 736 } 737 738 void CodeGeneratorARM::DumpFloatingPointRegister(std::ostream& stream, int reg) const { 739 stream << SRegister(reg); 740 } 741 742 size_t CodeGeneratorARM::SaveCoreRegister(size_t stack_index, uint32_t reg_id) { 743 __ StoreToOffset(kStoreWord, static_cast<Register>(reg_id), SP, stack_index); 744 return kArmWordSize; 745 } 746 747 size_t CodeGeneratorARM::RestoreCoreRegister(size_t stack_index, uint32_t reg_id) { 748 __ LoadFromOffset(kLoadWord, static_cast<Register>(reg_id), SP, stack_index); 749 return kArmWordSize; 750 } 751 752 size_t CodeGeneratorARM::SaveFloatingPointRegister(size_t stack_index, uint32_t reg_id) { 753 __ StoreSToOffset(static_cast<SRegister>(reg_id), SP, stack_index); 754 return kArmWordSize; 755 } 756 757 size_t CodeGeneratorARM::RestoreFloatingPointRegister(size_t stack_index, uint32_t reg_id) { 758 __ LoadSFromOffset(static_cast<SRegister>(reg_id), SP, stack_index); 759 return kArmWordSize; 760 } 761 762 CodeGeneratorARM::CodeGeneratorARM(HGraph* graph, 763 const ArmInstructionSetFeatures& isa_features, 764 const CompilerOptions& compiler_options, 765 OptimizingCompilerStats* stats) 766 : CodeGenerator(graph, 767 kNumberOfCoreRegisters, 768 kNumberOfSRegisters, 769 kNumberOfRegisterPairs, 770 ComputeRegisterMask(reinterpret_cast<const int*>(kCoreCalleeSaves), 771 arraysize(kCoreCalleeSaves)), 772 ComputeRegisterMask(reinterpret_cast<const int*>(kFpuCalleeSaves), 773 arraysize(kFpuCalleeSaves)), 774 compiler_options, 775 stats), 776 block_labels_(nullptr), 777 location_builder_(graph, this), 778 instruction_visitor_(graph, this), 779 move_resolver_(graph->GetArena(), this), 780 assembler_(graph->GetArena()), 781 isa_features_(isa_features), 782 uint32_literals_(std::less<uint32_t>(), 783 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 784 method_patches_(MethodReferenceComparator(), 785 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 786 call_patches_(MethodReferenceComparator(), 787 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 788 relative_call_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 789 pc_relative_dex_cache_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 790 boot_image_string_patches_(StringReferenceValueComparator(), 791 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 792 pc_relative_string_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 793 boot_image_address_patches_(std::less<uint32_t>(), 794 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)) { 795 // Always save the LR register to mimic Quick. 796 AddAllocatedRegister(Location::RegisterLocation(LR)); 797 } 798 799 void CodeGeneratorARM::Finalize(CodeAllocator* allocator) { 800 // Ensure that we fix up branches and literal loads and emit the literal pool. 801 __ FinalizeCode(); 802 803 // Adjust native pc offsets in stack maps. 804 for (size_t i = 0, num = stack_map_stream_.GetNumberOfStackMaps(); i != num; ++i) { 805 uint32_t old_position = stack_map_stream_.GetStackMap(i).native_pc_offset; 806 uint32_t new_position = __ GetAdjustedPosition(old_position); 807 stack_map_stream_.SetStackMapNativePcOffset(i, new_position); 808 } 809 // Adjust pc offsets for the disassembly information. 810 if (disasm_info_ != nullptr) { 811 GeneratedCodeInterval* frame_entry_interval = disasm_info_->GetFrameEntryInterval(); 812 frame_entry_interval->start = __ GetAdjustedPosition(frame_entry_interval->start); 813 frame_entry_interval->end = __ GetAdjustedPosition(frame_entry_interval->end); 814 for (auto& it : *disasm_info_->GetInstructionIntervals()) { 815 it.second.start = __ GetAdjustedPosition(it.second.start); 816 it.second.end = __ GetAdjustedPosition(it.second.end); 817 } 818 for (auto& it : *disasm_info_->GetSlowPathIntervals()) { 819 it.code_interval.start = __ GetAdjustedPosition(it.code_interval.start); 820 it.code_interval.end = __ GetAdjustedPosition(it.code_interval.end); 821 } 822 } 823 824 CodeGenerator::Finalize(allocator); 825 } 826 827 void CodeGeneratorARM::SetupBlockedRegisters() const { 828 // Don't allocate the dalvik style register pair passing. 829 blocked_register_pairs_[R1_R2] = true; 830 831 // Stack register, LR and PC are always reserved. 832 blocked_core_registers_[SP] = true; 833 blocked_core_registers_[LR] = true; 834 blocked_core_registers_[PC] = true; 835 836 // Reserve thread register. 837 blocked_core_registers_[TR] = true; 838 839 // Reserve temp register. 840 blocked_core_registers_[IP] = true; 841 842 if (GetGraph()->IsDebuggable()) { 843 // Stubs do not save callee-save floating point registers. If the graph 844 // is debuggable, we need to deal with these registers differently. For 845 // now, just block them. 846 for (size_t i = 0; i < arraysize(kFpuCalleeSaves); ++i) { 847 blocked_fpu_registers_[kFpuCalleeSaves[i]] = true; 848 } 849 } 850 851 UpdateBlockedPairRegisters(); 852 } 853 854 void CodeGeneratorARM::UpdateBlockedPairRegisters() const { 855 for (int i = 0; i < kNumberOfRegisterPairs; i++) { 856 ArmManagedRegister current = 857 ArmManagedRegister::FromRegisterPair(static_cast<RegisterPair>(i)); 858 if (blocked_core_registers_[current.AsRegisterPairLow()] 859 || blocked_core_registers_[current.AsRegisterPairHigh()]) { 860 blocked_register_pairs_[i] = true; 861 } 862 } 863 } 864 865 InstructionCodeGeneratorARM::InstructionCodeGeneratorARM(HGraph* graph, CodeGeneratorARM* codegen) 866 : InstructionCodeGenerator(graph, codegen), 867 assembler_(codegen->GetAssembler()), 868 codegen_(codegen) {} 869 870 void CodeGeneratorARM::ComputeSpillMask() { 871 core_spill_mask_ = allocated_registers_.GetCoreRegisters() & core_callee_save_mask_; 872 DCHECK_NE(core_spill_mask_, 0u) << "At least the return address register must be saved"; 873 // There is no easy instruction to restore just the PC on thumb2. We spill and 874 // restore another arbitrary register. 875 core_spill_mask_ |= (1 << kCoreAlwaysSpillRegister); 876 fpu_spill_mask_ = allocated_registers_.GetFloatingPointRegisters() & fpu_callee_save_mask_; 877 // We use vpush and vpop for saving and restoring floating point registers, which take 878 // a SRegister and the number of registers to save/restore after that SRegister. We 879 // therefore update the `fpu_spill_mask_` to also contain those registers not allocated, 880 // but in the range. 881 if (fpu_spill_mask_ != 0) { 882 uint32_t least_significant_bit = LeastSignificantBit(fpu_spill_mask_); 883 uint32_t most_significant_bit = MostSignificantBit(fpu_spill_mask_); 884 for (uint32_t i = least_significant_bit + 1 ; i < most_significant_bit; ++i) { 885 fpu_spill_mask_ |= (1 << i); 886 } 887 } 888 } 889 890 static dwarf::Reg DWARFReg(Register reg) { 891 return dwarf::Reg::ArmCore(static_cast<int>(reg)); 892 } 893 894 static dwarf::Reg DWARFReg(SRegister reg) { 895 return dwarf::Reg::ArmFp(static_cast<int>(reg)); 896 } 897 898 void CodeGeneratorARM::GenerateFrameEntry() { 899 bool skip_overflow_check = 900 IsLeafMethod() && !FrameNeedsStackCheck(GetFrameSize(), InstructionSet::kArm); 901 DCHECK(GetCompilerOptions().GetImplicitStackOverflowChecks()); 902 __ Bind(&frame_entry_label_); 903 904 if (HasEmptyFrame()) { 905 return; 906 } 907 908 if (!skip_overflow_check) { 909 __ AddConstant(IP, SP, -static_cast<int32_t>(GetStackOverflowReservedBytes(kArm))); 910 __ LoadFromOffset(kLoadWord, IP, IP, 0); 911 RecordPcInfo(nullptr, 0); 912 } 913 914 __ PushList(core_spill_mask_); 915 __ cfi().AdjustCFAOffset(kArmWordSize * POPCOUNT(core_spill_mask_)); 916 __ cfi().RelOffsetForMany(DWARFReg(kMethodRegisterArgument), 0, core_spill_mask_, kArmWordSize); 917 if (fpu_spill_mask_ != 0) { 918 SRegister start_register = SRegister(LeastSignificantBit(fpu_spill_mask_)); 919 __ vpushs(start_register, POPCOUNT(fpu_spill_mask_)); 920 __ cfi().AdjustCFAOffset(kArmWordSize * POPCOUNT(fpu_spill_mask_)); 921 __ cfi().RelOffsetForMany(DWARFReg(S0), 0, fpu_spill_mask_, kArmWordSize); 922 } 923 int adjust = GetFrameSize() - FrameEntrySpillSize(); 924 __ AddConstant(SP, -adjust); 925 __ cfi().AdjustCFAOffset(adjust); 926 __ StoreToOffset(kStoreWord, kMethodRegisterArgument, SP, 0); 927 } 928 929 void CodeGeneratorARM::GenerateFrameExit() { 930 if (HasEmptyFrame()) { 931 __ bx(LR); 932 return; 933 } 934 __ cfi().RememberState(); 935 int adjust = GetFrameSize() - FrameEntrySpillSize(); 936 __ AddConstant(SP, adjust); 937 __ cfi().AdjustCFAOffset(-adjust); 938 if (fpu_spill_mask_ != 0) { 939 SRegister start_register = SRegister(LeastSignificantBit(fpu_spill_mask_)); 940 __ vpops(start_register, POPCOUNT(fpu_spill_mask_)); 941 __ cfi().AdjustCFAOffset(-kArmPointerSize * POPCOUNT(fpu_spill_mask_)); 942 __ cfi().RestoreMany(DWARFReg(SRegister(0)), fpu_spill_mask_); 943 } 944 // Pop LR into PC to return. 945 DCHECK_NE(core_spill_mask_ & (1 << LR), 0U); 946 uint32_t pop_mask = (core_spill_mask_ & (~(1 << LR))) | 1 << PC; 947 __ PopList(pop_mask); 948 __ cfi().RestoreState(); 949 __ cfi().DefCFAOffset(GetFrameSize()); 950 } 951 952 void CodeGeneratorARM::Bind(HBasicBlock* block) { 953 Label* label = GetLabelOf(block); 954 __ BindTrackedLabel(label); 955 } 956 957 Location InvokeDexCallingConventionVisitorARM::GetNextLocation(Primitive::Type type) { 958 switch (type) { 959 case Primitive::kPrimBoolean: 960 case Primitive::kPrimByte: 961 case Primitive::kPrimChar: 962 case Primitive::kPrimShort: 963 case Primitive::kPrimInt: 964 case Primitive::kPrimNot: { 965 uint32_t index = gp_index_++; 966 uint32_t stack_index = stack_index_++; 967 if (index < calling_convention.GetNumberOfRegisters()) { 968 return Location::RegisterLocation(calling_convention.GetRegisterAt(index)); 969 } else { 970 return Location::StackSlot(calling_convention.GetStackOffsetOf(stack_index)); 971 } 972 } 973 974 case Primitive::kPrimLong: { 975 uint32_t index = gp_index_; 976 uint32_t stack_index = stack_index_; 977 gp_index_ += 2; 978 stack_index_ += 2; 979 if (index + 1 < calling_convention.GetNumberOfRegisters()) { 980 if (calling_convention.GetRegisterAt(index) == R1) { 981 // Skip R1, and use R2_R3 instead. 982 gp_index_++; 983 index++; 984 } 985 } 986 if (index + 1 < calling_convention.GetNumberOfRegisters()) { 987 DCHECK_EQ(calling_convention.GetRegisterAt(index) + 1, 988 calling_convention.GetRegisterAt(index + 1)); 989 990 return Location::RegisterPairLocation(calling_convention.GetRegisterAt(index), 991 calling_convention.GetRegisterAt(index + 1)); 992 } else { 993 return Location::DoubleStackSlot(calling_convention.GetStackOffsetOf(stack_index)); 994 } 995 } 996 997 case Primitive::kPrimFloat: { 998 uint32_t stack_index = stack_index_++; 999 if (float_index_ % 2 == 0) { 1000 float_index_ = std::max(double_index_, float_index_); 1001 } 1002 if (float_index_ < calling_convention.GetNumberOfFpuRegisters()) { 1003 return Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(float_index_++)); 1004 } else { 1005 return Location::StackSlot(calling_convention.GetStackOffsetOf(stack_index)); 1006 } 1007 } 1008 1009 case Primitive::kPrimDouble: { 1010 double_index_ = std::max(double_index_, RoundUp(float_index_, 2)); 1011 uint32_t stack_index = stack_index_; 1012 stack_index_ += 2; 1013 if (double_index_ + 1 < calling_convention.GetNumberOfFpuRegisters()) { 1014 uint32_t index = double_index_; 1015 double_index_ += 2; 1016 Location result = Location::FpuRegisterPairLocation( 1017 calling_convention.GetFpuRegisterAt(index), 1018 calling_convention.GetFpuRegisterAt(index + 1)); 1019 DCHECK(ExpectedPairLayout(result)); 1020 return result; 1021 } else { 1022 return Location::DoubleStackSlot(calling_convention.GetStackOffsetOf(stack_index)); 1023 } 1024 } 1025 1026 case Primitive::kPrimVoid: 1027 LOG(FATAL) << "Unexpected parameter type " << type; 1028 break; 1029 } 1030 return Location::NoLocation(); 1031 } 1032 1033 Location InvokeDexCallingConventionVisitorARM::GetReturnLocation(Primitive::Type type) const { 1034 switch (type) { 1035 case Primitive::kPrimBoolean: 1036 case Primitive::kPrimByte: 1037 case Primitive::kPrimChar: 1038 case Primitive::kPrimShort: 1039 case Primitive::kPrimInt: 1040 case Primitive::kPrimNot: { 1041 return Location::RegisterLocation(R0); 1042 } 1043 1044 case Primitive::kPrimFloat: { 1045 return Location::FpuRegisterLocation(S0); 1046 } 1047 1048 case Primitive::kPrimLong: { 1049 return Location::RegisterPairLocation(R0, R1); 1050 } 1051 1052 case Primitive::kPrimDouble: { 1053 return Location::FpuRegisterPairLocation(S0, S1); 1054 } 1055 1056 case Primitive::kPrimVoid: 1057 return Location::NoLocation(); 1058 } 1059 1060 UNREACHABLE(); 1061 } 1062 1063 Location InvokeDexCallingConventionVisitorARM::GetMethodLocation() const { 1064 return Location::RegisterLocation(kMethodRegisterArgument); 1065 } 1066 1067 void CodeGeneratorARM::Move32(Location destination, Location source) { 1068 if (source.Equals(destination)) { 1069 return; 1070 } 1071 if (destination.IsRegister()) { 1072 if (source.IsRegister()) { 1073 __ Mov(destination.AsRegister<Register>(), source.AsRegister<Register>()); 1074 } else if (source.IsFpuRegister()) { 1075 __ vmovrs(destination.AsRegister<Register>(), source.AsFpuRegister<SRegister>()); 1076 } else { 1077 __ LoadFromOffset(kLoadWord, destination.AsRegister<Register>(), SP, source.GetStackIndex()); 1078 } 1079 } else if (destination.IsFpuRegister()) { 1080 if (source.IsRegister()) { 1081 __ vmovsr(destination.AsFpuRegister<SRegister>(), source.AsRegister<Register>()); 1082 } else if (source.IsFpuRegister()) { 1083 __ vmovs(destination.AsFpuRegister<SRegister>(), source.AsFpuRegister<SRegister>()); 1084 } else { 1085 __ LoadSFromOffset(destination.AsFpuRegister<SRegister>(), SP, source.GetStackIndex()); 1086 } 1087 } else { 1088 DCHECK(destination.IsStackSlot()) << destination; 1089 if (source.IsRegister()) { 1090 __ StoreToOffset(kStoreWord, source.AsRegister<Register>(), SP, destination.GetStackIndex()); 1091 } else if (source.IsFpuRegister()) { 1092 __ StoreSToOffset(source.AsFpuRegister<SRegister>(), SP, destination.GetStackIndex()); 1093 } else { 1094 DCHECK(source.IsStackSlot()) << source; 1095 __ LoadFromOffset(kLoadWord, IP, SP, source.GetStackIndex()); 1096 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 1097 } 1098 } 1099 } 1100 1101 void CodeGeneratorARM::Move64(Location destination, Location source) { 1102 if (source.Equals(destination)) { 1103 return; 1104 } 1105 if (destination.IsRegisterPair()) { 1106 if (source.IsRegisterPair()) { 1107 EmitParallelMoves( 1108 Location::RegisterLocation(source.AsRegisterPairHigh<Register>()), 1109 Location::RegisterLocation(destination.AsRegisterPairHigh<Register>()), 1110 Primitive::kPrimInt, 1111 Location::RegisterLocation(source.AsRegisterPairLow<Register>()), 1112 Location::RegisterLocation(destination.AsRegisterPairLow<Register>()), 1113 Primitive::kPrimInt); 1114 } else if (source.IsFpuRegister()) { 1115 UNIMPLEMENTED(FATAL); 1116 } else if (source.IsFpuRegisterPair()) { 1117 __ vmovrrd(destination.AsRegisterPairLow<Register>(), 1118 destination.AsRegisterPairHigh<Register>(), 1119 FromLowSToD(source.AsFpuRegisterPairLow<SRegister>())); 1120 } else { 1121 DCHECK(source.IsDoubleStackSlot()); 1122 DCHECK(ExpectedPairLayout(destination)); 1123 __ LoadFromOffset(kLoadWordPair, destination.AsRegisterPairLow<Register>(), 1124 SP, source.GetStackIndex()); 1125 } 1126 } else if (destination.IsFpuRegisterPair()) { 1127 if (source.IsDoubleStackSlot()) { 1128 __ LoadDFromOffset(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 1129 SP, 1130 source.GetStackIndex()); 1131 } else if (source.IsRegisterPair()) { 1132 __ vmovdrr(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 1133 source.AsRegisterPairLow<Register>(), 1134 source.AsRegisterPairHigh<Register>()); 1135 } else { 1136 UNIMPLEMENTED(FATAL); 1137 } 1138 } else { 1139 DCHECK(destination.IsDoubleStackSlot()); 1140 if (source.IsRegisterPair()) { 1141 // No conflict possible, so just do the moves. 1142 if (source.AsRegisterPairLow<Register>() == R1) { 1143 DCHECK_EQ(source.AsRegisterPairHigh<Register>(), R2); 1144 __ StoreToOffset(kStoreWord, R1, SP, destination.GetStackIndex()); 1145 __ StoreToOffset(kStoreWord, R2, SP, destination.GetHighStackIndex(kArmWordSize)); 1146 } else { 1147 __ StoreToOffset(kStoreWordPair, source.AsRegisterPairLow<Register>(), 1148 SP, destination.GetStackIndex()); 1149 } 1150 } else if (source.IsFpuRegisterPair()) { 1151 __ StoreDToOffset(FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()), 1152 SP, 1153 destination.GetStackIndex()); 1154 } else { 1155 DCHECK(source.IsDoubleStackSlot()); 1156 EmitParallelMoves( 1157 Location::StackSlot(source.GetStackIndex()), 1158 Location::StackSlot(destination.GetStackIndex()), 1159 Primitive::kPrimInt, 1160 Location::StackSlot(source.GetHighStackIndex(kArmWordSize)), 1161 Location::StackSlot(destination.GetHighStackIndex(kArmWordSize)), 1162 Primitive::kPrimInt); 1163 } 1164 } 1165 } 1166 1167 void CodeGeneratorARM::MoveConstant(Location location, int32_t value) { 1168 DCHECK(location.IsRegister()); 1169 __ LoadImmediate(location.AsRegister<Register>(), value); 1170 } 1171 1172 void CodeGeneratorARM::MoveLocation(Location dst, Location src, Primitive::Type dst_type) { 1173 HParallelMove move(GetGraph()->GetArena()); 1174 move.AddMove(src, dst, dst_type, nullptr); 1175 GetMoveResolver()->EmitNativeCode(&move); 1176 } 1177 1178 void CodeGeneratorARM::AddLocationAsTemp(Location location, LocationSummary* locations) { 1179 if (location.IsRegister()) { 1180 locations->AddTemp(location); 1181 } else if (location.IsRegisterPair()) { 1182 locations->AddTemp(Location::RegisterLocation(location.AsRegisterPairLow<Register>())); 1183 locations->AddTemp(Location::RegisterLocation(location.AsRegisterPairHigh<Register>())); 1184 } else { 1185 UNIMPLEMENTED(FATAL) << "AddLocationAsTemp not implemented for location " << location; 1186 } 1187 } 1188 1189 void CodeGeneratorARM::InvokeRuntime(QuickEntrypointEnum entrypoint, 1190 HInstruction* instruction, 1191 uint32_t dex_pc, 1192 SlowPathCode* slow_path) { 1193 InvokeRuntime(GetThreadOffset<kArmWordSize>(entrypoint).Int32Value(), 1194 instruction, 1195 dex_pc, 1196 slow_path); 1197 } 1198 1199 void CodeGeneratorARM::InvokeRuntime(int32_t entry_point_offset, 1200 HInstruction* instruction, 1201 uint32_t dex_pc, 1202 SlowPathCode* slow_path) { 1203 ValidateInvokeRuntime(instruction, slow_path); 1204 __ LoadFromOffset(kLoadWord, LR, TR, entry_point_offset); 1205 __ blx(LR); 1206 RecordPcInfo(instruction, dex_pc, slow_path); 1207 } 1208 1209 void InstructionCodeGeneratorARM::HandleGoto(HInstruction* got, HBasicBlock* successor) { 1210 DCHECK(!successor->IsExitBlock()); 1211 1212 HBasicBlock* block = got->GetBlock(); 1213 HInstruction* previous = got->GetPrevious(); 1214 1215 HLoopInformation* info = block->GetLoopInformation(); 1216 if (info != nullptr && info->IsBackEdge(*block) && info->HasSuspendCheck()) { 1217 codegen_->ClearSpillSlotsFromLoopPhisInStackMap(info->GetSuspendCheck()); 1218 GenerateSuspendCheck(info->GetSuspendCheck(), successor); 1219 return; 1220 } 1221 1222 if (block->IsEntryBlock() && (previous != nullptr) && previous->IsSuspendCheck()) { 1223 GenerateSuspendCheck(previous->AsSuspendCheck(), nullptr); 1224 } 1225 if (!codegen_->GoesToNextBlock(got->GetBlock(), successor)) { 1226 __ b(codegen_->GetLabelOf(successor)); 1227 } 1228 } 1229 1230 void LocationsBuilderARM::VisitGoto(HGoto* got) { 1231 got->SetLocations(nullptr); 1232 } 1233 1234 void InstructionCodeGeneratorARM::VisitGoto(HGoto* got) { 1235 HandleGoto(got, got->GetSuccessor()); 1236 } 1237 1238 void LocationsBuilderARM::VisitTryBoundary(HTryBoundary* try_boundary) { 1239 try_boundary->SetLocations(nullptr); 1240 } 1241 1242 void InstructionCodeGeneratorARM::VisitTryBoundary(HTryBoundary* try_boundary) { 1243 HBasicBlock* successor = try_boundary->GetNormalFlowSuccessor(); 1244 if (!successor->IsExitBlock()) { 1245 HandleGoto(try_boundary, successor); 1246 } 1247 } 1248 1249 void LocationsBuilderARM::VisitExit(HExit* exit) { 1250 exit->SetLocations(nullptr); 1251 } 1252 1253 void InstructionCodeGeneratorARM::VisitExit(HExit* exit ATTRIBUTE_UNUSED) { 1254 } 1255 1256 void InstructionCodeGeneratorARM::GenerateFPJumps(HCondition* cond, 1257 Label* true_label, 1258 Label* false_label ATTRIBUTE_UNUSED) { 1259 __ vmstat(); // transfer FP status register to ARM APSR. 1260 __ b(true_label, ARMFPCondition(cond->GetCondition(), cond->IsGtBias())); 1261 } 1262 1263 void InstructionCodeGeneratorARM::GenerateLongComparesAndJumps(HCondition* cond, 1264 Label* true_label, 1265 Label* false_label) { 1266 LocationSummary* locations = cond->GetLocations(); 1267 Location left = locations->InAt(0); 1268 Location right = locations->InAt(1); 1269 IfCondition if_cond = cond->GetCondition(); 1270 1271 Register left_high = left.AsRegisterPairHigh<Register>(); 1272 Register left_low = left.AsRegisterPairLow<Register>(); 1273 IfCondition true_high_cond = if_cond; 1274 IfCondition false_high_cond = cond->GetOppositeCondition(); 1275 Condition final_condition = ARMUnsignedCondition(if_cond); // unsigned on lower part 1276 1277 // Set the conditions for the test, remembering that == needs to be 1278 // decided using the low words. 1279 // TODO: consider avoiding jumps with temporary and CMP low+SBC high 1280 switch (if_cond) { 1281 case kCondEQ: 1282 case kCondNE: 1283 // Nothing to do. 1284 break; 1285 case kCondLT: 1286 false_high_cond = kCondGT; 1287 break; 1288 case kCondLE: 1289 true_high_cond = kCondLT; 1290 break; 1291 case kCondGT: 1292 false_high_cond = kCondLT; 1293 break; 1294 case kCondGE: 1295 true_high_cond = kCondGT; 1296 break; 1297 case kCondB: 1298 false_high_cond = kCondA; 1299 break; 1300 case kCondBE: 1301 true_high_cond = kCondB; 1302 break; 1303 case kCondA: 1304 false_high_cond = kCondB; 1305 break; 1306 case kCondAE: 1307 true_high_cond = kCondA; 1308 break; 1309 } 1310 if (right.IsConstant()) { 1311 int64_t value = right.GetConstant()->AsLongConstant()->GetValue(); 1312 int32_t val_low = Low32Bits(value); 1313 int32_t val_high = High32Bits(value); 1314 1315 __ CmpConstant(left_high, val_high); 1316 if (if_cond == kCondNE) { 1317 __ b(true_label, ARMCondition(true_high_cond)); 1318 } else if (if_cond == kCondEQ) { 1319 __ b(false_label, ARMCondition(false_high_cond)); 1320 } else { 1321 __ b(true_label, ARMCondition(true_high_cond)); 1322 __ b(false_label, ARMCondition(false_high_cond)); 1323 } 1324 // Must be equal high, so compare the lows. 1325 __ CmpConstant(left_low, val_low); 1326 } else { 1327 Register right_high = right.AsRegisterPairHigh<Register>(); 1328 Register right_low = right.AsRegisterPairLow<Register>(); 1329 1330 __ cmp(left_high, ShifterOperand(right_high)); 1331 if (if_cond == kCondNE) { 1332 __ b(true_label, ARMCondition(true_high_cond)); 1333 } else if (if_cond == kCondEQ) { 1334 __ b(false_label, ARMCondition(false_high_cond)); 1335 } else { 1336 __ b(true_label, ARMCondition(true_high_cond)); 1337 __ b(false_label, ARMCondition(false_high_cond)); 1338 } 1339 // Must be equal high, so compare the lows. 1340 __ cmp(left_low, ShifterOperand(right_low)); 1341 } 1342 // The last comparison might be unsigned. 1343 // TODO: optimize cases where this is always true/false 1344 __ b(true_label, final_condition); 1345 } 1346 1347 void InstructionCodeGeneratorARM::GenerateCompareTestAndBranch(HCondition* condition, 1348 Label* true_target_in, 1349 Label* false_target_in) { 1350 // Generated branching requires both targets to be explicit. If either of the 1351 // targets is nullptr (fallthrough) use and bind `fallthrough_target` instead. 1352 Label fallthrough_target; 1353 Label* true_target = true_target_in == nullptr ? &fallthrough_target : true_target_in; 1354 Label* false_target = false_target_in == nullptr ? &fallthrough_target : false_target_in; 1355 1356 LocationSummary* locations = condition->GetLocations(); 1357 Location left = locations->InAt(0); 1358 Location right = locations->InAt(1); 1359 1360 Primitive::Type type = condition->InputAt(0)->GetType(); 1361 switch (type) { 1362 case Primitive::kPrimLong: 1363 GenerateLongComparesAndJumps(condition, true_target, false_target); 1364 break; 1365 case Primitive::kPrimFloat: 1366 __ vcmps(left.AsFpuRegister<SRegister>(), right.AsFpuRegister<SRegister>()); 1367 GenerateFPJumps(condition, true_target, false_target); 1368 break; 1369 case Primitive::kPrimDouble: 1370 __ vcmpd(FromLowSToD(left.AsFpuRegisterPairLow<SRegister>()), 1371 FromLowSToD(right.AsFpuRegisterPairLow<SRegister>())); 1372 GenerateFPJumps(condition, true_target, false_target); 1373 break; 1374 default: 1375 LOG(FATAL) << "Unexpected compare type " << type; 1376 } 1377 1378 if (false_target != &fallthrough_target) { 1379 __ b(false_target); 1380 } 1381 1382 if (fallthrough_target.IsLinked()) { 1383 __ Bind(&fallthrough_target); 1384 } 1385 } 1386 1387 void InstructionCodeGeneratorARM::GenerateTestAndBranch(HInstruction* instruction, 1388 size_t condition_input_index, 1389 Label* true_target, 1390 Label* false_target) { 1391 HInstruction* cond = instruction->InputAt(condition_input_index); 1392 1393 if (true_target == nullptr && false_target == nullptr) { 1394 // Nothing to do. The code always falls through. 1395 return; 1396 } else if (cond->IsIntConstant()) { 1397 // Constant condition, statically compared against "true" (integer value 1). 1398 if (cond->AsIntConstant()->IsTrue()) { 1399 if (true_target != nullptr) { 1400 __ b(true_target); 1401 } 1402 } else { 1403 DCHECK(cond->AsIntConstant()->IsFalse()) << cond->AsIntConstant()->GetValue(); 1404 if (false_target != nullptr) { 1405 __ b(false_target); 1406 } 1407 } 1408 return; 1409 } 1410 1411 // The following code generates these patterns: 1412 // (1) true_target == nullptr && false_target != nullptr 1413 // - opposite condition true => branch to false_target 1414 // (2) true_target != nullptr && false_target == nullptr 1415 // - condition true => branch to true_target 1416 // (3) true_target != nullptr && false_target != nullptr 1417 // - condition true => branch to true_target 1418 // - branch to false_target 1419 if (IsBooleanValueOrMaterializedCondition(cond)) { 1420 // Condition has been materialized, compare the output to 0. 1421 Location cond_val = instruction->GetLocations()->InAt(condition_input_index); 1422 DCHECK(cond_val.IsRegister()); 1423 if (true_target == nullptr) { 1424 __ CompareAndBranchIfZero(cond_val.AsRegister<Register>(), false_target); 1425 } else { 1426 __ CompareAndBranchIfNonZero(cond_val.AsRegister<Register>(), true_target); 1427 } 1428 } else { 1429 // Condition has not been materialized. Use its inputs as the comparison and 1430 // its condition as the branch condition. 1431 HCondition* condition = cond->AsCondition(); 1432 1433 // If this is a long or FP comparison that has been folded into 1434 // the HCondition, generate the comparison directly. 1435 Primitive::Type type = condition->InputAt(0)->GetType(); 1436 if (type == Primitive::kPrimLong || Primitive::IsFloatingPointType(type)) { 1437 GenerateCompareTestAndBranch(condition, true_target, false_target); 1438 return; 1439 } 1440 1441 LocationSummary* locations = cond->GetLocations(); 1442 DCHECK(locations->InAt(0).IsRegister()); 1443 Register left = locations->InAt(0).AsRegister<Register>(); 1444 Location right = locations->InAt(1); 1445 if (right.IsRegister()) { 1446 __ cmp(left, ShifterOperand(right.AsRegister<Register>())); 1447 } else { 1448 DCHECK(right.IsConstant()); 1449 __ CmpConstant(left, CodeGenerator::GetInt32ValueOf(right.GetConstant())); 1450 } 1451 if (true_target == nullptr) { 1452 __ b(false_target, ARMCondition(condition->GetOppositeCondition())); 1453 } else { 1454 __ b(true_target, ARMCondition(condition->GetCondition())); 1455 } 1456 } 1457 1458 // If neither branch falls through (case 3), the conditional branch to `true_target` 1459 // was already emitted (case 2) and we need to emit a jump to `false_target`. 1460 if (true_target != nullptr && false_target != nullptr) { 1461 __ b(false_target); 1462 } 1463 } 1464 1465 void LocationsBuilderARM::VisitIf(HIf* if_instr) { 1466 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(if_instr); 1467 if (IsBooleanValueOrMaterializedCondition(if_instr->InputAt(0))) { 1468 locations->SetInAt(0, Location::RequiresRegister()); 1469 } 1470 } 1471 1472 void InstructionCodeGeneratorARM::VisitIf(HIf* if_instr) { 1473 HBasicBlock* true_successor = if_instr->IfTrueSuccessor(); 1474 HBasicBlock* false_successor = if_instr->IfFalseSuccessor(); 1475 Label* true_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), true_successor) ? 1476 nullptr : codegen_->GetLabelOf(true_successor); 1477 Label* false_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), false_successor) ? 1478 nullptr : codegen_->GetLabelOf(false_successor); 1479 GenerateTestAndBranch(if_instr, /* condition_input_index */ 0, true_target, false_target); 1480 } 1481 1482 void LocationsBuilderARM::VisitDeoptimize(HDeoptimize* deoptimize) { 1483 LocationSummary* locations = new (GetGraph()->GetArena()) 1484 LocationSummary(deoptimize, LocationSummary::kCallOnSlowPath); 1485 if (IsBooleanValueOrMaterializedCondition(deoptimize->InputAt(0))) { 1486 locations->SetInAt(0, Location::RequiresRegister()); 1487 } 1488 } 1489 1490 void InstructionCodeGeneratorARM::VisitDeoptimize(HDeoptimize* deoptimize) { 1491 SlowPathCode* slow_path = deopt_slow_paths_.NewSlowPath<DeoptimizationSlowPathARM>(deoptimize); 1492 GenerateTestAndBranch(deoptimize, 1493 /* condition_input_index */ 0, 1494 slow_path->GetEntryLabel(), 1495 /* false_target */ nullptr); 1496 } 1497 1498 void LocationsBuilderARM::VisitSelect(HSelect* select) { 1499 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(select); 1500 if (Primitive::IsFloatingPointType(select->GetType())) { 1501 locations->SetInAt(0, Location::RequiresFpuRegister()); 1502 locations->SetInAt(1, Location::RequiresFpuRegister()); 1503 } else { 1504 locations->SetInAt(0, Location::RequiresRegister()); 1505 locations->SetInAt(1, Location::RequiresRegister()); 1506 } 1507 if (IsBooleanValueOrMaterializedCondition(select->GetCondition())) { 1508 locations->SetInAt(2, Location::RequiresRegister()); 1509 } 1510 locations->SetOut(Location::SameAsFirstInput()); 1511 } 1512 1513 void InstructionCodeGeneratorARM::VisitSelect(HSelect* select) { 1514 LocationSummary* locations = select->GetLocations(); 1515 Label false_target; 1516 GenerateTestAndBranch(select, 1517 /* condition_input_index */ 2, 1518 /* true_target */ nullptr, 1519 &false_target); 1520 codegen_->MoveLocation(locations->Out(), locations->InAt(1), select->GetType()); 1521 __ Bind(&false_target); 1522 } 1523 1524 void LocationsBuilderARM::VisitNativeDebugInfo(HNativeDebugInfo* info) { 1525 new (GetGraph()->GetArena()) LocationSummary(info); 1526 } 1527 1528 void InstructionCodeGeneratorARM::VisitNativeDebugInfo(HNativeDebugInfo*) { 1529 // MaybeRecordNativeDebugInfo is already called implicitly in CodeGenerator::Compile. 1530 } 1531 1532 void CodeGeneratorARM::GenerateNop() { 1533 __ nop(); 1534 } 1535 1536 void LocationsBuilderARM::HandleCondition(HCondition* cond) { 1537 LocationSummary* locations = 1538 new (GetGraph()->GetArena()) LocationSummary(cond, LocationSummary::kNoCall); 1539 // Handle the long/FP comparisons made in instruction simplification. 1540 switch (cond->InputAt(0)->GetType()) { 1541 case Primitive::kPrimLong: 1542 locations->SetInAt(0, Location::RequiresRegister()); 1543 locations->SetInAt(1, Location::RegisterOrConstant(cond->InputAt(1))); 1544 if (!cond->IsEmittedAtUseSite()) { 1545 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 1546 } 1547 break; 1548 1549 case Primitive::kPrimFloat: 1550 case Primitive::kPrimDouble: 1551 locations->SetInAt(0, Location::RequiresFpuRegister()); 1552 locations->SetInAt(1, Location::RequiresFpuRegister()); 1553 if (!cond->IsEmittedAtUseSite()) { 1554 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 1555 } 1556 break; 1557 1558 default: 1559 locations->SetInAt(0, Location::RequiresRegister()); 1560 locations->SetInAt(1, Location::RegisterOrConstant(cond->InputAt(1))); 1561 if (!cond->IsEmittedAtUseSite()) { 1562 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 1563 } 1564 } 1565 } 1566 1567 void InstructionCodeGeneratorARM::HandleCondition(HCondition* cond) { 1568 if (cond->IsEmittedAtUseSite()) { 1569 return; 1570 } 1571 1572 LocationSummary* locations = cond->GetLocations(); 1573 Location left = locations->InAt(0); 1574 Location right = locations->InAt(1); 1575 Register out = locations->Out().AsRegister<Register>(); 1576 Label true_label, false_label; 1577 1578 switch (cond->InputAt(0)->GetType()) { 1579 default: { 1580 // Integer case. 1581 if (right.IsRegister()) { 1582 __ cmp(left.AsRegister<Register>(), ShifterOperand(right.AsRegister<Register>())); 1583 } else { 1584 DCHECK(right.IsConstant()); 1585 __ CmpConstant(left.AsRegister<Register>(), 1586 CodeGenerator::GetInt32ValueOf(right.GetConstant())); 1587 } 1588 __ it(ARMCondition(cond->GetCondition()), kItElse); 1589 __ mov(locations->Out().AsRegister<Register>(), ShifterOperand(1), 1590 ARMCondition(cond->GetCondition())); 1591 __ mov(locations->Out().AsRegister<Register>(), ShifterOperand(0), 1592 ARMCondition(cond->GetOppositeCondition())); 1593 return; 1594 } 1595 case Primitive::kPrimLong: 1596 GenerateLongComparesAndJumps(cond, &true_label, &false_label); 1597 break; 1598 case Primitive::kPrimFloat: 1599 __ vcmps(left.AsFpuRegister<SRegister>(), right.AsFpuRegister<SRegister>()); 1600 GenerateFPJumps(cond, &true_label, &false_label); 1601 break; 1602 case Primitive::kPrimDouble: 1603 __ vcmpd(FromLowSToD(left.AsFpuRegisterPairLow<SRegister>()), 1604 FromLowSToD(right.AsFpuRegisterPairLow<SRegister>())); 1605 GenerateFPJumps(cond, &true_label, &false_label); 1606 break; 1607 } 1608 1609 // Convert the jumps into the result. 1610 Label done_label; 1611 1612 // False case: result = 0. 1613 __ Bind(&false_label); 1614 __ LoadImmediate(out, 0); 1615 __ b(&done_label); 1616 1617 // True case: result = 1. 1618 __ Bind(&true_label); 1619 __ LoadImmediate(out, 1); 1620 __ Bind(&done_label); 1621 } 1622 1623 void LocationsBuilderARM::VisitEqual(HEqual* comp) { 1624 HandleCondition(comp); 1625 } 1626 1627 void InstructionCodeGeneratorARM::VisitEqual(HEqual* comp) { 1628 HandleCondition(comp); 1629 } 1630 1631 void LocationsBuilderARM::VisitNotEqual(HNotEqual* comp) { 1632 HandleCondition(comp); 1633 } 1634 1635 void InstructionCodeGeneratorARM::VisitNotEqual(HNotEqual* comp) { 1636 HandleCondition(comp); 1637 } 1638 1639 void LocationsBuilderARM::VisitLessThan(HLessThan* comp) { 1640 HandleCondition(comp); 1641 } 1642 1643 void InstructionCodeGeneratorARM::VisitLessThan(HLessThan* comp) { 1644 HandleCondition(comp); 1645 } 1646 1647 void LocationsBuilderARM::VisitLessThanOrEqual(HLessThanOrEqual* comp) { 1648 HandleCondition(comp); 1649 } 1650 1651 void InstructionCodeGeneratorARM::VisitLessThanOrEqual(HLessThanOrEqual* comp) { 1652 HandleCondition(comp); 1653 } 1654 1655 void LocationsBuilderARM::VisitGreaterThan(HGreaterThan* comp) { 1656 HandleCondition(comp); 1657 } 1658 1659 void InstructionCodeGeneratorARM::VisitGreaterThan(HGreaterThan* comp) { 1660 HandleCondition(comp); 1661 } 1662 1663 void LocationsBuilderARM::VisitGreaterThanOrEqual(HGreaterThanOrEqual* comp) { 1664 HandleCondition(comp); 1665 } 1666 1667 void InstructionCodeGeneratorARM::VisitGreaterThanOrEqual(HGreaterThanOrEqual* comp) { 1668 HandleCondition(comp); 1669 } 1670 1671 void LocationsBuilderARM::VisitBelow(HBelow* comp) { 1672 HandleCondition(comp); 1673 } 1674 1675 void InstructionCodeGeneratorARM::VisitBelow(HBelow* comp) { 1676 HandleCondition(comp); 1677 } 1678 1679 void LocationsBuilderARM::VisitBelowOrEqual(HBelowOrEqual* comp) { 1680 HandleCondition(comp); 1681 } 1682 1683 void InstructionCodeGeneratorARM::VisitBelowOrEqual(HBelowOrEqual* comp) { 1684 HandleCondition(comp); 1685 } 1686 1687 void LocationsBuilderARM::VisitAbove(HAbove* comp) { 1688 HandleCondition(comp); 1689 } 1690 1691 void InstructionCodeGeneratorARM::VisitAbove(HAbove* comp) { 1692 HandleCondition(comp); 1693 } 1694 1695 void LocationsBuilderARM::VisitAboveOrEqual(HAboveOrEqual* comp) { 1696 HandleCondition(comp); 1697 } 1698 1699 void InstructionCodeGeneratorARM::VisitAboveOrEqual(HAboveOrEqual* comp) { 1700 HandleCondition(comp); 1701 } 1702 1703 void LocationsBuilderARM::VisitIntConstant(HIntConstant* constant) { 1704 LocationSummary* locations = 1705 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1706 locations->SetOut(Location::ConstantLocation(constant)); 1707 } 1708 1709 void InstructionCodeGeneratorARM::VisitIntConstant(HIntConstant* constant ATTRIBUTE_UNUSED) { 1710 // Will be generated at use site. 1711 } 1712 1713 void LocationsBuilderARM::VisitNullConstant(HNullConstant* constant) { 1714 LocationSummary* locations = 1715 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1716 locations->SetOut(Location::ConstantLocation(constant)); 1717 } 1718 1719 void InstructionCodeGeneratorARM::VisitNullConstant(HNullConstant* constant ATTRIBUTE_UNUSED) { 1720 // Will be generated at use site. 1721 } 1722 1723 void LocationsBuilderARM::VisitLongConstant(HLongConstant* constant) { 1724 LocationSummary* locations = 1725 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1726 locations->SetOut(Location::ConstantLocation(constant)); 1727 } 1728 1729 void InstructionCodeGeneratorARM::VisitLongConstant(HLongConstant* constant ATTRIBUTE_UNUSED) { 1730 // Will be generated at use site. 1731 } 1732 1733 void LocationsBuilderARM::VisitFloatConstant(HFloatConstant* constant) { 1734 LocationSummary* locations = 1735 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1736 locations->SetOut(Location::ConstantLocation(constant)); 1737 } 1738 1739 void InstructionCodeGeneratorARM::VisitFloatConstant(HFloatConstant* constant ATTRIBUTE_UNUSED) { 1740 // Will be generated at use site. 1741 } 1742 1743 void LocationsBuilderARM::VisitDoubleConstant(HDoubleConstant* constant) { 1744 LocationSummary* locations = 1745 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1746 locations->SetOut(Location::ConstantLocation(constant)); 1747 } 1748 1749 void InstructionCodeGeneratorARM::VisitDoubleConstant(HDoubleConstant* constant ATTRIBUTE_UNUSED) { 1750 // Will be generated at use site. 1751 } 1752 1753 void LocationsBuilderARM::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) { 1754 memory_barrier->SetLocations(nullptr); 1755 } 1756 1757 void InstructionCodeGeneratorARM::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) { 1758 codegen_->GenerateMemoryBarrier(memory_barrier->GetBarrierKind()); 1759 } 1760 1761 void LocationsBuilderARM::VisitReturnVoid(HReturnVoid* ret) { 1762 ret->SetLocations(nullptr); 1763 } 1764 1765 void InstructionCodeGeneratorARM::VisitReturnVoid(HReturnVoid* ret ATTRIBUTE_UNUSED) { 1766 codegen_->GenerateFrameExit(); 1767 } 1768 1769 void LocationsBuilderARM::VisitReturn(HReturn* ret) { 1770 LocationSummary* locations = 1771 new (GetGraph()->GetArena()) LocationSummary(ret, LocationSummary::kNoCall); 1772 locations->SetInAt(0, parameter_visitor_.GetReturnLocation(ret->InputAt(0)->GetType())); 1773 } 1774 1775 void InstructionCodeGeneratorARM::VisitReturn(HReturn* ret ATTRIBUTE_UNUSED) { 1776 codegen_->GenerateFrameExit(); 1777 } 1778 1779 void LocationsBuilderARM::VisitInvokeUnresolved(HInvokeUnresolved* invoke) { 1780 // The trampoline uses the same calling convention as dex calling conventions, 1781 // except instead of loading arg0/r0 with the target Method*, arg0/r0 will contain 1782 // the method_idx. 1783 HandleInvoke(invoke); 1784 } 1785 1786 void InstructionCodeGeneratorARM::VisitInvokeUnresolved(HInvokeUnresolved* invoke) { 1787 codegen_->GenerateInvokeUnresolvedRuntimeCall(invoke); 1788 } 1789 1790 void LocationsBuilderARM::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) { 1791 // Explicit clinit checks triggered by static invokes must have been pruned by 1792 // art::PrepareForRegisterAllocation. 1793 DCHECK(!invoke->IsStaticWithExplicitClinitCheck()); 1794 1795 IntrinsicLocationsBuilderARM intrinsic(GetGraph()->GetArena(), 1796 codegen_->GetAssembler(), 1797 codegen_->GetInstructionSetFeatures()); 1798 if (intrinsic.TryDispatch(invoke)) { 1799 if (invoke->GetLocations()->CanCall() && invoke->HasPcRelativeDexCache()) { 1800 invoke->GetLocations()->SetInAt(invoke->GetSpecialInputIndex(), Location::Any()); 1801 } 1802 return; 1803 } 1804 1805 HandleInvoke(invoke); 1806 1807 // For PC-relative dex cache the invoke has an extra input, the PC-relative address base. 1808 if (invoke->HasPcRelativeDexCache()) { 1809 invoke->GetLocations()->SetInAt(invoke->GetSpecialInputIndex(), Location::RequiresRegister()); 1810 } 1811 } 1812 1813 static bool TryGenerateIntrinsicCode(HInvoke* invoke, CodeGeneratorARM* codegen) { 1814 if (invoke->GetLocations()->Intrinsified()) { 1815 IntrinsicCodeGeneratorARM intrinsic(codegen); 1816 intrinsic.Dispatch(invoke); 1817 return true; 1818 } 1819 return false; 1820 } 1821 1822 void InstructionCodeGeneratorARM::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) { 1823 // Explicit clinit checks triggered by static invokes must have been pruned by 1824 // art::PrepareForRegisterAllocation. 1825 DCHECK(!invoke->IsStaticWithExplicitClinitCheck()); 1826 1827 if (TryGenerateIntrinsicCode(invoke, codegen_)) { 1828 return; 1829 } 1830 1831 LocationSummary* locations = invoke->GetLocations(); 1832 codegen_->GenerateStaticOrDirectCall( 1833 invoke, locations->HasTemps() ? locations->GetTemp(0) : Location::NoLocation()); 1834 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 1835 } 1836 1837 void LocationsBuilderARM::HandleInvoke(HInvoke* invoke) { 1838 InvokeDexCallingConventionVisitorARM calling_convention_visitor; 1839 CodeGenerator::CreateCommonInvokeLocationSummary(invoke, &calling_convention_visitor); 1840 } 1841 1842 void LocationsBuilderARM::VisitInvokeVirtual(HInvokeVirtual* invoke) { 1843 IntrinsicLocationsBuilderARM intrinsic(GetGraph()->GetArena(), 1844 codegen_->GetAssembler(), 1845 codegen_->GetInstructionSetFeatures()); 1846 if (intrinsic.TryDispatch(invoke)) { 1847 return; 1848 } 1849 1850 HandleInvoke(invoke); 1851 } 1852 1853 void InstructionCodeGeneratorARM::VisitInvokeVirtual(HInvokeVirtual* invoke) { 1854 if (TryGenerateIntrinsicCode(invoke, codegen_)) { 1855 return; 1856 } 1857 1858 codegen_->GenerateVirtualCall(invoke, invoke->GetLocations()->GetTemp(0)); 1859 DCHECK(!codegen_->IsLeafMethod()); 1860 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 1861 } 1862 1863 void LocationsBuilderARM::VisitInvokeInterface(HInvokeInterface* invoke) { 1864 HandleInvoke(invoke); 1865 // Add the hidden argument. 1866 invoke->GetLocations()->AddTemp(Location::RegisterLocation(R12)); 1867 } 1868 1869 void InstructionCodeGeneratorARM::VisitInvokeInterface(HInvokeInterface* invoke) { 1870 // TODO: b/18116999, our IMTs can miss an IncompatibleClassChangeError. 1871 LocationSummary* locations = invoke->GetLocations(); 1872 Register temp = locations->GetTemp(0).AsRegister<Register>(); 1873 Register hidden_reg = locations->GetTemp(1).AsRegister<Register>(); 1874 Location receiver = locations->InAt(0); 1875 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 1876 1877 // Set the hidden argument. This is safe to do this here, as R12 1878 // won't be modified thereafter, before the `blx` (call) instruction. 1879 DCHECK_EQ(R12, hidden_reg); 1880 __ LoadImmediate(hidden_reg, invoke->GetDexMethodIndex()); 1881 1882 if (receiver.IsStackSlot()) { 1883 __ LoadFromOffset(kLoadWord, temp, SP, receiver.GetStackIndex()); 1884 // /* HeapReference<Class> */ temp = temp->klass_ 1885 __ LoadFromOffset(kLoadWord, temp, temp, class_offset); 1886 } else { 1887 // /* HeapReference<Class> */ temp = receiver->klass_ 1888 __ LoadFromOffset(kLoadWord, temp, receiver.AsRegister<Register>(), class_offset); 1889 } 1890 codegen_->MaybeRecordImplicitNullCheck(invoke); 1891 // Instead of simply (possibly) unpoisoning `temp` here, we should 1892 // emit a read barrier for the previous class reference load. 1893 // However this is not required in practice, as this is an 1894 // intermediate/temporary reference and because the current 1895 // concurrent copying collector keeps the from-space memory 1896 // intact/accessible until the end of the marking phase (the 1897 // concurrent copying collector may not in the future). 1898 __ MaybeUnpoisonHeapReference(temp); 1899 __ LoadFromOffset(kLoadWord, temp, temp, 1900 mirror::Class::ImtPtrOffset(kArmPointerSize).Uint32Value()); 1901 uint32_t method_offset = static_cast<uint32_t>(ImTable::OffsetOfElement( 1902 invoke->GetImtIndex() % ImTable::kSize, kArmPointerSize)); 1903 // temp = temp->GetImtEntryAt(method_offset); 1904 __ LoadFromOffset(kLoadWord, temp, temp, method_offset); 1905 uint32_t entry_point = 1906 ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmWordSize).Int32Value(); 1907 // LR = temp->GetEntryPoint(); 1908 __ LoadFromOffset(kLoadWord, LR, temp, entry_point); 1909 // LR(); 1910 __ blx(LR); 1911 DCHECK(!codegen_->IsLeafMethod()); 1912 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 1913 } 1914 1915 void LocationsBuilderARM::VisitNeg(HNeg* neg) { 1916 LocationSummary* locations = 1917 new (GetGraph()->GetArena()) LocationSummary(neg, LocationSummary::kNoCall); 1918 switch (neg->GetResultType()) { 1919 case Primitive::kPrimInt: { 1920 locations->SetInAt(0, Location::RequiresRegister()); 1921 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 1922 break; 1923 } 1924 case Primitive::kPrimLong: { 1925 locations->SetInAt(0, Location::RequiresRegister()); 1926 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 1927 break; 1928 } 1929 1930 case Primitive::kPrimFloat: 1931 case Primitive::kPrimDouble: 1932 locations->SetInAt(0, Location::RequiresFpuRegister()); 1933 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 1934 break; 1935 1936 default: 1937 LOG(FATAL) << "Unexpected neg type " << neg->GetResultType(); 1938 } 1939 } 1940 1941 void InstructionCodeGeneratorARM::VisitNeg(HNeg* neg) { 1942 LocationSummary* locations = neg->GetLocations(); 1943 Location out = locations->Out(); 1944 Location in = locations->InAt(0); 1945 switch (neg->GetResultType()) { 1946 case Primitive::kPrimInt: 1947 DCHECK(in.IsRegister()); 1948 __ rsb(out.AsRegister<Register>(), in.AsRegister<Register>(), ShifterOperand(0)); 1949 break; 1950 1951 case Primitive::kPrimLong: 1952 DCHECK(in.IsRegisterPair()); 1953 // out.lo = 0 - in.lo (and update the carry/borrow (C) flag) 1954 __ rsbs(out.AsRegisterPairLow<Register>(), 1955 in.AsRegisterPairLow<Register>(), 1956 ShifterOperand(0)); 1957 // We cannot emit an RSC (Reverse Subtract with Carry) 1958 // instruction here, as it does not exist in the Thumb-2 1959 // instruction set. We use the following approach 1960 // using SBC and SUB instead. 1961 // 1962 // out.hi = -C 1963 __ sbc(out.AsRegisterPairHigh<Register>(), 1964 out.AsRegisterPairHigh<Register>(), 1965 ShifterOperand(out.AsRegisterPairHigh<Register>())); 1966 // out.hi = out.hi - in.hi 1967 __ sub(out.AsRegisterPairHigh<Register>(), 1968 out.AsRegisterPairHigh<Register>(), 1969 ShifterOperand(in.AsRegisterPairHigh<Register>())); 1970 break; 1971 1972 case Primitive::kPrimFloat: 1973 DCHECK(in.IsFpuRegister()); 1974 __ vnegs(out.AsFpuRegister<SRegister>(), in.AsFpuRegister<SRegister>()); 1975 break; 1976 1977 case Primitive::kPrimDouble: 1978 DCHECK(in.IsFpuRegisterPair()); 1979 __ vnegd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 1980 FromLowSToD(in.AsFpuRegisterPairLow<SRegister>())); 1981 break; 1982 1983 default: 1984 LOG(FATAL) << "Unexpected neg type " << neg->GetResultType(); 1985 } 1986 } 1987 1988 void LocationsBuilderARM::VisitTypeConversion(HTypeConversion* conversion) { 1989 Primitive::Type result_type = conversion->GetResultType(); 1990 Primitive::Type input_type = conversion->GetInputType(); 1991 DCHECK_NE(result_type, input_type); 1992 1993 // The float-to-long, double-to-long and long-to-float type conversions 1994 // rely on a call to the runtime. 1995 LocationSummary::CallKind call_kind = 1996 (((input_type == Primitive::kPrimFloat || input_type == Primitive::kPrimDouble) 1997 && result_type == Primitive::kPrimLong) 1998 || (input_type == Primitive::kPrimLong && result_type == Primitive::kPrimFloat)) 1999 ? LocationSummary::kCall 2000 : LocationSummary::kNoCall; 2001 LocationSummary* locations = 2002 new (GetGraph()->GetArena()) LocationSummary(conversion, call_kind); 2003 2004 // The Java language does not allow treating boolean as an integral type but 2005 // our bit representation makes it safe. 2006 2007 switch (result_type) { 2008 case Primitive::kPrimByte: 2009 switch (input_type) { 2010 case Primitive::kPrimLong: 2011 // Type conversion from long to byte is a result of code transformations. 2012 case Primitive::kPrimBoolean: 2013 // Boolean input is a result of code transformations. 2014 case Primitive::kPrimShort: 2015 case Primitive::kPrimInt: 2016 case Primitive::kPrimChar: 2017 // Processing a Dex `int-to-byte' instruction. 2018 locations->SetInAt(0, Location::RequiresRegister()); 2019 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2020 break; 2021 2022 default: 2023 LOG(FATAL) << "Unexpected type conversion from " << input_type 2024 << " to " << result_type; 2025 } 2026 break; 2027 2028 case Primitive::kPrimShort: 2029 switch (input_type) { 2030 case Primitive::kPrimLong: 2031 // Type conversion from long to short is a result of code transformations. 2032 case Primitive::kPrimBoolean: 2033 // Boolean input is a result of code transformations. 2034 case Primitive::kPrimByte: 2035 case Primitive::kPrimInt: 2036 case Primitive::kPrimChar: 2037 // Processing a Dex `int-to-short' instruction. 2038 locations->SetInAt(0, Location::RequiresRegister()); 2039 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2040 break; 2041 2042 default: 2043 LOG(FATAL) << "Unexpected type conversion from " << input_type 2044 << " to " << result_type; 2045 } 2046 break; 2047 2048 case Primitive::kPrimInt: 2049 switch (input_type) { 2050 case Primitive::kPrimLong: 2051 // Processing a Dex `long-to-int' instruction. 2052 locations->SetInAt(0, Location::Any()); 2053 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2054 break; 2055 2056 case Primitive::kPrimFloat: 2057 // Processing a Dex `float-to-int' instruction. 2058 locations->SetInAt(0, Location::RequiresFpuRegister()); 2059 locations->SetOut(Location::RequiresRegister()); 2060 locations->AddTemp(Location::RequiresFpuRegister()); 2061 break; 2062 2063 case Primitive::kPrimDouble: 2064 // Processing a Dex `double-to-int' instruction. 2065 locations->SetInAt(0, Location::RequiresFpuRegister()); 2066 locations->SetOut(Location::RequiresRegister()); 2067 locations->AddTemp(Location::RequiresFpuRegister()); 2068 break; 2069 2070 default: 2071 LOG(FATAL) << "Unexpected type conversion from " << input_type 2072 << " to " << result_type; 2073 } 2074 break; 2075 2076 case Primitive::kPrimLong: 2077 switch (input_type) { 2078 case Primitive::kPrimBoolean: 2079 // Boolean input is a result of code transformations. 2080 case Primitive::kPrimByte: 2081 case Primitive::kPrimShort: 2082 case Primitive::kPrimInt: 2083 case Primitive::kPrimChar: 2084 // Processing a Dex `int-to-long' instruction. 2085 locations->SetInAt(0, Location::RequiresRegister()); 2086 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2087 break; 2088 2089 case Primitive::kPrimFloat: { 2090 // Processing a Dex `float-to-long' instruction. 2091 InvokeRuntimeCallingConvention calling_convention; 2092 locations->SetInAt(0, Location::FpuRegisterLocation( 2093 calling_convention.GetFpuRegisterAt(0))); 2094 locations->SetOut(Location::RegisterPairLocation(R0, R1)); 2095 break; 2096 } 2097 2098 case Primitive::kPrimDouble: { 2099 // Processing a Dex `double-to-long' instruction. 2100 InvokeRuntimeCallingConvention calling_convention; 2101 locations->SetInAt(0, Location::FpuRegisterPairLocation( 2102 calling_convention.GetFpuRegisterAt(0), 2103 calling_convention.GetFpuRegisterAt(1))); 2104 locations->SetOut(Location::RegisterPairLocation(R0, R1)); 2105 break; 2106 } 2107 2108 default: 2109 LOG(FATAL) << "Unexpected type conversion from " << input_type 2110 << " to " << result_type; 2111 } 2112 break; 2113 2114 case Primitive::kPrimChar: 2115 switch (input_type) { 2116 case Primitive::kPrimLong: 2117 // Type conversion from long to char is a result of code transformations. 2118 case Primitive::kPrimBoolean: 2119 // Boolean input is a result of code transformations. 2120 case Primitive::kPrimByte: 2121 case Primitive::kPrimShort: 2122 case Primitive::kPrimInt: 2123 // Processing a Dex `int-to-char' instruction. 2124 locations->SetInAt(0, Location::RequiresRegister()); 2125 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2126 break; 2127 2128 default: 2129 LOG(FATAL) << "Unexpected type conversion from " << input_type 2130 << " to " << result_type; 2131 } 2132 break; 2133 2134 case Primitive::kPrimFloat: 2135 switch (input_type) { 2136 case Primitive::kPrimBoolean: 2137 // Boolean input is a result of code transformations. 2138 case Primitive::kPrimByte: 2139 case Primitive::kPrimShort: 2140 case Primitive::kPrimInt: 2141 case Primitive::kPrimChar: 2142 // Processing a Dex `int-to-float' instruction. 2143 locations->SetInAt(0, Location::RequiresRegister()); 2144 locations->SetOut(Location::RequiresFpuRegister()); 2145 break; 2146 2147 case Primitive::kPrimLong: { 2148 // Processing a Dex `long-to-float' instruction. 2149 InvokeRuntimeCallingConvention calling_convention; 2150 locations->SetInAt(0, Location::RegisterPairLocation( 2151 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1))); 2152 locations->SetOut(Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(0))); 2153 break; 2154 } 2155 2156 case Primitive::kPrimDouble: 2157 // Processing a Dex `double-to-float' instruction. 2158 locations->SetInAt(0, Location::RequiresFpuRegister()); 2159 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2160 break; 2161 2162 default: 2163 LOG(FATAL) << "Unexpected type conversion from " << input_type 2164 << " to " << result_type; 2165 }; 2166 break; 2167 2168 case Primitive::kPrimDouble: 2169 switch (input_type) { 2170 case Primitive::kPrimBoolean: 2171 // Boolean input is a result of code transformations. 2172 case Primitive::kPrimByte: 2173 case Primitive::kPrimShort: 2174 case Primitive::kPrimInt: 2175 case Primitive::kPrimChar: 2176 // Processing a Dex `int-to-double' instruction. 2177 locations->SetInAt(0, Location::RequiresRegister()); 2178 locations->SetOut(Location::RequiresFpuRegister()); 2179 break; 2180 2181 case Primitive::kPrimLong: 2182 // Processing a Dex `long-to-double' instruction. 2183 locations->SetInAt(0, Location::RequiresRegister()); 2184 locations->SetOut(Location::RequiresFpuRegister()); 2185 locations->AddTemp(Location::RequiresFpuRegister()); 2186 locations->AddTemp(Location::RequiresFpuRegister()); 2187 break; 2188 2189 case Primitive::kPrimFloat: 2190 // Processing a Dex `float-to-double' instruction. 2191 locations->SetInAt(0, Location::RequiresFpuRegister()); 2192 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2193 break; 2194 2195 default: 2196 LOG(FATAL) << "Unexpected type conversion from " << input_type 2197 << " to " << result_type; 2198 }; 2199 break; 2200 2201 default: 2202 LOG(FATAL) << "Unexpected type conversion from " << input_type 2203 << " to " << result_type; 2204 } 2205 } 2206 2207 void InstructionCodeGeneratorARM::VisitTypeConversion(HTypeConversion* conversion) { 2208 LocationSummary* locations = conversion->GetLocations(); 2209 Location out = locations->Out(); 2210 Location in = locations->InAt(0); 2211 Primitive::Type result_type = conversion->GetResultType(); 2212 Primitive::Type input_type = conversion->GetInputType(); 2213 DCHECK_NE(result_type, input_type); 2214 switch (result_type) { 2215 case Primitive::kPrimByte: 2216 switch (input_type) { 2217 case Primitive::kPrimLong: 2218 // Type conversion from long to byte is a result of code transformations. 2219 __ sbfx(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>(), 0, 8); 2220 break; 2221 case Primitive::kPrimBoolean: 2222 // Boolean input is a result of code transformations. 2223 case Primitive::kPrimShort: 2224 case Primitive::kPrimInt: 2225 case Primitive::kPrimChar: 2226 // Processing a Dex `int-to-byte' instruction. 2227 __ sbfx(out.AsRegister<Register>(), in.AsRegister<Register>(), 0, 8); 2228 break; 2229 2230 default: 2231 LOG(FATAL) << "Unexpected type conversion from " << input_type 2232 << " to " << result_type; 2233 } 2234 break; 2235 2236 case Primitive::kPrimShort: 2237 switch (input_type) { 2238 case Primitive::kPrimLong: 2239 // Type conversion from long to short is a result of code transformations. 2240 __ sbfx(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>(), 0, 16); 2241 break; 2242 case Primitive::kPrimBoolean: 2243 // Boolean input is a result of code transformations. 2244 case Primitive::kPrimByte: 2245 case Primitive::kPrimInt: 2246 case Primitive::kPrimChar: 2247 // Processing a Dex `int-to-short' instruction. 2248 __ sbfx(out.AsRegister<Register>(), in.AsRegister<Register>(), 0, 16); 2249 break; 2250 2251 default: 2252 LOG(FATAL) << "Unexpected type conversion from " << input_type 2253 << " to " << result_type; 2254 } 2255 break; 2256 2257 case Primitive::kPrimInt: 2258 switch (input_type) { 2259 case Primitive::kPrimLong: 2260 // Processing a Dex `long-to-int' instruction. 2261 DCHECK(out.IsRegister()); 2262 if (in.IsRegisterPair()) { 2263 __ Mov(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>()); 2264 } else if (in.IsDoubleStackSlot()) { 2265 __ LoadFromOffset(kLoadWord, out.AsRegister<Register>(), SP, in.GetStackIndex()); 2266 } else { 2267 DCHECK(in.IsConstant()); 2268 DCHECK(in.GetConstant()->IsLongConstant()); 2269 int64_t value = in.GetConstant()->AsLongConstant()->GetValue(); 2270 __ LoadImmediate(out.AsRegister<Register>(), static_cast<int32_t>(value)); 2271 } 2272 break; 2273 2274 case Primitive::kPrimFloat: { 2275 // Processing a Dex `float-to-int' instruction. 2276 SRegister temp = locations->GetTemp(0).AsFpuRegisterPairLow<SRegister>(); 2277 __ vmovs(temp, in.AsFpuRegister<SRegister>()); 2278 __ vcvtis(temp, temp); 2279 __ vmovrs(out.AsRegister<Register>(), temp); 2280 break; 2281 } 2282 2283 case Primitive::kPrimDouble: { 2284 // Processing a Dex `double-to-int' instruction. 2285 SRegister temp_s = locations->GetTemp(0).AsFpuRegisterPairLow<SRegister>(); 2286 DRegister temp_d = FromLowSToD(temp_s); 2287 __ vmovd(temp_d, FromLowSToD(in.AsFpuRegisterPairLow<SRegister>())); 2288 __ vcvtid(temp_s, temp_d); 2289 __ vmovrs(out.AsRegister<Register>(), temp_s); 2290 break; 2291 } 2292 2293 default: 2294 LOG(FATAL) << "Unexpected type conversion from " << input_type 2295 << " to " << result_type; 2296 } 2297 break; 2298 2299 case Primitive::kPrimLong: 2300 switch (input_type) { 2301 case Primitive::kPrimBoolean: 2302 // Boolean input is a result of code transformations. 2303 case Primitive::kPrimByte: 2304 case Primitive::kPrimShort: 2305 case Primitive::kPrimInt: 2306 case Primitive::kPrimChar: 2307 // Processing a Dex `int-to-long' instruction. 2308 DCHECK(out.IsRegisterPair()); 2309 DCHECK(in.IsRegister()); 2310 __ Mov(out.AsRegisterPairLow<Register>(), in.AsRegister<Register>()); 2311 // Sign extension. 2312 __ Asr(out.AsRegisterPairHigh<Register>(), 2313 out.AsRegisterPairLow<Register>(), 2314 31); 2315 break; 2316 2317 case Primitive::kPrimFloat: 2318 // Processing a Dex `float-to-long' instruction. 2319 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pF2l), 2320 conversion, 2321 conversion->GetDexPc(), 2322 nullptr); 2323 CheckEntrypointTypes<kQuickF2l, int64_t, float>(); 2324 break; 2325 2326 case Primitive::kPrimDouble: 2327 // Processing a Dex `double-to-long' instruction. 2328 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pD2l), 2329 conversion, 2330 conversion->GetDexPc(), 2331 nullptr); 2332 CheckEntrypointTypes<kQuickD2l, int64_t, double>(); 2333 break; 2334 2335 default: 2336 LOG(FATAL) << "Unexpected type conversion from " << input_type 2337 << " to " << result_type; 2338 } 2339 break; 2340 2341 case Primitive::kPrimChar: 2342 switch (input_type) { 2343 case Primitive::kPrimLong: 2344 // Type conversion from long to char is a result of code transformations. 2345 __ ubfx(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>(), 0, 16); 2346 break; 2347 case Primitive::kPrimBoolean: 2348 // Boolean input is a result of code transformations. 2349 case Primitive::kPrimByte: 2350 case Primitive::kPrimShort: 2351 case Primitive::kPrimInt: 2352 // Processing a Dex `int-to-char' instruction. 2353 __ ubfx(out.AsRegister<Register>(), in.AsRegister<Register>(), 0, 16); 2354 break; 2355 2356 default: 2357 LOG(FATAL) << "Unexpected type conversion from " << input_type 2358 << " to " << result_type; 2359 } 2360 break; 2361 2362 case Primitive::kPrimFloat: 2363 switch (input_type) { 2364 case Primitive::kPrimBoolean: 2365 // Boolean input is a result of code transformations. 2366 case Primitive::kPrimByte: 2367 case Primitive::kPrimShort: 2368 case Primitive::kPrimInt: 2369 case Primitive::kPrimChar: { 2370 // Processing a Dex `int-to-float' instruction. 2371 __ vmovsr(out.AsFpuRegister<SRegister>(), in.AsRegister<Register>()); 2372 __ vcvtsi(out.AsFpuRegister<SRegister>(), out.AsFpuRegister<SRegister>()); 2373 break; 2374 } 2375 2376 case Primitive::kPrimLong: 2377 // Processing a Dex `long-to-float' instruction. 2378 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pL2f), 2379 conversion, 2380 conversion->GetDexPc(), 2381 nullptr); 2382 CheckEntrypointTypes<kQuickL2f, float, int64_t>(); 2383 break; 2384 2385 case Primitive::kPrimDouble: 2386 // Processing a Dex `double-to-float' instruction. 2387 __ vcvtsd(out.AsFpuRegister<SRegister>(), 2388 FromLowSToD(in.AsFpuRegisterPairLow<SRegister>())); 2389 break; 2390 2391 default: 2392 LOG(FATAL) << "Unexpected type conversion from " << input_type 2393 << " to " << result_type; 2394 }; 2395 break; 2396 2397 case Primitive::kPrimDouble: 2398 switch (input_type) { 2399 case Primitive::kPrimBoolean: 2400 // Boolean input is a result of code transformations. 2401 case Primitive::kPrimByte: 2402 case Primitive::kPrimShort: 2403 case Primitive::kPrimInt: 2404 case Primitive::kPrimChar: { 2405 // Processing a Dex `int-to-double' instruction. 2406 __ vmovsr(out.AsFpuRegisterPairLow<SRegister>(), in.AsRegister<Register>()); 2407 __ vcvtdi(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2408 out.AsFpuRegisterPairLow<SRegister>()); 2409 break; 2410 } 2411 2412 case Primitive::kPrimLong: { 2413 // Processing a Dex `long-to-double' instruction. 2414 Register low = in.AsRegisterPairLow<Register>(); 2415 Register high = in.AsRegisterPairHigh<Register>(); 2416 SRegister out_s = out.AsFpuRegisterPairLow<SRegister>(); 2417 DRegister out_d = FromLowSToD(out_s); 2418 SRegister temp_s = locations->GetTemp(0).AsFpuRegisterPairLow<SRegister>(); 2419 DRegister temp_d = FromLowSToD(temp_s); 2420 SRegister constant_s = locations->GetTemp(1).AsFpuRegisterPairLow<SRegister>(); 2421 DRegister constant_d = FromLowSToD(constant_s); 2422 2423 // temp_d = int-to-double(high) 2424 __ vmovsr(temp_s, high); 2425 __ vcvtdi(temp_d, temp_s); 2426 // constant_d = k2Pow32EncodingForDouble 2427 __ LoadDImmediate(constant_d, bit_cast<double, int64_t>(k2Pow32EncodingForDouble)); 2428 // out_d = unsigned-to-double(low) 2429 __ vmovsr(out_s, low); 2430 __ vcvtdu(out_d, out_s); 2431 // out_d += temp_d * constant_d 2432 __ vmlad(out_d, temp_d, constant_d); 2433 break; 2434 } 2435 2436 case Primitive::kPrimFloat: 2437 // Processing a Dex `float-to-double' instruction. 2438 __ vcvtds(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2439 in.AsFpuRegister<SRegister>()); 2440 break; 2441 2442 default: 2443 LOG(FATAL) << "Unexpected type conversion from " << input_type 2444 << " to " << result_type; 2445 }; 2446 break; 2447 2448 default: 2449 LOG(FATAL) << "Unexpected type conversion from " << input_type 2450 << " to " << result_type; 2451 } 2452 } 2453 2454 void LocationsBuilderARM::VisitAdd(HAdd* add) { 2455 LocationSummary* locations = 2456 new (GetGraph()->GetArena()) LocationSummary(add, LocationSummary::kNoCall); 2457 switch (add->GetResultType()) { 2458 case Primitive::kPrimInt: { 2459 locations->SetInAt(0, Location::RequiresRegister()); 2460 locations->SetInAt(1, Location::RegisterOrConstant(add->InputAt(1))); 2461 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2462 break; 2463 } 2464 2465 case Primitive::kPrimLong: { 2466 locations->SetInAt(0, Location::RequiresRegister()); 2467 locations->SetInAt(1, Location::RequiresRegister()); 2468 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2469 break; 2470 } 2471 2472 case Primitive::kPrimFloat: 2473 case Primitive::kPrimDouble: { 2474 locations->SetInAt(0, Location::RequiresFpuRegister()); 2475 locations->SetInAt(1, Location::RequiresFpuRegister()); 2476 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2477 break; 2478 } 2479 2480 default: 2481 LOG(FATAL) << "Unexpected add type " << add->GetResultType(); 2482 } 2483 } 2484 2485 void InstructionCodeGeneratorARM::VisitAdd(HAdd* add) { 2486 LocationSummary* locations = add->GetLocations(); 2487 Location out = locations->Out(); 2488 Location first = locations->InAt(0); 2489 Location second = locations->InAt(1); 2490 switch (add->GetResultType()) { 2491 case Primitive::kPrimInt: 2492 if (second.IsRegister()) { 2493 __ add(out.AsRegister<Register>(), 2494 first.AsRegister<Register>(), 2495 ShifterOperand(second.AsRegister<Register>())); 2496 } else { 2497 __ AddConstant(out.AsRegister<Register>(), 2498 first.AsRegister<Register>(), 2499 second.GetConstant()->AsIntConstant()->GetValue()); 2500 } 2501 break; 2502 2503 case Primitive::kPrimLong: { 2504 DCHECK(second.IsRegisterPair()); 2505 __ adds(out.AsRegisterPairLow<Register>(), 2506 first.AsRegisterPairLow<Register>(), 2507 ShifterOperand(second.AsRegisterPairLow<Register>())); 2508 __ adc(out.AsRegisterPairHigh<Register>(), 2509 first.AsRegisterPairHigh<Register>(), 2510 ShifterOperand(second.AsRegisterPairHigh<Register>())); 2511 break; 2512 } 2513 2514 case Primitive::kPrimFloat: 2515 __ vadds(out.AsFpuRegister<SRegister>(), 2516 first.AsFpuRegister<SRegister>(), 2517 second.AsFpuRegister<SRegister>()); 2518 break; 2519 2520 case Primitive::kPrimDouble: 2521 __ vaddd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2522 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 2523 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 2524 break; 2525 2526 default: 2527 LOG(FATAL) << "Unexpected add type " << add->GetResultType(); 2528 } 2529 } 2530 2531 void LocationsBuilderARM::VisitSub(HSub* sub) { 2532 LocationSummary* locations = 2533 new (GetGraph()->GetArena()) LocationSummary(sub, LocationSummary::kNoCall); 2534 switch (sub->GetResultType()) { 2535 case Primitive::kPrimInt: { 2536 locations->SetInAt(0, Location::RequiresRegister()); 2537 locations->SetInAt(1, Location::RegisterOrConstant(sub->InputAt(1))); 2538 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2539 break; 2540 } 2541 2542 case Primitive::kPrimLong: { 2543 locations->SetInAt(0, Location::RequiresRegister()); 2544 locations->SetInAt(1, Location::RequiresRegister()); 2545 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2546 break; 2547 } 2548 case Primitive::kPrimFloat: 2549 case Primitive::kPrimDouble: { 2550 locations->SetInAt(0, Location::RequiresFpuRegister()); 2551 locations->SetInAt(1, Location::RequiresFpuRegister()); 2552 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2553 break; 2554 } 2555 default: 2556 LOG(FATAL) << "Unexpected sub type " << sub->GetResultType(); 2557 } 2558 } 2559 2560 void InstructionCodeGeneratorARM::VisitSub(HSub* sub) { 2561 LocationSummary* locations = sub->GetLocations(); 2562 Location out = locations->Out(); 2563 Location first = locations->InAt(0); 2564 Location second = locations->InAt(1); 2565 switch (sub->GetResultType()) { 2566 case Primitive::kPrimInt: { 2567 if (second.IsRegister()) { 2568 __ sub(out.AsRegister<Register>(), 2569 first.AsRegister<Register>(), 2570 ShifterOperand(second.AsRegister<Register>())); 2571 } else { 2572 __ AddConstant(out.AsRegister<Register>(), 2573 first.AsRegister<Register>(), 2574 -second.GetConstant()->AsIntConstant()->GetValue()); 2575 } 2576 break; 2577 } 2578 2579 case Primitive::kPrimLong: { 2580 DCHECK(second.IsRegisterPair()); 2581 __ subs(out.AsRegisterPairLow<Register>(), 2582 first.AsRegisterPairLow<Register>(), 2583 ShifterOperand(second.AsRegisterPairLow<Register>())); 2584 __ sbc(out.AsRegisterPairHigh<Register>(), 2585 first.AsRegisterPairHigh<Register>(), 2586 ShifterOperand(second.AsRegisterPairHigh<Register>())); 2587 break; 2588 } 2589 2590 case Primitive::kPrimFloat: { 2591 __ vsubs(out.AsFpuRegister<SRegister>(), 2592 first.AsFpuRegister<SRegister>(), 2593 second.AsFpuRegister<SRegister>()); 2594 break; 2595 } 2596 2597 case Primitive::kPrimDouble: { 2598 __ vsubd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2599 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 2600 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 2601 break; 2602 } 2603 2604 2605 default: 2606 LOG(FATAL) << "Unexpected sub type " << sub->GetResultType(); 2607 } 2608 } 2609 2610 void LocationsBuilderARM::VisitMul(HMul* mul) { 2611 LocationSummary* locations = 2612 new (GetGraph()->GetArena()) LocationSummary(mul, LocationSummary::kNoCall); 2613 switch (mul->GetResultType()) { 2614 case Primitive::kPrimInt: 2615 case Primitive::kPrimLong: { 2616 locations->SetInAt(0, Location::RequiresRegister()); 2617 locations->SetInAt(1, Location::RequiresRegister()); 2618 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2619 break; 2620 } 2621 2622 case Primitive::kPrimFloat: 2623 case Primitive::kPrimDouble: { 2624 locations->SetInAt(0, Location::RequiresFpuRegister()); 2625 locations->SetInAt(1, Location::RequiresFpuRegister()); 2626 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2627 break; 2628 } 2629 2630 default: 2631 LOG(FATAL) << "Unexpected mul type " << mul->GetResultType(); 2632 } 2633 } 2634 2635 void InstructionCodeGeneratorARM::VisitMul(HMul* mul) { 2636 LocationSummary* locations = mul->GetLocations(); 2637 Location out = locations->Out(); 2638 Location first = locations->InAt(0); 2639 Location second = locations->InAt(1); 2640 switch (mul->GetResultType()) { 2641 case Primitive::kPrimInt: { 2642 __ mul(out.AsRegister<Register>(), 2643 first.AsRegister<Register>(), 2644 second.AsRegister<Register>()); 2645 break; 2646 } 2647 case Primitive::kPrimLong: { 2648 Register out_hi = out.AsRegisterPairHigh<Register>(); 2649 Register out_lo = out.AsRegisterPairLow<Register>(); 2650 Register in1_hi = first.AsRegisterPairHigh<Register>(); 2651 Register in1_lo = first.AsRegisterPairLow<Register>(); 2652 Register in2_hi = second.AsRegisterPairHigh<Register>(); 2653 Register in2_lo = second.AsRegisterPairLow<Register>(); 2654 2655 // Extra checks to protect caused by the existence of R1_R2. 2656 // The algorithm is wrong if out.hi is either in1.lo or in2.lo: 2657 // (e.g. in1=r0_r1, in2=r2_r3 and out=r1_r2); 2658 DCHECK_NE(out_hi, in1_lo); 2659 DCHECK_NE(out_hi, in2_lo); 2660 2661 // input: in1 - 64 bits, in2 - 64 bits 2662 // output: out 2663 // formula: out.hi : out.lo = (in1.lo * in2.hi + in1.hi * in2.lo)* 2^32 + in1.lo * in2.lo 2664 // parts: out.hi = in1.lo * in2.hi + in1.hi * in2.lo + (in1.lo * in2.lo)[63:32] 2665 // parts: out.lo = (in1.lo * in2.lo)[31:0] 2666 2667 // IP <- in1.lo * in2.hi 2668 __ mul(IP, in1_lo, in2_hi); 2669 // out.hi <- in1.lo * in2.hi + in1.hi * in2.lo 2670 __ mla(out_hi, in1_hi, in2_lo, IP); 2671 // out.lo <- (in1.lo * in2.lo)[31:0]; 2672 __ umull(out_lo, IP, in1_lo, in2_lo); 2673 // out.hi <- in2.hi * in1.lo + in2.lo * in1.hi + (in1.lo * in2.lo)[63:32] 2674 __ add(out_hi, out_hi, ShifterOperand(IP)); 2675 break; 2676 } 2677 2678 case Primitive::kPrimFloat: { 2679 __ vmuls(out.AsFpuRegister<SRegister>(), 2680 first.AsFpuRegister<SRegister>(), 2681 second.AsFpuRegister<SRegister>()); 2682 break; 2683 } 2684 2685 case Primitive::kPrimDouble: { 2686 __ vmuld(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2687 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 2688 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 2689 break; 2690 } 2691 2692 default: 2693 LOG(FATAL) << "Unexpected mul type " << mul->GetResultType(); 2694 } 2695 } 2696 2697 void InstructionCodeGeneratorARM::DivRemOneOrMinusOne(HBinaryOperation* instruction) { 2698 DCHECK(instruction->IsDiv() || instruction->IsRem()); 2699 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 2700 2701 LocationSummary* locations = instruction->GetLocations(); 2702 Location second = locations->InAt(1); 2703 DCHECK(second.IsConstant()); 2704 2705 Register out = locations->Out().AsRegister<Register>(); 2706 Register dividend = locations->InAt(0).AsRegister<Register>(); 2707 int32_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 2708 DCHECK(imm == 1 || imm == -1); 2709 2710 if (instruction->IsRem()) { 2711 __ LoadImmediate(out, 0); 2712 } else { 2713 if (imm == 1) { 2714 __ Mov(out, dividend); 2715 } else { 2716 __ rsb(out, dividend, ShifterOperand(0)); 2717 } 2718 } 2719 } 2720 2721 void InstructionCodeGeneratorARM::DivRemByPowerOfTwo(HBinaryOperation* instruction) { 2722 DCHECK(instruction->IsDiv() || instruction->IsRem()); 2723 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 2724 2725 LocationSummary* locations = instruction->GetLocations(); 2726 Location second = locations->InAt(1); 2727 DCHECK(second.IsConstant()); 2728 2729 Register out = locations->Out().AsRegister<Register>(); 2730 Register dividend = locations->InAt(0).AsRegister<Register>(); 2731 Register temp = locations->GetTemp(0).AsRegister<Register>(); 2732 int32_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 2733 uint32_t abs_imm = static_cast<uint32_t>(AbsOrMin(imm)); 2734 int ctz_imm = CTZ(abs_imm); 2735 2736 if (ctz_imm == 1) { 2737 __ Lsr(temp, dividend, 32 - ctz_imm); 2738 } else { 2739 __ Asr(temp, dividend, 31); 2740 __ Lsr(temp, temp, 32 - ctz_imm); 2741 } 2742 __ add(out, temp, ShifterOperand(dividend)); 2743 2744 if (instruction->IsDiv()) { 2745 __ Asr(out, out, ctz_imm); 2746 if (imm < 0) { 2747 __ rsb(out, out, ShifterOperand(0)); 2748 } 2749 } else { 2750 __ ubfx(out, out, 0, ctz_imm); 2751 __ sub(out, out, ShifterOperand(temp)); 2752 } 2753 } 2754 2755 void InstructionCodeGeneratorARM::GenerateDivRemWithAnyConstant(HBinaryOperation* instruction) { 2756 DCHECK(instruction->IsDiv() || instruction->IsRem()); 2757 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 2758 2759 LocationSummary* locations = instruction->GetLocations(); 2760 Location second = locations->InAt(1); 2761 DCHECK(second.IsConstant()); 2762 2763 Register out = locations->Out().AsRegister<Register>(); 2764 Register dividend = locations->InAt(0).AsRegister<Register>(); 2765 Register temp1 = locations->GetTemp(0).AsRegister<Register>(); 2766 Register temp2 = locations->GetTemp(1).AsRegister<Register>(); 2767 int64_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 2768 2769 int64_t magic; 2770 int shift; 2771 CalculateMagicAndShiftForDivRem(imm, false /* is_long */, &magic, &shift); 2772 2773 __ LoadImmediate(temp1, magic); 2774 __ smull(temp2, temp1, dividend, temp1); 2775 2776 if (imm > 0 && magic < 0) { 2777 __ add(temp1, temp1, ShifterOperand(dividend)); 2778 } else if (imm < 0 && magic > 0) { 2779 __ sub(temp1, temp1, ShifterOperand(dividend)); 2780 } 2781 2782 if (shift != 0) { 2783 __ Asr(temp1, temp1, shift); 2784 } 2785 2786 if (instruction->IsDiv()) { 2787 __ sub(out, temp1, ShifterOperand(temp1, ASR, 31)); 2788 } else { 2789 __ sub(temp1, temp1, ShifterOperand(temp1, ASR, 31)); 2790 // TODO: Strength reduction for mls. 2791 __ LoadImmediate(temp2, imm); 2792 __ mls(out, temp1, temp2, dividend); 2793 } 2794 } 2795 2796 void InstructionCodeGeneratorARM::GenerateDivRemConstantIntegral(HBinaryOperation* instruction) { 2797 DCHECK(instruction->IsDiv() || instruction->IsRem()); 2798 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 2799 2800 LocationSummary* locations = instruction->GetLocations(); 2801 Location second = locations->InAt(1); 2802 DCHECK(second.IsConstant()); 2803 2804 int32_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 2805 if (imm == 0) { 2806 // Do not generate anything. DivZeroCheck would prevent any code to be executed. 2807 } else if (imm == 1 || imm == -1) { 2808 DivRemOneOrMinusOne(instruction); 2809 } else if (IsPowerOfTwo(AbsOrMin(imm))) { 2810 DivRemByPowerOfTwo(instruction); 2811 } else { 2812 DCHECK(imm <= -2 || imm >= 2); 2813 GenerateDivRemWithAnyConstant(instruction); 2814 } 2815 } 2816 2817 void LocationsBuilderARM::VisitDiv(HDiv* div) { 2818 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 2819 if (div->GetResultType() == Primitive::kPrimLong) { 2820 // pLdiv runtime call. 2821 call_kind = LocationSummary::kCall; 2822 } else if (div->GetResultType() == Primitive::kPrimInt && div->InputAt(1)->IsConstant()) { 2823 // sdiv will be replaced by other instruction sequence. 2824 } else if (div->GetResultType() == Primitive::kPrimInt && 2825 !codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 2826 // pIdivmod runtime call. 2827 call_kind = LocationSummary::kCall; 2828 } 2829 2830 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(div, call_kind); 2831 2832 switch (div->GetResultType()) { 2833 case Primitive::kPrimInt: { 2834 if (div->InputAt(1)->IsConstant()) { 2835 locations->SetInAt(0, Location::RequiresRegister()); 2836 locations->SetInAt(1, Location::ConstantLocation(div->InputAt(1)->AsConstant())); 2837 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2838 int32_t value = div->InputAt(1)->AsIntConstant()->GetValue(); 2839 if (value == 1 || value == 0 || value == -1) { 2840 // No temp register required. 2841 } else { 2842 locations->AddTemp(Location::RequiresRegister()); 2843 if (!IsPowerOfTwo(AbsOrMin(value))) { 2844 locations->AddTemp(Location::RequiresRegister()); 2845 } 2846 } 2847 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 2848 locations->SetInAt(0, Location::RequiresRegister()); 2849 locations->SetInAt(1, Location::RequiresRegister()); 2850 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2851 } else { 2852 InvokeRuntimeCallingConvention calling_convention; 2853 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 2854 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 2855 // Note: divrem will compute both the quotient and the remainder as the pair R0 and R1, but 2856 // we only need the former. 2857 locations->SetOut(Location::RegisterLocation(R0)); 2858 } 2859 break; 2860 } 2861 case Primitive::kPrimLong: { 2862 InvokeRuntimeCallingConvention calling_convention; 2863 locations->SetInAt(0, Location::RegisterPairLocation( 2864 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1))); 2865 locations->SetInAt(1, Location::RegisterPairLocation( 2866 calling_convention.GetRegisterAt(2), calling_convention.GetRegisterAt(3))); 2867 locations->SetOut(Location::RegisterPairLocation(R0, R1)); 2868 break; 2869 } 2870 case Primitive::kPrimFloat: 2871 case Primitive::kPrimDouble: { 2872 locations->SetInAt(0, Location::RequiresFpuRegister()); 2873 locations->SetInAt(1, Location::RequiresFpuRegister()); 2874 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2875 break; 2876 } 2877 2878 default: 2879 LOG(FATAL) << "Unexpected div type " << div->GetResultType(); 2880 } 2881 } 2882 2883 void InstructionCodeGeneratorARM::VisitDiv(HDiv* div) { 2884 LocationSummary* locations = div->GetLocations(); 2885 Location out = locations->Out(); 2886 Location first = locations->InAt(0); 2887 Location second = locations->InAt(1); 2888 2889 switch (div->GetResultType()) { 2890 case Primitive::kPrimInt: { 2891 if (second.IsConstant()) { 2892 GenerateDivRemConstantIntegral(div); 2893 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 2894 __ sdiv(out.AsRegister<Register>(), 2895 first.AsRegister<Register>(), 2896 second.AsRegister<Register>()); 2897 } else { 2898 InvokeRuntimeCallingConvention calling_convention; 2899 DCHECK_EQ(calling_convention.GetRegisterAt(0), first.AsRegister<Register>()); 2900 DCHECK_EQ(calling_convention.GetRegisterAt(1), second.AsRegister<Register>()); 2901 DCHECK_EQ(R0, out.AsRegister<Register>()); 2902 2903 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pIdivmod), div, div->GetDexPc(), nullptr); 2904 CheckEntrypointTypes<kQuickIdivmod, int32_t, int32_t, int32_t>(); 2905 } 2906 break; 2907 } 2908 2909 case Primitive::kPrimLong: { 2910 InvokeRuntimeCallingConvention calling_convention; 2911 DCHECK_EQ(calling_convention.GetRegisterAt(0), first.AsRegisterPairLow<Register>()); 2912 DCHECK_EQ(calling_convention.GetRegisterAt(1), first.AsRegisterPairHigh<Register>()); 2913 DCHECK_EQ(calling_convention.GetRegisterAt(2), second.AsRegisterPairLow<Register>()); 2914 DCHECK_EQ(calling_convention.GetRegisterAt(3), second.AsRegisterPairHigh<Register>()); 2915 DCHECK_EQ(R0, out.AsRegisterPairLow<Register>()); 2916 DCHECK_EQ(R1, out.AsRegisterPairHigh<Register>()); 2917 2918 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pLdiv), div, div->GetDexPc(), nullptr); 2919 CheckEntrypointTypes<kQuickLdiv, int64_t, int64_t, int64_t>(); 2920 break; 2921 } 2922 2923 case Primitive::kPrimFloat: { 2924 __ vdivs(out.AsFpuRegister<SRegister>(), 2925 first.AsFpuRegister<SRegister>(), 2926 second.AsFpuRegister<SRegister>()); 2927 break; 2928 } 2929 2930 case Primitive::kPrimDouble: { 2931 __ vdivd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2932 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 2933 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 2934 break; 2935 } 2936 2937 default: 2938 LOG(FATAL) << "Unexpected div type " << div->GetResultType(); 2939 } 2940 } 2941 2942 void LocationsBuilderARM::VisitRem(HRem* rem) { 2943 Primitive::Type type = rem->GetResultType(); 2944 2945 // Most remainders are implemented in the runtime. 2946 LocationSummary::CallKind call_kind = LocationSummary::kCall; 2947 if (rem->GetResultType() == Primitive::kPrimInt && rem->InputAt(1)->IsConstant()) { 2948 // sdiv will be replaced by other instruction sequence. 2949 call_kind = LocationSummary::kNoCall; 2950 } else if ((rem->GetResultType() == Primitive::kPrimInt) 2951 && codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 2952 // Have hardware divide instruction for int, do it with three instructions. 2953 call_kind = LocationSummary::kNoCall; 2954 } 2955 2956 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(rem, call_kind); 2957 2958 switch (type) { 2959 case Primitive::kPrimInt: { 2960 if (rem->InputAt(1)->IsConstant()) { 2961 locations->SetInAt(0, Location::RequiresRegister()); 2962 locations->SetInAt(1, Location::ConstantLocation(rem->InputAt(1)->AsConstant())); 2963 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2964 int32_t value = rem->InputAt(1)->AsIntConstant()->GetValue(); 2965 if (value == 1 || value == 0 || value == -1) { 2966 // No temp register required. 2967 } else { 2968 locations->AddTemp(Location::RequiresRegister()); 2969 if (!IsPowerOfTwo(AbsOrMin(value))) { 2970 locations->AddTemp(Location::RequiresRegister()); 2971 } 2972 } 2973 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 2974 locations->SetInAt(0, Location::RequiresRegister()); 2975 locations->SetInAt(1, Location::RequiresRegister()); 2976 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2977 locations->AddTemp(Location::RequiresRegister()); 2978 } else { 2979 InvokeRuntimeCallingConvention calling_convention; 2980 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 2981 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 2982 // Note: divrem will compute both the quotient and the remainder as the pair R0 and R1, but 2983 // we only need the latter. 2984 locations->SetOut(Location::RegisterLocation(R1)); 2985 } 2986 break; 2987 } 2988 case Primitive::kPrimLong: { 2989 InvokeRuntimeCallingConvention calling_convention; 2990 locations->SetInAt(0, Location::RegisterPairLocation( 2991 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1))); 2992 locations->SetInAt(1, Location::RegisterPairLocation( 2993 calling_convention.GetRegisterAt(2), calling_convention.GetRegisterAt(3))); 2994 // The runtime helper puts the output in R2,R3. 2995 locations->SetOut(Location::RegisterPairLocation(R2, R3)); 2996 break; 2997 } 2998 case Primitive::kPrimFloat: { 2999 InvokeRuntimeCallingConvention calling_convention; 3000 locations->SetInAt(0, Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(0))); 3001 locations->SetInAt(1, Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(1))); 3002 locations->SetOut(Location::FpuRegisterLocation(S0)); 3003 break; 3004 } 3005 3006 case Primitive::kPrimDouble: { 3007 InvokeRuntimeCallingConvention calling_convention; 3008 locations->SetInAt(0, Location::FpuRegisterPairLocation( 3009 calling_convention.GetFpuRegisterAt(0), calling_convention.GetFpuRegisterAt(1))); 3010 locations->SetInAt(1, Location::FpuRegisterPairLocation( 3011 calling_convention.GetFpuRegisterAt(2), calling_convention.GetFpuRegisterAt(3))); 3012 locations->SetOut(Location::Location::FpuRegisterPairLocation(S0, S1)); 3013 break; 3014 } 3015 3016 default: 3017 LOG(FATAL) << "Unexpected rem type " << type; 3018 } 3019 } 3020 3021 void InstructionCodeGeneratorARM::VisitRem(HRem* rem) { 3022 LocationSummary* locations = rem->GetLocations(); 3023 Location out = locations->Out(); 3024 Location first = locations->InAt(0); 3025 Location second = locations->InAt(1); 3026 3027 Primitive::Type type = rem->GetResultType(); 3028 switch (type) { 3029 case Primitive::kPrimInt: { 3030 if (second.IsConstant()) { 3031 GenerateDivRemConstantIntegral(rem); 3032 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 3033 Register reg1 = first.AsRegister<Register>(); 3034 Register reg2 = second.AsRegister<Register>(); 3035 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3036 3037 // temp = reg1 / reg2 (integer division) 3038 // dest = reg1 - temp * reg2 3039 __ sdiv(temp, reg1, reg2); 3040 __ mls(out.AsRegister<Register>(), temp, reg2, reg1); 3041 } else { 3042 InvokeRuntimeCallingConvention calling_convention; 3043 DCHECK_EQ(calling_convention.GetRegisterAt(0), first.AsRegister<Register>()); 3044 DCHECK_EQ(calling_convention.GetRegisterAt(1), second.AsRegister<Register>()); 3045 DCHECK_EQ(R1, out.AsRegister<Register>()); 3046 3047 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pIdivmod), rem, rem->GetDexPc(), nullptr); 3048 CheckEntrypointTypes<kQuickIdivmod, int32_t, int32_t, int32_t>(); 3049 } 3050 break; 3051 } 3052 3053 case Primitive::kPrimLong: { 3054 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pLmod), rem, rem->GetDexPc(), nullptr); 3055 CheckEntrypointTypes<kQuickLmod, int64_t, int64_t, int64_t>(); 3056 break; 3057 } 3058 3059 case Primitive::kPrimFloat: { 3060 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pFmodf), rem, rem->GetDexPc(), nullptr); 3061 CheckEntrypointTypes<kQuickFmodf, float, float, float>(); 3062 break; 3063 } 3064 3065 case Primitive::kPrimDouble: { 3066 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pFmod), rem, rem->GetDexPc(), nullptr); 3067 CheckEntrypointTypes<kQuickFmod, double, double, double>(); 3068 break; 3069 } 3070 3071 default: 3072 LOG(FATAL) << "Unexpected rem type " << type; 3073 } 3074 } 3075 3076 void LocationsBuilderARM::VisitDivZeroCheck(HDivZeroCheck* instruction) { 3077 LocationSummary::CallKind call_kind = instruction->CanThrowIntoCatchBlock() 3078 ? LocationSummary::kCallOnSlowPath 3079 : LocationSummary::kNoCall; 3080 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 3081 locations->SetInAt(0, Location::RegisterOrConstant(instruction->InputAt(0))); 3082 if (instruction->HasUses()) { 3083 locations->SetOut(Location::SameAsFirstInput()); 3084 } 3085 } 3086 3087 void InstructionCodeGeneratorARM::VisitDivZeroCheck(HDivZeroCheck* instruction) { 3088 SlowPathCode* slow_path = new (GetGraph()->GetArena()) DivZeroCheckSlowPathARM(instruction); 3089 codegen_->AddSlowPath(slow_path); 3090 3091 LocationSummary* locations = instruction->GetLocations(); 3092 Location value = locations->InAt(0); 3093 3094 switch (instruction->GetType()) { 3095 case Primitive::kPrimBoolean: 3096 case Primitive::kPrimByte: 3097 case Primitive::kPrimChar: 3098 case Primitive::kPrimShort: 3099 case Primitive::kPrimInt: { 3100 if (value.IsRegister()) { 3101 __ CompareAndBranchIfZero(value.AsRegister<Register>(), slow_path->GetEntryLabel()); 3102 } else { 3103 DCHECK(value.IsConstant()) << value; 3104 if (value.GetConstant()->AsIntConstant()->GetValue() == 0) { 3105 __ b(slow_path->GetEntryLabel()); 3106 } 3107 } 3108 break; 3109 } 3110 case Primitive::kPrimLong: { 3111 if (value.IsRegisterPair()) { 3112 __ orrs(IP, 3113 value.AsRegisterPairLow<Register>(), 3114 ShifterOperand(value.AsRegisterPairHigh<Register>())); 3115 __ b(slow_path->GetEntryLabel(), EQ); 3116 } else { 3117 DCHECK(value.IsConstant()) << value; 3118 if (value.GetConstant()->AsLongConstant()->GetValue() == 0) { 3119 __ b(slow_path->GetEntryLabel()); 3120 } 3121 } 3122 break; 3123 default: 3124 LOG(FATAL) << "Unexpected type for HDivZeroCheck " << instruction->GetType(); 3125 } 3126 } 3127 } 3128 3129 void InstructionCodeGeneratorARM::HandleIntegerRotate(LocationSummary* locations) { 3130 Register in = locations->InAt(0).AsRegister<Register>(); 3131 Location rhs = locations->InAt(1); 3132 Register out = locations->Out().AsRegister<Register>(); 3133 3134 if (rhs.IsConstant()) { 3135 // Arm32 and Thumb2 assemblers require a rotation on the interval [1,31], 3136 // so map all rotations to a +ve. equivalent in that range. 3137 // (e.g. left *or* right by -2 bits == 30 bits in the same direction.) 3138 uint32_t rot = CodeGenerator::GetInt32ValueOf(rhs.GetConstant()) & 0x1F; 3139 if (rot) { 3140 // Rotate, mapping left rotations to right equivalents if necessary. 3141 // (e.g. left by 2 bits == right by 30.) 3142 __ Ror(out, in, rot); 3143 } else if (out != in) { 3144 __ Mov(out, in); 3145 } 3146 } else { 3147 __ Ror(out, in, rhs.AsRegister<Register>()); 3148 } 3149 } 3150 3151 // Gain some speed by mapping all Long rotates onto equivalent pairs of Integer 3152 // rotates by swapping input regs (effectively rotating by the first 32-bits of 3153 // a larger rotation) or flipping direction (thus treating larger right/left 3154 // rotations as sub-word sized rotations in the other direction) as appropriate. 3155 void InstructionCodeGeneratorARM::HandleLongRotate(LocationSummary* locations) { 3156 Register in_reg_lo = locations->InAt(0).AsRegisterPairLow<Register>(); 3157 Register in_reg_hi = locations->InAt(0).AsRegisterPairHigh<Register>(); 3158 Location rhs = locations->InAt(1); 3159 Register out_reg_lo = locations->Out().AsRegisterPairLow<Register>(); 3160 Register out_reg_hi = locations->Out().AsRegisterPairHigh<Register>(); 3161 3162 if (rhs.IsConstant()) { 3163 uint64_t rot = CodeGenerator::GetInt64ValueOf(rhs.GetConstant()); 3164 // Map all rotations to +ve. equivalents on the interval [0,63]. 3165 rot &= kMaxLongShiftDistance; 3166 // For rotates over a word in size, 'pre-rotate' by 32-bits to keep rotate 3167 // logic below to a simple pair of binary orr. 3168 // (e.g. 34 bits == in_reg swap + 2 bits right.) 3169 if (rot >= kArmBitsPerWord) { 3170 rot -= kArmBitsPerWord; 3171 std::swap(in_reg_hi, in_reg_lo); 3172 } 3173 // Rotate, or mov to out for zero or word size rotations. 3174 if (rot != 0u) { 3175 __ Lsr(out_reg_hi, in_reg_hi, rot); 3176 __ orr(out_reg_hi, out_reg_hi, ShifterOperand(in_reg_lo, arm::LSL, kArmBitsPerWord - rot)); 3177 __ Lsr(out_reg_lo, in_reg_lo, rot); 3178 __ orr(out_reg_lo, out_reg_lo, ShifterOperand(in_reg_hi, arm::LSL, kArmBitsPerWord - rot)); 3179 } else { 3180 __ Mov(out_reg_lo, in_reg_lo); 3181 __ Mov(out_reg_hi, in_reg_hi); 3182 } 3183 } else { 3184 Register shift_right = locations->GetTemp(0).AsRegister<Register>(); 3185 Register shift_left = locations->GetTemp(1).AsRegister<Register>(); 3186 Label end; 3187 Label shift_by_32_plus_shift_right; 3188 3189 __ and_(shift_right, rhs.AsRegister<Register>(), ShifterOperand(0x1F)); 3190 __ Lsrs(shift_left, rhs.AsRegister<Register>(), 6); 3191 __ rsb(shift_left, shift_right, ShifterOperand(kArmBitsPerWord), AL, kCcKeep); 3192 __ b(&shift_by_32_plus_shift_right, CC); 3193 3194 // out_reg_hi = (reg_hi << shift_left) | (reg_lo >> shift_right). 3195 // out_reg_lo = (reg_lo << shift_left) | (reg_hi >> shift_right). 3196 __ Lsl(out_reg_hi, in_reg_hi, shift_left); 3197 __ Lsr(out_reg_lo, in_reg_lo, shift_right); 3198 __ add(out_reg_hi, out_reg_hi, ShifterOperand(out_reg_lo)); 3199 __ Lsl(out_reg_lo, in_reg_lo, shift_left); 3200 __ Lsr(shift_left, in_reg_hi, shift_right); 3201 __ add(out_reg_lo, out_reg_lo, ShifterOperand(shift_left)); 3202 __ b(&end); 3203 3204 __ Bind(&shift_by_32_plus_shift_right); // Shift by 32+shift_right. 3205 // out_reg_hi = (reg_hi >> shift_right) | (reg_lo << shift_left). 3206 // out_reg_lo = (reg_lo >> shift_right) | (reg_hi << shift_left). 3207 __ Lsr(out_reg_hi, in_reg_hi, shift_right); 3208 __ Lsl(out_reg_lo, in_reg_lo, shift_left); 3209 __ add(out_reg_hi, out_reg_hi, ShifterOperand(out_reg_lo)); 3210 __ Lsr(out_reg_lo, in_reg_lo, shift_right); 3211 __ Lsl(shift_right, in_reg_hi, shift_left); 3212 __ add(out_reg_lo, out_reg_lo, ShifterOperand(shift_right)); 3213 3214 __ Bind(&end); 3215 } 3216 } 3217 3218 void LocationsBuilderARM::VisitRor(HRor* ror) { 3219 LocationSummary* locations = 3220 new (GetGraph()->GetArena()) LocationSummary(ror, LocationSummary::kNoCall); 3221 switch (ror->GetResultType()) { 3222 case Primitive::kPrimInt: { 3223 locations->SetInAt(0, Location::RequiresRegister()); 3224 locations->SetInAt(1, Location::RegisterOrConstant(ror->InputAt(1))); 3225 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3226 break; 3227 } 3228 case Primitive::kPrimLong: { 3229 locations->SetInAt(0, Location::RequiresRegister()); 3230 if (ror->InputAt(1)->IsConstant()) { 3231 locations->SetInAt(1, Location::ConstantLocation(ror->InputAt(1)->AsConstant())); 3232 } else { 3233 locations->SetInAt(1, Location::RequiresRegister()); 3234 locations->AddTemp(Location::RequiresRegister()); 3235 locations->AddTemp(Location::RequiresRegister()); 3236 } 3237 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3238 break; 3239 } 3240 default: 3241 LOG(FATAL) << "Unexpected operation type " << ror->GetResultType(); 3242 } 3243 } 3244 3245 void InstructionCodeGeneratorARM::VisitRor(HRor* ror) { 3246 LocationSummary* locations = ror->GetLocations(); 3247 Primitive::Type type = ror->GetResultType(); 3248 switch (type) { 3249 case Primitive::kPrimInt: { 3250 HandleIntegerRotate(locations); 3251 break; 3252 } 3253 case Primitive::kPrimLong: { 3254 HandleLongRotate(locations); 3255 break; 3256 } 3257 default: 3258 LOG(FATAL) << "Unexpected operation type " << type; 3259 UNREACHABLE(); 3260 } 3261 } 3262 3263 void LocationsBuilderARM::HandleShift(HBinaryOperation* op) { 3264 DCHECK(op->IsShl() || op->IsShr() || op->IsUShr()); 3265 3266 LocationSummary* locations = 3267 new (GetGraph()->GetArena()) LocationSummary(op, LocationSummary::kNoCall); 3268 3269 switch (op->GetResultType()) { 3270 case Primitive::kPrimInt: { 3271 locations->SetInAt(0, Location::RequiresRegister()); 3272 if (op->InputAt(1)->IsConstant()) { 3273 locations->SetInAt(1, Location::ConstantLocation(op->InputAt(1)->AsConstant())); 3274 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3275 } else { 3276 locations->SetInAt(1, Location::RequiresRegister()); 3277 // Make the output overlap, as it will be used to hold the masked 3278 // second input. 3279 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3280 } 3281 break; 3282 } 3283 case Primitive::kPrimLong: { 3284 locations->SetInAt(0, Location::RequiresRegister()); 3285 if (op->InputAt(1)->IsConstant()) { 3286 locations->SetInAt(1, Location::ConstantLocation(op->InputAt(1)->AsConstant())); 3287 // For simplicity, use kOutputOverlap even though we only require that low registers 3288 // don't clash with high registers which the register allocator currently guarantees. 3289 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3290 } else { 3291 locations->SetInAt(1, Location::RequiresRegister()); 3292 locations->AddTemp(Location::RequiresRegister()); 3293 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3294 } 3295 break; 3296 } 3297 default: 3298 LOG(FATAL) << "Unexpected operation type " << op->GetResultType(); 3299 } 3300 } 3301 3302 void InstructionCodeGeneratorARM::HandleShift(HBinaryOperation* op) { 3303 DCHECK(op->IsShl() || op->IsShr() || op->IsUShr()); 3304 3305 LocationSummary* locations = op->GetLocations(); 3306 Location out = locations->Out(); 3307 Location first = locations->InAt(0); 3308 Location second = locations->InAt(1); 3309 3310 Primitive::Type type = op->GetResultType(); 3311 switch (type) { 3312 case Primitive::kPrimInt: { 3313 Register out_reg = out.AsRegister<Register>(); 3314 Register first_reg = first.AsRegister<Register>(); 3315 if (second.IsRegister()) { 3316 Register second_reg = second.AsRegister<Register>(); 3317 // ARM doesn't mask the shift count so we need to do it ourselves. 3318 __ and_(out_reg, second_reg, ShifterOperand(kMaxIntShiftDistance)); 3319 if (op->IsShl()) { 3320 __ Lsl(out_reg, first_reg, out_reg); 3321 } else if (op->IsShr()) { 3322 __ Asr(out_reg, first_reg, out_reg); 3323 } else { 3324 __ Lsr(out_reg, first_reg, out_reg); 3325 } 3326 } else { 3327 int32_t cst = second.GetConstant()->AsIntConstant()->GetValue(); 3328 uint32_t shift_value = cst & kMaxIntShiftDistance; 3329 if (shift_value == 0) { // ARM does not support shifting with 0 immediate. 3330 __ Mov(out_reg, first_reg); 3331 } else if (op->IsShl()) { 3332 __ Lsl(out_reg, first_reg, shift_value); 3333 } else if (op->IsShr()) { 3334 __ Asr(out_reg, first_reg, shift_value); 3335 } else { 3336 __ Lsr(out_reg, first_reg, shift_value); 3337 } 3338 } 3339 break; 3340 } 3341 case Primitive::kPrimLong: { 3342 Register o_h = out.AsRegisterPairHigh<Register>(); 3343 Register o_l = out.AsRegisterPairLow<Register>(); 3344 3345 Register high = first.AsRegisterPairHigh<Register>(); 3346 Register low = first.AsRegisterPairLow<Register>(); 3347 3348 if (second.IsRegister()) { 3349 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3350 3351 Register second_reg = second.AsRegister<Register>(); 3352 3353 if (op->IsShl()) { 3354 __ and_(o_l, second_reg, ShifterOperand(kMaxLongShiftDistance)); 3355 // Shift the high part 3356 __ Lsl(o_h, high, o_l); 3357 // Shift the low part and `or` what overflew on the high part 3358 __ rsb(temp, o_l, ShifterOperand(kArmBitsPerWord)); 3359 __ Lsr(temp, low, temp); 3360 __ orr(o_h, o_h, ShifterOperand(temp)); 3361 // If the shift is > 32 bits, override the high part 3362 __ subs(temp, o_l, ShifterOperand(kArmBitsPerWord)); 3363 __ it(PL); 3364 __ Lsl(o_h, low, temp, PL); 3365 // Shift the low part 3366 __ Lsl(o_l, low, o_l); 3367 } else if (op->IsShr()) { 3368 __ and_(o_h, second_reg, ShifterOperand(kMaxLongShiftDistance)); 3369 // Shift the low part 3370 __ Lsr(o_l, low, o_h); 3371 // Shift the high part and `or` what underflew on the low part 3372 __ rsb(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3373 __ Lsl(temp, high, temp); 3374 __ orr(o_l, o_l, ShifterOperand(temp)); 3375 // If the shift is > 32 bits, override the low part 3376 __ subs(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3377 __ it(PL); 3378 __ Asr(o_l, high, temp, PL); 3379 // Shift the high part 3380 __ Asr(o_h, high, o_h); 3381 } else { 3382 __ and_(o_h, second_reg, ShifterOperand(kMaxLongShiftDistance)); 3383 // same as Shr except we use `Lsr`s and not `Asr`s 3384 __ Lsr(o_l, low, o_h); 3385 __ rsb(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3386 __ Lsl(temp, high, temp); 3387 __ orr(o_l, o_l, ShifterOperand(temp)); 3388 __ subs(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3389 __ it(PL); 3390 __ Lsr(o_l, high, temp, PL); 3391 __ Lsr(o_h, high, o_h); 3392 } 3393 } else { 3394 // Register allocator doesn't create partial overlap. 3395 DCHECK_NE(o_l, high); 3396 DCHECK_NE(o_h, low); 3397 int32_t cst = second.GetConstant()->AsIntConstant()->GetValue(); 3398 uint32_t shift_value = cst & kMaxLongShiftDistance; 3399 if (shift_value > 32) { 3400 if (op->IsShl()) { 3401 __ Lsl(o_h, low, shift_value - 32); 3402 __ LoadImmediate(o_l, 0); 3403 } else if (op->IsShr()) { 3404 __ Asr(o_l, high, shift_value - 32); 3405 __ Asr(o_h, high, 31); 3406 } else { 3407 __ Lsr(o_l, high, shift_value - 32); 3408 __ LoadImmediate(o_h, 0); 3409 } 3410 } else if (shift_value == 32) { 3411 if (op->IsShl()) { 3412 __ mov(o_h, ShifterOperand(low)); 3413 __ LoadImmediate(o_l, 0); 3414 } else if (op->IsShr()) { 3415 __ mov(o_l, ShifterOperand(high)); 3416 __ Asr(o_h, high, 31); 3417 } else { 3418 __ mov(o_l, ShifterOperand(high)); 3419 __ LoadImmediate(o_h, 0); 3420 } 3421 } else if (shift_value == 1) { 3422 if (op->IsShl()) { 3423 __ Lsls(o_l, low, 1); 3424 __ adc(o_h, high, ShifterOperand(high)); 3425 } else if (op->IsShr()) { 3426 __ Asrs(o_h, high, 1); 3427 __ Rrx(o_l, low); 3428 } else { 3429 __ Lsrs(o_h, high, 1); 3430 __ Rrx(o_l, low); 3431 } 3432 } else { 3433 DCHECK(2 <= shift_value && shift_value < 32) << shift_value; 3434 if (op->IsShl()) { 3435 __ Lsl(o_h, high, shift_value); 3436 __ orr(o_h, o_h, ShifterOperand(low, LSR, 32 - shift_value)); 3437 __ Lsl(o_l, low, shift_value); 3438 } else if (op->IsShr()) { 3439 __ Lsr(o_l, low, shift_value); 3440 __ orr(o_l, o_l, ShifterOperand(high, LSL, 32 - shift_value)); 3441 __ Asr(o_h, high, shift_value); 3442 } else { 3443 __ Lsr(o_l, low, shift_value); 3444 __ orr(o_l, o_l, ShifterOperand(high, LSL, 32 - shift_value)); 3445 __ Lsr(o_h, high, shift_value); 3446 } 3447 } 3448 } 3449 break; 3450 } 3451 default: 3452 LOG(FATAL) << "Unexpected operation type " << type; 3453 UNREACHABLE(); 3454 } 3455 } 3456 3457 void LocationsBuilderARM::VisitShl(HShl* shl) { 3458 HandleShift(shl); 3459 } 3460 3461 void InstructionCodeGeneratorARM::VisitShl(HShl* shl) { 3462 HandleShift(shl); 3463 } 3464 3465 void LocationsBuilderARM::VisitShr(HShr* shr) { 3466 HandleShift(shr); 3467 } 3468 3469 void InstructionCodeGeneratorARM::VisitShr(HShr* shr) { 3470 HandleShift(shr); 3471 } 3472 3473 void LocationsBuilderARM::VisitUShr(HUShr* ushr) { 3474 HandleShift(ushr); 3475 } 3476 3477 void InstructionCodeGeneratorARM::VisitUShr(HUShr* ushr) { 3478 HandleShift(ushr); 3479 } 3480 3481 void LocationsBuilderARM::VisitNewInstance(HNewInstance* instruction) { 3482 LocationSummary* locations = 3483 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall); 3484 if (instruction->IsStringAlloc()) { 3485 locations->AddTemp(Location::RegisterLocation(kMethodRegisterArgument)); 3486 } else { 3487 InvokeRuntimeCallingConvention calling_convention; 3488 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 3489 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 3490 } 3491 locations->SetOut(Location::RegisterLocation(R0)); 3492 } 3493 3494 void InstructionCodeGeneratorARM::VisitNewInstance(HNewInstance* instruction) { 3495 // Note: if heap poisoning is enabled, the entry point takes cares 3496 // of poisoning the reference. 3497 if (instruction->IsStringAlloc()) { 3498 // String is allocated through StringFactory. Call NewEmptyString entry point. 3499 Register temp = instruction->GetLocations()->GetTemp(0).AsRegister<Register>(); 3500 MemberOffset code_offset = ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmWordSize); 3501 __ LoadFromOffset(kLoadWord, temp, TR, QUICK_ENTRY_POINT(pNewEmptyString)); 3502 __ LoadFromOffset(kLoadWord, LR, temp, code_offset.Int32Value()); 3503 __ blx(LR); 3504 codegen_->RecordPcInfo(instruction, instruction->GetDexPc()); 3505 } else { 3506 codegen_->InvokeRuntime(instruction->GetEntrypoint(), 3507 instruction, 3508 instruction->GetDexPc(), 3509 nullptr); 3510 CheckEntrypointTypes<kQuickAllocObjectWithAccessCheck, void*, uint32_t, ArtMethod*>(); 3511 } 3512 } 3513 3514 void LocationsBuilderARM::VisitNewArray(HNewArray* instruction) { 3515 LocationSummary* locations = 3516 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall); 3517 InvokeRuntimeCallingConvention calling_convention; 3518 locations->AddTemp(Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 3519 locations->SetOut(Location::RegisterLocation(R0)); 3520 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 3521 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(2))); 3522 } 3523 3524 void InstructionCodeGeneratorARM::VisitNewArray(HNewArray* instruction) { 3525 InvokeRuntimeCallingConvention calling_convention; 3526 __ LoadImmediate(calling_convention.GetRegisterAt(0), instruction->GetTypeIndex()); 3527 // Note: if heap poisoning is enabled, the entry point takes cares 3528 // of poisoning the reference. 3529 codegen_->InvokeRuntime(instruction->GetEntrypoint(), 3530 instruction, 3531 instruction->GetDexPc(), 3532 nullptr); 3533 CheckEntrypointTypes<kQuickAllocArrayWithAccessCheck, void*, uint32_t, int32_t, ArtMethod*>(); 3534 } 3535 3536 void LocationsBuilderARM::VisitParameterValue(HParameterValue* instruction) { 3537 LocationSummary* locations = 3538 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 3539 Location location = parameter_visitor_.GetNextLocation(instruction->GetType()); 3540 if (location.IsStackSlot()) { 3541 location = Location::StackSlot(location.GetStackIndex() + codegen_->GetFrameSize()); 3542 } else if (location.IsDoubleStackSlot()) { 3543 location = Location::DoubleStackSlot(location.GetStackIndex() + codegen_->GetFrameSize()); 3544 } 3545 locations->SetOut(location); 3546 } 3547 3548 void InstructionCodeGeneratorARM::VisitParameterValue( 3549 HParameterValue* instruction ATTRIBUTE_UNUSED) { 3550 // Nothing to do, the parameter is already at its location. 3551 } 3552 3553 void LocationsBuilderARM::VisitCurrentMethod(HCurrentMethod* instruction) { 3554 LocationSummary* locations = 3555 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 3556 locations->SetOut(Location::RegisterLocation(kMethodRegisterArgument)); 3557 } 3558 3559 void InstructionCodeGeneratorARM::VisitCurrentMethod(HCurrentMethod* instruction ATTRIBUTE_UNUSED) { 3560 // Nothing to do, the method is already at its location. 3561 } 3562 3563 void LocationsBuilderARM::VisitNot(HNot* not_) { 3564 LocationSummary* locations = 3565 new (GetGraph()->GetArena()) LocationSummary(not_, LocationSummary::kNoCall); 3566 locations->SetInAt(0, Location::RequiresRegister()); 3567 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3568 } 3569 3570 void InstructionCodeGeneratorARM::VisitNot(HNot* not_) { 3571 LocationSummary* locations = not_->GetLocations(); 3572 Location out = locations->Out(); 3573 Location in = locations->InAt(0); 3574 switch (not_->GetResultType()) { 3575 case Primitive::kPrimInt: 3576 __ mvn(out.AsRegister<Register>(), ShifterOperand(in.AsRegister<Register>())); 3577 break; 3578 3579 case Primitive::kPrimLong: 3580 __ mvn(out.AsRegisterPairLow<Register>(), 3581 ShifterOperand(in.AsRegisterPairLow<Register>())); 3582 __ mvn(out.AsRegisterPairHigh<Register>(), 3583 ShifterOperand(in.AsRegisterPairHigh<Register>())); 3584 break; 3585 3586 default: 3587 LOG(FATAL) << "Unimplemented type for not operation " << not_->GetResultType(); 3588 } 3589 } 3590 3591 void LocationsBuilderARM::VisitBooleanNot(HBooleanNot* bool_not) { 3592 LocationSummary* locations = 3593 new (GetGraph()->GetArena()) LocationSummary(bool_not, LocationSummary::kNoCall); 3594 locations->SetInAt(0, Location::RequiresRegister()); 3595 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3596 } 3597 3598 void InstructionCodeGeneratorARM::VisitBooleanNot(HBooleanNot* bool_not) { 3599 LocationSummary* locations = bool_not->GetLocations(); 3600 Location out = locations->Out(); 3601 Location in = locations->InAt(0); 3602 __ eor(out.AsRegister<Register>(), in.AsRegister<Register>(), ShifterOperand(1)); 3603 } 3604 3605 void LocationsBuilderARM::VisitCompare(HCompare* compare) { 3606 LocationSummary* locations = 3607 new (GetGraph()->GetArena()) LocationSummary(compare, LocationSummary::kNoCall); 3608 switch (compare->InputAt(0)->GetType()) { 3609 case Primitive::kPrimBoolean: 3610 case Primitive::kPrimByte: 3611 case Primitive::kPrimShort: 3612 case Primitive::kPrimChar: 3613 case Primitive::kPrimInt: 3614 case Primitive::kPrimLong: { 3615 locations->SetInAt(0, Location::RequiresRegister()); 3616 locations->SetInAt(1, Location::RequiresRegister()); 3617 // Output overlaps because it is written before doing the low comparison. 3618 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3619 break; 3620 } 3621 case Primitive::kPrimFloat: 3622 case Primitive::kPrimDouble: { 3623 locations->SetInAt(0, Location::RequiresFpuRegister()); 3624 locations->SetInAt(1, Location::RequiresFpuRegister()); 3625 locations->SetOut(Location::RequiresRegister()); 3626 break; 3627 } 3628 default: 3629 LOG(FATAL) << "Unexpected type for compare operation " << compare->InputAt(0)->GetType(); 3630 } 3631 } 3632 3633 void InstructionCodeGeneratorARM::VisitCompare(HCompare* compare) { 3634 LocationSummary* locations = compare->GetLocations(); 3635 Register out = locations->Out().AsRegister<Register>(); 3636 Location left = locations->InAt(0); 3637 Location right = locations->InAt(1); 3638 3639 Label less, greater, done; 3640 Primitive::Type type = compare->InputAt(0)->GetType(); 3641 Condition less_cond; 3642 switch (type) { 3643 case Primitive::kPrimBoolean: 3644 case Primitive::kPrimByte: 3645 case Primitive::kPrimShort: 3646 case Primitive::kPrimChar: 3647 case Primitive::kPrimInt: { 3648 __ LoadImmediate(out, 0); 3649 __ cmp(left.AsRegister<Register>(), 3650 ShifterOperand(right.AsRegister<Register>())); // Signed compare. 3651 less_cond = LT; 3652 break; 3653 } 3654 case Primitive::kPrimLong: { 3655 __ cmp(left.AsRegisterPairHigh<Register>(), 3656 ShifterOperand(right.AsRegisterPairHigh<Register>())); // Signed compare. 3657 __ b(&less, LT); 3658 __ b(&greater, GT); 3659 // Do LoadImmediate before the last `cmp`, as LoadImmediate might affect the status flags. 3660 __ LoadImmediate(out, 0); 3661 __ cmp(left.AsRegisterPairLow<Register>(), 3662 ShifterOperand(right.AsRegisterPairLow<Register>())); // Unsigned compare. 3663 less_cond = LO; 3664 break; 3665 } 3666 case Primitive::kPrimFloat: 3667 case Primitive::kPrimDouble: { 3668 __ LoadImmediate(out, 0); 3669 if (type == Primitive::kPrimFloat) { 3670 __ vcmps(left.AsFpuRegister<SRegister>(), right.AsFpuRegister<SRegister>()); 3671 } else { 3672 __ vcmpd(FromLowSToD(left.AsFpuRegisterPairLow<SRegister>()), 3673 FromLowSToD(right.AsFpuRegisterPairLow<SRegister>())); 3674 } 3675 __ vmstat(); // transfer FP status register to ARM APSR. 3676 less_cond = ARMFPCondition(kCondLT, compare->IsGtBias()); 3677 break; 3678 } 3679 default: 3680 LOG(FATAL) << "Unexpected compare type " << type; 3681 UNREACHABLE(); 3682 } 3683 3684 __ b(&done, EQ); 3685 __ b(&less, less_cond); 3686 3687 __ Bind(&greater); 3688 __ LoadImmediate(out, 1); 3689 __ b(&done); 3690 3691 __ Bind(&less); 3692 __ LoadImmediate(out, -1); 3693 3694 __ Bind(&done); 3695 } 3696 3697 void LocationsBuilderARM::VisitPhi(HPhi* instruction) { 3698 LocationSummary* locations = 3699 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 3700 for (size_t i = 0, e = instruction->InputCount(); i < e; ++i) { 3701 locations->SetInAt(i, Location::Any()); 3702 } 3703 locations->SetOut(Location::Any()); 3704 } 3705 3706 void InstructionCodeGeneratorARM::VisitPhi(HPhi* instruction ATTRIBUTE_UNUSED) { 3707 LOG(FATAL) << "Unreachable"; 3708 } 3709 3710 void CodeGeneratorARM::GenerateMemoryBarrier(MemBarrierKind kind) { 3711 // TODO (ported from quick): revisit ARM barrier kinds. 3712 DmbOptions flavor = DmbOptions::ISH; // Quiet C++ warnings. 3713 switch (kind) { 3714 case MemBarrierKind::kAnyStore: 3715 case MemBarrierKind::kLoadAny: 3716 case MemBarrierKind::kAnyAny: { 3717 flavor = DmbOptions::ISH; 3718 break; 3719 } 3720 case MemBarrierKind::kStoreStore: { 3721 flavor = DmbOptions::ISHST; 3722 break; 3723 } 3724 default: 3725 LOG(FATAL) << "Unexpected memory barrier " << kind; 3726 } 3727 __ dmb(flavor); 3728 } 3729 3730 void InstructionCodeGeneratorARM::GenerateWideAtomicLoad(Register addr, 3731 uint32_t offset, 3732 Register out_lo, 3733 Register out_hi) { 3734 if (offset != 0) { 3735 // Ensure `out_lo` is different from `addr`, so that loading 3736 // `offset` into `out_lo` does not clutter `addr`. 3737 DCHECK_NE(out_lo, addr); 3738 __ LoadImmediate(out_lo, offset); 3739 __ add(IP, addr, ShifterOperand(out_lo)); 3740 addr = IP; 3741 } 3742 __ ldrexd(out_lo, out_hi, addr); 3743 } 3744 3745 void InstructionCodeGeneratorARM::GenerateWideAtomicStore(Register addr, 3746 uint32_t offset, 3747 Register value_lo, 3748 Register value_hi, 3749 Register temp1, 3750 Register temp2, 3751 HInstruction* instruction) { 3752 Label fail; 3753 if (offset != 0) { 3754 __ LoadImmediate(temp1, offset); 3755 __ add(IP, addr, ShifterOperand(temp1)); 3756 addr = IP; 3757 } 3758 __ Bind(&fail); 3759 // We need a load followed by store. (The address used in a STREX instruction must 3760 // be the same as the address in the most recently executed LDREX instruction.) 3761 __ ldrexd(temp1, temp2, addr); 3762 codegen_->MaybeRecordImplicitNullCheck(instruction); 3763 __ strexd(temp1, value_lo, value_hi, addr); 3764 __ CompareAndBranchIfNonZero(temp1, &fail); 3765 } 3766 3767 void LocationsBuilderARM::HandleFieldSet(HInstruction* instruction, const FieldInfo& field_info) { 3768 DCHECK(instruction->IsInstanceFieldSet() || instruction->IsStaticFieldSet()); 3769 3770 LocationSummary* locations = 3771 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 3772 locations->SetInAt(0, Location::RequiresRegister()); 3773 3774 Primitive::Type field_type = field_info.GetFieldType(); 3775 if (Primitive::IsFloatingPointType(field_type)) { 3776 locations->SetInAt(1, Location::RequiresFpuRegister()); 3777 } else { 3778 locations->SetInAt(1, Location::RequiresRegister()); 3779 } 3780 3781 bool is_wide = field_type == Primitive::kPrimLong || field_type == Primitive::kPrimDouble; 3782 bool generate_volatile = field_info.IsVolatile() 3783 && is_wide 3784 && !codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 3785 bool needs_write_barrier = 3786 CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1)); 3787 // Temporary registers for the write barrier. 3788 // TODO: consider renaming StoreNeedsWriteBarrier to StoreNeedsGCMark. 3789 if (needs_write_barrier) { 3790 locations->AddTemp(Location::RequiresRegister()); // Possibly used for reference poisoning too. 3791 locations->AddTemp(Location::RequiresRegister()); 3792 } else if (generate_volatile) { 3793 // ARM encoding have some additional constraints for ldrexd/strexd: 3794 // - registers need to be consecutive 3795 // - the first register should be even but not R14. 3796 // We don't test for ARM yet, and the assertion makes sure that we 3797 // revisit this if we ever enable ARM encoding. 3798 DCHECK_EQ(InstructionSet::kThumb2, codegen_->GetInstructionSet()); 3799 3800 locations->AddTemp(Location::RequiresRegister()); 3801 locations->AddTemp(Location::RequiresRegister()); 3802 if (field_type == Primitive::kPrimDouble) { 3803 // For doubles we need two more registers to copy the value. 3804 locations->AddTemp(Location::RegisterLocation(R2)); 3805 locations->AddTemp(Location::RegisterLocation(R3)); 3806 } 3807 } 3808 } 3809 3810 void InstructionCodeGeneratorARM::HandleFieldSet(HInstruction* instruction, 3811 const FieldInfo& field_info, 3812 bool value_can_be_null) { 3813 DCHECK(instruction->IsInstanceFieldSet() || instruction->IsStaticFieldSet()); 3814 3815 LocationSummary* locations = instruction->GetLocations(); 3816 Register base = locations->InAt(0).AsRegister<Register>(); 3817 Location value = locations->InAt(1); 3818 3819 bool is_volatile = field_info.IsVolatile(); 3820 bool atomic_ldrd_strd = codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 3821 Primitive::Type field_type = field_info.GetFieldType(); 3822 uint32_t offset = field_info.GetFieldOffset().Uint32Value(); 3823 bool needs_write_barrier = 3824 CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1)); 3825 3826 if (is_volatile) { 3827 codegen_->GenerateMemoryBarrier(MemBarrierKind::kAnyStore); 3828 } 3829 3830 switch (field_type) { 3831 case Primitive::kPrimBoolean: 3832 case Primitive::kPrimByte: { 3833 __ StoreToOffset(kStoreByte, value.AsRegister<Register>(), base, offset); 3834 break; 3835 } 3836 3837 case Primitive::kPrimShort: 3838 case Primitive::kPrimChar: { 3839 __ StoreToOffset(kStoreHalfword, value.AsRegister<Register>(), base, offset); 3840 break; 3841 } 3842 3843 case Primitive::kPrimInt: 3844 case Primitive::kPrimNot: { 3845 if (kPoisonHeapReferences && needs_write_barrier) { 3846 // Note that in the case where `value` is a null reference, 3847 // we do not enter this block, as a null reference does not 3848 // need poisoning. 3849 DCHECK_EQ(field_type, Primitive::kPrimNot); 3850 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3851 __ Mov(temp, value.AsRegister<Register>()); 3852 __ PoisonHeapReference(temp); 3853 __ StoreToOffset(kStoreWord, temp, base, offset); 3854 } else { 3855 __ StoreToOffset(kStoreWord, value.AsRegister<Register>(), base, offset); 3856 } 3857 break; 3858 } 3859 3860 case Primitive::kPrimLong: { 3861 if (is_volatile && !atomic_ldrd_strd) { 3862 GenerateWideAtomicStore(base, offset, 3863 value.AsRegisterPairLow<Register>(), 3864 value.AsRegisterPairHigh<Register>(), 3865 locations->GetTemp(0).AsRegister<Register>(), 3866 locations->GetTemp(1).AsRegister<Register>(), 3867 instruction); 3868 } else { 3869 __ StoreToOffset(kStoreWordPair, value.AsRegisterPairLow<Register>(), base, offset); 3870 codegen_->MaybeRecordImplicitNullCheck(instruction); 3871 } 3872 break; 3873 } 3874 3875 case Primitive::kPrimFloat: { 3876 __ StoreSToOffset(value.AsFpuRegister<SRegister>(), base, offset); 3877 break; 3878 } 3879 3880 case Primitive::kPrimDouble: { 3881 DRegister value_reg = FromLowSToD(value.AsFpuRegisterPairLow<SRegister>()); 3882 if (is_volatile && !atomic_ldrd_strd) { 3883 Register value_reg_lo = locations->GetTemp(0).AsRegister<Register>(); 3884 Register value_reg_hi = locations->GetTemp(1).AsRegister<Register>(); 3885 3886 __ vmovrrd(value_reg_lo, value_reg_hi, value_reg); 3887 3888 GenerateWideAtomicStore(base, offset, 3889 value_reg_lo, 3890 value_reg_hi, 3891 locations->GetTemp(2).AsRegister<Register>(), 3892 locations->GetTemp(3).AsRegister<Register>(), 3893 instruction); 3894 } else { 3895 __ StoreDToOffset(value_reg, base, offset); 3896 codegen_->MaybeRecordImplicitNullCheck(instruction); 3897 } 3898 break; 3899 } 3900 3901 case Primitive::kPrimVoid: 3902 LOG(FATAL) << "Unreachable type " << field_type; 3903 UNREACHABLE(); 3904 } 3905 3906 // Longs and doubles are handled in the switch. 3907 if (field_type != Primitive::kPrimLong && field_type != Primitive::kPrimDouble) { 3908 codegen_->MaybeRecordImplicitNullCheck(instruction); 3909 } 3910 3911 if (CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1))) { 3912 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3913 Register card = locations->GetTemp(1).AsRegister<Register>(); 3914 codegen_->MarkGCCard( 3915 temp, card, base, value.AsRegister<Register>(), value_can_be_null); 3916 } 3917 3918 if (is_volatile) { 3919 codegen_->GenerateMemoryBarrier(MemBarrierKind::kAnyAny); 3920 } 3921 } 3922 3923 void LocationsBuilderARM::HandleFieldGet(HInstruction* instruction, const FieldInfo& field_info) { 3924 DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet()); 3925 3926 bool object_field_get_with_read_barrier = 3927 kEmitCompilerReadBarrier && (field_info.GetFieldType() == Primitive::kPrimNot); 3928 LocationSummary* locations = 3929 new (GetGraph()->GetArena()) LocationSummary(instruction, 3930 object_field_get_with_read_barrier ? 3931 LocationSummary::kCallOnSlowPath : 3932 LocationSummary::kNoCall); 3933 locations->SetInAt(0, Location::RequiresRegister()); 3934 3935 bool volatile_for_double = field_info.IsVolatile() 3936 && (field_info.GetFieldType() == Primitive::kPrimDouble) 3937 && !codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 3938 // The output overlaps in case of volatile long: we don't want the 3939 // code generated by GenerateWideAtomicLoad to overwrite the 3940 // object's location. Likewise, in the case of an object field get 3941 // with read barriers enabled, we do not want the load to overwrite 3942 // the object's location, as we need it to emit the read barrier. 3943 bool overlap = (field_info.IsVolatile() && (field_info.GetFieldType() == Primitive::kPrimLong)) || 3944 object_field_get_with_read_barrier; 3945 3946 if (Primitive::IsFloatingPointType(instruction->GetType())) { 3947 locations->SetOut(Location::RequiresFpuRegister()); 3948 } else { 3949 locations->SetOut(Location::RequiresRegister(), 3950 (overlap ? Location::kOutputOverlap : Location::kNoOutputOverlap)); 3951 } 3952 if (volatile_for_double) { 3953 // ARM encoding have some additional constraints for ldrexd/strexd: 3954 // - registers need to be consecutive 3955 // - the first register should be even but not R14. 3956 // We don't test for ARM yet, and the assertion makes sure that we 3957 // revisit this if we ever enable ARM encoding. 3958 DCHECK_EQ(InstructionSet::kThumb2, codegen_->GetInstructionSet()); 3959 locations->AddTemp(Location::RequiresRegister()); 3960 locations->AddTemp(Location::RequiresRegister()); 3961 } else if (object_field_get_with_read_barrier && kUseBakerReadBarrier) { 3962 // We need a temporary register for the read barrier marking slow 3963 // path in CodeGeneratorARM::GenerateFieldLoadWithBakerReadBarrier. 3964 locations->AddTemp(Location::RequiresRegister()); 3965 } 3966 } 3967 3968 Location LocationsBuilderARM::ArmEncodableConstantOrRegister(HInstruction* constant, 3969 Opcode opcode) { 3970 DCHECK(!Primitive::IsFloatingPointType(constant->GetType())); 3971 if (constant->IsConstant() && 3972 CanEncodeConstantAsImmediate(constant->AsConstant(), opcode)) { 3973 return Location::ConstantLocation(constant->AsConstant()); 3974 } 3975 return Location::RequiresRegister(); 3976 } 3977 3978 bool LocationsBuilderARM::CanEncodeConstantAsImmediate(HConstant* input_cst, 3979 Opcode opcode) { 3980 uint64_t value = static_cast<uint64_t>(Int64FromConstant(input_cst)); 3981 if (Primitive::Is64BitType(input_cst->GetType())) { 3982 return CanEncodeConstantAsImmediate(Low32Bits(value), opcode) && 3983 CanEncodeConstantAsImmediate(High32Bits(value), opcode); 3984 } else { 3985 return CanEncodeConstantAsImmediate(Low32Bits(value), opcode); 3986 } 3987 } 3988 3989 bool LocationsBuilderARM::CanEncodeConstantAsImmediate(uint32_t value, Opcode opcode) { 3990 ShifterOperand so; 3991 ArmAssembler* assembler = codegen_->GetAssembler(); 3992 if (assembler->ShifterOperandCanHold(kNoRegister, kNoRegister, opcode, value, &so)) { 3993 return true; 3994 } 3995 Opcode neg_opcode = kNoOperand; 3996 switch (opcode) { 3997 case AND: 3998 neg_opcode = BIC; 3999 break; 4000 case ORR: 4001 neg_opcode = ORN; 4002 break; 4003 default: 4004 return false; 4005 } 4006 return assembler->ShifterOperandCanHold(kNoRegister, kNoRegister, neg_opcode, ~value, &so); 4007 } 4008 4009 void InstructionCodeGeneratorARM::HandleFieldGet(HInstruction* instruction, 4010 const FieldInfo& field_info) { 4011 DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet()); 4012 4013 LocationSummary* locations = instruction->GetLocations(); 4014 Location base_loc = locations->InAt(0); 4015 Register base = base_loc.AsRegister<Register>(); 4016 Location out = locations->Out(); 4017 bool is_volatile = field_info.IsVolatile(); 4018 bool atomic_ldrd_strd = codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 4019 Primitive::Type field_type = field_info.GetFieldType(); 4020 uint32_t offset = field_info.GetFieldOffset().Uint32Value(); 4021 4022 switch (field_type) { 4023 case Primitive::kPrimBoolean: 4024 __ LoadFromOffset(kLoadUnsignedByte, out.AsRegister<Register>(), base, offset); 4025 break; 4026 4027 case Primitive::kPrimByte: 4028 __ LoadFromOffset(kLoadSignedByte, out.AsRegister<Register>(), base, offset); 4029 break; 4030 4031 case Primitive::kPrimShort: 4032 __ LoadFromOffset(kLoadSignedHalfword, out.AsRegister<Register>(), base, offset); 4033 break; 4034 4035 case Primitive::kPrimChar: 4036 __ LoadFromOffset(kLoadUnsignedHalfword, out.AsRegister<Register>(), base, offset); 4037 break; 4038 4039 case Primitive::kPrimInt: 4040 __ LoadFromOffset(kLoadWord, out.AsRegister<Register>(), base, offset); 4041 break; 4042 4043 case Primitive::kPrimNot: { 4044 // /* HeapReference<Object> */ out = *(base + offset) 4045 if (kEmitCompilerReadBarrier && kUseBakerReadBarrier) { 4046 Location temp_loc = locations->GetTemp(0); 4047 // Note that a potential implicit null check is handled in this 4048 // CodeGeneratorARM::GenerateFieldLoadWithBakerReadBarrier call. 4049 codegen_->GenerateFieldLoadWithBakerReadBarrier( 4050 instruction, out, base, offset, temp_loc, /* needs_null_check */ true); 4051 if (is_volatile) { 4052 codegen_->GenerateMemoryBarrier(MemBarrierKind::kLoadAny); 4053 } 4054 } else { 4055 __ LoadFromOffset(kLoadWord, out.AsRegister<Register>(), base, offset); 4056 codegen_->MaybeRecordImplicitNullCheck(instruction); 4057 if (is_volatile) { 4058 codegen_->GenerateMemoryBarrier(MemBarrierKind::kLoadAny); 4059 } 4060 // If read barriers are enabled, emit read barriers other than 4061 // Baker's using a slow path (and also unpoison the loaded 4062 // reference, if heap poisoning is enabled). 4063 codegen_->MaybeGenerateReadBarrierSlow(instruction, out, out, base_loc, offset); 4064 } 4065 break; 4066 } 4067 4068 case Primitive::kPrimLong: 4069 if (is_volatile && !atomic_ldrd_strd) { 4070 GenerateWideAtomicLoad(base, offset, 4071 out.AsRegisterPairLow<Register>(), 4072 out.AsRegisterPairHigh<Register>()); 4073 } else { 4074 __ LoadFromOffset(kLoadWordPair, out.AsRegisterPairLow<Register>(), base, offset); 4075 } 4076 break; 4077 4078 case Primitive::kPrimFloat: 4079 __ LoadSFromOffset(out.AsFpuRegister<SRegister>(), base, offset); 4080 break; 4081 4082 case Primitive::kPrimDouble: { 4083 DRegister out_reg = FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()); 4084 if (is_volatile && !atomic_ldrd_strd) { 4085 Register lo = locations->GetTemp(0).AsRegister<Register>(); 4086 Register hi = locations->GetTemp(1).AsRegister<Register>(); 4087 GenerateWideAtomicLoad(base, offset, lo, hi); 4088 codegen_->MaybeRecordImplicitNullCheck(instruction); 4089 __ vmovdrr(out_reg, lo, hi); 4090 } else { 4091 __ LoadDFromOffset(out_reg, base, offset); 4092 codegen_->MaybeRecordImplicitNullCheck(instruction); 4093 } 4094 break; 4095 } 4096 4097 case Primitive::kPrimVoid: 4098 LOG(FATAL) << "Unreachable type " << field_type; 4099 UNREACHABLE(); 4100 } 4101 4102 if (field_type == Primitive::kPrimNot || field_type == Primitive::kPrimDouble) { 4103 // Potential implicit null checks, in the case of reference or 4104 // double fields, are handled in the previous switch statement. 4105 } else { 4106 codegen_->MaybeRecordImplicitNullCheck(instruction); 4107 } 4108 4109 if (is_volatile) { 4110 if (field_type == Primitive::kPrimNot) { 4111 // Memory barriers, in the case of references, are also handled 4112 // in the previous switch statement. 4113 } else { 4114 codegen_->GenerateMemoryBarrier(MemBarrierKind::kLoadAny); 4115 } 4116 } 4117 } 4118 4119 void LocationsBuilderARM::VisitInstanceFieldSet(HInstanceFieldSet* instruction) { 4120 HandleFieldSet(instruction, instruction->GetFieldInfo()); 4121 } 4122 4123 void InstructionCodeGeneratorARM::VisitInstanceFieldSet(HInstanceFieldSet* instruction) { 4124 HandleFieldSet(instruction, instruction->GetFieldInfo(), instruction->GetValueCanBeNull()); 4125 } 4126 4127 void LocationsBuilderARM::VisitInstanceFieldGet(HInstanceFieldGet* instruction) { 4128 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4129 } 4130 4131 void InstructionCodeGeneratorARM::VisitInstanceFieldGet(HInstanceFieldGet* instruction) { 4132 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4133 } 4134 4135 void LocationsBuilderARM::VisitStaticFieldGet(HStaticFieldGet* instruction) { 4136 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4137 } 4138 4139 void InstructionCodeGeneratorARM::VisitStaticFieldGet(HStaticFieldGet* instruction) { 4140 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4141 } 4142 4143 void LocationsBuilderARM::VisitStaticFieldSet(HStaticFieldSet* instruction) { 4144 HandleFieldSet(instruction, instruction->GetFieldInfo()); 4145 } 4146 4147 void InstructionCodeGeneratorARM::VisitStaticFieldSet(HStaticFieldSet* instruction) { 4148 HandleFieldSet(instruction, instruction->GetFieldInfo(), instruction->GetValueCanBeNull()); 4149 } 4150 4151 void LocationsBuilderARM::VisitUnresolvedInstanceFieldGet( 4152 HUnresolvedInstanceFieldGet* instruction) { 4153 FieldAccessCallingConventionARM calling_convention; 4154 codegen_->CreateUnresolvedFieldLocationSummary( 4155 instruction, instruction->GetFieldType(), calling_convention); 4156 } 4157 4158 void InstructionCodeGeneratorARM::VisitUnresolvedInstanceFieldGet( 4159 HUnresolvedInstanceFieldGet* instruction) { 4160 FieldAccessCallingConventionARM calling_convention; 4161 codegen_->GenerateUnresolvedFieldAccess(instruction, 4162 instruction->GetFieldType(), 4163 instruction->GetFieldIndex(), 4164 instruction->GetDexPc(), 4165 calling_convention); 4166 } 4167 4168 void LocationsBuilderARM::VisitUnresolvedInstanceFieldSet( 4169 HUnresolvedInstanceFieldSet* instruction) { 4170 FieldAccessCallingConventionARM calling_convention; 4171 codegen_->CreateUnresolvedFieldLocationSummary( 4172 instruction, instruction->GetFieldType(), calling_convention); 4173 } 4174 4175 void InstructionCodeGeneratorARM::VisitUnresolvedInstanceFieldSet( 4176 HUnresolvedInstanceFieldSet* instruction) { 4177 FieldAccessCallingConventionARM calling_convention; 4178 codegen_->GenerateUnresolvedFieldAccess(instruction, 4179 instruction->GetFieldType(), 4180 instruction->GetFieldIndex(), 4181 instruction->GetDexPc(), 4182 calling_convention); 4183 } 4184 4185 void LocationsBuilderARM::VisitUnresolvedStaticFieldGet( 4186 HUnresolvedStaticFieldGet* instruction) { 4187 FieldAccessCallingConventionARM calling_convention; 4188 codegen_->CreateUnresolvedFieldLocationSummary( 4189 instruction, instruction->GetFieldType(), calling_convention); 4190 } 4191 4192 void InstructionCodeGeneratorARM::VisitUnresolvedStaticFieldGet( 4193 HUnresolvedStaticFieldGet* instruction) { 4194 FieldAccessCallingConventionARM calling_convention; 4195 codegen_->GenerateUnresolvedFieldAccess(instruction, 4196 instruction->GetFieldType(), 4197 instruction->GetFieldIndex(), 4198 instruction->GetDexPc(), 4199 calling_convention); 4200 } 4201 4202 void LocationsBuilderARM::VisitUnresolvedStaticFieldSet( 4203 HUnresolvedStaticFieldSet* instruction) { 4204 FieldAccessCallingConventionARM calling_convention; 4205 codegen_->CreateUnresolvedFieldLocationSummary( 4206 instruction, instruction->GetFieldType(), calling_convention); 4207 } 4208 4209 void InstructionCodeGeneratorARM::VisitUnresolvedStaticFieldSet( 4210 HUnresolvedStaticFieldSet* instruction) { 4211 FieldAccessCallingConventionARM calling_convention; 4212 codegen_->GenerateUnresolvedFieldAccess(instruction, 4213 instruction->GetFieldType(), 4214 instruction->GetFieldIndex(), 4215 instruction->GetDexPc(), 4216 calling_convention); 4217 } 4218 4219 void LocationsBuilderARM::VisitNullCheck(HNullCheck* instruction) { 4220 LocationSummary::CallKind call_kind = instruction->CanThrowIntoCatchBlock() 4221 ? LocationSummary::kCallOnSlowPath 4222 : LocationSummary::kNoCall; 4223 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 4224 locations->SetInAt(0, Location::RequiresRegister()); 4225 if (instruction->HasUses()) { 4226 locations->SetOut(Location::SameAsFirstInput()); 4227 } 4228 } 4229 4230 void CodeGeneratorARM::GenerateImplicitNullCheck(HNullCheck* instruction) { 4231 if (CanMoveNullCheckToUser(instruction)) { 4232 return; 4233 } 4234 Location obj = instruction->GetLocations()->InAt(0); 4235 4236 __ LoadFromOffset(kLoadWord, IP, obj.AsRegister<Register>(), 0); 4237 RecordPcInfo(instruction, instruction->GetDexPc()); 4238 } 4239 4240 void CodeGeneratorARM::GenerateExplicitNullCheck(HNullCheck* instruction) { 4241 SlowPathCode* slow_path = new (GetGraph()->GetArena()) NullCheckSlowPathARM(instruction); 4242 AddSlowPath(slow_path); 4243 4244 LocationSummary* locations = instruction->GetLocations(); 4245 Location obj = locations->InAt(0); 4246 4247 __ CompareAndBranchIfZero(obj.AsRegister<Register>(), slow_path->GetEntryLabel()); 4248 } 4249 4250 void InstructionCodeGeneratorARM::VisitNullCheck(HNullCheck* instruction) { 4251 codegen_->GenerateNullCheck(instruction); 4252 } 4253 4254 void LocationsBuilderARM::VisitArrayGet(HArrayGet* instruction) { 4255 bool object_array_get_with_read_barrier = 4256 kEmitCompilerReadBarrier && (instruction->GetType() == Primitive::kPrimNot); 4257 LocationSummary* locations = 4258 new (GetGraph()->GetArena()) LocationSummary(instruction, 4259 object_array_get_with_read_barrier ? 4260 LocationSummary::kCallOnSlowPath : 4261 LocationSummary::kNoCall); 4262 locations->SetInAt(0, Location::RequiresRegister()); 4263 locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1))); 4264 if (Primitive::IsFloatingPointType(instruction->GetType())) { 4265 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 4266 } else { 4267 // The output overlaps in the case of an object array get with 4268 // read barriers enabled: we do not want the move to overwrite the 4269 // array's location, as we need it to emit the read barrier. 4270 locations->SetOut( 4271 Location::RequiresRegister(), 4272 object_array_get_with_read_barrier ? Location::kOutputOverlap : Location::kNoOutputOverlap); 4273 } 4274 // We need a temporary register for the read barrier marking slow 4275 // path in CodeGeneratorARM::GenerateArrayLoadWithBakerReadBarrier. 4276 if (object_array_get_with_read_barrier && kUseBakerReadBarrier) { 4277 locations->AddTemp(Location::RequiresRegister()); 4278 } 4279 } 4280 4281 void InstructionCodeGeneratorARM::VisitArrayGet(HArrayGet* instruction) { 4282 LocationSummary* locations = instruction->GetLocations(); 4283 Location obj_loc = locations->InAt(0); 4284 Register obj = obj_loc.AsRegister<Register>(); 4285 Location index = locations->InAt(1); 4286 Location out_loc = locations->Out(); 4287 4288 Primitive::Type type = instruction->GetType(); 4289 switch (type) { 4290 case Primitive::kPrimBoolean: { 4291 uint32_t data_offset = mirror::Array::DataOffset(sizeof(uint8_t)).Uint32Value(); 4292 Register out = out_loc.AsRegister<Register>(); 4293 if (index.IsConstant()) { 4294 size_t offset = 4295 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_1) + data_offset; 4296 __ LoadFromOffset(kLoadUnsignedByte, out, obj, offset); 4297 } else { 4298 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>())); 4299 __ LoadFromOffset(kLoadUnsignedByte, out, IP, data_offset); 4300 } 4301 break; 4302 } 4303 4304 case Primitive::kPrimByte: { 4305 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int8_t)).Uint32Value(); 4306 Register out = out_loc.AsRegister<Register>(); 4307 if (index.IsConstant()) { 4308 size_t offset = 4309 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_1) + data_offset; 4310 __ LoadFromOffset(kLoadSignedByte, out, obj, offset); 4311 } else { 4312 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>())); 4313 __ LoadFromOffset(kLoadSignedByte, out, IP, data_offset); 4314 } 4315 break; 4316 } 4317 4318 case Primitive::kPrimShort: { 4319 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int16_t)).Uint32Value(); 4320 Register out = out_loc.AsRegister<Register>(); 4321 if (index.IsConstant()) { 4322 size_t offset = 4323 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_2) + data_offset; 4324 __ LoadFromOffset(kLoadSignedHalfword, out, obj, offset); 4325 } else { 4326 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_2)); 4327 __ LoadFromOffset(kLoadSignedHalfword, out, IP, data_offset); 4328 } 4329 break; 4330 } 4331 4332 case Primitive::kPrimChar: { 4333 uint32_t data_offset = mirror::Array::DataOffset(sizeof(uint16_t)).Uint32Value(); 4334 Register out = out_loc.AsRegister<Register>(); 4335 if (index.IsConstant()) { 4336 size_t offset = 4337 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_2) + data_offset; 4338 __ LoadFromOffset(kLoadUnsignedHalfword, out, obj, offset); 4339 } else { 4340 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_2)); 4341 __ LoadFromOffset(kLoadUnsignedHalfword, out, IP, data_offset); 4342 } 4343 break; 4344 } 4345 4346 case Primitive::kPrimInt: { 4347 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int32_t)).Uint32Value(); 4348 Register out = out_loc.AsRegister<Register>(); 4349 if (index.IsConstant()) { 4350 size_t offset = 4351 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4352 __ LoadFromOffset(kLoadWord, out, obj, offset); 4353 } else { 4354 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 4355 __ LoadFromOffset(kLoadWord, out, IP, data_offset); 4356 } 4357 break; 4358 } 4359 4360 case Primitive::kPrimNot: { 4361 static_assert( 4362 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 4363 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 4364 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int32_t)).Uint32Value(); 4365 // /* HeapReference<Object> */ out = 4366 // *(obj + data_offset + index * sizeof(HeapReference<Object>)) 4367 if (kEmitCompilerReadBarrier && kUseBakerReadBarrier) { 4368 Location temp = locations->GetTemp(0); 4369 // Note that a potential implicit null check is handled in this 4370 // CodeGeneratorARM::GenerateArrayLoadWithBakerReadBarrier call. 4371 codegen_->GenerateArrayLoadWithBakerReadBarrier( 4372 instruction, out_loc, obj, data_offset, index, temp, /* needs_null_check */ true); 4373 } else { 4374 Register out = out_loc.AsRegister<Register>(); 4375 if (index.IsConstant()) { 4376 size_t offset = 4377 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4378 __ LoadFromOffset(kLoadWord, out, obj, offset); 4379 codegen_->MaybeRecordImplicitNullCheck(instruction); 4380 // If read barriers are enabled, emit read barriers other than 4381 // Baker's using a slow path (and also unpoison the loaded 4382 // reference, if heap poisoning is enabled). 4383 codegen_->MaybeGenerateReadBarrierSlow(instruction, out_loc, out_loc, obj_loc, offset); 4384 } else { 4385 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 4386 __ LoadFromOffset(kLoadWord, out, IP, data_offset); 4387 codegen_->MaybeRecordImplicitNullCheck(instruction); 4388 // If read barriers are enabled, emit read barriers other than 4389 // Baker's using a slow path (and also unpoison the loaded 4390 // reference, if heap poisoning is enabled). 4391 codegen_->MaybeGenerateReadBarrierSlow( 4392 instruction, out_loc, out_loc, obj_loc, data_offset, index); 4393 } 4394 } 4395 break; 4396 } 4397 4398 case Primitive::kPrimLong: { 4399 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int64_t)).Uint32Value(); 4400 if (index.IsConstant()) { 4401 size_t offset = 4402 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 4403 __ LoadFromOffset(kLoadWordPair, out_loc.AsRegisterPairLow<Register>(), obj, offset); 4404 } else { 4405 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 4406 __ LoadFromOffset(kLoadWordPair, out_loc.AsRegisterPairLow<Register>(), IP, data_offset); 4407 } 4408 break; 4409 } 4410 4411 case Primitive::kPrimFloat: { 4412 uint32_t data_offset = mirror::Array::DataOffset(sizeof(float)).Uint32Value(); 4413 SRegister out = out_loc.AsFpuRegister<SRegister>(); 4414 if (index.IsConstant()) { 4415 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4416 __ LoadSFromOffset(out, obj, offset); 4417 } else { 4418 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 4419 __ LoadSFromOffset(out, IP, data_offset); 4420 } 4421 break; 4422 } 4423 4424 case Primitive::kPrimDouble: { 4425 uint32_t data_offset = mirror::Array::DataOffset(sizeof(double)).Uint32Value(); 4426 SRegister out = out_loc.AsFpuRegisterPairLow<SRegister>(); 4427 if (index.IsConstant()) { 4428 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 4429 __ LoadDFromOffset(FromLowSToD(out), obj, offset); 4430 } else { 4431 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 4432 __ LoadDFromOffset(FromLowSToD(out), IP, data_offset); 4433 } 4434 break; 4435 } 4436 4437 case Primitive::kPrimVoid: 4438 LOG(FATAL) << "Unreachable type " << type; 4439 UNREACHABLE(); 4440 } 4441 4442 if (type == Primitive::kPrimNot) { 4443 // Potential implicit null checks, in the case of reference 4444 // arrays, are handled in the previous switch statement. 4445 } else { 4446 codegen_->MaybeRecordImplicitNullCheck(instruction); 4447 } 4448 } 4449 4450 void LocationsBuilderARM::VisitArraySet(HArraySet* instruction) { 4451 Primitive::Type value_type = instruction->GetComponentType(); 4452 4453 bool needs_write_barrier = 4454 CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue()); 4455 bool may_need_runtime_call_for_type_check = instruction->NeedsTypeCheck(); 4456 bool object_array_set_with_read_barrier = 4457 kEmitCompilerReadBarrier && (value_type == Primitive::kPrimNot); 4458 4459 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary( 4460 instruction, 4461 (may_need_runtime_call_for_type_check || object_array_set_with_read_barrier) ? 4462 LocationSummary::kCallOnSlowPath : 4463 LocationSummary::kNoCall); 4464 4465 locations->SetInAt(0, Location::RequiresRegister()); 4466 locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1))); 4467 if (Primitive::IsFloatingPointType(value_type)) { 4468 locations->SetInAt(2, Location::RequiresFpuRegister()); 4469 } else { 4470 locations->SetInAt(2, Location::RequiresRegister()); 4471 } 4472 if (needs_write_barrier) { 4473 // Temporary registers for the write barrier. 4474 locations->AddTemp(Location::RequiresRegister()); // Possibly used for ref. poisoning too. 4475 locations->AddTemp(Location::RequiresRegister()); 4476 } 4477 } 4478 4479 void InstructionCodeGeneratorARM::VisitArraySet(HArraySet* instruction) { 4480 LocationSummary* locations = instruction->GetLocations(); 4481 Location array_loc = locations->InAt(0); 4482 Register array = array_loc.AsRegister<Register>(); 4483 Location index = locations->InAt(1); 4484 Primitive::Type value_type = instruction->GetComponentType(); 4485 bool may_need_runtime_call_for_type_check = instruction->NeedsTypeCheck(); 4486 bool needs_write_barrier = 4487 CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue()); 4488 4489 switch (value_type) { 4490 case Primitive::kPrimBoolean: 4491 case Primitive::kPrimByte: { 4492 uint32_t data_offset = mirror::Array::DataOffset(sizeof(uint8_t)).Uint32Value(); 4493 Register value = locations->InAt(2).AsRegister<Register>(); 4494 if (index.IsConstant()) { 4495 size_t offset = 4496 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_1) + data_offset; 4497 __ StoreToOffset(kStoreByte, value, array, offset); 4498 } else { 4499 __ add(IP, array, ShifterOperand(index.AsRegister<Register>())); 4500 __ StoreToOffset(kStoreByte, value, IP, data_offset); 4501 } 4502 break; 4503 } 4504 4505 case Primitive::kPrimShort: 4506 case Primitive::kPrimChar: { 4507 uint32_t data_offset = mirror::Array::DataOffset(sizeof(uint16_t)).Uint32Value(); 4508 Register value = locations->InAt(2).AsRegister<Register>(); 4509 if (index.IsConstant()) { 4510 size_t offset = 4511 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_2) + data_offset; 4512 __ StoreToOffset(kStoreHalfword, value, array, offset); 4513 } else { 4514 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_2)); 4515 __ StoreToOffset(kStoreHalfword, value, IP, data_offset); 4516 } 4517 break; 4518 } 4519 4520 case Primitive::kPrimNot: { 4521 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int32_t)).Uint32Value(); 4522 Location value_loc = locations->InAt(2); 4523 Register value = value_loc.AsRegister<Register>(); 4524 Register source = value; 4525 4526 if (instruction->InputAt(2)->IsNullConstant()) { 4527 // Just setting null. 4528 if (index.IsConstant()) { 4529 size_t offset = 4530 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4531 __ StoreToOffset(kStoreWord, source, array, offset); 4532 } else { 4533 DCHECK(index.IsRegister()) << index; 4534 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 4535 __ StoreToOffset(kStoreWord, source, IP, data_offset); 4536 } 4537 codegen_->MaybeRecordImplicitNullCheck(instruction); 4538 DCHECK(!needs_write_barrier); 4539 DCHECK(!may_need_runtime_call_for_type_check); 4540 break; 4541 } 4542 4543 DCHECK(needs_write_barrier); 4544 Register temp1 = locations->GetTemp(0).AsRegister<Register>(); 4545 Register temp2 = locations->GetTemp(1).AsRegister<Register>(); 4546 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 4547 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 4548 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 4549 Label done; 4550 SlowPathCode* slow_path = nullptr; 4551 4552 if (may_need_runtime_call_for_type_check) { 4553 slow_path = new (GetGraph()->GetArena()) ArraySetSlowPathARM(instruction); 4554 codegen_->AddSlowPath(slow_path); 4555 if (instruction->GetValueCanBeNull()) { 4556 Label non_zero; 4557 __ CompareAndBranchIfNonZero(value, &non_zero); 4558 if (index.IsConstant()) { 4559 size_t offset = 4560 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4561 __ StoreToOffset(kStoreWord, value, array, offset); 4562 } else { 4563 DCHECK(index.IsRegister()) << index; 4564 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 4565 __ StoreToOffset(kStoreWord, value, IP, data_offset); 4566 } 4567 codegen_->MaybeRecordImplicitNullCheck(instruction); 4568 __ b(&done); 4569 __ Bind(&non_zero); 4570 } 4571 4572 if (kEmitCompilerReadBarrier) { 4573 // When read barriers are enabled, the type checking 4574 // instrumentation requires two read barriers: 4575 // 4576 // __ Mov(temp2, temp1); 4577 // // /* HeapReference<Class> */ temp1 = temp1->component_type_ 4578 // __ LoadFromOffset(kLoadWord, temp1, temp1, component_offset); 4579 // codegen_->GenerateReadBarrierSlow( 4580 // instruction, temp1_loc, temp1_loc, temp2_loc, component_offset); 4581 // 4582 // // /* HeapReference<Class> */ temp2 = value->klass_ 4583 // __ LoadFromOffset(kLoadWord, temp2, value, class_offset); 4584 // codegen_->GenerateReadBarrierSlow( 4585 // instruction, temp2_loc, temp2_loc, value_loc, class_offset, temp1_loc); 4586 // 4587 // __ cmp(temp1, ShifterOperand(temp2)); 4588 // 4589 // However, the second read barrier may trash `temp`, as it 4590 // is a temporary register, and as such would not be saved 4591 // along with live registers before calling the runtime (nor 4592 // restored afterwards). So in this case, we bail out and 4593 // delegate the work to the array set slow path. 4594 // 4595 // TODO: Extend the register allocator to support a new 4596 // "(locally) live temp" location so as to avoid always 4597 // going into the slow path when read barriers are enabled. 4598 __ b(slow_path->GetEntryLabel()); 4599 } else { 4600 // /* HeapReference<Class> */ temp1 = array->klass_ 4601 __ LoadFromOffset(kLoadWord, temp1, array, class_offset); 4602 codegen_->MaybeRecordImplicitNullCheck(instruction); 4603 __ MaybeUnpoisonHeapReference(temp1); 4604 4605 // /* HeapReference<Class> */ temp1 = temp1->component_type_ 4606 __ LoadFromOffset(kLoadWord, temp1, temp1, component_offset); 4607 // /* HeapReference<Class> */ temp2 = value->klass_ 4608 __ LoadFromOffset(kLoadWord, temp2, value, class_offset); 4609 // If heap poisoning is enabled, no need to unpoison `temp1` 4610 // nor `temp2`, as we are comparing two poisoned references. 4611 __ cmp(temp1, ShifterOperand(temp2)); 4612 4613 if (instruction->StaticTypeOfArrayIsObjectArray()) { 4614 Label do_put; 4615 __ b(&do_put, EQ); 4616 // If heap poisoning is enabled, the `temp1` reference has 4617 // not been unpoisoned yet; unpoison it now. 4618 __ MaybeUnpoisonHeapReference(temp1); 4619 4620 // /* HeapReference<Class> */ temp1 = temp1->super_class_ 4621 __ LoadFromOffset(kLoadWord, temp1, temp1, super_offset); 4622 // If heap poisoning is enabled, no need to unpoison 4623 // `temp1`, as we are comparing against null below. 4624 __ CompareAndBranchIfNonZero(temp1, slow_path->GetEntryLabel()); 4625 __ Bind(&do_put); 4626 } else { 4627 __ b(slow_path->GetEntryLabel(), NE); 4628 } 4629 } 4630 } 4631 4632 if (kPoisonHeapReferences) { 4633 // Note that in the case where `value` is a null reference, 4634 // we do not enter this block, as a null reference does not 4635 // need poisoning. 4636 DCHECK_EQ(value_type, Primitive::kPrimNot); 4637 __ Mov(temp1, value); 4638 __ PoisonHeapReference(temp1); 4639 source = temp1; 4640 } 4641 4642 if (index.IsConstant()) { 4643 size_t offset = 4644 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4645 __ StoreToOffset(kStoreWord, source, array, offset); 4646 } else { 4647 DCHECK(index.IsRegister()) << index; 4648 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 4649 __ StoreToOffset(kStoreWord, source, IP, data_offset); 4650 } 4651 4652 if (!may_need_runtime_call_for_type_check) { 4653 codegen_->MaybeRecordImplicitNullCheck(instruction); 4654 } 4655 4656 codegen_->MarkGCCard(temp1, temp2, array, value, instruction->GetValueCanBeNull()); 4657 4658 if (done.IsLinked()) { 4659 __ Bind(&done); 4660 } 4661 4662 if (slow_path != nullptr) { 4663 __ Bind(slow_path->GetExitLabel()); 4664 } 4665 4666 break; 4667 } 4668 4669 case Primitive::kPrimInt: { 4670 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int32_t)).Uint32Value(); 4671 Register value = locations->InAt(2).AsRegister<Register>(); 4672 if (index.IsConstant()) { 4673 size_t offset = 4674 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4675 __ StoreToOffset(kStoreWord, value, array, offset); 4676 } else { 4677 DCHECK(index.IsRegister()) << index; 4678 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 4679 __ StoreToOffset(kStoreWord, value, IP, data_offset); 4680 } 4681 break; 4682 } 4683 4684 case Primitive::kPrimLong: { 4685 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int64_t)).Uint32Value(); 4686 Location value = locations->InAt(2); 4687 if (index.IsConstant()) { 4688 size_t offset = 4689 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 4690 __ StoreToOffset(kStoreWordPair, value.AsRegisterPairLow<Register>(), array, offset); 4691 } else { 4692 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 4693 __ StoreToOffset(kStoreWordPair, value.AsRegisterPairLow<Register>(), IP, data_offset); 4694 } 4695 break; 4696 } 4697 4698 case Primitive::kPrimFloat: { 4699 uint32_t data_offset = mirror::Array::DataOffset(sizeof(float)).Uint32Value(); 4700 Location value = locations->InAt(2); 4701 DCHECK(value.IsFpuRegister()); 4702 if (index.IsConstant()) { 4703 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4704 __ StoreSToOffset(value.AsFpuRegister<SRegister>(), array, offset); 4705 } else { 4706 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 4707 __ StoreSToOffset(value.AsFpuRegister<SRegister>(), IP, data_offset); 4708 } 4709 break; 4710 } 4711 4712 case Primitive::kPrimDouble: { 4713 uint32_t data_offset = mirror::Array::DataOffset(sizeof(double)).Uint32Value(); 4714 Location value = locations->InAt(2); 4715 DCHECK(value.IsFpuRegisterPair()); 4716 if (index.IsConstant()) { 4717 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 4718 __ StoreDToOffset(FromLowSToD(value.AsFpuRegisterPairLow<SRegister>()), array, offset); 4719 } else { 4720 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 4721 __ StoreDToOffset(FromLowSToD(value.AsFpuRegisterPairLow<SRegister>()), IP, data_offset); 4722 } 4723 4724 break; 4725 } 4726 4727 case Primitive::kPrimVoid: 4728 LOG(FATAL) << "Unreachable type " << value_type; 4729 UNREACHABLE(); 4730 } 4731 4732 // Objects are handled in the switch. 4733 if (value_type != Primitive::kPrimNot) { 4734 codegen_->MaybeRecordImplicitNullCheck(instruction); 4735 } 4736 } 4737 4738 void LocationsBuilderARM::VisitArrayLength(HArrayLength* instruction) { 4739 LocationSummary* locations = 4740 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 4741 locations->SetInAt(0, Location::RequiresRegister()); 4742 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 4743 } 4744 4745 void InstructionCodeGeneratorARM::VisitArrayLength(HArrayLength* instruction) { 4746 LocationSummary* locations = instruction->GetLocations(); 4747 uint32_t offset = mirror::Array::LengthOffset().Uint32Value(); 4748 Register obj = locations->InAt(0).AsRegister<Register>(); 4749 Register out = locations->Out().AsRegister<Register>(); 4750 __ LoadFromOffset(kLoadWord, out, obj, offset); 4751 codegen_->MaybeRecordImplicitNullCheck(instruction); 4752 } 4753 4754 void LocationsBuilderARM::VisitBoundsCheck(HBoundsCheck* instruction) { 4755 LocationSummary::CallKind call_kind = instruction->CanThrowIntoCatchBlock() 4756 ? LocationSummary::kCallOnSlowPath 4757 : LocationSummary::kNoCall; 4758 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 4759 locations->SetInAt(0, Location::RequiresRegister()); 4760 locations->SetInAt(1, Location::RequiresRegister()); 4761 if (instruction->HasUses()) { 4762 locations->SetOut(Location::SameAsFirstInput()); 4763 } 4764 } 4765 4766 void InstructionCodeGeneratorARM::VisitBoundsCheck(HBoundsCheck* instruction) { 4767 LocationSummary* locations = instruction->GetLocations(); 4768 SlowPathCode* slow_path = 4769 new (GetGraph()->GetArena()) BoundsCheckSlowPathARM(instruction); 4770 codegen_->AddSlowPath(slow_path); 4771 4772 Register index = locations->InAt(0).AsRegister<Register>(); 4773 Register length = locations->InAt(1).AsRegister<Register>(); 4774 4775 __ cmp(index, ShifterOperand(length)); 4776 __ b(slow_path->GetEntryLabel(), HS); 4777 } 4778 4779 void CodeGeneratorARM::MarkGCCard(Register temp, 4780 Register card, 4781 Register object, 4782 Register value, 4783 bool can_be_null) { 4784 Label is_null; 4785 if (can_be_null) { 4786 __ CompareAndBranchIfZero(value, &is_null); 4787 } 4788 __ LoadFromOffset(kLoadWord, card, TR, Thread::CardTableOffset<kArmWordSize>().Int32Value()); 4789 __ Lsr(temp, object, gc::accounting::CardTable::kCardShift); 4790 __ strb(card, Address(card, temp)); 4791 if (can_be_null) { 4792 __ Bind(&is_null); 4793 } 4794 } 4795 4796 void LocationsBuilderARM::VisitParallelMove(HParallelMove* instruction ATTRIBUTE_UNUSED) { 4797 LOG(FATAL) << "Unreachable"; 4798 } 4799 4800 void InstructionCodeGeneratorARM::VisitParallelMove(HParallelMove* instruction) { 4801 codegen_->GetMoveResolver()->EmitNativeCode(instruction); 4802 } 4803 4804 void LocationsBuilderARM::VisitSuspendCheck(HSuspendCheck* instruction) { 4805 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnSlowPath); 4806 } 4807 4808 void InstructionCodeGeneratorARM::VisitSuspendCheck(HSuspendCheck* instruction) { 4809 HBasicBlock* block = instruction->GetBlock(); 4810 if (block->GetLoopInformation() != nullptr) { 4811 DCHECK(block->GetLoopInformation()->GetSuspendCheck() == instruction); 4812 // The back edge will generate the suspend check. 4813 return; 4814 } 4815 if (block->IsEntryBlock() && instruction->GetNext()->IsGoto()) { 4816 // The goto will generate the suspend check. 4817 return; 4818 } 4819 GenerateSuspendCheck(instruction, nullptr); 4820 } 4821 4822 void InstructionCodeGeneratorARM::GenerateSuspendCheck(HSuspendCheck* instruction, 4823 HBasicBlock* successor) { 4824 SuspendCheckSlowPathARM* slow_path = 4825 down_cast<SuspendCheckSlowPathARM*>(instruction->GetSlowPath()); 4826 if (slow_path == nullptr) { 4827 slow_path = new (GetGraph()->GetArena()) SuspendCheckSlowPathARM(instruction, successor); 4828 instruction->SetSlowPath(slow_path); 4829 codegen_->AddSlowPath(slow_path); 4830 if (successor != nullptr) { 4831 DCHECK(successor->IsLoopHeader()); 4832 codegen_->ClearSpillSlotsFromLoopPhisInStackMap(instruction); 4833 } 4834 } else { 4835 DCHECK_EQ(slow_path->GetSuccessor(), successor); 4836 } 4837 4838 __ LoadFromOffset( 4839 kLoadUnsignedHalfword, IP, TR, Thread::ThreadFlagsOffset<kArmWordSize>().Int32Value()); 4840 if (successor == nullptr) { 4841 __ CompareAndBranchIfNonZero(IP, slow_path->GetEntryLabel()); 4842 __ Bind(slow_path->GetReturnLabel()); 4843 } else { 4844 __ CompareAndBranchIfZero(IP, codegen_->GetLabelOf(successor)); 4845 __ b(slow_path->GetEntryLabel()); 4846 } 4847 } 4848 4849 ArmAssembler* ParallelMoveResolverARM::GetAssembler() const { 4850 return codegen_->GetAssembler(); 4851 } 4852 4853 void ParallelMoveResolverARM::EmitMove(size_t index) { 4854 MoveOperands* move = moves_[index]; 4855 Location source = move->GetSource(); 4856 Location destination = move->GetDestination(); 4857 4858 if (source.IsRegister()) { 4859 if (destination.IsRegister()) { 4860 __ Mov(destination.AsRegister<Register>(), source.AsRegister<Register>()); 4861 } else if (destination.IsFpuRegister()) { 4862 __ vmovsr(destination.AsFpuRegister<SRegister>(), source.AsRegister<Register>()); 4863 } else { 4864 DCHECK(destination.IsStackSlot()); 4865 __ StoreToOffset(kStoreWord, source.AsRegister<Register>(), 4866 SP, destination.GetStackIndex()); 4867 } 4868 } else if (source.IsStackSlot()) { 4869 if (destination.IsRegister()) { 4870 __ LoadFromOffset(kLoadWord, destination.AsRegister<Register>(), 4871 SP, source.GetStackIndex()); 4872 } else if (destination.IsFpuRegister()) { 4873 __ LoadSFromOffset(destination.AsFpuRegister<SRegister>(), SP, source.GetStackIndex()); 4874 } else { 4875 DCHECK(destination.IsStackSlot()); 4876 __ LoadFromOffset(kLoadWord, IP, SP, source.GetStackIndex()); 4877 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 4878 } 4879 } else if (source.IsFpuRegister()) { 4880 if (destination.IsRegister()) { 4881 __ vmovrs(destination.AsRegister<Register>(), source.AsFpuRegister<SRegister>()); 4882 } else if (destination.IsFpuRegister()) { 4883 __ vmovs(destination.AsFpuRegister<SRegister>(), source.AsFpuRegister<SRegister>()); 4884 } else { 4885 DCHECK(destination.IsStackSlot()); 4886 __ StoreSToOffset(source.AsFpuRegister<SRegister>(), SP, destination.GetStackIndex()); 4887 } 4888 } else if (source.IsDoubleStackSlot()) { 4889 if (destination.IsDoubleStackSlot()) { 4890 __ LoadDFromOffset(DTMP, SP, source.GetStackIndex()); 4891 __ StoreDToOffset(DTMP, SP, destination.GetStackIndex()); 4892 } else if (destination.IsRegisterPair()) { 4893 DCHECK(ExpectedPairLayout(destination)); 4894 __ LoadFromOffset( 4895 kLoadWordPair, destination.AsRegisterPairLow<Register>(), SP, source.GetStackIndex()); 4896 } else { 4897 DCHECK(destination.IsFpuRegisterPair()) << destination; 4898 __ LoadDFromOffset(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 4899 SP, 4900 source.GetStackIndex()); 4901 } 4902 } else if (source.IsRegisterPair()) { 4903 if (destination.IsRegisterPair()) { 4904 __ Mov(destination.AsRegisterPairLow<Register>(), source.AsRegisterPairLow<Register>()); 4905 __ Mov(destination.AsRegisterPairHigh<Register>(), source.AsRegisterPairHigh<Register>()); 4906 } else if (destination.IsFpuRegisterPair()) { 4907 __ vmovdrr(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 4908 source.AsRegisterPairLow<Register>(), 4909 source.AsRegisterPairHigh<Register>()); 4910 } else { 4911 DCHECK(destination.IsDoubleStackSlot()) << destination; 4912 DCHECK(ExpectedPairLayout(source)); 4913 __ StoreToOffset( 4914 kStoreWordPair, source.AsRegisterPairLow<Register>(), SP, destination.GetStackIndex()); 4915 } 4916 } else if (source.IsFpuRegisterPair()) { 4917 if (destination.IsRegisterPair()) { 4918 __ vmovrrd(destination.AsRegisterPairLow<Register>(), 4919 destination.AsRegisterPairHigh<Register>(), 4920 FromLowSToD(source.AsFpuRegisterPairLow<SRegister>())); 4921 } else if (destination.IsFpuRegisterPair()) { 4922 __ vmovd(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 4923 FromLowSToD(source.AsFpuRegisterPairLow<SRegister>())); 4924 } else { 4925 DCHECK(destination.IsDoubleStackSlot()) << destination; 4926 __ StoreDToOffset(FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()), 4927 SP, 4928 destination.GetStackIndex()); 4929 } 4930 } else { 4931 DCHECK(source.IsConstant()) << source; 4932 HConstant* constant = source.GetConstant(); 4933 if (constant->IsIntConstant() || constant->IsNullConstant()) { 4934 int32_t value = CodeGenerator::GetInt32ValueOf(constant); 4935 if (destination.IsRegister()) { 4936 __ LoadImmediate(destination.AsRegister<Register>(), value); 4937 } else { 4938 DCHECK(destination.IsStackSlot()); 4939 __ LoadImmediate(IP, value); 4940 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 4941 } 4942 } else if (constant->IsLongConstant()) { 4943 int64_t value = constant->AsLongConstant()->GetValue(); 4944 if (destination.IsRegisterPair()) { 4945 __ LoadImmediate(destination.AsRegisterPairLow<Register>(), Low32Bits(value)); 4946 __ LoadImmediate(destination.AsRegisterPairHigh<Register>(), High32Bits(value)); 4947 } else { 4948 DCHECK(destination.IsDoubleStackSlot()) << destination; 4949 __ LoadImmediate(IP, Low32Bits(value)); 4950 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 4951 __ LoadImmediate(IP, High32Bits(value)); 4952 __ StoreToOffset(kStoreWord, IP, SP, destination.GetHighStackIndex(kArmWordSize)); 4953 } 4954 } else if (constant->IsDoubleConstant()) { 4955 double value = constant->AsDoubleConstant()->GetValue(); 4956 if (destination.IsFpuRegisterPair()) { 4957 __ LoadDImmediate(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), value); 4958 } else { 4959 DCHECK(destination.IsDoubleStackSlot()) << destination; 4960 uint64_t int_value = bit_cast<uint64_t, double>(value); 4961 __ LoadImmediate(IP, Low32Bits(int_value)); 4962 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 4963 __ LoadImmediate(IP, High32Bits(int_value)); 4964 __ StoreToOffset(kStoreWord, IP, SP, destination.GetHighStackIndex(kArmWordSize)); 4965 } 4966 } else { 4967 DCHECK(constant->IsFloatConstant()) << constant->DebugName(); 4968 float value = constant->AsFloatConstant()->GetValue(); 4969 if (destination.IsFpuRegister()) { 4970 __ LoadSImmediate(destination.AsFpuRegister<SRegister>(), value); 4971 } else { 4972 DCHECK(destination.IsStackSlot()); 4973 __ LoadImmediate(IP, bit_cast<int32_t, float>(value)); 4974 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 4975 } 4976 } 4977 } 4978 } 4979 4980 void ParallelMoveResolverARM::Exchange(Register reg, int mem) { 4981 __ Mov(IP, reg); 4982 __ LoadFromOffset(kLoadWord, reg, SP, mem); 4983 __ StoreToOffset(kStoreWord, IP, SP, mem); 4984 } 4985 4986 void ParallelMoveResolverARM::Exchange(int mem1, int mem2) { 4987 ScratchRegisterScope ensure_scratch(this, IP, R0, codegen_->GetNumberOfCoreRegisters()); 4988 int stack_offset = ensure_scratch.IsSpilled() ? kArmWordSize : 0; 4989 __ LoadFromOffset(kLoadWord, static_cast<Register>(ensure_scratch.GetRegister()), 4990 SP, mem1 + stack_offset); 4991 __ LoadFromOffset(kLoadWord, IP, SP, mem2 + stack_offset); 4992 __ StoreToOffset(kStoreWord, static_cast<Register>(ensure_scratch.GetRegister()), 4993 SP, mem2 + stack_offset); 4994 __ StoreToOffset(kStoreWord, IP, SP, mem1 + stack_offset); 4995 } 4996 4997 void ParallelMoveResolverARM::EmitSwap(size_t index) { 4998 MoveOperands* move = moves_[index]; 4999 Location source = move->GetSource(); 5000 Location destination = move->GetDestination(); 5001 5002 if (source.IsRegister() && destination.IsRegister()) { 5003 DCHECK_NE(source.AsRegister<Register>(), IP); 5004 DCHECK_NE(destination.AsRegister<Register>(), IP); 5005 __ Mov(IP, source.AsRegister<Register>()); 5006 __ Mov(source.AsRegister<Register>(), destination.AsRegister<Register>()); 5007 __ Mov(destination.AsRegister<Register>(), IP); 5008 } else if (source.IsRegister() && destination.IsStackSlot()) { 5009 Exchange(source.AsRegister<Register>(), destination.GetStackIndex()); 5010 } else if (source.IsStackSlot() && destination.IsRegister()) { 5011 Exchange(destination.AsRegister<Register>(), source.GetStackIndex()); 5012 } else if (source.IsStackSlot() && destination.IsStackSlot()) { 5013 Exchange(source.GetStackIndex(), destination.GetStackIndex()); 5014 } else if (source.IsFpuRegister() && destination.IsFpuRegister()) { 5015 __ vmovrs(IP, source.AsFpuRegister<SRegister>()); 5016 __ vmovs(source.AsFpuRegister<SRegister>(), destination.AsFpuRegister<SRegister>()); 5017 __ vmovsr(destination.AsFpuRegister<SRegister>(), IP); 5018 } else if (source.IsRegisterPair() && destination.IsRegisterPair()) { 5019 __ vmovdrr(DTMP, source.AsRegisterPairLow<Register>(), source.AsRegisterPairHigh<Register>()); 5020 __ Mov(source.AsRegisterPairLow<Register>(), destination.AsRegisterPairLow<Register>()); 5021 __ Mov(source.AsRegisterPairHigh<Register>(), destination.AsRegisterPairHigh<Register>()); 5022 __ vmovrrd(destination.AsRegisterPairLow<Register>(), 5023 destination.AsRegisterPairHigh<Register>(), 5024 DTMP); 5025 } else if (source.IsRegisterPair() || destination.IsRegisterPair()) { 5026 Register low_reg = source.IsRegisterPair() 5027 ? source.AsRegisterPairLow<Register>() 5028 : destination.AsRegisterPairLow<Register>(); 5029 int mem = source.IsRegisterPair() 5030 ? destination.GetStackIndex() 5031 : source.GetStackIndex(); 5032 DCHECK(ExpectedPairLayout(source.IsRegisterPair() ? source : destination)); 5033 __ vmovdrr(DTMP, low_reg, static_cast<Register>(low_reg + 1)); 5034 __ LoadFromOffset(kLoadWordPair, low_reg, SP, mem); 5035 __ StoreDToOffset(DTMP, SP, mem); 5036 } else if (source.IsFpuRegisterPair() && destination.IsFpuRegisterPair()) { 5037 DRegister first = FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()); 5038 DRegister second = FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()); 5039 __ vmovd(DTMP, first); 5040 __ vmovd(first, second); 5041 __ vmovd(second, DTMP); 5042 } else if (source.IsFpuRegisterPair() || destination.IsFpuRegisterPair()) { 5043 DRegister reg = source.IsFpuRegisterPair() 5044 ? FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()) 5045 : FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()); 5046 int mem = source.IsFpuRegisterPair() 5047 ? destination.GetStackIndex() 5048 : source.GetStackIndex(); 5049 __ vmovd(DTMP, reg); 5050 __ LoadDFromOffset(reg, SP, mem); 5051 __ StoreDToOffset(DTMP, SP, mem); 5052 } else if (source.IsFpuRegister() || destination.IsFpuRegister()) { 5053 SRegister reg = source.IsFpuRegister() ? source.AsFpuRegister<SRegister>() 5054 : destination.AsFpuRegister<SRegister>(); 5055 int mem = source.IsFpuRegister() 5056 ? destination.GetStackIndex() 5057 : source.GetStackIndex(); 5058 5059 __ vmovrs(IP, reg); 5060 __ LoadSFromOffset(reg, SP, mem); 5061 __ StoreToOffset(kStoreWord, IP, SP, mem); 5062 } else if (source.IsDoubleStackSlot() && destination.IsDoubleStackSlot()) { 5063 Exchange(source.GetStackIndex(), destination.GetStackIndex()); 5064 Exchange(source.GetHighStackIndex(kArmWordSize), destination.GetHighStackIndex(kArmWordSize)); 5065 } else { 5066 LOG(FATAL) << "Unimplemented" << source << " <-> " << destination; 5067 } 5068 } 5069 5070 void ParallelMoveResolverARM::SpillScratch(int reg) { 5071 __ Push(static_cast<Register>(reg)); 5072 } 5073 5074 void ParallelMoveResolverARM::RestoreScratch(int reg) { 5075 __ Pop(static_cast<Register>(reg)); 5076 } 5077 5078 void LocationsBuilderARM::VisitLoadClass(HLoadClass* cls) { 5079 InvokeRuntimeCallingConvention calling_convention; 5080 CodeGenerator::CreateLoadClassLocationSummary( 5081 cls, 5082 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 5083 Location::RegisterLocation(R0), 5084 /* code_generator_supports_read_barrier */ true); 5085 } 5086 5087 void InstructionCodeGeneratorARM::VisitLoadClass(HLoadClass* cls) { 5088 LocationSummary* locations = cls->GetLocations(); 5089 if (cls->NeedsAccessCheck()) { 5090 codegen_->MoveConstant(locations->GetTemp(0), cls->GetTypeIndex()); 5091 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pInitializeTypeAndVerifyAccess), 5092 cls, 5093 cls->GetDexPc(), 5094 nullptr); 5095 CheckEntrypointTypes<kQuickInitializeTypeAndVerifyAccess, void*, uint32_t>(); 5096 return; 5097 } 5098 5099 Location out_loc = locations->Out(); 5100 Register out = out_loc.AsRegister<Register>(); 5101 Register current_method = locations->InAt(0).AsRegister<Register>(); 5102 5103 if (cls->IsReferrersClass()) { 5104 DCHECK(!cls->CanCallRuntime()); 5105 DCHECK(!cls->MustGenerateClinitCheck()); 5106 // /* GcRoot<mirror::Class> */ out = current_method->declaring_class_ 5107 GenerateGcRootFieldLoad( 5108 cls, out_loc, current_method, ArtMethod::DeclaringClassOffset().Int32Value()); 5109 } else { 5110 // /* GcRoot<mirror::Class>[] */ out = 5111 // current_method.ptr_sized_fields_->dex_cache_resolved_types_ 5112 __ LoadFromOffset(kLoadWord, 5113 out, 5114 current_method, 5115 ArtMethod::DexCacheResolvedTypesOffset(kArmPointerSize).Int32Value()); 5116 // /* GcRoot<mirror::Class> */ out = out[type_index] 5117 GenerateGcRootFieldLoad(cls, out_loc, out, CodeGenerator::GetCacheOffset(cls->GetTypeIndex())); 5118 5119 if (!cls->IsInDexCache() || cls->MustGenerateClinitCheck()) { 5120 DCHECK(cls->CanCallRuntime()); 5121 SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathARM( 5122 cls, cls, cls->GetDexPc(), cls->MustGenerateClinitCheck()); 5123 codegen_->AddSlowPath(slow_path); 5124 if (!cls->IsInDexCache()) { 5125 __ CompareAndBranchIfZero(out, slow_path->GetEntryLabel()); 5126 } 5127 if (cls->MustGenerateClinitCheck()) { 5128 GenerateClassInitializationCheck(slow_path, out); 5129 } else { 5130 __ Bind(slow_path->GetExitLabel()); 5131 } 5132 } 5133 } 5134 } 5135 5136 void LocationsBuilderARM::VisitClinitCheck(HClinitCheck* check) { 5137 LocationSummary* locations = 5138 new (GetGraph()->GetArena()) LocationSummary(check, LocationSummary::kCallOnSlowPath); 5139 locations->SetInAt(0, Location::RequiresRegister()); 5140 if (check->HasUses()) { 5141 locations->SetOut(Location::SameAsFirstInput()); 5142 } 5143 } 5144 5145 void InstructionCodeGeneratorARM::VisitClinitCheck(HClinitCheck* check) { 5146 // We assume the class is not null. 5147 SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathARM( 5148 check->GetLoadClass(), check, check->GetDexPc(), true); 5149 codegen_->AddSlowPath(slow_path); 5150 GenerateClassInitializationCheck(slow_path, 5151 check->GetLocations()->InAt(0).AsRegister<Register>()); 5152 } 5153 5154 void InstructionCodeGeneratorARM::GenerateClassInitializationCheck( 5155 SlowPathCode* slow_path, Register class_reg) { 5156 __ LoadFromOffset(kLoadWord, IP, class_reg, mirror::Class::StatusOffset().Int32Value()); 5157 __ cmp(IP, ShifterOperand(mirror::Class::kStatusInitialized)); 5158 __ b(slow_path->GetEntryLabel(), LT); 5159 // Even if the initialized flag is set, we may be in a situation where caches are not synced 5160 // properly. Therefore, we do a memory fence. 5161 __ dmb(ISH); 5162 __ Bind(slow_path->GetExitLabel()); 5163 } 5164 5165 HLoadString::LoadKind CodeGeneratorARM::GetSupportedLoadStringKind( 5166 HLoadString::LoadKind desired_string_load_kind) { 5167 if (kEmitCompilerReadBarrier) { 5168 switch (desired_string_load_kind) { 5169 case HLoadString::LoadKind::kBootImageLinkTimeAddress: 5170 case HLoadString::LoadKind::kBootImageLinkTimePcRelative: 5171 case HLoadString::LoadKind::kBootImageAddress: 5172 // TODO: Implement for read barrier. 5173 return HLoadString::LoadKind::kDexCacheViaMethod; 5174 default: 5175 break; 5176 } 5177 } 5178 switch (desired_string_load_kind) { 5179 case HLoadString::LoadKind::kBootImageLinkTimeAddress: 5180 DCHECK(!GetCompilerOptions().GetCompilePic()); 5181 break; 5182 case HLoadString::LoadKind::kBootImageLinkTimePcRelative: 5183 DCHECK(GetCompilerOptions().GetCompilePic()); 5184 break; 5185 case HLoadString::LoadKind::kBootImageAddress: 5186 break; 5187 case HLoadString::LoadKind::kDexCacheAddress: 5188 DCHECK(Runtime::Current()->UseJitCompilation()); 5189 break; 5190 case HLoadString::LoadKind::kDexCachePcRelative: 5191 DCHECK(!Runtime::Current()->UseJitCompilation()); 5192 // We disable pc-relative load when there is an irreducible loop, as the optimization 5193 // is incompatible with it. 5194 // TODO: Create as many ArmDexCacheArraysBase instructions as needed for methods 5195 // with irreducible loops. 5196 if (GetGraph()->HasIrreducibleLoops()) { 5197 return HLoadString::LoadKind::kDexCacheViaMethod; 5198 } 5199 break; 5200 case HLoadString::LoadKind::kDexCacheViaMethod: 5201 break; 5202 } 5203 return desired_string_load_kind; 5204 } 5205 5206 void LocationsBuilderARM::VisitLoadString(HLoadString* load) { 5207 LocationSummary::CallKind call_kind = (load->NeedsEnvironment() || kEmitCompilerReadBarrier) 5208 ? LocationSummary::kCallOnSlowPath 5209 : LocationSummary::kNoCall; 5210 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(load, call_kind); 5211 HLoadString::LoadKind load_kind = load->GetLoadKind(); 5212 if (load_kind == HLoadString::LoadKind::kDexCacheViaMethod || 5213 load_kind == HLoadString::LoadKind::kDexCachePcRelative) { 5214 locations->SetInAt(0, Location::RequiresRegister()); 5215 } 5216 locations->SetOut(Location::RequiresRegister()); 5217 } 5218 5219 void InstructionCodeGeneratorARM::VisitLoadString(HLoadString* load) { 5220 LocationSummary* locations = load->GetLocations(); 5221 Location out_loc = locations->Out(); 5222 Register out = out_loc.AsRegister<Register>(); 5223 5224 switch (load->GetLoadKind()) { 5225 case HLoadString::LoadKind::kBootImageLinkTimeAddress: { 5226 DCHECK(!kEmitCompilerReadBarrier); 5227 __ LoadLiteral(out, codegen_->DeduplicateBootImageStringLiteral(load->GetDexFile(), 5228 load->GetStringIndex())); 5229 return; // No dex cache slow path. 5230 } 5231 case HLoadString::LoadKind::kBootImageLinkTimePcRelative: { 5232 DCHECK(!kEmitCompilerReadBarrier); 5233 CodeGeneratorARM::PcRelativePatchInfo* labels = 5234 codegen_->NewPcRelativeStringPatch(load->GetDexFile(), load->GetStringIndex()); 5235 __ BindTrackedLabel(&labels->movw_label); 5236 __ movw(out, /* placeholder */ 0u); 5237 __ BindTrackedLabel(&labels->movt_label); 5238 __ movt(out, /* placeholder */ 0u); 5239 __ BindTrackedLabel(&labels->add_pc_label); 5240 __ add(out, out, ShifterOperand(PC)); 5241 return; // No dex cache slow path. 5242 } 5243 case HLoadString::LoadKind::kBootImageAddress: { 5244 DCHECK(!kEmitCompilerReadBarrier); 5245 DCHECK_NE(load->GetAddress(), 0u); 5246 uint32_t address = dchecked_integral_cast<uint32_t>(load->GetAddress()); 5247 __ LoadLiteral(out, codegen_->DeduplicateBootImageAddressLiteral(address)); 5248 return; // No dex cache slow path. 5249 } 5250 case HLoadString::LoadKind::kDexCacheAddress: { 5251 DCHECK_NE(load->GetAddress(), 0u); 5252 uint32_t address = dchecked_integral_cast<uint32_t>(load->GetAddress()); 5253 // 16-bit LDR immediate has a 5-bit offset multiplied by the size and that gives 5254 // a 128B range. To try and reduce the number of literals if we load multiple strings, 5255 // simply split the dex cache address to a 128B aligned base loaded from a literal 5256 // and the remaining offset embedded in the load. 5257 static_assert(sizeof(GcRoot<mirror::String>) == 4u, "Expected GC root to be 4 bytes."); 5258 DCHECK_ALIGNED(load->GetAddress(), 4u); 5259 constexpr size_t offset_bits = /* encoded bits */ 5 + /* scale */ 2; 5260 uint32_t base_address = address & ~MaxInt<uint32_t>(offset_bits); 5261 uint32_t offset = address & MaxInt<uint32_t>(offset_bits); 5262 __ LoadLiteral(out, codegen_->DeduplicateDexCacheAddressLiteral(base_address)); 5263 GenerateGcRootFieldLoad(load, out_loc, out, offset); 5264 break; 5265 } 5266 case HLoadString::LoadKind::kDexCachePcRelative: { 5267 Register base_reg = locations->InAt(0).AsRegister<Register>(); 5268 HArmDexCacheArraysBase* base = load->InputAt(0)->AsArmDexCacheArraysBase(); 5269 int32_t offset = load->GetDexCacheElementOffset() - base->GetElementOffset(); 5270 GenerateGcRootFieldLoad(load, out_loc, base_reg, offset); 5271 break; 5272 } 5273 case HLoadString::LoadKind::kDexCacheViaMethod: { 5274 Register current_method = locations->InAt(0).AsRegister<Register>(); 5275 5276 // /* GcRoot<mirror::Class> */ out = current_method->declaring_class_ 5277 GenerateGcRootFieldLoad( 5278 load, out_loc, current_method, ArtMethod::DeclaringClassOffset().Int32Value()); 5279 // /* GcRoot<mirror::String>[] */ out = out->dex_cache_strings_ 5280 __ LoadFromOffset(kLoadWord, out, out, mirror::Class::DexCacheStringsOffset().Int32Value()); 5281 // /* GcRoot<mirror::String> */ out = out[string_index] 5282 GenerateGcRootFieldLoad( 5283 load, out_loc, out, CodeGenerator::GetCacheOffset(load->GetStringIndex())); 5284 break; 5285 } 5286 default: 5287 LOG(FATAL) << "Unexpected load kind: " << load->GetLoadKind(); 5288 UNREACHABLE(); 5289 } 5290 5291 if (!load->IsInDexCache()) { 5292 SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadStringSlowPathARM(load); 5293 codegen_->AddSlowPath(slow_path); 5294 __ CompareAndBranchIfZero(out, slow_path->GetEntryLabel()); 5295 __ Bind(slow_path->GetExitLabel()); 5296 } 5297 } 5298 5299 static int32_t GetExceptionTlsOffset() { 5300 return Thread::ExceptionOffset<kArmWordSize>().Int32Value(); 5301 } 5302 5303 void LocationsBuilderARM::VisitLoadException(HLoadException* load) { 5304 LocationSummary* locations = 5305 new (GetGraph()->GetArena()) LocationSummary(load, LocationSummary::kNoCall); 5306 locations->SetOut(Location::RequiresRegister()); 5307 } 5308 5309 void InstructionCodeGeneratorARM::VisitLoadException(HLoadException* load) { 5310 Register out = load->GetLocations()->Out().AsRegister<Register>(); 5311 __ LoadFromOffset(kLoadWord, out, TR, GetExceptionTlsOffset()); 5312 } 5313 5314 void LocationsBuilderARM::VisitClearException(HClearException* clear) { 5315 new (GetGraph()->GetArena()) LocationSummary(clear, LocationSummary::kNoCall); 5316 } 5317 5318 void InstructionCodeGeneratorARM::VisitClearException(HClearException* clear ATTRIBUTE_UNUSED) { 5319 __ LoadImmediate(IP, 0); 5320 __ StoreToOffset(kStoreWord, IP, TR, GetExceptionTlsOffset()); 5321 } 5322 5323 void LocationsBuilderARM::VisitThrow(HThrow* instruction) { 5324 LocationSummary* locations = 5325 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall); 5326 InvokeRuntimeCallingConvention calling_convention; 5327 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 5328 } 5329 5330 void InstructionCodeGeneratorARM::VisitThrow(HThrow* instruction) { 5331 codegen_->InvokeRuntime( 5332 QUICK_ENTRY_POINT(pDeliverException), instruction, instruction->GetDexPc(), nullptr); 5333 CheckEntrypointTypes<kQuickDeliverException, void, mirror::Object*>(); 5334 } 5335 5336 static bool TypeCheckNeedsATemporary(TypeCheckKind type_check_kind) { 5337 return kEmitCompilerReadBarrier && 5338 (kUseBakerReadBarrier || 5339 type_check_kind == TypeCheckKind::kAbstractClassCheck || 5340 type_check_kind == TypeCheckKind::kClassHierarchyCheck || 5341 type_check_kind == TypeCheckKind::kArrayObjectCheck); 5342 } 5343 5344 void LocationsBuilderARM::VisitInstanceOf(HInstanceOf* instruction) { 5345 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 5346 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 5347 switch (type_check_kind) { 5348 case TypeCheckKind::kExactCheck: 5349 case TypeCheckKind::kAbstractClassCheck: 5350 case TypeCheckKind::kClassHierarchyCheck: 5351 case TypeCheckKind::kArrayObjectCheck: 5352 call_kind = 5353 kEmitCompilerReadBarrier ? LocationSummary::kCallOnSlowPath : LocationSummary::kNoCall; 5354 break; 5355 case TypeCheckKind::kArrayCheck: 5356 case TypeCheckKind::kUnresolvedCheck: 5357 case TypeCheckKind::kInterfaceCheck: 5358 call_kind = LocationSummary::kCallOnSlowPath; 5359 break; 5360 } 5361 5362 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 5363 locations->SetInAt(0, Location::RequiresRegister()); 5364 locations->SetInAt(1, Location::RequiresRegister()); 5365 // The "out" register is used as a temporary, so it overlaps with the inputs. 5366 // Note that TypeCheckSlowPathARM uses this register too. 5367 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 5368 // When read barriers are enabled, we need a temporary register for 5369 // some cases. 5370 if (TypeCheckNeedsATemporary(type_check_kind)) { 5371 locations->AddTemp(Location::RequiresRegister()); 5372 } 5373 } 5374 5375 void InstructionCodeGeneratorARM::VisitInstanceOf(HInstanceOf* instruction) { 5376 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 5377 LocationSummary* locations = instruction->GetLocations(); 5378 Location obj_loc = locations->InAt(0); 5379 Register obj = obj_loc.AsRegister<Register>(); 5380 Register cls = locations->InAt(1).AsRegister<Register>(); 5381 Location out_loc = locations->Out(); 5382 Register out = out_loc.AsRegister<Register>(); 5383 Location maybe_temp_loc = TypeCheckNeedsATemporary(type_check_kind) ? 5384 locations->GetTemp(0) : 5385 Location::NoLocation(); 5386 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 5387 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 5388 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 5389 uint32_t primitive_offset = mirror::Class::PrimitiveTypeOffset().Int32Value(); 5390 Label done, zero; 5391 SlowPathCode* slow_path = nullptr; 5392 5393 // Return 0 if `obj` is null. 5394 // avoid null check if we know obj is not null. 5395 if (instruction->MustDoNullCheck()) { 5396 __ CompareAndBranchIfZero(obj, &zero); 5397 } 5398 5399 // /* HeapReference<Class> */ out = obj->klass_ 5400 GenerateReferenceLoadTwoRegisters(instruction, out_loc, obj_loc, class_offset, maybe_temp_loc); 5401 5402 switch (type_check_kind) { 5403 case TypeCheckKind::kExactCheck: { 5404 __ cmp(out, ShifterOperand(cls)); 5405 // Classes must be equal for the instanceof to succeed. 5406 __ b(&zero, NE); 5407 __ LoadImmediate(out, 1); 5408 __ b(&done); 5409 break; 5410 } 5411 5412 case TypeCheckKind::kAbstractClassCheck: { 5413 // If the class is abstract, we eagerly fetch the super class of the 5414 // object to avoid doing a comparison we know will fail. 5415 Label loop; 5416 __ Bind(&loop); 5417 // /* HeapReference<Class> */ out = out->super_class_ 5418 GenerateReferenceLoadOneRegister(instruction, out_loc, super_offset, maybe_temp_loc); 5419 // If `out` is null, we use it for the result, and jump to `done`. 5420 __ CompareAndBranchIfZero(out, &done); 5421 __ cmp(out, ShifterOperand(cls)); 5422 __ b(&loop, NE); 5423 __ LoadImmediate(out, 1); 5424 if (zero.IsLinked()) { 5425 __ b(&done); 5426 } 5427 break; 5428 } 5429 5430 case TypeCheckKind::kClassHierarchyCheck: { 5431 // Walk over the class hierarchy to find a match. 5432 Label loop, success; 5433 __ Bind(&loop); 5434 __ cmp(out, ShifterOperand(cls)); 5435 __ b(&success, EQ); 5436 // /* HeapReference<Class> */ out = out->super_class_ 5437 GenerateReferenceLoadOneRegister(instruction, out_loc, super_offset, maybe_temp_loc); 5438 __ CompareAndBranchIfNonZero(out, &loop); 5439 // If `out` is null, we use it for the result, and jump to `done`. 5440 __ b(&done); 5441 __ Bind(&success); 5442 __ LoadImmediate(out, 1); 5443 if (zero.IsLinked()) { 5444 __ b(&done); 5445 } 5446 break; 5447 } 5448 5449 case TypeCheckKind::kArrayObjectCheck: { 5450 // Do an exact check. 5451 Label exact_check; 5452 __ cmp(out, ShifterOperand(cls)); 5453 __ b(&exact_check, EQ); 5454 // Otherwise, we need to check that the object's class is a non-primitive array. 5455 // /* HeapReference<Class> */ out = out->component_type_ 5456 GenerateReferenceLoadOneRegister(instruction, out_loc, component_offset, maybe_temp_loc); 5457 // If `out` is null, we use it for the result, and jump to `done`. 5458 __ CompareAndBranchIfZero(out, &done); 5459 __ LoadFromOffset(kLoadUnsignedHalfword, out, out, primitive_offset); 5460 static_assert(Primitive::kPrimNot == 0, "Expected 0 for kPrimNot"); 5461 __ CompareAndBranchIfNonZero(out, &zero); 5462 __ Bind(&exact_check); 5463 __ LoadImmediate(out, 1); 5464 __ b(&done); 5465 break; 5466 } 5467 5468 case TypeCheckKind::kArrayCheck: { 5469 __ cmp(out, ShifterOperand(cls)); 5470 DCHECK(locations->OnlyCallsOnSlowPath()); 5471 slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathARM(instruction, 5472 /* is_fatal */ false); 5473 codegen_->AddSlowPath(slow_path); 5474 __ b(slow_path->GetEntryLabel(), NE); 5475 __ LoadImmediate(out, 1); 5476 if (zero.IsLinked()) { 5477 __ b(&done); 5478 } 5479 break; 5480 } 5481 5482 case TypeCheckKind::kUnresolvedCheck: 5483 case TypeCheckKind::kInterfaceCheck: { 5484 // Note that we indeed only call on slow path, but we always go 5485 // into the slow path for the unresolved and interface check 5486 // cases. 5487 // 5488 // We cannot directly call the InstanceofNonTrivial runtime 5489 // entry point without resorting to a type checking slow path 5490 // here (i.e. by calling InvokeRuntime directly), as it would 5491 // require to assign fixed registers for the inputs of this 5492 // HInstanceOf instruction (following the runtime calling 5493 // convention), which might be cluttered by the potential first 5494 // read barrier emission at the beginning of this method. 5495 // 5496 // TODO: Introduce a new runtime entry point taking the object 5497 // to test (instead of its class) as argument, and let it deal 5498 // with the read barrier issues. This will let us refactor this 5499 // case of the `switch` code as it was previously (with a direct 5500 // call to the runtime not using a type checking slow path). 5501 // This should also be beneficial for the other cases above. 5502 DCHECK(locations->OnlyCallsOnSlowPath()); 5503 slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathARM(instruction, 5504 /* is_fatal */ false); 5505 codegen_->AddSlowPath(slow_path); 5506 __ b(slow_path->GetEntryLabel()); 5507 if (zero.IsLinked()) { 5508 __ b(&done); 5509 } 5510 break; 5511 } 5512 } 5513 5514 if (zero.IsLinked()) { 5515 __ Bind(&zero); 5516 __ LoadImmediate(out, 0); 5517 } 5518 5519 if (done.IsLinked()) { 5520 __ Bind(&done); 5521 } 5522 5523 if (slow_path != nullptr) { 5524 __ Bind(slow_path->GetExitLabel()); 5525 } 5526 } 5527 5528 void LocationsBuilderARM::VisitCheckCast(HCheckCast* instruction) { 5529 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 5530 bool throws_into_catch = instruction->CanThrowIntoCatchBlock(); 5531 5532 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 5533 switch (type_check_kind) { 5534 case TypeCheckKind::kExactCheck: 5535 case TypeCheckKind::kAbstractClassCheck: 5536 case TypeCheckKind::kClassHierarchyCheck: 5537 case TypeCheckKind::kArrayObjectCheck: 5538 call_kind = (throws_into_catch || kEmitCompilerReadBarrier) ? 5539 LocationSummary::kCallOnSlowPath : 5540 LocationSummary::kNoCall; // In fact, call on a fatal (non-returning) slow path. 5541 break; 5542 case TypeCheckKind::kArrayCheck: 5543 case TypeCheckKind::kUnresolvedCheck: 5544 case TypeCheckKind::kInterfaceCheck: 5545 call_kind = LocationSummary::kCallOnSlowPath; 5546 break; 5547 } 5548 5549 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 5550 locations->SetInAt(0, Location::RequiresRegister()); 5551 locations->SetInAt(1, Location::RequiresRegister()); 5552 // Note that TypeCheckSlowPathARM uses this "temp" register too. 5553 locations->AddTemp(Location::RequiresRegister()); 5554 // When read barriers are enabled, we need an additional temporary 5555 // register for some cases. 5556 if (TypeCheckNeedsATemporary(type_check_kind)) { 5557 locations->AddTemp(Location::RequiresRegister()); 5558 } 5559 } 5560 5561 void InstructionCodeGeneratorARM::VisitCheckCast(HCheckCast* instruction) { 5562 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 5563 LocationSummary* locations = instruction->GetLocations(); 5564 Location obj_loc = locations->InAt(0); 5565 Register obj = obj_loc.AsRegister<Register>(); 5566 Register cls = locations->InAt(1).AsRegister<Register>(); 5567 Location temp_loc = locations->GetTemp(0); 5568 Register temp = temp_loc.AsRegister<Register>(); 5569 Location maybe_temp2_loc = TypeCheckNeedsATemporary(type_check_kind) ? 5570 locations->GetTemp(1) : 5571 Location::NoLocation(); 5572 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 5573 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 5574 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 5575 uint32_t primitive_offset = mirror::Class::PrimitiveTypeOffset().Int32Value(); 5576 5577 bool is_type_check_slow_path_fatal = 5578 (type_check_kind == TypeCheckKind::kExactCheck || 5579 type_check_kind == TypeCheckKind::kAbstractClassCheck || 5580 type_check_kind == TypeCheckKind::kClassHierarchyCheck || 5581 type_check_kind == TypeCheckKind::kArrayObjectCheck) && 5582 !instruction->CanThrowIntoCatchBlock(); 5583 SlowPathCode* type_check_slow_path = 5584 new (GetGraph()->GetArena()) TypeCheckSlowPathARM(instruction, 5585 is_type_check_slow_path_fatal); 5586 codegen_->AddSlowPath(type_check_slow_path); 5587 5588 Label done; 5589 // Avoid null check if we know obj is not null. 5590 if (instruction->MustDoNullCheck()) { 5591 __ CompareAndBranchIfZero(obj, &done); 5592 } 5593 5594 // /* HeapReference<Class> */ temp = obj->klass_ 5595 GenerateReferenceLoadTwoRegisters(instruction, temp_loc, obj_loc, class_offset, maybe_temp2_loc); 5596 5597 switch (type_check_kind) { 5598 case TypeCheckKind::kExactCheck: 5599 case TypeCheckKind::kArrayCheck: { 5600 __ cmp(temp, ShifterOperand(cls)); 5601 // Jump to slow path for throwing the exception or doing a 5602 // more involved array check. 5603 __ b(type_check_slow_path->GetEntryLabel(), NE); 5604 break; 5605 } 5606 5607 case TypeCheckKind::kAbstractClassCheck: { 5608 // If the class is abstract, we eagerly fetch the super class of the 5609 // object to avoid doing a comparison we know will fail. 5610 Label loop, compare_classes; 5611 __ Bind(&loop); 5612 // /* HeapReference<Class> */ temp = temp->super_class_ 5613 GenerateReferenceLoadOneRegister(instruction, temp_loc, super_offset, maybe_temp2_loc); 5614 5615 // If the class reference currently in `temp` is not null, jump 5616 // to the `compare_classes` label to compare it with the checked 5617 // class. 5618 __ CompareAndBranchIfNonZero(temp, &compare_classes); 5619 // Otherwise, jump to the slow path to throw the exception. 5620 // 5621 // But before, move back the object's class into `temp` before 5622 // going into the slow path, as it has been overwritten in the 5623 // meantime. 5624 // /* HeapReference<Class> */ temp = obj->klass_ 5625 GenerateReferenceLoadTwoRegisters( 5626 instruction, temp_loc, obj_loc, class_offset, maybe_temp2_loc); 5627 __ b(type_check_slow_path->GetEntryLabel()); 5628 5629 __ Bind(&compare_classes); 5630 __ cmp(temp, ShifterOperand(cls)); 5631 __ b(&loop, NE); 5632 break; 5633 } 5634 5635 case TypeCheckKind::kClassHierarchyCheck: { 5636 // Walk over the class hierarchy to find a match. 5637 Label loop; 5638 __ Bind(&loop); 5639 __ cmp(temp, ShifterOperand(cls)); 5640 __ b(&done, EQ); 5641 5642 // /* HeapReference<Class> */ temp = temp->super_class_ 5643 GenerateReferenceLoadOneRegister(instruction, temp_loc, super_offset, maybe_temp2_loc); 5644 5645 // If the class reference currently in `temp` is not null, jump 5646 // back at the beginning of the loop. 5647 __ CompareAndBranchIfNonZero(temp, &loop); 5648 // Otherwise, jump to the slow path to throw the exception. 5649 // 5650 // But before, move back the object's class into `temp` before 5651 // going into the slow path, as it has been overwritten in the 5652 // meantime. 5653 // /* HeapReference<Class> */ temp = obj->klass_ 5654 GenerateReferenceLoadTwoRegisters( 5655 instruction, temp_loc, obj_loc, class_offset, maybe_temp2_loc); 5656 __ b(type_check_slow_path->GetEntryLabel()); 5657 break; 5658 } 5659 5660 case TypeCheckKind::kArrayObjectCheck: { 5661 // Do an exact check. 5662 Label check_non_primitive_component_type; 5663 __ cmp(temp, ShifterOperand(cls)); 5664 __ b(&done, EQ); 5665 5666 // Otherwise, we need to check that the object's class is a non-primitive array. 5667 // /* HeapReference<Class> */ temp = temp->component_type_ 5668 GenerateReferenceLoadOneRegister(instruction, temp_loc, component_offset, maybe_temp2_loc); 5669 5670 // If the component type is not null (i.e. the object is indeed 5671 // an array), jump to label `check_non_primitive_component_type` 5672 // to further check that this component type is not a primitive 5673 // type. 5674 __ CompareAndBranchIfNonZero(temp, &check_non_primitive_component_type); 5675 // Otherwise, jump to the slow path to throw the exception. 5676 // 5677 // But before, move back the object's class into `temp` before 5678 // going into the slow path, as it has been overwritten in the 5679 // meantime. 5680 // /* HeapReference<Class> */ temp = obj->klass_ 5681 GenerateReferenceLoadTwoRegisters( 5682 instruction, temp_loc, obj_loc, class_offset, maybe_temp2_loc); 5683 __ b(type_check_slow_path->GetEntryLabel()); 5684 5685 __ Bind(&check_non_primitive_component_type); 5686 __ LoadFromOffset(kLoadUnsignedHalfword, temp, temp, primitive_offset); 5687 static_assert(Primitive::kPrimNot == 0, "Expected 0 for art::Primitive::kPrimNot"); 5688 __ CompareAndBranchIfZero(temp, &done); 5689 // Same comment as above regarding `temp` and the slow path. 5690 // /* HeapReference<Class> */ temp = obj->klass_ 5691 GenerateReferenceLoadTwoRegisters( 5692 instruction, temp_loc, obj_loc, class_offset, maybe_temp2_loc); 5693 __ b(type_check_slow_path->GetEntryLabel()); 5694 break; 5695 } 5696 5697 case TypeCheckKind::kUnresolvedCheck: 5698 case TypeCheckKind::kInterfaceCheck: 5699 // We always go into the type check slow path for the unresolved 5700 // and interface check cases. 5701 // 5702 // We cannot directly call the CheckCast runtime entry point 5703 // without resorting to a type checking slow path here (i.e. by 5704 // calling InvokeRuntime directly), as it would require to 5705 // assign fixed registers for the inputs of this HInstanceOf 5706 // instruction (following the runtime calling convention), which 5707 // might be cluttered by the potential first read barrier 5708 // emission at the beginning of this method. 5709 // 5710 // TODO: Introduce a new runtime entry point taking the object 5711 // to test (instead of its class) as argument, and let it deal 5712 // with the read barrier issues. This will let us refactor this 5713 // case of the `switch` code as it was previously (with a direct 5714 // call to the runtime not using a type checking slow path). 5715 // This should also be beneficial for the other cases above. 5716 __ b(type_check_slow_path->GetEntryLabel()); 5717 break; 5718 } 5719 __ Bind(&done); 5720 5721 __ Bind(type_check_slow_path->GetExitLabel()); 5722 } 5723 5724 void LocationsBuilderARM::VisitMonitorOperation(HMonitorOperation* instruction) { 5725 LocationSummary* locations = 5726 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall); 5727 InvokeRuntimeCallingConvention calling_convention; 5728 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 5729 } 5730 5731 void InstructionCodeGeneratorARM::VisitMonitorOperation(HMonitorOperation* instruction) { 5732 codegen_->InvokeRuntime(instruction->IsEnter() 5733 ? QUICK_ENTRY_POINT(pLockObject) : QUICK_ENTRY_POINT(pUnlockObject), 5734 instruction, 5735 instruction->GetDexPc(), 5736 nullptr); 5737 if (instruction->IsEnter()) { 5738 CheckEntrypointTypes<kQuickLockObject, void, mirror::Object*>(); 5739 } else { 5740 CheckEntrypointTypes<kQuickUnlockObject, void, mirror::Object*>(); 5741 } 5742 } 5743 5744 void LocationsBuilderARM::VisitAnd(HAnd* instruction) { HandleBitwiseOperation(instruction, AND); } 5745 void LocationsBuilderARM::VisitOr(HOr* instruction) { HandleBitwiseOperation(instruction, ORR); } 5746 void LocationsBuilderARM::VisitXor(HXor* instruction) { HandleBitwiseOperation(instruction, EOR); } 5747 5748 void LocationsBuilderARM::HandleBitwiseOperation(HBinaryOperation* instruction, Opcode opcode) { 5749 LocationSummary* locations = 5750 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 5751 DCHECK(instruction->GetResultType() == Primitive::kPrimInt 5752 || instruction->GetResultType() == Primitive::kPrimLong); 5753 // Note: GVN reorders commutative operations to have the constant on the right hand side. 5754 locations->SetInAt(0, Location::RequiresRegister()); 5755 locations->SetInAt(1, ArmEncodableConstantOrRegister(instruction->InputAt(1), opcode)); 5756 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 5757 } 5758 5759 void InstructionCodeGeneratorARM::VisitAnd(HAnd* instruction) { 5760 HandleBitwiseOperation(instruction); 5761 } 5762 5763 void InstructionCodeGeneratorARM::VisitOr(HOr* instruction) { 5764 HandleBitwiseOperation(instruction); 5765 } 5766 5767 void InstructionCodeGeneratorARM::VisitXor(HXor* instruction) { 5768 HandleBitwiseOperation(instruction); 5769 } 5770 5771 5772 void LocationsBuilderARM::VisitBitwiseNegatedRight(HBitwiseNegatedRight* instruction) { 5773 LocationSummary* locations = 5774 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 5775 DCHECK(instruction->GetResultType() == Primitive::kPrimInt 5776 || instruction->GetResultType() == Primitive::kPrimLong); 5777 5778 locations->SetInAt(0, Location::RequiresRegister()); 5779 locations->SetInAt(1, Location::RequiresRegister()); 5780 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 5781 } 5782 5783 void InstructionCodeGeneratorARM::VisitBitwiseNegatedRight(HBitwiseNegatedRight* instruction) { 5784 LocationSummary* locations = instruction->GetLocations(); 5785 Location first = locations->InAt(0); 5786 Location second = locations->InAt(1); 5787 Location out = locations->Out(); 5788 5789 if (instruction->GetResultType() == Primitive::kPrimInt) { 5790 Register first_reg = first.AsRegister<Register>(); 5791 ShifterOperand second_reg(second.AsRegister<Register>()); 5792 Register out_reg = out.AsRegister<Register>(); 5793 5794 switch (instruction->GetOpKind()) { 5795 case HInstruction::kAnd: 5796 __ bic(out_reg, first_reg, second_reg); 5797 break; 5798 case HInstruction::kOr: 5799 __ orn(out_reg, first_reg, second_reg); 5800 break; 5801 // There is no EON on arm. 5802 case HInstruction::kXor: 5803 default: 5804 LOG(FATAL) << "Unexpected instruction " << instruction->DebugName(); 5805 UNREACHABLE(); 5806 } 5807 return; 5808 5809 } else { 5810 DCHECK_EQ(instruction->GetResultType(), Primitive::kPrimLong); 5811 Register first_low = first.AsRegisterPairLow<Register>(); 5812 Register first_high = first.AsRegisterPairHigh<Register>(); 5813 ShifterOperand second_low(second.AsRegisterPairLow<Register>()); 5814 ShifterOperand second_high(second.AsRegisterPairHigh<Register>()); 5815 Register out_low = out.AsRegisterPairLow<Register>(); 5816 Register out_high = out.AsRegisterPairHigh<Register>(); 5817 5818 switch (instruction->GetOpKind()) { 5819 case HInstruction::kAnd: 5820 __ bic(out_low, first_low, second_low); 5821 __ bic(out_high, first_high, second_high); 5822 break; 5823 case HInstruction::kOr: 5824 __ orn(out_low, first_low, second_low); 5825 __ orn(out_high, first_high, second_high); 5826 break; 5827 // There is no EON on arm. 5828 case HInstruction::kXor: 5829 default: 5830 LOG(FATAL) << "Unexpected instruction " << instruction->DebugName(); 5831 UNREACHABLE(); 5832 } 5833 } 5834 } 5835 5836 void InstructionCodeGeneratorARM::GenerateAndConst(Register out, Register first, uint32_t value) { 5837 // Optimize special cases for individual halfs of `and-long` (`and` is simplified earlier). 5838 if (value == 0xffffffffu) { 5839 if (out != first) { 5840 __ mov(out, ShifterOperand(first)); 5841 } 5842 return; 5843 } 5844 if (value == 0u) { 5845 __ mov(out, ShifterOperand(0)); 5846 return; 5847 } 5848 ShifterOperand so; 5849 if (__ ShifterOperandCanHold(kNoRegister, kNoRegister, AND, value, &so)) { 5850 __ and_(out, first, so); 5851 } else { 5852 DCHECK(__ ShifterOperandCanHold(kNoRegister, kNoRegister, BIC, ~value, &so)); 5853 __ bic(out, first, ShifterOperand(~value)); 5854 } 5855 } 5856 5857 void InstructionCodeGeneratorARM::GenerateOrrConst(Register out, Register first, uint32_t value) { 5858 // Optimize special cases for individual halfs of `or-long` (`or` is simplified earlier). 5859 if (value == 0u) { 5860 if (out != first) { 5861 __ mov(out, ShifterOperand(first)); 5862 } 5863 return; 5864 } 5865 if (value == 0xffffffffu) { 5866 __ mvn(out, ShifterOperand(0)); 5867 return; 5868 } 5869 ShifterOperand so; 5870 if (__ ShifterOperandCanHold(kNoRegister, kNoRegister, ORR, value, &so)) { 5871 __ orr(out, first, so); 5872 } else { 5873 DCHECK(__ ShifterOperandCanHold(kNoRegister, kNoRegister, ORN, ~value, &so)); 5874 __ orn(out, first, ShifterOperand(~value)); 5875 } 5876 } 5877 5878 void InstructionCodeGeneratorARM::GenerateEorConst(Register out, Register first, uint32_t value) { 5879 // Optimize special case for individual halfs of `xor-long` (`xor` is simplified earlier). 5880 if (value == 0u) { 5881 if (out != first) { 5882 __ mov(out, ShifterOperand(first)); 5883 } 5884 return; 5885 } 5886 __ eor(out, first, ShifterOperand(value)); 5887 } 5888 5889 void InstructionCodeGeneratorARM::HandleBitwiseOperation(HBinaryOperation* instruction) { 5890 LocationSummary* locations = instruction->GetLocations(); 5891 Location first = locations->InAt(0); 5892 Location second = locations->InAt(1); 5893 Location out = locations->Out(); 5894 5895 if (second.IsConstant()) { 5896 uint64_t value = static_cast<uint64_t>(Int64FromConstant(second.GetConstant())); 5897 uint32_t value_low = Low32Bits(value); 5898 if (instruction->GetResultType() == Primitive::kPrimInt) { 5899 Register first_reg = first.AsRegister<Register>(); 5900 Register out_reg = out.AsRegister<Register>(); 5901 if (instruction->IsAnd()) { 5902 GenerateAndConst(out_reg, first_reg, value_low); 5903 } else if (instruction->IsOr()) { 5904 GenerateOrrConst(out_reg, first_reg, value_low); 5905 } else { 5906 DCHECK(instruction->IsXor()); 5907 GenerateEorConst(out_reg, first_reg, value_low); 5908 } 5909 } else { 5910 DCHECK_EQ(instruction->GetResultType(), Primitive::kPrimLong); 5911 uint32_t value_high = High32Bits(value); 5912 Register first_low = first.AsRegisterPairLow<Register>(); 5913 Register first_high = first.AsRegisterPairHigh<Register>(); 5914 Register out_low = out.AsRegisterPairLow<Register>(); 5915 Register out_high = out.AsRegisterPairHigh<Register>(); 5916 if (instruction->IsAnd()) { 5917 GenerateAndConst(out_low, first_low, value_low); 5918 GenerateAndConst(out_high, first_high, value_high); 5919 } else if (instruction->IsOr()) { 5920 GenerateOrrConst(out_low, first_low, value_low); 5921 GenerateOrrConst(out_high, first_high, value_high); 5922 } else { 5923 DCHECK(instruction->IsXor()); 5924 GenerateEorConst(out_low, first_low, value_low); 5925 GenerateEorConst(out_high, first_high, value_high); 5926 } 5927 } 5928 return; 5929 } 5930 5931 if (instruction->GetResultType() == Primitive::kPrimInt) { 5932 Register first_reg = first.AsRegister<Register>(); 5933 ShifterOperand second_reg(second.AsRegister<Register>()); 5934 Register out_reg = out.AsRegister<Register>(); 5935 if (instruction->IsAnd()) { 5936 __ and_(out_reg, first_reg, second_reg); 5937 } else if (instruction->IsOr()) { 5938 __ orr(out_reg, first_reg, second_reg); 5939 } else { 5940 DCHECK(instruction->IsXor()); 5941 __ eor(out_reg, first_reg, second_reg); 5942 } 5943 } else { 5944 DCHECK_EQ(instruction->GetResultType(), Primitive::kPrimLong); 5945 Register first_low = first.AsRegisterPairLow<Register>(); 5946 Register first_high = first.AsRegisterPairHigh<Register>(); 5947 ShifterOperand second_low(second.AsRegisterPairLow<Register>()); 5948 ShifterOperand second_high(second.AsRegisterPairHigh<Register>()); 5949 Register out_low = out.AsRegisterPairLow<Register>(); 5950 Register out_high = out.AsRegisterPairHigh<Register>(); 5951 if (instruction->IsAnd()) { 5952 __ and_(out_low, first_low, second_low); 5953 __ and_(out_high, first_high, second_high); 5954 } else if (instruction->IsOr()) { 5955 __ orr(out_low, first_low, second_low); 5956 __ orr(out_high, first_high, second_high); 5957 } else { 5958 DCHECK(instruction->IsXor()); 5959 __ eor(out_low, first_low, second_low); 5960 __ eor(out_high, first_high, second_high); 5961 } 5962 } 5963 } 5964 5965 void InstructionCodeGeneratorARM::GenerateReferenceLoadOneRegister(HInstruction* instruction, 5966 Location out, 5967 uint32_t offset, 5968 Location maybe_temp) { 5969 Register out_reg = out.AsRegister<Register>(); 5970 if (kEmitCompilerReadBarrier) { 5971 DCHECK(maybe_temp.IsRegister()) << maybe_temp; 5972 if (kUseBakerReadBarrier) { 5973 // Load with fast path based Baker's read barrier. 5974 // /* HeapReference<Object> */ out = *(out + offset) 5975 codegen_->GenerateFieldLoadWithBakerReadBarrier( 5976 instruction, out, out_reg, offset, maybe_temp, /* needs_null_check */ false); 5977 } else { 5978 // Load with slow path based read barrier. 5979 // Save the value of `out` into `maybe_temp` before overwriting it 5980 // in the following move operation, as we will need it for the 5981 // read barrier below. 5982 __ Mov(maybe_temp.AsRegister<Register>(), out_reg); 5983 // /* HeapReference<Object> */ out = *(out + offset) 5984 __ LoadFromOffset(kLoadWord, out_reg, out_reg, offset); 5985 codegen_->GenerateReadBarrierSlow(instruction, out, out, maybe_temp, offset); 5986 } 5987 } else { 5988 // Plain load with no read barrier. 5989 // /* HeapReference<Object> */ out = *(out + offset) 5990 __ LoadFromOffset(kLoadWord, out_reg, out_reg, offset); 5991 __ MaybeUnpoisonHeapReference(out_reg); 5992 } 5993 } 5994 5995 void InstructionCodeGeneratorARM::GenerateReferenceLoadTwoRegisters(HInstruction* instruction, 5996 Location out, 5997 Location obj, 5998 uint32_t offset, 5999 Location maybe_temp) { 6000 Register out_reg = out.AsRegister<Register>(); 6001 Register obj_reg = obj.AsRegister<Register>(); 6002 if (kEmitCompilerReadBarrier) { 6003 if (kUseBakerReadBarrier) { 6004 DCHECK(maybe_temp.IsRegister()) << maybe_temp; 6005 // Load with fast path based Baker's read barrier. 6006 // /* HeapReference<Object> */ out = *(obj + offset) 6007 codegen_->GenerateFieldLoadWithBakerReadBarrier( 6008 instruction, out, obj_reg, offset, maybe_temp, /* needs_null_check */ false); 6009 } else { 6010 // Load with slow path based read barrier. 6011 // /* HeapReference<Object> */ out = *(obj + offset) 6012 __ LoadFromOffset(kLoadWord, out_reg, obj_reg, offset); 6013 codegen_->GenerateReadBarrierSlow(instruction, out, out, obj, offset); 6014 } 6015 } else { 6016 // Plain load with no read barrier. 6017 // /* HeapReference<Object> */ out = *(obj + offset) 6018 __ LoadFromOffset(kLoadWord, out_reg, obj_reg, offset); 6019 __ MaybeUnpoisonHeapReference(out_reg); 6020 } 6021 } 6022 6023 void InstructionCodeGeneratorARM::GenerateGcRootFieldLoad(HInstruction* instruction, 6024 Location root, 6025 Register obj, 6026 uint32_t offset) { 6027 Register root_reg = root.AsRegister<Register>(); 6028 if (kEmitCompilerReadBarrier) { 6029 if (kUseBakerReadBarrier) { 6030 // Fast path implementation of art::ReadBarrier::BarrierForRoot when 6031 // Baker's read barrier are used: 6032 // 6033 // root = obj.field; 6034 // if (Thread::Current()->GetIsGcMarking()) { 6035 // root = ReadBarrier::Mark(root) 6036 // } 6037 6038 // /* GcRoot<mirror::Object> */ root = *(obj + offset) 6039 __ LoadFromOffset(kLoadWord, root_reg, obj, offset); 6040 static_assert( 6041 sizeof(mirror::CompressedReference<mirror::Object>) == sizeof(GcRoot<mirror::Object>), 6042 "art::mirror::CompressedReference<mirror::Object> and art::GcRoot<mirror::Object> " 6043 "have different sizes."); 6044 static_assert(sizeof(mirror::CompressedReference<mirror::Object>) == sizeof(int32_t), 6045 "art::mirror::CompressedReference<mirror::Object> and int32_t " 6046 "have different sizes."); 6047 6048 // Slow path used to mark the GC root `root`. 6049 SlowPathCode* slow_path = 6050 new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM(instruction, root, root); 6051 codegen_->AddSlowPath(slow_path); 6052 6053 // IP = Thread::Current()->GetIsGcMarking() 6054 __ LoadFromOffset( 6055 kLoadWord, IP, TR, Thread::IsGcMarkingOffset<kArmWordSize>().Int32Value()); 6056 __ CompareAndBranchIfNonZero(IP, slow_path->GetEntryLabel()); 6057 __ Bind(slow_path->GetExitLabel()); 6058 } else { 6059 // GC root loaded through a slow path for read barriers other 6060 // than Baker's. 6061 // /* GcRoot<mirror::Object>* */ root = obj + offset 6062 __ AddConstant(root_reg, obj, offset); 6063 // /* mirror::Object* */ root = root->Read() 6064 codegen_->GenerateReadBarrierForRootSlow(instruction, root, root); 6065 } 6066 } else { 6067 // Plain GC root load with no read barrier. 6068 // /* GcRoot<mirror::Object> */ root = *(obj + offset) 6069 __ LoadFromOffset(kLoadWord, root_reg, obj, offset); 6070 // Note that GC roots are not affected by heap poisoning, thus we 6071 // do not have to unpoison `root_reg` here. 6072 } 6073 } 6074 6075 void CodeGeneratorARM::GenerateFieldLoadWithBakerReadBarrier(HInstruction* instruction, 6076 Location ref, 6077 Register obj, 6078 uint32_t offset, 6079 Location temp, 6080 bool needs_null_check) { 6081 DCHECK(kEmitCompilerReadBarrier); 6082 DCHECK(kUseBakerReadBarrier); 6083 6084 // /* HeapReference<Object> */ ref = *(obj + offset) 6085 Location no_index = Location::NoLocation(); 6086 GenerateReferenceLoadWithBakerReadBarrier( 6087 instruction, ref, obj, offset, no_index, temp, needs_null_check); 6088 } 6089 6090 void CodeGeneratorARM::GenerateArrayLoadWithBakerReadBarrier(HInstruction* instruction, 6091 Location ref, 6092 Register obj, 6093 uint32_t data_offset, 6094 Location index, 6095 Location temp, 6096 bool needs_null_check) { 6097 DCHECK(kEmitCompilerReadBarrier); 6098 DCHECK(kUseBakerReadBarrier); 6099 6100 // /* HeapReference<Object> */ ref = 6101 // *(obj + data_offset + index * sizeof(HeapReference<Object>)) 6102 GenerateReferenceLoadWithBakerReadBarrier( 6103 instruction, ref, obj, data_offset, index, temp, needs_null_check); 6104 } 6105 6106 void CodeGeneratorARM::GenerateReferenceLoadWithBakerReadBarrier(HInstruction* instruction, 6107 Location ref, 6108 Register obj, 6109 uint32_t offset, 6110 Location index, 6111 Location temp, 6112 bool needs_null_check) { 6113 DCHECK(kEmitCompilerReadBarrier); 6114 DCHECK(kUseBakerReadBarrier); 6115 6116 // In slow path based read barriers, the read barrier call is 6117 // inserted after the original load. However, in fast path based 6118 // Baker's read barriers, we need to perform the load of 6119 // mirror::Object::monitor_ *before* the original reference load. 6120 // This load-load ordering is required by the read barrier. 6121 // The fast path/slow path (for Baker's algorithm) should look like: 6122 // 6123 // uint32_t rb_state = Lockword(obj->monitor_).ReadBarrierState(); 6124 // lfence; // Load fence or artificial data dependency to prevent load-load reordering 6125 // HeapReference<Object> ref = *src; // Original reference load. 6126 // bool is_gray = (rb_state == ReadBarrier::gray_ptr_); 6127 // if (is_gray) { 6128 // ref = ReadBarrier::Mark(ref); // Performed by runtime entrypoint slow path. 6129 // } 6130 // 6131 // Note: the original implementation in ReadBarrier::Barrier is 6132 // slightly more complex as it performs additional checks that we do 6133 // not do here for performance reasons. 6134 6135 Register ref_reg = ref.AsRegister<Register>(); 6136 Register temp_reg = temp.AsRegister<Register>(); 6137 uint32_t monitor_offset = mirror::Object::MonitorOffset().Int32Value(); 6138 6139 // /* int32_t */ monitor = obj->monitor_ 6140 __ LoadFromOffset(kLoadWord, temp_reg, obj, monitor_offset); 6141 if (needs_null_check) { 6142 MaybeRecordImplicitNullCheck(instruction); 6143 } 6144 // /* LockWord */ lock_word = LockWord(monitor) 6145 static_assert(sizeof(LockWord) == sizeof(int32_t), 6146 "art::LockWord and int32_t have different sizes."); 6147 // /* uint32_t */ rb_state = lock_word.ReadBarrierState() 6148 __ Lsr(temp_reg, temp_reg, LockWord::kReadBarrierStateShift); 6149 __ and_(temp_reg, temp_reg, ShifterOperand(LockWord::kReadBarrierStateMask)); 6150 static_assert( 6151 LockWord::kReadBarrierStateMask == ReadBarrier::rb_ptr_mask_, 6152 "art::LockWord::kReadBarrierStateMask is not equal to art::ReadBarrier::rb_ptr_mask_."); 6153 6154 // Introduce a dependency on the high bits of rb_state, which shall 6155 // be all zeroes, to prevent load-load reordering, and without using 6156 // a memory barrier (which would be more expensive). 6157 // IP = rb_state & ~LockWord::kReadBarrierStateMask = 0 6158 __ bic(IP, temp_reg, ShifterOperand(LockWord::kReadBarrierStateMask)); 6159 // obj is unchanged by this operation, but its value now depends on 6160 // IP, which depends on temp_reg. 6161 __ add(obj, obj, ShifterOperand(IP)); 6162 6163 // The actual reference load. 6164 if (index.IsValid()) { 6165 static_assert( 6166 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 6167 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 6168 // /* HeapReference<Object> */ ref = 6169 // *(obj + offset + index * sizeof(HeapReference<Object>)) 6170 if (index.IsConstant()) { 6171 size_t computed_offset = 6172 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + offset; 6173 __ LoadFromOffset(kLoadWord, ref_reg, obj, computed_offset); 6174 } else { 6175 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 6176 __ LoadFromOffset(kLoadWord, ref_reg, IP, offset); 6177 } 6178 } else { 6179 // /* HeapReference<Object> */ ref = *(obj + offset) 6180 __ LoadFromOffset(kLoadWord, ref_reg, obj, offset); 6181 } 6182 6183 // Object* ref = ref_addr->AsMirrorPtr() 6184 __ MaybeUnpoisonHeapReference(ref_reg); 6185 6186 // Slow path used to mark the object `ref` when it is gray. 6187 SlowPathCode* slow_path = 6188 new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM(instruction, ref, ref); 6189 AddSlowPath(slow_path); 6190 6191 // if (rb_state == ReadBarrier::gray_ptr_) 6192 // ref = ReadBarrier::Mark(ref); 6193 __ cmp(temp_reg, ShifterOperand(ReadBarrier::gray_ptr_)); 6194 __ b(slow_path->GetEntryLabel(), EQ); 6195 __ Bind(slow_path->GetExitLabel()); 6196 } 6197 6198 void CodeGeneratorARM::GenerateReadBarrierSlow(HInstruction* instruction, 6199 Location out, 6200 Location ref, 6201 Location obj, 6202 uint32_t offset, 6203 Location index) { 6204 DCHECK(kEmitCompilerReadBarrier); 6205 6206 // Insert a slow path based read barrier *after* the reference load. 6207 // 6208 // If heap poisoning is enabled, the unpoisoning of the loaded 6209 // reference will be carried out by the runtime within the slow 6210 // path. 6211 // 6212 // Note that `ref` currently does not get unpoisoned (when heap 6213 // poisoning is enabled), which is alright as the `ref` argument is 6214 // not used by the artReadBarrierSlow entry point. 6215 // 6216 // TODO: Unpoison `ref` when it is used by artReadBarrierSlow. 6217 SlowPathCode* slow_path = new (GetGraph()->GetArena()) 6218 ReadBarrierForHeapReferenceSlowPathARM(instruction, out, ref, obj, offset, index); 6219 AddSlowPath(slow_path); 6220 6221 __ b(slow_path->GetEntryLabel()); 6222 __ Bind(slow_path->GetExitLabel()); 6223 } 6224 6225 void CodeGeneratorARM::MaybeGenerateReadBarrierSlow(HInstruction* instruction, 6226 Location out, 6227 Location ref, 6228 Location obj, 6229 uint32_t offset, 6230 Location index) { 6231 if (kEmitCompilerReadBarrier) { 6232 // Baker's read barriers shall be handled by the fast path 6233 // (CodeGeneratorARM::GenerateReferenceLoadWithBakerReadBarrier). 6234 DCHECK(!kUseBakerReadBarrier); 6235 // If heap poisoning is enabled, unpoisoning will be taken care of 6236 // by the runtime within the slow path. 6237 GenerateReadBarrierSlow(instruction, out, ref, obj, offset, index); 6238 } else if (kPoisonHeapReferences) { 6239 __ UnpoisonHeapReference(out.AsRegister<Register>()); 6240 } 6241 } 6242 6243 void CodeGeneratorARM::GenerateReadBarrierForRootSlow(HInstruction* instruction, 6244 Location out, 6245 Location root) { 6246 DCHECK(kEmitCompilerReadBarrier); 6247 6248 // Insert a slow path based read barrier *after* the GC root load. 6249 // 6250 // Note that GC roots are not affected by heap poisoning, so we do 6251 // not need to do anything special for this here. 6252 SlowPathCode* slow_path = 6253 new (GetGraph()->GetArena()) ReadBarrierForRootSlowPathARM(instruction, out, root); 6254 AddSlowPath(slow_path); 6255 6256 __ b(slow_path->GetEntryLabel()); 6257 __ Bind(slow_path->GetExitLabel()); 6258 } 6259 6260 HInvokeStaticOrDirect::DispatchInfo CodeGeneratorARM::GetSupportedInvokeStaticOrDirectDispatch( 6261 const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info, 6262 MethodReference target_method) { 6263 HInvokeStaticOrDirect::DispatchInfo dispatch_info = desired_dispatch_info; 6264 // We disable pc-relative load when there is an irreducible loop, as the optimization 6265 // is incompatible with it. 6266 // TODO: Create as many ArmDexCacheArraysBase instructions as needed for methods 6267 // with irreducible loops. 6268 if (GetGraph()->HasIrreducibleLoops() && 6269 (dispatch_info.method_load_kind == 6270 HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative)) { 6271 dispatch_info.method_load_kind = HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod; 6272 } 6273 6274 if (dispatch_info.code_ptr_location == HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative) { 6275 const DexFile& outer_dex_file = GetGraph()->GetDexFile(); 6276 if (&outer_dex_file != target_method.dex_file) { 6277 // Calls across dex files are more likely to exceed the available BL range, 6278 // so use absolute patch with fixup if available and kCallArtMethod otherwise. 6279 HInvokeStaticOrDirect::CodePtrLocation code_ptr_location = 6280 (desired_dispatch_info.method_load_kind == 6281 HInvokeStaticOrDirect::MethodLoadKind::kDirectAddressWithFixup) 6282 ? HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup 6283 : HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod; 6284 return HInvokeStaticOrDirect::DispatchInfo { 6285 dispatch_info.method_load_kind, 6286 code_ptr_location, 6287 dispatch_info.method_load_data, 6288 0u 6289 }; 6290 } 6291 } 6292 return dispatch_info; 6293 } 6294 6295 Register CodeGeneratorARM::GetInvokeStaticOrDirectExtraParameter(HInvokeStaticOrDirect* invoke, 6296 Register temp) { 6297 DCHECK_EQ(invoke->InputCount(), invoke->GetNumberOfArguments() + 1u); 6298 Location location = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 6299 if (!invoke->GetLocations()->Intrinsified()) { 6300 return location.AsRegister<Register>(); 6301 } 6302 // For intrinsics we allow any location, so it may be on the stack. 6303 if (!location.IsRegister()) { 6304 __ LoadFromOffset(kLoadWord, temp, SP, location.GetStackIndex()); 6305 return temp; 6306 } 6307 // For register locations, check if the register was saved. If so, get it from the stack. 6308 // Note: There is a chance that the register was saved but not overwritten, so we could 6309 // save one load. However, since this is just an intrinsic slow path we prefer this 6310 // simple and more robust approach rather that trying to determine if that's the case. 6311 SlowPathCode* slow_path = GetCurrentSlowPath(); 6312 DCHECK(slow_path != nullptr); // For intrinsified invokes the call is emitted on the slow path. 6313 if (slow_path->IsCoreRegisterSaved(location.AsRegister<Register>())) { 6314 int stack_offset = slow_path->GetStackOffsetOfCoreRegister(location.AsRegister<Register>()); 6315 __ LoadFromOffset(kLoadWord, temp, SP, stack_offset); 6316 return temp; 6317 } 6318 return location.AsRegister<Register>(); 6319 } 6320 6321 void CodeGeneratorARM::GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) { 6322 // For better instruction scheduling we load the direct code pointer before the method pointer. 6323 switch (invoke->GetCodePtrLocation()) { 6324 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup: 6325 // LR = code address from literal pool with link-time patch. 6326 __ LoadLiteral(LR, DeduplicateMethodCodeLiteral(invoke->GetTargetMethod())); 6327 break; 6328 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect: 6329 // LR = invoke->GetDirectCodePtr(); 6330 __ LoadImmediate(LR, invoke->GetDirectCodePtr()); 6331 break; 6332 default: 6333 break; 6334 } 6335 6336 Location callee_method = temp; // For all kinds except kRecursive, callee will be in temp. 6337 switch (invoke->GetMethodLoadKind()) { 6338 case HInvokeStaticOrDirect::MethodLoadKind::kStringInit: 6339 // temp = thread->string_init_entrypoint 6340 __ LoadFromOffset(kLoadWord, temp.AsRegister<Register>(), TR, invoke->GetStringInitOffset()); 6341 break; 6342 case HInvokeStaticOrDirect::MethodLoadKind::kRecursive: 6343 callee_method = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 6344 break; 6345 case HInvokeStaticOrDirect::MethodLoadKind::kDirectAddress: 6346 __ LoadImmediate(temp.AsRegister<Register>(), invoke->GetMethodAddress()); 6347 break; 6348 case HInvokeStaticOrDirect::MethodLoadKind::kDirectAddressWithFixup: 6349 __ LoadLiteral(temp.AsRegister<Register>(), 6350 DeduplicateMethodAddressLiteral(invoke->GetTargetMethod())); 6351 break; 6352 case HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative: { 6353 HArmDexCacheArraysBase* base = 6354 invoke->InputAt(invoke->GetSpecialInputIndex())->AsArmDexCacheArraysBase(); 6355 Register base_reg = GetInvokeStaticOrDirectExtraParameter(invoke, 6356 temp.AsRegister<Register>()); 6357 int32_t offset = invoke->GetDexCacheArrayOffset() - base->GetElementOffset(); 6358 __ LoadFromOffset(kLoadWord, temp.AsRegister<Register>(), base_reg, offset); 6359 break; 6360 } 6361 case HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod: { 6362 Location current_method = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 6363 Register method_reg; 6364 Register reg = temp.AsRegister<Register>(); 6365 if (current_method.IsRegister()) { 6366 method_reg = current_method.AsRegister<Register>(); 6367 } else { 6368 DCHECK(invoke->GetLocations()->Intrinsified()); 6369 DCHECK(!current_method.IsValid()); 6370 method_reg = reg; 6371 __ LoadFromOffset(kLoadWord, reg, SP, kCurrentMethodStackOffset); 6372 } 6373 // /* ArtMethod*[] */ temp = temp.ptr_sized_fields_->dex_cache_resolved_methods_; 6374 __ LoadFromOffset(kLoadWord, 6375 reg, 6376 method_reg, 6377 ArtMethod::DexCacheResolvedMethodsOffset(kArmPointerSize).Int32Value()); 6378 // temp = temp[index_in_cache]; 6379 // Note: Don't use invoke->GetTargetMethod() as it may point to a different dex file. 6380 uint32_t index_in_cache = invoke->GetDexMethodIndex(); 6381 __ LoadFromOffset(kLoadWord, reg, reg, CodeGenerator::GetCachePointerOffset(index_in_cache)); 6382 break; 6383 } 6384 } 6385 6386 switch (invoke->GetCodePtrLocation()) { 6387 case HInvokeStaticOrDirect::CodePtrLocation::kCallSelf: 6388 __ bl(GetFrameEntryLabel()); 6389 break; 6390 case HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative: 6391 relative_call_patches_.emplace_back(invoke->GetTargetMethod()); 6392 __ BindTrackedLabel(&relative_call_patches_.back().label); 6393 // Arbitrarily branch to the BL itself, override at link time. 6394 __ bl(&relative_call_patches_.back().label); 6395 break; 6396 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup: 6397 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect: 6398 // LR prepared above for better instruction scheduling. 6399 // LR() 6400 __ blx(LR); 6401 break; 6402 case HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod: 6403 // LR = callee_method->entry_point_from_quick_compiled_code_ 6404 __ LoadFromOffset( 6405 kLoadWord, LR, callee_method.AsRegister<Register>(), 6406 ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmWordSize).Int32Value()); 6407 // LR() 6408 __ blx(LR); 6409 break; 6410 } 6411 6412 DCHECK(!IsLeafMethod()); 6413 } 6414 6415 void CodeGeneratorARM::GenerateVirtualCall(HInvokeVirtual* invoke, Location temp_location) { 6416 Register temp = temp_location.AsRegister<Register>(); 6417 uint32_t method_offset = mirror::Class::EmbeddedVTableEntryOffset( 6418 invoke->GetVTableIndex(), kArmPointerSize).Uint32Value(); 6419 6420 // Use the calling convention instead of the location of the receiver, as 6421 // intrinsics may have put the receiver in a different register. In the intrinsics 6422 // slow path, the arguments have been moved to the right place, so here we are 6423 // guaranteed that the receiver is the first register of the calling convention. 6424 InvokeDexCallingConvention calling_convention; 6425 Register receiver = calling_convention.GetRegisterAt(0); 6426 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 6427 // /* HeapReference<Class> */ temp = receiver->klass_ 6428 __ LoadFromOffset(kLoadWord, temp, receiver, class_offset); 6429 MaybeRecordImplicitNullCheck(invoke); 6430 // Instead of simply (possibly) unpoisoning `temp` here, we should 6431 // emit a read barrier for the previous class reference load. 6432 // However this is not required in practice, as this is an 6433 // intermediate/temporary reference and because the current 6434 // concurrent copying collector keeps the from-space memory 6435 // intact/accessible until the end of the marking phase (the 6436 // concurrent copying collector may not in the future). 6437 __ MaybeUnpoisonHeapReference(temp); 6438 // temp = temp->GetMethodAt(method_offset); 6439 uint32_t entry_point = ArtMethod::EntryPointFromQuickCompiledCodeOffset( 6440 kArmWordSize).Int32Value(); 6441 __ LoadFromOffset(kLoadWord, temp, temp, method_offset); 6442 // LR = temp->GetEntryPoint(); 6443 __ LoadFromOffset(kLoadWord, LR, temp, entry_point); 6444 // LR(); 6445 __ blx(LR); 6446 } 6447 6448 CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewPcRelativeStringPatch( 6449 const DexFile& dex_file, uint32_t string_index) { 6450 return NewPcRelativePatch(dex_file, string_index, &pc_relative_string_patches_); 6451 } 6452 6453 CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewPcRelativeDexCacheArrayPatch( 6454 const DexFile& dex_file, uint32_t element_offset) { 6455 return NewPcRelativePatch(dex_file, element_offset, &pc_relative_dex_cache_patches_); 6456 } 6457 6458 CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewPcRelativePatch( 6459 const DexFile& dex_file, uint32_t offset_or_index, ArenaDeque<PcRelativePatchInfo>* patches) { 6460 patches->emplace_back(dex_file, offset_or_index); 6461 return &patches->back(); 6462 } 6463 6464 Literal* CodeGeneratorARM::DeduplicateBootImageStringLiteral(const DexFile& dex_file, 6465 uint32_t string_index) { 6466 return boot_image_string_patches_.GetOrCreate( 6467 StringReference(&dex_file, string_index), 6468 [this]() { return __ NewLiteral<uint32_t>(/* placeholder */ 0u); }); 6469 } 6470 6471 Literal* CodeGeneratorARM::DeduplicateBootImageAddressLiteral(uint32_t address) { 6472 bool needs_patch = GetCompilerOptions().GetIncludePatchInformation(); 6473 Uint32ToLiteralMap* map = needs_patch ? &boot_image_address_patches_ : &uint32_literals_; 6474 return DeduplicateUint32Literal(dchecked_integral_cast<uint32_t>(address), map); 6475 } 6476 6477 Literal* CodeGeneratorARM::DeduplicateDexCacheAddressLiteral(uint32_t address) { 6478 return DeduplicateUint32Literal(address, &uint32_literals_); 6479 } 6480 6481 void CodeGeneratorARM::EmitLinkerPatches(ArenaVector<LinkerPatch>* linker_patches) { 6482 DCHECK(linker_patches->empty()); 6483 size_t size = 6484 method_patches_.size() + 6485 call_patches_.size() + 6486 relative_call_patches_.size() + 6487 /* MOVW+MOVT for each base */ 2u * pc_relative_dex_cache_patches_.size() + 6488 boot_image_string_patches_.size() + 6489 /* MOVW+MOVT for each base */ 2u * pc_relative_string_patches_.size() + 6490 boot_image_address_patches_.size(); 6491 linker_patches->reserve(size); 6492 for (const auto& entry : method_patches_) { 6493 const MethodReference& target_method = entry.first; 6494 Literal* literal = entry.second; 6495 DCHECK(literal->GetLabel()->IsBound()); 6496 uint32_t literal_offset = literal->GetLabel()->Position(); 6497 linker_patches->push_back(LinkerPatch::MethodPatch(literal_offset, 6498 target_method.dex_file, 6499 target_method.dex_method_index)); 6500 } 6501 for (const auto& entry : call_patches_) { 6502 const MethodReference& target_method = entry.first; 6503 Literal* literal = entry.second; 6504 DCHECK(literal->GetLabel()->IsBound()); 6505 uint32_t literal_offset = literal->GetLabel()->Position(); 6506 linker_patches->push_back(LinkerPatch::CodePatch(literal_offset, 6507 target_method.dex_file, 6508 target_method.dex_method_index)); 6509 } 6510 for (const MethodPatchInfo<Label>& info : relative_call_patches_) { 6511 uint32_t literal_offset = info.label.Position(); 6512 linker_patches->push_back(LinkerPatch::RelativeCodePatch(literal_offset, 6513 info.target_method.dex_file, 6514 info.target_method.dex_method_index)); 6515 } 6516 for (const PcRelativePatchInfo& info : pc_relative_dex_cache_patches_) { 6517 const DexFile& dex_file = info.target_dex_file; 6518 size_t base_element_offset = info.offset_or_index; 6519 DCHECK(info.add_pc_label.IsBound()); 6520 uint32_t add_pc_offset = dchecked_integral_cast<uint32_t>(info.add_pc_label.Position()); 6521 // Add MOVW patch. 6522 DCHECK(info.movw_label.IsBound()); 6523 uint32_t movw_offset = dchecked_integral_cast<uint32_t>(info.movw_label.Position()); 6524 linker_patches->push_back(LinkerPatch::DexCacheArrayPatch(movw_offset, 6525 &dex_file, 6526 add_pc_offset, 6527 base_element_offset)); 6528 // Add MOVT patch. 6529 DCHECK(info.movt_label.IsBound()); 6530 uint32_t movt_offset = dchecked_integral_cast<uint32_t>(info.movt_label.Position()); 6531 linker_patches->push_back(LinkerPatch::DexCacheArrayPatch(movt_offset, 6532 &dex_file, 6533 add_pc_offset, 6534 base_element_offset)); 6535 } 6536 for (const auto& entry : boot_image_string_patches_) { 6537 const StringReference& target_string = entry.first; 6538 Literal* literal = entry.second; 6539 DCHECK(literal->GetLabel()->IsBound()); 6540 uint32_t literal_offset = literal->GetLabel()->Position(); 6541 linker_patches->push_back(LinkerPatch::StringPatch(literal_offset, 6542 target_string.dex_file, 6543 target_string.string_index)); 6544 } 6545 for (const PcRelativePatchInfo& info : pc_relative_string_patches_) { 6546 const DexFile& dex_file = info.target_dex_file; 6547 uint32_t string_index = info.offset_or_index; 6548 DCHECK(info.add_pc_label.IsBound()); 6549 uint32_t add_pc_offset = dchecked_integral_cast<uint32_t>(info.add_pc_label.Position()); 6550 // Add MOVW patch. 6551 DCHECK(info.movw_label.IsBound()); 6552 uint32_t movw_offset = dchecked_integral_cast<uint32_t>(info.movw_label.Position()); 6553 linker_patches->push_back(LinkerPatch::RelativeStringPatch(movw_offset, 6554 &dex_file, 6555 add_pc_offset, 6556 string_index)); 6557 // Add MOVT patch. 6558 DCHECK(info.movt_label.IsBound()); 6559 uint32_t movt_offset = dchecked_integral_cast<uint32_t>(info.movt_label.Position()); 6560 linker_patches->push_back(LinkerPatch::RelativeStringPatch(movt_offset, 6561 &dex_file, 6562 add_pc_offset, 6563 string_index)); 6564 } 6565 for (const auto& entry : boot_image_address_patches_) { 6566 DCHECK(GetCompilerOptions().GetIncludePatchInformation()); 6567 Literal* literal = entry.second; 6568 DCHECK(literal->GetLabel()->IsBound()); 6569 uint32_t literal_offset = literal->GetLabel()->Position(); 6570 linker_patches->push_back(LinkerPatch::RecordPosition(literal_offset)); 6571 } 6572 } 6573 6574 Literal* CodeGeneratorARM::DeduplicateUint32Literal(uint32_t value, Uint32ToLiteralMap* map) { 6575 return map->GetOrCreate( 6576 value, 6577 [this, value]() { return __ NewLiteral<uint32_t>(value); }); 6578 } 6579 6580 Literal* CodeGeneratorARM::DeduplicateMethodLiteral(MethodReference target_method, 6581 MethodToLiteralMap* map) { 6582 return map->GetOrCreate( 6583 target_method, 6584 [this]() { return __ NewLiteral<uint32_t>(/* placeholder */ 0u); }); 6585 } 6586 6587 Literal* CodeGeneratorARM::DeduplicateMethodAddressLiteral(MethodReference target_method) { 6588 return DeduplicateMethodLiteral(target_method, &method_patches_); 6589 } 6590 6591 Literal* CodeGeneratorARM::DeduplicateMethodCodeLiteral(MethodReference target_method) { 6592 return DeduplicateMethodLiteral(target_method, &call_patches_); 6593 } 6594 6595 void LocationsBuilderARM::VisitMultiplyAccumulate(HMultiplyAccumulate* instr) { 6596 LocationSummary* locations = 6597 new (GetGraph()->GetArena()) LocationSummary(instr, LocationSummary::kNoCall); 6598 locations->SetInAt(HMultiplyAccumulate::kInputAccumulatorIndex, 6599 Location::RequiresRegister()); 6600 locations->SetInAt(HMultiplyAccumulate::kInputMulLeftIndex, Location::RequiresRegister()); 6601 locations->SetInAt(HMultiplyAccumulate::kInputMulRightIndex, Location::RequiresRegister()); 6602 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 6603 } 6604 6605 void InstructionCodeGeneratorARM::VisitMultiplyAccumulate(HMultiplyAccumulate* instr) { 6606 LocationSummary* locations = instr->GetLocations(); 6607 Register res = locations->Out().AsRegister<Register>(); 6608 Register accumulator = 6609 locations->InAt(HMultiplyAccumulate::kInputAccumulatorIndex).AsRegister<Register>(); 6610 Register mul_left = 6611 locations->InAt(HMultiplyAccumulate::kInputMulLeftIndex).AsRegister<Register>(); 6612 Register mul_right = 6613 locations->InAt(HMultiplyAccumulate::kInputMulRightIndex).AsRegister<Register>(); 6614 6615 if (instr->GetOpKind() == HInstruction::kAdd) { 6616 __ mla(res, mul_left, mul_right, accumulator); 6617 } else { 6618 __ mls(res, mul_left, mul_right, accumulator); 6619 } 6620 } 6621 6622 void LocationsBuilderARM::VisitBoundType(HBoundType* instruction ATTRIBUTE_UNUSED) { 6623 // Nothing to do, this should be removed during prepare for register allocator. 6624 LOG(FATAL) << "Unreachable"; 6625 } 6626 6627 void InstructionCodeGeneratorARM::VisitBoundType(HBoundType* instruction ATTRIBUTE_UNUSED) { 6628 // Nothing to do, this should be removed during prepare for register allocator. 6629 LOG(FATAL) << "Unreachable"; 6630 } 6631 6632 // Simple implementation of packed switch - generate cascaded compare/jumps. 6633 void LocationsBuilderARM::VisitPackedSwitch(HPackedSwitch* switch_instr) { 6634 LocationSummary* locations = 6635 new (GetGraph()->GetArena()) LocationSummary(switch_instr, LocationSummary::kNoCall); 6636 locations->SetInAt(0, Location::RequiresRegister()); 6637 if (switch_instr->GetNumEntries() > kPackedSwitchCompareJumpThreshold && 6638 codegen_->GetAssembler()->IsThumb()) { 6639 locations->AddTemp(Location::RequiresRegister()); // We need a temp for the table base. 6640 if (switch_instr->GetStartValue() != 0) { 6641 locations->AddTemp(Location::RequiresRegister()); // We need a temp for the bias. 6642 } 6643 } 6644 } 6645 6646 void InstructionCodeGeneratorARM::VisitPackedSwitch(HPackedSwitch* switch_instr) { 6647 int32_t lower_bound = switch_instr->GetStartValue(); 6648 uint32_t num_entries = switch_instr->GetNumEntries(); 6649 LocationSummary* locations = switch_instr->GetLocations(); 6650 Register value_reg = locations->InAt(0).AsRegister<Register>(); 6651 HBasicBlock* default_block = switch_instr->GetDefaultBlock(); 6652 6653 if (num_entries <= kPackedSwitchCompareJumpThreshold || !codegen_->GetAssembler()->IsThumb()) { 6654 // Create a series of compare/jumps. 6655 Register temp_reg = IP; 6656 // Note: It is fine for the below AddConstantSetFlags() using IP register to temporarily store 6657 // the immediate, because IP is used as the destination register. For the other 6658 // AddConstantSetFlags() and GenerateCompareWithImmediate(), the immediate values are constant, 6659 // and they can be encoded in the instruction without making use of IP register. 6660 __ AddConstantSetFlags(temp_reg, value_reg, -lower_bound); 6661 6662 const ArenaVector<HBasicBlock*>& successors = switch_instr->GetBlock()->GetSuccessors(); 6663 // Jump to successors[0] if value == lower_bound. 6664 __ b(codegen_->GetLabelOf(successors[0]), EQ); 6665 int32_t last_index = 0; 6666 for (; num_entries - last_index > 2; last_index += 2) { 6667 __ AddConstantSetFlags(temp_reg, temp_reg, -2); 6668 // Jump to successors[last_index + 1] if value < case_value[last_index + 2]. 6669 __ b(codegen_->GetLabelOf(successors[last_index + 1]), LO); 6670 // Jump to successors[last_index + 2] if value == case_value[last_index + 2]. 6671 __ b(codegen_->GetLabelOf(successors[last_index + 2]), EQ); 6672 } 6673 if (num_entries - last_index == 2) { 6674 // The last missing case_value. 6675 __ CmpConstant(temp_reg, 1); 6676 __ b(codegen_->GetLabelOf(successors[last_index + 1]), EQ); 6677 } 6678 6679 // And the default for any other value. 6680 if (!codegen_->GoesToNextBlock(switch_instr->GetBlock(), default_block)) { 6681 __ b(codegen_->GetLabelOf(default_block)); 6682 } 6683 } else { 6684 // Create a table lookup. 6685 Register temp_reg = locations->GetTemp(0).AsRegister<Register>(); 6686 6687 // Materialize a pointer to the switch table 6688 std::vector<Label*> labels(num_entries); 6689 const ArenaVector<HBasicBlock*>& successors = switch_instr->GetBlock()->GetSuccessors(); 6690 for (uint32_t i = 0; i < num_entries; i++) { 6691 labels[i] = codegen_->GetLabelOf(successors[i]); 6692 } 6693 JumpTable* table = __ CreateJumpTable(std::move(labels), temp_reg); 6694 6695 // Remove the bias. 6696 Register key_reg; 6697 if (lower_bound != 0) { 6698 key_reg = locations->GetTemp(1).AsRegister<Register>(); 6699 __ AddConstant(key_reg, value_reg, -lower_bound); 6700 } else { 6701 key_reg = value_reg; 6702 } 6703 6704 // Check whether the value is in the table, jump to default block if not. 6705 __ CmpConstant(key_reg, num_entries - 1); 6706 __ b(codegen_->GetLabelOf(default_block), Condition::HI); 6707 6708 // Load the displacement from the table. 6709 __ ldr(temp_reg, Address(temp_reg, key_reg, Shift::LSL, 2)); 6710 6711 // Dispatch is a direct add to the PC (for Thumb2). 6712 __ EmitJumpTableDispatch(table, temp_reg); 6713 } 6714 } 6715 6716 void LocationsBuilderARM::VisitArmDexCacheArraysBase(HArmDexCacheArraysBase* base) { 6717 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(base); 6718 locations->SetOut(Location::RequiresRegister()); 6719 } 6720 6721 void InstructionCodeGeneratorARM::VisitArmDexCacheArraysBase(HArmDexCacheArraysBase* base) { 6722 Register base_reg = base->GetLocations()->Out().AsRegister<Register>(); 6723 CodeGeneratorARM::PcRelativePatchInfo* labels = 6724 codegen_->NewPcRelativeDexCacheArrayPatch(base->GetDexFile(), base->GetElementOffset()); 6725 __ BindTrackedLabel(&labels->movw_label); 6726 __ movw(base_reg, /* placeholder */ 0u); 6727 __ BindTrackedLabel(&labels->movt_label); 6728 __ movt(base_reg, /* placeholder */ 0u); 6729 __ BindTrackedLabel(&labels->add_pc_label); 6730 __ add(base_reg, base_reg, ShifterOperand(PC)); 6731 } 6732 6733 void CodeGeneratorARM::MoveFromReturnRegister(Location trg, Primitive::Type type) { 6734 if (!trg.IsValid()) { 6735 DCHECK_EQ(type, Primitive::kPrimVoid); 6736 return; 6737 } 6738 6739 DCHECK_NE(type, Primitive::kPrimVoid); 6740 6741 Location return_loc = InvokeDexCallingConventionVisitorARM().GetReturnLocation(type); 6742 if (return_loc.Equals(trg)) { 6743 return; 6744 } 6745 6746 // TODO: Consider pairs in the parallel move resolver, then this could be nicely merged 6747 // with the last branch. 6748 if (type == Primitive::kPrimLong) { 6749 HParallelMove parallel_move(GetGraph()->GetArena()); 6750 parallel_move.AddMove(return_loc.ToLow(), trg.ToLow(), Primitive::kPrimInt, nullptr); 6751 parallel_move.AddMove(return_loc.ToHigh(), trg.ToHigh(), Primitive::kPrimInt, nullptr); 6752 GetMoveResolver()->EmitNativeCode(¶llel_move); 6753 } else if (type == Primitive::kPrimDouble) { 6754 HParallelMove parallel_move(GetGraph()->GetArena()); 6755 parallel_move.AddMove(return_loc.ToLow(), trg.ToLow(), Primitive::kPrimFloat, nullptr); 6756 parallel_move.AddMove(return_loc.ToHigh(), trg.ToHigh(), Primitive::kPrimFloat, nullptr); 6757 GetMoveResolver()->EmitNativeCode(¶llel_move); 6758 } else { 6759 // Let the parallel move resolver take care of all of this. 6760 HParallelMove parallel_move(GetGraph()->GetArena()); 6761 parallel_move.AddMove(return_loc, trg, type, nullptr); 6762 GetMoveResolver()->EmitNativeCode(¶llel_move); 6763 } 6764 } 6765 6766 void LocationsBuilderARM::VisitClassTableGet(HClassTableGet* instruction) { 6767 LocationSummary* locations = 6768 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 6769 locations->SetInAt(0, Location::RequiresRegister()); 6770 locations->SetOut(Location::RequiresRegister()); 6771 } 6772 6773 void InstructionCodeGeneratorARM::VisitClassTableGet(HClassTableGet* instruction) { 6774 LocationSummary* locations = instruction->GetLocations(); 6775 if (instruction->GetTableKind() == HClassTableGet::TableKind::kVTable) { 6776 uint32_t method_offset = mirror::Class::EmbeddedVTableEntryOffset( 6777 instruction->GetIndex(), kArmPointerSize).SizeValue(); 6778 __ LoadFromOffset(kLoadWord, 6779 locations->Out().AsRegister<Register>(), 6780 locations->InAt(0).AsRegister<Register>(), 6781 method_offset); 6782 } else { 6783 uint32_t method_offset = static_cast<uint32_t>(ImTable::OffsetOfElement( 6784 instruction->GetIndex() % ImTable::kSize, kArmPointerSize)); 6785 __ LoadFromOffset(kLoadWord, 6786 locations->Out().AsRegister<Register>(), 6787 locations->InAt(0).AsRegister<Register>(), 6788 mirror::Class::ImtPtrOffset(kArmPointerSize).Uint32Value()); 6789 __ LoadFromOffset(kLoadWord, 6790 locations->Out().AsRegister<Register>(), 6791 locations->Out().AsRegister<Register>(), 6792 method_offset); 6793 } 6794 } 6795 6796 #undef __ 6797 #undef QUICK_ENTRY_POINT 6798 6799 } // namespace arm 6800 } // namespace art 6801