code_generator_arm.cc revision d4ceecc85a5aab2ec23ea1bd010692ba8c8aaa0c
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 35namespace art { 36 37template<class MirrorType> 38class GcRoot; 39 40namespace arm { 41 42static 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 47static constexpr int kCurrentMethodStackOffset = 0; 48static constexpr Register kMethodRegisterArgument = R0; 49 50static constexpr Register kCoreAlwaysSpillRegister = R5; 51static constexpr Register kCoreCalleeSaves[] = 52 { R5, R6, R7, R8, R10, R11, LR }; 53static 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. 58static constexpr DRegister DTMP = D31; 59 60static constexpr uint32_t kPackedSwitchCompareJumpThreshold = 7; 61 62// NOLINT on __ macro to suppress wrong warning/fix from clang-tidy. 63#define __ down_cast<ArmAssembler*>(codegen->GetAssembler())-> // NOLINT 64#define QUICK_ENTRY_POINT(x) QUICK_ENTRYPOINT_OFFSET(kArmWordSize, x).Int32Value() 65 66class NullCheckSlowPathARM : public SlowPathCode { 67 public: 68 explicit NullCheckSlowPathARM(HNullCheck* instruction) : SlowPathCode(instruction) {} 69 70 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 71 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 72 __ Bind(GetEntryLabel()); 73 if (instruction_->CanThrowIntoCatchBlock()) { 74 // Live registers will be restored in the catch block if caught. 75 SaveLiveRegisters(codegen, instruction_->GetLocations()); 76 } 77 arm_codegen->InvokeRuntime( 78 QUICK_ENTRY_POINT(pThrowNullPointer), instruction_, instruction_->GetDexPc(), this); 79 CheckEntrypointTypes<kQuickThrowNullPointer, void, void>(); 80 } 81 82 bool IsFatal() const OVERRIDE { return true; } 83 84 const char* GetDescription() const OVERRIDE { return "NullCheckSlowPathARM"; } 85 86 private: 87 DISALLOW_COPY_AND_ASSIGN(NullCheckSlowPathARM); 88}; 89 90class DivZeroCheckSlowPathARM : public SlowPathCode { 91 public: 92 explicit DivZeroCheckSlowPathARM(HDivZeroCheck* instruction) : SlowPathCode(instruction) {} 93 94 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 95 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 96 __ Bind(GetEntryLabel()); 97 if (instruction_->CanThrowIntoCatchBlock()) { 98 // Live registers will be restored in the catch block if caught. 99 SaveLiveRegisters(codegen, instruction_->GetLocations()); 100 } 101 arm_codegen->InvokeRuntime( 102 QUICK_ENTRY_POINT(pThrowDivZero), instruction_, instruction_->GetDexPc(), this); 103 CheckEntrypointTypes<kQuickThrowDivZero, void, void>(); 104 } 105 106 bool IsFatal() const OVERRIDE { return true; } 107 108 const char* GetDescription() const OVERRIDE { return "DivZeroCheckSlowPathARM"; } 109 110 private: 111 DISALLOW_COPY_AND_ASSIGN(DivZeroCheckSlowPathARM); 112}; 113 114class SuspendCheckSlowPathARM : public SlowPathCode { 115 public: 116 SuspendCheckSlowPathARM(HSuspendCheck* instruction, HBasicBlock* successor) 117 : SlowPathCode(instruction), successor_(successor) {} 118 119 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 120 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 121 __ Bind(GetEntryLabel()); 122 SaveLiveRegisters(codegen, instruction_->GetLocations()); 123 arm_codegen->InvokeRuntime( 124 QUICK_ENTRY_POINT(pTestSuspend), instruction_, instruction_->GetDexPc(), this); 125 CheckEntrypointTypes<kQuickTestSuspend, void, void>(); 126 RestoreLiveRegisters(codegen, instruction_->GetLocations()); 127 if (successor_ == nullptr) { 128 __ b(GetReturnLabel()); 129 } else { 130 __ b(arm_codegen->GetLabelOf(successor_)); 131 } 132 } 133 134 Label* GetReturnLabel() { 135 DCHECK(successor_ == nullptr); 136 return &return_label_; 137 } 138 139 HBasicBlock* GetSuccessor() const { 140 return successor_; 141 } 142 143 const char* GetDescription() const OVERRIDE { return "SuspendCheckSlowPathARM"; } 144 145 private: 146 // If not null, the block to branch to after the suspend check. 147 HBasicBlock* const successor_; 148 149 // If `successor_` is null, the label to branch to after the suspend check. 150 Label return_label_; 151 152 DISALLOW_COPY_AND_ASSIGN(SuspendCheckSlowPathARM); 153}; 154 155class BoundsCheckSlowPathARM : public SlowPathCode { 156 public: 157 explicit BoundsCheckSlowPathARM(HBoundsCheck* instruction) 158 : SlowPathCode(instruction) {} 159 160 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 161 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 162 LocationSummary* locations = instruction_->GetLocations(); 163 164 __ Bind(GetEntryLabel()); 165 if (instruction_->CanThrowIntoCatchBlock()) { 166 // Live registers will be restored in the catch block if caught. 167 SaveLiveRegisters(codegen, instruction_->GetLocations()); 168 } 169 // We're moving two locations to locations that could overlap, so we need a parallel 170 // move resolver. 171 InvokeRuntimeCallingConvention calling_convention; 172 codegen->EmitParallelMoves( 173 locations->InAt(0), 174 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 175 Primitive::kPrimInt, 176 locations->InAt(1), 177 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 178 Primitive::kPrimInt); 179 arm_codegen->InvokeRuntime( 180 QUICK_ENTRY_POINT(pThrowArrayBounds), instruction_, instruction_->GetDexPc(), this); 181 CheckEntrypointTypes<kQuickThrowArrayBounds, void, int32_t, int32_t>(); 182 } 183 184 bool IsFatal() const OVERRIDE { return true; } 185 186 const char* GetDescription() const OVERRIDE { return "BoundsCheckSlowPathARM"; } 187 188 private: 189 DISALLOW_COPY_AND_ASSIGN(BoundsCheckSlowPathARM); 190}; 191 192class LoadClassSlowPathARM : public SlowPathCode { 193 public: 194 LoadClassSlowPathARM(HLoadClass* cls, 195 HInstruction* at, 196 uint32_t dex_pc, 197 bool do_clinit) 198 : SlowPathCode(at), cls_(cls), at_(at), dex_pc_(dex_pc), do_clinit_(do_clinit) { 199 DCHECK(at->IsLoadClass() || at->IsClinitCheck()); 200 } 201 202 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 203 LocationSummary* locations = at_->GetLocations(); 204 205 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 206 __ Bind(GetEntryLabel()); 207 SaveLiveRegisters(codegen, locations); 208 209 InvokeRuntimeCallingConvention calling_convention; 210 __ LoadImmediate(calling_convention.GetRegisterAt(0), cls_->GetTypeIndex()); 211 int32_t entry_point_offset = do_clinit_ 212 ? QUICK_ENTRY_POINT(pInitializeStaticStorage) 213 : QUICK_ENTRY_POINT(pInitializeType); 214 arm_codegen->InvokeRuntime(entry_point_offset, at_, dex_pc_, this); 215 if (do_clinit_) { 216 CheckEntrypointTypes<kQuickInitializeStaticStorage, void*, uint32_t>(); 217 } else { 218 CheckEntrypointTypes<kQuickInitializeType, void*, uint32_t>(); 219 } 220 221 // Move the class to the desired location. 222 Location out = locations->Out(); 223 if (out.IsValid()) { 224 DCHECK(out.IsRegister() && !locations->GetLiveRegisters()->ContainsCoreRegister(out.reg())); 225 arm_codegen->Move32(locations->Out(), Location::RegisterLocation(R0)); 226 } 227 RestoreLiveRegisters(codegen, locations); 228 __ b(GetExitLabel()); 229 } 230 231 const char* GetDescription() const OVERRIDE { return "LoadClassSlowPathARM"; } 232 233 private: 234 // The class this slow path will load. 235 HLoadClass* const cls_; 236 237 // The instruction where this slow path is happening. 238 // (Might be the load class or an initialization check). 239 HInstruction* const at_; 240 241 // The dex PC of `at_`. 242 const uint32_t dex_pc_; 243 244 // Whether to initialize the class. 245 const bool do_clinit_; 246 247 DISALLOW_COPY_AND_ASSIGN(LoadClassSlowPathARM); 248}; 249 250class LoadStringSlowPathARM : public SlowPathCode { 251 public: 252 explicit LoadStringSlowPathARM(HLoadString* instruction) : SlowPathCode(instruction) {} 253 254 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 255 LocationSummary* locations = instruction_->GetLocations(); 256 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg())); 257 258 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 259 __ Bind(GetEntryLabel()); 260 SaveLiveRegisters(codegen, locations); 261 262 InvokeRuntimeCallingConvention calling_convention; 263 const uint32_t string_index = instruction_->AsLoadString()->GetStringIndex(); 264 __ LoadImmediate(calling_convention.GetRegisterAt(0), string_index); 265 arm_codegen->InvokeRuntime( 266 QUICK_ENTRY_POINT(pResolveString), instruction_, instruction_->GetDexPc(), this); 267 CheckEntrypointTypes<kQuickResolveString, void*, uint32_t>(); 268 arm_codegen->Move32(locations->Out(), Location::RegisterLocation(R0)); 269 270 RestoreLiveRegisters(codegen, locations); 271 __ b(GetExitLabel()); 272 } 273 274 const char* GetDescription() const OVERRIDE { return "LoadStringSlowPathARM"; } 275 276 private: 277 DISALLOW_COPY_AND_ASSIGN(LoadStringSlowPathARM); 278}; 279 280class TypeCheckSlowPathARM : public SlowPathCode { 281 public: 282 TypeCheckSlowPathARM(HInstruction* instruction, bool is_fatal) 283 : SlowPathCode(instruction), is_fatal_(is_fatal) {} 284 285 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 286 LocationSummary* locations = instruction_->GetLocations(); 287 Location object_class = instruction_->IsCheckCast() ? locations->GetTemp(0) 288 : locations->Out(); 289 DCHECK(instruction_->IsCheckCast() 290 || !locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg())); 291 292 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 293 __ Bind(GetEntryLabel()); 294 295 if (!is_fatal_) { 296 SaveLiveRegisters(codegen, locations); 297 } 298 299 // We're moving two locations to locations that could overlap, so we need a parallel 300 // move resolver. 301 InvokeRuntimeCallingConvention calling_convention; 302 codegen->EmitParallelMoves( 303 locations->InAt(1), 304 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 305 Primitive::kPrimNot, 306 object_class, 307 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 308 Primitive::kPrimNot); 309 310 if (instruction_->IsInstanceOf()) { 311 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pInstanceofNonTrivial), 312 instruction_, 313 instruction_->GetDexPc(), 314 this); 315 CheckEntrypointTypes< 316 kQuickInstanceofNonTrivial, uint32_t, const mirror::Class*, const mirror::Class*>(); 317 arm_codegen->Move32(locations->Out(), Location::RegisterLocation(R0)); 318 } else { 319 DCHECK(instruction_->IsCheckCast()); 320 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pCheckCast), 321 instruction_, 322 instruction_->GetDexPc(), 323 this); 324 CheckEntrypointTypes<kQuickCheckCast, void, const mirror::Class*, const mirror::Class*>(); 325 } 326 327 if (!is_fatal_) { 328 RestoreLiveRegisters(codegen, locations); 329 __ b(GetExitLabel()); 330 } 331 } 332 333 const char* GetDescription() const OVERRIDE { return "TypeCheckSlowPathARM"; } 334 335 bool IsFatal() const OVERRIDE { return is_fatal_; } 336 337 private: 338 const bool is_fatal_; 339 340 DISALLOW_COPY_AND_ASSIGN(TypeCheckSlowPathARM); 341}; 342 343class DeoptimizationSlowPathARM : public SlowPathCode { 344 public: 345 explicit DeoptimizationSlowPathARM(HDeoptimize* instruction) 346 : SlowPathCode(instruction) {} 347 348 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 349 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 350 __ Bind(GetEntryLabel()); 351 SaveLiveRegisters(codegen, instruction_->GetLocations()); 352 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pDeoptimize), 353 instruction_, 354 instruction_->GetDexPc(), 355 this); 356 CheckEntrypointTypes<kQuickDeoptimize, void, void>(); 357 } 358 359 const char* GetDescription() const OVERRIDE { return "DeoptimizationSlowPathARM"; } 360 361 private: 362 DISALLOW_COPY_AND_ASSIGN(DeoptimizationSlowPathARM); 363}; 364 365class ArraySetSlowPathARM : public SlowPathCode { 366 public: 367 explicit ArraySetSlowPathARM(HInstruction* instruction) : SlowPathCode(instruction) {} 368 369 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 370 LocationSummary* locations = instruction_->GetLocations(); 371 __ Bind(GetEntryLabel()); 372 SaveLiveRegisters(codegen, locations); 373 374 InvokeRuntimeCallingConvention calling_convention; 375 HParallelMove parallel_move(codegen->GetGraph()->GetArena()); 376 parallel_move.AddMove( 377 locations->InAt(0), 378 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 379 Primitive::kPrimNot, 380 nullptr); 381 parallel_move.AddMove( 382 locations->InAt(1), 383 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 384 Primitive::kPrimInt, 385 nullptr); 386 parallel_move.AddMove( 387 locations->InAt(2), 388 Location::RegisterLocation(calling_convention.GetRegisterAt(2)), 389 Primitive::kPrimNot, 390 nullptr); 391 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 392 393 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 394 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pAputObject), 395 instruction_, 396 instruction_->GetDexPc(), 397 this); 398 CheckEntrypointTypes<kQuickAputObject, void, mirror::Array*, int32_t, mirror::Object*>(); 399 RestoreLiveRegisters(codegen, locations); 400 __ b(GetExitLabel()); 401 } 402 403 const char* GetDescription() const OVERRIDE { return "ArraySetSlowPathARM"; } 404 405 private: 406 DISALLOW_COPY_AND_ASSIGN(ArraySetSlowPathARM); 407}; 408 409// Slow path marking an object during a read barrier. 410class ReadBarrierMarkSlowPathARM : public SlowPathCode { 411 public: 412 ReadBarrierMarkSlowPathARM(HInstruction* instruction, Location out, Location obj) 413 : SlowPathCode(instruction), out_(out), obj_(obj) { 414 DCHECK(kEmitCompilerReadBarrier); 415 } 416 417 const char* GetDescription() const OVERRIDE { return "ReadBarrierMarkSlowPathARM"; } 418 419 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 420 LocationSummary* locations = instruction_->GetLocations(); 421 Register reg_out = out_.AsRegister<Register>(); 422 DCHECK(locations->CanCall()); 423 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg_out)); 424 DCHECK(instruction_->IsInstanceFieldGet() || 425 instruction_->IsStaticFieldGet() || 426 instruction_->IsArrayGet() || 427 instruction_->IsLoadClass() || 428 instruction_->IsLoadString() || 429 instruction_->IsInstanceOf() || 430 instruction_->IsCheckCast()) 431 << "Unexpected instruction in read barrier marking slow path: " 432 << instruction_->DebugName(); 433 434 __ Bind(GetEntryLabel()); 435 SaveLiveRegisters(codegen, locations); 436 437 InvokeRuntimeCallingConvention calling_convention; 438 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 439 arm_codegen->Move32(Location::RegisterLocation(calling_convention.GetRegisterAt(0)), obj_); 440 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierMark), 441 instruction_, 442 instruction_->GetDexPc(), 443 this); 444 CheckEntrypointTypes<kQuickReadBarrierMark, mirror::Object*, mirror::Object*>(); 445 arm_codegen->Move32(out_, Location::RegisterLocation(R0)); 446 447 RestoreLiveRegisters(codegen, locations); 448 __ b(GetExitLabel()); 449 } 450 451 private: 452 const Location out_; 453 const Location obj_; 454 455 DISALLOW_COPY_AND_ASSIGN(ReadBarrierMarkSlowPathARM); 456}; 457 458// Slow path generating a read barrier for a heap reference. 459class ReadBarrierForHeapReferenceSlowPathARM : public SlowPathCode { 460 public: 461 ReadBarrierForHeapReferenceSlowPathARM(HInstruction* instruction, 462 Location out, 463 Location ref, 464 Location obj, 465 uint32_t offset, 466 Location index) 467 : SlowPathCode(instruction), 468 out_(out), 469 ref_(ref), 470 obj_(obj), 471 offset_(offset), 472 index_(index) { 473 DCHECK(kEmitCompilerReadBarrier); 474 // If `obj` is equal to `out` or `ref`, it means the initial object 475 // has been overwritten by (or after) the heap object reference load 476 // to be instrumented, e.g.: 477 // 478 // __ LoadFromOffset(kLoadWord, out, out, offset); 479 // codegen_->GenerateReadBarrierSlow(instruction, out_loc, out_loc, out_loc, offset); 480 // 481 // In that case, we have lost the information about the original 482 // object, and the emitted read barrier cannot work properly. 483 DCHECK(!obj.Equals(out)) << "obj=" << obj << " out=" << out; 484 DCHECK(!obj.Equals(ref)) << "obj=" << obj << " ref=" << ref; 485 } 486 487 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 488 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 489 LocationSummary* locations = instruction_->GetLocations(); 490 Register reg_out = out_.AsRegister<Register>(); 491 DCHECK(locations->CanCall()); 492 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg_out)); 493 DCHECK(!instruction_->IsInvoke() || 494 (instruction_->IsInvokeStaticOrDirect() && 495 instruction_->GetLocations()->Intrinsified())) 496 << "Unexpected instruction in read barrier for heap reference slow path: " 497 << instruction_->DebugName(); 498 499 __ Bind(GetEntryLabel()); 500 SaveLiveRegisters(codegen, locations); 501 502 // We may have to change the index's value, but as `index_` is a 503 // constant member (like other "inputs" of this slow path), 504 // introduce a copy of it, `index`. 505 Location index = index_; 506 if (index_.IsValid()) { 507 // Handle `index_` for HArrayGet and intrinsic UnsafeGetObject. 508 if (instruction_->IsArrayGet()) { 509 // Compute the actual memory offset and store it in `index`. 510 Register index_reg = index_.AsRegister<Register>(); 511 DCHECK(locations->GetLiveRegisters()->ContainsCoreRegister(index_reg)); 512 if (codegen->IsCoreCalleeSaveRegister(index_reg)) { 513 // We are about to change the value of `index_reg` (see the 514 // calls to art::arm::Thumb2Assembler::Lsl and 515 // art::arm::Thumb2Assembler::AddConstant below), but it has 516 // not been saved by the previous call to 517 // art::SlowPathCode::SaveLiveRegisters, as it is a 518 // callee-save register -- 519 // art::SlowPathCode::SaveLiveRegisters does not consider 520 // callee-save registers, as it has been designed with the 521 // assumption that callee-save registers are supposed to be 522 // handled by the called function. So, as a callee-save 523 // register, `index_reg` _would_ eventually be saved onto 524 // the stack, but it would be too late: we would have 525 // changed its value earlier. Therefore, we manually save 526 // it here into another freely available register, 527 // `free_reg`, chosen of course among the caller-save 528 // registers (as a callee-save `free_reg` register would 529 // exhibit the same problem). 530 // 531 // Note we could have requested a temporary register from 532 // the register allocator instead; but we prefer not to, as 533 // this is a slow path, and we know we can find a 534 // caller-save register that is available. 535 Register free_reg = FindAvailableCallerSaveRegister(codegen); 536 __ Mov(free_reg, index_reg); 537 index_reg = free_reg; 538 index = Location::RegisterLocation(index_reg); 539 } else { 540 // The initial register stored in `index_` has already been 541 // saved in the call to art::SlowPathCode::SaveLiveRegisters 542 // (as it is not a callee-save register), so we can freely 543 // use it. 544 } 545 // Shifting the index value contained in `index_reg` by the scale 546 // factor (2) cannot overflow in practice, as the runtime is 547 // unable to allocate object arrays with a size larger than 548 // 2^26 - 1 (that is, 2^28 - 4 bytes). 549 __ Lsl(index_reg, index_reg, TIMES_4); 550 static_assert( 551 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 552 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 553 __ AddConstant(index_reg, index_reg, offset_); 554 } else { 555 DCHECK(instruction_->IsInvoke()); 556 DCHECK(instruction_->GetLocations()->Intrinsified()); 557 DCHECK((instruction_->AsInvoke()->GetIntrinsic() == Intrinsics::kUnsafeGetObject) || 558 (instruction_->AsInvoke()->GetIntrinsic() == Intrinsics::kUnsafeGetObjectVolatile)) 559 << instruction_->AsInvoke()->GetIntrinsic(); 560 DCHECK_EQ(offset_, 0U); 561 DCHECK(index_.IsRegisterPair()); 562 // UnsafeGet's offset location is a register pair, the low 563 // part contains the correct offset. 564 index = index_.ToLow(); 565 } 566 } 567 568 // We're moving two or three locations to locations that could 569 // overlap, so we need a parallel move resolver. 570 InvokeRuntimeCallingConvention calling_convention; 571 HParallelMove parallel_move(codegen->GetGraph()->GetArena()); 572 parallel_move.AddMove(ref_, 573 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 574 Primitive::kPrimNot, 575 nullptr); 576 parallel_move.AddMove(obj_, 577 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 578 Primitive::kPrimNot, 579 nullptr); 580 if (index.IsValid()) { 581 parallel_move.AddMove(index, 582 Location::RegisterLocation(calling_convention.GetRegisterAt(2)), 583 Primitive::kPrimInt, 584 nullptr); 585 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 586 } else { 587 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 588 __ LoadImmediate(calling_convention.GetRegisterAt(2), offset_); 589 } 590 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierSlow), 591 instruction_, 592 instruction_->GetDexPc(), 593 this); 594 CheckEntrypointTypes< 595 kQuickReadBarrierSlow, mirror::Object*, mirror::Object*, mirror::Object*, uint32_t>(); 596 arm_codegen->Move32(out_, Location::RegisterLocation(R0)); 597 598 RestoreLiveRegisters(codegen, locations); 599 __ b(GetExitLabel()); 600 } 601 602 const char* GetDescription() const OVERRIDE { return "ReadBarrierForHeapReferenceSlowPathARM"; } 603 604 private: 605 Register FindAvailableCallerSaveRegister(CodeGenerator* codegen) { 606 size_t ref = static_cast<int>(ref_.AsRegister<Register>()); 607 size_t obj = static_cast<int>(obj_.AsRegister<Register>()); 608 for (size_t i = 0, e = codegen->GetNumberOfCoreRegisters(); i < e; ++i) { 609 if (i != ref && i != obj && !codegen->IsCoreCalleeSaveRegister(i)) { 610 return static_cast<Register>(i); 611 } 612 } 613 // We shall never fail to find a free caller-save register, as 614 // there are more than two core caller-save registers on ARM 615 // (meaning it is possible to find one which is different from 616 // `ref` and `obj`). 617 DCHECK_GT(codegen->GetNumberOfCoreCallerSaveRegisters(), 2u); 618 LOG(FATAL) << "Could not find a free caller-save register"; 619 UNREACHABLE(); 620 } 621 622 const Location out_; 623 const Location ref_; 624 const Location obj_; 625 const uint32_t offset_; 626 // An additional location containing an index to an array. 627 // Only used for HArrayGet and the UnsafeGetObject & 628 // UnsafeGetObjectVolatile intrinsics. 629 const Location index_; 630 631 DISALLOW_COPY_AND_ASSIGN(ReadBarrierForHeapReferenceSlowPathARM); 632}; 633 634// Slow path generating a read barrier for a GC root. 635class ReadBarrierForRootSlowPathARM : public SlowPathCode { 636 public: 637 ReadBarrierForRootSlowPathARM(HInstruction* instruction, Location out, Location root) 638 : SlowPathCode(instruction), out_(out), root_(root) { 639 DCHECK(kEmitCompilerReadBarrier); 640 } 641 642 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 643 LocationSummary* locations = instruction_->GetLocations(); 644 Register reg_out = out_.AsRegister<Register>(); 645 DCHECK(locations->CanCall()); 646 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg_out)); 647 DCHECK(instruction_->IsLoadClass() || instruction_->IsLoadString()) 648 << "Unexpected instruction in read barrier for GC root slow path: " 649 << instruction_->DebugName(); 650 651 __ Bind(GetEntryLabel()); 652 SaveLiveRegisters(codegen, locations); 653 654 InvokeRuntimeCallingConvention calling_convention; 655 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 656 arm_codegen->Move32(Location::RegisterLocation(calling_convention.GetRegisterAt(0)), root_); 657 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierForRootSlow), 658 instruction_, 659 instruction_->GetDexPc(), 660 this); 661 CheckEntrypointTypes<kQuickReadBarrierForRootSlow, mirror::Object*, GcRoot<mirror::Object>*>(); 662 arm_codegen->Move32(out_, Location::RegisterLocation(R0)); 663 664 RestoreLiveRegisters(codegen, locations); 665 __ b(GetExitLabel()); 666 } 667 668 const char* GetDescription() const OVERRIDE { return "ReadBarrierForRootSlowPathARM"; } 669 670 private: 671 const Location out_; 672 const Location root_; 673 674 DISALLOW_COPY_AND_ASSIGN(ReadBarrierForRootSlowPathARM); 675}; 676 677#undef __ 678// NOLINT on __ macro to suppress wrong warning/fix from clang-tidy. 679#define __ down_cast<ArmAssembler*>(GetAssembler())-> // NOLINT 680 681inline Condition ARMCondition(IfCondition cond) { 682 switch (cond) { 683 case kCondEQ: return EQ; 684 case kCondNE: return NE; 685 case kCondLT: return LT; 686 case kCondLE: return LE; 687 case kCondGT: return GT; 688 case kCondGE: return GE; 689 case kCondB: return LO; 690 case kCondBE: return LS; 691 case kCondA: return HI; 692 case kCondAE: return HS; 693 } 694 LOG(FATAL) << "Unreachable"; 695 UNREACHABLE(); 696} 697 698// Maps signed condition to unsigned condition. 699inline Condition ARMUnsignedCondition(IfCondition cond) { 700 switch (cond) { 701 case kCondEQ: return EQ; 702 case kCondNE: return NE; 703 // Signed to unsigned. 704 case kCondLT: return LO; 705 case kCondLE: return LS; 706 case kCondGT: return HI; 707 case kCondGE: return HS; 708 // Unsigned remain unchanged. 709 case kCondB: return LO; 710 case kCondBE: return LS; 711 case kCondA: return HI; 712 case kCondAE: return HS; 713 } 714 LOG(FATAL) << "Unreachable"; 715 UNREACHABLE(); 716} 717 718inline Condition ARMFPCondition(IfCondition cond, bool gt_bias) { 719 // The ARM condition codes can express all the necessary branches, see the 720 // "Meaning (floating-point)" column in the table A8-1 of the ARMv7 reference manual. 721 // There is no dex instruction or HIR that would need the missing conditions 722 // "equal or unordered" or "not equal". 723 switch (cond) { 724 case kCondEQ: return EQ; 725 case kCondNE: return NE /* unordered */; 726 case kCondLT: return gt_bias ? CC : LT /* unordered */; 727 case kCondLE: return gt_bias ? LS : LE /* unordered */; 728 case kCondGT: return gt_bias ? HI /* unordered */ : GT; 729 case kCondGE: return gt_bias ? CS /* unordered */ : GE; 730 default: 731 LOG(FATAL) << "UNREACHABLE"; 732 UNREACHABLE(); 733 } 734} 735 736void CodeGeneratorARM::DumpCoreRegister(std::ostream& stream, int reg) const { 737 stream << Register(reg); 738} 739 740void CodeGeneratorARM::DumpFloatingPointRegister(std::ostream& stream, int reg) const { 741 stream << SRegister(reg); 742} 743 744size_t CodeGeneratorARM::SaveCoreRegister(size_t stack_index, uint32_t reg_id) { 745 __ StoreToOffset(kStoreWord, static_cast<Register>(reg_id), SP, stack_index); 746 return kArmWordSize; 747} 748 749size_t CodeGeneratorARM::RestoreCoreRegister(size_t stack_index, uint32_t reg_id) { 750 __ LoadFromOffset(kLoadWord, static_cast<Register>(reg_id), SP, stack_index); 751 return kArmWordSize; 752} 753 754size_t CodeGeneratorARM::SaveFloatingPointRegister(size_t stack_index, uint32_t reg_id) { 755 __ StoreSToOffset(static_cast<SRegister>(reg_id), SP, stack_index); 756 return kArmWordSize; 757} 758 759size_t CodeGeneratorARM::RestoreFloatingPointRegister(size_t stack_index, uint32_t reg_id) { 760 __ LoadSFromOffset(static_cast<SRegister>(reg_id), SP, stack_index); 761 return kArmWordSize; 762} 763 764CodeGeneratorARM::CodeGeneratorARM(HGraph* graph, 765 const ArmInstructionSetFeatures& isa_features, 766 const CompilerOptions& compiler_options, 767 OptimizingCompilerStats* stats) 768 : CodeGenerator(graph, 769 kNumberOfCoreRegisters, 770 kNumberOfSRegisters, 771 kNumberOfRegisterPairs, 772 ComputeRegisterMask(reinterpret_cast<const int*>(kCoreCalleeSaves), 773 arraysize(kCoreCalleeSaves)), 774 ComputeRegisterMask(reinterpret_cast<const int*>(kFpuCalleeSaves), 775 arraysize(kFpuCalleeSaves)), 776 compiler_options, 777 stats), 778 block_labels_(nullptr), 779 location_builder_(graph, this), 780 instruction_visitor_(graph, this), 781 move_resolver_(graph->GetArena(), this), 782 assembler_(graph->GetArena()), 783 isa_features_(isa_features), 784 uint32_literals_(std::less<uint32_t>(), 785 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 786 method_patches_(MethodReferenceComparator(), 787 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 788 call_patches_(MethodReferenceComparator(), 789 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 790 relative_call_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 791 pc_relative_dex_cache_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 792 boot_image_string_patches_(StringReferenceValueComparator(), 793 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 794 pc_relative_string_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 795 boot_image_address_patches_(std::less<uint32_t>(), 796 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)) { 797 // Always save the LR register to mimic Quick. 798 AddAllocatedRegister(Location::RegisterLocation(LR)); 799} 800 801void CodeGeneratorARM::Finalize(CodeAllocator* allocator) { 802 // Ensure that we fix up branches and literal loads and emit the literal pool. 803 __ FinalizeCode(); 804 805 // Adjust native pc offsets in stack maps. 806 for (size_t i = 0, num = stack_map_stream_.GetNumberOfStackMaps(); i != num; ++i) { 807 uint32_t old_position = stack_map_stream_.GetStackMap(i).native_pc_offset; 808 uint32_t new_position = __ GetAdjustedPosition(old_position); 809 stack_map_stream_.SetStackMapNativePcOffset(i, new_position); 810 } 811 // Adjust pc offsets for the disassembly information. 812 if (disasm_info_ != nullptr) { 813 GeneratedCodeInterval* frame_entry_interval = disasm_info_->GetFrameEntryInterval(); 814 frame_entry_interval->start = __ GetAdjustedPosition(frame_entry_interval->start); 815 frame_entry_interval->end = __ GetAdjustedPosition(frame_entry_interval->end); 816 for (auto& it : *disasm_info_->GetInstructionIntervals()) { 817 it.second.start = __ GetAdjustedPosition(it.second.start); 818 it.second.end = __ GetAdjustedPosition(it.second.end); 819 } 820 for (auto& it : *disasm_info_->GetSlowPathIntervals()) { 821 it.code_interval.start = __ GetAdjustedPosition(it.code_interval.start); 822 it.code_interval.end = __ GetAdjustedPosition(it.code_interval.end); 823 } 824 } 825 826 CodeGenerator::Finalize(allocator); 827} 828 829void CodeGeneratorARM::SetupBlockedRegisters() const { 830 // Don't allocate the dalvik style register pair passing. 831 blocked_register_pairs_[R1_R2] = true; 832 833 // Stack register, LR and PC are always reserved. 834 blocked_core_registers_[SP] = true; 835 blocked_core_registers_[LR] = true; 836 blocked_core_registers_[PC] = true; 837 838 // Reserve thread register. 839 blocked_core_registers_[TR] = true; 840 841 // Reserve temp register. 842 blocked_core_registers_[IP] = true; 843 844 if (GetGraph()->IsDebuggable()) { 845 // Stubs do not save callee-save floating point registers. If the graph 846 // is debuggable, we need to deal with these registers differently. For 847 // now, just block them. 848 for (size_t i = 0; i < arraysize(kFpuCalleeSaves); ++i) { 849 blocked_fpu_registers_[kFpuCalleeSaves[i]] = true; 850 } 851 } 852 853 UpdateBlockedPairRegisters(); 854} 855 856void CodeGeneratorARM::UpdateBlockedPairRegisters() const { 857 for (int i = 0; i < kNumberOfRegisterPairs; i++) { 858 ArmManagedRegister current = 859 ArmManagedRegister::FromRegisterPair(static_cast<RegisterPair>(i)); 860 if (blocked_core_registers_[current.AsRegisterPairLow()] 861 || blocked_core_registers_[current.AsRegisterPairHigh()]) { 862 blocked_register_pairs_[i] = true; 863 } 864 } 865} 866 867InstructionCodeGeneratorARM::InstructionCodeGeneratorARM(HGraph* graph, CodeGeneratorARM* codegen) 868 : InstructionCodeGenerator(graph, codegen), 869 assembler_(codegen->GetAssembler()), 870 codegen_(codegen) {} 871 872void CodeGeneratorARM::ComputeSpillMask() { 873 core_spill_mask_ = allocated_registers_.GetCoreRegisters() & core_callee_save_mask_; 874 DCHECK_NE(core_spill_mask_, 0u) << "At least the return address register must be saved"; 875 // There is no easy instruction to restore just the PC on thumb2. We spill and 876 // restore another arbitrary register. 877 core_spill_mask_ |= (1 << kCoreAlwaysSpillRegister); 878 fpu_spill_mask_ = allocated_registers_.GetFloatingPointRegisters() & fpu_callee_save_mask_; 879 // We use vpush and vpop for saving and restoring floating point registers, which take 880 // a SRegister and the number of registers to save/restore after that SRegister. We 881 // therefore update the `fpu_spill_mask_` to also contain those registers not allocated, 882 // but in the range. 883 if (fpu_spill_mask_ != 0) { 884 uint32_t least_significant_bit = LeastSignificantBit(fpu_spill_mask_); 885 uint32_t most_significant_bit = MostSignificantBit(fpu_spill_mask_); 886 for (uint32_t i = least_significant_bit + 1 ; i < most_significant_bit; ++i) { 887 fpu_spill_mask_ |= (1 << i); 888 } 889 } 890} 891 892static dwarf::Reg DWARFReg(Register reg) { 893 return dwarf::Reg::ArmCore(static_cast<int>(reg)); 894} 895 896static dwarf::Reg DWARFReg(SRegister reg) { 897 return dwarf::Reg::ArmFp(static_cast<int>(reg)); 898} 899 900void CodeGeneratorARM::GenerateFrameEntry() { 901 bool skip_overflow_check = 902 IsLeafMethod() && !FrameNeedsStackCheck(GetFrameSize(), InstructionSet::kArm); 903 DCHECK(GetCompilerOptions().GetImplicitStackOverflowChecks()); 904 __ Bind(&frame_entry_label_); 905 906 if (HasEmptyFrame()) { 907 return; 908 } 909 910 if (!skip_overflow_check) { 911 __ AddConstant(IP, SP, -static_cast<int32_t>(GetStackOverflowReservedBytes(kArm))); 912 __ LoadFromOffset(kLoadWord, IP, IP, 0); 913 RecordPcInfo(nullptr, 0); 914 } 915 916 __ PushList(core_spill_mask_); 917 __ cfi().AdjustCFAOffset(kArmWordSize * POPCOUNT(core_spill_mask_)); 918 __ cfi().RelOffsetForMany(DWARFReg(kMethodRegisterArgument), 0, core_spill_mask_, kArmWordSize); 919 if (fpu_spill_mask_ != 0) { 920 SRegister start_register = SRegister(LeastSignificantBit(fpu_spill_mask_)); 921 __ vpushs(start_register, POPCOUNT(fpu_spill_mask_)); 922 __ cfi().AdjustCFAOffset(kArmWordSize * POPCOUNT(fpu_spill_mask_)); 923 __ cfi().RelOffsetForMany(DWARFReg(S0), 0, fpu_spill_mask_, kArmWordSize); 924 } 925 int adjust = GetFrameSize() - FrameEntrySpillSize(); 926 __ AddConstant(SP, -adjust); 927 __ cfi().AdjustCFAOffset(adjust); 928 __ StoreToOffset(kStoreWord, kMethodRegisterArgument, SP, 0); 929} 930 931void CodeGeneratorARM::GenerateFrameExit() { 932 if (HasEmptyFrame()) { 933 __ bx(LR); 934 return; 935 } 936 __ cfi().RememberState(); 937 int adjust = GetFrameSize() - FrameEntrySpillSize(); 938 __ AddConstant(SP, adjust); 939 __ cfi().AdjustCFAOffset(-adjust); 940 if (fpu_spill_mask_ != 0) { 941 SRegister start_register = SRegister(LeastSignificantBit(fpu_spill_mask_)); 942 __ vpops(start_register, POPCOUNT(fpu_spill_mask_)); 943 __ cfi().AdjustCFAOffset(-kArmPointerSize * POPCOUNT(fpu_spill_mask_)); 944 __ cfi().RestoreMany(DWARFReg(SRegister(0)), fpu_spill_mask_); 945 } 946 // Pop LR into PC to return. 947 DCHECK_NE(core_spill_mask_ & (1 << LR), 0U); 948 uint32_t pop_mask = (core_spill_mask_ & (~(1 << LR))) | 1 << PC; 949 __ PopList(pop_mask); 950 __ cfi().RestoreState(); 951 __ cfi().DefCFAOffset(GetFrameSize()); 952} 953 954void CodeGeneratorARM::Bind(HBasicBlock* block) { 955 Label* label = GetLabelOf(block); 956 __ BindTrackedLabel(label); 957} 958 959Location InvokeDexCallingConventionVisitorARM::GetNextLocation(Primitive::Type type) { 960 switch (type) { 961 case Primitive::kPrimBoolean: 962 case Primitive::kPrimByte: 963 case Primitive::kPrimChar: 964 case Primitive::kPrimShort: 965 case Primitive::kPrimInt: 966 case Primitive::kPrimNot: { 967 uint32_t index = gp_index_++; 968 uint32_t stack_index = stack_index_++; 969 if (index < calling_convention.GetNumberOfRegisters()) { 970 return Location::RegisterLocation(calling_convention.GetRegisterAt(index)); 971 } else { 972 return Location::StackSlot(calling_convention.GetStackOffsetOf(stack_index)); 973 } 974 } 975 976 case Primitive::kPrimLong: { 977 uint32_t index = gp_index_; 978 uint32_t stack_index = stack_index_; 979 gp_index_ += 2; 980 stack_index_ += 2; 981 if (index + 1 < calling_convention.GetNumberOfRegisters()) { 982 if (calling_convention.GetRegisterAt(index) == R1) { 983 // Skip R1, and use R2_R3 instead. 984 gp_index_++; 985 index++; 986 } 987 } 988 if (index + 1 < calling_convention.GetNumberOfRegisters()) { 989 DCHECK_EQ(calling_convention.GetRegisterAt(index) + 1, 990 calling_convention.GetRegisterAt(index + 1)); 991 992 return Location::RegisterPairLocation(calling_convention.GetRegisterAt(index), 993 calling_convention.GetRegisterAt(index + 1)); 994 } else { 995 return Location::DoubleStackSlot(calling_convention.GetStackOffsetOf(stack_index)); 996 } 997 } 998 999 case Primitive::kPrimFloat: { 1000 uint32_t stack_index = stack_index_++; 1001 if (float_index_ % 2 == 0) { 1002 float_index_ = std::max(double_index_, float_index_); 1003 } 1004 if (float_index_ < calling_convention.GetNumberOfFpuRegisters()) { 1005 return Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(float_index_++)); 1006 } else { 1007 return Location::StackSlot(calling_convention.GetStackOffsetOf(stack_index)); 1008 } 1009 } 1010 1011 case Primitive::kPrimDouble: { 1012 double_index_ = std::max(double_index_, RoundUp(float_index_, 2)); 1013 uint32_t stack_index = stack_index_; 1014 stack_index_ += 2; 1015 if (double_index_ + 1 < calling_convention.GetNumberOfFpuRegisters()) { 1016 uint32_t index = double_index_; 1017 double_index_ += 2; 1018 Location result = Location::FpuRegisterPairLocation( 1019 calling_convention.GetFpuRegisterAt(index), 1020 calling_convention.GetFpuRegisterAt(index + 1)); 1021 DCHECK(ExpectedPairLayout(result)); 1022 return result; 1023 } else { 1024 return Location::DoubleStackSlot(calling_convention.GetStackOffsetOf(stack_index)); 1025 } 1026 } 1027 1028 case Primitive::kPrimVoid: 1029 LOG(FATAL) << "Unexpected parameter type " << type; 1030 break; 1031 } 1032 return Location::NoLocation(); 1033} 1034 1035Location InvokeDexCallingConventionVisitorARM::GetReturnLocation(Primitive::Type type) const { 1036 switch (type) { 1037 case Primitive::kPrimBoolean: 1038 case Primitive::kPrimByte: 1039 case Primitive::kPrimChar: 1040 case Primitive::kPrimShort: 1041 case Primitive::kPrimInt: 1042 case Primitive::kPrimNot: { 1043 return Location::RegisterLocation(R0); 1044 } 1045 1046 case Primitive::kPrimFloat: { 1047 return Location::FpuRegisterLocation(S0); 1048 } 1049 1050 case Primitive::kPrimLong: { 1051 return Location::RegisterPairLocation(R0, R1); 1052 } 1053 1054 case Primitive::kPrimDouble: { 1055 return Location::FpuRegisterPairLocation(S0, S1); 1056 } 1057 1058 case Primitive::kPrimVoid: 1059 return Location::NoLocation(); 1060 } 1061 1062 UNREACHABLE(); 1063} 1064 1065Location InvokeDexCallingConventionVisitorARM::GetMethodLocation() const { 1066 return Location::RegisterLocation(kMethodRegisterArgument); 1067} 1068 1069void CodeGeneratorARM::Move32(Location destination, Location source) { 1070 if (source.Equals(destination)) { 1071 return; 1072 } 1073 if (destination.IsRegister()) { 1074 if (source.IsRegister()) { 1075 __ Mov(destination.AsRegister<Register>(), source.AsRegister<Register>()); 1076 } else if (source.IsFpuRegister()) { 1077 __ vmovrs(destination.AsRegister<Register>(), source.AsFpuRegister<SRegister>()); 1078 } else { 1079 __ LoadFromOffset(kLoadWord, destination.AsRegister<Register>(), SP, source.GetStackIndex()); 1080 } 1081 } else if (destination.IsFpuRegister()) { 1082 if (source.IsRegister()) { 1083 __ vmovsr(destination.AsFpuRegister<SRegister>(), source.AsRegister<Register>()); 1084 } else if (source.IsFpuRegister()) { 1085 __ vmovs(destination.AsFpuRegister<SRegister>(), source.AsFpuRegister<SRegister>()); 1086 } else { 1087 __ LoadSFromOffset(destination.AsFpuRegister<SRegister>(), SP, source.GetStackIndex()); 1088 } 1089 } else { 1090 DCHECK(destination.IsStackSlot()) << destination; 1091 if (source.IsRegister()) { 1092 __ StoreToOffset(kStoreWord, source.AsRegister<Register>(), SP, destination.GetStackIndex()); 1093 } else if (source.IsFpuRegister()) { 1094 __ StoreSToOffset(source.AsFpuRegister<SRegister>(), SP, destination.GetStackIndex()); 1095 } else { 1096 DCHECK(source.IsStackSlot()) << source; 1097 __ LoadFromOffset(kLoadWord, IP, SP, source.GetStackIndex()); 1098 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 1099 } 1100 } 1101} 1102 1103void CodeGeneratorARM::Move64(Location destination, Location source) { 1104 if (source.Equals(destination)) { 1105 return; 1106 } 1107 if (destination.IsRegisterPair()) { 1108 if (source.IsRegisterPair()) { 1109 EmitParallelMoves( 1110 Location::RegisterLocation(source.AsRegisterPairHigh<Register>()), 1111 Location::RegisterLocation(destination.AsRegisterPairHigh<Register>()), 1112 Primitive::kPrimInt, 1113 Location::RegisterLocation(source.AsRegisterPairLow<Register>()), 1114 Location::RegisterLocation(destination.AsRegisterPairLow<Register>()), 1115 Primitive::kPrimInt); 1116 } else if (source.IsFpuRegister()) { 1117 UNIMPLEMENTED(FATAL); 1118 } else if (source.IsFpuRegisterPair()) { 1119 __ vmovrrd(destination.AsRegisterPairLow<Register>(), 1120 destination.AsRegisterPairHigh<Register>(), 1121 FromLowSToD(source.AsFpuRegisterPairLow<SRegister>())); 1122 } else { 1123 DCHECK(source.IsDoubleStackSlot()); 1124 DCHECK(ExpectedPairLayout(destination)); 1125 __ LoadFromOffset(kLoadWordPair, destination.AsRegisterPairLow<Register>(), 1126 SP, source.GetStackIndex()); 1127 } 1128 } else if (destination.IsFpuRegisterPair()) { 1129 if (source.IsDoubleStackSlot()) { 1130 __ LoadDFromOffset(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 1131 SP, 1132 source.GetStackIndex()); 1133 } else if (source.IsRegisterPair()) { 1134 __ vmovdrr(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 1135 source.AsRegisterPairLow<Register>(), 1136 source.AsRegisterPairHigh<Register>()); 1137 } else { 1138 UNIMPLEMENTED(FATAL); 1139 } 1140 } else { 1141 DCHECK(destination.IsDoubleStackSlot()); 1142 if (source.IsRegisterPair()) { 1143 // No conflict possible, so just do the moves. 1144 if (source.AsRegisterPairLow<Register>() == R1) { 1145 DCHECK_EQ(source.AsRegisterPairHigh<Register>(), R2); 1146 __ StoreToOffset(kStoreWord, R1, SP, destination.GetStackIndex()); 1147 __ StoreToOffset(kStoreWord, R2, SP, destination.GetHighStackIndex(kArmWordSize)); 1148 } else { 1149 __ StoreToOffset(kStoreWordPair, source.AsRegisterPairLow<Register>(), 1150 SP, destination.GetStackIndex()); 1151 } 1152 } else if (source.IsFpuRegisterPair()) { 1153 __ StoreDToOffset(FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()), 1154 SP, 1155 destination.GetStackIndex()); 1156 } else { 1157 DCHECK(source.IsDoubleStackSlot()); 1158 EmitParallelMoves( 1159 Location::StackSlot(source.GetStackIndex()), 1160 Location::StackSlot(destination.GetStackIndex()), 1161 Primitive::kPrimInt, 1162 Location::StackSlot(source.GetHighStackIndex(kArmWordSize)), 1163 Location::StackSlot(destination.GetHighStackIndex(kArmWordSize)), 1164 Primitive::kPrimInt); 1165 } 1166 } 1167} 1168 1169void CodeGeneratorARM::MoveConstant(Location location, int32_t value) { 1170 DCHECK(location.IsRegister()); 1171 __ LoadImmediate(location.AsRegister<Register>(), value); 1172} 1173 1174void CodeGeneratorARM::MoveLocation(Location dst, Location src, Primitive::Type dst_type) { 1175 HParallelMove move(GetGraph()->GetArena()); 1176 move.AddMove(src, dst, dst_type, nullptr); 1177 GetMoveResolver()->EmitNativeCode(&move); 1178} 1179 1180void CodeGeneratorARM::AddLocationAsTemp(Location location, LocationSummary* locations) { 1181 if (location.IsRegister()) { 1182 locations->AddTemp(location); 1183 } else if (location.IsRegisterPair()) { 1184 locations->AddTemp(Location::RegisterLocation(location.AsRegisterPairLow<Register>())); 1185 locations->AddTemp(Location::RegisterLocation(location.AsRegisterPairHigh<Register>())); 1186 } else { 1187 UNIMPLEMENTED(FATAL) << "AddLocationAsTemp not implemented for location " << location; 1188 } 1189} 1190 1191void CodeGeneratorARM::InvokeRuntime(QuickEntrypointEnum entrypoint, 1192 HInstruction* instruction, 1193 uint32_t dex_pc, 1194 SlowPathCode* slow_path) { 1195 InvokeRuntime(GetThreadOffset<kArmWordSize>(entrypoint).Int32Value(), 1196 instruction, 1197 dex_pc, 1198 slow_path); 1199} 1200 1201void CodeGeneratorARM::InvokeRuntime(int32_t entry_point_offset, 1202 HInstruction* instruction, 1203 uint32_t dex_pc, 1204 SlowPathCode* slow_path) { 1205 ValidateInvokeRuntime(instruction, slow_path); 1206 __ LoadFromOffset(kLoadWord, LR, TR, entry_point_offset); 1207 __ blx(LR); 1208 RecordPcInfo(instruction, dex_pc, slow_path); 1209} 1210 1211void InstructionCodeGeneratorARM::HandleGoto(HInstruction* got, HBasicBlock* successor) { 1212 DCHECK(!successor->IsExitBlock()); 1213 1214 HBasicBlock* block = got->GetBlock(); 1215 HInstruction* previous = got->GetPrevious(); 1216 1217 HLoopInformation* info = block->GetLoopInformation(); 1218 if (info != nullptr && info->IsBackEdge(*block) && info->HasSuspendCheck()) { 1219 codegen_->ClearSpillSlotsFromLoopPhisInStackMap(info->GetSuspendCheck()); 1220 GenerateSuspendCheck(info->GetSuspendCheck(), successor); 1221 return; 1222 } 1223 1224 if (block->IsEntryBlock() && (previous != nullptr) && previous->IsSuspendCheck()) { 1225 GenerateSuspendCheck(previous->AsSuspendCheck(), nullptr); 1226 } 1227 if (!codegen_->GoesToNextBlock(got->GetBlock(), successor)) { 1228 __ b(codegen_->GetLabelOf(successor)); 1229 } 1230} 1231 1232void LocationsBuilderARM::VisitGoto(HGoto* got) { 1233 got->SetLocations(nullptr); 1234} 1235 1236void InstructionCodeGeneratorARM::VisitGoto(HGoto* got) { 1237 HandleGoto(got, got->GetSuccessor()); 1238} 1239 1240void LocationsBuilderARM::VisitTryBoundary(HTryBoundary* try_boundary) { 1241 try_boundary->SetLocations(nullptr); 1242} 1243 1244void InstructionCodeGeneratorARM::VisitTryBoundary(HTryBoundary* try_boundary) { 1245 HBasicBlock* successor = try_boundary->GetNormalFlowSuccessor(); 1246 if (!successor->IsExitBlock()) { 1247 HandleGoto(try_boundary, successor); 1248 } 1249} 1250 1251void LocationsBuilderARM::VisitExit(HExit* exit) { 1252 exit->SetLocations(nullptr); 1253} 1254 1255void InstructionCodeGeneratorARM::VisitExit(HExit* exit ATTRIBUTE_UNUSED) { 1256} 1257 1258void InstructionCodeGeneratorARM::GenerateFPJumps(HCondition* cond, 1259 Label* true_label, 1260 Label* false_label ATTRIBUTE_UNUSED) { 1261 __ vmstat(); // transfer FP status register to ARM APSR. 1262 __ b(true_label, ARMFPCondition(cond->GetCondition(), cond->IsGtBias())); 1263} 1264 1265void InstructionCodeGeneratorARM::GenerateLongComparesAndJumps(HCondition* cond, 1266 Label* true_label, 1267 Label* false_label) { 1268 LocationSummary* locations = cond->GetLocations(); 1269 Location left = locations->InAt(0); 1270 Location right = locations->InAt(1); 1271 IfCondition if_cond = cond->GetCondition(); 1272 1273 Register left_high = left.AsRegisterPairHigh<Register>(); 1274 Register left_low = left.AsRegisterPairLow<Register>(); 1275 IfCondition true_high_cond = if_cond; 1276 IfCondition false_high_cond = cond->GetOppositeCondition(); 1277 Condition final_condition = ARMUnsignedCondition(if_cond); // unsigned on lower part 1278 1279 // Set the conditions for the test, remembering that == needs to be 1280 // decided using the low words. 1281 // TODO: consider avoiding jumps with temporary and CMP low+SBC high 1282 switch (if_cond) { 1283 case kCondEQ: 1284 case kCondNE: 1285 // Nothing to do. 1286 break; 1287 case kCondLT: 1288 false_high_cond = kCondGT; 1289 break; 1290 case kCondLE: 1291 true_high_cond = kCondLT; 1292 break; 1293 case kCondGT: 1294 false_high_cond = kCondLT; 1295 break; 1296 case kCondGE: 1297 true_high_cond = kCondGT; 1298 break; 1299 case kCondB: 1300 false_high_cond = kCondA; 1301 break; 1302 case kCondBE: 1303 true_high_cond = kCondB; 1304 break; 1305 case kCondA: 1306 false_high_cond = kCondB; 1307 break; 1308 case kCondAE: 1309 true_high_cond = kCondA; 1310 break; 1311 } 1312 if (right.IsConstant()) { 1313 int64_t value = right.GetConstant()->AsLongConstant()->GetValue(); 1314 int32_t val_low = Low32Bits(value); 1315 int32_t val_high = High32Bits(value); 1316 1317 __ CmpConstant(left_high, val_high); 1318 if (if_cond == kCondNE) { 1319 __ b(true_label, ARMCondition(true_high_cond)); 1320 } else if (if_cond == kCondEQ) { 1321 __ b(false_label, ARMCondition(false_high_cond)); 1322 } else { 1323 __ b(true_label, ARMCondition(true_high_cond)); 1324 __ b(false_label, ARMCondition(false_high_cond)); 1325 } 1326 // Must be equal high, so compare the lows. 1327 __ CmpConstant(left_low, val_low); 1328 } else { 1329 Register right_high = right.AsRegisterPairHigh<Register>(); 1330 Register right_low = right.AsRegisterPairLow<Register>(); 1331 1332 __ cmp(left_high, ShifterOperand(right_high)); 1333 if (if_cond == kCondNE) { 1334 __ b(true_label, ARMCondition(true_high_cond)); 1335 } else if (if_cond == kCondEQ) { 1336 __ b(false_label, ARMCondition(false_high_cond)); 1337 } else { 1338 __ b(true_label, ARMCondition(true_high_cond)); 1339 __ b(false_label, ARMCondition(false_high_cond)); 1340 } 1341 // Must be equal high, so compare the lows. 1342 __ cmp(left_low, ShifterOperand(right_low)); 1343 } 1344 // The last comparison might be unsigned. 1345 // TODO: optimize cases where this is always true/false 1346 __ b(true_label, final_condition); 1347} 1348 1349void InstructionCodeGeneratorARM::GenerateCompareTestAndBranch(HCondition* condition, 1350 Label* true_target_in, 1351 Label* false_target_in) { 1352 // Generated branching requires both targets to be explicit. If either of the 1353 // targets is nullptr (fallthrough) use and bind `fallthrough_target` instead. 1354 Label fallthrough_target; 1355 Label* true_target = true_target_in == nullptr ? &fallthrough_target : true_target_in; 1356 Label* false_target = false_target_in == nullptr ? &fallthrough_target : false_target_in; 1357 1358 LocationSummary* locations = condition->GetLocations(); 1359 Location left = locations->InAt(0); 1360 Location right = locations->InAt(1); 1361 1362 Primitive::Type type = condition->InputAt(0)->GetType(); 1363 switch (type) { 1364 case Primitive::kPrimLong: 1365 GenerateLongComparesAndJumps(condition, true_target, false_target); 1366 break; 1367 case Primitive::kPrimFloat: 1368 __ vcmps(left.AsFpuRegister<SRegister>(), right.AsFpuRegister<SRegister>()); 1369 GenerateFPJumps(condition, true_target, false_target); 1370 break; 1371 case Primitive::kPrimDouble: 1372 __ vcmpd(FromLowSToD(left.AsFpuRegisterPairLow<SRegister>()), 1373 FromLowSToD(right.AsFpuRegisterPairLow<SRegister>())); 1374 GenerateFPJumps(condition, true_target, false_target); 1375 break; 1376 default: 1377 LOG(FATAL) << "Unexpected compare type " << type; 1378 } 1379 1380 if (false_target != &fallthrough_target) { 1381 __ b(false_target); 1382 } 1383 1384 if (fallthrough_target.IsLinked()) { 1385 __ Bind(&fallthrough_target); 1386 } 1387} 1388 1389void InstructionCodeGeneratorARM::GenerateTestAndBranch(HInstruction* instruction, 1390 size_t condition_input_index, 1391 Label* true_target, 1392 Label* false_target) { 1393 HInstruction* cond = instruction->InputAt(condition_input_index); 1394 1395 if (true_target == nullptr && false_target == nullptr) { 1396 // Nothing to do. The code always falls through. 1397 return; 1398 } else if (cond->IsIntConstant()) { 1399 // Constant condition, statically compared against "true" (integer value 1). 1400 if (cond->AsIntConstant()->IsTrue()) { 1401 if (true_target != nullptr) { 1402 __ b(true_target); 1403 } 1404 } else { 1405 DCHECK(cond->AsIntConstant()->IsFalse()) << cond->AsIntConstant()->GetValue(); 1406 if (false_target != nullptr) { 1407 __ b(false_target); 1408 } 1409 } 1410 return; 1411 } 1412 1413 // The following code generates these patterns: 1414 // (1) true_target == nullptr && false_target != nullptr 1415 // - opposite condition true => branch to false_target 1416 // (2) true_target != nullptr && false_target == nullptr 1417 // - condition true => branch to true_target 1418 // (3) true_target != nullptr && false_target != nullptr 1419 // - condition true => branch to true_target 1420 // - branch to false_target 1421 if (IsBooleanValueOrMaterializedCondition(cond)) { 1422 // Condition has been materialized, compare the output to 0. 1423 Location cond_val = instruction->GetLocations()->InAt(condition_input_index); 1424 DCHECK(cond_val.IsRegister()); 1425 if (true_target == nullptr) { 1426 __ CompareAndBranchIfZero(cond_val.AsRegister<Register>(), false_target); 1427 } else { 1428 __ CompareAndBranchIfNonZero(cond_val.AsRegister<Register>(), true_target); 1429 } 1430 } else { 1431 // Condition has not been materialized. Use its inputs as the comparison and 1432 // its condition as the branch condition. 1433 HCondition* condition = cond->AsCondition(); 1434 1435 // If this is a long or FP comparison that has been folded into 1436 // the HCondition, generate the comparison directly. 1437 Primitive::Type type = condition->InputAt(0)->GetType(); 1438 if (type == Primitive::kPrimLong || Primitive::IsFloatingPointType(type)) { 1439 GenerateCompareTestAndBranch(condition, true_target, false_target); 1440 return; 1441 } 1442 1443 LocationSummary* locations = cond->GetLocations(); 1444 DCHECK(locations->InAt(0).IsRegister()); 1445 Register left = locations->InAt(0).AsRegister<Register>(); 1446 Location right = locations->InAt(1); 1447 if (right.IsRegister()) { 1448 __ cmp(left, ShifterOperand(right.AsRegister<Register>())); 1449 } else { 1450 DCHECK(right.IsConstant()); 1451 __ CmpConstant(left, CodeGenerator::GetInt32ValueOf(right.GetConstant())); 1452 } 1453 if (true_target == nullptr) { 1454 __ b(false_target, ARMCondition(condition->GetOppositeCondition())); 1455 } else { 1456 __ b(true_target, ARMCondition(condition->GetCondition())); 1457 } 1458 } 1459 1460 // If neither branch falls through (case 3), the conditional branch to `true_target` 1461 // was already emitted (case 2) and we need to emit a jump to `false_target`. 1462 if (true_target != nullptr && false_target != nullptr) { 1463 __ b(false_target); 1464 } 1465} 1466 1467void LocationsBuilderARM::VisitIf(HIf* if_instr) { 1468 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(if_instr); 1469 if (IsBooleanValueOrMaterializedCondition(if_instr->InputAt(0))) { 1470 locations->SetInAt(0, Location::RequiresRegister()); 1471 } 1472} 1473 1474void InstructionCodeGeneratorARM::VisitIf(HIf* if_instr) { 1475 HBasicBlock* true_successor = if_instr->IfTrueSuccessor(); 1476 HBasicBlock* false_successor = if_instr->IfFalseSuccessor(); 1477 Label* true_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), true_successor) ? 1478 nullptr : codegen_->GetLabelOf(true_successor); 1479 Label* false_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), false_successor) ? 1480 nullptr : codegen_->GetLabelOf(false_successor); 1481 GenerateTestAndBranch(if_instr, /* condition_input_index */ 0, true_target, false_target); 1482} 1483 1484void LocationsBuilderARM::VisitDeoptimize(HDeoptimize* deoptimize) { 1485 LocationSummary* locations = new (GetGraph()->GetArena()) 1486 LocationSummary(deoptimize, LocationSummary::kCallOnSlowPath); 1487 if (IsBooleanValueOrMaterializedCondition(deoptimize->InputAt(0))) { 1488 locations->SetInAt(0, Location::RequiresRegister()); 1489 } 1490} 1491 1492void InstructionCodeGeneratorARM::VisitDeoptimize(HDeoptimize* deoptimize) { 1493 SlowPathCode* slow_path = deopt_slow_paths_.NewSlowPath<DeoptimizationSlowPathARM>(deoptimize); 1494 GenerateTestAndBranch(deoptimize, 1495 /* condition_input_index */ 0, 1496 slow_path->GetEntryLabel(), 1497 /* false_target */ nullptr); 1498} 1499 1500void LocationsBuilderARM::VisitSelect(HSelect* select) { 1501 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(select); 1502 if (Primitive::IsFloatingPointType(select->GetType())) { 1503 locations->SetInAt(0, Location::RequiresFpuRegister()); 1504 locations->SetInAt(1, Location::RequiresFpuRegister()); 1505 } else { 1506 locations->SetInAt(0, Location::RequiresRegister()); 1507 locations->SetInAt(1, Location::RequiresRegister()); 1508 } 1509 if (IsBooleanValueOrMaterializedCondition(select->GetCondition())) { 1510 locations->SetInAt(2, Location::RequiresRegister()); 1511 } 1512 locations->SetOut(Location::SameAsFirstInput()); 1513} 1514 1515void InstructionCodeGeneratorARM::VisitSelect(HSelect* select) { 1516 LocationSummary* locations = select->GetLocations(); 1517 Label false_target; 1518 GenerateTestAndBranch(select, 1519 /* condition_input_index */ 2, 1520 /* true_target */ nullptr, 1521 &false_target); 1522 codegen_->MoveLocation(locations->Out(), locations->InAt(1), select->GetType()); 1523 __ Bind(&false_target); 1524} 1525 1526void LocationsBuilderARM::VisitNativeDebugInfo(HNativeDebugInfo* info) { 1527 new (GetGraph()->GetArena()) LocationSummary(info); 1528} 1529 1530void InstructionCodeGeneratorARM::VisitNativeDebugInfo(HNativeDebugInfo*) { 1531 // MaybeRecordNativeDebugInfo is already called implicitly in CodeGenerator::Compile. 1532} 1533 1534void CodeGeneratorARM::GenerateNop() { 1535 __ nop(); 1536} 1537 1538void LocationsBuilderARM::HandleCondition(HCondition* cond) { 1539 LocationSummary* locations = 1540 new (GetGraph()->GetArena()) LocationSummary(cond, LocationSummary::kNoCall); 1541 // Handle the long/FP comparisons made in instruction simplification. 1542 switch (cond->InputAt(0)->GetType()) { 1543 case Primitive::kPrimLong: 1544 locations->SetInAt(0, Location::RequiresRegister()); 1545 locations->SetInAt(1, Location::RegisterOrConstant(cond->InputAt(1))); 1546 if (!cond->IsEmittedAtUseSite()) { 1547 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 1548 } 1549 break; 1550 1551 case Primitive::kPrimFloat: 1552 case Primitive::kPrimDouble: 1553 locations->SetInAt(0, Location::RequiresFpuRegister()); 1554 locations->SetInAt(1, Location::RequiresFpuRegister()); 1555 if (!cond->IsEmittedAtUseSite()) { 1556 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 1557 } 1558 break; 1559 1560 default: 1561 locations->SetInAt(0, Location::RequiresRegister()); 1562 locations->SetInAt(1, Location::RegisterOrConstant(cond->InputAt(1))); 1563 if (!cond->IsEmittedAtUseSite()) { 1564 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 1565 } 1566 } 1567} 1568 1569void InstructionCodeGeneratorARM::HandleCondition(HCondition* cond) { 1570 if (cond->IsEmittedAtUseSite()) { 1571 return; 1572 } 1573 1574 LocationSummary* locations = cond->GetLocations(); 1575 Location left = locations->InAt(0); 1576 Location right = locations->InAt(1); 1577 Register out = locations->Out().AsRegister<Register>(); 1578 Label true_label, false_label; 1579 1580 switch (cond->InputAt(0)->GetType()) { 1581 default: { 1582 // Integer case. 1583 if (right.IsRegister()) { 1584 __ cmp(left.AsRegister<Register>(), ShifterOperand(right.AsRegister<Register>())); 1585 } else { 1586 DCHECK(right.IsConstant()); 1587 __ CmpConstant(left.AsRegister<Register>(), 1588 CodeGenerator::GetInt32ValueOf(right.GetConstant())); 1589 } 1590 __ it(ARMCondition(cond->GetCondition()), kItElse); 1591 __ mov(locations->Out().AsRegister<Register>(), ShifterOperand(1), 1592 ARMCondition(cond->GetCondition())); 1593 __ mov(locations->Out().AsRegister<Register>(), ShifterOperand(0), 1594 ARMCondition(cond->GetOppositeCondition())); 1595 return; 1596 } 1597 case Primitive::kPrimLong: 1598 GenerateLongComparesAndJumps(cond, &true_label, &false_label); 1599 break; 1600 case Primitive::kPrimFloat: 1601 __ vcmps(left.AsFpuRegister<SRegister>(), right.AsFpuRegister<SRegister>()); 1602 GenerateFPJumps(cond, &true_label, &false_label); 1603 break; 1604 case Primitive::kPrimDouble: 1605 __ vcmpd(FromLowSToD(left.AsFpuRegisterPairLow<SRegister>()), 1606 FromLowSToD(right.AsFpuRegisterPairLow<SRegister>())); 1607 GenerateFPJumps(cond, &true_label, &false_label); 1608 break; 1609 } 1610 1611 // Convert the jumps into the result. 1612 Label done_label; 1613 1614 // False case: result = 0. 1615 __ Bind(&false_label); 1616 __ LoadImmediate(out, 0); 1617 __ b(&done_label); 1618 1619 // True case: result = 1. 1620 __ Bind(&true_label); 1621 __ LoadImmediate(out, 1); 1622 __ Bind(&done_label); 1623} 1624 1625void LocationsBuilderARM::VisitEqual(HEqual* comp) { 1626 HandleCondition(comp); 1627} 1628 1629void InstructionCodeGeneratorARM::VisitEqual(HEqual* comp) { 1630 HandleCondition(comp); 1631} 1632 1633void LocationsBuilderARM::VisitNotEqual(HNotEqual* comp) { 1634 HandleCondition(comp); 1635} 1636 1637void InstructionCodeGeneratorARM::VisitNotEqual(HNotEqual* comp) { 1638 HandleCondition(comp); 1639} 1640 1641void LocationsBuilderARM::VisitLessThan(HLessThan* comp) { 1642 HandleCondition(comp); 1643} 1644 1645void InstructionCodeGeneratorARM::VisitLessThan(HLessThan* comp) { 1646 HandleCondition(comp); 1647} 1648 1649void LocationsBuilderARM::VisitLessThanOrEqual(HLessThanOrEqual* comp) { 1650 HandleCondition(comp); 1651} 1652 1653void InstructionCodeGeneratorARM::VisitLessThanOrEqual(HLessThanOrEqual* comp) { 1654 HandleCondition(comp); 1655} 1656 1657void LocationsBuilderARM::VisitGreaterThan(HGreaterThan* comp) { 1658 HandleCondition(comp); 1659} 1660 1661void InstructionCodeGeneratorARM::VisitGreaterThan(HGreaterThan* comp) { 1662 HandleCondition(comp); 1663} 1664 1665void LocationsBuilderARM::VisitGreaterThanOrEqual(HGreaterThanOrEqual* comp) { 1666 HandleCondition(comp); 1667} 1668 1669void InstructionCodeGeneratorARM::VisitGreaterThanOrEqual(HGreaterThanOrEqual* comp) { 1670 HandleCondition(comp); 1671} 1672 1673void LocationsBuilderARM::VisitBelow(HBelow* comp) { 1674 HandleCondition(comp); 1675} 1676 1677void InstructionCodeGeneratorARM::VisitBelow(HBelow* comp) { 1678 HandleCondition(comp); 1679} 1680 1681void LocationsBuilderARM::VisitBelowOrEqual(HBelowOrEqual* comp) { 1682 HandleCondition(comp); 1683} 1684 1685void InstructionCodeGeneratorARM::VisitBelowOrEqual(HBelowOrEqual* comp) { 1686 HandleCondition(comp); 1687} 1688 1689void LocationsBuilderARM::VisitAbove(HAbove* comp) { 1690 HandleCondition(comp); 1691} 1692 1693void InstructionCodeGeneratorARM::VisitAbove(HAbove* comp) { 1694 HandleCondition(comp); 1695} 1696 1697void LocationsBuilderARM::VisitAboveOrEqual(HAboveOrEqual* comp) { 1698 HandleCondition(comp); 1699} 1700 1701void InstructionCodeGeneratorARM::VisitAboveOrEqual(HAboveOrEqual* comp) { 1702 HandleCondition(comp); 1703} 1704 1705void LocationsBuilderARM::VisitIntConstant(HIntConstant* constant) { 1706 LocationSummary* locations = 1707 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1708 locations->SetOut(Location::ConstantLocation(constant)); 1709} 1710 1711void InstructionCodeGeneratorARM::VisitIntConstant(HIntConstant* constant ATTRIBUTE_UNUSED) { 1712 // Will be generated at use site. 1713} 1714 1715void LocationsBuilderARM::VisitNullConstant(HNullConstant* constant) { 1716 LocationSummary* locations = 1717 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1718 locations->SetOut(Location::ConstantLocation(constant)); 1719} 1720 1721void InstructionCodeGeneratorARM::VisitNullConstant(HNullConstant* constant ATTRIBUTE_UNUSED) { 1722 // Will be generated at use site. 1723} 1724 1725void LocationsBuilderARM::VisitLongConstant(HLongConstant* constant) { 1726 LocationSummary* locations = 1727 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1728 locations->SetOut(Location::ConstantLocation(constant)); 1729} 1730 1731void InstructionCodeGeneratorARM::VisitLongConstant(HLongConstant* constant ATTRIBUTE_UNUSED) { 1732 // Will be generated at use site. 1733} 1734 1735void LocationsBuilderARM::VisitFloatConstant(HFloatConstant* constant) { 1736 LocationSummary* locations = 1737 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1738 locations->SetOut(Location::ConstantLocation(constant)); 1739} 1740 1741void InstructionCodeGeneratorARM::VisitFloatConstant(HFloatConstant* constant ATTRIBUTE_UNUSED) { 1742 // Will be generated at use site. 1743} 1744 1745void LocationsBuilderARM::VisitDoubleConstant(HDoubleConstant* constant) { 1746 LocationSummary* locations = 1747 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1748 locations->SetOut(Location::ConstantLocation(constant)); 1749} 1750 1751void InstructionCodeGeneratorARM::VisitDoubleConstant(HDoubleConstant* constant ATTRIBUTE_UNUSED) { 1752 // Will be generated at use site. 1753} 1754 1755void LocationsBuilderARM::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) { 1756 memory_barrier->SetLocations(nullptr); 1757} 1758 1759void InstructionCodeGeneratorARM::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) { 1760 codegen_->GenerateMemoryBarrier(memory_barrier->GetBarrierKind()); 1761} 1762 1763void LocationsBuilderARM::VisitReturnVoid(HReturnVoid* ret) { 1764 ret->SetLocations(nullptr); 1765} 1766 1767void InstructionCodeGeneratorARM::VisitReturnVoid(HReturnVoid* ret ATTRIBUTE_UNUSED) { 1768 codegen_->GenerateFrameExit(); 1769} 1770 1771void LocationsBuilderARM::VisitReturn(HReturn* ret) { 1772 LocationSummary* locations = 1773 new (GetGraph()->GetArena()) LocationSummary(ret, LocationSummary::kNoCall); 1774 locations->SetInAt(0, parameter_visitor_.GetReturnLocation(ret->InputAt(0)->GetType())); 1775} 1776 1777void InstructionCodeGeneratorARM::VisitReturn(HReturn* ret ATTRIBUTE_UNUSED) { 1778 codegen_->GenerateFrameExit(); 1779} 1780 1781void LocationsBuilderARM::VisitInvokeUnresolved(HInvokeUnresolved* invoke) { 1782 // The trampoline uses the same calling convention as dex calling conventions, 1783 // except instead of loading arg0/r0 with the target Method*, arg0/r0 will contain 1784 // the method_idx. 1785 HandleInvoke(invoke); 1786} 1787 1788void InstructionCodeGeneratorARM::VisitInvokeUnresolved(HInvokeUnresolved* invoke) { 1789 codegen_->GenerateInvokeUnresolvedRuntimeCall(invoke); 1790} 1791 1792void LocationsBuilderARM::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) { 1793 // Explicit clinit checks triggered by static invokes must have been pruned by 1794 // art::PrepareForRegisterAllocation. 1795 DCHECK(!invoke->IsStaticWithExplicitClinitCheck()); 1796 1797 IntrinsicLocationsBuilderARM intrinsic(GetGraph()->GetArena(), 1798 codegen_->GetAssembler(), 1799 codegen_->GetInstructionSetFeatures()); 1800 if (intrinsic.TryDispatch(invoke)) { 1801 if (invoke->GetLocations()->CanCall() && invoke->HasPcRelativeDexCache()) { 1802 invoke->GetLocations()->SetInAt(invoke->GetSpecialInputIndex(), Location::Any()); 1803 } 1804 return; 1805 } 1806 1807 HandleInvoke(invoke); 1808 1809 // For PC-relative dex cache the invoke has an extra input, the PC-relative address base. 1810 if (invoke->HasPcRelativeDexCache()) { 1811 invoke->GetLocations()->SetInAt(invoke->GetSpecialInputIndex(), Location::RequiresRegister()); 1812 } 1813} 1814 1815static bool TryGenerateIntrinsicCode(HInvoke* invoke, CodeGeneratorARM* codegen) { 1816 if (invoke->GetLocations()->Intrinsified()) { 1817 IntrinsicCodeGeneratorARM intrinsic(codegen); 1818 intrinsic.Dispatch(invoke); 1819 return true; 1820 } 1821 return false; 1822} 1823 1824void InstructionCodeGeneratorARM::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) { 1825 // Explicit clinit checks triggered by static invokes must have been pruned by 1826 // art::PrepareForRegisterAllocation. 1827 DCHECK(!invoke->IsStaticWithExplicitClinitCheck()); 1828 1829 if (TryGenerateIntrinsicCode(invoke, codegen_)) { 1830 return; 1831 } 1832 1833 LocationSummary* locations = invoke->GetLocations(); 1834 codegen_->GenerateStaticOrDirectCall( 1835 invoke, locations->HasTemps() ? locations->GetTemp(0) : Location::NoLocation()); 1836 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 1837} 1838 1839void LocationsBuilderARM::HandleInvoke(HInvoke* invoke) { 1840 InvokeDexCallingConventionVisitorARM calling_convention_visitor; 1841 CodeGenerator::CreateCommonInvokeLocationSummary(invoke, &calling_convention_visitor); 1842} 1843 1844void LocationsBuilderARM::VisitInvokeVirtual(HInvokeVirtual* invoke) { 1845 IntrinsicLocationsBuilderARM intrinsic(GetGraph()->GetArena(), 1846 codegen_->GetAssembler(), 1847 codegen_->GetInstructionSetFeatures()); 1848 if (intrinsic.TryDispatch(invoke)) { 1849 return; 1850 } 1851 1852 HandleInvoke(invoke); 1853} 1854 1855void InstructionCodeGeneratorARM::VisitInvokeVirtual(HInvokeVirtual* invoke) { 1856 if (TryGenerateIntrinsicCode(invoke, codegen_)) { 1857 return; 1858 } 1859 1860 codegen_->GenerateVirtualCall(invoke, invoke->GetLocations()->GetTemp(0)); 1861 DCHECK(!codegen_->IsLeafMethod()); 1862 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 1863} 1864 1865void LocationsBuilderARM::VisitInvokeInterface(HInvokeInterface* invoke) { 1866 HandleInvoke(invoke); 1867 // Add the hidden argument. 1868 invoke->GetLocations()->AddTemp(Location::RegisterLocation(R12)); 1869} 1870 1871void InstructionCodeGeneratorARM::VisitInvokeInterface(HInvokeInterface* invoke) { 1872 // TODO: b/18116999, our IMTs can miss an IncompatibleClassChangeError. 1873 LocationSummary* locations = invoke->GetLocations(); 1874 Register temp = locations->GetTemp(0).AsRegister<Register>(); 1875 Register hidden_reg = locations->GetTemp(1).AsRegister<Register>(); 1876 Location receiver = locations->InAt(0); 1877 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 1878 1879 // Set the hidden argument. This is safe to do this here, as R12 1880 // won't be modified thereafter, before the `blx` (call) instruction. 1881 DCHECK_EQ(R12, hidden_reg); 1882 __ LoadImmediate(hidden_reg, invoke->GetDexMethodIndex()); 1883 1884 if (receiver.IsStackSlot()) { 1885 __ LoadFromOffset(kLoadWord, temp, SP, receiver.GetStackIndex()); 1886 // /* HeapReference<Class> */ temp = temp->klass_ 1887 __ LoadFromOffset(kLoadWord, temp, temp, class_offset); 1888 } else { 1889 // /* HeapReference<Class> */ temp = receiver->klass_ 1890 __ LoadFromOffset(kLoadWord, temp, receiver.AsRegister<Register>(), class_offset); 1891 } 1892 codegen_->MaybeRecordImplicitNullCheck(invoke); 1893 // Instead of simply (possibly) unpoisoning `temp` here, we should 1894 // emit a read barrier for the previous class reference load. 1895 // However this is not required in practice, as this is an 1896 // intermediate/temporary reference and because the current 1897 // concurrent copying collector keeps the from-space memory 1898 // intact/accessible until the end of the marking phase (the 1899 // concurrent copying collector may not in the future). 1900 __ MaybeUnpoisonHeapReference(temp); 1901 __ LoadFromOffset(kLoadWord, temp, temp, 1902 mirror::Class::ImtPtrOffset(kArmPointerSize).Uint32Value()); 1903 uint32_t method_offset = static_cast<uint32_t>(ImTable::OffsetOfElement( 1904 invoke->GetImtIndex() % ImTable::kSize, kArmPointerSize)); 1905 // temp = temp->GetImtEntryAt(method_offset); 1906 __ LoadFromOffset(kLoadWord, temp, temp, method_offset); 1907 uint32_t entry_point = 1908 ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmWordSize).Int32Value(); 1909 // LR = temp->GetEntryPoint(); 1910 __ LoadFromOffset(kLoadWord, LR, temp, entry_point); 1911 // LR(); 1912 __ blx(LR); 1913 DCHECK(!codegen_->IsLeafMethod()); 1914 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 1915} 1916 1917void LocationsBuilderARM::VisitNeg(HNeg* neg) { 1918 LocationSummary* locations = 1919 new (GetGraph()->GetArena()) LocationSummary(neg, LocationSummary::kNoCall); 1920 switch (neg->GetResultType()) { 1921 case Primitive::kPrimInt: { 1922 locations->SetInAt(0, Location::RequiresRegister()); 1923 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 1924 break; 1925 } 1926 case Primitive::kPrimLong: { 1927 locations->SetInAt(0, Location::RequiresRegister()); 1928 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 1929 break; 1930 } 1931 1932 case Primitive::kPrimFloat: 1933 case Primitive::kPrimDouble: 1934 locations->SetInAt(0, Location::RequiresFpuRegister()); 1935 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 1936 break; 1937 1938 default: 1939 LOG(FATAL) << "Unexpected neg type " << neg->GetResultType(); 1940 } 1941} 1942 1943void InstructionCodeGeneratorARM::VisitNeg(HNeg* neg) { 1944 LocationSummary* locations = neg->GetLocations(); 1945 Location out = locations->Out(); 1946 Location in = locations->InAt(0); 1947 switch (neg->GetResultType()) { 1948 case Primitive::kPrimInt: 1949 DCHECK(in.IsRegister()); 1950 __ rsb(out.AsRegister<Register>(), in.AsRegister<Register>(), ShifterOperand(0)); 1951 break; 1952 1953 case Primitive::kPrimLong: 1954 DCHECK(in.IsRegisterPair()); 1955 // out.lo = 0 - in.lo (and update the carry/borrow (C) flag) 1956 __ rsbs(out.AsRegisterPairLow<Register>(), 1957 in.AsRegisterPairLow<Register>(), 1958 ShifterOperand(0)); 1959 // We cannot emit an RSC (Reverse Subtract with Carry) 1960 // instruction here, as it does not exist in the Thumb-2 1961 // instruction set. We use the following approach 1962 // using SBC and SUB instead. 1963 // 1964 // out.hi = -C 1965 __ sbc(out.AsRegisterPairHigh<Register>(), 1966 out.AsRegisterPairHigh<Register>(), 1967 ShifterOperand(out.AsRegisterPairHigh<Register>())); 1968 // out.hi = out.hi - in.hi 1969 __ sub(out.AsRegisterPairHigh<Register>(), 1970 out.AsRegisterPairHigh<Register>(), 1971 ShifterOperand(in.AsRegisterPairHigh<Register>())); 1972 break; 1973 1974 case Primitive::kPrimFloat: 1975 DCHECK(in.IsFpuRegister()); 1976 __ vnegs(out.AsFpuRegister<SRegister>(), in.AsFpuRegister<SRegister>()); 1977 break; 1978 1979 case Primitive::kPrimDouble: 1980 DCHECK(in.IsFpuRegisterPair()); 1981 __ vnegd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 1982 FromLowSToD(in.AsFpuRegisterPairLow<SRegister>())); 1983 break; 1984 1985 default: 1986 LOG(FATAL) << "Unexpected neg type " << neg->GetResultType(); 1987 } 1988} 1989 1990void LocationsBuilderARM::VisitTypeConversion(HTypeConversion* conversion) { 1991 Primitive::Type result_type = conversion->GetResultType(); 1992 Primitive::Type input_type = conversion->GetInputType(); 1993 DCHECK_NE(result_type, input_type); 1994 1995 // The float-to-long, double-to-long and long-to-float type conversions 1996 // rely on a call to the runtime. 1997 LocationSummary::CallKind call_kind = 1998 (((input_type == Primitive::kPrimFloat || input_type == Primitive::kPrimDouble) 1999 && result_type == Primitive::kPrimLong) 2000 || (input_type == Primitive::kPrimLong && result_type == Primitive::kPrimFloat)) 2001 ? LocationSummary::kCall 2002 : LocationSummary::kNoCall; 2003 LocationSummary* locations = 2004 new (GetGraph()->GetArena()) LocationSummary(conversion, call_kind); 2005 2006 // The Java language does not allow treating boolean as an integral type but 2007 // our bit representation makes it safe. 2008 2009 switch (result_type) { 2010 case Primitive::kPrimByte: 2011 switch (input_type) { 2012 case Primitive::kPrimLong: 2013 // Type conversion from long to byte is a result of code transformations. 2014 case Primitive::kPrimBoolean: 2015 // Boolean input is a result of code transformations. 2016 case Primitive::kPrimShort: 2017 case Primitive::kPrimInt: 2018 case Primitive::kPrimChar: 2019 // Processing a Dex `int-to-byte' instruction. 2020 locations->SetInAt(0, Location::RequiresRegister()); 2021 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2022 break; 2023 2024 default: 2025 LOG(FATAL) << "Unexpected type conversion from " << input_type 2026 << " to " << result_type; 2027 } 2028 break; 2029 2030 case Primitive::kPrimShort: 2031 switch (input_type) { 2032 case Primitive::kPrimLong: 2033 // Type conversion from long to short is a result of code transformations. 2034 case Primitive::kPrimBoolean: 2035 // Boolean input is a result of code transformations. 2036 case Primitive::kPrimByte: 2037 case Primitive::kPrimInt: 2038 case Primitive::kPrimChar: 2039 // Processing a Dex `int-to-short' instruction. 2040 locations->SetInAt(0, Location::RequiresRegister()); 2041 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2042 break; 2043 2044 default: 2045 LOG(FATAL) << "Unexpected type conversion from " << input_type 2046 << " to " << result_type; 2047 } 2048 break; 2049 2050 case Primitive::kPrimInt: 2051 switch (input_type) { 2052 case Primitive::kPrimLong: 2053 // Processing a Dex `long-to-int' instruction. 2054 locations->SetInAt(0, Location::Any()); 2055 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2056 break; 2057 2058 case Primitive::kPrimFloat: 2059 // Processing a Dex `float-to-int' instruction. 2060 locations->SetInAt(0, Location::RequiresFpuRegister()); 2061 locations->SetOut(Location::RequiresRegister()); 2062 locations->AddTemp(Location::RequiresFpuRegister()); 2063 break; 2064 2065 case Primitive::kPrimDouble: 2066 // Processing a Dex `double-to-int' instruction. 2067 locations->SetInAt(0, Location::RequiresFpuRegister()); 2068 locations->SetOut(Location::RequiresRegister()); 2069 locations->AddTemp(Location::RequiresFpuRegister()); 2070 break; 2071 2072 default: 2073 LOG(FATAL) << "Unexpected type conversion from " << input_type 2074 << " to " << result_type; 2075 } 2076 break; 2077 2078 case Primitive::kPrimLong: 2079 switch (input_type) { 2080 case Primitive::kPrimBoolean: 2081 // Boolean input is a result of code transformations. 2082 case Primitive::kPrimByte: 2083 case Primitive::kPrimShort: 2084 case Primitive::kPrimInt: 2085 case Primitive::kPrimChar: 2086 // Processing a Dex `int-to-long' instruction. 2087 locations->SetInAt(0, Location::RequiresRegister()); 2088 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2089 break; 2090 2091 case Primitive::kPrimFloat: { 2092 // Processing a Dex `float-to-long' instruction. 2093 InvokeRuntimeCallingConvention calling_convention; 2094 locations->SetInAt(0, Location::FpuRegisterLocation( 2095 calling_convention.GetFpuRegisterAt(0))); 2096 locations->SetOut(Location::RegisterPairLocation(R0, R1)); 2097 break; 2098 } 2099 2100 case Primitive::kPrimDouble: { 2101 // Processing a Dex `double-to-long' instruction. 2102 InvokeRuntimeCallingConvention calling_convention; 2103 locations->SetInAt(0, Location::FpuRegisterPairLocation( 2104 calling_convention.GetFpuRegisterAt(0), 2105 calling_convention.GetFpuRegisterAt(1))); 2106 locations->SetOut(Location::RegisterPairLocation(R0, R1)); 2107 break; 2108 } 2109 2110 default: 2111 LOG(FATAL) << "Unexpected type conversion from " << input_type 2112 << " to " << result_type; 2113 } 2114 break; 2115 2116 case Primitive::kPrimChar: 2117 switch (input_type) { 2118 case Primitive::kPrimLong: 2119 // Type conversion from long to char is a result of code transformations. 2120 case Primitive::kPrimBoolean: 2121 // Boolean input is a result of code transformations. 2122 case Primitive::kPrimByte: 2123 case Primitive::kPrimShort: 2124 case Primitive::kPrimInt: 2125 // Processing a Dex `int-to-char' instruction. 2126 locations->SetInAt(0, Location::RequiresRegister()); 2127 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2128 break; 2129 2130 default: 2131 LOG(FATAL) << "Unexpected type conversion from " << input_type 2132 << " to " << result_type; 2133 } 2134 break; 2135 2136 case Primitive::kPrimFloat: 2137 switch (input_type) { 2138 case Primitive::kPrimBoolean: 2139 // Boolean input is a result of code transformations. 2140 case Primitive::kPrimByte: 2141 case Primitive::kPrimShort: 2142 case Primitive::kPrimInt: 2143 case Primitive::kPrimChar: 2144 // Processing a Dex `int-to-float' instruction. 2145 locations->SetInAt(0, Location::RequiresRegister()); 2146 locations->SetOut(Location::RequiresFpuRegister()); 2147 break; 2148 2149 case Primitive::kPrimLong: { 2150 // Processing a Dex `long-to-float' instruction. 2151 InvokeRuntimeCallingConvention calling_convention; 2152 locations->SetInAt(0, Location::RegisterPairLocation( 2153 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1))); 2154 locations->SetOut(Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(0))); 2155 break; 2156 } 2157 2158 case Primitive::kPrimDouble: 2159 // Processing a Dex `double-to-float' instruction. 2160 locations->SetInAt(0, Location::RequiresFpuRegister()); 2161 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2162 break; 2163 2164 default: 2165 LOG(FATAL) << "Unexpected type conversion from " << input_type 2166 << " to " << result_type; 2167 }; 2168 break; 2169 2170 case Primitive::kPrimDouble: 2171 switch (input_type) { 2172 case Primitive::kPrimBoolean: 2173 // Boolean input is a result of code transformations. 2174 case Primitive::kPrimByte: 2175 case Primitive::kPrimShort: 2176 case Primitive::kPrimInt: 2177 case Primitive::kPrimChar: 2178 // Processing a Dex `int-to-double' instruction. 2179 locations->SetInAt(0, Location::RequiresRegister()); 2180 locations->SetOut(Location::RequiresFpuRegister()); 2181 break; 2182 2183 case Primitive::kPrimLong: 2184 // Processing a Dex `long-to-double' instruction. 2185 locations->SetInAt(0, Location::RequiresRegister()); 2186 locations->SetOut(Location::RequiresFpuRegister()); 2187 locations->AddTemp(Location::RequiresFpuRegister()); 2188 locations->AddTemp(Location::RequiresFpuRegister()); 2189 break; 2190 2191 case Primitive::kPrimFloat: 2192 // Processing a Dex `float-to-double' instruction. 2193 locations->SetInAt(0, Location::RequiresFpuRegister()); 2194 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2195 break; 2196 2197 default: 2198 LOG(FATAL) << "Unexpected type conversion from " << input_type 2199 << " to " << result_type; 2200 }; 2201 break; 2202 2203 default: 2204 LOG(FATAL) << "Unexpected type conversion from " << input_type 2205 << " to " << result_type; 2206 } 2207} 2208 2209void InstructionCodeGeneratorARM::VisitTypeConversion(HTypeConversion* conversion) { 2210 LocationSummary* locations = conversion->GetLocations(); 2211 Location out = locations->Out(); 2212 Location in = locations->InAt(0); 2213 Primitive::Type result_type = conversion->GetResultType(); 2214 Primitive::Type input_type = conversion->GetInputType(); 2215 DCHECK_NE(result_type, input_type); 2216 switch (result_type) { 2217 case Primitive::kPrimByte: 2218 switch (input_type) { 2219 case Primitive::kPrimLong: 2220 // Type conversion from long to byte is a result of code transformations. 2221 __ sbfx(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>(), 0, 8); 2222 break; 2223 case Primitive::kPrimBoolean: 2224 // Boolean input is a result of code transformations. 2225 case Primitive::kPrimShort: 2226 case Primitive::kPrimInt: 2227 case Primitive::kPrimChar: 2228 // Processing a Dex `int-to-byte' instruction. 2229 __ sbfx(out.AsRegister<Register>(), in.AsRegister<Register>(), 0, 8); 2230 break; 2231 2232 default: 2233 LOG(FATAL) << "Unexpected type conversion from " << input_type 2234 << " to " << result_type; 2235 } 2236 break; 2237 2238 case Primitive::kPrimShort: 2239 switch (input_type) { 2240 case Primitive::kPrimLong: 2241 // Type conversion from long to short is a result of code transformations. 2242 __ sbfx(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>(), 0, 16); 2243 break; 2244 case Primitive::kPrimBoolean: 2245 // Boolean input is a result of code transformations. 2246 case Primitive::kPrimByte: 2247 case Primitive::kPrimInt: 2248 case Primitive::kPrimChar: 2249 // Processing a Dex `int-to-short' instruction. 2250 __ sbfx(out.AsRegister<Register>(), in.AsRegister<Register>(), 0, 16); 2251 break; 2252 2253 default: 2254 LOG(FATAL) << "Unexpected type conversion from " << input_type 2255 << " to " << result_type; 2256 } 2257 break; 2258 2259 case Primitive::kPrimInt: 2260 switch (input_type) { 2261 case Primitive::kPrimLong: 2262 // Processing a Dex `long-to-int' instruction. 2263 DCHECK(out.IsRegister()); 2264 if (in.IsRegisterPair()) { 2265 __ Mov(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>()); 2266 } else if (in.IsDoubleStackSlot()) { 2267 __ LoadFromOffset(kLoadWord, out.AsRegister<Register>(), SP, in.GetStackIndex()); 2268 } else { 2269 DCHECK(in.IsConstant()); 2270 DCHECK(in.GetConstant()->IsLongConstant()); 2271 int64_t value = in.GetConstant()->AsLongConstant()->GetValue(); 2272 __ LoadImmediate(out.AsRegister<Register>(), static_cast<int32_t>(value)); 2273 } 2274 break; 2275 2276 case Primitive::kPrimFloat: { 2277 // Processing a Dex `float-to-int' instruction. 2278 SRegister temp = locations->GetTemp(0).AsFpuRegisterPairLow<SRegister>(); 2279 __ vmovs(temp, in.AsFpuRegister<SRegister>()); 2280 __ vcvtis(temp, temp); 2281 __ vmovrs(out.AsRegister<Register>(), temp); 2282 break; 2283 } 2284 2285 case Primitive::kPrimDouble: { 2286 // Processing a Dex `double-to-int' instruction. 2287 SRegister temp_s = locations->GetTemp(0).AsFpuRegisterPairLow<SRegister>(); 2288 DRegister temp_d = FromLowSToD(temp_s); 2289 __ vmovd(temp_d, FromLowSToD(in.AsFpuRegisterPairLow<SRegister>())); 2290 __ vcvtid(temp_s, temp_d); 2291 __ vmovrs(out.AsRegister<Register>(), temp_s); 2292 break; 2293 } 2294 2295 default: 2296 LOG(FATAL) << "Unexpected type conversion from " << input_type 2297 << " to " << result_type; 2298 } 2299 break; 2300 2301 case Primitive::kPrimLong: 2302 switch (input_type) { 2303 case Primitive::kPrimBoolean: 2304 // Boolean input is a result of code transformations. 2305 case Primitive::kPrimByte: 2306 case Primitive::kPrimShort: 2307 case Primitive::kPrimInt: 2308 case Primitive::kPrimChar: 2309 // Processing a Dex `int-to-long' instruction. 2310 DCHECK(out.IsRegisterPair()); 2311 DCHECK(in.IsRegister()); 2312 __ Mov(out.AsRegisterPairLow<Register>(), in.AsRegister<Register>()); 2313 // Sign extension. 2314 __ Asr(out.AsRegisterPairHigh<Register>(), 2315 out.AsRegisterPairLow<Register>(), 2316 31); 2317 break; 2318 2319 case Primitive::kPrimFloat: 2320 // Processing a Dex `float-to-long' instruction. 2321 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pF2l), 2322 conversion, 2323 conversion->GetDexPc(), 2324 nullptr); 2325 CheckEntrypointTypes<kQuickF2l, int64_t, float>(); 2326 break; 2327 2328 case Primitive::kPrimDouble: 2329 // Processing a Dex `double-to-long' instruction. 2330 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pD2l), 2331 conversion, 2332 conversion->GetDexPc(), 2333 nullptr); 2334 CheckEntrypointTypes<kQuickD2l, int64_t, double>(); 2335 break; 2336 2337 default: 2338 LOG(FATAL) << "Unexpected type conversion from " << input_type 2339 << " to " << result_type; 2340 } 2341 break; 2342 2343 case Primitive::kPrimChar: 2344 switch (input_type) { 2345 case Primitive::kPrimLong: 2346 // Type conversion from long to char is a result of code transformations. 2347 __ ubfx(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>(), 0, 16); 2348 break; 2349 case Primitive::kPrimBoolean: 2350 // Boolean input is a result of code transformations. 2351 case Primitive::kPrimByte: 2352 case Primitive::kPrimShort: 2353 case Primitive::kPrimInt: 2354 // Processing a Dex `int-to-char' instruction. 2355 __ ubfx(out.AsRegister<Register>(), in.AsRegister<Register>(), 0, 16); 2356 break; 2357 2358 default: 2359 LOG(FATAL) << "Unexpected type conversion from " << input_type 2360 << " to " << result_type; 2361 } 2362 break; 2363 2364 case Primitive::kPrimFloat: 2365 switch (input_type) { 2366 case Primitive::kPrimBoolean: 2367 // Boolean input is a result of code transformations. 2368 case Primitive::kPrimByte: 2369 case Primitive::kPrimShort: 2370 case Primitive::kPrimInt: 2371 case Primitive::kPrimChar: { 2372 // Processing a Dex `int-to-float' instruction. 2373 __ vmovsr(out.AsFpuRegister<SRegister>(), in.AsRegister<Register>()); 2374 __ vcvtsi(out.AsFpuRegister<SRegister>(), out.AsFpuRegister<SRegister>()); 2375 break; 2376 } 2377 2378 case Primitive::kPrimLong: 2379 // Processing a Dex `long-to-float' instruction. 2380 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pL2f), 2381 conversion, 2382 conversion->GetDexPc(), 2383 nullptr); 2384 CheckEntrypointTypes<kQuickL2f, float, int64_t>(); 2385 break; 2386 2387 case Primitive::kPrimDouble: 2388 // Processing a Dex `double-to-float' instruction. 2389 __ vcvtsd(out.AsFpuRegister<SRegister>(), 2390 FromLowSToD(in.AsFpuRegisterPairLow<SRegister>())); 2391 break; 2392 2393 default: 2394 LOG(FATAL) << "Unexpected type conversion from " << input_type 2395 << " to " << result_type; 2396 }; 2397 break; 2398 2399 case Primitive::kPrimDouble: 2400 switch (input_type) { 2401 case Primitive::kPrimBoolean: 2402 // Boolean input is a result of code transformations. 2403 case Primitive::kPrimByte: 2404 case Primitive::kPrimShort: 2405 case Primitive::kPrimInt: 2406 case Primitive::kPrimChar: { 2407 // Processing a Dex `int-to-double' instruction. 2408 __ vmovsr(out.AsFpuRegisterPairLow<SRegister>(), in.AsRegister<Register>()); 2409 __ vcvtdi(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2410 out.AsFpuRegisterPairLow<SRegister>()); 2411 break; 2412 } 2413 2414 case Primitive::kPrimLong: { 2415 // Processing a Dex `long-to-double' instruction. 2416 Register low = in.AsRegisterPairLow<Register>(); 2417 Register high = in.AsRegisterPairHigh<Register>(); 2418 SRegister out_s = out.AsFpuRegisterPairLow<SRegister>(); 2419 DRegister out_d = FromLowSToD(out_s); 2420 SRegister temp_s = locations->GetTemp(0).AsFpuRegisterPairLow<SRegister>(); 2421 DRegister temp_d = FromLowSToD(temp_s); 2422 SRegister constant_s = locations->GetTemp(1).AsFpuRegisterPairLow<SRegister>(); 2423 DRegister constant_d = FromLowSToD(constant_s); 2424 2425 // temp_d = int-to-double(high) 2426 __ vmovsr(temp_s, high); 2427 __ vcvtdi(temp_d, temp_s); 2428 // constant_d = k2Pow32EncodingForDouble 2429 __ LoadDImmediate(constant_d, bit_cast<double, int64_t>(k2Pow32EncodingForDouble)); 2430 // out_d = unsigned-to-double(low) 2431 __ vmovsr(out_s, low); 2432 __ vcvtdu(out_d, out_s); 2433 // out_d += temp_d * constant_d 2434 __ vmlad(out_d, temp_d, constant_d); 2435 break; 2436 } 2437 2438 case Primitive::kPrimFloat: 2439 // Processing a Dex `float-to-double' instruction. 2440 __ vcvtds(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2441 in.AsFpuRegister<SRegister>()); 2442 break; 2443 2444 default: 2445 LOG(FATAL) << "Unexpected type conversion from " << input_type 2446 << " to " << result_type; 2447 }; 2448 break; 2449 2450 default: 2451 LOG(FATAL) << "Unexpected type conversion from " << input_type 2452 << " to " << result_type; 2453 } 2454} 2455 2456void LocationsBuilderARM::VisitAdd(HAdd* add) { 2457 LocationSummary* locations = 2458 new (GetGraph()->GetArena()) LocationSummary(add, LocationSummary::kNoCall); 2459 switch (add->GetResultType()) { 2460 case Primitive::kPrimInt: { 2461 locations->SetInAt(0, Location::RequiresRegister()); 2462 locations->SetInAt(1, Location::RegisterOrConstant(add->InputAt(1))); 2463 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2464 break; 2465 } 2466 2467 case Primitive::kPrimLong: { 2468 locations->SetInAt(0, Location::RequiresRegister()); 2469 locations->SetInAt(1, Location::RequiresRegister()); 2470 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2471 break; 2472 } 2473 2474 case Primitive::kPrimFloat: 2475 case Primitive::kPrimDouble: { 2476 locations->SetInAt(0, Location::RequiresFpuRegister()); 2477 locations->SetInAt(1, Location::RequiresFpuRegister()); 2478 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2479 break; 2480 } 2481 2482 default: 2483 LOG(FATAL) << "Unexpected add type " << add->GetResultType(); 2484 } 2485} 2486 2487void InstructionCodeGeneratorARM::VisitAdd(HAdd* add) { 2488 LocationSummary* locations = add->GetLocations(); 2489 Location out = locations->Out(); 2490 Location first = locations->InAt(0); 2491 Location second = locations->InAt(1); 2492 switch (add->GetResultType()) { 2493 case Primitive::kPrimInt: 2494 if (second.IsRegister()) { 2495 __ add(out.AsRegister<Register>(), 2496 first.AsRegister<Register>(), 2497 ShifterOperand(second.AsRegister<Register>())); 2498 } else { 2499 __ AddConstant(out.AsRegister<Register>(), 2500 first.AsRegister<Register>(), 2501 second.GetConstant()->AsIntConstant()->GetValue()); 2502 } 2503 break; 2504 2505 case Primitive::kPrimLong: { 2506 DCHECK(second.IsRegisterPair()); 2507 __ adds(out.AsRegisterPairLow<Register>(), 2508 first.AsRegisterPairLow<Register>(), 2509 ShifterOperand(second.AsRegisterPairLow<Register>())); 2510 __ adc(out.AsRegisterPairHigh<Register>(), 2511 first.AsRegisterPairHigh<Register>(), 2512 ShifterOperand(second.AsRegisterPairHigh<Register>())); 2513 break; 2514 } 2515 2516 case Primitive::kPrimFloat: 2517 __ vadds(out.AsFpuRegister<SRegister>(), 2518 first.AsFpuRegister<SRegister>(), 2519 second.AsFpuRegister<SRegister>()); 2520 break; 2521 2522 case Primitive::kPrimDouble: 2523 __ vaddd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2524 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 2525 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 2526 break; 2527 2528 default: 2529 LOG(FATAL) << "Unexpected add type " << add->GetResultType(); 2530 } 2531} 2532 2533void LocationsBuilderARM::VisitSub(HSub* sub) { 2534 LocationSummary* locations = 2535 new (GetGraph()->GetArena()) LocationSummary(sub, LocationSummary::kNoCall); 2536 switch (sub->GetResultType()) { 2537 case Primitive::kPrimInt: { 2538 locations->SetInAt(0, Location::RequiresRegister()); 2539 locations->SetInAt(1, Location::RegisterOrConstant(sub->InputAt(1))); 2540 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2541 break; 2542 } 2543 2544 case Primitive::kPrimLong: { 2545 locations->SetInAt(0, Location::RequiresRegister()); 2546 locations->SetInAt(1, Location::RequiresRegister()); 2547 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2548 break; 2549 } 2550 case Primitive::kPrimFloat: 2551 case Primitive::kPrimDouble: { 2552 locations->SetInAt(0, Location::RequiresFpuRegister()); 2553 locations->SetInAt(1, Location::RequiresFpuRegister()); 2554 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2555 break; 2556 } 2557 default: 2558 LOG(FATAL) << "Unexpected sub type " << sub->GetResultType(); 2559 } 2560} 2561 2562void InstructionCodeGeneratorARM::VisitSub(HSub* sub) { 2563 LocationSummary* locations = sub->GetLocations(); 2564 Location out = locations->Out(); 2565 Location first = locations->InAt(0); 2566 Location second = locations->InAt(1); 2567 switch (sub->GetResultType()) { 2568 case Primitive::kPrimInt: { 2569 if (second.IsRegister()) { 2570 __ sub(out.AsRegister<Register>(), 2571 first.AsRegister<Register>(), 2572 ShifterOperand(second.AsRegister<Register>())); 2573 } else { 2574 __ AddConstant(out.AsRegister<Register>(), 2575 first.AsRegister<Register>(), 2576 -second.GetConstant()->AsIntConstant()->GetValue()); 2577 } 2578 break; 2579 } 2580 2581 case Primitive::kPrimLong: { 2582 DCHECK(second.IsRegisterPair()); 2583 __ subs(out.AsRegisterPairLow<Register>(), 2584 first.AsRegisterPairLow<Register>(), 2585 ShifterOperand(second.AsRegisterPairLow<Register>())); 2586 __ sbc(out.AsRegisterPairHigh<Register>(), 2587 first.AsRegisterPairHigh<Register>(), 2588 ShifterOperand(second.AsRegisterPairHigh<Register>())); 2589 break; 2590 } 2591 2592 case Primitive::kPrimFloat: { 2593 __ vsubs(out.AsFpuRegister<SRegister>(), 2594 first.AsFpuRegister<SRegister>(), 2595 second.AsFpuRegister<SRegister>()); 2596 break; 2597 } 2598 2599 case Primitive::kPrimDouble: { 2600 __ vsubd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2601 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 2602 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 2603 break; 2604 } 2605 2606 2607 default: 2608 LOG(FATAL) << "Unexpected sub type " << sub->GetResultType(); 2609 } 2610} 2611 2612void LocationsBuilderARM::VisitMul(HMul* mul) { 2613 LocationSummary* locations = 2614 new (GetGraph()->GetArena()) LocationSummary(mul, LocationSummary::kNoCall); 2615 switch (mul->GetResultType()) { 2616 case Primitive::kPrimInt: 2617 case Primitive::kPrimLong: { 2618 locations->SetInAt(0, Location::RequiresRegister()); 2619 locations->SetInAt(1, Location::RequiresRegister()); 2620 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2621 break; 2622 } 2623 2624 case Primitive::kPrimFloat: 2625 case Primitive::kPrimDouble: { 2626 locations->SetInAt(0, Location::RequiresFpuRegister()); 2627 locations->SetInAt(1, Location::RequiresFpuRegister()); 2628 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2629 break; 2630 } 2631 2632 default: 2633 LOG(FATAL) << "Unexpected mul type " << mul->GetResultType(); 2634 } 2635} 2636 2637void InstructionCodeGeneratorARM::VisitMul(HMul* mul) { 2638 LocationSummary* locations = mul->GetLocations(); 2639 Location out = locations->Out(); 2640 Location first = locations->InAt(0); 2641 Location second = locations->InAt(1); 2642 switch (mul->GetResultType()) { 2643 case Primitive::kPrimInt: { 2644 __ mul(out.AsRegister<Register>(), 2645 first.AsRegister<Register>(), 2646 second.AsRegister<Register>()); 2647 break; 2648 } 2649 case Primitive::kPrimLong: { 2650 Register out_hi = out.AsRegisterPairHigh<Register>(); 2651 Register out_lo = out.AsRegisterPairLow<Register>(); 2652 Register in1_hi = first.AsRegisterPairHigh<Register>(); 2653 Register in1_lo = first.AsRegisterPairLow<Register>(); 2654 Register in2_hi = second.AsRegisterPairHigh<Register>(); 2655 Register in2_lo = second.AsRegisterPairLow<Register>(); 2656 2657 // Extra checks to protect caused by the existence of R1_R2. 2658 // The algorithm is wrong if out.hi is either in1.lo or in2.lo: 2659 // (e.g. in1=r0_r1, in2=r2_r3 and out=r1_r2); 2660 DCHECK_NE(out_hi, in1_lo); 2661 DCHECK_NE(out_hi, in2_lo); 2662 2663 // input: in1 - 64 bits, in2 - 64 bits 2664 // output: out 2665 // formula: out.hi : out.lo = (in1.lo * in2.hi + in1.hi * in2.lo)* 2^32 + in1.lo * in2.lo 2666 // parts: out.hi = in1.lo * in2.hi + in1.hi * in2.lo + (in1.lo * in2.lo)[63:32] 2667 // parts: out.lo = (in1.lo * in2.lo)[31:0] 2668 2669 // IP <- in1.lo * in2.hi 2670 __ mul(IP, in1_lo, in2_hi); 2671 // out.hi <- in1.lo * in2.hi + in1.hi * in2.lo 2672 __ mla(out_hi, in1_hi, in2_lo, IP); 2673 // out.lo <- (in1.lo * in2.lo)[31:0]; 2674 __ umull(out_lo, IP, in1_lo, in2_lo); 2675 // out.hi <- in2.hi * in1.lo + in2.lo * in1.hi + (in1.lo * in2.lo)[63:32] 2676 __ add(out_hi, out_hi, ShifterOperand(IP)); 2677 break; 2678 } 2679 2680 case Primitive::kPrimFloat: { 2681 __ vmuls(out.AsFpuRegister<SRegister>(), 2682 first.AsFpuRegister<SRegister>(), 2683 second.AsFpuRegister<SRegister>()); 2684 break; 2685 } 2686 2687 case Primitive::kPrimDouble: { 2688 __ vmuld(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2689 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 2690 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 2691 break; 2692 } 2693 2694 default: 2695 LOG(FATAL) << "Unexpected mul type " << mul->GetResultType(); 2696 } 2697} 2698 2699void InstructionCodeGeneratorARM::DivRemOneOrMinusOne(HBinaryOperation* instruction) { 2700 DCHECK(instruction->IsDiv() || instruction->IsRem()); 2701 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 2702 2703 LocationSummary* locations = instruction->GetLocations(); 2704 Location second = locations->InAt(1); 2705 DCHECK(second.IsConstant()); 2706 2707 Register out = locations->Out().AsRegister<Register>(); 2708 Register dividend = locations->InAt(0).AsRegister<Register>(); 2709 int32_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 2710 DCHECK(imm == 1 || imm == -1); 2711 2712 if (instruction->IsRem()) { 2713 __ LoadImmediate(out, 0); 2714 } else { 2715 if (imm == 1) { 2716 __ Mov(out, dividend); 2717 } else { 2718 __ rsb(out, dividend, ShifterOperand(0)); 2719 } 2720 } 2721} 2722 2723void InstructionCodeGeneratorARM::DivRemByPowerOfTwo(HBinaryOperation* instruction) { 2724 DCHECK(instruction->IsDiv() || instruction->IsRem()); 2725 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 2726 2727 LocationSummary* locations = instruction->GetLocations(); 2728 Location second = locations->InAt(1); 2729 DCHECK(second.IsConstant()); 2730 2731 Register out = locations->Out().AsRegister<Register>(); 2732 Register dividend = locations->InAt(0).AsRegister<Register>(); 2733 Register temp = locations->GetTemp(0).AsRegister<Register>(); 2734 int32_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 2735 uint32_t abs_imm = static_cast<uint32_t>(AbsOrMin(imm)); 2736 int ctz_imm = CTZ(abs_imm); 2737 2738 if (ctz_imm == 1) { 2739 __ Lsr(temp, dividend, 32 - ctz_imm); 2740 } else { 2741 __ Asr(temp, dividend, 31); 2742 __ Lsr(temp, temp, 32 - ctz_imm); 2743 } 2744 __ add(out, temp, ShifterOperand(dividend)); 2745 2746 if (instruction->IsDiv()) { 2747 __ Asr(out, out, ctz_imm); 2748 if (imm < 0) { 2749 __ rsb(out, out, ShifterOperand(0)); 2750 } 2751 } else { 2752 __ ubfx(out, out, 0, ctz_imm); 2753 __ sub(out, out, ShifterOperand(temp)); 2754 } 2755} 2756 2757void InstructionCodeGeneratorARM::GenerateDivRemWithAnyConstant(HBinaryOperation* instruction) { 2758 DCHECK(instruction->IsDiv() || instruction->IsRem()); 2759 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 2760 2761 LocationSummary* locations = instruction->GetLocations(); 2762 Location second = locations->InAt(1); 2763 DCHECK(second.IsConstant()); 2764 2765 Register out = locations->Out().AsRegister<Register>(); 2766 Register dividend = locations->InAt(0).AsRegister<Register>(); 2767 Register temp1 = locations->GetTemp(0).AsRegister<Register>(); 2768 Register temp2 = locations->GetTemp(1).AsRegister<Register>(); 2769 int64_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 2770 2771 int64_t magic; 2772 int shift; 2773 CalculateMagicAndShiftForDivRem(imm, false /* is_long */, &magic, &shift); 2774 2775 __ LoadImmediate(temp1, magic); 2776 __ smull(temp2, temp1, dividend, temp1); 2777 2778 if (imm > 0 && magic < 0) { 2779 __ add(temp1, temp1, ShifterOperand(dividend)); 2780 } else if (imm < 0 && magic > 0) { 2781 __ sub(temp1, temp1, ShifterOperand(dividend)); 2782 } 2783 2784 if (shift != 0) { 2785 __ Asr(temp1, temp1, shift); 2786 } 2787 2788 if (instruction->IsDiv()) { 2789 __ sub(out, temp1, ShifterOperand(temp1, ASR, 31)); 2790 } else { 2791 __ sub(temp1, temp1, ShifterOperand(temp1, ASR, 31)); 2792 // TODO: Strength reduction for mls. 2793 __ LoadImmediate(temp2, imm); 2794 __ mls(out, temp1, temp2, dividend); 2795 } 2796} 2797 2798void InstructionCodeGeneratorARM::GenerateDivRemConstantIntegral(HBinaryOperation* instruction) { 2799 DCHECK(instruction->IsDiv() || instruction->IsRem()); 2800 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 2801 2802 LocationSummary* locations = instruction->GetLocations(); 2803 Location second = locations->InAt(1); 2804 DCHECK(second.IsConstant()); 2805 2806 int32_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 2807 if (imm == 0) { 2808 // Do not generate anything. DivZeroCheck would prevent any code to be executed. 2809 } else if (imm == 1 || imm == -1) { 2810 DivRemOneOrMinusOne(instruction); 2811 } else if (IsPowerOfTwo(AbsOrMin(imm))) { 2812 DivRemByPowerOfTwo(instruction); 2813 } else { 2814 DCHECK(imm <= -2 || imm >= 2); 2815 GenerateDivRemWithAnyConstant(instruction); 2816 } 2817} 2818 2819void LocationsBuilderARM::VisitDiv(HDiv* div) { 2820 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 2821 if (div->GetResultType() == Primitive::kPrimLong) { 2822 // pLdiv runtime call. 2823 call_kind = LocationSummary::kCall; 2824 } else if (div->GetResultType() == Primitive::kPrimInt && div->InputAt(1)->IsConstant()) { 2825 // sdiv will be replaced by other instruction sequence. 2826 } else if (div->GetResultType() == Primitive::kPrimInt && 2827 !codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 2828 // pIdivmod runtime call. 2829 call_kind = LocationSummary::kCall; 2830 } 2831 2832 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(div, call_kind); 2833 2834 switch (div->GetResultType()) { 2835 case Primitive::kPrimInt: { 2836 if (div->InputAt(1)->IsConstant()) { 2837 locations->SetInAt(0, Location::RequiresRegister()); 2838 locations->SetInAt(1, Location::ConstantLocation(div->InputAt(1)->AsConstant())); 2839 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2840 int32_t value = div->InputAt(1)->AsIntConstant()->GetValue(); 2841 if (value == 1 || value == 0 || value == -1) { 2842 // No temp register required. 2843 } else { 2844 locations->AddTemp(Location::RequiresRegister()); 2845 if (!IsPowerOfTwo(AbsOrMin(value))) { 2846 locations->AddTemp(Location::RequiresRegister()); 2847 } 2848 } 2849 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 2850 locations->SetInAt(0, Location::RequiresRegister()); 2851 locations->SetInAt(1, Location::RequiresRegister()); 2852 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2853 } else { 2854 InvokeRuntimeCallingConvention calling_convention; 2855 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 2856 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 2857 // Note: divrem will compute both the quotient and the remainder as the pair R0 and R1, but 2858 // we only need the former. 2859 locations->SetOut(Location::RegisterLocation(R0)); 2860 } 2861 break; 2862 } 2863 case Primitive::kPrimLong: { 2864 InvokeRuntimeCallingConvention calling_convention; 2865 locations->SetInAt(0, Location::RegisterPairLocation( 2866 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1))); 2867 locations->SetInAt(1, Location::RegisterPairLocation( 2868 calling_convention.GetRegisterAt(2), calling_convention.GetRegisterAt(3))); 2869 locations->SetOut(Location::RegisterPairLocation(R0, R1)); 2870 break; 2871 } 2872 case Primitive::kPrimFloat: 2873 case Primitive::kPrimDouble: { 2874 locations->SetInAt(0, Location::RequiresFpuRegister()); 2875 locations->SetInAt(1, Location::RequiresFpuRegister()); 2876 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2877 break; 2878 } 2879 2880 default: 2881 LOG(FATAL) << "Unexpected div type " << div->GetResultType(); 2882 } 2883} 2884 2885void InstructionCodeGeneratorARM::VisitDiv(HDiv* div) { 2886 LocationSummary* locations = div->GetLocations(); 2887 Location out = locations->Out(); 2888 Location first = locations->InAt(0); 2889 Location second = locations->InAt(1); 2890 2891 switch (div->GetResultType()) { 2892 case Primitive::kPrimInt: { 2893 if (second.IsConstant()) { 2894 GenerateDivRemConstantIntegral(div); 2895 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 2896 __ sdiv(out.AsRegister<Register>(), 2897 first.AsRegister<Register>(), 2898 second.AsRegister<Register>()); 2899 } else { 2900 InvokeRuntimeCallingConvention calling_convention; 2901 DCHECK_EQ(calling_convention.GetRegisterAt(0), first.AsRegister<Register>()); 2902 DCHECK_EQ(calling_convention.GetRegisterAt(1), second.AsRegister<Register>()); 2903 DCHECK_EQ(R0, out.AsRegister<Register>()); 2904 2905 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pIdivmod), div, div->GetDexPc(), nullptr); 2906 CheckEntrypointTypes<kQuickIdivmod, int32_t, int32_t, int32_t>(); 2907 } 2908 break; 2909 } 2910 2911 case Primitive::kPrimLong: { 2912 InvokeRuntimeCallingConvention calling_convention; 2913 DCHECK_EQ(calling_convention.GetRegisterAt(0), first.AsRegisterPairLow<Register>()); 2914 DCHECK_EQ(calling_convention.GetRegisterAt(1), first.AsRegisterPairHigh<Register>()); 2915 DCHECK_EQ(calling_convention.GetRegisterAt(2), second.AsRegisterPairLow<Register>()); 2916 DCHECK_EQ(calling_convention.GetRegisterAt(3), second.AsRegisterPairHigh<Register>()); 2917 DCHECK_EQ(R0, out.AsRegisterPairLow<Register>()); 2918 DCHECK_EQ(R1, out.AsRegisterPairHigh<Register>()); 2919 2920 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pLdiv), div, div->GetDexPc(), nullptr); 2921 CheckEntrypointTypes<kQuickLdiv, int64_t, int64_t, int64_t>(); 2922 break; 2923 } 2924 2925 case Primitive::kPrimFloat: { 2926 __ vdivs(out.AsFpuRegister<SRegister>(), 2927 first.AsFpuRegister<SRegister>(), 2928 second.AsFpuRegister<SRegister>()); 2929 break; 2930 } 2931 2932 case Primitive::kPrimDouble: { 2933 __ vdivd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2934 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 2935 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 2936 break; 2937 } 2938 2939 default: 2940 LOG(FATAL) << "Unexpected div type " << div->GetResultType(); 2941 } 2942} 2943 2944void LocationsBuilderARM::VisitRem(HRem* rem) { 2945 Primitive::Type type = rem->GetResultType(); 2946 2947 // Most remainders are implemented in the runtime. 2948 LocationSummary::CallKind call_kind = LocationSummary::kCall; 2949 if (rem->GetResultType() == Primitive::kPrimInt && rem->InputAt(1)->IsConstant()) { 2950 // sdiv will be replaced by other instruction sequence. 2951 call_kind = LocationSummary::kNoCall; 2952 } else if ((rem->GetResultType() == Primitive::kPrimInt) 2953 && codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 2954 // Have hardware divide instruction for int, do it with three instructions. 2955 call_kind = LocationSummary::kNoCall; 2956 } 2957 2958 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(rem, call_kind); 2959 2960 switch (type) { 2961 case Primitive::kPrimInt: { 2962 if (rem->InputAt(1)->IsConstant()) { 2963 locations->SetInAt(0, Location::RequiresRegister()); 2964 locations->SetInAt(1, Location::ConstantLocation(rem->InputAt(1)->AsConstant())); 2965 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2966 int32_t value = rem->InputAt(1)->AsIntConstant()->GetValue(); 2967 if (value == 1 || value == 0 || value == -1) { 2968 // No temp register required. 2969 } else { 2970 locations->AddTemp(Location::RequiresRegister()); 2971 if (!IsPowerOfTwo(AbsOrMin(value))) { 2972 locations->AddTemp(Location::RequiresRegister()); 2973 } 2974 } 2975 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 2976 locations->SetInAt(0, Location::RequiresRegister()); 2977 locations->SetInAt(1, Location::RequiresRegister()); 2978 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2979 locations->AddTemp(Location::RequiresRegister()); 2980 } else { 2981 InvokeRuntimeCallingConvention calling_convention; 2982 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 2983 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 2984 // Note: divrem will compute both the quotient and the remainder as the pair R0 and R1, but 2985 // we only need the latter. 2986 locations->SetOut(Location::RegisterLocation(R1)); 2987 } 2988 break; 2989 } 2990 case Primitive::kPrimLong: { 2991 InvokeRuntimeCallingConvention calling_convention; 2992 locations->SetInAt(0, Location::RegisterPairLocation( 2993 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1))); 2994 locations->SetInAt(1, Location::RegisterPairLocation( 2995 calling_convention.GetRegisterAt(2), calling_convention.GetRegisterAt(3))); 2996 // The runtime helper puts the output in R2,R3. 2997 locations->SetOut(Location::RegisterPairLocation(R2, R3)); 2998 break; 2999 } 3000 case Primitive::kPrimFloat: { 3001 InvokeRuntimeCallingConvention calling_convention; 3002 locations->SetInAt(0, Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(0))); 3003 locations->SetInAt(1, Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(1))); 3004 locations->SetOut(Location::FpuRegisterLocation(S0)); 3005 break; 3006 } 3007 3008 case Primitive::kPrimDouble: { 3009 InvokeRuntimeCallingConvention calling_convention; 3010 locations->SetInAt(0, Location::FpuRegisterPairLocation( 3011 calling_convention.GetFpuRegisterAt(0), calling_convention.GetFpuRegisterAt(1))); 3012 locations->SetInAt(1, Location::FpuRegisterPairLocation( 3013 calling_convention.GetFpuRegisterAt(2), calling_convention.GetFpuRegisterAt(3))); 3014 locations->SetOut(Location::Location::FpuRegisterPairLocation(S0, S1)); 3015 break; 3016 } 3017 3018 default: 3019 LOG(FATAL) << "Unexpected rem type " << type; 3020 } 3021} 3022 3023void InstructionCodeGeneratorARM::VisitRem(HRem* rem) { 3024 LocationSummary* locations = rem->GetLocations(); 3025 Location out = locations->Out(); 3026 Location first = locations->InAt(0); 3027 Location second = locations->InAt(1); 3028 3029 Primitive::Type type = rem->GetResultType(); 3030 switch (type) { 3031 case Primitive::kPrimInt: { 3032 if (second.IsConstant()) { 3033 GenerateDivRemConstantIntegral(rem); 3034 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 3035 Register reg1 = first.AsRegister<Register>(); 3036 Register reg2 = second.AsRegister<Register>(); 3037 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3038 3039 // temp = reg1 / reg2 (integer division) 3040 // dest = reg1 - temp * reg2 3041 __ sdiv(temp, reg1, reg2); 3042 __ mls(out.AsRegister<Register>(), temp, reg2, reg1); 3043 } else { 3044 InvokeRuntimeCallingConvention calling_convention; 3045 DCHECK_EQ(calling_convention.GetRegisterAt(0), first.AsRegister<Register>()); 3046 DCHECK_EQ(calling_convention.GetRegisterAt(1), second.AsRegister<Register>()); 3047 DCHECK_EQ(R1, out.AsRegister<Register>()); 3048 3049 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pIdivmod), rem, rem->GetDexPc(), nullptr); 3050 CheckEntrypointTypes<kQuickIdivmod, int32_t, int32_t, int32_t>(); 3051 } 3052 break; 3053 } 3054 3055 case Primitive::kPrimLong: { 3056 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pLmod), rem, rem->GetDexPc(), nullptr); 3057 CheckEntrypointTypes<kQuickLmod, int64_t, int64_t, int64_t>(); 3058 break; 3059 } 3060 3061 case Primitive::kPrimFloat: { 3062 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pFmodf), rem, rem->GetDexPc(), nullptr); 3063 CheckEntrypointTypes<kQuickFmodf, float, float, float>(); 3064 break; 3065 } 3066 3067 case Primitive::kPrimDouble: { 3068 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pFmod), rem, rem->GetDexPc(), nullptr); 3069 CheckEntrypointTypes<kQuickFmod, double, double, double>(); 3070 break; 3071 } 3072 3073 default: 3074 LOG(FATAL) << "Unexpected rem type " << type; 3075 } 3076} 3077 3078void LocationsBuilderARM::VisitDivZeroCheck(HDivZeroCheck* instruction) { 3079 LocationSummary::CallKind call_kind = instruction->CanThrowIntoCatchBlock() 3080 ? LocationSummary::kCallOnSlowPath 3081 : LocationSummary::kNoCall; 3082 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 3083 locations->SetInAt(0, Location::RegisterOrConstant(instruction->InputAt(0))); 3084 if (instruction->HasUses()) { 3085 locations->SetOut(Location::SameAsFirstInput()); 3086 } 3087} 3088 3089void InstructionCodeGeneratorARM::VisitDivZeroCheck(HDivZeroCheck* instruction) { 3090 SlowPathCode* slow_path = new (GetGraph()->GetArena()) DivZeroCheckSlowPathARM(instruction); 3091 codegen_->AddSlowPath(slow_path); 3092 3093 LocationSummary* locations = instruction->GetLocations(); 3094 Location value = locations->InAt(0); 3095 3096 switch (instruction->GetType()) { 3097 case Primitive::kPrimBoolean: 3098 case Primitive::kPrimByte: 3099 case Primitive::kPrimChar: 3100 case Primitive::kPrimShort: 3101 case Primitive::kPrimInt: { 3102 if (value.IsRegister()) { 3103 __ CompareAndBranchIfZero(value.AsRegister<Register>(), slow_path->GetEntryLabel()); 3104 } else { 3105 DCHECK(value.IsConstant()) << value; 3106 if (value.GetConstant()->AsIntConstant()->GetValue() == 0) { 3107 __ b(slow_path->GetEntryLabel()); 3108 } 3109 } 3110 break; 3111 } 3112 case Primitive::kPrimLong: { 3113 if (value.IsRegisterPair()) { 3114 __ orrs(IP, 3115 value.AsRegisterPairLow<Register>(), 3116 ShifterOperand(value.AsRegisterPairHigh<Register>())); 3117 __ b(slow_path->GetEntryLabel(), EQ); 3118 } else { 3119 DCHECK(value.IsConstant()) << value; 3120 if (value.GetConstant()->AsLongConstant()->GetValue() == 0) { 3121 __ b(slow_path->GetEntryLabel()); 3122 } 3123 } 3124 break; 3125 default: 3126 LOG(FATAL) << "Unexpected type for HDivZeroCheck " << instruction->GetType(); 3127 } 3128 } 3129} 3130 3131void InstructionCodeGeneratorARM::HandleIntegerRotate(LocationSummary* locations) { 3132 Register in = locations->InAt(0).AsRegister<Register>(); 3133 Location rhs = locations->InAt(1); 3134 Register out = locations->Out().AsRegister<Register>(); 3135 3136 if (rhs.IsConstant()) { 3137 // Arm32 and Thumb2 assemblers require a rotation on the interval [1,31], 3138 // so map all rotations to a +ve. equivalent in that range. 3139 // (e.g. left *or* right by -2 bits == 30 bits in the same direction.) 3140 uint32_t rot = CodeGenerator::GetInt32ValueOf(rhs.GetConstant()) & 0x1F; 3141 if (rot) { 3142 // Rotate, mapping left rotations to right equivalents if necessary. 3143 // (e.g. left by 2 bits == right by 30.) 3144 __ Ror(out, in, rot); 3145 } else if (out != in) { 3146 __ Mov(out, in); 3147 } 3148 } else { 3149 __ Ror(out, in, rhs.AsRegister<Register>()); 3150 } 3151} 3152 3153// Gain some speed by mapping all Long rotates onto equivalent pairs of Integer 3154// rotates by swapping input regs (effectively rotating by the first 32-bits of 3155// a larger rotation) or flipping direction (thus treating larger right/left 3156// rotations as sub-word sized rotations in the other direction) as appropriate. 3157void InstructionCodeGeneratorARM::HandleLongRotate(LocationSummary* locations) { 3158 Register in_reg_lo = locations->InAt(0).AsRegisterPairLow<Register>(); 3159 Register in_reg_hi = locations->InAt(0).AsRegisterPairHigh<Register>(); 3160 Location rhs = locations->InAt(1); 3161 Register out_reg_lo = locations->Out().AsRegisterPairLow<Register>(); 3162 Register out_reg_hi = locations->Out().AsRegisterPairHigh<Register>(); 3163 3164 if (rhs.IsConstant()) { 3165 uint64_t rot = CodeGenerator::GetInt64ValueOf(rhs.GetConstant()); 3166 // Map all rotations to +ve. equivalents on the interval [0,63]. 3167 rot &= kMaxLongShiftDistance; 3168 // For rotates over a word in size, 'pre-rotate' by 32-bits to keep rotate 3169 // logic below to a simple pair of binary orr. 3170 // (e.g. 34 bits == in_reg swap + 2 bits right.) 3171 if (rot >= kArmBitsPerWord) { 3172 rot -= kArmBitsPerWord; 3173 std::swap(in_reg_hi, in_reg_lo); 3174 } 3175 // Rotate, or mov to out for zero or word size rotations. 3176 if (rot != 0u) { 3177 __ Lsr(out_reg_hi, in_reg_hi, rot); 3178 __ orr(out_reg_hi, out_reg_hi, ShifterOperand(in_reg_lo, arm::LSL, kArmBitsPerWord - rot)); 3179 __ Lsr(out_reg_lo, in_reg_lo, rot); 3180 __ orr(out_reg_lo, out_reg_lo, ShifterOperand(in_reg_hi, arm::LSL, kArmBitsPerWord - rot)); 3181 } else { 3182 __ Mov(out_reg_lo, in_reg_lo); 3183 __ Mov(out_reg_hi, in_reg_hi); 3184 } 3185 } else { 3186 Register shift_right = locations->GetTemp(0).AsRegister<Register>(); 3187 Register shift_left = locations->GetTemp(1).AsRegister<Register>(); 3188 Label end; 3189 Label shift_by_32_plus_shift_right; 3190 3191 __ and_(shift_right, rhs.AsRegister<Register>(), ShifterOperand(0x1F)); 3192 __ Lsrs(shift_left, rhs.AsRegister<Register>(), 6); 3193 __ rsb(shift_left, shift_right, ShifterOperand(kArmBitsPerWord), AL, kCcKeep); 3194 __ b(&shift_by_32_plus_shift_right, CC); 3195 3196 // out_reg_hi = (reg_hi << shift_left) | (reg_lo >> shift_right). 3197 // out_reg_lo = (reg_lo << shift_left) | (reg_hi >> shift_right). 3198 __ Lsl(out_reg_hi, in_reg_hi, shift_left); 3199 __ Lsr(out_reg_lo, in_reg_lo, shift_right); 3200 __ add(out_reg_hi, out_reg_hi, ShifterOperand(out_reg_lo)); 3201 __ Lsl(out_reg_lo, in_reg_lo, shift_left); 3202 __ Lsr(shift_left, in_reg_hi, shift_right); 3203 __ add(out_reg_lo, out_reg_lo, ShifterOperand(shift_left)); 3204 __ b(&end); 3205 3206 __ Bind(&shift_by_32_plus_shift_right); // Shift by 32+shift_right. 3207 // out_reg_hi = (reg_hi >> shift_right) | (reg_lo << shift_left). 3208 // out_reg_lo = (reg_lo >> shift_right) | (reg_hi << shift_left). 3209 __ Lsr(out_reg_hi, in_reg_hi, shift_right); 3210 __ Lsl(out_reg_lo, in_reg_lo, shift_left); 3211 __ add(out_reg_hi, out_reg_hi, ShifterOperand(out_reg_lo)); 3212 __ Lsr(out_reg_lo, in_reg_lo, shift_right); 3213 __ Lsl(shift_right, in_reg_hi, shift_left); 3214 __ add(out_reg_lo, out_reg_lo, ShifterOperand(shift_right)); 3215 3216 __ Bind(&end); 3217 } 3218} 3219 3220void LocationsBuilderARM::VisitRor(HRor* ror) { 3221 LocationSummary* locations = 3222 new (GetGraph()->GetArena()) LocationSummary(ror, LocationSummary::kNoCall); 3223 switch (ror->GetResultType()) { 3224 case Primitive::kPrimInt: { 3225 locations->SetInAt(0, Location::RequiresRegister()); 3226 locations->SetInAt(1, Location::RegisterOrConstant(ror->InputAt(1))); 3227 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3228 break; 3229 } 3230 case Primitive::kPrimLong: { 3231 locations->SetInAt(0, Location::RequiresRegister()); 3232 if (ror->InputAt(1)->IsConstant()) { 3233 locations->SetInAt(1, Location::ConstantLocation(ror->InputAt(1)->AsConstant())); 3234 } else { 3235 locations->SetInAt(1, Location::RequiresRegister()); 3236 locations->AddTemp(Location::RequiresRegister()); 3237 locations->AddTemp(Location::RequiresRegister()); 3238 } 3239 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3240 break; 3241 } 3242 default: 3243 LOG(FATAL) << "Unexpected operation type " << ror->GetResultType(); 3244 } 3245} 3246 3247void InstructionCodeGeneratorARM::VisitRor(HRor* ror) { 3248 LocationSummary* locations = ror->GetLocations(); 3249 Primitive::Type type = ror->GetResultType(); 3250 switch (type) { 3251 case Primitive::kPrimInt: { 3252 HandleIntegerRotate(locations); 3253 break; 3254 } 3255 case Primitive::kPrimLong: { 3256 HandleLongRotate(locations); 3257 break; 3258 } 3259 default: 3260 LOG(FATAL) << "Unexpected operation type " << type; 3261 UNREACHABLE(); 3262 } 3263} 3264 3265void LocationsBuilderARM::HandleShift(HBinaryOperation* op) { 3266 DCHECK(op->IsShl() || op->IsShr() || op->IsUShr()); 3267 3268 LocationSummary* locations = 3269 new (GetGraph()->GetArena()) LocationSummary(op, LocationSummary::kNoCall); 3270 3271 switch (op->GetResultType()) { 3272 case Primitive::kPrimInt: { 3273 locations->SetInAt(0, Location::RequiresRegister()); 3274 if (op->InputAt(1)->IsConstant()) { 3275 locations->SetInAt(1, Location::ConstantLocation(op->InputAt(1)->AsConstant())); 3276 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3277 } else { 3278 locations->SetInAt(1, Location::RequiresRegister()); 3279 // Make the output overlap, as it will be used to hold the masked 3280 // second input. 3281 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3282 } 3283 break; 3284 } 3285 case Primitive::kPrimLong: { 3286 locations->SetInAt(0, Location::RequiresRegister()); 3287 if (op->InputAt(1)->IsConstant()) { 3288 locations->SetInAt(1, Location::ConstantLocation(op->InputAt(1)->AsConstant())); 3289 // For simplicity, use kOutputOverlap even though we only require that low registers 3290 // don't clash with high registers which the register allocator currently guarantees. 3291 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3292 } else { 3293 locations->SetInAt(1, Location::RequiresRegister()); 3294 locations->AddTemp(Location::RequiresRegister()); 3295 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3296 } 3297 break; 3298 } 3299 default: 3300 LOG(FATAL) << "Unexpected operation type " << op->GetResultType(); 3301 } 3302} 3303 3304void InstructionCodeGeneratorARM::HandleShift(HBinaryOperation* op) { 3305 DCHECK(op->IsShl() || op->IsShr() || op->IsUShr()); 3306 3307 LocationSummary* locations = op->GetLocations(); 3308 Location out = locations->Out(); 3309 Location first = locations->InAt(0); 3310 Location second = locations->InAt(1); 3311 3312 Primitive::Type type = op->GetResultType(); 3313 switch (type) { 3314 case Primitive::kPrimInt: { 3315 Register out_reg = out.AsRegister<Register>(); 3316 Register first_reg = first.AsRegister<Register>(); 3317 if (second.IsRegister()) { 3318 Register second_reg = second.AsRegister<Register>(); 3319 // ARM doesn't mask the shift count so we need to do it ourselves. 3320 __ and_(out_reg, second_reg, ShifterOperand(kMaxIntShiftDistance)); 3321 if (op->IsShl()) { 3322 __ Lsl(out_reg, first_reg, out_reg); 3323 } else if (op->IsShr()) { 3324 __ Asr(out_reg, first_reg, out_reg); 3325 } else { 3326 __ Lsr(out_reg, first_reg, out_reg); 3327 } 3328 } else { 3329 int32_t cst = second.GetConstant()->AsIntConstant()->GetValue(); 3330 uint32_t shift_value = cst & kMaxIntShiftDistance; 3331 if (shift_value == 0) { // ARM does not support shifting with 0 immediate. 3332 __ Mov(out_reg, first_reg); 3333 } else if (op->IsShl()) { 3334 __ Lsl(out_reg, first_reg, shift_value); 3335 } else if (op->IsShr()) { 3336 __ Asr(out_reg, first_reg, shift_value); 3337 } else { 3338 __ Lsr(out_reg, first_reg, shift_value); 3339 } 3340 } 3341 break; 3342 } 3343 case Primitive::kPrimLong: { 3344 Register o_h = out.AsRegisterPairHigh<Register>(); 3345 Register o_l = out.AsRegisterPairLow<Register>(); 3346 3347 Register high = first.AsRegisterPairHigh<Register>(); 3348 Register low = first.AsRegisterPairLow<Register>(); 3349 3350 if (second.IsRegister()) { 3351 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3352 3353 Register second_reg = second.AsRegister<Register>(); 3354 3355 if (op->IsShl()) { 3356 __ and_(o_l, second_reg, ShifterOperand(kMaxLongShiftDistance)); 3357 // Shift the high part 3358 __ Lsl(o_h, high, o_l); 3359 // Shift the low part and `or` what overflew on the high part 3360 __ rsb(temp, o_l, ShifterOperand(kArmBitsPerWord)); 3361 __ Lsr(temp, low, temp); 3362 __ orr(o_h, o_h, ShifterOperand(temp)); 3363 // If the shift is > 32 bits, override the high part 3364 __ subs(temp, o_l, ShifterOperand(kArmBitsPerWord)); 3365 __ it(PL); 3366 __ Lsl(o_h, low, temp, PL); 3367 // Shift the low part 3368 __ Lsl(o_l, low, o_l); 3369 } else if (op->IsShr()) { 3370 __ and_(o_h, second_reg, ShifterOperand(kMaxLongShiftDistance)); 3371 // Shift the low part 3372 __ Lsr(o_l, low, o_h); 3373 // Shift the high part and `or` what underflew on the low part 3374 __ rsb(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3375 __ Lsl(temp, high, temp); 3376 __ orr(o_l, o_l, ShifterOperand(temp)); 3377 // If the shift is > 32 bits, override the low part 3378 __ subs(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3379 __ it(PL); 3380 __ Asr(o_l, high, temp, PL); 3381 // Shift the high part 3382 __ Asr(o_h, high, o_h); 3383 } else { 3384 __ and_(o_h, second_reg, ShifterOperand(kMaxLongShiftDistance)); 3385 // same as Shr except we use `Lsr`s and not `Asr`s 3386 __ Lsr(o_l, low, o_h); 3387 __ rsb(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3388 __ Lsl(temp, high, temp); 3389 __ orr(o_l, o_l, ShifterOperand(temp)); 3390 __ subs(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3391 __ it(PL); 3392 __ Lsr(o_l, high, temp, PL); 3393 __ Lsr(o_h, high, o_h); 3394 } 3395 } else { 3396 // Register allocator doesn't create partial overlap. 3397 DCHECK_NE(o_l, high); 3398 DCHECK_NE(o_h, low); 3399 int32_t cst = second.GetConstant()->AsIntConstant()->GetValue(); 3400 uint32_t shift_value = cst & kMaxLongShiftDistance; 3401 if (shift_value > 32) { 3402 if (op->IsShl()) { 3403 __ Lsl(o_h, low, shift_value - 32); 3404 __ LoadImmediate(o_l, 0); 3405 } else if (op->IsShr()) { 3406 __ Asr(o_l, high, shift_value - 32); 3407 __ Asr(o_h, high, 31); 3408 } else { 3409 __ Lsr(o_l, high, shift_value - 32); 3410 __ LoadImmediate(o_h, 0); 3411 } 3412 } else if (shift_value == 32) { 3413 if (op->IsShl()) { 3414 __ mov(o_h, ShifterOperand(low)); 3415 __ LoadImmediate(o_l, 0); 3416 } else if (op->IsShr()) { 3417 __ mov(o_l, ShifterOperand(high)); 3418 __ Asr(o_h, high, 31); 3419 } else { 3420 __ mov(o_l, ShifterOperand(high)); 3421 __ LoadImmediate(o_h, 0); 3422 } 3423 } else if (shift_value == 1) { 3424 if (op->IsShl()) { 3425 __ Lsls(o_l, low, 1); 3426 __ adc(o_h, high, ShifterOperand(high)); 3427 } else if (op->IsShr()) { 3428 __ Asrs(o_h, high, 1); 3429 __ Rrx(o_l, low); 3430 } else { 3431 __ Lsrs(o_h, high, 1); 3432 __ Rrx(o_l, low); 3433 } 3434 } else { 3435 DCHECK(2 <= shift_value && shift_value < 32) << shift_value; 3436 if (op->IsShl()) { 3437 __ Lsl(o_h, high, shift_value); 3438 __ orr(o_h, o_h, ShifterOperand(low, LSR, 32 - shift_value)); 3439 __ Lsl(o_l, low, shift_value); 3440 } else if (op->IsShr()) { 3441 __ Lsr(o_l, low, shift_value); 3442 __ orr(o_l, o_l, ShifterOperand(high, LSL, 32 - shift_value)); 3443 __ Asr(o_h, high, shift_value); 3444 } else { 3445 __ Lsr(o_l, low, shift_value); 3446 __ orr(o_l, o_l, ShifterOperand(high, LSL, 32 - shift_value)); 3447 __ Lsr(o_h, high, shift_value); 3448 } 3449 } 3450 } 3451 break; 3452 } 3453 default: 3454 LOG(FATAL) << "Unexpected operation type " << type; 3455 UNREACHABLE(); 3456 } 3457} 3458 3459void LocationsBuilderARM::VisitShl(HShl* shl) { 3460 HandleShift(shl); 3461} 3462 3463void InstructionCodeGeneratorARM::VisitShl(HShl* shl) { 3464 HandleShift(shl); 3465} 3466 3467void LocationsBuilderARM::VisitShr(HShr* shr) { 3468 HandleShift(shr); 3469} 3470 3471void InstructionCodeGeneratorARM::VisitShr(HShr* shr) { 3472 HandleShift(shr); 3473} 3474 3475void LocationsBuilderARM::VisitUShr(HUShr* ushr) { 3476 HandleShift(ushr); 3477} 3478 3479void InstructionCodeGeneratorARM::VisitUShr(HUShr* ushr) { 3480 HandleShift(ushr); 3481} 3482 3483void LocationsBuilderARM::VisitNewInstance(HNewInstance* instruction) { 3484 LocationSummary* locations = 3485 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall); 3486 if (instruction->IsStringAlloc()) { 3487 locations->AddTemp(Location::RegisterLocation(kMethodRegisterArgument)); 3488 } else { 3489 InvokeRuntimeCallingConvention calling_convention; 3490 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 3491 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 3492 } 3493 locations->SetOut(Location::RegisterLocation(R0)); 3494} 3495 3496void InstructionCodeGeneratorARM::VisitNewInstance(HNewInstance* instruction) { 3497 // Note: if heap poisoning is enabled, the entry point takes cares 3498 // of poisoning the reference. 3499 if (instruction->IsStringAlloc()) { 3500 // String is allocated through StringFactory. Call NewEmptyString entry point. 3501 Register temp = instruction->GetLocations()->GetTemp(0).AsRegister<Register>(); 3502 MemberOffset code_offset = ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmWordSize); 3503 __ LoadFromOffset(kLoadWord, temp, TR, QUICK_ENTRY_POINT(pNewEmptyString)); 3504 __ LoadFromOffset(kLoadWord, LR, temp, code_offset.Int32Value()); 3505 __ blx(LR); 3506 codegen_->RecordPcInfo(instruction, instruction->GetDexPc()); 3507 } else { 3508 codegen_->InvokeRuntime(instruction->GetEntrypoint(), 3509 instruction, 3510 instruction->GetDexPc(), 3511 nullptr); 3512 CheckEntrypointTypes<kQuickAllocObjectWithAccessCheck, void*, uint32_t, ArtMethod*>(); 3513 } 3514} 3515 3516void LocationsBuilderARM::VisitNewArray(HNewArray* instruction) { 3517 LocationSummary* locations = 3518 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall); 3519 InvokeRuntimeCallingConvention calling_convention; 3520 locations->AddTemp(Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 3521 locations->SetOut(Location::RegisterLocation(R0)); 3522 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 3523 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(2))); 3524} 3525 3526void InstructionCodeGeneratorARM::VisitNewArray(HNewArray* instruction) { 3527 InvokeRuntimeCallingConvention calling_convention; 3528 __ LoadImmediate(calling_convention.GetRegisterAt(0), instruction->GetTypeIndex()); 3529 // Note: if heap poisoning is enabled, the entry point takes cares 3530 // of poisoning the reference. 3531 codegen_->InvokeRuntime(instruction->GetEntrypoint(), 3532 instruction, 3533 instruction->GetDexPc(), 3534 nullptr); 3535 CheckEntrypointTypes<kQuickAllocArrayWithAccessCheck, void*, uint32_t, int32_t, ArtMethod*>(); 3536} 3537 3538void LocationsBuilderARM::VisitParameterValue(HParameterValue* instruction) { 3539 LocationSummary* locations = 3540 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 3541 Location location = parameter_visitor_.GetNextLocation(instruction->GetType()); 3542 if (location.IsStackSlot()) { 3543 location = Location::StackSlot(location.GetStackIndex() + codegen_->GetFrameSize()); 3544 } else if (location.IsDoubleStackSlot()) { 3545 location = Location::DoubleStackSlot(location.GetStackIndex() + codegen_->GetFrameSize()); 3546 } 3547 locations->SetOut(location); 3548} 3549 3550void InstructionCodeGeneratorARM::VisitParameterValue( 3551 HParameterValue* instruction ATTRIBUTE_UNUSED) { 3552 // Nothing to do, the parameter is already at its location. 3553} 3554 3555void LocationsBuilderARM::VisitCurrentMethod(HCurrentMethod* instruction) { 3556 LocationSummary* locations = 3557 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 3558 locations->SetOut(Location::RegisterLocation(kMethodRegisterArgument)); 3559} 3560 3561void InstructionCodeGeneratorARM::VisitCurrentMethod(HCurrentMethod* instruction ATTRIBUTE_UNUSED) { 3562 // Nothing to do, the method is already at its location. 3563} 3564 3565void LocationsBuilderARM::VisitNot(HNot* not_) { 3566 LocationSummary* locations = 3567 new (GetGraph()->GetArena()) LocationSummary(not_, LocationSummary::kNoCall); 3568 locations->SetInAt(0, Location::RequiresRegister()); 3569 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3570} 3571 3572void InstructionCodeGeneratorARM::VisitNot(HNot* not_) { 3573 LocationSummary* locations = not_->GetLocations(); 3574 Location out = locations->Out(); 3575 Location in = locations->InAt(0); 3576 switch (not_->GetResultType()) { 3577 case Primitive::kPrimInt: 3578 __ mvn(out.AsRegister<Register>(), ShifterOperand(in.AsRegister<Register>())); 3579 break; 3580 3581 case Primitive::kPrimLong: 3582 __ mvn(out.AsRegisterPairLow<Register>(), 3583 ShifterOperand(in.AsRegisterPairLow<Register>())); 3584 __ mvn(out.AsRegisterPairHigh<Register>(), 3585 ShifterOperand(in.AsRegisterPairHigh<Register>())); 3586 break; 3587 3588 default: 3589 LOG(FATAL) << "Unimplemented type for not operation " << not_->GetResultType(); 3590 } 3591} 3592 3593void LocationsBuilderARM::VisitBooleanNot(HBooleanNot* bool_not) { 3594 LocationSummary* locations = 3595 new (GetGraph()->GetArena()) LocationSummary(bool_not, LocationSummary::kNoCall); 3596 locations->SetInAt(0, Location::RequiresRegister()); 3597 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3598} 3599 3600void InstructionCodeGeneratorARM::VisitBooleanNot(HBooleanNot* bool_not) { 3601 LocationSummary* locations = bool_not->GetLocations(); 3602 Location out = locations->Out(); 3603 Location in = locations->InAt(0); 3604 __ eor(out.AsRegister<Register>(), in.AsRegister<Register>(), ShifterOperand(1)); 3605} 3606 3607void LocationsBuilderARM::VisitCompare(HCompare* compare) { 3608 LocationSummary* locations = 3609 new (GetGraph()->GetArena()) LocationSummary(compare, LocationSummary::kNoCall); 3610 switch (compare->InputAt(0)->GetType()) { 3611 case Primitive::kPrimBoolean: 3612 case Primitive::kPrimByte: 3613 case Primitive::kPrimShort: 3614 case Primitive::kPrimChar: 3615 case Primitive::kPrimInt: 3616 case Primitive::kPrimLong: { 3617 locations->SetInAt(0, Location::RequiresRegister()); 3618 locations->SetInAt(1, Location::RequiresRegister()); 3619 // Output overlaps because it is written before doing the low comparison. 3620 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3621 break; 3622 } 3623 case Primitive::kPrimFloat: 3624 case Primitive::kPrimDouble: { 3625 locations->SetInAt(0, Location::RequiresFpuRegister()); 3626 locations->SetInAt(1, Location::RequiresFpuRegister()); 3627 locations->SetOut(Location::RequiresRegister()); 3628 break; 3629 } 3630 default: 3631 LOG(FATAL) << "Unexpected type for compare operation " << compare->InputAt(0)->GetType(); 3632 } 3633} 3634 3635void InstructionCodeGeneratorARM::VisitCompare(HCompare* compare) { 3636 LocationSummary* locations = compare->GetLocations(); 3637 Register out = locations->Out().AsRegister<Register>(); 3638 Location left = locations->InAt(0); 3639 Location right = locations->InAt(1); 3640 3641 Label less, greater, done; 3642 Primitive::Type type = compare->InputAt(0)->GetType(); 3643 Condition less_cond; 3644 switch (type) { 3645 case Primitive::kPrimBoolean: 3646 case Primitive::kPrimByte: 3647 case Primitive::kPrimShort: 3648 case Primitive::kPrimChar: 3649 case Primitive::kPrimInt: { 3650 __ LoadImmediate(out, 0); 3651 __ cmp(left.AsRegister<Register>(), 3652 ShifterOperand(right.AsRegister<Register>())); // Signed compare. 3653 less_cond = LT; 3654 break; 3655 } 3656 case Primitive::kPrimLong: { 3657 __ cmp(left.AsRegisterPairHigh<Register>(), 3658 ShifterOperand(right.AsRegisterPairHigh<Register>())); // Signed compare. 3659 __ b(&less, LT); 3660 __ b(&greater, GT); 3661 // Do LoadImmediate before the last `cmp`, as LoadImmediate might affect the status flags. 3662 __ LoadImmediate(out, 0); 3663 __ cmp(left.AsRegisterPairLow<Register>(), 3664 ShifterOperand(right.AsRegisterPairLow<Register>())); // Unsigned compare. 3665 less_cond = LO; 3666 break; 3667 } 3668 case Primitive::kPrimFloat: 3669 case Primitive::kPrimDouble: { 3670 __ LoadImmediate(out, 0); 3671 if (type == Primitive::kPrimFloat) { 3672 __ vcmps(left.AsFpuRegister<SRegister>(), right.AsFpuRegister<SRegister>()); 3673 } else { 3674 __ vcmpd(FromLowSToD(left.AsFpuRegisterPairLow<SRegister>()), 3675 FromLowSToD(right.AsFpuRegisterPairLow<SRegister>())); 3676 } 3677 __ vmstat(); // transfer FP status register to ARM APSR. 3678 less_cond = ARMFPCondition(kCondLT, compare->IsGtBias()); 3679 break; 3680 } 3681 default: 3682 LOG(FATAL) << "Unexpected compare type " << type; 3683 UNREACHABLE(); 3684 } 3685 3686 __ b(&done, EQ); 3687 __ b(&less, less_cond); 3688 3689 __ Bind(&greater); 3690 __ LoadImmediate(out, 1); 3691 __ b(&done); 3692 3693 __ Bind(&less); 3694 __ LoadImmediate(out, -1); 3695 3696 __ Bind(&done); 3697} 3698 3699void LocationsBuilderARM::VisitPhi(HPhi* instruction) { 3700 LocationSummary* locations = 3701 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 3702 for (size_t i = 0, e = locations->GetInputCount(); i < e; ++i) { 3703 locations->SetInAt(i, Location::Any()); 3704 } 3705 locations->SetOut(Location::Any()); 3706} 3707 3708void InstructionCodeGeneratorARM::VisitPhi(HPhi* instruction ATTRIBUTE_UNUSED) { 3709 LOG(FATAL) << "Unreachable"; 3710} 3711 3712void CodeGeneratorARM::GenerateMemoryBarrier(MemBarrierKind kind) { 3713 // TODO (ported from quick): revisit ARM barrier kinds. 3714 DmbOptions flavor = DmbOptions::ISH; // Quiet C++ warnings. 3715 switch (kind) { 3716 case MemBarrierKind::kAnyStore: 3717 case MemBarrierKind::kLoadAny: 3718 case MemBarrierKind::kAnyAny: { 3719 flavor = DmbOptions::ISH; 3720 break; 3721 } 3722 case MemBarrierKind::kStoreStore: { 3723 flavor = DmbOptions::ISHST; 3724 break; 3725 } 3726 default: 3727 LOG(FATAL) << "Unexpected memory barrier " << kind; 3728 } 3729 __ dmb(flavor); 3730} 3731 3732void InstructionCodeGeneratorARM::GenerateWideAtomicLoad(Register addr, 3733 uint32_t offset, 3734 Register out_lo, 3735 Register out_hi) { 3736 if (offset != 0) { 3737 // Ensure `out_lo` is different from `addr`, so that loading 3738 // `offset` into `out_lo` does not clutter `addr`. 3739 DCHECK_NE(out_lo, addr); 3740 __ LoadImmediate(out_lo, offset); 3741 __ add(IP, addr, ShifterOperand(out_lo)); 3742 addr = IP; 3743 } 3744 __ ldrexd(out_lo, out_hi, addr); 3745} 3746 3747void InstructionCodeGeneratorARM::GenerateWideAtomicStore(Register addr, 3748 uint32_t offset, 3749 Register value_lo, 3750 Register value_hi, 3751 Register temp1, 3752 Register temp2, 3753 HInstruction* instruction) { 3754 Label fail; 3755 if (offset != 0) { 3756 __ LoadImmediate(temp1, offset); 3757 __ add(IP, addr, ShifterOperand(temp1)); 3758 addr = IP; 3759 } 3760 __ Bind(&fail); 3761 // We need a load followed by store. (The address used in a STREX instruction must 3762 // be the same as the address in the most recently executed LDREX instruction.) 3763 __ ldrexd(temp1, temp2, addr); 3764 codegen_->MaybeRecordImplicitNullCheck(instruction); 3765 __ strexd(temp1, value_lo, value_hi, addr); 3766 __ CompareAndBranchIfNonZero(temp1, &fail); 3767} 3768 3769void LocationsBuilderARM::HandleFieldSet(HInstruction* instruction, const FieldInfo& field_info) { 3770 DCHECK(instruction->IsInstanceFieldSet() || instruction->IsStaticFieldSet()); 3771 3772 LocationSummary* locations = 3773 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 3774 locations->SetInAt(0, Location::RequiresRegister()); 3775 3776 Primitive::Type field_type = field_info.GetFieldType(); 3777 if (Primitive::IsFloatingPointType(field_type)) { 3778 locations->SetInAt(1, Location::RequiresFpuRegister()); 3779 } else { 3780 locations->SetInAt(1, Location::RequiresRegister()); 3781 } 3782 3783 bool is_wide = field_type == Primitive::kPrimLong || field_type == Primitive::kPrimDouble; 3784 bool generate_volatile = field_info.IsVolatile() 3785 && is_wide 3786 && !codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 3787 bool needs_write_barrier = 3788 CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1)); 3789 // Temporary registers for the write barrier. 3790 // TODO: consider renaming StoreNeedsWriteBarrier to StoreNeedsGCMark. 3791 if (needs_write_barrier) { 3792 locations->AddTemp(Location::RequiresRegister()); // Possibly used for reference poisoning too. 3793 locations->AddTemp(Location::RequiresRegister()); 3794 } else if (generate_volatile) { 3795 // ARM encoding have some additional constraints for ldrexd/strexd: 3796 // - registers need to be consecutive 3797 // - the first register should be even but not R14. 3798 // We don't test for ARM yet, and the assertion makes sure that we 3799 // revisit this if we ever enable ARM encoding. 3800 DCHECK_EQ(InstructionSet::kThumb2, codegen_->GetInstructionSet()); 3801 3802 locations->AddTemp(Location::RequiresRegister()); 3803 locations->AddTemp(Location::RequiresRegister()); 3804 if (field_type == Primitive::kPrimDouble) { 3805 // For doubles we need two more registers to copy the value. 3806 locations->AddTemp(Location::RegisterLocation(R2)); 3807 locations->AddTemp(Location::RegisterLocation(R3)); 3808 } 3809 } 3810} 3811 3812void InstructionCodeGeneratorARM::HandleFieldSet(HInstruction* instruction, 3813 const FieldInfo& field_info, 3814 bool value_can_be_null) { 3815 DCHECK(instruction->IsInstanceFieldSet() || instruction->IsStaticFieldSet()); 3816 3817 LocationSummary* locations = instruction->GetLocations(); 3818 Register base = locations->InAt(0).AsRegister<Register>(); 3819 Location value = locations->InAt(1); 3820 3821 bool is_volatile = field_info.IsVolatile(); 3822 bool atomic_ldrd_strd = codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 3823 Primitive::Type field_type = field_info.GetFieldType(); 3824 uint32_t offset = field_info.GetFieldOffset().Uint32Value(); 3825 bool needs_write_barrier = 3826 CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1)); 3827 3828 if (is_volatile) { 3829 codegen_->GenerateMemoryBarrier(MemBarrierKind::kAnyStore); 3830 } 3831 3832 switch (field_type) { 3833 case Primitive::kPrimBoolean: 3834 case Primitive::kPrimByte: { 3835 __ StoreToOffset(kStoreByte, value.AsRegister<Register>(), base, offset); 3836 break; 3837 } 3838 3839 case Primitive::kPrimShort: 3840 case Primitive::kPrimChar: { 3841 __ StoreToOffset(kStoreHalfword, value.AsRegister<Register>(), base, offset); 3842 break; 3843 } 3844 3845 case Primitive::kPrimInt: 3846 case Primitive::kPrimNot: { 3847 if (kPoisonHeapReferences && needs_write_barrier) { 3848 // Note that in the case where `value` is a null reference, 3849 // we do not enter this block, as a null reference does not 3850 // need poisoning. 3851 DCHECK_EQ(field_type, Primitive::kPrimNot); 3852 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3853 __ Mov(temp, value.AsRegister<Register>()); 3854 __ PoisonHeapReference(temp); 3855 __ StoreToOffset(kStoreWord, temp, base, offset); 3856 } else { 3857 __ StoreToOffset(kStoreWord, value.AsRegister<Register>(), base, offset); 3858 } 3859 break; 3860 } 3861 3862 case Primitive::kPrimLong: { 3863 if (is_volatile && !atomic_ldrd_strd) { 3864 GenerateWideAtomicStore(base, offset, 3865 value.AsRegisterPairLow<Register>(), 3866 value.AsRegisterPairHigh<Register>(), 3867 locations->GetTemp(0).AsRegister<Register>(), 3868 locations->GetTemp(1).AsRegister<Register>(), 3869 instruction); 3870 } else { 3871 __ StoreToOffset(kStoreWordPair, value.AsRegisterPairLow<Register>(), base, offset); 3872 codegen_->MaybeRecordImplicitNullCheck(instruction); 3873 } 3874 break; 3875 } 3876 3877 case Primitive::kPrimFloat: { 3878 __ StoreSToOffset(value.AsFpuRegister<SRegister>(), base, offset); 3879 break; 3880 } 3881 3882 case Primitive::kPrimDouble: { 3883 DRegister value_reg = FromLowSToD(value.AsFpuRegisterPairLow<SRegister>()); 3884 if (is_volatile && !atomic_ldrd_strd) { 3885 Register value_reg_lo = locations->GetTemp(0).AsRegister<Register>(); 3886 Register value_reg_hi = locations->GetTemp(1).AsRegister<Register>(); 3887 3888 __ vmovrrd(value_reg_lo, value_reg_hi, value_reg); 3889 3890 GenerateWideAtomicStore(base, offset, 3891 value_reg_lo, 3892 value_reg_hi, 3893 locations->GetTemp(2).AsRegister<Register>(), 3894 locations->GetTemp(3).AsRegister<Register>(), 3895 instruction); 3896 } else { 3897 __ StoreDToOffset(value_reg, base, offset); 3898 codegen_->MaybeRecordImplicitNullCheck(instruction); 3899 } 3900 break; 3901 } 3902 3903 case Primitive::kPrimVoid: 3904 LOG(FATAL) << "Unreachable type " << field_type; 3905 UNREACHABLE(); 3906 } 3907 3908 // Longs and doubles are handled in the switch. 3909 if (field_type != Primitive::kPrimLong && field_type != Primitive::kPrimDouble) { 3910 codegen_->MaybeRecordImplicitNullCheck(instruction); 3911 } 3912 3913 if (CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1))) { 3914 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3915 Register card = locations->GetTemp(1).AsRegister<Register>(); 3916 codegen_->MarkGCCard( 3917 temp, card, base, value.AsRegister<Register>(), value_can_be_null); 3918 } 3919 3920 if (is_volatile) { 3921 codegen_->GenerateMemoryBarrier(MemBarrierKind::kAnyAny); 3922 } 3923} 3924 3925void LocationsBuilderARM::HandleFieldGet(HInstruction* instruction, const FieldInfo& field_info) { 3926 DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet()); 3927 3928 bool object_field_get_with_read_barrier = 3929 kEmitCompilerReadBarrier && (field_info.GetFieldType() == Primitive::kPrimNot); 3930 LocationSummary* locations = 3931 new (GetGraph()->GetArena()) LocationSummary(instruction, 3932 object_field_get_with_read_barrier ? 3933 LocationSummary::kCallOnSlowPath : 3934 LocationSummary::kNoCall); 3935 locations->SetInAt(0, Location::RequiresRegister()); 3936 3937 bool volatile_for_double = field_info.IsVolatile() 3938 && (field_info.GetFieldType() == Primitive::kPrimDouble) 3939 && !codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 3940 // The output overlaps in case of volatile long: we don't want the 3941 // code generated by GenerateWideAtomicLoad to overwrite the 3942 // object's location. Likewise, in the case of an object field get 3943 // with read barriers enabled, we do not want the load to overwrite 3944 // the object's location, as we need it to emit the read barrier. 3945 bool overlap = (field_info.IsVolatile() && (field_info.GetFieldType() == Primitive::kPrimLong)) || 3946 object_field_get_with_read_barrier; 3947 3948 if (Primitive::IsFloatingPointType(instruction->GetType())) { 3949 locations->SetOut(Location::RequiresFpuRegister()); 3950 } else { 3951 locations->SetOut(Location::RequiresRegister(), 3952 (overlap ? Location::kOutputOverlap : Location::kNoOutputOverlap)); 3953 } 3954 if (volatile_for_double) { 3955 // ARM encoding have some additional constraints for ldrexd/strexd: 3956 // - registers need to be consecutive 3957 // - the first register should be even but not R14. 3958 // We don't test for ARM yet, and the assertion makes sure that we 3959 // revisit this if we ever enable ARM encoding. 3960 DCHECK_EQ(InstructionSet::kThumb2, codegen_->GetInstructionSet()); 3961 locations->AddTemp(Location::RequiresRegister()); 3962 locations->AddTemp(Location::RequiresRegister()); 3963 } else if (object_field_get_with_read_barrier && kUseBakerReadBarrier) { 3964 // We need a temporary register for the read barrier marking slow 3965 // path in CodeGeneratorARM::GenerateFieldLoadWithBakerReadBarrier. 3966 locations->AddTemp(Location::RequiresRegister()); 3967 } 3968} 3969 3970Location LocationsBuilderARM::ArmEncodableConstantOrRegister(HInstruction* constant, 3971 Opcode opcode) { 3972 DCHECK(!Primitive::IsFloatingPointType(constant->GetType())); 3973 if (constant->IsConstant() && 3974 CanEncodeConstantAsImmediate(constant->AsConstant(), opcode)) { 3975 return Location::ConstantLocation(constant->AsConstant()); 3976 } 3977 return Location::RequiresRegister(); 3978} 3979 3980bool LocationsBuilderARM::CanEncodeConstantAsImmediate(HConstant* input_cst, 3981 Opcode opcode) { 3982 uint64_t value = static_cast<uint64_t>(Int64FromConstant(input_cst)); 3983 if (Primitive::Is64BitType(input_cst->GetType())) { 3984 return CanEncodeConstantAsImmediate(Low32Bits(value), opcode) && 3985 CanEncodeConstantAsImmediate(High32Bits(value), opcode); 3986 } else { 3987 return CanEncodeConstantAsImmediate(Low32Bits(value), opcode); 3988 } 3989} 3990 3991bool LocationsBuilderARM::CanEncodeConstantAsImmediate(uint32_t value, Opcode opcode) { 3992 ShifterOperand so; 3993 ArmAssembler* assembler = codegen_->GetAssembler(); 3994 if (assembler->ShifterOperandCanHold(kNoRegister, kNoRegister, opcode, value, &so)) { 3995 return true; 3996 } 3997 Opcode neg_opcode = kNoOperand; 3998 switch (opcode) { 3999 case AND: 4000 neg_opcode = BIC; 4001 break; 4002 case ORR: 4003 neg_opcode = ORN; 4004 break; 4005 default: 4006 return false; 4007 } 4008 return assembler->ShifterOperandCanHold(kNoRegister, kNoRegister, neg_opcode, ~value, &so); 4009} 4010 4011void InstructionCodeGeneratorARM::HandleFieldGet(HInstruction* instruction, 4012 const FieldInfo& field_info) { 4013 DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet()); 4014 4015 LocationSummary* locations = instruction->GetLocations(); 4016 Location base_loc = locations->InAt(0); 4017 Register base = base_loc.AsRegister<Register>(); 4018 Location out = locations->Out(); 4019 bool is_volatile = field_info.IsVolatile(); 4020 bool atomic_ldrd_strd = codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 4021 Primitive::Type field_type = field_info.GetFieldType(); 4022 uint32_t offset = field_info.GetFieldOffset().Uint32Value(); 4023 4024 switch (field_type) { 4025 case Primitive::kPrimBoolean: 4026 __ LoadFromOffset(kLoadUnsignedByte, out.AsRegister<Register>(), base, offset); 4027 break; 4028 4029 case Primitive::kPrimByte: 4030 __ LoadFromOffset(kLoadSignedByte, out.AsRegister<Register>(), base, offset); 4031 break; 4032 4033 case Primitive::kPrimShort: 4034 __ LoadFromOffset(kLoadSignedHalfword, out.AsRegister<Register>(), base, offset); 4035 break; 4036 4037 case Primitive::kPrimChar: 4038 __ LoadFromOffset(kLoadUnsignedHalfword, out.AsRegister<Register>(), base, offset); 4039 break; 4040 4041 case Primitive::kPrimInt: 4042 __ LoadFromOffset(kLoadWord, out.AsRegister<Register>(), base, offset); 4043 break; 4044 4045 case Primitive::kPrimNot: { 4046 // /* HeapReference<Object> */ out = *(base + offset) 4047 if (kEmitCompilerReadBarrier && kUseBakerReadBarrier) { 4048 Location temp_loc = locations->GetTemp(0); 4049 // Note that a potential implicit null check is handled in this 4050 // CodeGeneratorARM::GenerateFieldLoadWithBakerReadBarrier call. 4051 codegen_->GenerateFieldLoadWithBakerReadBarrier( 4052 instruction, out, base, offset, temp_loc, /* needs_null_check */ true); 4053 if (is_volatile) { 4054 codegen_->GenerateMemoryBarrier(MemBarrierKind::kLoadAny); 4055 } 4056 } else { 4057 __ LoadFromOffset(kLoadWord, out.AsRegister<Register>(), base, offset); 4058 codegen_->MaybeRecordImplicitNullCheck(instruction); 4059 if (is_volatile) { 4060 codegen_->GenerateMemoryBarrier(MemBarrierKind::kLoadAny); 4061 } 4062 // If read barriers are enabled, emit read barriers other than 4063 // Baker's using a slow path (and also unpoison the loaded 4064 // reference, if heap poisoning is enabled). 4065 codegen_->MaybeGenerateReadBarrierSlow(instruction, out, out, base_loc, offset); 4066 } 4067 break; 4068 } 4069 4070 case Primitive::kPrimLong: 4071 if (is_volatile && !atomic_ldrd_strd) { 4072 GenerateWideAtomicLoad(base, offset, 4073 out.AsRegisterPairLow<Register>(), 4074 out.AsRegisterPairHigh<Register>()); 4075 } else { 4076 __ LoadFromOffset(kLoadWordPair, out.AsRegisterPairLow<Register>(), base, offset); 4077 } 4078 break; 4079 4080 case Primitive::kPrimFloat: 4081 __ LoadSFromOffset(out.AsFpuRegister<SRegister>(), base, offset); 4082 break; 4083 4084 case Primitive::kPrimDouble: { 4085 DRegister out_reg = FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()); 4086 if (is_volatile && !atomic_ldrd_strd) { 4087 Register lo = locations->GetTemp(0).AsRegister<Register>(); 4088 Register hi = locations->GetTemp(1).AsRegister<Register>(); 4089 GenerateWideAtomicLoad(base, offset, lo, hi); 4090 codegen_->MaybeRecordImplicitNullCheck(instruction); 4091 __ vmovdrr(out_reg, lo, hi); 4092 } else { 4093 __ LoadDFromOffset(out_reg, base, offset); 4094 codegen_->MaybeRecordImplicitNullCheck(instruction); 4095 } 4096 break; 4097 } 4098 4099 case Primitive::kPrimVoid: 4100 LOG(FATAL) << "Unreachable type " << field_type; 4101 UNREACHABLE(); 4102 } 4103 4104 if (field_type == Primitive::kPrimNot || field_type == Primitive::kPrimDouble) { 4105 // Potential implicit null checks, in the case of reference or 4106 // double fields, are handled in the previous switch statement. 4107 } else { 4108 codegen_->MaybeRecordImplicitNullCheck(instruction); 4109 } 4110 4111 if (is_volatile) { 4112 if (field_type == Primitive::kPrimNot) { 4113 // Memory barriers, in the case of references, are also handled 4114 // in the previous switch statement. 4115 } else { 4116 codegen_->GenerateMemoryBarrier(MemBarrierKind::kLoadAny); 4117 } 4118 } 4119} 4120 4121void LocationsBuilderARM::VisitInstanceFieldSet(HInstanceFieldSet* instruction) { 4122 HandleFieldSet(instruction, instruction->GetFieldInfo()); 4123} 4124 4125void InstructionCodeGeneratorARM::VisitInstanceFieldSet(HInstanceFieldSet* instruction) { 4126 HandleFieldSet(instruction, instruction->GetFieldInfo(), instruction->GetValueCanBeNull()); 4127} 4128 4129void LocationsBuilderARM::VisitInstanceFieldGet(HInstanceFieldGet* instruction) { 4130 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4131} 4132 4133void InstructionCodeGeneratorARM::VisitInstanceFieldGet(HInstanceFieldGet* instruction) { 4134 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4135} 4136 4137void LocationsBuilderARM::VisitStaticFieldGet(HStaticFieldGet* instruction) { 4138 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4139} 4140 4141void InstructionCodeGeneratorARM::VisitStaticFieldGet(HStaticFieldGet* instruction) { 4142 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4143} 4144 4145void LocationsBuilderARM::VisitStaticFieldSet(HStaticFieldSet* instruction) { 4146 HandleFieldSet(instruction, instruction->GetFieldInfo()); 4147} 4148 4149void InstructionCodeGeneratorARM::VisitStaticFieldSet(HStaticFieldSet* instruction) { 4150 HandleFieldSet(instruction, instruction->GetFieldInfo(), instruction->GetValueCanBeNull()); 4151} 4152 4153void LocationsBuilderARM::VisitUnresolvedInstanceFieldGet( 4154 HUnresolvedInstanceFieldGet* instruction) { 4155 FieldAccessCallingConventionARM calling_convention; 4156 codegen_->CreateUnresolvedFieldLocationSummary( 4157 instruction, instruction->GetFieldType(), calling_convention); 4158} 4159 4160void InstructionCodeGeneratorARM::VisitUnresolvedInstanceFieldGet( 4161 HUnresolvedInstanceFieldGet* instruction) { 4162 FieldAccessCallingConventionARM calling_convention; 4163 codegen_->GenerateUnresolvedFieldAccess(instruction, 4164 instruction->GetFieldType(), 4165 instruction->GetFieldIndex(), 4166 instruction->GetDexPc(), 4167 calling_convention); 4168} 4169 4170void LocationsBuilderARM::VisitUnresolvedInstanceFieldSet( 4171 HUnresolvedInstanceFieldSet* instruction) { 4172 FieldAccessCallingConventionARM calling_convention; 4173 codegen_->CreateUnresolvedFieldLocationSummary( 4174 instruction, instruction->GetFieldType(), calling_convention); 4175} 4176 4177void InstructionCodeGeneratorARM::VisitUnresolvedInstanceFieldSet( 4178 HUnresolvedInstanceFieldSet* instruction) { 4179 FieldAccessCallingConventionARM calling_convention; 4180 codegen_->GenerateUnresolvedFieldAccess(instruction, 4181 instruction->GetFieldType(), 4182 instruction->GetFieldIndex(), 4183 instruction->GetDexPc(), 4184 calling_convention); 4185} 4186 4187void LocationsBuilderARM::VisitUnresolvedStaticFieldGet( 4188 HUnresolvedStaticFieldGet* instruction) { 4189 FieldAccessCallingConventionARM calling_convention; 4190 codegen_->CreateUnresolvedFieldLocationSummary( 4191 instruction, instruction->GetFieldType(), calling_convention); 4192} 4193 4194void InstructionCodeGeneratorARM::VisitUnresolvedStaticFieldGet( 4195 HUnresolvedStaticFieldGet* instruction) { 4196 FieldAccessCallingConventionARM calling_convention; 4197 codegen_->GenerateUnresolvedFieldAccess(instruction, 4198 instruction->GetFieldType(), 4199 instruction->GetFieldIndex(), 4200 instruction->GetDexPc(), 4201 calling_convention); 4202} 4203 4204void LocationsBuilderARM::VisitUnresolvedStaticFieldSet( 4205 HUnresolvedStaticFieldSet* instruction) { 4206 FieldAccessCallingConventionARM calling_convention; 4207 codegen_->CreateUnresolvedFieldLocationSummary( 4208 instruction, instruction->GetFieldType(), calling_convention); 4209} 4210 4211void InstructionCodeGeneratorARM::VisitUnresolvedStaticFieldSet( 4212 HUnresolvedStaticFieldSet* instruction) { 4213 FieldAccessCallingConventionARM calling_convention; 4214 codegen_->GenerateUnresolvedFieldAccess(instruction, 4215 instruction->GetFieldType(), 4216 instruction->GetFieldIndex(), 4217 instruction->GetDexPc(), 4218 calling_convention); 4219} 4220 4221void LocationsBuilderARM::VisitNullCheck(HNullCheck* instruction) { 4222 LocationSummary::CallKind call_kind = instruction->CanThrowIntoCatchBlock() 4223 ? LocationSummary::kCallOnSlowPath 4224 : LocationSummary::kNoCall; 4225 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 4226 locations->SetInAt(0, Location::RequiresRegister()); 4227 if (instruction->HasUses()) { 4228 locations->SetOut(Location::SameAsFirstInput()); 4229 } 4230} 4231 4232void CodeGeneratorARM::GenerateImplicitNullCheck(HNullCheck* instruction) { 4233 if (CanMoveNullCheckToUser(instruction)) { 4234 return; 4235 } 4236 Location obj = instruction->GetLocations()->InAt(0); 4237 4238 __ LoadFromOffset(kLoadWord, IP, obj.AsRegister<Register>(), 0); 4239 RecordPcInfo(instruction, instruction->GetDexPc()); 4240} 4241 4242void CodeGeneratorARM::GenerateExplicitNullCheck(HNullCheck* instruction) { 4243 SlowPathCode* slow_path = new (GetGraph()->GetArena()) NullCheckSlowPathARM(instruction); 4244 AddSlowPath(slow_path); 4245 4246 LocationSummary* locations = instruction->GetLocations(); 4247 Location obj = locations->InAt(0); 4248 4249 __ CompareAndBranchIfZero(obj.AsRegister<Register>(), slow_path->GetEntryLabel()); 4250} 4251 4252void InstructionCodeGeneratorARM::VisitNullCheck(HNullCheck* instruction) { 4253 codegen_->GenerateNullCheck(instruction); 4254} 4255 4256void LocationsBuilderARM::VisitArrayGet(HArrayGet* instruction) { 4257 bool object_array_get_with_read_barrier = 4258 kEmitCompilerReadBarrier && (instruction->GetType() == Primitive::kPrimNot); 4259 LocationSummary* locations = 4260 new (GetGraph()->GetArena()) LocationSummary(instruction, 4261 object_array_get_with_read_barrier ? 4262 LocationSummary::kCallOnSlowPath : 4263 LocationSummary::kNoCall); 4264 locations->SetInAt(0, Location::RequiresRegister()); 4265 locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1))); 4266 if (Primitive::IsFloatingPointType(instruction->GetType())) { 4267 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 4268 } else { 4269 // The output overlaps in the case of an object array get with 4270 // read barriers enabled: we do not want the move to overwrite the 4271 // array's location, as we need it to emit the read barrier. 4272 locations->SetOut( 4273 Location::RequiresRegister(), 4274 object_array_get_with_read_barrier ? Location::kOutputOverlap : Location::kNoOutputOverlap); 4275 } 4276 // We need a temporary register for the read barrier marking slow 4277 // path in CodeGeneratorARM::GenerateArrayLoadWithBakerReadBarrier. 4278 if (object_array_get_with_read_barrier && kUseBakerReadBarrier) { 4279 locations->AddTemp(Location::RequiresRegister()); 4280 } 4281} 4282 4283void InstructionCodeGeneratorARM::VisitArrayGet(HArrayGet* instruction) { 4284 LocationSummary* locations = instruction->GetLocations(); 4285 Location obj_loc = locations->InAt(0); 4286 Register obj = obj_loc.AsRegister<Register>(); 4287 Location index = locations->InAt(1); 4288 Location out_loc = locations->Out(); 4289 4290 Primitive::Type type = instruction->GetType(); 4291 switch (type) { 4292 case Primitive::kPrimBoolean: { 4293 uint32_t data_offset = mirror::Array::DataOffset(sizeof(uint8_t)).Uint32Value(); 4294 Register out = out_loc.AsRegister<Register>(); 4295 if (index.IsConstant()) { 4296 size_t offset = 4297 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_1) + data_offset; 4298 __ LoadFromOffset(kLoadUnsignedByte, out, obj, offset); 4299 } else { 4300 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>())); 4301 __ LoadFromOffset(kLoadUnsignedByte, out, IP, data_offset); 4302 } 4303 break; 4304 } 4305 4306 case Primitive::kPrimByte: { 4307 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int8_t)).Uint32Value(); 4308 Register out = out_loc.AsRegister<Register>(); 4309 if (index.IsConstant()) { 4310 size_t offset = 4311 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_1) + data_offset; 4312 __ LoadFromOffset(kLoadSignedByte, out, obj, offset); 4313 } else { 4314 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>())); 4315 __ LoadFromOffset(kLoadSignedByte, out, IP, data_offset); 4316 } 4317 break; 4318 } 4319 4320 case Primitive::kPrimShort: { 4321 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int16_t)).Uint32Value(); 4322 Register out = out_loc.AsRegister<Register>(); 4323 if (index.IsConstant()) { 4324 size_t offset = 4325 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_2) + data_offset; 4326 __ LoadFromOffset(kLoadSignedHalfword, out, obj, offset); 4327 } else { 4328 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_2)); 4329 __ LoadFromOffset(kLoadSignedHalfword, out, IP, data_offset); 4330 } 4331 break; 4332 } 4333 4334 case Primitive::kPrimChar: { 4335 uint32_t data_offset = mirror::Array::DataOffset(sizeof(uint16_t)).Uint32Value(); 4336 Register out = out_loc.AsRegister<Register>(); 4337 if (index.IsConstant()) { 4338 size_t offset = 4339 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_2) + data_offset; 4340 __ LoadFromOffset(kLoadUnsignedHalfword, out, obj, offset); 4341 } else { 4342 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_2)); 4343 __ LoadFromOffset(kLoadUnsignedHalfword, out, IP, data_offset); 4344 } 4345 break; 4346 } 4347 4348 case Primitive::kPrimInt: { 4349 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int32_t)).Uint32Value(); 4350 Register out = out_loc.AsRegister<Register>(); 4351 if (index.IsConstant()) { 4352 size_t offset = 4353 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4354 __ LoadFromOffset(kLoadWord, out, obj, offset); 4355 } else { 4356 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 4357 __ LoadFromOffset(kLoadWord, out, IP, data_offset); 4358 } 4359 break; 4360 } 4361 4362 case Primitive::kPrimNot: { 4363 static_assert( 4364 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 4365 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 4366 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int32_t)).Uint32Value(); 4367 // /* HeapReference<Object> */ out = 4368 // *(obj + data_offset + index * sizeof(HeapReference<Object>)) 4369 if (kEmitCompilerReadBarrier && kUseBakerReadBarrier) { 4370 Location temp = locations->GetTemp(0); 4371 // Note that a potential implicit null check is handled in this 4372 // CodeGeneratorARM::GenerateArrayLoadWithBakerReadBarrier call. 4373 codegen_->GenerateArrayLoadWithBakerReadBarrier( 4374 instruction, out_loc, obj, data_offset, index, temp, /* needs_null_check */ true); 4375 } else { 4376 Register out = out_loc.AsRegister<Register>(); 4377 if (index.IsConstant()) { 4378 size_t offset = 4379 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4380 __ LoadFromOffset(kLoadWord, out, obj, offset); 4381 codegen_->MaybeRecordImplicitNullCheck(instruction); 4382 // If read barriers are enabled, emit read barriers other than 4383 // Baker's using a slow path (and also unpoison the loaded 4384 // reference, if heap poisoning is enabled). 4385 codegen_->MaybeGenerateReadBarrierSlow(instruction, out_loc, out_loc, obj_loc, offset); 4386 } else { 4387 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 4388 __ LoadFromOffset(kLoadWord, out, IP, data_offset); 4389 codegen_->MaybeRecordImplicitNullCheck(instruction); 4390 // If read barriers are enabled, emit read barriers other than 4391 // Baker's using a slow path (and also unpoison the loaded 4392 // reference, if heap poisoning is enabled). 4393 codegen_->MaybeGenerateReadBarrierSlow( 4394 instruction, out_loc, out_loc, obj_loc, data_offset, index); 4395 } 4396 } 4397 break; 4398 } 4399 4400 case Primitive::kPrimLong: { 4401 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int64_t)).Uint32Value(); 4402 if (index.IsConstant()) { 4403 size_t offset = 4404 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 4405 __ LoadFromOffset(kLoadWordPair, out_loc.AsRegisterPairLow<Register>(), obj, offset); 4406 } else { 4407 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 4408 __ LoadFromOffset(kLoadWordPair, out_loc.AsRegisterPairLow<Register>(), IP, data_offset); 4409 } 4410 break; 4411 } 4412 4413 case Primitive::kPrimFloat: { 4414 uint32_t data_offset = mirror::Array::DataOffset(sizeof(float)).Uint32Value(); 4415 SRegister out = out_loc.AsFpuRegister<SRegister>(); 4416 if (index.IsConstant()) { 4417 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4418 __ LoadSFromOffset(out, obj, offset); 4419 } else { 4420 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 4421 __ LoadSFromOffset(out, IP, data_offset); 4422 } 4423 break; 4424 } 4425 4426 case Primitive::kPrimDouble: { 4427 uint32_t data_offset = mirror::Array::DataOffset(sizeof(double)).Uint32Value(); 4428 SRegister out = out_loc.AsFpuRegisterPairLow<SRegister>(); 4429 if (index.IsConstant()) { 4430 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 4431 __ LoadDFromOffset(FromLowSToD(out), obj, offset); 4432 } else { 4433 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 4434 __ LoadDFromOffset(FromLowSToD(out), IP, data_offset); 4435 } 4436 break; 4437 } 4438 4439 case Primitive::kPrimVoid: 4440 LOG(FATAL) << "Unreachable type " << type; 4441 UNREACHABLE(); 4442 } 4443 4444 if (type == Primitive::kPrimNot) { 4445 // Potential implicit null checks, in the case of reference 4446 // arrays, are handled in the previous switch statement. 4447 } else { 4448 codegen_->MaybeRecordImplicitNullCheck(instruction); 4449 } 4450} 4451 4452void LocationsBuilderARM::VisitArraySet(HArraySet* instruction) { 4453 Primitive::Type value_type = instruction->GetComponentType(); 4454 4455 bool needs_write_barrier = 4456 CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue()); 4457 bool may_need_runtime_call_for_type_check = instruction->NeedsTypeCheck(); 4458 bool object_array_set_with_read_barrier = 4459 kEmitCompilerReadBarrier && (value_type == Primitive::kPrimNot); 4460 4461 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary( 4462 instruction, 4463 (may_need_runtime_call_for_type_check || object_array_set_with_read_barrier) ? 4464 LocationSummary::kCallOnSlowPath : 4465 LocationSummary::kNoCall); 4466 4467 locations->SetInAt(0, Location::RequiresRegister()); 4468 locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1))); 4469 if (Primitive::IsFloatingPointType(value_type)) { 4470 locations->SetInAt(2, Location::RequiresFpuRegister()); 4471 } else { 4472 locations->SetInAt(2, Location::RequiresRegister()); 4473 } 4474 if (needs_write_barrier) { 4475 // Temporary registers for the write barrier. 4476 locations->AddTemp(Location::RequiresRegister()); // Possibly used for ref. poisoning too. 4477 locations->AddTemp(Location::RequiresRegister()); 4478 } 4479} 4480 4481void InstructionCodeGeneratorARM::VisitArraySet(HArraySet* instruction) { 4482 LocationSummary* locations = instruction->GetLocations(); 4483 Location array_loc = locations->InAt(0); 4484 Register array = array_loc.AsRegister<Register>(); 4485 Location index = locations->InAt(1); 4486 Primitive::Type value_type = instruction->GetComponentType(); 4487 bool may_need_runtime_call_for_type_check = instruction->NeedsTypeCheck(); 4488 bool needs_write_barrier = 4489 CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue()); 4490 4491 switch (value_type) { 4492 case Primitive::kPrimBoolean: 4493 case Primitive::kPrimByte: { 4494 uint32_t data_offset = mirror::Array::DataOffset(sizeof(uint8_t)).Uint32Value(); 4495 Register value = locations->InAt(2).AsRegister<Register>(); 4496 if (index.IsConstant()) { 4497 size_t offset = 4498 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_1) + data_offset; 4499 __ StoreToOffset(kStoreByte, value, array, offset); 4500 } else { 4501 __ add(IP, array, ShifterOperand(index.AsRegister<Register>())); 4502 __ StoreToOffset(kStoreByte, value, IP, data_offset); 4503 } 4504 break; 4505 } 4506 4507 case Primitive::kPrimShort: 4508 case Primitive::kPrimChar: { 4509 uint32_t data_offset = mirror::Array::DataOffset(sizeof(uint16_t)).Uint32Value(); 4510 Register value = locations->InAt(2).AsRegister<Register>(); 4511 if (index.IsConstant()) { 4512 size_t offset = 4513 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_2) + data_offset; 4514 __ StoreToOffset(kStoreHalfword, value, array, offset); 4515 } else { 4516 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_2)); 4517 __ StoreToOffset(kStoreHalfword, value, IP, data_offset); 4518 } 4519 break; 4520 } 4521 4522 case Primitive::kPrimNot: { 4523 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int32_t)).Uint32Value(); 4524 Location value_loc = locations->InAt(2); 4525 Register value = value_loc.AsRegister<Register>(); 4526 Register source = value; 4527 4528 if (instruction->InputAt(2)->IsNullConstant()) { 4529 // Just setting null. 4530 if (index.IsConstant()) { 4531 size_t offset = 4532 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4533 __ StoreToOffset(kStoreWord, source, array, offset); 4534 } else { 4535 DCHECK(index.IsRegister()) << index; 4536 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 4537 __ StoreToOffset(kStoreWord, source, IP, data_offset); 4538 } 4539 codegen_->MaybeRecordImplicitNullCheck(instruction); 4540 DCHECK(!needs_write_barrier); 4541 DCHECK(!may_need_runtime_call_for_type_check); 4542 break; 4543 } 4544 4545 DCHECK(needs_write_barrier); 4546 Register temp1 = locations->GetTemp(0).AsRegister<Register>(); 4547 Register temp2 = locations->GetTemp(1).AsRegister<Register>(); 4548 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 4549 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 4550 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 4551 Label done; 4552 SlowPathCode* slow_path = nullptr; 4553 4554 if (may_need_runtime_call_for_type_check) { 4555 slow_path = new (GetGraph()->GetArena()) ArraySetSlowPathARM(instruction); 4556 codegen_->AddSlowPath(slow_path); 4557 if (instruction->GetValueCanBeNull()) { 4558 Label non_zero; 4559 __ CompareAndBranchIfNonZero(value, &non_zero); 4560 if (index.IsConstant()) { 4561 size_t offset = 4562 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4563 __ StoreToOffset(kStoreWord, value, array, offset); 4564 } else { 4565 DCHECK(index.IsRegister()) << index; 4566 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 4567 __ StoreToOffset(kStoreWord, value, IP, data_offset); 4568 } 4569 codegen_->MaybeRecordImplicitNullCheck(instruction); 4570 __ b(&done); 4571 __ Bind(&non_zero); 4572 } 4573 4574 if (kEmitCompilerReadBarrier) { 4575 // When read barriers are enabled, the type checking 4576 // instrumentation requires two read barriers: 4577 // 4578 // __ Mov(temp2, temp1); 4579 // // /* HeapReference<Class> */ temp1 = temp1->component_type_ 4580 // __ LoadFromOffset(kLoadWord, temp1, temp1, component_offset); 4581 // codegen_->GenerateReadBarrierSlow( 4582 // instruction, temp1_loc, temp1_loc, temp2_loc, component_offset); 4583 // 4584 // // /* HeapReference<Class> */ temp2 = value->klass_ 4585 // __ LoadFromOffset(kLoadWord, temp2, value, class_offset); 4586 // codegen_->GenerateReadBarrierSlow( 4587 // instruction, temp2_loc, temp2_loc, value_loc, class_offset, temp1_loc); 4588 // 4589 // __ cmp(temp1, ShifterOperand(temp2)); 4590 // 4591 // However, the second read barrier may trash `temp`, as it 4592 // is a temporary register, and as such would not be saved 4593 // along with live registers before calling the runtime (nor 4594 // restored afterwards). So in this case, we bail out and 4595 // delegate the work to the array set slow path. 4596 // 4597 // TODO: Extend the register allocator to support a new 4598 // "(locally) live temp" location so as to avoid always 4599 // going into the slow path when read barriers are enabled. 4600 __ b(slow_path->GetEntryLabel()); 4601 } else { 4602 // /* HeapReference<Class> */ temp1 = array->klass_ 4603 __ LoadFromOffset(kLoadWord, temp1, array, class_offset); 4604 codegen_->MaybeRecordImplicitNullCheck(instruction); 4605 __ MaybeUnpoisonHeapReference(temp1); 4606 4607 // /* HeapReference<Class> */ temp1 = temp1->component_type_ 4608 __ LoadFromOffset(kLoadWord, temp1, temp1, component_offset); 4609 // /* HeapReference<Class> */ temp2 = value->klass_ 4610 __ LoadFromOffset(kLoadWord, temp2, value, class_offset); 4611 // If heap poisoning is enabled, no need to unpoison `temp1` 4612 // nor `temp2`, as we are comparing two poisoned references. 4613 __ cmp(temp1, ShifterOperand(temp2)); 4614 4615 if (instruction->StaticTypeOfArrayIsObjectArray()) { 4616 Label do_put; 4617 __ b(&do_put, EQ); 4618 // If heap poisoning is enabled, the `temp1` reference has 4619 // not been unpoisoned yet; unpoison it now. 4620 __ MaybeUnpoisonHeapReference(temp1); 4621 4622 // /* HeapReference<Class> */ temp1 = temp1->super_class_ 4623 __ LoadFromOffset(kLoadWord, temp1, temp1, super_offset); 4624 // If heap poisoning is enabled, no need to unpoison 4625 // `temp1`, as we are comparing against null below. 4626 __ CompareAndBranchIfNonZero(temp1, slow_path->GetEntryLabel()); 4627 __ Bind(&do_put); 4628 } else { 4629 __ b(slow_path->GetEntryLabel(), NE); 4630 } 4631 } 4632 } 4633 4634 if (kPoisonHeapReferences) { 4635 // Note that in the case where `value` is a null reference, 4636 // we do not enter this block, as a null reference does not 4637 // need poisoning. 4638 DCHECK_EQ(value_type, Primitive::kPrimNot); 4639 __ Mov(temp1, value); 4640 __ PoisonHeapReference(temp1); 4641 source = temp1; 4642 } 4643 4644 if (index.IsConstant()) { 4645 size_t offset = 4646 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4647 __ StoreToOffset(kStoreWord, source, array, offset); 4648 } else { 4649 DCHECK(index.IsRegister()) << index; 4650 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 4651 __ StoreToOffset(kStoreWord, source, IP, data_offset); 4652 } 4653 4654 if (!may_need_runtime_call_for_type_check) { 4655 codegen_->MaybeRecordImplicitNullCheck(instruction); 4656 } 4657 4658 codegen_->MarkGCCard(temp1, temp2, array, value, instruction->GetValueCanBeNull()); 4659 4660 if (done.IsLinked()) { 4661 __ Bind(&done); 4662 } 4663 4664 if (slow_path != nullptr) { 4665 __ Bind(slow_path->GetExitLabel()); 4666 } 4667 4668 break; 4669 } 4670 4671 case Primitive::kPrimInt: { 4672 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int32_t)).Uint32Value(); 4673 Register value = locations->InAt(2).AsRegister<Register>(); 4674 if (index.IsConstant()) { 4675 size_t offset = 4676 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4677 __ StoreToOffset(kStoreWord, value, array, offset); 4678 } else { 4679 DCHECK(index.IsRegister()) << index; 4680 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 4681 __ StoreToOffset(kStoreWord, value, IP, data_offset); 4682 } 4683 break; 4684 } 4685 4686 case Primitive::kPrimLong: { 4687 uint32_t data_offset = mirror::Array::DataOffset(sizeof(int64_t)).Uint32Value(); 4688 Location value = locations->InAt(2); 4689 if (index.IsConstant()) { 4690 size_t offset = 4691 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 4692 __ StoreToOffset(kStoreWordPair, value.AsRegisterPairLow<Register>(), array, offset); 4693 } else { 4694 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 4695 __ StoreToOffset(kStoreWordPair, value.AsRegisterPairLow<Register>(), IP, data_offset); 4696 } 4697 break; 4698 } 4699 4700 case Primitive::kPrimFloat: { 4701 uint32_t data_offset = mirror::Array::DataOffset(sizeof(float)).Uint32Value(); 4702 Location value = locations->InAt(2); 4703 DCHECK(value.IsFpuRegister()); 4704 if (index.IsConstant()) { 4705 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4706 __ StoreSToOffset(value.AsFpuRegister<SRegister>(), array, offset); 4707 } else { 4708 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 4709 __ StoreSToOffset(value.AsFpuRegister<SRegister>(), IP, data_offset); 4710 } 4711 break; 4712 } 4713 4714 case Primitive::kPrimDouble: { 4715 uint32_t data_offset = mirror::Array::DataOffset(sizeof(double)).Uint32Value(); 4716 Location value = locations->InAt(2); 4717 DCHECK(value.IsFpuRegisterPair()); 4718 if (index.IsConstant()) { 4719 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 4720 __ StoreDToOffset(FromLowSToD(value.AsFpuRegisterPairLow<SRegister>()), array, offset); 4721 } else { 4722 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 4723 __ StoreDToOffset(FromLowSToD(value.AsFpuRegisterPairLow<SRegister>()), IP, data_offset); 4724 } 4725 4726 break; 4727 } 4728 4729 case Primitive::kPrimVoid: 4730 LOG(FATAL) << "Unreachable type " << value_type; 4731 UNREACHABLE(); 4732 } 4733 4734 // Objects are handled in the switch. 4735 if (value_type != Primitive::kPrimNot) { 4736 codegen_->MaybeRecordImplicitNullCheck(instruction); 4737 } 4738} 4739 4740void LocationsBuilderARM::VisitArrayLength(HArrayLength* instruction) { 4741 LocationSummary* locations = 4742 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 4743 locations->SetInAt(0, Location::RequiresRegister()); 4744 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 4745} 4746 4747void InstructionCodeGeneratorARM::VisitArrayLength(HArrayLength* instruction) { 4748 LocationSummary* locations = instruction->GetLocations(); 4749 uint32_t offset = CodeGenerator::GetArrayLengthOffset(instruction); 4750 Register obj = locations->InAt(0).AsRegister<Register>(); 4751 Register out = locations->Out().AsRegister<Register>(); 4752 __ LoadFromOffset(kLoadWord, out, obj, offset); 4753 codegen_->MaybeRecordImplicitNullCheck(instruction); 4754} 4755 4756void LocationsBuilderARM::VisitBoundsCheck(HBoundsCheck* instruction) { 4757 LocationSummary::CallKind call_kind = instruction->CanThrowIntoCatchBlock() 4758 ? LocationSummary::kCallOnSlowPath 4759 : LocationSummary::kNoCall; 4760 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 4761 locations->SetInAt(0, Location::RequiresRegister()); 4762 locations->SetInAt(1, Location::RequiresRegister()); 4763 if (instruction->HasUses()) { 4764 locations->SetOut(Location::SameAsFirstInput()); 4765 } 4766} 4767 4768void InstructionCodeGeneratorARM::VisitBoundsCheck(HBoundsCheck* instruction) { 4769 LocationSummary* locations = instruction->GetLocations(); 4770 SlowPathCode* slow_path = 4771 new (GetGraph()->GetArena()) BoundsCheckSlowPathARM(instruction); 4772 codegen_->AddSlowPath(slow_path); 4773 4774 Register index = locations->InAt(0).AsRegister<Register>(); 4775 Register length = locations->InAt(1).AsRegister<Register>(); 4776 4777 __ cmp(index, ShifterOperand(length)); 4778 __ b(slow_path->GetEntryLabel(), HS); 4779} 4780 4781void CodeGeneratorARM::MarkGCCard(Register temp, 4782 Register card, 4783 Register object, 4784 Register value, 4785 bool can_be_null) { 4786 Label is_null; 4787 if (can_be_null) { 4788 __ CompareAndBranchIfZero(value, &is_null); 4789 } 4790 __ LoadFromOffset(kLoadWord, card, TR, Thread::CardTableOffset<kArmWordSize>().Int32Value()); 4791 __ Lsr(temp, object, gc::accounting::CardTable::kCardShift); 4792 __ strb(card, Address(card, temp)); 4793 if (can_be_null) { 4794 __ Bind(&is_null); 4795 } 4796} 4797 4798void LocationsBuilderARM::VisitParallelMove(HParallelMove* instruction ATTRIBUTE_UNUSED) { 4799 LOG(FATAL) << "Unreachable"; 4800} 4801 4802void InstructionCodeGeneratorARM::VisitParallelMove(HParallelMove* instruction) { 4803 codegen_->GetMoveResolver()->EmitNativeCode(instruction); 4804} 4805 4806void LocationsBuilderARM::VisitSuspendCheck(HSuspendCheck* instruction) { 4807 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnSlowPath); 4808} 4809 4810void InstructionCodeGeneratorARM::VisitSuspendCheck(HSuspendCheck* instruction) { 4811 HBasicBlock* block = instruction->GetBlock(); 4812 if (block->GetLoopInformation() != nullptr) { 4813 DCHECK(block->GetLoopInformation()->GetSuspendCheck() == instruction); 4814 // The back edge will generate the suspend check. 4815 return; 4816 } 4817 if (block->IsEntryBlock() && instruction->GetNext()->IsGoto()) { 4818 // The goto will generate the suspend check. 4819 return; 4820 } 4821 GenerateSuspendCheck(instruction, nullptr); 4822} 4823 4824void InstructionCodeGeneratorARM::GenerateSuspendCheck(HSuspendCheck* instruction, 4825 HBasicBlock* successor) { 4826 SuspendCheckSlowPathARM* slow_path = 4827 down_cast<SuspendCheckSlowPathARM*>(instruction->GetSlowPath()); 4828 if (slow_path == nullptr) { 4829 slow_path = new (GetGraph()->GetArena()) SuspendCheckSlowPathARM(instruction, successor); 4830 instruction->SetSlowPath(slow_path); 4831 codegen_->AddSlowPath(slow_path); 4832 if (successor != nullptr) { 4833 DCHECK(successor->IsLoopHeader()); 4834 codegen_->ClearSpillSlotsFromLoopPhisInStackMap(instruction); 4835 } 4836 } else { 4837 DCHECK_EQ(slow_path->GetSuccessor(), successor); 4838 } 4839 4840 __ LoadFromOffset( 4841 kLoadUnsignedHalfword, IP, TR, Thread::ThreadFlagsOffset<kArmWordSize>().Int32Value()); 4842 if (successor == nullptr) { 4843 __ CompareAndBranchIfNonZero(IP, slow_path->GetEntryLabel()); 4844 __ Bind(slow_path->GetReturnLabel()); 4845 } else { 4846 __ CompareAndBranchIfZero(IP, codegen_->GetLabelOf(successor)); 4847 __ b(slow_path->GetEntryLabel()); 4848 } 4849} 4850 4851ArmAssembler* ParallelMoveResolverARM::GetAssembler() const { 4852 return codegen_->GetAssembler(); 4853} 4854 4855void ParallelMoveResolverARM::EmitMove(size_t index) { 4856 MoveOperands* move = moves_[index]; 4857 Location source = move->GetSource(); 4858 Location destination = move->GetDestination(); 4859 4860 if (source.IsRegister()) { 4861 if (destination.IsRegister()) { 4862 __ Mov(destination.AsRegister<Register>(), source.AsRegister<Register>()); 4863 } else if (destination.IsFpuRegister()) { 4864 __ vmovsr(destination.AsFpuRegister<SRegister>(), source.AsRegister<Register>()); 4865 } else { 4866 DCHECK(destination.IsStackSlot()); 4867 __ StoreToOffset(kStoreWord, source.AsRegister<Register>(), 4868 SP, destination.GetStackIndex()); 4869 } 4870 } else if (source.IsStackSlot()) { 4871 if (destination.IsRegister()) { 4872 __ LoadFromOffset(kLoadWord, destination.AsRegister<Register>(), 4873 SP, source.GetStackIndex()); 4874 } else if (destination.IsFpuRegister()) { 4875 __ LoadSFromOffset(destination.AsFpuRegister<SRegister>(), SP, source.GetStackIndex()); 4876 } else { 4877 DCHECK(destination.IsStackSlot()); 4878 __ LoadFromOffset(kLoadWord, IP, SP, source.GetStackIndex()); 4879 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 4880 } 4881 } else if (source.IsFpuRegister()) { 4882 if (destination.IsRegister()) { 4883 __ vmovrs(destination.AsRegister<Register>(), source.AsFpuRegister<SRegister>()); 4884 } else if (destination.IsFpuRegister()) { 4885 __ vmovs(destination.AsFpuRegister<SRegister>(), source.AsFpuRegister<SRegister>()); 4886 } else { 4887 DCHECK(destination.IsStackSlot()); 4888 __ StoreSToOffset(source.AsFpuRegister<SRegister>(), SP, destination.GetStackIndex()); 4889 } 4890 } else if (source.IsDoubleStackSlot()) { 4891 if (destination.IsDoubleStackSlot()) { 4892 __ LoadDFromOffset(DTMP, SP, source.GetStackIndex()); 4893 __ StoreDToOffset(DTMP, SP, destination.GetStackIndex()); 4894 } else if (destination.IsRegisterPair()) { 4895 DCHECK(ExpectedPairLayout(destination)); 4896 __ LoadFromOffset( 4897 kLoadWordPair, destination.AsRegisterPairLow<Register>(), SP, source.GetStackIndex()); 4898 } else { 4899 DCHECK(destination.IsFpuRegisterPair()) << destination; 4900 __ LoadDFromOffset(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 4901 SP, 4902 source.GetStackIndex()); 4903 } 4904 } else if (source.IsRegisterPair()) { 4905 if (destination.IsRegisterPair()) { 4906 __ Mov(destination.AsRegisterPairLow<Register>(), source.AsRegisterPairLow<Register>()); 4907 __ Mov(destination.AsRegisterPairHigh<Register>(), source.AsRegisterPairHigh<Register>()); 4908 } else if (destination.IsFpuRegisterPair()) { 4909 __ vmovdrr(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 4910 source.AsRegisterPairLow<Register>(), 4911 source.AsRegisterPairHigh<Register>()); 4912 } else { 4913 DCHECK(destination.IsDoubleStackSlot()) << destination; 4914 DCHECK(ExpectedPairLayout(source)); 4915 __ StoreToOffset( 4916 kStoreWordPair, source.AsRegisterPairLow<Register>(), SP, destination.GetStackIndex()); 4917 } 4918 } else if (source.IsFpuRegisterPair()) { 4919 if (destination.IsRegisterPair()) { 4920 __ vmovrrd(destination.AsRegisterPairLow<Register>(), 4921 destination.AsRegisterPairHigh<Register>(), 4922 FromLowSToD(source.AsFpuRegisterPairLow<SRegister>())); 4923 } else if (destination.IsFpuRegisterPair()) { 4924 __ vmovd(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 4925 FromLowSToD(source.AsFpuRegisterPairLow<SRegister>())); 4926 } else { 4927 DCHECK(destination.IsDoubleStackSlot()) << destination; 4928 __ StoreDToOffset(FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()), 4929 SP, 4930 destination.GetStackIndex()); 4931 } 4932 } else { 4933 DCHECK(source.IsConstant()) << source; 4934 HConstant* constant = source.GetConstant(); 4935 if (constant->IsIntConstant() || constant->IsNullConstant()) { 4936 int32_t value = CodeGenerator::GetInt32ValueOf(constant); 4937 if (destination.IsRegister()) { 4938 __ LoadImmediate(destination.AsRegister<Register>(), value); 4939 } else { 4940 DCHECK(destination.IsStackSlot()); 4941 __ LoadImmediate(IP, value); 4942 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 4943 } 4944 } else if (constant->IsLongConstant()) { 4945 int64_t value = constant->AsLongConstant()->GetValue(); 4946 if (destination.IsRegisterPair()) { 4947 __ LoadImmediate(destination.AsRegisterPairLow<Register>(), Low32Bits(value)); 4948 __ LoadImmediate(destination.AsRegisterPairHigh<Register>(), High32Bits(value)); 4949 } else { 4950 DCHECK(destination.IsDoubleStackSlot()) << destination; 4951 __ LoadImmediate(IP, Low32Bits(value)); 4952 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 4953 __ LoadImmediate(IP, High32Bits(value)); 4954 __ StoreToOffset(kStoreWord, IP, SP, destination.GetHighStackIndex(kArmWordSize)); 4955 } 4956 } else if (constant->IsDoubleConstant()) { 4957 double value = constant->AsDoubleConstant()->GetValue(); 4958 if (destination.IsFpuRegisterPair()) { 4959 __ LoadDImmediate(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), value); 4960 } else { 4961 DCHECK(destination.IsDoubleStackSlot()) << destination; 4962 uint64_t int_value = bit_cast<uint64_t, double>(value); 4963 __ LoadImmediate(IP, Low32Bits(int_value)); 4964 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 4965 __ LoadImmediate(IP, High32Bits(int_value)); 4966 __ StoreToOffset(kStoreWord, IP, SP, destination.GetHighStackIndex(kArmWordSize)); 4967 } 4968 } else { 4969 DCHECK(constant->IsFloatConstant()) << constant->DebugName(); 4970 float value = constant->AsFloatConstant()->GetValue(); 4971 if (destination.IsFpuRegister()) { 4972 __ LoadSImmediate(destination.AsFpuRegister<SRegister>(), value); 4973 } else { 4974 DCHECK(destination.IsStackSlot()); 4975 __ LoadImmediate(IP, bit_cast<int32_t, float>(value)); 4976 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 4977 } 4978 } 4979 } 4980} 4981 4982void ParallelMoveResolverARM::Exchange(Register reg, int mem) { 4983 __ Mov(IP, reg); 4984 __ LoadFromOffset(kLoadWord, reg, SP, mem); 4985 __ StoreToOffset(kStoreWord, IP, SP, mem); 4986} 4987 4988void ParallelMoveResolverARM::Exchange(int mem1, int mem2) { 4989 ScratchRegisterScope ensure_scratch(this, IP, R0, codegen_->GetNumberOfCoreRegisters()); 4990 int stack_offset = ensure_scratch.IsSpilled() ? kArmWordSize : 0; 4991 __ LoadFromOffset(kLoadWord, static_cast<Register>(ensure_scratch.GetRegister()), 4992 SP, mem1 + stack_offset); 4993 __ LoadFromOffset(kLoadWord, IP, SP, mem2 + stack_offset); 4994 __ StoreToOffset(kStoreWord, static_cast<Register>(ensure_scratch.GetRegister()), 4995 SP, mem2 + stack_offset); 4996 __ StoreToOffset(kStoreWord, IP, SP, mem1 + stack_offset); 4997} 4998 4999void ParallelMoveResolverARM::EmitSwap(size_t index) { 5000 MoveOperands* move = moves_[index]; 5001 Location source = move->GetSource(); 5002 Location destination = move->GetDestination(); 5003 5004 if (source.IsRegister() && destination.IsRegister()) { 5005 DCHECK_NE(source.AsRegister<Register>(), IP); 5006 DCHECK_NE(destination.AsRegister<Register>(), IP); 5007 __ Mov(IP, source.AsRegister<Register>()); 5008 __ Mov(source.AsRegister<Register>(), destination.AsRegister<Register>()); 5009 __ Mov(destination.AsRegister<Register>(), IP); 5010 } else if (source.IsRegister() && destination.IsStackSlot()) { 5011 Exchange(source.AsRegister<Register>(), destination.GetStackIndex()); 5012 } else if (source.IsStackSlot() && destination.IsRegister()) { 5013 Exchange(destination.AsRegister<Register>(), source.GetStackIndex()); 5014 } else if (source.IsStackSlot() && destination.IsStackSlot()) { 5015 Exchange(source.GetStackIndex(), destination.GetStackIndex()); 5016 } else if (source.IsFpuRegister() && destination.IsFpuRegister()) { 5017 __ vmovrs(IP, source.AsFpuRegister<SRegister>()); 5018 __ vmovs(source.AsFpuRegister<SRegister>(), destination.AsFpuRegister<SRegister>()); 5019 __ vmovsr(destination.AsFpuRegister<SRegister>(), IP); 5020 } else if (source.IsRegisterPair() && destination.IsRegisterPair()) { 5021 __ vmovdrr(DTMP, source.AsRegisterPairLow<Register>(), source.AsRegisterPairHigh<Register>()); 5022 __ Mov(source.AsRegisterPairLow<Register>(), destination.AsRegisterPairLow<Register>()); 5023 __ Mov(source.AsRegisterPairHigh<Register>(), destination.AsRegisterPairHigh<Register>()); 5024 __ vmovrrd(destination.AsRegisterPairLow<Register>(), 5025 destination.AsRegisterPairHigh<Register>(), 5026 DTMP); 5027 } else if (source.IsRegisterPair() || destination.IsRegisterPair()) { 5028 Register low_reg = source.IsRegisterPair() 5029 ? source.AsRegisterPairLow<Register>() 5030 : destination.AsRegisterPairLow<Register>(); 5031 int mem = source.IsRegisterPair() 5032 ? destination.GetStackIndex() 5033 : source.GetStackIndex(); 5034 DCHECK(ExpectedPairLayout(source.IsRegisterPair() ? source : destination)); 5035 __ vmovdrr(DTMP, low_reg, static_cast<Register>(low_reg + 1)); 5036 __ LoadFromOffset(kLoadWordPair, low_reg, SP, mem); 5037 __ StoreDToOffset(DTMP, SP, mem); 5038 } else if (source.IsFpuRegisterPair() && destination.IsFpuRegisterPair()) { 5039 DRegister first = FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()); 5040 DRegister second = FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()); 5041 __ vmovd(DTMP, first); 5042 __ vmovd(first, second); 5043 __ vmovd(second, DTMP); 5044 } else if (source.IsFpuRegisterPair() || destination.IsFpuRegisterPair()) { 5045 DRegister reg = source.IsFpuRegisterPair() 5046 ? FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()) 5047 : FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()); 5048 int mem = source.IsFpuRegisterPair() 5049 ? destination.GetStackIndex() 5050 : source.GetStackIndex(); 5051 __ vmovd(DTMP, reg); 5052 __ LoadDFromOffset(reg, SP, mem); 5053 __ StoreDToOffset(DTMP, SP, mem); 5054 } else if (source.IsFpuRegister() || destination.IsFpuRegister()) { 5055 SRegister reg = source.IsFpuRegister() ? source.AsFpuRegister<SRegister>() 5056 : destination.AsFpuRegister<SRegister>(); 5057 int mem = source.IsFpuRegister() 5058 ? destination.GetStackIndex() 5059 : source.GetStackIndex(); 5060 5061 __ vmovrs(IP, reg); 5062 __ LoadSFromOffset(reg, SP, mem); 5063 __ StoreToOffset(kStoreWord, IP, SP, mem); 5064 } else if (source.IsDoubleStackSlot() && destination.IsDoubleStackSlot()) { 5065 Exchange(source.GetStackIndex(), destination.GetStackIndex()); 5066 Exchange(source.GetHighStackIndex(kArmWordSize), destination.GetHighStackIndex(kArmWordSize)); 5067 } else { 5068 LOG(FATAL) << "Unimplemented" << source << " <-> " << destination; 5069 } 5070} 5071 5072void ParallelMoveResolverARM::SpillScratch(int reg) { 5073 __ Push(static_cast<Register>(reg)); 5074} 5075 5076void ParallelMoveResolverARM::RestoreScratch(int reg) { 5077 __ Pop(static_cast<Register>(reg)); 5078} 5079 5080void LocationsBuilderARM::VisitLoadClass(HLoadClass* cls) { 5081 InvokeRuntimeCallingConvention calling_convention; 5082 CodeGenerator::CreateLoadClassLocationSummary( 5083 cls, 5084 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 5085 Location::RegisterLocation(R0), 5086 /* code_generator_supports_read_barrier */ true); 5087} 5088 5089void InstructionCodeGeneratorARM::VisitLoadClass(HLoadClass* cls) { 5090 LocationSummary* locations = cls->GetLocations(); 5091 if (cls->NeedsAccessCheck()) { 5092 codegen_->MoveConstant(locations->GetTemp(0), cls->GetTypeIndex()); 5093 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pInitializeTypeAndVerifyAccess), 5094 cls, 5095 cls->GetDexPc(), 5096 nullptr); 5097 CheckEntrypointTypes<kQuickInitializeTypeAndVerifyAccess, void*, uint32_t>(); 5098 return; 5099 } 5100 5101 Location out_loc = locations->Out(); 5102 Register out = out_loc.AsRegister<Register>(); 5103 Register current_method = locations->InAt(0).AsRegister<Register>(); 5104 5105 if (cls->IsReferrersClass()) { 5106 DCHECK(!cls->CanCallRuntime()); 5107 DCHECK(!cls->MustGenerateClinitCheck()); 5108 // /* GcRoot<mirror::Class> */ out = current_method->declaring_class_ 5109 GenerateGcRootFieldLoad( 5110 cls, out_loc, current_method, ArtMethod::DeclaringClassOffset().Int32Value()); 5111 } else { 5112 // /* GcRoot<mirror::Class>[] */ out = 5113 // current_method.ptr_sized_fields_->dex_cache_resolved_types_ 5114 __ LoadFromOffset(kLoadWord, 5115 out, 5116 current_method, 5117 ArtMethod::DexCacheResolvedTypesOffset(kArmPointerSize).Int32Value()); 5118 // /* GcRoot<mirror::Class> */ out = out[type_index] 5119 GenerateGcRootFieldLoad(cls, out_loc, out, CodeGenerator::GetCacheOffset(cls->GetTypeIndex())); 5120 5121 if (!cls->IsInDexCache() || cls->MustGenerateClinitCheck()) { 5122 DCHECK(cls->CanCallRuntime()); 5123 SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathARM( 5124 cls, cls, cls->GetDexPc(), cls->MustGenerateClinitCheck()); 5125 codegen_->AddSlowPath(slow_path); 5126 if (!cls->IsInDexCache()) { 5127 __ CompareAndBranchIfZero(out, slow_path->GetEntryLabel()); 5128 } 5129 if (cls->MustGenerateClinitCheck()) { 5130 GenerateClassInitializationCheck(slow_path, out); 5131 } else { 5132 __ Bind(slow_path->GetExitLabel()); 5133 } 5134 } 5135 } 5136} 5137 5138void LocationsBuilderARM::VisitClinitCheck(HClinitCheck* check) { 5139 LocationSummary* locations = 5140 new (GetGraph()->GetArena()) LocationSummary(check, LocationSummary::kCallOnSlowPath); 5141 locations->SetInAt(0, Location::RequiresRegister()); 5142 if (check->HasUses()) { 5143 locations->SetOut(Location::SameAsFirstInput()); 5144 } 5145} 5146 5147void InstructionCodeGeneratorARM::VisitClinitCheck(HClinitCheck* check) { 5148 // We assume the class is not null. 5149 SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathARM( 5150 check->GetLoadClass(), check, check->GetDexPc(), true); 5151 codegen_->AddSlowPath(slow_path); 5152 GenerateClassInitializationCheck(slow_path, 5153 check->GetLocations()->InAt(0).AsRegister<Register>()); 5154} 5155 5156void InstructionCodeGeneratorARM::GenerateClassInitializationCheck( 5157 SlowPathCode* slow_path, Register class_reg) { 5158 __ LoadFromOffset(kLoadWord, IP, class_reg, mirror::Class::StatusOffset().Int32Value()); 5159 __ cmp(IP, ShifterOperand(mirror::Class::kStatusInitialized)); 5160 __ b(slow_path->GetEntryLabel(), LT); 5161 // Even if the initialized flag is set, we may be in a situation where caches are not synced 5162 // properly. Therefore, we do a memory fence. 5163 __ dmb(ISH); 5164 __ Bind(slow_path->GetExitLabel()); 5165} 5166 5167HLoadString::LoadKind CodeGeneratorARM::GetSupportedLoadStringKind( 5168 HLoadString::LoadKind desired_string_load_kind) { 5169 if (kEmitCompilerReadBarrier) { 5170 switch (desired_string_load_kind) { 5171 case HLoadString::LoadKind::kBootImageLinkTimeAddress: 5172 case HLoadString::LoadKind::kBootImageLinkTimePcRelative: 5173 case HLoadString::LoadKind::kBootImageAddress: 5174 // TODO: Implement for read barrier. 5175 return HLoadString::LoadKind::kDexCacheViaMethod; 5176 default: 5177 break; 5178 } 5179 } 5180 switch (desired_string_load_kind) { 5181 case HLoadString::LoadKind::kBootImageLinkTimeAddress: 5182 DCHECK(!GetCompilerOptions().GetCompilePic()); 5183 break; 5184 case HLoadString::LoadKind::kBootImageLinkTimePcRelative: 5185 DCHECK(GetCompilerOptions().GetCompilePic()); 5186 break; 5187 case HLoadString::LoadKind::kBootImageAddress: 5188 break; 5189 case HLoadString::LoadKind::kDexCacheAddress: 5190 DCHECK(Runtime::Current()->UseJitCompilation()); 5191 break; 5192 case HLoadString::LoadKind::kDexCachePcRelative: 5193 DCHECK(!Runtime::Current()->UseJitCompilation()); 5194 // We disable pc-relative load when there is an irreducible loop, as the optimization 5195 // is incompatible with it. 5196 // TODO: Create as many ArmDexCacheArraysBase instructions as needed for methods 5197 // with irreducible loops. 5198 if (GetGraph()->HasIrreducibleLoops()) { 5199 return HLoadString::LoadKind::kDexCacheViaMethod; 5200 } 5201 break; 5202 case HLoadString::LoadKind::kDexCacheViaMethod: 5203 break; 5204 } 5205 return desired_string_load_kind; 5206} 5207 5208void LocationsBuilderARM::VisitLoadString(HLoadString* load) { 5209 LocationSummary::CallKind call_kind = (load->NeedsEnvironment() || kEmitCompilerReadBarrier) 5210 ? LocationSummary::kCallOnSlowPath 5211 : LocationSummary::kNoCall; 5212 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(load, call_kind); 5213 HLoadString::LoadKind load_kind = load->GetLoadKind(); 5214 if (load_kind == HLoadString::LoadKind::kDexCacheViaMethod || 5215 load_kind == HLoadString::LoadKind::kDexCachePcRelative) { 5216 locations->SetInAt(0, Location::RequiresRegister()); 5217 } 5218 locations->SetOut(Location::RequiresRegister()); 5219} 5220 5221void InstructionCodeGeneratorARM::VisitLoadString(HLoadString* load) { 5222 LocationSummary* locations = load->GetLocations(); 5223 Location out_loc = locations->Out(); 5224 Register out = out_loc.AsRegister<Register>(); 5225 5226 switch (load->GetLoadKind()) { 5227 case HLoadString::LoadKind::kBootImageLinkTimeAddress: { 5228 DCHECK(!kEmitCompilerReadBarrier); 5229 __ LoadLiteral(out, codegen_->DeduplicateBootImageStringLiteral(load->GetDexFile(), 5230 load->GetStringIndex())); 5231 return; // No dex cache slow path. 5232 } 5233 case HLoadString::LoadKind::kBootImageLinkTimePcRelative: { 5234 DCHECK(!kEmitCompilerReadBarrier); 5235 CodeGeneratorARM::PcRelativePatchInfo* labels = 5236 codegen_->NewPcRelativeStringPatch(load->GetDexFile(), load->GetStringIndex()); 5237 __ BindTrackedLabel(&labels->movw_label); 5238 __ movw(out, /* placeholder */ 0u); 5239 __ BindTrackedLabel(&labels->movt_label); 5240 __ movt(out, /* placeholder */ 0u); 5241 __ BindTrackedLabel(&labels->add_pc_label); 5242 __ add(out, out, ShifterOperand(PC)); 5243 return; // No dex cache slow path. 5244 } 5245 case HLoadString::LoadKind::kBootImageAddress: { 5246 DCHECK(!kEmitCompilerReadBarrier); 5247 DCHECK_NE(load->GetAddress(), 0u); 5248 uint32_t address = dchecked_integral_cast<uint32_t>(load->GetAddress()); 5249 __ LoadLiteral(out, codegen_->DeduplicateBootImageAddressLiteral(address)); 5250 return; // No dex cache slow path. 5251 } 5252 case HLoadString::LoadKind::kDexCacheAddress: { 5253 DCHECK_NE(load->GetAddress(), 0u); 5254 uint32_t address = dchecked_integral_cast<uint32_t>(load->GetAddress()); 5255 // 16-bit LDR immediate has a 5-bit offset multiplied by the size and that gives 5256 // a 128B range. To try and reduce the number of literals if we load multiple strings, 5257 // simply split the dex cache address to a 128B aligned base loaded from a literal 5258 // and the remaining offset embedded in the load. 5259 static_assert(sizeof(GcRoot<mirror::String>) == 4u, "Expected GC root to be 4 bytes."); 5260 DCHECK_ALIGNED(load->GetAddress(), 4u); 5261 constexpr size_t offset_bits = /* encoded bits */ 5 + /* scale */ 2; 5262 uint32_t base_address = address & ~MaxInt<uint32_t>(offset_bits); 5263 uint32_t offset = address & MaxInt<uint32_t>(offset_bits); 5264 __ LoadLiteral(out, codegen_->DeduplicateDexCacheAddressLiteral(base_address)); 5265 GenerateGcRootFieldLoad(load, out_loc, out, offset); 5266 break; 5267 } 5268 case HLoadString::LoadKind::kDexCachePcRelative: { 5269 Register base_reg = locations->InAt(0).AsRegister<Register>(); 5270 HArmDexCacheArraysBase* base = load->InputAt(0)->AsArmDexCacheArraysBase(); 5271 int32_t offset = load->GetDexCacheElementOffset() - base->GetElementOffset(); 5272 GenerateGcRootFieldLoad(load, out_loc, base_reg, offset); 5273 break; 5274 } 5275 case HLoadString::LoadKind::kDexCacheViaMethod: { 5276 Register current_method = locations->InAt(0).AsRegister<Register>(); 5277 5278 // /* GcRoot<mirror::Class> */ out = current_method->declaring_class_ 5279 GenerateGcRootFieldLoad( 5280 load, out_loc, current_method, ArtMethod::DeclaringClassOffset().Int32Value()); 5281 // /* GcRoot<mirror::String>[] */ out = out->dex_cache_strings_ 5282 __ LoadFromOffset(kLoadWord, out, out, mirror::Class::DexCacheStringsOffset().Int32Value()); 5283 // /* GcRoot<mirror::String> */ out = out[string_index] 5284 GenerateGcRootFieldLoad( 5285 load, out_loc, out, CodeGenerator::GetCacheOffset(load->GetStringIndex())); 5286 break; 5287 } 5288 default: 5289 LOG(FATAL) << "Unexpected load kind: " << load->GetLoadKind(); 5290 UNREACHABLE(); 5291 } 5292 5293 if (!load->IsInDexCache()) { 5294 SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadStringSlowPathARM(load); 5295 codegen_->AddSlowPath(slow_path); 5296 __ CompareAndBranchIfZero(out, slow_path->GetEntryLabel()); 5297 __ Bind(slow_path->GetExitLabel()); 5298 } 5299} 5300 5301static int32_t GetExceptionTlsOffset() { 5302 return Thread::ExceptionOffset<kArmWordSize>().Int32Value(); 5303} 5304 5305void LocationsBuilderARM::VisitLoadException(HLoadException* load) { 5306 LocationSummary* locations = 5307 new (GetGraph()->GetArena()) LocationSummary(load, LocationSummary::kNoCall); 5308 locations->SetOut(Location::RequiresRegister()); 5309} 5310 5311void InstructionCodeGeneratorARM::VisitLoadException(HLoadException* load) { 5312 Register out = load->GetLocations()->Out().AsRegister<Register>(); 5313 __ LoadFromOffset(kLoadWord, out, TR, GetExceptionTlsOffset()); 5314} 5315 5316void LocationsBuilderARM::VisitClearException(HClearException* clear) { 5317 new (GetGraph()->GetArena()) LocationSummary(clear, LocationSummary::kNoCall); 5318} 5319 5320void InstructionCodeGeneratorARM::VisitClearException(HClearException* clear ATTRIBUTE_UNUSED) { 5321 __ LoadImmediate(IP, 0); 5322 __ StoreToOffset(kStoreWord, IP, TR, GetExceptionTlsOffset()); 5323} 5324 5325void LocationsBuilderARM::VisitThrow(HThrow* instruction) { 5326 LocationSummary* locations = 5327 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall); 5328 InvokeRuntimeCallingConvention calling_convention; 5329 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 5330} 5331 5332void InstructionCodeGeneratorARM::VisitThrow(HThrow* instruction) { 5333 codegen_->InvokeRuntime( 5334 QUICK_ENTRY_POINT(pDeliverException), instruction, instruction->GetDexPc(), nullptr); 5335 CheckEntrypointTypes<kQuickDeliverException, void, mirror::Object*>(); 5336} 5337 5338static bool TypeCheckNeedsATemporary(TypeCheckKind type_check_kind) { 5339 return kEmitCompilerReadBarrier && 5340 (kUseBakerReadBarrier || 5341 type_check_kind == TypeCheckKind::kAbstractClassCheck || 5342 type_check_kind == TypeCheckKind::kClassHierarchyCheck || 5343 type_check_kind == TypeCheckKind::kArrayObjectCheck); 5344} 5345 5346void LocationsBuilderARM::VisitInstanceOf(HInstanceOf* instruction) { 5347 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 5348 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 5349 switch (type_check_kind) { 5350 case TypeCheckKind::kExactCheck: 5351 case TypeCheckKind::kAbstractClassCheck: 5352 case TypeCheckKind::kClassHierarchyCheck: 5353 case TypeCheckKind::kArrayObjectCheck: 5354 call_kind = 5355 kEmitCompilerReadBarrier ? LocationSummary::kCallOnSlowPath : LocationSummary::kNoCall; 5356 break; 5357 case TypeCheckKind::kArrayCheck: 5358 case TypeCheckKind::kUnresolvedCheck: 5359 case TypeCheckKind::kInterfaceCheck: 5360 call_kind = LocationSummary::kCallOnSlowPath; 5361 break; 5362 } 5363 5364 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 5365 locations->SetInAt(0, Location::RequiresRegister()); 5366 locations->SetInAt(1, Location::RequiresRegister()); 5367 // The "out" register is used as a temporary, so it overlaps with the inputs. 5368 // Note that TypeCheckSlowPathARM uses this register too. 5369 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 5370 // When read barriers are enabled, we need a temporary register for 5371 // some cases. 5372 if (TypeCheckNeedsATemporary(type_check_kind)) { 5373 locations->AddTemp(Location::RequiresRegister()); 5374 } 5375} 5376 5377void InstructionCodeGeneratorARM::VisitInstanceOf(HInstanceOf* instruction) { 5378 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 5379 LocationSummary* locations = instruction->GetLocations(); 5380 Location obj_loc = locations->InAt(0); 5381 Register obj = obj_loc.AsRegister<Register>(); 5382 Register cls = locations->InAt(1).AsRegister<Register>(); 5383 Location out_loc = locations->Out(); 5384 Register out = out_loc.AsRegister<Register>(); 5385 Location maybe_temp_loc = TypeCheckNeedsATemporary(type_check_kind) ? 5386 locations->GetTemp(0) : 5387 Location::NoLocation(); 5388 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 5389 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 5390 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 5391 uint32_t primitive_offset = mirror::Class::PrimitiveTypeOffset().Int32Value(); 5392 Label done, zero; 5393 SlowPathCode* slow_path = nullptr; 5394 5395 // Return 0 if `obj` is null. 5396 // avoid null check if we know obj is not null. 5397 if (instruction->MustDoNullCheck()) { 5398 __ CompareAndBranchIfZero(obj, &zero); 5399 } 5400 5401 // /* HeapReference<Class> */ out = obj->klass_ 5402 GenerateReferenceLoadTwoRegisters(instruction, out_loc, obj_loc, class_offset, maybe_temp_loc); 5403 5404 switch (type_check_kind) { 5405 case TypeCheckKind::kExactCheck: { 5406 __ cmp(out, ShifterOperand(cls)); 5407 // Classes must be equal for the instanceof to succeed. 5408 __ b(&zero, NE); 5409 __ LoadImmediate(out, 1); 5410 __ b(&done); 5411 break; 5412 } 5413 5414 case TypeCheckKind::kAbstractClassCheck: { 5415 // If the class is abstract, we eagerly fetch the super class of the 5416 // object to avoid doing a comparison we know will fail. 5417 Label loop; 5418 __ Bind(&loop); 5419 // /* HeapReference<Class> */ out = out->super_class_ 5420 GenerateReferenceLoadOneRegister(instruction, out_loc, super_offset, maybe_temp_loc); 5421 // If `out` is null, we use it for the result, and jump to `done`. 5422 __ CompareAndBranchIfZero(out, &done); 5423 __ cmp(out, ShifterOperand(cls)); 5424 __ b(&loop, NE); 5425 __ LoadImmediate(out, 1); 5426 if (zero.IsLinked()) { 5427 __ b(&done); 5428 } 5429 break; 5430 } 5431 5432 case TypeCheckKind::kClassHierarchyCheck: { 5433 // Walk over the class hierarchy to find a match. 5434 Label loop, success; 5435 __ Bind(&loop); 5436 __ cmp(out, ShifterOperand(cls)); 5437 __ b(&success, EQ); 5438 // /* HeapReference<Class> */ out = out->super_class_ 5439 GenerateReferenceLoadOneRegister(instruction, out_loc, super_offset, maybe_temp_loc); 5440 __ CompareAndBranchIfNonZero(out, &loop); 5441 // If `out` is null, we use it for the result, and jump to `done`. 5442 __ b(&done); 5443 __ Bind(&success); 5444 __ LoadImmediate(out, 1); 5445 if (zero.IsLinked()) { 5446 __ b(&done); 5447 } 5448 break; 5449 } 5450 5451 case TypeCheckKind::kArrayObjectCheck: { 5452 // Do an exact check. 5453 Label exact_check; 5454 __ cmp(out, ShifterOperand(cls)); 5455 __ b(&exact_check, EQ); 5456 // Otherwise, we need to check that the object's class is a non-primitive array. 5457 // /* HeapReference<Class> */ out = out->component_type_ 5458 GenerateReferenceLoadOneRegister(instruction, out_loc, component_offset, maybe_temp_loc); 5459 // If `out` is null, we use it for the result, and jump to `done`. 5460 __ CompareAndBranchIfZero(out, &done); 5461 __ LoadFromOffset(kLoadUnsignedHalfword, out, out, primitive_offset); 5462 static_assert(Primitive::kPrimNot == 0, "Expected 0 for kPrimNot"); 5463 __ CompareAndBranchIfNonZero(out, &zero); 5464 __ Bind(&exact_check); 5465 __ LoadImmediate(out, 1); 5466 __ b(&done); 5467 break; 5468 } 5469 5470 case TypeCheckKind::kArrayCheck: { 5471 __ cmp(out, ShifterOperand(cls)); 5472 DCHECK(locations->OnlyCallsOnSlowPath()); 5473 slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathARM(instruction, 5474 /* is_fatal */ false); 5475 codegen_->AddSlowPath(slow_path); 5476 __ b(slow_path->GetEntryLabel(), NE); 5477 __ LoadImmediate(out, 1); 5478 if (zero.IsLinked()) { 5479 __ b(&done); 5480 } 5481 break; 5482 } 5483 5484 case TypeCheckKind::kUnresolvedCheck: 5485 case TypeCheckKind::kInterfaceCheck: { 5486 // Note that we indeed only call on slow path, but we always go 5487 // into the slow path for the unresolved and interface check 5488 // cases. 5489 // 5490 // We cannot directly call the InstanceofNonTrivial runtime 5491 // entry point without resorting to a type checking slow path 5492 // here (i.e. by calling InvokeRuntime directly), as it would 5493 // require to assign fixed registers for the inputs of this 5494 // HInstanceOf instruction (following the runtime calling 5495 // convention), which might be cluttered by the potential first 5496 // read barrier emission at the beginning of this method. 5497 // 5498 // TODO: Introduce a new runtime entry point taking the object 5499 // to test (instead of its class) as argument, and let it deal 5500 // with the read barrier issues. This will let us refactor this 5501 // case of the `switch` code as it was previously (with a direct 5502 // call to the runtime not using a type checking slow path). 5503 // This should also be beneficial for the other cases above. 5504 DCHECK(locations->OnlyCallsOnSlowPath()); 5505 slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathARM(instruction, 5506 /* is_fatal */ false); 5507 codegen_->AddSlowPath(slow_path); 5508 __ b(slow_path->GetEntryLabel()); 5509 if (zero.IsLinked()) { 5510 __ b(&done); 5511 } 5512 break; 5513 } 5514 } 5515 5516 if (zero.IsLinked()) { 5517 __ Bind(&zero); 5518 __ LoadImmediate(out, 0); 5519 } 5520 5521 if (done.IsLinked()) { 5522 __ Bind(&done); 5523 } 5524 5525 if (slow_path != nullptr) { 5526 __ Bind(slow_path->GetExitLabel()); 5527 } 5528} 5529 5530void LocationsBuilderARM::VisitCheckCast(HCheckCast* instruction) { 5531 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 5532 bool throws_into_catch = instruction->CanThrowIntoCatchBlock(); 5533 5534 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 5535 switch (type_check_kind) { 5536 case TypeCheckKind::kExactCheck: 5537 case TypeCheckKind::kAbstractClassCheck: 5538 case TypeCheckKind::kClassHierarchyCheck: 5539 case TypeCheckKind::kArrayObjectCheck: 5540 call_kind = (throws_into_catch || kEmitCompilerReadBarrier) ? 5541 LocationSummary::kCallOnSlowPath : 5542 LocationSummary::kNoCall; // In fact, call on a fatal (non-returning) slow path. 5543 break; 5544 case TypeCheckKind::kArrayCheck: 5545 case TypeCheckKind::kUnresolvedCheck: 5546 case TypeCheckKind::kInterfaceCheck: 5547 call_kind = LocationSummary::kCallOnSlowPath; 5548 break; 5549 } 5550 5551 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 5552 locations->SetInAt(0, Location::RequiresRegister()); 5553 locations->SetInAt(1, Location::RequiresRegister()); 5554 // Note that TypeCheckSlowPathARM uses this "temp" register too. 5555 locations->AddTemp(Location::RequiresRegister()); 5556 // When read barriers are enabled, we need an additional temporary 5557 // register for some cases. 5558 if (TypeCheckNeedsATemporary(type_check_kind)) { 5559 locations->AddTemp(Location::RequiresRegister()); 5560 } 5561} 5562 5563void InstructionCodeGeneratorARM::VisitCheckCast(HCheckCast* instruction) { 5564 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 5565 LocationSummary* locations = instruction->GetLocations(); 5566 Location obj_loc = locations->InAt(0); 5567 Register obj = obj_loc.AsRegister<Register>(); 5568 Register cls = locations->InAt(1).AsRegister<Register>(); 5569 Location temp_loc = locations->GetTemp(0); 5570 Register temp = temp_loc.AsRegister<Register>(); 5571 Location maybe_temp2_loc = TypeCheckNeedsATemporary(type_check_kind) ? 5572 locations->GetTemp(1) : 5573 Location::NoLocation(); 5574 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 5575 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 5576 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 5577 uint32_t primitive_offset = mirror::Class::PrimitiveTypeOffset().Int32Value(); 5578 5579 bool is_type_check_slow_path_fatal = 5580 (type_check_kind == TypeCheckKind::kExactCheck || 5581 type_check_kind == TypeCheckKind::kAbstractClassCheck || 5582 type_check_kind == TypeCheckKind::kClassHierarchyCheck || 5583 type_check_kind == TypeCheckKind::kArrayObjectCheck) && 5584 !instruction->CanThrowIntoCatchBlock(); 5585 SlowPathCode* type_check_slow_path = 5586 new (GetGraph()->GetArena()) TypeCheckSlowPathARM(instruction, 5587 is_type_check_slow_path_fatal); 5588 codegen_->AddSlowPath(type_check_slow_path); 5589 5590 Label done; 5591 // Avoid null check if we know obj is not null. 5592 if (instruction->MustDoNullCheck()) { 5593 __ CompareAndBranchIfZero(obj, &done); 5594 } 5595 5596 // /* HeapReference<Class> */ temp = obj->klass_ 5597 GenerateReferenceLoadTwoRegisters(instruction, temp_loc, obj_loc, class_offset, maybe_temp2_loc); 5598 5599 switch (type_check_kind) { 5600 case TypeCheckKind::kExactCheck: 5601 case TypeCheckKind::kArrayCheck: { 5602 __ cmp(temp, ShifterOperand(cls)); 5603 // Jump to slow path for throwing the exception or doing a 5604 // more involved array check. 5605 __ b(type_check_slow_path->GetEntryLabel(), NE); 5606 break; 5607 } 5608 5609 case TypeCheckKind::kAbstractClassCheck: { 5610 // If the class is abstract, we eagerly fetch the super class of the 5611 // object to avoid doing a comparison we know will fail. 5612 Label loop, compare_classes; 5613 __ Bind(&loop); 5614 // /* HeapReference<Class> */ temp = temp->super_class_ 5615 GenerateReferenceLoadOneRegister(instruction, temp_loc, super_offset, maybe_temp2_loc); 5616 5617 // If the class reference currently in `temp` is not null, jump 5618 // to the `compare_classes` label to compare it with the checked 5619 // class. 5620 __ CompareAndBranchIfNonZero(temp, &compare_classes); 5621 // Otherwise, jump to the slow path to throw the exception. 5622 // 5623 // But before, move back the object's class into `temp` before 5624 // going into the slow path, as it has been overwritten in the 5625 // meantime. 5626 // /* HeapReference<Class> */ temp = obj->klass_ 5627 GenerateReferenceLoadTwoRegisters( 5628 instruction, temp_loc, obj_loc, class_offset, maybe_temp2_loc); 5629 __ b(type_check_slow_path->GetEntryLabel()); 5630 5631 __ Bind(&compare_classes); 5632 __ cmp(temp, ShifterOperand(cls)); 5633 __ b(&loop, NE); 5634 break; 5635 } 5636 5637 case TypeCheckKind::kClassHierarchyCheck: { 5638 // Walk over the class hierarchy to find a match. 5639 Label loop; 5640 __ Bind(&loop); 5641 __ cmp(temp, ShifterOperand(cls)); 5642 __ b(&done, EQ); 5643 5644 // /* HeapReference<Class> */ temp = temp->super_class_ 5645 GenerateReferenceLoadOneRegister(instruction, temp_loc, super_offset, maybe_temp2_loc); 5646 5647 // If the class reference currently in `temp` is not null, jump 5648 // back at the beginning of the loop. 5649 __ CompareAndBranchIfNonZero(temp, &loop); 5650 // Otherwise, jump to the slow path to throw the exception. 5651 // 5652 // But before, move back the object's class into `temp` before 5653 // going into the slow path, as it has been overwritten in the 5654 // meantime. 5655 // /* HeapReference<Class> */ temp = obj->klass_ 5656 GenerateReferenceLoadTwoRegisters( 5657 instruction, temp_loc, obj_loc, class_offset, maybe_temp2_loc); 5658 __ b(type_check_slow_path->GetEntryLabel()); 5659 break; 5660 } 5661 5662 case TypeCheckKind::kArrayObjectCheck: { 5663 // Do an exact check. 5664 Label check_non_primitive_component_type; 5665 __ cmp(temp, ShifterOperand(cls)); 5666 __ b(&done, EQ); 5667 5668 // Otherwise, we need to check that the object's class is a non-primitive array. 5669 // /* HeapReference<Class> */ temp = temp->component_type_ 5670 GenerateReferenceLoadOneRegister(instruction, temp_loc, component_offset, maybe_temp2_loc); 5671 5672 // If the component type is not null (i.e. the object is indeed 5673 // an array), jump to label `check_non_primitive_component_type` 5674 // to further check that this component type is not a primitive 5675 // type. 5676 __ CompareAndBranchIfNonZero(temp, &check_non_primitive_component_type); 5677 // Otherwise, jump to the slow path to throw the exception. 5678 // 5679 // But before, move back the object's class into `temp` before 5680 // going into the slow path, as it has been overwritten in the 5681 // meantime. 5682 // /* HeapReference<Class> */ temp = obj->klass_ 5683 GenerateReferenceLoadTwoRegisters( 5684 instruction, temp_loc, obj_loc, class_offset, maybe_temp2_loc); 5685 __ b(type_check_slow_path->GetEntryLabel()); 5686 5687 __ Bind(&check_non_primitive_component_type); 5688 __ LoadFromOffset(kLoadUnsignedHalfword, temp, temp, primitive_offset); 5689 static_assert(Primitive::kPrimNot == 0, "Expected 0 for art::Primitive::kPrimNot"); 5690 __ CompareAndBranchIfZero(temp, &done); 5691 // Same comment as above regarding `temp` and the slow path. 5692 // /* HeapReference<Class> */ temp = obj->klass_ 5693 GenerateReferenceLoadTwoRegisters( 5694 instruction, temp_loc, obj_loc, class_offset, maybe_temp2_loc); 5695 __ b(type_check_slow_path->GetEntryLabel()); 5696 break; 5697 } 5698 5699 case TypeCheckKind::kUnresolvedCheck: 5700 case TypeCheckKind::kInterfaceCheck: 5701 // We always go into the type check slow path for the unresolved 5702 // and interface check cases. 5703 // 5704 // We cannot directly call the CheckCast runtime entry point 5705 // without resorting to a type checking slow path here (i.e. by 5706 // calling InvokeRuntime directly), as it would require to 5707 // assign fixed registers for the inputs of this HInstanceOf 5708 // instruction (following the runtime calling convention), which 5709 // might be cluttered by the potential first read barrier 5710 // emission at the beginning of this method. 5711 // 5712 // TODO: Introduce a new runtime entry point taking the object 5713 // to test (instead of its class) as argument, and let it deal 5714 // with the read barrier issues. This will let us refactor this 5715 // case of the `switch` code as it was previously (with a direct 5716 // call to the runtime not using a type checking slow path). 5717 // This should also be beneficial for the other cases above. 5718 __ b(type_check_slow_path->GetEntryLabel()); 5719 break; 5720 } 5721 __ Bind(&done); 5722 5723 __ Bind(type_check_slow_path->GetExitLabel()); 5724} 5725 5726void LocationsBuilderARM::VisitMonitorOperation(HMonitorOperation* instruction) { 5727 LocationSummary* locations = 5728 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall); 5729 InvokeRuntimeCallingConvention calling_convention; 5730 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 5731} 5732 5733void InstructionCodeGeneratorARM::VisitMonitorOperation(HMonitorOperation* instruction) { 5734 codegen_->InvokeRuntime(instruction->IsEnter() 5735 ? QUICK_ENTRY_POINT(pLockObject) : QUICK_ENTRY_POINT(pUnlockObject), 5736 instruction, 5737 instruction->GetDexPc(), 5738 nullptr); 5739 if (instruction->IsEnter()) { 5740 CheckEntrypointTypes<kQuickLockObject, void, mirror::Object*>(); 5741 } else { 5742 CheckEntrypointTypes<kQuickUnlockObject, void, mirror::Object*>(); 5743 } 5744} 5745 5746void LocationsBuilderARM::VisitAnd(HAnd* instruction) { HandleBitwiseOperation(instruction, AND); } 5747void LocationsBuilderARM::VisitOr(HOr* instruction) { HandleBitwiseOperation(instruction, ORR); } 5748void LocationsBuilderARM::VisitXor(HXor* instruction) { HandleBitwiseOperation(instruction, EOR); } 5749 5750void LocationsBuilderARM::HandleBitwiseOperation(HBinaryOperation* instruction, Opcode opcode) { 5751 LocationSummary* locations = 5752 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 5753 DCHECK(instruction->GetResultType() == Primitive::kPrimInt 5754 || instruction->GetResultType() == Primitive::kPrimLong); 5755 // Note: GVN reorders commutative operations to have the constant on the right hand side. 5756 locations->SetInAt(0, Location::RequiresRegister()); 5757 locations->SetInAt(1, ArmEncodableConstantOrRegister(instruction->InputAt(1), opcode)); 5758 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 5759} 5760 5761void InstructionCodeGeneratorARM::VisitAnd(HAnd* instruction) { 5762 HandleBitwiseOperation(instruction); 5763} 5764 5765void InstructionCodeGeneratorARM::VisitOr(HOr* instruction) { 5766 HandleBitwiseOperation(instruction); 5767} 5768 5769void InstructionCodeGeneratorARM::VisitXor(HXor* instruction) { 5770 HandleBitwiseOperation(instruction); 5771} 5772 5773 5774void LocationsBuilderARM::VisitBitwiseNegatedRight(HBitwiseNegatedRight* instruction) { 5775 LocationSummary* locations = 5776 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 5777 DCHECK(instruction->GetResultType() == Primitive::kPrimInt 5778 || instruction->GetResultType() == Primitive::kPrimLong); 5779 5780 locations->SetInAt(0, Location::RequiresRegister()); 5781 locations->SetInAt(1, Location::RequiresRegister()); 5782 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 5783} 5784 5785void InstructionCodeGeneratorARM::VisitBitwiseNegatedRight(HBitwiseNegatedRight* instruction) { 5786 LocationSummary* locations = instruction->GetLocations(); 5787 Location first = locations->InAt(0); 5788 Location second = locations->InAt(1); 5789 Location out = locations->Out(); 5790 5791 if (instruction->GetResultType() == Primitive::kPrimInt) { 5792 Register first_reg = first.AsRegister<Register>(); 5793 ShifterOperand second_reg(second.AsRegister<Register>()); 5794 Register out_reg = out.AsRegister<Register>(); 5795 5796 switch (instruction->GetOpKind()) { 5797 case HInstruction::kAnd: 5798 __ bic(out_reg, first_reg, second_reg); 5799 break; 5800 case HInstruction::kOr: 5801 __ orn(out_reg, first_reg, second_reg); 5802 break; 5803 // There is no EON on arm. 5804 case HInstruction::kXor: 5805 default: 5806 LOG(FATAL) << "Unexpected instruction " << instruction->DebugName(); 5807 UNREACHABLE(); 5808 } 5809 return; 5810 5811 } else { 5812 DCHECK_EQ(instruction->GetResultType(), Primitive::kPrimLong); 5813 Register first_low = first.AsRegisterPairLow<Register>(); 5814 Register first_high = first.AsRegisterPairHigh<Register>(); 5815 ShifterOperand second_low(second.AsRegisterPairLow<Register>()); 5816 ShifterOperand second_high(second.AsRegisterPairHigh<Register>()); 5817 Register out_low = out.AsRegisterPairLow<Register>(); 5818 Register out_high = out.AsRegisterPairHigh<Register>(); 5819 5820 switch (instruction->GetOpKind()) { 5821 case HInstruction::kAnd: 5822 __ bic(out_low, first_low, second_low); 5823 __ bic(out_high, first_high, second_high); 5824 break; 5825 case HInstruction::kOr: 5826 __ orn(out_low, first_low, second_low); 5827 __ orn(out_high, first_high, second_high); 5828 break; 5829 // There is no EON on arm. 5830 case HInstruction::kXor: 5831 default: 5832 LOG(FATAL) << "Unexpected instruction " << instruction->DebugName(); 5833 UNREACHABLE(); 5834 } 5835 } 5836} 5837 5838void InstructionCodeGeneratorARM::GenerateAndConst(Register out, Register first, uint32_t value) { 5839 // Optimize special cases for individual halfs of `and-long` (`and` is simplified earlier). 5840 if (value == 0xffffffffu) { 5841 if (out != first) { 5842 __ mov(out, ShifterOperand(first)); 5843 } 5844 return; 5845 } 5846 if (value == 0u) { 5847 __ mov(out, ShifterOperand(0)); 5848 return; 5849 } 5850 ShifterOperand so; 5851 if (__ ShifterOperandCanHold(kNoRegister, kNoRegister, AND, value, &so)) { 5852 __ and_(out, first, so); 5853 } else { 5854 DCHECK(__ ShifterOperandCanHold(kNoRegister, kNoRegister, BIC, ~value, &so)); 5855 __ bic(out, first, ShifterOperand(~value)); 5856 } 5857} 5858 5859void InstructionCodeGeneratorARM::GenerateOrrConst(Register out, Register first, uint32_t value) { 5860 // Optimize special cases for individual halfs of `or-long` (`or` is simplified earlier). 5861 if (value == 0u) { 5862 if (out != first) { 5863 __ mov(out, ShifterOperand(first)); 5864 } 5865 return; 5866 } 5867 if (value == 0xffffffffu) { 5868 __ mvn(out, ShifterOperand(0)); 5869 return; 5870 } 5871 ShifterOperand so; 5872 if (__ ShifterOperandCanHold(kNoRegister, kNoRegister, ORR, value, &so)) { 5873 __ orr(out, first, so); 5874 } else { 5875 DCHECK(__ ShifterOperandCanHold(kNoRegister, kNoRegister, ORN, ~value, &so)); 5876 __ orn(out, first, ShifterOperand(~value)); 5877 } 5878} 5879 5880void InstructionCodeGeneratorARM::GenerateEorConst(Register out, Register first, uint32_t value) { 5881 // Optimize special case for individual halfs of `xor-long` (`xor` is simplified earlier). 5882 if (value == 0u) { 5883 if (out != first) { 5884 __ mov(out, ShifterOperand(first)); 5885 } 5886 return; 5887 } 5888 __ eor(out, first, ShifterOperand(value)); 5889} 5890 5891void InstructionCodeGeneratorARM::HandleBitwiseOperation(HBinaryOperation* instruction) { 5892 LocationSummary* locations = instruction->GetLocations(); 5893 Location first = locations->InAt(0); 5894 Location second = locations->InAt(1); 5895 Location out = locations->Out(); 5896 5897 if (second.IsConstant()) { 5898 uint64_t value = static_cast<uint64_t>(Int64FromConstant(second.GetConstant())); 5899 uint32_t value_low = Low32Bits(value); 5900 if (instruction->GetResultType() == Primitive::kPrimInt) { 5901 Register first_reg = first.AsRegister<Register>(); 5902 Register out_reg = out.AsRegister<Register>(); 5903 if (instruction->IsAnd()) { 5904 GenerateAndConst(out_reg, first_reg, value_low); 5905 } else if (instruction->IsOr()) { 5906 GenerateOrrConst(out_reg, first_reg, value_low); 5907 } else { 5908 DCHECK(instruction->IsXor()); 5909 GenerateEorConst(out_reg, first_reg, value_low); 5910 } 5911 } else { 5912 DCHECK_EQ(instruction->GetResultType(), Primitive::kPrimLong); 5913 uint32_t value_high = High32Bits(value); 5914 Register first_low = first.AsRegisterPairLow<Register>(); 5915 Register first_high = first.AsRegisterPairHigh<Register>(); 5916 Register out_low = out.AsRegisterPairLow<Register>(); 5917 Register out_high = out.AsRegisterPairHigh<Register>(); 5918 if (instruction->IsAnd()) { 5919 GenerateAndConst(out_low, first_low, value_low); 5920 GenerateAndConst(out_high, first_high, value_high); 5921 } else if (instruction->IsOr()) { 5922 GenerateOrrConst(out_low, first_low, value_low); 5923 GenerateOrrConst(out_high, first_high, value_high); 5924 } else { 5925 DCHECK(instruction->IsXor()); 5926 GenerateEorConst(out_low, first_low, value_low); 5927 GenerateEorConst(out_high, first_high, value_high); 5928 } 5929 } 5930 return; 5931 } 5932 5933 if (instruction->GetResultType() == Primitive::kPrimInt) { 5934 Register first_reg = first.AsRegister<Register>(); 5935 ShifterOperand second_reg(second.AsRegister<Register>()); 5936 Register out_reg = out.AsRegister<Register>(); 5937 if (instruction->IsAnd()) { 5938 __ and_(out_reg, first_reg, second_reg); 5939 } else if (instruction->IsOr()) { 5940 __ orr(out_reg, first_reg, second_reg); 5941 } else { 5942 DCHECK(instruction->IsXor()); 5943 __ eor(out_reg, first_reg, second_reg); 5944 } 5945 } else { 5946 DCHECK_EQ(instruction->GetResultType(), Primitive::kPrimLong); 5947 Register first_low = first.AsRegisterPairLow<Register>(); 5948 Register first_high = first.AsRegisterPairHigh<Register>(); 5949 ShifterOperand second_low(second.AsRegisterPairLow<Register>()); 5950 ShifterOperand second_high(second.AsRegisterPairHigh<Register>()); 5951 Register out_low = out.AsRegisterPairLow<Register>(); 5952 Register out_high = out.AsRegisterPairHigh<Register>(); 5953 if (instruction->IsAnd()) { 5954 __ and_(out_low, first_low, second_low); 5955 __ and_(out_high, first_high, second_high); 5956 } else if (instruction->IsOr()) { 5957 __ orr(out_low, first_low, second_low); 5958 __ orr(out_high, first_high, second_high); 5959 } else { 5960 DCHECK(instruction->IsXor()); 5961 __ eor(out_low, first_low, second_low); 5962 __ eor(out_high, first_high, second_high); 5963 } 5964 } 5965} 5966 5967void InstructionCodeGeneratorARM::GenerateReferenceLoadOneRegister(HInstruction* instruction, 5968 Location out, 5969 uint32_t offset, 5970 Location maybe_temp) { 5971 Register out_reg = out.AsRegister<Register>(); 5972 if (kEmitCompilerReadBarrier) { 5973 DCHECK(maybe_temp.IsRegister()) << maybe_temp; 5974 if (kUseBakerReadBarrier) { 5975 // Load with fast path based Baker's read barrier. 5976 // /* HeapReference<Object> */ out = *(out + offset) 5977 codegen_->GenerateFieldLoadWithBakerReadBarrier( 5978 instruction, out, out_reg, offset, maybe_temp, /* needs_null_check */ false); 5979 } else { 5980 // Load with slow path based read barrier. 5981 // Save the value of `out` into `maybe_temp` before overwriting it 5982 // in the following move operation, as we will need it for the 5983 // read barrier below. 5984 __ Mov(maybe_temp.AsRegister<Register>(), out_reg); 5985 // /* HeapReference<Object> */ out = *(out + offset) 5986 __ LoadFromOffset(kLoadWord, out_reg, out_reg, offset); 5987 codegen_->GenerateReadBarrierSlow(instruction, out, out, maybe_temp, offset); 5988 } 5989 } else { 5990 // Plain load with no read barrier. 5991 // /* HeapReference<Object> */ out = *(out + offset) 5992 __ LoadFromOffset(kLoadWord, out_reg, out_reg, offset); 5993 __ MaybeUnpoisonHeapReference(out_reg); 5994 } 5995} 5996 5997void InstructionCodeGeneratorARM::GenerateReferenceLoadTwoRegisters(HInstruction* instruction, 5998 Location out, 5999 Location obj, 6000 uint32_t offset, 6001 Location maybe_temp) { 6002 Register out_reg = out.AsRegister<Register>(); 6003 Register obj_reg = obj.AsRegister<Register>(); 6004 if (kEmitCompilerReadBarrier) { 6005 if (kUseBakerReadBarrier) { 6006 DCHECK(maybe_temp.IsRegister()) << maybe_temp; 6007 // Load with fast path based Baker's read barrier. 6008 // /* HeapReference<Object> */ out = *(obj + offset) 6009 codegen_->GenerateFieldLoadWithBakerReadBarrier( 6010 instruction, out, obj_reg, offset, maybe_temp, /* needs_null_check */ false); 6011 } else { 6012 // Load with slow path based read barrier. 6013 // /* HeapReference<Object> */ out = *(obj + offset) 6014 __ LoadFromOffset(kLoadWord, out_reg, obj_reg, offset); 6015 codegen_->GenerateReadBarrierSlow(instruction, out, out, obj, offset); 6016 } 6017 } else { 6018 // Plain load with no read barrier. 6019 // /* HeapReference<Object> */ out = *(obj + offset) 6020 __ LoadFromOffset(kLoadWord, out_reg, obj_reg, offset); 6021 __ MaybeUnpoisonHeapReference(out_reg); 6022 } 6023} 6024 6025void InstructionCodeGeneratorARM::GenerateGcRootFieldLoad(HInstruction* instruction, 6026 Location root, 6027 Register obj, 6028 uint32_t offset) { 6029 Register root_reg = root.AsRegister<Register>(); 6030 if (kEmitCompilerReadBarrier) { 6031 if (kUseBakerReadBarrier) { 6032 // Fast path implementation of art::ReadBarrier::BarrierForRoot when 6033 // Baker's read barrier are used: 6034 // 6035 // root = obj.field; 6036 // if (Thread::Current()->GetIsGcMarking()) { 6037 // root = ReadBarrier::Mark(root) 6038 // } 6039 6040 // /* GcRoot<mirror::Object> */ root = *(obj + offset) 6041 __ LoadFromOffset(kLoadWord, root_reg, obj, offset); 6042 static_assert( 6043 sizeof(mirror::CompressedReference<mirror::Object>) == sizeof(GcRoot<mirror::Object>), 6044 "art::mirror::CompressedReference<mirror::Object> and art::GcRoot<mirror::Object> " 6045 "have different sizes."); 6046 static_assert(sizeof(mirror::CompressedReference<mirror::Object>) == sizeof(int32_t), 6047 "art::mirror::CompressedReference<mirror::Object> and int32_t " 6048 "have different sizes."); 6049 6050 // Slow path used to mark the GC root `root`. 6051 SlowPathCode* slow_path = 6052 new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM(instruction, root, root); 6053 codegen_->AddSlowPath(slow_path); 6054 6055 // IP = Thread::Current()->GetIsGcMarking() 6056 __ LoadFromOffset( 6057 kLoadWord, IP, TR, Thread::IsGcMarkingOffset<kArmWordSize>().Int32Value()); 6058 __ CompareAndBranchIfNonZero(IP, slow_path->GetEntryLabel()); 6059 __ Bind(slow_path->GetExitLabel()); 6060 } else { 6061 // GC root loaded through a slow path for read barriers other 6062 // than Baker's. 6063 // /* GcRoot<mirror::Object>* */ root = obj + offset 6064 __ AddConstant(root_reg, obj, offset); 6065 // /* mirror::Object* */ root = root->Read() 6066 codegen_->GenerateReadBarrierForRootSlow(instruction, root, root); 6067 } 6068 } else { 6069 // Plain GC root load with no read barrier. 6070 // /* GcRoot<mirror::Object> */ root = *(obj + offset) 6071 __ LoadFromOffset(kLoadWord, root_reg, obj, offset); 6072 // Note that GC roots are not affected by heap poisoning, thus we 6073 // do not have to unpoison `root_reg` here. 6074 } 6075} 6076 6077void CodeGeneratorARM::GenerateFieldLoadWithBakerReadBarrier(HInstruction* instruction, 6078 Location ref, 6079 Register obj, 6080 uint32_t offset, 6081 Location temp, 6082 bool needs_null_check) { 6083 DCHECK(kEmitCompilerReadBarrier); 6084 DCHECK(kUseBakerReadBarrier); 6085 6086 // /* HeapReference<Object> */ ref = *(obj + offset) 6087 Location no_index = Location::NoLocation(); 6088 GenerateReferenceLoadWithBakerReadBarrier( 6089 instruction, ref, obj, offset, no_index, temp, needs_null_check); 6090} 6091 6092void CodeGeneratorARM::GenerateArrayLoadWithBakerReadBarrier(HInstruction* instruction, 6093 Location ref, 6094 Register obj, 6095 uint32_t data_offset, 6096 Location index, 6097 Location temp, 6098 bool needs_null_check) { 6099 DCHECK(kEmitCompilerReadBarrier); 6100 DCHECK(kUseBakerReadBarrier); 6101 6102 // /* HeapReference<Object> */ ref = 6103 // *(obj + data_offset + index * sizeof(HeapReference<Object>)) 6104 GenerateReferenceLoadWithBakerReadBarrier( 6105 instruction, ref, obj, data_offset, index, temp, needs_null_check); 6106} 6107 6108void CodeGeneratorARM::GenerateReferenceLoadWithBakerReadBarrier(HInstruction* instruction, 6109 Location ref, 6110 Register obj, 6111 uint32_t offset, 6112 Location index, 6113 Location temp, 6114 bool needs_null_check) { 6115 DCHECK(kEmitCompilerReadBarrier); 6116 DCHECK(kUseBakerReadBarrier); 6117 6118 // In slow path based read barriers, the read barrier call is 6119 // inserted after the original load. However, in fast path based 6120 // Baker's read barriers, we need to perform the load of 6121 // mirror::Object::monitor_ *before* the original reference load. 6122 // This load-load ordering is required by the read barrier. 6123 // The fast path/slow path (for Baker's algorithm) should look like: 6124 // 6125 // uint32_t rb_state = Lockword(obj->monitor_).ReadBarrierState(); 6126 // lfence; // Load fence or artificial data dependency to prevent load-load reordering 6127 // HeapReference<Object> ref = *src; // Original reference load. 6128 // bool is_gray = (rb_state == ReadBarrier::gray_ptr_); 6129 // if (is_gray) { 6130 // ref = ReadBarrier::Mark(ref); // Performed by runtime entrypoint slow path. 6131 // } 6132 // 6133 // Note: the original implementation in ReadBarrier::Barrier is 6134 // slightly more complex as it performs additional checks that we do 6135 // not do here for performance reasons. 6136 6137 Register ref_reg = ref.AsRegister<Register>(); 6138 Register temp_reg = temp.AsRegister<Register>(); 6139 uint32_t monitor_offset = mirror::Object::MonitorOffset().Int32Value(); 6140 6141 // /* int32_t */ monitor = obj->monitor_ 6142 __ LoadFromOffset(kLoadWord, temp_reg, obj, monitor_offset); 6143 if (needs_null_check) { 6144 MaybeRecordImplicitNullCheck(instruction); 6145 } 6146 // /* LockWord */ lock_word = LockWord(monitor) 6147 static_assert(sizeof(LockWord) == sizeof(int32_t), 6148 "art::LockWord and int32_t have different sizes."); 6149 // /* uint32_t */ rb_state = lock_word.ReadBarrierState() 6150 __ Lsr(temp_reg, temp_reg, LockWord::kReadBarrierStateShift); 6151 __ and_(temp_reg, temp_reg, ShifterOperand(LockWord::kReadBarrierStateMask)); 6152 static_assert( 6153 LockWord::kReadBarrierStateMask == ReadBarrier::rb_ptr_mask_, 6154 "art::LockWord::kReadBarrierStateMask is not equal to art::ReadBarrier::rb_ptr_mask_."); 6155 6156 // Introduce a dependency on the high bits of rb_state, which shall 6157 // be all zeroes, to prevent load-load reordering, and without using 6158 // a memory barrier (which would be more expensive). 6159 // IP = rb_state & ~LockWord::kReadBarrierStateMask = 0 6160 __ bic(IP, temp_reg, ShifterOperand(LockWord::kReadBarrierStateMask)); 6161 // obj is unchanged by this operation, but its value now depends on 6162 // IP, which depends on temp_reg. 6163 __ add(obj, obj, ShifterOperand(IP)); 6164 6165 // The actual reference load. 6166 if (index.IsValid()) { 6167 static_assert( 6168 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 6169 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 6170 // /* HeapReference<Object> */ ref = 6171 // *(obj + offset + index * sizeof(HeapReference<Object>)) 6172 if (index.IsConstant()) { 6173 size_t computed_offset = 6174 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + offset; 6175 __ LoadFromOffset(kLoadWord, ref_reg, obj, computed_offset); 6176 } else { 6177 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 6178 __ LoadFromOffset(kLoadWord, ref_reg, IP, offset); 6179 } 6180 } else { 6181 // /* HeapReference<Object> */ ref = *(obj + offset) 6182 __ LoadFromOffset(kLoadWord, ref_reg, obj, offset); 6183 } 6184 6185 // Object* ref = ref_addr->AsMirrorPtr() 6186 __ MaybeUnpoisonHeapReference(ref_reg); 6187 6188 // Slow path used to mark the object `ref` when it is gray. 6189 SlowPathCode* slow_path = 6190 new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM(instruction, ref, ref); 6191 AddSlowPath(slow_path); 6192 6193 // if (rb_state == ReadBarrier::gray_ptr_) 6194 // ref = ReadBarrier::Mark(ref); 6195 __ cmp(temp_reg, ShifterOperand(ReadBarrier::gray_ptr_)); 6196 __ b(slow_path->GetEntryLabel(), EQ); 6197 __ Bind(slow_path->GetExitLabel()); 6198} 6199 6200void CodeGeneratorARM::GenerateReadBarrierSlow(HInstruction* instruction, 6201 Location out, 6202 Location ref, 6203 Location obj, 6204 uint32_t offset, 6205 Location index) { 6206 DCHECK(kEmitCompilerReadBarrier); 6207 6208 // Insert a slow path based read barrier *after* the reference load. 6209 // 6210 // If heap poisoning is enabled, the unpoisoning of the loaded 6211 // reference will be carried out by the runtime within the slow 6212 // path. 6213 // 6214 // Note that `ref` currently does not get unpoisoned (when heap 6215 // poisoning is enabled), which is alright as the `ref` argument is 6216 // not used by the artReadBarrierSlow entry point. 6217 // 6218 // TODO: Unpoison `ref` when it is used by artReadBarrierSlow. 6219 SlowPathCode* slow_path = new (GetGraph()->GetArena()) 6220 ReadBarrierForHeapReferenceSlowPathARM(instruction, out, ref, obj, offset, index); 6221 AddSlowPath(slow_path); 6222 6223 __ b(slow_path->GetEntryLabel()); 6224 __ Bind(slow_path->GetExitLabel()); 6225} 6226 6227void CodeGeneratorARM::MaybeGenerateReadBarrierSlow(HInstruction* instruction, 6228 Location out, 6229 Location ref, 6230 Location obj, 6231 uint32_t offset, 6232 Location index) { 6233 if (kEmitCompilerReadBarrier) { 6234 // Baker's read barriers shall be handled by the fast path 6235 // (CodeGeneratorARM::GenerateReferenceLoadWithBakerReadBarrier). 6236 DCHECK(!kUseBakerReadBarrier); 6237 // If heap poisoning is enabled, unpoisoning will be taken care of 6238 // by the runtime within the slow path. 6239 GenerateReadBarrierSlow(instruction, out, ref, obj, offset, index); 6240 } else if (kPoisonHeapReferences) { 6241 __ UnpoisonHeapReference(out.AsRegister<Register>()); 6242 } 6243} 6244 6245void CodeGeneratorARM::GenerateReadBarrierForRootSlow(HInstruction* instruction, 6246 Location out, 6247 Location root) { 6248 DCHECK(kEmitCompilerReadBarrier); 6249 6250 // Insert a slow path based read barrier *after* the GC root load. 6251 // 6252 // Note that GC roots are not affected by heap poisoning, so we do 6253 // not need to do anything special for this here. 6254 SlowPathCode* slow_path = 6255 new (GetGraph()->GetArena()) ReadBarrierForRootSlowPathARM(instruction, out, root); 6256 AddSlowPath(slow_path); 6257 6258 __ b(slow_path->GetEntryLabel()); 6259 __ Bind(slow_path->GetExitLabel()); 6260} 6261 6262HInvokeStaticOrDirect::DispatchInfo CodeGeneratorARM::GetSupportedInvokeStaticOrDirectDispatch( 6263 const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info, 6264 MethodReference target_method) { 6265 HInvokeStaticOrDirect::DispatchInfo dispatch_info = desired_dispatch_info; 6266 // We disable pc-relative load when there is an irreducible loop, as the optimization 6267 // is incompatible with it. 6268 // TODO: Create as many ArmDexCacheArraysBase instructions as needed for methods 6269 // with irreducible loops. 6270 if (GetGraph()->HasIrreducibleLoops() && 6271 (dispatch_info.method_load_kind == 6272 HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative)) { 6273 dispatch_info.method_load_kind = HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod; 6274 } 6275 6276 if (dispatch_info.code_ptr_location == HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative) { 6277 const DexFile& outer_dex_file = GetGraph()->GetDexFile(); 6278 if (&outer_dex_file != target_method.dex_file) { 6279 // Calls across dex files are more likely to exceed the available BL range, 6280 // so use absolute patch with fixup if available and kCallArtMethod otherwise. 6281 HInvokeStaticOrDirect::CodePtrLocation code_ptr_location = 6282 (desired_dispatch_info.method_load_kind == 6283 HInvokeStaticOrDirect::MethodLoadKind::kDirectAddressWithFixup) 6284 ? HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup 6285 : HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod; 6286 return HInvokeStaticOrDirect::DispatchInfo { 6287 dispatch_info.method_load_kind, 6288 code_ptr_location, 6289 dispatch_info.method_load_data, 6290 0u 6291 }; 6292 } 6293 } 6294 return dispatch_info; 6295} 6296 6297Register CodeGeneratorARM::GetInvokeStaticOrDirectExtraParameter(HInvokeStaticOrDirect* invoke, 6298 Register temp) { 6299 DCHECK_EQ(invoke->InputCount(), invoke->GetNumberOfArguments() + 1u); 6300 Location location = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 6301 if (!invoke->GetLocations()->Intrinsified()) { 6302 return location.AsRegister<Register>(); 6303 } 6304 // For intrinsics we allow any location, so it may be on the stack. 6305 if (!location.IsRegister()) { 6306 __ LoadFromOffset(kLoadWord, temp, SP, location.GetStackIndex()); 6307 return temp; 6308 } 6309 // For register locations, check if the register was saved. If so, get it from the stack. 6310 // Note: There is a chance that the register was saved but not overwritten, so we could 6311 // save one load. However, since this is just an intrinsic slow path we prefer this 6312 // simple and more robust approach rather that trying to determine if that's the case. 6313 SlowPathCode* slow_path = GetCurrentSlowPath(); 6314 DCHECK(slow_path != nullptr); // For intrinsified invokes the call is emitted on the slow path. 6315 if (slow_path->IsCoreRegisterSaved(location.AsRegister<Register>())) { 6316 int stack_offset = slow_path->GetStackOffsetOfCoreRegister(location.AsRegister<Register>()); 6317 __ LoadFromOffset(kLoadWord, temp, SP, stack_offset); 6318 return temp; 6319 } 6320 return location.AsRegister<Register>(); 6321} 6322 6323void CodeGeneratorARM::GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) { 6324 // For better instruction scheduling we load the direct code pointer before the method pointer. 6325 switch (invoke->GetCodePtrLocation()) { 6326 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup: 6327 // LR = code address from literal pool with link-time patch. 6328 __ LoadLiteral(LR, DeduplicateMethodCodeLiteral(invoke->GetTargetMethod())); 6329 break; 6330 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect: 6331 // LR = invoke->GetDirectCodePtr(); 6332 __ LoadImmediate(LR, invoke->GetDirectCodePtr()); 6333 break; 6334 default: 6335 break; 6336 } 6337 6338 Location callee_method = temp; // For all kinds except kRecursive, callee will be in temp. 6339 switch (invoke->GetMethodLoadKind()) { 6340 case HInvokeStaticOrDirect::MethodLoadKind::kStringInit: 6341 // temp = thread->string_init_entrypoint 6342 __ LoadFromOffset(kLoadWord, temp.AsRegister<Register>(), TR, invoke->GetStringInitOffset()); 6343 break; 6344 case HInvokeStaticOrDirect::MethodLoadKind::kRecursive: 6345 callee_method = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 6346 break; 6347 case HInvokeStaticOrDirect::MethodLoadKind::kDirectAddress: 6348 __ LoadImmediate(temp.AsRegister<Register>(), invoke->GetMethodAddress()); 6349 break; 6350 case HInvokeStaticOrDirect::MethodLoadKind::kDirectAddressWithFixup: 6351 __ LoadLiteral(temp.AsRegister<Register>(), 6352 DeduplicateMethodAddressLiteral(invoke->GetTargetMethod())); 6353 break; 6354 case HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative: { 6355 HArmDexCacheArraysBase* base = 6356 invoke->InputAt(invoke->GetSpecialInputIndex())->AsArmDexCacheArraysBase(); 6357 Register base_reg = GetInvokeStaticOrDirectExtraParameter(invoke, 6358 temp.AsRegister<Register>()); 6359 int32_t offset = invoke->GetDexCacheArrayOffset() - base->GetElementOffset(); 6360 __ LoadFromOffset(kLoadWord, temp.AsRegister<Register>(), base_reg, offset); 6361 break; 6362 } 6363 case HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod: { 6364 Location current_method = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 6365 Register method_reg; 6366 Register reg = temp.AsRegister<Register>(); 6367 if (current_method.IsRegister()) { 6368 method_reg = current_method.AsRegister<Register>(); 6369 } else { 6370 DCHECK(invoke->GetLocations()->Intrinsified()); 6371 DCHECK(!current_method.IsValid()); 6372 method_reg = reg; 6373 __ LoadFromOffset(kLoadWord, reg, SP, kCurrentMethodStackOffset); 6374 } 6375 // /* ArtMethod*[] */ temp = temp.ptr_sized_fields_->dex_cache_resolved_methods_; 6376 __ LoadFromOffset(kLoadWord, 6377 reg, 6378 method_reg, 6379 ArtMethod::DexCacheResolvedMethodsOffset(kArmPointerSize).Int32Value()); 6380 // temp = temp[index_in_cache]; 6381 // Note: Don't use invoke->GetTargetMethod() as it may point to a different dex file. 6382 uint32_t index_in_cache = invoke->GetDexMethodIndex(); 6383 __ LoadFromOffset(kLoadWord, reg, reg, CodeGenerator::GetCachePointerOffset(index_in_cache)); 6384 break; 6385 } 6386 } 6387 6388 switch (invoke->GetCodePtrLocation()) { 6389 case HInvokeStaticOrDirect::CodePtrLocation::kCallSelf: 6390 __ bl(GetFrameEntryLabel()); 6391 break; 6392 case HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative: 6393 relative_call_patches_.emplace_back(invoke->GetTargetMethod()); 6394 __ BindTrackedLabel(&relative_call_patches_.back().label); 6395 // Arbitrarily branch to the BL itself, override at link time. 6396 __ bl(&relative_call_patches_.back().label); 6397 break; 6398 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup: 6399 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect: 6400 // LR prepared above for better instruction scheduling. 6401 // LR() 6402 __ blx(LR); 6403 break; 6404 case HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod: 6405 // LR = callee_method->entry_point_from_quick_compiled_code_ 6406 __ LoadFromOffset( 6407 kLoadWord, LR, callee_method.AsRegister<Register>(), 6408 ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmWordSize).Int32Value()); 6409 // LR() 6410 __ blx(LR); 6411 break; 6412 } 6413 6414 DCHECK(!IsLeafMethod()); 6415} 6416 6417void CodeGeneratorARM::GenerateVirtualCall(HInvokeVirtual* invoke, Location temp_location) { 6418 Register temp = temp_location.AsRegister<Register>(); 6419 uint32_t method_offset = mirror::Class::EmbeddedVTableEntryOffset( 6420 invoke->GetVTableIndex(), kArmPointerSize).Uint32Value(); 6421 6422 // Use the calling convention instead of the location of the receiver, as 6423 // intrinsics may have put the receiver in a different register. In the intrinsics 6424 // slow path, the arguments have been moved to the right place, so here we are 6425 // guaranteed that the receiver is the first register of the calling convention. 6426 InvokeDexCallingConvention calling_convention; 6427 Register receiver = calling_convention.GetRegisterAt(0); 6428 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 6429 // /* HeapReference<Class> */ temp = receiver->klass_ 6430 __ LoadFromOffset(kLoadWord, temp, receiver, class_offset); 6431 MaybeRecordImplicitNullCheck(invoke); 6432 // Instead of simply (possibly) unpoisoning `temp` here, we should 6433 // emit a read barrier for the previous class reference load. 6434 // However this is not required in practice, as this is an 6435 // intermediate/temporary reference and because the current 6436 // concurrent copying collector keeps the from-space memory 6437 // intact/accessible until the end of the marking phase (the 6438 // concurrent copying collector may not in the future). 6439 __ MaybeUnpoisonHeapReference(temp); 6440 // temp = temp->GetMethodAt(method_offset); 6441 uint32_t entry_point = ArtMethod::EntryPointFromQuickCompiledCodeOffset( 6442 kArmWordSize).Int32Value(); 6443 __ LoadFromOffset(kLoadWord, temp, temp, method_offset); 6444 // LR = temp->GetEntryPoint(); 6445 __ LoadFromOffset(kLoadWord, LR, temp, entry_point); 6446 // LR(); 6447 __ blx(LR); 6448} 6449 6450CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewPcRelativeStringPatch( 6451 const DexFile& dex_file, uint32_t string_index) { 6452 return NewPcRelativePatch(dex_file, string_index, &pc_relative_string_patches_); 6453} 6454 6455CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewPcRelativeDexCacheArrayPatch( 6456 const DexFile& dex_file, uint32_t element_offset) { 6457 return NewPcRelativePatch(dex_file, element_offset, &pc_relative_dex_cache_patches_); 6458} 6459 6460CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewPcRelativePatch( 6461 const DexFile& dex_file, uint32_t offset_or_index, ArenaDeque<PcRelativePatchInfo>* patches) { 6462 patches->emplace_back(dex_file, offset_or_index); 6463 return &patches->back(); 6464} 6465 6466Literal* CodeGeneratorARM::DeduplicateBootImageStringLiteral(const DexFile& dex_file, 6467 uint32_t string_index) { 6468 return boot_image_string_patches_.GetOrCreate( 6469 StringReference(&dex_file, string_index), 6470 [this]() { return __ NewLiteral<uint32_t>(/* placeholder */ 0u); }); 6471} 6472 6473Literal* CodeGeneratorARM::DeduplicateBootImageAddressLiteral(uint32_t address) { 6474 bool needs_patch = GetCompilerOptions().GetIncludePatchInformation(); 6475 Uint32ToLiteralMap* map = needs_patch ? &boot_image_address_patches_ : &uint32_literals_; 6476 return DeduplicateUint32Literal(dchecked_integral_cast<uint32_t>(address), map); 6477} 6478 6479Literal* CodeGeneratorARM::DeduplicateDexCacheAddressLiteral(uint32_t address) { 6480 return DeduplicateUint32Literal(address, &uint32_literals_); 6481} 6482 6483void CodeGeneratorARM::EmitLinkerPatches(ArenaVector<LinkerPatch>* linker_patches) { 6484 DCHECK(linker_patches->empty()); 6485 size_t size = 6486 method_patches_.size() + 6487 call_patches_.size() + 6488 relative_call_patches_.size() + 6489 /* MOVW+MOVT for each base */ 2u * pc_relative_dex_cache_patches_.size() + 6490 boot_image_string_patches_.size() + 6491 /* MOVW+MOVT for each base */ 2u * pc_relative_string_patches_.size() + 6492 boot_image_address_patches_.size(); 6493 linker_patches->reserve(size); 6494 for (const auto& entry : method_patches_) { 6495 const MethodReference& target_method = entry.first; 6496 Literal* literal = entry.second; 6497 DCHECK(literal->GetLabel()->IsBound()); 6498 uint32_t literal_offset = literal->GetLabel()->Position(); 6499 linker_patches->push_back(LinkerPatch::MethodPatch(literal_offset, 6500 target_method.dex_file, 6501 target_method.dex_method_index)); 6502 } 6503 for (const auto& entry : call_patches_) { 6504 const MethodReference& target_method = entry.first; 6505 Literal* literal = entry.second; 6506 DCHECK(literal->GetLabel()->IsBound()); 6507 uint32_t literal_offset = literal->GetLabel()->Position(); 6508 linker_patches->push_back(LinkerPatch::CodePatch(literal_offset, 6509 target_method.dex_file, 6510 target_method.dex_method_index)); 6511 } 6512 for (const MethodPatchInfo<Label>& info : relative_call_patches_) { 6513 uint32_t literal_offset = info.label.Position(); 6514 linker_patches->push_back(LinkerPatch::RelativeCodePatch(literal_offset, 6515 info.target_method.dex_file, 6516 info.target_method.dex_method_index)); 6517 } 6518 for (const PcRelativePatchInfo& info : pc_relative_dex_cache_patches_) { 6519 const DexFile& dex_file = info.target_dex_file; 6520 size_t base_element_offset = info.offset_or_index; 6521 DCHECK(info.add_pc_label.IsBound()); 6522 uint32_t add_pc_offset = dchecked_integral_cast<uint32_t>(info.add_pc_label.Position()); 6523 // Add MOVW patch. 6524 DCHECK(info.movw_label.IsBound()); 6525 uint32_t movw_offset = dchecked_integral_cast<uint32_t>(info.movw_label.Position()); 6526 linker_patches->push_back(LinkerPatch::DexCacheArrayPatch(movw_offset, 6527 &dex_file, 6528 add_pc_offset, 6529 base_element_offset)); 6530 // Add MOVT patch. 6531 DCHECK(info.movt_label.IsBound()); 6532 uint32_t movt_offset = dchecked_integral_cast<uint32_t>(info.movt_label.Position()); 6533 linker_patches->push_back(LinkerPatch::DexCacheArrayPatch(movt_offset, 6534 &dex_file, 6535 add_pc_offset, 6536 base_element_offset)); 6537 } 6538 for (const auto& entry : boot_image_string_patches_) { 6539 const StringReference& target_string = entry.first; 6540 Literal* literal = entry.second; 6541 DCHECK(literal->GetLabel()->IsBound()); 6542 uint32_t literal_offset = literal->GetLabel()->Position(); 6543 linker_patches->push_back(LinkerPatch::StringPatch(literal_offset, 6544 target_string.dex_file, 6545 target_string.string_index)); 6546 } 6547 for (const PcRelativePatchInfo& info : pc_relative_string_patches_) { 6548 const DexFile& dex_file = info.target_dex_file; 6549 uint32_t string_index = info.offset_or_index; 6550 DCHECK(info.add_pc_label.IsBound()); 6551 uint32_t add_pc_offset = dchecked_integral_cast<uint32_t>(info.add_pc_label.Position()); 6552 // Add MOVW patch. 6553 DCHECK(info.movw_label.IsBound()); 6554 uint32_t movw_offset = dchecked_integral_cast<uint32_t>(info.movw_label.Position()); 6555 linker_patches->push_back(LinkerPatch::RelativeStringPatch(movw_offset, 6556 &dex_file, 6557 add_pc_offset, 6558 string_index)); 6559 // Add MOVT patch. 6560 DCHECK(info.movt_label.IsBound()); 6561 uint32_t movt_offset = dchecked_integral_cast<uint32_t>(info.movt_label.Position()); 6562 linker_patches->push_back(LinkerPatch::RelativeStringPatch(movt_offset, 6563 &dex_file, 6564 add_pc_offset, 6565 string_index)); 6566 } 6567 for (const auto& entry : boot_image_address_patches_) { 6568 DCHECK(GetCompilerOptions().GetIncludePatchInformation()); 6569 Literal* literal = entry.second; 6570 DCHECK(literal->GetLabel()->IsBound()); 6571 uint32_t literal_offset = literal->GetLabel()->Position(); 6572 linker_patches->push_back(LinkerPatch::RecordPosition(literal_offset)); 6573 } 6574} 6575 6576Literal* CodeGeneratorARM::DeduplicateUint32Literal(uint32_t value, Uint32ToLiteralMap* map) { 6577 return map->GetOrCreate( 6578 value, 6579 [this, value]() { return __ NewLiteral<uint32_t>(value); }); 6580} 6581 6582Literal* CodeGeneratorARM::DeduplicateMethodLiteral(MethodReference target_method, 6583 MethodToLiteralMap* map) { 6584 return map->GetOrCreate( 6585 target_method, 6586 [this]() { return __ NewLiteral<uint32_t>(/* placeholder */ 0u); }); 6587} 6588 6589Literal* CodeGeneratorARM::DeduplicateMethodAddressLiteral(MethodReference target_method) { 6590 return DeduplicateMethodLiteral(target_method, &method_patches_); 6591} 6592 6593Literal* CodeGeneratorARM::DeduplicateMethodCodeLiteral(MethodReference target_method) { 6594 return DeduplicateMethodLiteral(target_method, &call_patches_); 6595} 6596 6597void LocationsBuilderARM::VisitMultiplyAccumulate(HMultiplyAccumulate* instr) { 6598 LocationSummary* locations = 6599 new (GetGraph()->GetArena()) LocationSummary(instr, LocationSummary::kNoCall); 6600 locations->SetInAt(HMultiplyAccumulate::kInputAccumulatorIndex, 6601 Location::RequiresRegister()); 6602 locations->SetInAt(HMultiplyAccumulate::kInputMulLeftIndex, Location::RequiresRegister()); 6603 locations->SetInAt(HMultiplyAccumulate::kInputMulRightIndex, Location::RequiresRegister()); 6604 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 6605} 6606 6607void InstructionCodeGeneratorARM::VisitMultiplyAccumulate(HMultiplyAccumulate* instr) { 6608 LocationSummary* locations = instr->GetLocations(); 6609 Register res = locations->Out().AsRegister<Register>(); 6610 Register accumulator = 6611 locations->InAt(HMultiplyAccumulate::kInputAccumulatorIndex).AsRegister<Register>(); 6612 Register mul_left = 6613 locations->InAt(HMultiplyAccumulate::kInputMulLeftIndex).AsRegister<Register>(); 6614 Register mul_right = 6615 locations->InAt(HMultiplyAccumulate::kInputMulRightIndex).AsRegister<Register>(); 6616 6617 if (instr->GetOpKind() == HInstruction::kAdd) { 6618 __ mla(res, mul_left, mul_right, accumulator); 6619 } else { 6620 __ mls(res, mul_left, mul_right, accumulator); 6621 } 6622} 6623 6624void LocationsBuilderARM::VisitBoundType(HBoundType* instruction ATTRIBUTE_UNUSED) { 6625 // Nothing to do, this should be removed during prepare for register allocator. 6626 LOG(FATAL) << "Unreachable"; 6627} 6628 6629void InstructionCodeGeneratorARM::VisitBoundType(HBoundType* instruction ATTRIBUTE_UNUSED) { 6630 // Nothing to do, this should be removed during prepare for register allocator. 6631 LOG(FATAL) << "Unreachable"; 6632} 6633 6634// Simple implementation of packed switch - generate cascaded compare/jumps. 6635void LocationsBuilderARM::VisitPackedSwitch(HPackedSwitch* switch_instr) { 6636 LocationSummary* locations = 6637 new (GetGraph()->GetArena()) LocationSummary(switch_instr, LocationSummary::kNoCall); 6638 locations->SetInAt(0, Location::RequiresRegister()); 6639 if (switch_instr->GetNumEntries() > kPackedSwitchCompareJumpThreshold && 6640 codegen_->GetAssembler()->IsThumb()) { 6641 locations->AddTemp(Location::RequiresRegister()); // We need a temp for the table base. 6642 if (switch_instr->GetStartValue() != 0) { 6643 locations->AddTemp(Location::RequiresRegister()); // We need a temp for the bias. 6644 } 6645 } 6646} 6647 6648void InstructionCodeGeneratorARM::VisitPackedSwitch(HPackedSwitch* switch_instr) { 6649 int32_t lower_bound = switch_instr->GetStartValue(); 6650 uint32_t num_entries = switch_instr->GetNumEntries(); 6651 LocationSummary* locations = switch_instr->GetLocations(); 6652 Register value_reg = locations->InAt(0).AsRegister<Register>(); 6653 HBasicBlock* default_block = switch_instr->GetDefaultBlock(); 6654 6655 if (num_entries <= kPackedSwitchCompareJumpThreshold || !codegen_->GetAssembler()->IsThumb()) { 6656 // Create a series of compare/jumps. 6657 Register temp_reg = IP; 6658 // Note: It is fine for the below AddConstantSetFlags() using IP register to temporarily store 6659 // the immediate, because IP is used as the destination register. For the other 6660 // AddConstantSetFlags() and GenerateCompareWithImmediate(), the immediate values are constant, 6661 // and they can be encoded in the instruction without making use of IP register. 6662 __ AddConstantSetFlags(temp_reg, value_reg, -lower_bound); 6663 6664 const ArenaVector<HBasicBlock*>& successors = switch_instr->GetBlock()->GetSuccessors(); 6665 // Jump to successors[0] if value == lower_bound. 6666 __ b(codegen_->GetLabelOf(successors[0]), EQ); 6667 int32_t last_index = 0; 6668 for (; num_entries - last_index > 2; last_index += 2) { 6669 __ AddConstantSetFlags(temp_reg, temp_reg, -2); 6670 // Jump to successors[last_index + 1] if value < case_value[last_index + 2]. 6671 __ b(codegen_->GetLabelOf(successors[last_index + 1]), LO); 6672 // Jump to successors[last_index + 2] if value == case_value[last_index + 2]. 6673 __ b(codegen_->GetLabelOf(successors[last_index + 2]), EQ); 6674 } 6675 if (num_entries - last_index == 2) { 6676 // The last missing case_value. 6677 __ CmpConstant(temp_reg, 1); 6678 __ b(codegen_->GetLabelOf(successors[last_index + 1]), EQ); 6679 } 6680 6681 // And the default for any other value. 6682 if (!codegen_->GoesToNextBlock(switch_instr->GetBlock(), default_block)) { 6683 __ b(codegen_->GetLabelOf(default_block)); 6684 } 6685 } else { 6686 // Create a table lookup. 6687 Register temp_reg = locations->GetTemp(0).AsRegister<Register>(); 6688 6689 // Materialize a pointer to the switch table 6690 std::vector<Label*> labels(num_entries); 6691 const ArenaVector<HBasicBlock*>& successors = switch_instr->GetBlock()->GetSuccessors(); 6692 for (uint32_t i = 0; i < num_entries; i++) { 6693 labels[i] = codegen_->GetLabelOf(successors[i]); 6694 } 6695 JumpTable* table = __ CreateJumpTable(std::move(labels), temp_reg); 6696 6697 // Remove the bias. 6698 Register key_reg; 6699 if (lower_bound != 0) { 6700 key_reg = locations->GetTemp(1).AsRegister<Register>(); 6701 __ AddConstant(key_reg, value_reg, -lower_bound); 6702 } else { 6703 key_reg = value_reg; 6704 } 6705 6706 // Check whether the value is in the table, jump to default block if not. 6707 __ CmpConstant(key_reg, num_entries - 1); 6708 __ b(codegen_->GetLabelOf(default_block), Condition::HI); 6709 6710 // Load the displacement from the table. 6711 __ ldr(temp_reg, Address(temp_reg, key_reg, Shift::LSL, 2)); 6712 6713 // Dispatch is a direct add to the PC (for Thumb2). 6714 __ EmitJumpTableDispatch(table, temp_reg); 6715 } 6716} 6717 6718void LocationsBuilderARM::VisitArmDexCacheArraysBase(HArmDexCacheArraysBase* base) { 6719 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(base); 6720 locations->SetOut(Location::RequiresRegister()); 6721} 6722 6723void InstructionCodeGeneratorARM::VisitArmDexCacheArraysBase(HArmDexCacheArraysBase* base) { 6724 Register base_reg = base->GetLocations()->Out().AsRegister<Register>(); 6725 CodeGeneratorARM::PcRelativePatchInfo* labels = 6726 codegen_->NewPcRelativeDexCacheArrayPatch(base->GetDexFile(), base->GetElementOffset()); 6727 __ BindTrackedLabel(&labels->movw_label); 6728 __ movw(base_reg, /* placeholder */ 0u); 6729 __ BindTrackedLabel(&labels->movt_label); 6730 __ movt(base_reg, /* placeholder */ 0u); 6731 __ BindTrackedLabel(&labels->add_pc_label); 6732 __ add(base_reg, base_reg, ShifterOperand(PC)); 6733} 6734 6735void CodeGeneratorARM::MoveFromReturnRegister(Location trg, Primitive::Type type) { 6736 if (!trg.IsValid()) { 6737 DCHECK_EQ(type, Primitive::kPrimVoid); 6738 return; 6739 } 6740 6741 DCHECK_NE(type, Primitive::kPrimVoid); 6742 6743 Location return_loc = InvokeDexCallingConventionVisitorARM().GetReturnLocation(type); 6744 if (return_loc.Equals(trg)) { 6745 return; 6746 } 6747 6748 // TODO: Consider pairs in the parallel move resolver, then this could be nicely merged 6749 // with the last branch. 6750 if (type == Primitive::kPrimLong) { 6751 HParallelMove parallel_move(GetGraph()->GetArena()); 6752 parallel_move.AddMove(return_loc.ToLow(), trg.ToLow(), Primitive::kPrimInt, nullptr); 6753 parallel_move.AddMove(return_loc.ToHigh(), trg.ToHigh(), Primitive::kPrimInt, nullptr); 6754 GetMoveResolver()->EmitNativeCode(¶llel_move); 6755 } else if (type == Primitive::kPrimDouble) { 6756 HParallelMove parallel_move(GetGraph()->GetArena()); 6757 parallel_move.AddMove(return_loc.ToLow(), trg.ToLow(), Primitive::kPrimFloat, nullptr); 6758 parallel_move.AddMove(return_loc.ToHigh(), trg.ToHigh(), Primitive::kPrimFloat, nullptr); 6759 GetMoveResolver()->EmitNativeCode(¶llel_move); 6760 } else { 6761 // Let the parallel move resolver take care of all of this. 6762 HParallelMove parallel_move(GetGraph()->GetArena()); 6763 parallel_move.AddMove(return_loc, trg, type, nullptr); 6764 GetMoveResolver()->EmitNativeCode(¶llel_move); 6765 } 6766} 6767 6768void LocationsBuilderARM::VisitClassTableGet(HClassTableGet* instruction) { 6769 LocationSummary* locations = 6770 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 6771 locations->SetInAt(0, Location::RequiresRegister()); 6772 locations->SetOut(Location::RequiresRegister()); 6773} 6774 6775void InstructionCodeGeneratorARM::VisitClassTableGet(HClassTableGet* instruction) { 6776 LocationSummary* locations = instruction->GetLocations(); 6777 uint32_t method_offset = 0; 6778 if (instruction->GetTableKind() == HClassTableGet::TableKind::kVTable) { 6779 method_offset = mirror::Class::EmbeddedVTableEntryOffset( 6780 instruction->GetIndex(), kArmPointerSize).SizeValue(); 6781 } else { 6782 __ LoadFromOffset(kLoadWord, locations->Out().AsRegister<Register>(), 6783 locations->InAt(0).AsRegister<Register>(), 6784 mirror::Class::ImtPtrOffset(kArmPointerSize).Uint32Value()); 6785 method_offset = static_cast<uint32_t>(ImTable::OffsetOfElement( 6786 instruction->GetIndex() % ImTable::kSize, kArmPointerSize)); 6787 } 6788 __ LoadFromOffset(kLoadWord, 6789 locations->Out().AsRegister<Register>(), 6790 locations->InAt(0).AsRegister<Register>(), 6791 method_offset); 6792} 6793 6794#undef __ 6795#undef QUICK_ENTRY_POINT 6796 6797} // namespace arm 6798} // namespace art 6799