code_generator_arm.cc revision 68c981fad87720fae9c799b240141ce3c12cd5bf
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 (misc-macro-parentheses) from clang-tidy. 63#define __ down_cast<ArmAssembler*>(codegen->GetAssembler())-> // NOLINT 64#define QUICK_ENTRY_POINT(x) QUICK_ENTRYPOINT_OFFSET(kArmPointerSize, 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 arm_codegen->InvokeRuntime( 123 QUICK_ENTRY_POINT(pTestSuspend), instruction_, instruction_->GetDexPc(), this); 124 CheckEntrypointTypes<kQuickTestSuspend, void, void>(); 125 if (successor_ == nullptr) { 126 __ b(GetReturnLabel()); 127 } else { 128 __ b(arm_codegen->GetLabelOf(successor_)); 129 } 130 } 131 132 Label* GetReturnLabel() { 133 DCHECK(successor_ == nullptr); 134 return &return_label_; 135 } 136 137 HBasicBlock* GetSuccessor() const { 138 return successor_; 139 } 140 141 const char* GetDescription() const OVERRIDE { return "SuspendCheckSlowPathARM"; } 142 143 private: 144 // If not null, the block to branch to after the suspend check. 145 HBasicBlock* const successor_; 146 147 // If `successor_` is null, the label to branch to after the suspend check. 148 Label return_label_; 149 150 DISALLOW_COPY_AND_ASSIGN(SuspendCheckSlowPathARM); 151}; 152 153class BoundsCheckSlowPathARM : public SlowPathCode { 154 public: 155 explicit BoundsCheckSlowPathARM(HBoundsCheck* instruction) 156 : SlowPathCode(instruction) {} 157 158 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 159 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 160 LocationSummary* locations = instruction_->GetLocations(); 161 162 __ Bind(GetEntryLabel()); 163 if (instruction_->CanThrowIntoCatchBlock()) { 164 // Live registers will be restored in the catch block if caught. 165 SaveLiveRegisters(codegen, instruction_->GetLocations()); 166 } 167 // We're moving two locations to locations that could overlap, so we need a parallel 168 // move resolver. 169 InvokeRuntimeCallingConvention calling_convention; 170 codegen->EmitParallelMoves( 171 locations->InAt(0), 172 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 173 Primitive::kPrimInt, 174 locations->InAt(1), 175 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 176 Primitive::kPrimInt); 177 uint32_t entry_point_offset = instruction_->AsBoundsCheck()->IsStringCharAt() 178 ? QUICK_ENTRY_POINT(pThrowStringBounds) 179 : QUICK_ENTRY_POINT(pThrowArrayBounds); 180 arm_codegen->InvokeRuntime(entry_point_offset, instruction_, instruction_->GetDexPc(), this); 181 CheckEntrypointTypes<kQuickThrowStringBounds, void, int32_t, int32_t>(); 182 CheckEntrypointTypes<kQuickThrowArrayBounds, void, int32_t, int32_t>(); 183 } 184 185 bool IsFatal() const OVERRIDE { return true; } 186 187 const char* GetDescription() const OVERRIDE { return "BoundsCheckSlowPathARM"; } 188 189 private: 190 DISALLOW_COPY_AND_ASSIGN(BoundsCheckSlowPathARM); 191}; 192 193class LoadClassSlowPathARM : public SlowPathCode { 194 public: 195 LoadClassSlowPathARM(HLoadClass* cls, 196 HInstruction* at, 197 uint32_t dex_pc, 198 bool do_clinit) 199 : SlowPathCode(at), cls_(cls), at_(at), dex_pc_(dex_pc), do_clinit_(do_clinit) { 200 DCHECK(at->IsLoadClass() || at->IsClinitCheck()); 201 } 202 203 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 204 LocationSummary* locations = at_->GetLocations(); 205 206 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 207 __ Bind(GetEntryLabel()); 208 SaveLiveRegisters(codegen, locations); 209 210 InvokeRuntimeCallingConvention calling_convention; 211 __ LoadImmediate(calling_convention.GetRegisterAt(0), cls_->GetTypeIndex()); 212 int32_t entry_point_offset = do_clinit_ 213 ? QUICK_ENTRY_POINT(pInitializeStaticStorage) 214 : QUICK_ENTRY_POINT(pInitializeType); 215 arm_codegen->InvokeRuntime(entry_point_offset, at_, dex_pc_, this); 216 if (do_clinit_) { 217 CheckEntrypointTypes<kQuickInitializeStaticStorage, void*, uint32_t>(); 218 } else { 219 CheckEntrypointTypes<kQuickInitializeType, void*, uint32_t>(); 220 } 221 222 // Move the class to the desired location. 223 Location out = locations->Out(); 224 if (out.IsValid()) { 225 DCHECK(out.IsRegister() && !locations->GetLiveRegisters()->ContainsCoreRegister(out.reg())); 226 arm_codegen->Move32(locations->Out(), Location::RegisterLocation(R0)); 227 } 228 RestoreLiveRegisters(codegen, locations); 229 __ b(GetExitLabel()); 230 } 231 232 const char* GetDescription() const OVERRIDE { return "LoadClassSlowPathARM"; } 233 234 private: 235 // The class this slow path will load. 236 HLoadClass* const cls_; 237 238 // The instruction where this slow path is happening. 239 // (Might be the load class or an initialization check). 240 HInstruction* const at_; 241 242 // The dex PC of `at_`. 243 const uint32_t dex_pc_; 244 245 // Whether to initialize the class. 246 const bool do_clinit_; 247 248 DISALLOW_COPY_AND_ASSIGN(LoadClassSlowPathARM); 249}; 250 251class LoadStringSlowPathARM : public SlowPathCode { 252 public: 253 explicit LoadStringSlowPathARM(HLoadString* instruction) : SlowPathCode(instruction) {} 254 255 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 256 LocationSummary* locations = instruction_->GetLocations(); 257 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg())); 258 259 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 260 __ Bind(GetEntryLabel()); 261 SaveLiveRegisters(codegen, locations); 262 263 InvokeRuntimeCallingConvention calling_convention; 264 const uint32_t string_index = instruction_->AsLoadString()->GetStringIndex(); 265 __ LoadImmediate(calling_convention.GetRegisterAt(0), string_index); 266 arm_codegen->InvokeRuntime( 267 QUICK_ENTRY_POINT(pResolveString), instruction_, instruction_->GetDexPc(), this); 268 CheckEntrypointTypes<kQuickResolveString, void*, uint32_t>(); 269 arm_codegen->Move32(locations->Out(), Location::RegisterLocation(R0)); 270 271 RestoreLiveRegisters(codegen, locations); 272 __ b(GetExitLabel()); 273 } 274 275 const char* GetDescription() const OVERRIDE { return "LoadStringSlowPathARM"; } 276 277 private: 278 DISALLOW_COPY_AND_ASSIGN(LoadStringSlowPathARM); 279}; 280 281class TypeCheckSlowPathARM : public SlowPathCode { 282 public: 283 TypeCheckSlowPathARM(HInstruction* instruction, bool is_fatal) 284 : SlowPathCode(instruction), is_fatal_(is_fatal) {} 285 286 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 287 LocationSummary* locations = instruction_->GetLocations(); 288 Location object_class = instruction_->IsCheckCast() ? locations->GetTemp(0) 289 : locations->Out(); 290 DCHECK(instruction_->IsCheckCast() 291 || !locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg())); 292 293 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 294 __ Bind(GetEntryLabel()); 295 296 if (!is_fatal_) { 297 SaveLiveRegisters(codegen, locations); 298 } 299 300 // We're moving two locations to locations that could overlap, so we need a parallel 301 // move resolver. 302 InvokeRuntimeCallingConvention calling_convention; 303 codegen->EmitParallelMoves( 304 locations->InAt(1), 305 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 306 Primitive::kPrimNot, 307 object_class, 308 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 309 Primitive::kPrimNot); 310 311 if (instruction_->IsInstanceOf()) { 312 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pInstanceofNonTrivial), 313 instruction_, 314 instruction_->GetDexPc(), 315 this); 316 CheckEntrypointTypes< 317 kQuickInstanceofNonTrivial, size_t, const mirror::Class*, const mirror::Class*>(); 318 arm_codegen->Move32(locations->Out(), Location::RegisterLocation(R0)); 319 } else { 320 DCHECK(instruction_->IsCheckCast()); 321 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pCheckCast), 322 instruction_, 323 instruction_->GetDexPc(), 324 this); 325 CheckEntrypointTypes<kQuickCheckCast, void, const mirror::Class*, const mirror::Class*>(); 326 } 327 328 if (!is_fatal_) { 329 RestoreLiveRegisters(codegen, locations); 330 __ b(GetExitLabel()); 331 } 332 } 333 334 const char* GetDescription() const OVERRIDE { return "TypeCheckSlowPathARM"; } 335 336 bool IsFatal() const OVERRIDE { return is_fatal_; } 337 338 private: 339 const bool is_fatal_; 340 341 DISALLOW_COPY_AND_ASSIGN(TypeCheckSlowPathARM); 342}; 343 344class DeoptimizationSlowPathARM : public SlowPathCode { 345 public: 346 explicit DeoptimizationSlowPathARM(HDeoptimize* instruction) 347 : SlowPathCode(instruction) {} 348 349 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 350 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 351 __ Bind(GetEntryLabel()); 352 SaveLiveRegisters(codegen, instruction_->GetLocations()); 353 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pDeoptimize), 354 instruction_, 355 instruction_->GetDexPc(), 356 this); 357 CheckEntrypointTypes<kQuickDeoptimize, void, void>(); 358 } 359 360 const char* GetDescription() const OVERRIDE { return "DeoptimizationSlowPathARM"; } 361 362 private: 363 DISALLOW_COPY_AND_ASSIGN(DeoptimizationSlowPathARM); 364}; 365 366class ArraySetSlowPathARM : public SlowPathCode { 367 public: 368 explicit ArraySetSlowPathARM(HInstruction* instruction) : SlowPathCode(instruction) {} 369 370 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 371 LocationSummary* locations = instruction_->GetLocations(); 372 __ Bind(GetEntryLabel()); 373 SaveLiveRegisters(codegen, locations); 374 375 InvokeRuntimeCallingConvention calling_convention; 376 HParallelMove parallel_move(codegen->GetGraph()->GetArena()); 377 parallel_move.AddMove( 378 locations->InAt(0), 379 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 380 Primitive::kPrimNot, 381 nullptr); 382 parallel_move.AddMove( 383 locations->InAt(1), 384 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 385 Primitive::kPrimInt, 386 nullptr); 387 parallel_move.AddMove( 388 locations->InAt(2), 389 Location::RegisterLocation(calling_convention.GetRegisterAt(2)), 390 Primitive::kPrimNot, 391 nullptr); 392 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 393 394 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 395 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pAputObject), 396 instruction_, 397 instruction_->GetDexPc(), 398 this); 399 CheckEntrypointTypes<kQuickAputObject, void, mirror::Array*, int32_t, mirror::Object*>(); 400 RestoreLiveRegisters(codegen, locations); 401 __ b(GetExitLabel()); 402 } 403 404 const char* GetDescription() const OVERRIDE { return "ArraySetSlowPathARM"; } 405 406 private: 407 DISALLOW_COPY_AND_ASSIGN(ArraySetSlowPathARM); 408}; 409 410// Slow path marking an object during a read barrier. 411class ReadBarrierMarkSlowPathARM : public SlowPathCode { 412 public: 413 ReadBarrierMarkSlowPathARM(HInstruction* instruction, Location obj) 414 : SlowPathCode(instruction), obj_(obj) { 415 DCHECK(kEmitCompilerReadBarrier); 416 } 417 418 const char* GetDescription() const OVERRIDE { return "ReadBarrierMarkSlowPathARM"; } 419 420 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 421 LocationSummary* locations = instruction_->GetLocations(); 422 Register reg = obj_.AsRegister<Register>(); 423 DCHECK(locations->CanCall()); 424 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg)); 425 DCHECK(instruction_->IsInstanceFieldGet() || 426 instruction_->IsStaticFieldGet() || 427 instruction_->IsArrayGet() || 428 instruction_->IsArraySet() || 429 instruction_->IsLoadClass() || 430 instruction_->IsLoadString() || 431 instruction_->IsInstanceOf() || 432 instruction_->IsCheckCast() || 433 (instruction_->IsInvokeVirtual() && instruction_->GetLocations()->Intrinsified()) || 434 (instruction_->IsInvokeStaticOrDirect() && instruction_->GetLocations()->Intrinsified())) 435 << "Unexpected instruction in read barrier marking slow path: " 436 << instruction_->DebugName(); 437 438 __ Bind(GetEntryLabel()); 439 // No need to save live registers; it's taken care of by the 440 // entrypoint. Also, there is no need to update the stack mask, 441 // as this runtime call will not trigger a garbage collection. 442 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 443 DCHECK_NE(reg, SP); 444 DCHECK_NE(reg, LR); 445 DCHECK_NE(reg, PC); 446 // IP is used internally by the ReadBarrierMarkRegX entry point 447 // as a temporary, it cannot be the entry point's input/output. 448 DCHECK_NE(reg, IP); 449 DCHECK(0 <= reg && reg < kNumberOfCoreRegisters) << reg; 450 // "Compact" slow path, saving two moves. 451 // 452 // Instead of using the standard runtime calling convention (input 453 // and output in R0): 454 // 455 // R0 <- obj 456 // R0 <- ReadBarrierMark(R0) 457 // obj <- R0 458 // 459 // we just use rX (the register holding `obj`) as input and output 460 // of a dedicated entrypoint: 461 // 462 // rX <- ReadBarrierMarkRegX(rX) 463 // 464 int32_t entry_point_offset = 465 CodeGenerator::GetReadBarrierMarkEntryPointsOffset<kArmPointerSize>(reg); 466 // This runtime call does not require a stack map. 467 arm_codegen->InvokeRuntimeWithoutRecordingPcInfo(entry_point_offset, instruction_, this); 468 __ b(GetExitLabel()); 469 } 470 471 private: 472 const Location obj_; 473 474 DISALLOW_COPY_AND_ASSIGN(ReadBarrierMarkSlowPathARM); 475}; 476 477// Slow path generating a read barrier for a heap reference. 478class ReadBarrierForHeapReferenceSlowPathARM : public SlowPathCode { 479 public: 480 ReadBarrierForHeapReferenceSlowPathARM(HInstruction* instruction, 481 Location out, 482 Location ref, 483 Location obj, 484 uint32_t offset, 485 Location index) 486 : SlowPathCode(instruction), 487 out_(out), 488 ref_(ref), 489 obj_(obj), 490 offset_(offset), 491 index_(index) { 492 DCHECK(kEmitCompilerReadBarrier); 493 // If `obj` is equal to `out` or `ref`, it means the initial object 494 // has been overwritten by (or after) the heap object reference load 495 // to be instrumented, e.g.: 496 // 497 // __ LoadFromOffset(kLoadWord, out, out, offset); 498 // codegen_->GenerateReadBarrierSlow(instruction, out_loc, out_loc, out_loc, offset); 499 // 500 // In that case, we have lost the information about the original 501 // object, and the emitted read barrier cannot work properly. 502 DCHECK(!obj.Equals(out)) << "obj=" << obj << " out=" << out; 503 DCHECK(!obj.Equals(ref)) << "obj=" << obj << " ref=" << ref; 504 } 505 506 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 507 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 508 LocationSummary* locations = instruction_->GetLocations(); 509 Register reg_out = out_.AsRegister<Register>(); 510 DCHECK(locations->CanCall()); 511 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg_out)); 512 DCHECK(instruction_->IsInstanceFieldGet() || 513 instruction_->IsStaticFieldGet() || 514 instruction_->IsArrayGet() || 515 instruction_->IsInstanceOf() || 516 instruction_->IsCheckCast() || 517 (instruction_->IsInvokeVirtual()) && instruction_->GetLocations()->Intrinsified()) 518 << "Unexpected instruction in read barrier for heap reference slow path: " 519 << instruction_->DebugName(); 520 521 __ Bind(GetEntryLabel()); 522 SaveLiveRegisters(codegen, locations); 523 524 // We may have to change the index's value, but as `index_` is a 525 // constant member (like other "inputs" of this slow path), 526 // introduce a copy of it, `index`. 527 Location index = index_; 528 if (index_.IsValid()) { 529 // Handle `index_` for HArrayGet and UnsafeGetObject/UnsafeGetObjectVolatile intrinsics. 530 if (instruction_->IsArrayGet()) { 531 // Compute the actual memory offset and store it in `index`. 532 Register index_reg = index_.AsRegister<Register>(); 533 DCHECK(locations->GetLiveRegisters()->ContainsCoreRegister(index_reg)); 534 if (codegen->IsCoreCalleeSaveRegister(index_reg)) { 535 // We are about to change the value of `index_reg` (see the 536 // calls to art::arm::Thumb2Assembler::Lsl and 537 // art::arm::Thumb2Assembler::AddConstant below), but it has 538 // not been saved by the previous call to 539 // art::SlowPathCode::SaveLiveRegisters, as it is a 540 // callee-save register -- 541 // art::SlowPathCode::SaveLiveRegisters does not consider 542 // callee-save registers, as it has been designed with the 543 // assumption that callee-save registers are supposed to be 544 // handled by the called function. So, as a callee-save 545 // register, `index_reg` _would_ eventually be saved onto 546 // the stack, but it would be too late: we would have 547 // changed its value earlier. Therefore, we manually save 548 // it here into another freely available register, 549 // `free_reg`, chosen of course among the caller-save 550 // registers (as a callee-save `free_reg` register would 551 // exhibit the same problem). 552 // 553 // Note we could have requested a temporary register from 554 // the register allocator instead; but we prefer not to, as 555 // this is a slow path, and we know we can find a 556 // caller-save register that is available. 557 Register free_reg = FindAvailableCallerSaveRegister(codegen); 558 __ Mov(free_reg, index_reg); 559 index_reg = free_reg; 560 index = Location::RegisterLocation(index_reg); 561 } else { 562 // The initial register stored in `index_` has already been 563 // saved in the call to art::SlowPathCode::SaveLiveRegisters 564 // (as it is not a callee-save register), so we can freely 565 // use it. 566 } 567 // Shifting the index value contained in `index_reg` by the scale 568 // factor (2) cannot overflow in practice, as the runtime is 569 // unable to allocate object arrays with a size larger than 570 // 2^26 - 1 (that is, 2^28 - 4 bytes). 571 __ Lsl(index_reg, index_reg, TIMES_4); 572 static_assert( 573 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 574 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 575 __ AddConstant(index_reg, index_reg, offset_); 576 } else { 577 // In the case of the UnsafeGetObject/UnsafeGetObjectVolatile 578 // intrinsics, `index_` is not shifted by a scale factor of 2 579 // (as in the case of ArrayGet), as it is actually an offset 580 // to an object field within an object. 581 DCHECK(instruction_->IsInvoke()) << instruction_->DebugName(); 582 DCHECK(instruction_->GetLocations()->Intrinsified()); 583 DCHECK((instruction_->AsInvoke()->GetIntrinsic() == Intrinsics::kUnsafeGetObject) || 584 (instruction_->AsInvoke()->GetIntrinsic() == Intrinsics::kUnsafeGetObjectVolatile)) 585 << instruction_->AsInvoke()->GetIntrinsic(); 586 DCHECK_EQ(offset_, 0U); 587 DCHECK(index_.IsRegisterPair()); 588 // UnsafeGet's offset location is a register pair, the low 589 // part contains the correct offset. 590 index = index_.ToLow(); 591 } 592 } 593 594 // We're moving two or three locations to locations that could 595 // overlap, so we need a parallel move resolver. 596 InvokeRuntimeCallingConvention calling_convention; 597 HParallelMove parallel_move(codegen->GetGraph()->GetArena()); 598 parallel_move.AddMove(ref_, 599 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 600 Primitive::kPrimNot, 601 nullptr); 602 parallel_move.AddMove(obj_, 603 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 604 Primitive::kPrimNot, 605 nullptr); 606 if (index.IsValid()) { 607 parallel_move.AddMove(index, 608 Location::RegisterLocation(calling_convention.GetRegisterAt(2)), 609 Primitive::kPrimInt, 610 nullptr); 611 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 612 } else { 613 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 614 __ LoadImmediate(calling_convention.GetRegisterAt(2), offset_); 615 } 616 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierSlow), 617 instruction_, 618 instruction_->GetDexPc(), 619 this); 620 CheckEntrypointTypes< 621 kQuickReadBarrierSlow, mirror::Object*, mirror::Object*, mirror::Object*, uint32_t>(); 622 arm_codegen->Move32(out_, Location::RegisterLocation(R0)); 623 624 RestoreLiveRegisters(codegen, locations); 625 __ b(GetExitLabel()); 626 } 627 628 const char* GetDescription() const OVERRIDE { return "ReadBarrierForHeapReferenceSlowPathARM"; } 629 630 private: 631 Register FindAvailableCallerSaveRegister(CodeGenerator* codegen) { 632 size_t ref = static_cast<int>(ref_.AsRegister<Register>()); 633 size_t obj = static_cast<int>(obj_.AsRegister<Register>()); 634 for (size_t i = 0, e = codegen->GetNumberOfCoreRegisters(); i < e; ++i) { 635 if (i != ref && i != obj && !codegen->IsCoreCalleeSaveRegister(i)) { 636 return static_cast<Register>(i); 637 } 638 } 639 // We shall never fail to find a free caller-save register, as 640 // there are more than two core caller-save registers on ARM 641 // (meaning it is possible to find one which is different from 642 // `ref` and `obj`). 643 DCHECK_GT(codegen->GetNumberOfCoreCallerSaveRegisters(), 2u); 644 LOG(FATAL) << "Could not find a free caller-save register"; 645 UNREACHABLE(); 646 } 647 648 const Location out_; 649 const Location ref_; 650 const Location obj_; 651 const uint32_t offset_; 652 // An additional location containing an index to an array. 653 // Only used for HArrayGet and the UnsafeGetObject & 654 // UnsafeGetObjectVolatile intrinsics. 655 const Location index_; 656 657 DISALLOW_COPY_AND_ASSIGN(ReadBarrierForHeapReferenceSlowPathARM); 658}; 659 660// Slow path generating a read barrier for a GC root. 661class ReadBarrierForRootSlowPathARM : public SlowPathCode { 662 public: 663 ReadBarrierForRootSlowPathARM(HInstruction* instruction, Location out, Location root) 664 : SlowPathCode(instruction), out_(out), root_(root) { 665 DCHECK(kEmitCompilerReadBarrier); 666 } 667 668 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 669 LocationSummary* locations = instruction_->GetLocations(); 670 Register reg_out = out_.AsRegister<Register>(); 671 DCHECK(locations->CanCall()); 672 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg_out)); 673 DCHECK(instruction_->IsLoadClass() || instruction_->IsLoadString()) 674 << "Unexpected instruction in read barrier for GC root slow path: " 675 << instruction_->DebugName(); 676 677 __ Bind(GetEntryLabel()); 678 SaveLiveRegisters(codegen, locations); 679 680 InvokeRuntimeCallingConvention calling_convention; 681 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 682 arm_codegen->Move32(Location::RegisterLocation(calling_convention.GetRegisterAt(0)), root_); 683 arm_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierForRootSlow), 684 instruction_, 685 instruction_->GetDexPc(), 686 this); 687 CheckEntrypointTypes<kQuickReadBarrierForRootSlow, mirror::Object*, GcRoot<mirror::Object>*>(); 688 arm_codegen->Move32(out_, Location::RegisterLocation(R0)); 689 690 RestoreLiveRegisters(codegen, locations); 691 __ b(GetExitLabel()); 692 } 693 694 const char* GetDescription() const OVERRIDE { return "ReadBarrierForRootSlowPathARM"; } 695 696 private: 697 const Location out_; 698 const Location root_; 699 700 DISALLOW_COPY_AND_ASSIGN(ReadBarrierForRootSlowPathARM); 701}; 702 703#undef __ 704// NOLINT on __ macro to suppress wrong warning/fix (misc-macro-parentheses) from clang-tidy. 705#define __ down_cast<ArmAssembler*>(GetAssembler())-> // NOLINT 706 707inline Condition ARMCondition(IfCondition cond) { 708 switch (cond) { 709 case kCondEQ: return EQ; 710 case kCondNE: return NE; 711 case kCondLT: return LT; 712 case kCondLE: return LE; 713 case kCondGT: return GT; 714 case kCondGE: return GE; 715 case kCondB: return LO; 716 case kCondBE: return LS; 717 case kCondA: return HI; 718 case kCondAE: return HS; 719 } 720 LOG(FATAL) << "Unreachable"; 721 UNREACHABLE(); 722} 723 724// Maps signed condition to unsigned condition. 725inline Condition ARMUnsignedCondition(IfCondition cond) { 726 switch (cond) { 727 case kCondEQ: return EQ; 728 case kCondNE: return NE; 729 // Signed to unsigned. 730 case kCondLT: return LO; 731 case kCondLE: return LS; 732 case kCondGT: return HI; 733 case kCondGE: return HS; 734 // Unsigned remain unchanged. 735 case kCondB: return LO; 736 case kCondBE: return LS; 737 case kCondA: return HI; 738 case kCondAE: return HS; 739 } 740 LOG(FATAL) << "Unreachable"; 741 UNREACHABLE(); 742} 743 744inline Condition ARMFPCondition(IfCondition cond, bool gt_bias) { 745 // The ARM condition codes can express all the necessary branches, see the 746 // "Meaning (floating-point)" column in the table A8-1 of the ARMv7 reference manual. 747 // There is no dex instruction or HIR that would need the missing conditions 748 // "equal or unordered" or "not equal". 749 switch (cond) { 750 case kCondEQ: return EQ; 751 case kCondNE: return NE /* unordered */; 752 case kCondLT: return gt_bias ? CC : LT /* unordered */; 753 case kCondLE: return gt_bias ? LS : LE /* unordered */; 754 case kCondGT: return gt_bias ? HI /* unordered */ : GT; 755 case kCondGE: return gt_bias ? CS /* unordered */ : GE; 756 default: 757 LOG(FATAL) << "UNREACHABLE"; 758 UNREACHABLE(); 759 } 760} 761 762void CodeGeneratorARM::DumpCoreRegister(std::ostream& stream, int reg) const { 763 stream << Register(reg); 764} 765 766void CodeGeneratorARM::DumpFloatingPointRegister(std::ostream& stream, int reg) const { 767 stream << SRegister(reg); 768} 769 770size_t CodeGeneratorARM::SaveCoreRegister(size_t stack_index, uint32_t reg_id) { 771 __ StoreToOffset(kStoreWord, static_cast<Register>(reg_id), SP, stack_index); 772 return kArmWordSize; 773} 774 775size_t CodeGeneratorARM::RestoreCoreRegister(size_t stack_index, uint32_t reg_id) { 776 __ LoadFromOffset(kLoadWord, static_cast<Register>(reg_id), SP, stack_index); 777 return kArmWordSize; 778} 779 780size_t CodeGeneratorARM::SaveFloatingPointRegister(size_t stack_index, uint32_t reg_id) { 781 __ StoreSToOffset(static_cast<SRegister>(reg_id), SP, stack_index); 782 return kArmWordSize; 783} 784 785size_t CodeGeneratorARM::RestoreFloatingPointRegister(size_t stack_index, uint32_t reg_id) { 786 __ LoadSFromOffset(static_cast<SRegister>(reg_id), SP, stack_index); 787 return kArmWordSize; 788} 789 790CodeGeneratorARM::CodeGeneratorARM(HGraph* graph, 791 const ArmInstructionSetFeatures& isa_features, 792 const CompilerOptions& compiler_options, 793 OptimizingCompilerStats* stats) 794 : CodeGenerator(graph, 795 kNumberOfCoreRegisters, 796 kNumberOfSRegisters, 797 kNumberOfRegisterPairs, 798 ComputeRegisterMask(reinterpret_cast<const int*>(kCoreCalleeSaves), 799 arraysize(kCoreCalleeSaves)), 800 ComputeRegisterMask(reinterpret_cast<const int*>(kFpuCalleeSaves), 801 arraysize(kFpuCalleeSaves)), 802 compiler_options, 803 stats), 804 block_labels_(nullptr), 805 location_builder_(graph, this), 806 instruction_visitor_(graph, this), 807 move_resolver_(graph->GetArena(), this), 808 assembler_(graph->GetArena()), 809 isa_features_(isa_features), 810 uint32_literals_(std::less<uint32_t>(), 811 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 812 method_patches_(MethodReferenceComparator(), 813 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 814 call_patches_(MethodReferenceComparator(), 815 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 816 relative_call_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 817 pc_relative_dex_cache_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 818 boot_image_string_patches_(StringReferenceValueComparator(), 819 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 820 pc_relative_string_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 821 boot_image_type_patches_(TypeReferenceValueComparator(), 822 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 823 pc_relative_type_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 824 boot_image_address_patches_(std::less<uint32_t>(), 825 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)) { 826 // Always save the LR register to mimic Quick. 827 AddAllocatedRegister(Location::RegisterLocation(LR)); 828} 829 830void CodeGeneratorARM::Finalize(CodeAllocator* allocator) { 831 // Ensure that we fix up branches and literal loads and emit the literal pool. 832 __ FinalizeCode(); 833 834 // Adjust native pc offsets in stack maps. 835 for (size_t i = 0, num = stack_map_stream_.GetNumberOfStackMaps(); i != num; ++i) { 836 uint32_t old_position = stack_map_stream_.GetStackMap(i).native_pc_offset; 837 uint32_t new_position = __ GetAdjustedPosition(old_position); 838 stack_map_stream_.SetStackMapNativePcOffset(i, new_position); 839 } 840 // Adjust pc offsets for the disassembly information. 841 if (disasm_info_ != nullptr) { 842 GeneratedCodeInterval* frame_entry_interval = disasm_info_->GetFrameEntryInterval(); 843 frame_entry_interval->start = __ GetAdjustedPosition(frame_entry_interval->start); 844 frame_entry_interval->end = __ GetAdjustedPosition(frame_entry_interval->end); 845 for (auto& it : *disasm_info_->GetInstructionIntervals()) { 846 it.second.start = __ GetAdjustedPosition(it.second.start); 847 it.second.end = __ GetAdjustedPosition(it.second.end); 848 } 849 for (auto& it : *disasm_info_->GetSlowPathIntervals()) { 850 it.code_interval.start = __ GetAdjustedPosition(it.code_interval.start); 851 it.code_interval.end = __ GetAdjustedPosition(it.code_interval.end); 852 } 853 } 854 855 CodeGenerator::Finalize(allocator); 856} 857 858void CodeGeneratorARM::SetupBlockedRegisters() const { 859 // Don't allocate the dalvik style register pair passing. 860 blocked_register_pairs_[R1_R2] = true; 861 862 // Stack register, LR and PC are always reserved. 863 blocked_core_registers_[SP] = true; 864 blocked_core_registers_[LR] = true; 865 blocked_core_registers_[PC] = true; 866 867 // Reserve thread register. 868 blocked_core_registers_[TR] = true; 869 870 // Reserve temp register. 871 blocked_core_registers_[IP] = true; 872 873 if (GetGraph()->IsDebuggable()) { 874 // Stubs do not save callee-save floating point registers. If the graph 875 // is debuggable, we need to deal with these registers differently. For 876 // now, just block them. 877 for (size_t i = 0; i < arraysize(kFpuCalleeSaves); ++i) { 878 blocked_fpu_registers_[kFpuCalleeSaves[i]] = true; 879 } 880 } 881 882 UpdateBlockedPairRegisters(); 883} 884 885void CodeGeneratorARM::UpdateBlockedPairRegisters() const { 886 for (int i = 0; i < kNumberOfRegisterPairs; i++) { 887 ArmManagedRegister current = 888 ArmManagedRegister::FromRegisterPair(static_cast<RegisterPair>(i)); 889 if (blocked_core_registers_[current.AsRegisterPairLow()] 890 || blocked_core_registers_[current.AsRegisterPairHigh()]) { 891 blocked_register_pairs_[i] = true; 892 } 893 } 894} 895 896InstructionCodeGeneratorARM::InstructionCodeGeneratorARM(HGraph* graph, CodeGeneratorARM* codegen) 897 : InstructionCodeGenerator(graph, codegen), 898 assembler_(codegen->GetAssembler()), 899 codegen_(codegen) {} 900 901void CodeGeneratorARM::ComputeSpillMask() { 902 core_spill_mask_ = allocated_registers_.GetCoreRegisters() & core_callee_save_mask_; 903 DCHECK_NE(core_spill_mask_, 0u) << "At least the return address register must be saved"; 904 // There is no easy instruction to restore just the PC on thumb2. We spill and 905 // restore another arbitrary register. 906 core_spill_mask_ |= (1 << kCoreAlwaysSpillRegister); 907 fpu_spill_mask_ = allocated_registers_.GetFloatingPointRegisters() & fpu_callee_save_mask_; 908 // We use vpush and vpop for saving and restoring floating point registers, which take 909 // a SRegister and the number of registers to save/restore after that SRegister. We 910 // therefore update the `fpu_spill_mask_` to also contain those registers not allocated, 911 // but in the range. 912 if (fpu_spill_mask_ != 0) { 913 uint32_t least_significant_bit = LeastSignificantBit(fpu_spill_mask_); 914 uint32_t most_significant_bit = MostSignificantBit(fpu_spill_mask_); 915 for (uint32_t i = least_significant_bit + 1 ; i < most_significant_bit; ++i) { 916 fpu_spill_mask_ |= (1 << i); 917 } 918 } 919} 920 921static dwarf::Reg DWARFReg(Register reg) { 922 return dwarf::Reg::ArmCore(static_cast<int>(reg)); 923} 924 925static dwarf::Reg DWARFReg(SRegister reg) { 926 return dwarf::Reg::ArmFp(static_cast<int>(reg)); 927} 928 929void CodeGeneratorARM::GenerateFrameEntry() { 930 bool skip_overflow_check = 931 IsLeafMethod() && !FrameNeedsStackCheck(GetFrameSize(), InstructionSet::kArm); 932 DCHECK(GetCompilerOptions().GetImplicitStackOverflowChecks()); 933 __ Bind(&frame_entry_label_); 934 935 if (HasEmptyFrame()) { 936 return; 937 } 938 939 if (!skip_overflow_check) { 940 __ AddConstant(IP, SP, -static_cast<int32_t>(GetStackOverflowReservedBytes(kArm))); 941 __ LoadFromOffset(kLoadWord, IP, IP, 0); 942 RecordPcInfo(nullptr, 0); 943 } 944 945 __ PushList(core_spill_mask_); 946 __ cfi().AdjustCFAOffset(kArmWordSize * POPCOUNT(core_spill_mask_)); 947 __ cfi().RelOffsetForMany(DWARFReg(kMethodRegisterArgument), 0, core_spill_mask_, kArmWordSize); 948 if (fpu_spill_mask_ != 0) { 949 SRegister start_register = SRegister(LeastSignificantBit(fpu_spill_mask_)); 950 __ vpushs(start_register, POPCOUNT(fpu_spill_mask_)); 951 __ cfi().AdjustCFAOffset(kArmWordSize * POPCOUNT(fpu_spill_mask_)); 952 __ cfi().RelOffsetForMany(DWARFReg(S0), 0, fpu_spill_mask_, kArmWordSize); 953 } 954 int adjust = GetFrameSize() - FrameEntrySpillSize(); 955 __ AddConstant(SP, -adjust); 956 __ cfi().AdjustCFAOffset(adjust); 957 __ StoreToOffset(kStoreWord, kMethodRegisterArgument, SP, 0); 958} 959 960void CodeGeneratorARM::GenerateFrameExit() { 961 if (HasEmptyFrame()) { 962 __ bx(LR); 963 return; 964 } 965 __ cfi().RememberState(); 966 int adjust = GetFrameSize() - FrameEntrySpillSize(); 967 __ AddConstant(SP, adjust); 968 __ cfi().AdjustCFAOffset(-adjust); 969 if (fpu_spill_mask_ != 0) { 970 SRegister start_register = SRegister(LeastSignificantBit(fpu_spill_mask_)); 971 __ vpops(start_register, POPCOUNT(fpu_spill_mask_)); 972 __ cfi().AdjustCFAOffset(-static_cast<int>(kArmPointerSize) * POPCOUNT(fpu_spill_mask_)); 973 __ cfi().RestoreMany(DWARFReg(SRegister(0)), fpu_spill_mask_); 974 } 975 // Pop LR into PC to return. 976 DCHECK_NE(core_spill_mask_ & (1 << LR), 0U); 977 uint32_t pop_mask = (core_spill_mask_ & (~(1 << LR))) | 1 << PC; 978 __ PopList(pop_mask); 979 __ cfi().RestoreState(); 980 __ cfi().DefCFAOffset(GetFrameSize()); 981} 982 983void CodeGeneratorARM::Bind(HBasicBlock* block) { 984 Label* label = GetLabelOf(block); 985 __ BindTrackedLabel(label); 986} 987 988Location InvokeDexCallingConventionVisitorARM::GetNextLocation(Primitive::Type type) { 989 switch (type) { 990 case Primitive::kPrimBoolean: 991 case Primitive::kPrimByte: 992 case Primitive::kPrimChar: 993 case Primitive::kPrimShort: 994 case Primitive::kPrimInt: 995 case Primitive::kPrimNot: { 996 uint32_t index = gp_index_++; 997 uint32_t stack_index = stack_index_++; 998 if (index < calling_convention.GetNumberOfRegisters()) { 999 return Location::RegisterLocation(calling_convention.GetRegisterAt(index)); 1000 } else { 1001 return Location::StackSlot(calling_convention.GetStackOffsetOf(stack_index)); 1002 } 1003 } 1004 1005 case Primitive::kPrimLong: { 1006 uint32_t index = gp_index_; 1007 uint32_t stack_index = stack_index_; 1008 gp_index_ += 2; 1009 stack_index_ += 2; 1010 if (index + 1 < calling_convention.GetNumberOfRegisters()) { 1011 if (calling_convention.GetRegisterAt(index) == R1) { 1012 // Skip R1, and use R2_R3 instead. 1013 gp_index_++; 1014 index++; 1015 } 1016 } 1017 if (index + 1 < calling_convention.GetNumberOfRegisters()) { 1018 DCHECK_EQ(calling_convention.GetRegisterAt(index) + 1, 1019 calling_convention.GetRegisterAt(index + 1)); 1020 1021 return Location::RegisterPairLocation(calling_convention.GetRegisterAt(index), 1022 calling_convention.GetRegisterAt(index + 1)); 1023 } else { 1024 return Location::DoubleStackSlot(calling_convention.GetStackOffsetOf(stack_index)); 1025 } 1026 } 1027 1028 case Primitive::kPrimFloat: { 1029 uint32_t stack_index = stack_index_++; 1030 if (float_index_ % 2 == 0) { 1031 float_index_ = std::max(double_index_, float_index_); 1032 } 1033 if (float_index_ < calling_convention.GetNumberOfFpuRegisters()) { 1034 return Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(float_index_++)); 1035 } else { 1036 return Location::StackSlot(calling_convention.GetStackOffsetOf(stack_index)); 1037 } 1038 } 1039 1040 case Primitive::kPrimDouble: { 1041 double_index_ = std::max(double_index_, RoundUp(float_index_, 2)); 1042 uint32_t stack_index = stack_index_; 1043 stack_index_ += 2; 1044 if (double_index_ + 1 < calling_convention.GetNumberOfFpuRegisters()) { 1045 uint32_t index = double_index_; 1046 double_index_ += 2; 1047 Location result = Location::FpuRegisterPairLocation( 1048 calling_convention.GetFpuRegisterAt(index), 1049 calling_convention.GetFpuRegisterAt(index + 1)); 1050 DCHECK(ExpectedPairLayout(result)); 1051 return result; 1052 } else { 1053 return Location::DoubleStackSlot(calling_convention.GetStackOffsetOf(stack_index)); 1054 } 1055 } 1056 1057 case Primitive::kPrimVoid: 1058 LOG(FATAL) << "Unexpected parameter type " << type; 1059 break; 1060 } 1061 return Location::NoLocation(); 1062} 1063 1064Location InvokeDexCallingConventionVisitorARM::GetReturnLocation(Primitive::Type type) const { 1065 switch (type) { 1066 case Primitive::kPrimBoolean: 1067 case Primitive::kPrimByte: 1068 case Primitive::kPrimChar: 1069 case Primitive::kPrimShort: 1070 case Primitive::kPrimInt: 1071 case Primitive::kPrimNot: { 1072 return Location::RegisterLocation(R0); 1073 } 1074 1075 case Primitive::kPrimFloat: { 1076 return Location::FpuRegisterLocation(S0); 1077 } 1078 1079 case Primitive::kPrimLong: { 1080 return Location::RegisterPairLocation(R0, R1); 1081 } 1082 1083 case Primitive::kPrimDouble: { 1084 return Location::FpuRegisterPairLocation(S0, S1); 1085 } 1086 1087 case Primitive::kPrimVoid: 1088 return Location::NoLocation(); 1089 } 1090 1091 UNREACHABLE(); 1092} 1093 1094Location InvokeDexCallingConventionVisitorARM::GetMethodLocation() const { 1095 return Location::RegisterLocation(kMethodRegisterArgument); 1096} 1097 1098void CodeGeneratorARM::Move32(Location destination, Location source) { 1099 if (source.Equals(destination)) { 1100 return; 1101 } 1102 if (destination.IsRegister()) { 1103 if (source.IsRegister()) { 1104 __ Mov(destination.AsRegister<Register>(), source.AsRegister<Register>()); 1105 } else if (source.IsFpuRegister()) { 1106 __ vmovrs(destination.AsRegister<Register>(), source.AsFpuRegister<SRegister>()); 1107 } else { 1108 __ LoadFromOffset(kLoadWord, destination.AsRegister<Register>(), SP, source.GetStackIndex()); 1109 } 1110 } else if (destination.IsFpuRegister()) { 1111 if (source.IsRegister()) { 1112 __ vmovsr(destination.AsFpuRegister<SRegister>(), source.AsRegister<Register>()); 1113 } else if (source.IsFpuRegister()) { 1114 __ vmovs(destination.AsFpuRegister<SRegister>(), source.AsFpuRegister<SRegister>()); 1115 } else { 1116 __ LoadSFromOffset(destination.AsFpuRegister<SRegister>(), SP, source.GetStackIndex()); 1117 } 1118 } else { 1119 DCHECK(destination.IsStackSlot()) << destination; 1120 if (source.IsRegister()) { 1121 __ StoreToOffset(kStoreWord, source.AsRegister<Register>(), SP, destination.GetStackIndex()); 1122 } else if (source.IsFpuRegister()) { 1123 __ StoreSToOffset(source.AsFpuRegister<SRegister>(), SP, destination.GetStackIndex()); 1124 } else { 1125 DCHECK(source.IsStackSlot()) << source; 1126 __ LoadFromOffset(kLoadWord, IP, SP, source.GetStackIndex()); 1127 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 1128 } 1129 } 1130} 1131 1132void CodeGeneratorARM::Move64(Location destination, Location source) { 1133 if (source.Equals(destination)) { 1134 return; 1135 } 1136 if (destination.IsRegisterPair()) { 1137 if (source.IsRegisterPair()) { 1138 EmitParallelMoves( 1139 Location::RegisterLocation(source.AsRegisterPairHigh<Register>()), 1140 Location::RegisterLocation(destination.AsRegisterPairHigh<Register>()), 1141 Primitive::kPrimInt, 1142 Location::RegisterLocation(source.AsRegisterPairLow<Register>()), 1143 Location::RegisterLocation(destination.AsRegisterPairLow<Register>()), 1144 Primitive::kPrimInt); 1145 } else if (source.IsFpuRegister()) { 1146 UNIMPLEMENTED(FATAL); 1147 } else if (source.IsFpuRegisterPair()) { 1148 __ vmovrrd(destination.AsRegisterPairLow<Register>(), 1149 destination.AsRegisterPairHigh<Register>(), 1150 FromLowSToD(source.AsFpuRegisterPairLow<SRegister>())); 1151 } else { 1152 DCHECK(source.IsDoubleStackSlot()); 1153 DCHECK(ExpectedPairLayout(destination)); 1154 __ LoadFromOffset(kLoadWordPair, destination.AsRegisterPairLow<Register>(), 1155 SP, source.GetStackIndex()); 1156 } 1157 } else if (destination.IsFpuRegisterPair()) { 1158 if (source.IsDoubleStackSlot()) { 1159 __ LoadDFromOffset(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 1160 SP, 1161 source.GetStackIndex()); 1162 } else if (source.IsRegisterPair()) { 1163 __ vmovdrr(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 1164 source.AsRegisterPairLow<Register>(), 1165 source.AsRegisterPairHigh<Register>()); 1166 } else { 1167 UNIMPLEMENTED(FATAL); 1168 } 1169 } else { 1170 DCHECK(destination.IsDoubleStackSlot()); 1171 if (source.IsRegisterPair()) { 1172 // No conflict possible, so just do the moves. 1173 if (source.AsRegisterPairLow<Register>() == R1) { 1174 DCHECK_EQ(source.AsRegisterPairHigh<Register>(), R2); 1175 __ StoreToOffset(kStoreWord, R1, SP, destination.GetStackIndex()); 1176 __ StoreToOffset(kStoreWord, R2, SP, destination.GetHighStackIndex(kArmWordSize)); 1177 } else { 1178 __ StoreToOffset(kStoreWordPair, source.AsRegisterPairLow<Register>(), 1179 SP, destination.GetStackIndex()); 1180 } 1181 } else if (source.IsFpuRegisterPair()) { 1182 __ StoreDToOffset(FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()), 1183 SP, 1184 destination.GetStackIndex()); 1185 } else { 1186 DCHECK(source.IsDoubleStackSlot()); 1187 EmitParallelMoves( 1188 Location::StackSlot(source.GetStackIndex()), 1189 Location::StackSlot(destination.GetStackIndex()), 1190 Primitive::kPrimInt, 1191 Location::StackSlot(source.GetHighStackIndex(kArmWordSize)), 1192 Location::StackSlot(destination.GetHighStackIndex(kArmWordSize)), 1193 Primitive::kPrimInt); 1194 } 1195 } 1196} 1197 1198void CodeGeneratorARM::MoveConstant(Location location, int32_t value) { 1199 DCHECK(location.IsRegister()); 1200 __ LoadImmediate(location.AsRegister<Register>(), value); 1201} 1202 1203void CodeGeneratorARM::MoveLocation(Location dst, Location src, Primitive::Type dst_type) { 1204 HParallelMove move(GetGraph()->GetArena()); 1205 move.AddMove(src, dst, dst_type, nullptr); 1206 GetMoveResolver()->EmitNativeCode(&move); 1207} 1208 1209void CodeGeneratorARM::AddLocationAsTemp(Location location, LocationSummary* locations) { 1210 if (location.IsRegister()) { 1211 locations->AddTemp(location); 1212 } else if (location.IsRegisterPair()) { 1213 locations->AddTemp(Location::RegisterLocation(location.AsRegisterPairLow<Register>())); 1214 locations->AddTemp(Location::RegisterLocation(location.AsRegisterPairHigh<Register>())); 1215 } else { 1216 UNIMPLEMENTED(FATAL) << "AddLocationAsTemp not implemented for location " << location; 1217 } 1218} 1219 1220void CodeGeneratorARM::InvokeRuntime(QuickEntrypointEnum entrypoint, 1221 HInstruction* instruction, 1222 uint32_t dex_pc, 1223 SlowPathCode* slow_path) { 1224 InvokeRuntime(GetThreadOffset<kArmPointerSize>(entrypoint).Int32Value(), 1225 instruction, 1226 dex_pc, 1227 slow_path); 1228} 1229 1230void CodeGeneratorARM::InvokeRuntime(int32_t entry_point_offset, 1231 HInstruction* instruction, 1232 uint32_t dex_pc, 1233 SlowPathCode* slow_path) { 1234 ValidateInvokeRuntime(instruction, slow_path); 1235 __ LoadFromOffset(kLoadWord, LR, TR, entry_point_offset); 1236 __ blx(LR); 1237 RecordPcInfo(instruction, dex_pc, slow_path); 1238} 1239 1240void CodeGeneratorARM::InvokeRuntimeWithoutRecordingPcInfo(int32_t entry_point_offset, 1241 HInstruction* instruction, 1242 SlowPathCode* slow_path) { 1243 ValidateInvokeRuntimeWithoutRecordingPcInfo(instruction, slow_path); 1244 __ LoadFromOffset(kLoadWord, LR, TR, entry_point_offset); 1245 __ blx(LR); 1246} 1247 1248void InstructionCodeGeneratorARM::HandleGoto(HInstruction* got, HBasicBlock* successor) { 1249 DCHECK(!successor->IsExitBlock()); 1250 1251 HBasicBlock* block = got->GetBlock(); 1252 HInstruction* previous = got->GetPrevious(); 1253 1254 HLoopInformation* info = block->GetLoopInformation(); 1255 if (info != nullptr && info->IsBackEdge(*block) && info->HasSuspendCheck()) { 1256 codegen_->ClearSpillSlotsFromLoopPhisInStackMap(info->GetSuspendCheck()); 1257 GenerateSuspendCheck(info->GetSuspendCheck(), successor); 1258 return; 1259 } 1260 1261 if (block->IsEntryBlock() && (previous != nullptr) && previous->IsSuspendCheck()) { 1262 GenerateSuspendCheck(previous->AsSuspendCheck(), nullptr); 1263 } 1264 if (!codegen_->GoesToNextBlock(got->GetBlock(), successor)) { 1265 __ b(codegen_->GetLabelOf(successor)); 1266 } 1267} 1268 1269void LocationsBuilderARM::VisitGoto(HGoto* got) { 1270 got->SetLocations(nullptr); 1271} 1272 1273void InstructionCodeGeneratorARM::VisitGoto(HGoto* got) { 1274 HandleGoto(got, got->GetSuccessor()); 1275} 1276 1277void LocationsBuilderARM::VisitTryBoundary(HTryBoundary* try_boundary) { 1278 try_boundary->SetLocations(nullptr); 1279} 1280 1281void InstructionCodeGeneratorARM::VisitTryBoundary(HTryBoundary* try_boundary) { 1282 HBasicBlock* successor = try_boundary->GetNormalFlowSuccessor(); 1283 if (!successor->IsExitBlock()) { 1284 HandleGoto(try_boundary, successor); 1285 } 1286} 1287 1288void LocationsBuilderARM::VisitExit(HExit* exit) { 1289 exit->SetLocations(nullptr); 1290} 1291 1292void InstructionCodeGeneratorARM::VisitExit(HExit* exit ATTRIBUTE_UNUSED) { 1293} 1294 1295void InstructionCodeGeneratorARM::GenerateVcmp(HInstruction* instruction) { 1296 Primitive::Type type = instruction->InputAt(0)->GetType(); 1297 Location lhs_loc = instruction->GetLocations()->InAt(0); 1298 Location rhs_loc = instruction->GetLocations()->InAt(1); 1299 if (rhs_loc.IsConstant()) { 1300 // 0.0 is the only immediate that can be encoded directly in 1301 // a VCMP instruction. 1302 // 1303 // Both the JLS (section 15.20.1) and the JVMS (section 6.5) 1304 // specify that in a floating-point comparison, positive zero 1305 // and negative zero are considered equal, so we can use the 1306 // literal 0.0 for both cases here. 1307 // 1308 // Note however that some methods (Float.equal, Float.compare, 1309 // Float.compareTo, Double.equal, Double.compare, 1310 // Double.compareTo, Math.max, Math.min, StrictMath.max, 1311 // StrictMath.min) consider 0.0 to be (strictly) greater than 1312 // -0.0. So if we ever translate calls to these methods into a 1313 // HCompare instruction, we must handle the -0.0 case with 1314 // care here. 1315 DCHECK(rhs_loc.GetConstant()->IsArithmeticZero()); 1316 if (type == Primitive::kPrimFloat) { 1317 __ vcmpsz(lhs_loc.AsFpuRegister<SRegister>()); 1318 } else { 1319 DCHECK_EQ(type, Primitive::kPrimDouble); 1320 __ vcmpdz(FromLowSToD(lhs_loc.AsFpuRegisterPairLow<SRegister>())); 1321 } 1322 } else { 1323 if (type == Primitive::kPrimFloat) { 1324 __ vcmps(lhs_loc.AsFpuRegister<SRegister>(), rhs_loc.AsFpuRegister<SRegister>()); 1325 } else { 1326 DCHECK_EQ(type, Primitive::kPrimDouble); 1327 __ vcmpd(FromLowSToD(lhs_loc.AsFpuRegisterPairLow<SRegister>()), 1328 FromLowSToD(rhs_loc.AsFpuRegisterPairLow<SRegister>())); 1329 } 1330 } 1331} 1332 1333void InstructionCodeGeneratorARM::GenerateFPJumps(HCondition* cond, 1334 Label* true_label, 1335 Label* false_label ATTRIBUTE_UNUSED) { 1336 __ vmstat(); // transfer FP status register to ARM APSR. 1337 __ b(true_label, ARMFPCondition(cond->GetCondition(), cond->IsGtBias())); 1338} 1339 1340void InstructionCodeGeneratorARM::GenerateLongComparesAndJumps(HCondition* cond, 1341 Label* true_label, 1342 Label* false_label) { 1343 LocationSummary* locations = cond->GetLocations(); 1344 Location left = locations->InAt(0); 1345 Location right = locations->InAt(1); 1346 IfCondition if_cond = cond->GetCondition(); 1347 1348 Register left_high = left.AsRegisterPairHigh<Register>(); 1349 Register left_low = left.AsRegisterPairLow<Register>(); 1350 IfCondition true_high_cond = if_cond; 1351 IfCondition false_high_cond = cond->GetOppositeCondition(); 1352 Condition final_condition = ARMUnsignedCondition(if_cond); // unsigned on lower part 1353 1354 // Set the conditions for the test, remembering that == needs to be 1355 // decided using the low words. 1356 // TODO: consider avoiding jumps with temporary and CMP low+SBC high 1357 switch (if_cond) { 1358 case kCondEQ: 1359 case kCondNE: 1360 // Nothing to do. 1361 break; 1362 case kCondLT: 1363 false_high_cond = kCondGT; 1364 break; 1365 case kCondLE: 1366 true_high_cond = kCondLT; 1367 break; 1368 case kCondGT: 1369 false_high_cond = kCondLT; 1370 break; 1371 case kCondGE: 1372 true_high_cond = kCondGT; 1373 break; 1374 case kCondB: 1375 false_high_cond = kCondA; 1376 break; 1377 case kCondBE: 1378 true_high_cond = kCondB; 1379 break; 1380 case kCondA: 1381 false_high_cond = kCondB; 1382 break; 1383 case kCondAE: 1384 true_high_cond = kCondA; 1385 break; 1386 } 1387 if (right.IsConstant()) { 1388 int64_t value = right.GetConstant()->AsLongConstant()->GetValue(); 1389 int32_t val_low = Low32Bits(value); 1390 int32_t val_high = High32Bits(value); 1391 1392 __ CmpConstant(left_high, val_high); 1393 if (if_cond == kCondNE) { 1394 __ b(true_label, ARMCondition(true_high_cond)); 1395 } else if (if_cond == kCondEQ) { 1396 __ b(false_label, ARMCondition(false_high_cond)); 1397 } else { 1398 __ b(true_label, ARMCondition(true_high_cond)); 1399 __ b(false_label, ARMCondition(false_high_cond)); 1400 } 1401 // Must be equal high, so compare the lows. 1402 __ CmpConstant(left_low, val_low); 1403 } else { 1404 Register right_high = right.AsRegisterPairHigh<Register>(); 1405 Register right_low = right.AsRegisterPairLow<Register>(); 1406 1407 __ cmp(left_high, ShifterOperand(right_high)); 1408 if (if_cond == kCondNE) { 1409 __ b(true_label, ARMCondition(true_high_cond)); 1410 } else if (if_cond == kCondEQ) { 1411 __ b(false_label, ARMCondition(false_high_cond)); 1412 } else { 1413 __ b(true_label, ARMCondition(true_high_cond)); 1414 __ b(false_label, ARMCondition(false_high_cond)); 1415 } 1416 // Must be equal high, so compare the lows. 1417 __ cmp(left_low, ShifterOperand(right_low)); 1418 } 1419 // The last comparison might be unsigned. 1420 // TODO: optimize cases where this is always true/false 1421 __ b(true_label, final_condition); 1422} 1423 1424void InstructionCodeGeneratorARM::GenerateCompareTestAndBranch(HCondition* condition, 1425 Label* true_target_in, 1426 Label* false_target_in) { 1427 // Generated branching requires both targets to be explicit. If either of the 1428 // targets is nullptr (fallthrough) use and bind `fallthrough_target` instead. 1429 Label fallthrough_target; 1430 Label* true_target = true_target_in == nullptr ? &fallthrough_target : true_target_in; 1431 Label* false_target = false_target_in == nullptr ? &fallthrough_target : false_target_in; 1432 1433 Primitive::Type type = condition->InputAt(0)->GetType(); 1434 switch (type) { 1435 case Primitive::kPrimLong: 1436 GenerateLongComparesAndJumps(condition, true_target, false_target); 1437 break; 1438 case Primitive::kPrimFloat: 1439 case Primitive::kPrimDouble: 1440 GenerateVcmp(condition); 1441 GenerateFPJumps(condition, true_target, false_target); 1442 break; 1443 default: 1444 LOG(FATAL) << "Unexpected compare type " << type; 1445 } 1446 1447 if (false_target != &fallthrough_target) { 1448 __ b(false_target); 1449 } 1450 1451 if (fallthrough_target.IsLinked()) { 1452 __ Bind(&fallthrough_target); 1453 } 1454} 1455 1456void InstructionCodeGeneratorARM::GenerateTestAndBranch(HInstruction* instruction, 1457 size_t condition_input_index, 1458 Label* true_target, 1459 Label* false_target) { 1460 HInstruction* cond = instruction->InputAt(condition_input_index); 1461 1462 if (true_target == nullptr && false_target == nullptr) { 1463 // Nothing to do. The code always falls through. 1464 return; 1465 } else if (cond->IsIntConstant()) { 1466 // Constant condition, statically compared against "true" (integer value 1). 1467 if (cond->AsIntConstant()->IsTrue()) { 1468 if (true_target != nullptr) { 1469 __ b(true_target); 1470 } 1471 } else { 1472 DCHECK(cond->AsIntConstant()->IsFalse()) << cond->AsIntConstant()->GetValue(); 1473 if (false_target != nullptr) { 1474 __ b(false_target); 1475 } 1476 } 1477 return; 1478 } 1479 1480 // The following code generates these patterns: 1481 // (1) true_target == nullptr && false_target != nullptr 1482 // - opposite condition true => branch to false_target 1483 // (2) true_target != nullptr && false_target == nullptr 1484 // - condition true => branch to true_target 1485 // (3) true_target != nullptr && false_target != nullptr 1486 // - condition true => branch to true_target 1487 // - branch to false_target 1488 if (IsBooleanValueOrMaterializedCondition(cond)) { 1489 // Condition has been materialized, compare the output to 0. 1490 Location cond_val = instruction->GetLocations()->InAt(condition_input_index); 1491 DCHECK(cond_val.IsRegister()); 1492 if (true_target == nullptr) { 1493 __ CompareAndBranchIfZero(cond_val.AsRegister<Register>(), false_target); 1494 } else { 1495 __ CompareAndBranchIfNonZero(cond_val.AsRegister<Register>(), true_target); 1496 } 1497 } else { 1498 // Condition has not been materialized. Use its inputs as the comparison and 1499 // its condition as the branch condition. 1500 HCondition* condition = cond->AsCondition(); 1501 1502 // If this is a long or FP comparison that has been folded into 1503 // the HCondition, generate the comparison directly. 1504 Primitive::Type type = condition->InputAt(0)->GetType(); 1505 if (type == Primitive::kPrimLong || Primitive::IsFloatingPointType(type)) { 1506 GenerateCompareTestAndBranch(condition, true_target, false_target); 1507 return; 1508 } 1509 1510 LocationSummary* locations = cond->GetLocations(); 1511 DCHECK(locations->InAt(0).IsRegister()); 1512 Register left = locations->InAt(0).AsRegister<Register>(); 1513 Location right = locations->InAt(1); 1514 if (right.IsRegister()) { 1515 __ cmp(left, ShifterOperand(right.AsRegister<Register>())); 1516 } else { 1517 DCHECK(right.IsConstant()); 1518 __ CmpConstant(left, CodeGenerator::GetInt32ValueOf(right.GetConstant())); 1519 } 1520 if (true_target == nullptr) { 1521 __ b(false_target, ARMCondition(condition->GetOppositeCondition())); 1522 } else { 1523 __ b(true_target, ARMCondition(condition->GetCondition())); 1524 } 1525 } 1526 1527 // If neither branch falls through (case 3), the conditional branch to `true_target` 1528 // was already emitted (case 2) and we need to emit a jump to `false_target`. 1529 if (true_target != nullptr && false_target != nullptr) { 1530 __ b(false_target); 1531 } 1532} 1533 1534void LocationsBuilderARM::VisitIf(HIf* if_instr) { 1535 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(if_instr); 1536 if (IsBooleanValueOrMaterializedCondition(if_instr->InputAt(0))) { 1537 locations->SetInAt(0, Location::RequiresRegister()); 1538 } 1539} 1540 1541void InstructionCodeGeneratorARM::VisitIf(HIf* if_instr) { 1542 HBasicBlock* true_successor = if_instr->IfTrueSuccessor(); 1543 HBasicBlock* false_successor = if_instr->IfFalseSuccessor(); 1544 Label* true_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), true_successor) ? 1545 nullptr : codegen_->GetLabelOf(true_successor); 1546 Label* false_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), false_successor) ? 1547 nullptr : codegen_->GetLabelOf(false_successor); 1548 GenerateTestAndBranch(if_instr, /* condition_input_index */ 0, true_target, false_target); 1549} 1550 1551void LocationsBuilderARM::VisitDeoptimize(HDeoptimize* deoptimize) { 1552 LocationSummary* locations = new (GetGraph()->GetArena()) 1553 LocationSummary(deoptimize, LocationSummary::kCallOnSlowPath); 1554 if (IsBooleanValueOrMaterializedCondition(deoptimize->InputAt(0))) { 1555 locations->SetInAt(0, Location::RequiresRegister()); 1556 } 1557} 1558 1559void InstructionCodeGeneratorARM::VisitDeoptimize(HDeoptimize* deoptimize) { 1560 SlowPathCode* slow_path = deopt_slow_paths_.NewSlowPath<DeoptimizationSlowPathARM>(deoptimize); 1561 GenerateTestAndBranch(deoptimize, 1562 /* condition_input_index */ 0, 1563 slow_path->GetEntryLabel(), 1564 /* false_target */ nullptr); 1565} 1566 1567void LocationsBuilderARM::VisitSelect(HSelect* select) { 1568 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(select); 1569 if (Primitive::IsFloatingPointType(select->GetType())) { 1570 locations->SetInAt(0, Location::RequiresFpuRegister()); 1571 locations->SetInAt(1, Location::RequiresFpuRegister()); 1572 } else { 1573 locations->SetInAt(0, Location::RequiresRegister()); 1574 locations->SetInAt(1, Location::RequiresRegister()); 1575 } 1576 if (IsBooleanValueOrMaterializedCondition(select->GetCondition())) { 1577 locations->SetInAt(2, Location::RequiresRegister()); 1578 } 1579 locations->SetOut(Location::SameAsFirstInput()); 1580} 1581 1582void InstructionCodeGeneratorARM::VisitSelect(HSelect* select) { 1583 LocationSummary* locations = select->GetLocations(); 1584 Label false_target; 1585 GenerateTestAndBranch(select, 1586 /* condition_input_index */ 2, 1587 /* true_target */ nullptr, 1588 &false_target); 1589 codegen_->MoveLocation(locations->Out(), locations->InAt(1), select->GetType()); 1590 __ Bind(&false_target); 1591} 1592 1593void LocationsBuilderARM::VisitNativeDebugInfo(HNativeDebugInfo* info) { 1594 new (GetGraph()->GetArena()) LocationSummary(info); 1595} 1596 1597void InstructionCodeGeneratorARM::VisitNativeDebugInfo(HNativeDebugInfo*) { 1598 // MaybeRecordNativeDebugInfo is already called implicitly in CodeGenerator::Compile. 1599} 1600 1601void CodeGeneratorARM::GenerateNop() { 1602 __ nop(); 1603} 1604 1605void LocationsBuilderARM::HandleCondition(HCondition* cond) { 1606 LocationSummary* locations = 1607 new (GetGraph()->GetArena()) LocationSummary(cond, LocationSummary::kNoCall); 1608 // Handle the long/FP comparisons made in instruction simplification. 1609 switch (cond->InputAt(0)->GetType()) { 1610 case Primitive::kPrimLong: 1611 locations->SetInAt(0, Location::RequiresRegister()); 1612 locations->SetInAt(1, Location::RegisterOrConstant(cond->InputAt(1))); 1613 if (!cond->IsEmittedAtUseSite()) { 1614 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 1615 } 1616 break; 1617 1618 case Primitive::kPrimFloat: 1619 case Primitive::kPrimDouble: 1620 locations->SetInAt(0, Location::RequiresFpuRegister()); 1621 locations->SetInAt(1, ArithmeticZeroOrFpuRegister(cond->InputAt(1))); 1622 if (!cond->IsEmittedAtUseSite()) { 1623 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 1624 } 1625 break; 1626 1627 default: 1628 locations->SetInAt(0, Location::RequiresRegister()); 1629 locations->SetInAt(1, Location::RegisterOrConstant(cond->InputAt(1))); 1630 if (!cond->IsEmittedAtUseSite()) { 1631 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 1632 } 1633 } 1634} 1635 1636void InstructionCodeGeneratorARM::HandleCondition(HCondition* cond) { 1637 if (cond->IsEmittedAtUseSite()) { 1638 return; 1639 } 1640 1641 LocationSummary* locations = cond->GetLocations(); 1642 Location left = locations->InAt(0); 1643 Location right = locations->InAt(1); 1644 Register out = locations->Out().AsRegister<Register>(); 1645 Label true_label, false_label; 1646 1647 switch (cond->InputAt(0)->GetType()) { 1648 default: { 1649 // Integer case. 1650 if (right.IsRegister()) { 1651 __ cmp(left.AsRegister<Register>(), ShifterOperand(right.AsRegister<Register>())); 1652 } else { 1653 DCHECK(right.IsConstant()); 1654 __ CmpConstant(left.AsRegister<Register>(), 1655 CodeGenerator::GetInt32ValueOf(right.GetConstant())); 1656 } 1657 __ it(ARMCondition(cond->GetCondition()), kItElse); 1658 __ mov(locations->Out().AsRegister<Register>(), ShifterOperand(1), 1659 ARMCondition(cond->GetCondition())); 1660 __ mov(locations->Out().AsRegister<Register>(), ShifterOperand(0), 1661 ARMCondition(cond->GetOppositeCondition())); 1662 return; 1663 } 1664 case Primitive::kPrimLong: 1665 GenerateLongComparesAndJumps(cond, &true_label, &false_label); 1666 break; 1667 case Primitive::kPrimFloat: 1668 case Primitive::kPrimDouble: 1669 GenerateVcmp(cond); 1670 GenerateFPJumps(cond, &true_label, &false_label); 1671 break; 1672 } 1673 1674 // Convert the jumps into the result. 1675 Label done_label; 1676 1677 // False case: result = 0. 1678 __ Bind(&false_label); 1679 __ LoadImmediate(out, 0); 1680 __ b(&done_label); 1681 1682 // True case: result = 1. 1683 __ Bind(&true_label); 1684 __ LoadImmediate(out, 1); 1685 __ Bind(&done_label); 1686} 1687 1688void LocationsBuilderARM::VisitEqual(HEqual* comp) { 1689 HandleCondition(comp); 1690} 1691 1692void InstructionCodeGeneratorARM::VisitEqual(HEqual* comp) { 1693 HandleCondition(comp); 1694} 1695 1696void LocationsBuilderARM::VisitNotEqual(HNotEqual* comp) { 1697 HandleCondition(comp); 1698} 1699 1700void InstructionCodeGeneratorARM::VisitNotEqual(HNotEqual* comp) { 1701 HandleCondition(comp); 1702} 1703 1704void LocationsBuilderARM::VisitLessThan(HLessThan* comp) { 1705 HandleCondition(comp); 1706} 1707 1708void InstructionCodeGeneratorARM::VisitLessThan(HLessThan* comp) { 1709 HandleCondition(comp); 1710} 1711 1712void LocationsBuilderARM::VisitLessThanOrEqual(HLessThanOrEqual* comp) { 1713 HandleCondition(comp); 1714} 1715 1716void InstructionCodeGeneratorARM::VisitLessThanOrEqual(HLessThanOrEqual* comp) { 1717 HandleCondition(comp); 1718} 1719 1720void LocationsBuilderARM::VisitGreaterThan(HGreaterThan* comp) { 1721 HandleCondition(comp); 1722} 1723 1724void InstructionCodeGeneratorARM::VisitGreaterThan(HGreaterThan* comp) { 1725 HandleCondition(comp); 1726} 1727 1728void LocationsBuilderARM::VisitGreaterThanOrEqual(HGreaterThanOrEqual* comp) { 1729 HandleCondition(comp); 1730} 1731 1732void InstructionCodeGeneratorARM::VisitGreaterThanOrEqual(HGreaterThanOrEqual* comp) { 1733 HandleCondition(comp); 1734} 1735 1736void LocationsBuilderARM::VisitBelow(HBelow* comp) { 1737 HandleCondition(comp); 1738} 1739 1740void InstructionCodeGeneratorARM::VisitBelow(HBelow* comp) { 1741 HandleCondition(comp); 1742} 1743 1744void LocationsBuilderARM::VisitBelowOrEqual(HBelowOrEqual* comp) { 1745 HandleCondition(comp); 1746} 1747 1748void InstructionCodeGeneratorARM::VisitBelowOrEqual(HBelowOrEqual* comp) { 1749 HandleCondition(comp); 1750} 1751 1752void LocationsBuilderARM::VisitAbove(HAbove* comp) { 1753 HandleCondition(comp); 1754} 1755 1756void InstructionCodeGeneratorARM::VisitAbove(HAbove* comp) { 1757 HandleCondition(comp); 1758} 1759 1760void LocationsBuilderARM::VisitAboveOrEqual(HAboveOrEqual* comp) { 1761 HandleCondition(comp); 1762} 1763 1764void InstructionCodeGeneratorARM::VisitAboveOrEqual(HAboveOrEqual* comp) { 1765 HandleCondition(comp); 1766} 1767 1768void LocationsBuilderARM::VisitIntConstant(HIntConstant* constant) { 1769 LocationSummary* locations = 1770 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1771 locations->SetOut(Location::ConstantLocation(constant)); 1772} 1773 1774void InstructionCodeGeneratorARM::VisitIntConstant(HIntConstant* constant ATTRIBUTE_UNUSED) { 1775 // Will be generated at use site. 1776} 1777 1778void LocationsBuilderARM::VisitNullConstant(HNullConstant* constant) { 1779 LocationSummary* locations = 1780 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1781 locations->SetOut(Location::ConstantLocation(constant)); 1782} 1783 1784void InstructionCodeGeneratorARM::VisitNullConstant(HNullConstant* constant ATTRIBUTE_UNUSED) { 1785 // Will be generated at use site. 1786} 1787 1788void LocationsBuilderARM::VisitLongConstant(HLongConstant* constant) { 1789 LocationSummary* locations = 1790 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1791 locations->SetOut(Location::ConstantLocation(constant)); 1792} 1793 1794void InstructionCodeGeneratorARM::VisitLongConstant(HLongConstant* constant ATTRIBUTE_UNUSED) { 1795 // Will be generated at use site. 1796} 1797 1798void LocationsBuilderARM::VisitFloatConstant(HFloatConstant* constant) { 1799 LocationSummary* locations = 1800 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1801 locations->SetOut(Location::ConstantLocation(constant)); 1802} 1803 1804void InstructionCodeGeneratorARM::VisitFloatConstant(HFloatConstant* constant ATTRIBUTE_UNUSED) { 1805 // Will be generated at use site. 1806} 1807 1808void LocationsBuilderARM::VisitDoubleConstant(HDoubleConstant* constant) { 1809 LocationSummary* locations = 1810 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 1811 locations->SetOut(Location::ConstantLocation(constant)); 1812} 1813 1814void InstructionCodeGeneratorARM::VisitDoubleConstant(HDoubleConstant* constant ATTRIBUTE_UNUSED) { 1815 // Will be generated at use site. 1816} 1817 1818void LocationsBuilderARM::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) { 1819 memory_barrier->SetLocations(nullptr); 1820} 1821 1822void InstructionCodeGeneratorARM::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) { 1823 codegen_->GenerateMemoryBarrier(memory_barrier->GetBarrierKind()); 1824} 1825 1826void LocationsBuilderARM::VisitReturnVoid(HReturnVoid* ret) { 1827 ret->SetLocations(nullptr); 1828} 1829 1830void InstructionCodeGeneratorARM::VisitReturnVoid(HReturnVoid* ret ATTRIBUTE_UNUSED) { 1831 codegen_->GenerateFrameExit(); 1832} 1833 1834void LocationsBuilderARM::VisitReturn(HReturn* ret) { 1835 LocationSummary* locations = 1836 new (GetGraph()->GetArena()) LocationSummary(ret, LocationSummary::kNoCall); 1837 locations->SetInAt(0, parameter_visitor_.GetReturnLocation(ret->InputAt(0)->GetType())); 1838} 1839 1840void InstructionCodeGeneratorARM::VisitReturn(HReturn* ret ATTRIBUTE_UNUSED) { 1841 codegen_->GenerateFrameExit(); 1842} 1843 1844void LocationsBuilderARM::VisitInvokeUnresolved(HInvokeUnresolved* invoke) { 1845 // The trampoline uses the same calling convention as dex calling conventions, 1846 // except instead of loading arg0/r0 with the target Method*, arg0/r0 will contain 1847 // the method_idx. 1848 HandleInvoke(invoke); 1849} 1850 1851void InstructionCodeGeneratorARM::VisitInvokeUnresolved(HInvokeUnresolved* invoke) { 1852 codegen_->GenerateInvokeUnresolvedRuntimeCall(invoke); 1853} 1854 1855void LocationsBuilderARM::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) { 1856 // Explicit clinit checks triggered by static invokes must have been pruned by 1857 // art::PrepareForRegisterAllocation. 1858 DCHECK(!invoke->IsStaticWithExplicitClinitCheck()); 1859 1860 IntrinsicLocationsBuilderARM intrinsic(codegen_); 1861 if (intrinsic.TryDispatch(invoke)) { 1862 if (invoke->GetLocations()->CanCall() && invoke->HasPcRelativeDexCache()) { 1863 invoke->GetLocations()->SetInAt(invoke->GetSpecialInputIndex(), Location::Any()); 1864 } 1865 return; 1866 } 1867 1868 HandleInvoke(invoke); 1869 1870 // For PC-relative dex cache the invoke has an extra input, the PC-relative address base. 1871 if (invoke->HasPcRelativeDexCache()) { 1872 invoke->GetLocations()->SetInAt(invoke->GetSpecialInputIndex(), Location::RequiresRegister()); 1873 } 1874} 1875 1876static bool TryGenerateIntrinsicCode(HInvoke* invoke, CodeGeneratorARM* codegen) { 1877 if (invoke->GetLocations()->Intrinsified()) { 1878 IntrinsicCodeGeneratorARM intrinsic(codegen); 1879 intrinsic.Dispatch(invoke); 1880 return true; 1881 } 1882 return false; 1883} 1884 1885void InstructionCodeGeneratorARM::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) { 1886 // Explicit clinit checks triggered by static invokes must have been pruned by 1887 // art::PrepareForRegisterAllocation. 1888 DCHECK(!invoke->IsStaticWithExplicitClinitCheck()); 1889 1890 if (TryGenerateIntrinsicCode(invoke, codegen_)) { 1891 return; 1892 } 1893 1894 LocationSummary* locations = invoke->GetLocations(); 1895 codegen_->GenerateStaticOrDirectCall( 1896 invoke, locations->HasTemps() ? locations->GetTemp(0) : Location::NoLocation()); 1897 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 1898} 1899 1900void LocationsBuilderARM::HandleInvoke(HInvoke* invoke) { 1901 InvokeDexCallingConventionVisitorARM calling_convention_visitor; 1902 CodeGenerator::CreateCommonInvokeLocationSummary(invoke, &calling_convention_visitor); 1903} 1904 1905void LocationsBuilderARM::VisitInvokeVirtual(HInvokeVirtual* invoke) { 1906 IntrinsicLocationsBuilderARM intrinsic(codegen_); 1907 if (intrinsic.TryDispatch(invoke)) { 1908 return; 1909 } 1910 1911 HandleInvoke(invoke); 1912} 1913 1914void InstructionCodeGeneratorARM::VisitInvokeVirtual(HInvokeVirtual* invoke) { 1915 if (TryGenerateIntrinsicCode(invoke, codegen_)) { 1916 return; 1917 } 1918 1919 codegen_->GenerateVirtualCall(invoke, invoke->GetLocations()->GetTemp(0)); 1920 DCHECK(!codegen_->IsLeafMethod()); 1921 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 1922} 1923 1924void LocationsBuilderARM::VisitInvokeInterface(HInvokeInterface* invoke) { 1925 HandleInvoke(invoke); 1926 // Add the hidden argument. 1927 invoke->GetLocations()->AddTemp(Location::RegisterLocation(R12)); 1928} 1929 1930void InstructionCodeGeneratorARM::VisitInvokeInterface(HInvokeInterface* invoke) { 1931 // TODO: b/18116999, our IMTs can miss an IncompatibleClassChangeError. 1932 LocationSummary* locations = invoke->GetLocations(); 1933 Register temp = locations->GetTemp(0).AsRegister<Register>(); 1934 Register hidden_reg = locations->GetTemp(1).AsRegister<Register>(); 1935 Location receiver = locations->InAt(0); 1936 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 1937 1938 // Set the hidden argument. This is safe to do this here, as R12 1939 // won't be modified thereafter, before the `blx` (call) instruction. 1940 DCHECK_EQ(R12, hidden_reg); 1941 __ LoadImmediate(hidden_reg, invoke->GetDexMethodIndex()); 1942 1943 if (receiver.IsStackSlot()) { 1944 __ LoadFromOffset(kLoadWord, temp, SP, receiver.GetStackIndex()); 1945 // /* HeapReference<Class> */ temp = temp->klass_ 1946 __ LoadFromOffset(kLoadWord, temp, temp, class_offset); 1947 } else { 1948 // /* HeapReference<Class> */ temp = receiver->klass_ 1949 __ LoadFromOffset(kLoadWord, temp, receiver.AsRegister<Register>(), class_offset); 1950 } 1951 codegen_->MaybeRecordImplicitNullCheck(invoke); 1952 // Instead of simply (possibly) unpoisoning `temp` here, we should 1953 // emit a read barrier for the previous class reference load. 1954 // However this is not required in practice, as this is an 1955 // intermediate/temporary reference and because the current 1956 // concurrent copying collector keeps the from-space memory 1957 // intact/accessible until the end of the marking phase (the 1958 // concurrent copying collector may not in the future). 1959 __ MaybeUnpoisonHeapReference(temp); 1960 __ LoadFromOffset(kLoadWord, temp, temp, 1961 mirror::Class::ImtPtrOffset(kArmPointerSize).Uint32Value()); 1962 uint32_t method_offset = static_cast<uint32_t>(ImTable::OffsetOfElement( 1963 invoke->GetImtIndex(), kArmPointerSize)); 1964 // temp = temp->GetImtEntryAt(method_offset); 1965 __ LoadFromOffset(kLoadWord, temp, temp, method_offset); 1966 uint32_t entry_point = 1967 ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmPointerSize).Int32Value(); 1968 // LR = temp->GetEntryPoint(); 1969 __ LoadFromOffset(kLoadWord, LR, temp, entry_point); 1970 // LR(); 1971 __ blx(LR); 1972 DCHECK(!codegen_->IsLeafMethod()); 1973 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 1974} 1975 1976void LocationsBuilderARM::VisitNeg(HNeg* neg) { 1977 LocationSummary* locations = 1978 new (GetGraph()->GetArena()) LocationSummary(neg, LocationSummary::kNoCall); 1979 switch (neg->GetResultType()) { 1980 case Primitive::kPrimInt: { 1981 locations->SetInAt(0, Location::RequiresRegister()); 1982 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 1983 break; 1984 } 1985 case Primitive::kPrimLong: { 1986 locations->SetInAt(0, Location::RequiresRegister()); 1987 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 1988 break; 1989 } 1990 1991 case Primitive::kPrimFloat: 1992 case Primitive::kPrimDouble: 1993 locations->SetInAt(0, Location::RequiresFpuRegister()); 1994 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 1995 break; 1996 1997 default: 1998 LOG(FATAL) << "Unexpected neg type " << neg->GetResultType(); 1999 } 2000} 2001 2002void InstructionCodeGeneratorARM::VisitNeg(HNeg* neg) { 2003 LocationSummary* locations = neg->GetLocations(); 2004 Location out = locations->Out(); 2005 Location in = locations->InAt(0); 2006 switch (neg->GetResultType()) { 2007 case Primitive::kPrimInt: 2008 DCHECK(in.IsRegister()); 2009 __ rsb(out.AsRegister<Register>(), in.AsRegister<Register>(), ShifterOperand(0)); 2010 break; 2011 2012 case Primitive::kPrimLong: 2013 DCHECK(in.IsRegisterPair()); 2014 // out.lo = 0 - in.lo (and update the carry/borrow (C) flag) 2015 __ rsbs(out.AsRegisterPairLow<Register>(), 2016 in.AsRegisterPairLow<Register>(), 2017 ShifterOperand(0)); 2018 // We cannot emit an RSC (Reverse Subtract with Carry) 2019 // instruction here, as it does not exist in the Thumb-2 2020 // instruction set. We use the following approach 2021 // using SBC and SUB instead. 2022 // 2023 // out.hi = -C 2024 __ sbc(out.AsRegisterPairHigh<Register>(), 2025 out.AsRegisterPairHigh<Register>(), 2026 ShifterOperand(out.AsRegisterPairHigh<Register>())); 2027 // out.hi = out.hi - in.hi 2028 __ sub(out.AsRegisterPairHigh<Register>(), 2029 out.AsRegisterPairHigh<Register>(), 2030 ShifterOperand(in.AsRegisterPairHigh<Register>())); 2031 break; 2032 2033 case Primitive::kPrimFloat: 2034 DCHECK(in.IsFpuRegister()); 2035 __ vnegs(out.AsFpuRegister<SRegister>(), in.AsFpuRegister<SRegister>()); 2036 break; 2037 2038 case Primitive::kPrimDouble: 2039 DCHECK(in.IsFpuRegisterPair()); 2040 __ vnegd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2041 FromLowSToD(in.AsFpuRegisterPairLow<SRegister>())); 2042 break; 2043 2044 default: 2045 LOG(FATAL) << "Unexpected neg type " << neg->GetResultType(); 2046 } 2047} 2048 2049void LocationsBuilderARM::VisitTypeConversion(HTypeConversion* conversion) { 2050 Primitive::Type result_type = conversion->GetResultType(); 2051 Primitive::Type input_type = conversion->GetInputType(); 2052 DCHECK_NE(result_type, input_type); 2053 2054 // The float-to-long, double-to-long and long-to-float type conversions 2055 // rely on a call to the runtime. 2056 LocationSummary::CallKind call_kind = 2057 (((input_type == Primitive::kPrimFloat || input_type == Primitive::kPrimDouble) 2058 && result_type == Primitive::kPrimLong) 2059 || (input_type == Primitive::kPrimLong && result_type == Primitive::kPrimFloat)) 2060 ? LocationSummary::kCallOnMainOnly 2061 : LocationSummary::kNoCall; 2062 LocationSummary* locations = 2063 new (GetGraph()->GetArena()) LocationSummary(conversion, call_kind); 2064 2065 // The Java language does not allow treating boolean as an integral type but 2066 // our bit representation makes it safe. 2067 2068 switch (result_type) { 2069 case Primitive::kPrimByte: 2070 switch (input_type) { 2071 case Primitive::kPrimLong: 2072 // Type conversion from long to byte is a result of code transformations. 2073 case Primitive::kPrimBoolean: 2074 // Boolean input is a result of code transformations. 2075 case Primitive::kPrimShort: 2076 case Primitive::kPrimInt: 2077 case Primitive::kPrimChar: 2078 // Processing a Dex `int-to-byte' instruction. 2079 locations->SetInAt(0, Location::RequiresRegister()); 2080 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2081 break; 2082 2083 default: 2084 LOG(FATAL) << "Unexpected type conversion from " << input_type 2085 << " to " << result_type; 2086 } 2087 break; 2088 2089 case Primitive::kPrimShort: 2090 switch (input_type) { 2091 case Primitive::kPrimLong: 2092 // Type conversion from long to short is a result of code transformations. 2093 case Primitive::kPrimBoolean: 2094 // Boolean input is a result of code transformations. 2095 case Primitive::kPrimByte: 2096 case Primitive::kPrimInt: 2097 case Primitive::kPrimChar: 2098 // Processing a Dex `int-to-short' instruction. 2099 locations->SetInAt(0, Location::RequiresRegister()); 2100 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2101 break; 2102 2103 default: 2104 LOG(FATAL) << "Unexpected type conversion from " << input_type 2105 << " to " << result_type; 2106 } 2107 break; 2108 2109 case Primitive::kPrimInt: 2110 switch (input_type) { 2111 case Primitive::kPrimLong: 2112 // Processing a Dex `long-to-int' instruction. 2113 locations->SetInAt(0, Location::Any()); 2114 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2115 break; 2116 2117 case Primitive::kPrimFloat: 2118 // Processing a Dex `float-to-int' instruction. 2119 locations->SetInAt(0, Location::RequiresFpuRegister()); 2120 locations->SetOut(Location::RequiresRegister()); 2121 locations->AddTemp(Location::RequiresFpuRegister()); 2122 break; 2123 2124 case Primitive::kPrimDouble: 2125 // Processing a Dex `double-to-int' instruction. 2126 locations->SetInAt(0, Location::RequiresFpuRegister()); 2127 locations->SetOut(Location::RequiresRegister()); 2128 locations->AddTemp(Location::RequiresFpuRegister()); 2129 break; 2130 2131 default: 2132 LOG(FATAL) << "Unexpected type conversion from " << input_type 2133 << " to " << result_type; 2134 } 2135 break; 2136 2137 case Primitive::kPrimLong: 2138 switch (input_type) { 2139 case Primitive::kPrimBoolean: 2140 // Boolean input is a result of code transformations. 2141 case Primitive::kPrimByte: 2142 case Primitive::kPrimShort: 2143 case Primitive::kPrimInt: 2144 case Primitive::kPrimChar: 2145 // Processing a Dex `int-to-long' instruction. 2146 locations->SetInAt(0, Location::RequiresRegister()); 2147 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2148 break; 2149 2150 case Primitive::kPrimFloat: { 2151 // Processing a Dex `float-to-long' instruction. 2152 InvokeRuntimeCallingConvention calling_convention; 2153 locations->SetInAt(0, Location::FpuRegisterLocation( 2154 calling_convention.GetFpuRegisterAt(0))); 2155 locations->SetOut(Location::RegisterPairLocation(R0, R1)); 2156 break; 2157 } 2158 2159 case Primitive::kPrimDouble: { 2160 // Processing a Dex `double-to-long' instruction. 2161 InvokeRuntimeCallingConvention calling_convention; 2162 locations->SetInAt(0, Location::FpuRegisterPairLocation( 2163 calling_convention.GetFpuRegisterAt(0), 2164 calling_convention.GetFpuRegisterAt(1))); 2165 locations->SetOut(Location::RegisterPairLocation(R0, R1)); 2166 break; 2167 } 2168 2169 default: 2170 LOG(FATAL) << "Unexpected type conversion from " << input_type 2171 << " to " << result_type; 2172 } 2173 break; 2174 2175 case Primitive::kPrimChar: 2176 switch (input_type) { 2177 case Primitive::kPrimLong: 2178 // Type conversion from long to char is a result of code transformations. 2179 case Primitive::kPrimBoolean: 2180 // Boolean input is a result of code transformations. 2181 case Primitive::kPrimByte: 2182 case Primitive::kPrimShort: 2183 case Primitive::kPrimInt: 2184 // Processing a Dex `int-to-char' instruction. 2185 locations->SetInAt(0, Location::RequiresRegister()); 2186 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2187 break; 2188 2189 default: 2190 LOG(FATAL) << "Unexpected type conversion from " << input_type 2191 << " to " << result_type; 2192 } 2193 break; 2194 2195 case Primitive::kPrimFloat: 2196 switch (input_type) { 2197 case Primitive::kPrimBoolean: 2198 // Boolean input is a result of code transformations. 2199 case Primitive::kPrimByte: 2200 case Primitive::kPrimShort: 2201 case Primitive::kPrimInt: 2202 case Primitive::kPrimChar: 2203 // Processing a Dex `int-to-float' instruction. 2204 locations->SetInAt(0, Location::RequiresRegister()); 2205 locations->SetOut(Location::RequiresFpuRegister()); 2206 break; 2207 2208 case Primitive::kPrimLong: { 2209 // Processing a Dex `long-to-float' instruction. 2210 InvokeRuntimeCallingConvention calling_convention; 2211 locations->SetInAt(0, Location::RegisterPairLocation( 2212 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1))); 2213 locations->SetOut(Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(0))); 2214 break; 2215 } 2216 2217 case Primitive::kPrimDouble: 2218 // Processing a Dex `double-to-float' instruction. 2219 locations->SetInAt(0, Location::RequiresFpuRegister()); 2220 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2221 break; 2222 2223 default: 2224 LOG(FATAL) << "Unexpected type conversion from " << input_type 2225 << " to " << result_type; 2226 }; 2227 break; 2228 2229 case Primitive::kPrimDouble: 2230 switch (input_type) { 2231 case Primitive::kPrimBoolean: 2232 // Boolean input is a result of code transformations. 2233 case Primitive::kPrimByte: 2234 case Primitive::kPrimShort: 2235 case Primitive::kPrimInt: 2236 case Primitive::kPrimChar: 2237 // Processing a Dex `int-to-double' instruction. 2238 locations->SetInAt(0, Location::RequiresRegister()); 2239 locations->SetOut(Location::RequiresFpuRegister()); 2240 break; 2241 2242 case Primitive::kPrimLong: 2243 // Processing a Dex `long-to-double' instruction. 2244 locations->SetInAt(0, Location::RequiresRegister()); 2245 locations->SetOut(Location::RequiresFpuRegister()); 2246 locations->AddTemp(Location::RequiresFpuRegister()); 2247 locations->AddTemp(Location::RequiresFpuRegister()); 2248 break; 2249 2250 case Primitive::kPrimFloat: 2251 // Processing a Dex `float-to-double' instruction. 2252 locations->SetInAt(0, Location::RequiresFpuRegister()); 2253 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2254 break; 2255 2256 default: 2257 LOG(FATAL) << "Unexpected type conversion from " << input_type 2258 << " to " << result_type; 2259 }; 2260 break; 2261 2262 default: 2263 LOG(FATAL) << "Unexpected type conversion from " << input_type 2264 << " to " << result_type; 2265 } 2266} 2267 2268void InstructionCodeGeneratorARM::VisitTypeConversion(HTypeConversion* conversion) { 2269 LocationSummary* locations = conversion->GetLocations(); 2270 Location out = locations->Out(); 2271 Location in = locations->InAt(0); 2272 Primitive::Type result_type = conversion->GetResultType(); 2273 Primitive::Type input_type = conversion->GetInputType(); 2274 DCHECK_NE(result_type, input_type); 2275 switch (result_type) { 2276 case Primitive::kPrimByte: 2277 switch (input_type) { 2278 case Primitive::kPrimLong: 2279 // Type conversion from long to byte is a result of code transformations. 2280 __ sbfx(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>(), 0, 8); 2281 break; 2282 case Primitive::kPrimBoolean: 2283 // Boolean input is a result of code transformations. 2284 case Primitive::kPrimShort: 2285 case Primitive::kPrimInt: 2286 case Primitive::kPrimChar: 2287 // Processing a Dex `int-to-byte' instruction. 2288 __ sbfx(out.AsRegister<Register>(), in.AsRegister<Register>(), 0, 8); 2289 break; 2290 2291 default: 2292 LOG(FATAL) << "Unexpected type conversion from " << input_type 2293 << " to " << result_type; 2294 } 2295 break; 2296 2297 case Primitive::kPrimShort: 2298 switch (input_type) { 2299 case Primitive::kPrimLong: 2300 // Type conversion from long to short is a result of code transformations. 2301 __ sbfx(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>(), 0, 16); 2302 break; 2303 case Primitive::kPrimBoolean: 2304 // Boolean input is a result of code transformations. 2305 case Primitive::kPrimByte: 2306 case Primitive::kPrimInt: 2307 case Primitive::kPrimChar: 2308 // Processing a Dex `int-to-short' instruction. 2309 __ sbfx(out.AsRegister<Register>(), in.AsRegister<Register>(), 0, 16); 2310 break; 2311 2312 default: 2313 LOG(FATAL) << "Unexpected type conversion from " << input_type 2314 << " to " << result_type; 2315 } 2316 break; 2317 2318 case Primitive::kPrimInt: 2319 switch (input_type) { 2320 case Primitive::kPrimLong: 2321 // Processing a Dex `long-to-int' instruction. 2322 DCHECK(out.IsRegister()); 2323 if (in.IsRegisterPair()) { 2324 __ Mov(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>()); 2325 } else if (in.IsDoubleStackSlot()) { 2326 __ LoadFromOffset(kLoadWord, out.AsRegister<Register>(), SP, in.GetStackIndex()); 2327 } else { 2328 DCHECK(in.IsConstant()); 2329 DCHECK(in.GetConstant()->IsLongConstant()); 2330 int64_t value = in.GetConstant()->AsLongConstant()->GetValue(); 2331 __ LoadImmediate(out.AsRegister<Register>(), static_cast<int32_t>(value)); 2332 } 2333 break; 2334 2335 case Primitive::kPrimFloat: { 2336 // Processing a Dex `float-to-int' instruction. 2337 SRegister temp = locations->GetTemp(0).AsFpuRegisterPairLow<SRegister>(); 2338 __ vcvtis(temp, in.AsFpuRegister<SRegister>()); 2339 __ vmovrs(out.AsRegister<Register>(), temp); 2340 break; 2341 } 2342 2343 case Primitive::kPrimDouble: { 2344 // Processing a Dex `double-to-int' instruction. 2345 SRegister temp_s = locations->GetTemp(0).AsFpuRegisterPairLow<SRegister>(); 2346 __ vcvtid(temp_s, FromLowSToD(in.AsFpuRegisterPairLow<SRegister>())); 2347 __ vmovrs(out.AsRegister<Register>(), temp_s); 2348 break; 2349 } 2350 2351 default: 2352 LOG(FATAL) << "Unexpected type conversion from " << input_type 2353 << " to " << result_type; 2354 } 2355 break; 2356 2357 case Primitive::kPrimLong: 2358 switch (input_type) { 2359 case Primitive::kPrimBoolean: 2360 // Boolean input is a result of code transformations. 2361 case Primitive::kPrimByte: 2362 case Primitive::kPrimShort: 2363 case Primitive::kPrimInt: 2364 case Primitive::kPrimChar: 2365 // Processing a Dex `int-to-long' instruction. 2366 DCHECK(out.IsRegisterPair()); 2367 DCHECK(in.IsRegister()); 2368 __ Mov(out.AsRegisterPairLow<Register>(), in.AsRegister<Register>()); 2369 // Sign extension. 2370 __ Asr(out.AsRegisterPairHigh<Register>(), 2371 out.AsRegisterPairLow<Register>(), 2372 31); 2373 break; 2374 2375 case Primitive::kPrimFloat: 2376 // Processing a Dex `float-to-long' instruction. 2377 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pF2l), 2378 conversion, 2379 conversion->GetDexPc(), 2380 nullptr); 2381 CheckEntrypointTypes<kQuickF2l, int64_t, float>(); 2382 break; 2383 2384 case Primitive::kPrimDouble: 2385 // Processing a Dex `double-to-long' instruction. 2386 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pD2l), 2387 conversion, 2388 conversion->GetDexPc(), 2389 nullptr); 2390 CheckEntrypointTypes<kQuickD2l, int64_t, double>(); 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::kPrimChar: 2400 switch (input_type) { 2401 case Primitive::kPrimLong: 2402 // Type conversion from long to char is a result of code transformations. 2403 __ ubfx(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>(), 0, 16); 2404 break; 2405 case Primitive::kPrimBoolean: 2406 // Boolean input is a result of code transformations. 2407 case Primitive::kPrimByte: 2408 case Primitive::kPrimShort: 2409 case Primitive::kPrimInt: 2410 // Processing a Dex `int-to-char' instruction. 2411 __ ubfx(out.AsRegister<Register>(), in.AsRegister<Register>(), 0, 16); 2412 break; 2413 2414 default: 2415 LOG(FATAL) << "Unexpected type conversion from " << input_type 2416 << " to " << result_type; 2417 } 2418 break; 2419 2420 case Primitive::kPrimFloat: 2421 switch (input_type) { 2422 case Primitive::kPrimBoolean: 2423 // Boolean input is a result of code transformations. 2424 case Primitive::kPrimByte: 2425 case Primitive::kPrimShort: 2426 case Primitive::kPrimInt: 2427 case Primitive::kPrimChar: { 2428 // Processing a Dex `int-to-float' instruction. 2429 __ vmovsr(out.AsFpuRegister<SRegister>(), in.AsRegister<Register>()); 2430 __ vcvtsi(out.AsFpuRegister<SRegister>(), out.AsFpuRegister<SRegister>()); 2431 break; 2432 } 2433 2434 case Primitive::kPrimLong: 2435 // Processing a Dex `long-to-float' instruction. 2436 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pL2f), 2437 conversion, 2438 conversion->GetDexPc(), 2439 nullptr); 2440 CheckEntrypointTypes<kQuickL2f, float, int64_t>(); 2441 break; 2442 2443 case Primitive::kPrimDouble: 2444 // Processing a Dex `double-to-float' instruction. 2445 __ vcvtsd(out.AsFpuRegister<SRegister>(), 2446 FromLowSToD(in.AsFpuRegisterPairLow<SRegister>())); 2447 break; 2448 2449 default: 2450 LOG(FATAL) << "Unexpected type conversion from " << input_type 2451 << " to " << result_type; 2452 }; 2453 break; 2454 2455 case Primitive::kPrimDouble: 2456 switch (input_type) { 2457 case Primitive::kPrimBoolean: 2458 // Boolean input is a result of code transformations. 2459 case Primitive::kPrimByte: 2460 case Primitive::kPrimShort: 2461 case Primitive::kPrimInt: 2462 case Primitive::kPrimChar: { 2463 // Processing a Dex `int-to-double' instruction. 2464 __ vmovsr(out.AsFpuRegisterPairLow<SRegister>(), in.AsRegister<Register>()); 2465 __ vcvtdi(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2466 out.AsFpuRegisterPairLow<SRegister>()); 2467 break; 2468 } 2469 2470 case Primitive::kPrimLong: { 2471 // Processing a Dex `long-to-double' instruction. 2472 Register low = in.AsRegisterPairLow<Register>(); 2473 Register high = in.AsRegisterPairHigh<Register>(); 2474 SRegister out_s = out.AsFpuRegisterPairLow<SRegister>(); 2475 DRegister out_d = FromLowSToD(out_s); 2476 SRegister temp_s = locations->GetTemp(0).AsFpuRegisterPairLow<SRegister>(); 2477 DRegister temp_d = FromLowSToD(temp_s); 2478 SRegister constant_s = locations->GetTemp(1).AsFpuRegisterPairLow<SRegister>(); 2479 DRegister constant_d = FromLowSToD(constant_s); 2480 2481 // temp_d = int-to-double(high) 2482 __ vmovsr(temp_s, high); 2483 __ vcvtdi(temp_d, temp_s); 2484 // constant_d = k2Pow32EncodingForDouble 2485 __ LoadDImmediate(constant_d, bit_cast<double, int64_t>(k2Pow32EncodingForDouble)); 2486 // out_d = unsigned-to-double(low) 2487 __ vmovsr(out_s, low); 2488 __ vcvtdu(out_d, out_s); 2489 // out_d += temp_d * constant_d 2490 __ vmlad(out_d, temp_d, constant_d); 2491 break; 2492 } 2493 2494 case Primitive::kPrimFloat: 2495 // Processing a Dex `float-to-double' instruction. 2496 __ vcvtds(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2497 in.AsFpuRegister<SRegister>()); 2498 break; 2499 2500 default: 2501 LOG(FATAL) << "Unexpected type conversion from " << input_type 2502 << " to " << result_type; 2503 }; 2504 break; 2505 2506 default: 2507 LOG(FATAL) << "Unexpected type conversion from " << input_type 2508 << " to " << result_type; 2509 } 2510} 2511 2512void LocationsBuilderARM::VisitAdd(HAdd* add) { 2513 LocationSummary* locations = 2514 new (GetGraph()->GetArena()) LocationSummary(add, LocationSummary::kNoCall); 2515 switch (add->GetResultType()) { 2516 case Primitive::kPrimInt: { 2517 locations->SetInAt(0, Location::RequiresRegister()); 2518 locations->SetInAt(1, Location::RegisterOrConstant(add->InputAt(1))); 2519 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2520 break; 2521 } 2522 2523 case Primitive::kPrimLong: { 2524 locations->SetInAt(0, Location::RequiresRegister()); 2525 locations->SetInAt(1, ArmEncodableConstantOrRegister(add->InputAt(1), ADD)); 2526 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2527 break; 2528 } 2529 2530 case Primitive::kPrimFloat: 2531 case Primitive::kPrimDouble: { 2532 locations->SetInAt(0, Location::RequiresFpuRegister()); 2533 locations->SetInAt(1, Location::RequiresFpuRegister()); 2534 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2535 break; 2536 } 2537 2538 default: 2539 LOG(FATAL) << "Unexpected add type " << add->GetResultType(); 2540 } 2541} 2542 2543void InstructionCodeGeneratorARM::VisitAdd(HAdd* add) { 2544 LocationSummary* locations = add->GetLocations(); 2545 Location out = locations->Out(); 2546 Location first = locations->InAt(0); 2547 Location second = locations->InAt(1); 2548 switch (add->GetResultType()) { 2549 case Primitive::kPrimInt: 2550 if (second.IsRegister()) { 2551 __ add(out.AsRegister<Register>(), 2552 first.AsRegister<Register>(), 2553 ShifterOperand(second.AsRegister<Register>())); 2554 } else { 2555 __ AddConstant(out.AsRegister<Register>(), 2556 first.AsRegister<Register>(), 2557 second.GetConstant()->AsIntConstant()->GetValue()); 2558 } 2559 break; 2560 2561 case Primitive::kPrimLong: { 2562 if (second.IsConstant()) { 2563 uint64_t value = static_cast<uint64_t>(Int64FromConstant(second.GetConstant())); 2564 GenerateAddLongConst(out, first, value); 2565 } else { 2566 DCHECK(second.IsRegisterPair()); 2567 __ adds(out.AsRegisterPairLow<Register>(), 2568 first.AsRegisterPairLow<Register>(), 2569 ShifterOperand(second.AsRegisterPairLow<Register>())); 2570 __ adc(out.AsRegisterPairHigh<Register>(), 2571 first.AsRegisterPairHigh<Register>(), 2572 ShifterOperand(second.AsRegisterPairHigh<Register>())); 2573 } 2574 break; 2575 } 2576 2577 case Primitive::kPrimFloat: 2578 __ vadds(out.AsFpuRegister<SRegister>(), 2579 first.AsFpuRegister<SRegister>(), 2580 second.AsFpuRegister<SRegister>()); 2581 break; 2582 2583 case Primitive::kPrimDouble: 2584 __ vaddd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2585 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 2586 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 2587 break; 2588 2589 default: 2590 LOG(FATAL) << "Unexpected add type " << add->GetResultType(); 2591 } 2592} 2593 2594void LocationsBuilderARM::VisitSub(HSub* sub) { 2595 LocationSummary* locations = 2596 new (GetGraph()->GetArena()) LocationSummary(sub, LocationSummary::kNoCall); 2597 switch (sub->GetResultType()) { 2598 case Primitive::kPrimInt: { 2599 locations->SetInAt(0, Location::RequiresRegister()); 2600 locations->SetInAt(1, Location::RegisterOrConstant(sub->InputAt(1))); 2601 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2602 break; 2603 } 2604 2605 case Primitive::kPrimLong: { 2606 locations->SetInAt(0, Location::RequiresRegister()); 2607 locations->SetInAt(1, ArmEncodableConstantOrRegister(sub->InputAt(1), SUB)); 2608 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2609 break; 2610 } 2611 case Primitive::kPrimFloat: 2612 case Primitive::kPrimDouble: { 2613 locations->SetInAt(0, Location::RequiresFpuRegister()); 2614 locations->SetInAt(1, Location::RequiresFpuRegister()); 2615 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2616 break; 2617 } 2618 default: 2619 LOG(FATAL) << "Unexpected sub type " << sub->GetResultType(); 2620 } 2621} 2622 2623void InstructionCodeGeneratorARM::VisitSub(HSub* sub) { 2624 LocationSummary* locations = sub->GetLocations(); 2625 Location out = locations->Out(); 2626 Location first = locations->InAt(0); 2627 Location second = locations->InAt(1); 2628 switch (sub->GetResultType()) { 2629 case Primitive::kPrimInt: { 2630 if (second.IsRegister()) { 2631 __ sub(out.AsRegister<Register>(), 2632 first.AsRegister<Register>(), 2633 ShifterOperand(second.AsRegister<Register>())); 2634 } else { 2635 __ AddConstant(out.AsRegister<Register>(), 2636 first.AsRegister<Register>(), 2637 -second.GetConstant()->AsIntConstant()->GetValue()); 2638 } 2639 break; 2640 } 2641 2642 case Primitive::kPrimLong: { 2643 if (second.IsConstant()) { 2644 uint64_t value = static_cast<uint64_t>(Int64FromConstant(second.GetConstant())); 2645 GenerateAddLongConst(out, first, -value); 2646 } else { 2647 DCHECK(second.IsRegisterPair()); 2648 __ subs(out.AsRegisterPairLow<Register>(), 2649 first.AsRegisterPairLow<Register>(), 2650 ShifterOperand(second.AsRegisterPairLow<Register>())); 2651 __ sbc(out.AsRegisterPairHigh<Register>(), 2652 first.AsRegisterPairHigh<Register>(), 2653 ShifterOperand(second.AsRegisterPairHigh<Register>())); 2654 } 2655 break; 2656 } 2657 2658 case Primitive::kPrimFloat: { 2659 __ vsubs(out.AsFpuRegister<SRegister>(), 2660 first.AsFpuRegister<SRegister>(), 2661 second.AsFpuRegister<SRegister>()); 2662 break; 2663 } 2664 2665 case Primitive::kPrimDouble: { 2666 __ vsubd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2667 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 2668 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 2669 break; 2670 } 2671 2672 2673 default: 2674 LOG(FATAL) << "Unexpected sub type " << sub->GetResultType(); 2675 } 2676} 2677 2678void LocationsBuilderARM::VisitMul(HMul* mul) { 2679 LocationSummary* locations = 2680 new (GetGraph()->GetArena()) LocationSummary(mul, LocationSummary::kNoCall); 2681 switch (mul->GetResultType()) { 2682 case Primitive::kPrimInt: 2683 case Primitive::kPrimLong: { 2684 locations->SetInAt(0, Location::RequiresRegister()); 2685 locations->SetInAt(1, Location::RequiresRegister()); 2686 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2687 break; 2688 } 2689 2690 case Primitive::kPrimFloat: 2691 case Primitive::kPrimDouble: { 2692 locations->SetInAt(0, Location::RequiresFpuRegister()); 2693 locations->SetInAt(1, Location::RequiresFpuRegister()); 2694 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2695 break; 2696 } 2697 2698 default: 2699 LOG(FATAL) << "Unexpected mul type " << mul->GetResultType(); 2700 } 2701} 2702 2703void InstructionCodeGeneratorARM::VisitMul(HMul* mul) { 2704 LocationSummary* locations = mul->GetLocations(); 2705 Location out = locations->Out(); 2706 Location first = locations->InAt(0); 2707 Location second = locations->InAt(1); 2708 switch (mul->GetResultType()) { 2709 case Primitive::kPrimInt: { 2710 __ mul(out.AsRegister<Register>(), 2711 first.AsRegister<Register>(), 2712 second.AsRegister<Register>()); 2713 break; 2714 } 2715 case Primitive::kPrimLong: { 2716 Register out_hi = out.AsRegisterPairHigh<Register>(); 2717 Register out_lo = out.AsRegisterPairLow<Register>(); 2718 Register in1_hi = first.AsRegisterPairHigh<Register>(); 2719 Register in1_lo = first.AsRegisterPairLow<Register>(); 2720 Register in2_hi = second.AsRegisterPairHigh<Register>(); 2721 Register in2_lo = second.AsRegisterPairLow<Register>(); 2722 2723 // Extra checks to protect caused by the existence of R1_R2. 2724 // The algorithm is wrong if out.hi is either in1.lo or in2.lo: 2725 // (e.g. in1=r0_r1, in2=r2_r3 and out=r1_r2); 2726 DCHECK_NE(out_hi, in1_lo); 2727 DCHECK_NE(out_hi, in2_lo); 2728 2729 // input: in1 - 64 bits, in2 - 64 bits 2730 // output: out 2731 // formula: out.hi : out.lo = (in1.lo * in2.hi + in1.hi * in2.lo)* 2^32 + in1.lo * in2.lo 2732 // parts: out.hi = in1.lo * in2.hi + in1.hi * in2.lo + (in1.lo * in2.lo)[63:32] 2733 // parts: out.lo = (in1.lo * in2.lo)[31:0] 2734 2735 // IP <- in1.lo * in2.hi 2736 __ mul(IP, in1_lo, in2_hi); 2737 // out.hi <- in1.lo * in2.hi + in1.hi * in2.lo 2738 __ mla(out_hi, in1_hi, in2_lo, IP); 2739 // out.lo <- (in1.lo * in2.lo)[31:0]; 2740 __ umull(out_lo, IP, in1_lo, in2_lo); 2741 // out.hi <- in2.hi * in1.lo + in2.lo * in1.hi + (in1.lo * in2.lo)[63:32] 2742 __ add(out_hi, out_hi, ShifterOperand(IP)); 2743 break; 2744 } 2745 2746 case Primitive::kPrimFloat: { 2747 __ vmuls(out.AsFpuRegister<SRegister>(), 2748 first.AsFpuRegister<SRegister>(), 2749 second.AsFpuRegister<SRegister>()); 2750 break; 2751 } 2752 2753 case Primitive::kPrimDouble: { 2754 __ vmuld(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2755 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 2756 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 2757 break; 2758 } 2759 2760 default: 2761 LOG(FATAL) << "Unexpected mul type " << mul->GetResultType(); 2762 } 2763} 2764 2765void InstructionCodeGeneratorARM::DivRemOneOrMinusOne(HBinaryOperation* instruction) { 2766 DCHECK(instruction->IsDiv() || instruction->IsRem()); 2767 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 2768 2769 LocationSummary* locations = instruction->GetLocations(); 2770 Location second = locations->InAt(1); 2771 DCHECK(second.IsConstant()); 2772 2773 Register out = locations->Out().AsRegister<Register>(); 2774 Register dividend = locations->InAt(0).AsRegister<Register>(); 2775 int32_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 2776 DCHECK(imm == 1 || imm == -1); 2777 2778 if (instruction->IsRem()) { 2779 __ LoadImmediate(out, 0); 2780 } else { 2781 if (imm == 1) { 2782 __ Mov(out, dividend); 2783 } else { 2784 __ rsb(out, dividend, ShifterOperand(0)); 2785 } 2786 } 2787} 2788 2789void InstructionCodeGeneratorARM::DivRemByPowerOfTwo(HBinaryOperation* instruction) { 2790 DCHECK(instruction->IsDiv() || instruction->IsRem()); 2791 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 2792 2793 LocationSummary* locations = instruction->GetLocations(); 2794 Location second = locations->InAt(1); 2795 DCHECK(second.IsConstant()); 2796 2797 Register out = locations->Out().AsRegister<Register>(); 2798 Register dividend = locations->InAt(0).AsRegister<Register>(); 2799 Register temp = locations->GetTemp(0).AsRegister<Register>(); 2800 int32_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 2801 uint32_t abs_imm = static_cast<uint32_t>(AbsOrMin(imm)); 2802 int ctz_imm = CTZ(abs_imm); 2803 2804 if (ctz_imm == 1) { 2805 __ Lsr(temp, dividend, 32 - ctz_imm); 2806 } else { 2807 __ Asr(temp, dividend, 31); 2808 __ Lsr(temp, temp, 32 - ctz_imm); 2809 } 2810 __ add(out, temp, ShifterOperand(dividend)); 2811 2812 if (instruction->IsDiv()) { 2813 __ Asr(out, out, ctz_imm); 2814 if (imm < 0) { 2815 __ rsb(out, out, ShifterOperand(0)); 2816 } 2817 } else { 2818 __ ubfx(out, out, 0, ctz_imm); 2819 __ sub(out, out, ShifterOperand(temp)); 2820 } 2821} 2822 2823void InstructionCodeGeneratorARM::GenerateDivRemWithAnyConstant(HBinaryOperation* instruction) { 2824 DCHECK(instruction->IsDiv() || instruction->IsRem()); 2825 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 2826 2827 LocationSummary* locations = instruction->GetLocations(); 2828 Location second = locations->InAt(1); 2829 DCHECK(second.IsConstant()); 2830 2831 Register out = locations->Out().AsRegister<Register>(); 2832 Register dividend = locations->InAt(0).AsRegister<Register>(); 2833 Register temp1 = locations->GetTemp(0).AsRegister<Register>(); 2834 Register temp2 = locations->GetTemp(1).AsRegister<Register>(); 2835 int64_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 2836 2837 int64_t magic; 2838 int shift; 2839 CalculateMagicAndShiftForDivRem(imm, false /* is_long */, &magic, &shift); 2840 2841 __ LoadImmediate(temp1, magic); 2842 __ smull(temp2, temp1, dividend, temp1); 2843 2844 if (imm > 0 && magic < 0) { 2845 __ add(temp1, temp1, ShifterOperand(dividend)); 2846 } else if (imm < 0 && magic > 0) { 2847 __ sub(temp1, temp1, ShifterOperand(dividend)); 2848 } 2849 2850 if (shift != 0) { 2851 __ Asr(temp1, temp1, shift); 2852 } 2853 2854 if (instruction->IsDiv()) { 2855 __ sub(out, temp1, ShifterOperand(temp1, ASR, 31)); 2856 } else { 2857 __ sub(temp1, temp1, ShifterOperand(temp1, ASR, 31)); 2858 // TODO: Strength reduction for mls. 2859 __ LoadImmediate(temp2, imm); 2860 __ mls(out, temp1, temp2, dividend); 2861 } 2862} 2863 2864void InstructionCodeGeneratorARM::GenerateDivRemConstantIntegral(HBinaryOperation* instruction) { 2865 DCHECK(instruction->IsDiv() || instruction->IsRem()); 2866 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 2867 2868 LocationSummary* locations = instruction->GetLocations(); 2869 Location second = locations->InAt(1); 2870 DCHECK(second.IsConstant()); 2871 2872 int32_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 2873 if (imm == 0) { 2874 // Do not generate anything. DivZeroCheck would prevent any code to be executed. 2875 } else if (imm == 1 || imm == -1) { 2876 DivRemOneOrMinusOne(instruction); 2877 } else if (IsPowerOfTwo(AbsOrMin(imm))) { 2878 DivRemByPowerOfTwo(instruction); 2879 } else { 2880 DCHECK(imm <= -2 || imm >= 2); 2881 GenerateDivRemWithAnyConstant(instruction); 2882 } 2883} 2884 2885void LocationsBuilderARM::VisitDiv(HDiv* div) { 2886 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 2887 if (div->GetResultType() == Primitive::kPrimLong) { 2888 // pLdiv runtime call. 2889 call_kind = LocationSummary::kCallOnMainOnly; 2890 } else if (div->GetResultType() == Primitive::kPrimInt && div->InputAt(1)->IsConstant()) { 2891 // sdiv will be replaced by other instruction sequence. 2892 } else if (div->GetResultType() == Primitive::kPrimInt && 2893 !codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 2894 // pIdivmod runtime call. 2895 call_kind = LocationSummary::kCallOnMainOnly; 2896 } 2897 2898 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(div, call_kind); 2899 2900 switch (div->GetResultType()) { 2901 case Primitive::kPrimInt: { 2902 if (div->InputAt(1)->IsConstant()) { 2903 locations->SetInAt(0, Location::RequiresRegister()); 2904 locations->SetInAt(1, Location::ConstantLocation(div->InputAt(1)->AsConstant())); 2905 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2906 int32_t value = div->InputAt(1)->AsIntConstant()->GetValue(); 2907 if (value == 1 || value == 0 || value == -1) { 2908 // No temp register required. 2909 } else { 2910 locations->AddTemp(Location::RequiresRegister()); 2911 if (!IsPowerOfTwo(AbsOrMin(value))) { 2912 locations->AddTemp(Location::RequiresRegister()); 2913 } 2914 } 2915 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 2916 locations->SetInAt(0, Location::RequiresRegister()); 2917 locations->SetInAt(1, Location::RequiresRegister()); 2918 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2919 } else { 2920 InvokeRuntimeCallingConvention calling_convention; 2921 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 2922 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 2923 // Note: divrem will compute both the quotient and the remainder as the pair R0 and R1, but 2924 // we only need the former. 2925 locations->SetOut(Location::RegisterLocation(R0)); 2926 } 2927 break; 2928 } 2929 case Primitive::kPrimLong: { 2930 InvokeRuntimeCallingConvention calling_convention; 2931 locations->SetInAt(0, Location::RegisterPairLocation( 2932 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1))); 2933 locations->SetInAt(1, Location::RegisterPairLocation( 2934 calling_convention.GetRegisterAt(2), calling_convention.GetRegisterAt(3))); 2935 locations->SetOut(Location::RegisterPairLocation(R0, R1)); 2936 break; 2937 } 2938 case Primitive::kPrimFloat: 2939 case Primitive::kPrimDouble: { 2940 locations->SetInAt(0, Location::RequiresFpuRegister()); 2941 locations->SetInAt(1, Location::RequiresFpuRegister()); 2942 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2943 break; 2944 } 2945 2946 default: 2947 LOG(FATAL) << "Unexpected div type " << div->GetResultType(); 2948 } 2949} 2950 2951void InstructionCodeGeneratorARM::VisitDiv(HDiv* div) { 2952 LocationSummary* locations = div->GetLocations(); 2953 Location out = locations->Out(); 2954 Location first = locations->InAt(0); 2955 Location second = locations->InAt(1); 2956 2957 switch (div->GetResultType()) { 2958 case Primitive::kPrimInt: { 2959 if (second.IsConstant()) { 2960 GenerateDivRemConstantIntegral(div); 2961 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 2962 __ sdiv(out.AsRegister<Register>(), 2963 first.AsRegister<Register>(), 2964 second.AsRegister<Register>()); 2965 } else { 2966 InvokeRuntimeCallingConvention calling_convention; 2967 DCHECK_EQ(calling_convention.GetRegisterAt(0), first.AsRegister<Register>()); 2968 DCHECK_EQ(calling_convention.GetRegisterAt(1), second.AsRegister<Register>()); 2969 DCHECK_EQ(R0, out.AsRegister<Register>()); 2970 2971 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pIdivmod), div, div->GetDexPc(), nullptr); 2972 CheckEntrypointTypes<kQuickIdivmod, int32_t, int32_t, int32_t>(); 2973 } 2974 break; 2975 } 2976 2977 case Primitive::kPrimLong: { 2978 InvokeRuntimeCallingConvention calling_convention; 2979 DCHECK_EQ(calling_convention.GetRegisterAt(0), first.AsRegisterPairLow<Register>()); 2980 DCHECK_EQ(calling_convention.GetRegisterAt(1), first.AsRegisterPairHigh<Register>()); 2981 DCHECK_EQ(calling_convention.GetRegisterAt(2), second.AsRegisterPairLow<Register>()); 2982 DCHECK_EQ(calling_convention.GetRegisterAt(3), second.AsRegisterPairHigh<Register>()); 2983 DCHECK_EQ(R0, out.AsRegisterPairLow<Register>()); 2984 DCHECK_EQ(R1, out.AsRegisterPairHigh<Register>()); 2985 2986 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pLdiv), div, div->GetDexPc(), nullptr); 2987 CheckEntrypointTypes<kQuickLdiv, int64_t, int64_t, int64_t>(); 2988 break; 2989 } 2990 2991 case Primitive::kPrimFloat: { 2992 __ vdivs(out.AsFpuRegister<SRegister>(), 2993 first.AsFpuRegister<SRegister>(), 2994 second.AsFpuRegister<SRegister>()); 2995 break; 2996 } 2997 2998 case Primitive::kPrimDouble: { 2999 __ vdivd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 3000 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 3001 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 3002 break; 3003 } 3004 3005 default: 3006 LOG(FATAL) << "Unexpected div type " << div->GetResultType(); 3007 } 3008} 3009 3010void LocationsBuilderARM::VisitRem(HRem* rem) { 3011 Primitive::Type type = rem->GetResultType(); 3012 3013 // Most remainders are implemented in the runtime. 3014 LocationSummary::CallKind call_kind = LocationSummary::kCallOnMainOnly; 3015 if (rem->GetResultType() == Primitive::kPrimInt && rem->InputAt(1)->IsConstant()) { 3016 // sdiv will be replaced by other instruction sequence. 3017 call_kind = LocationSummary::kNoCall; 3018 } else if ((rem->GetResultType() == Primitive::kPrimInt) 3019 && codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 3020 // Have hardware divide instruction for int, do it with three instructions. 3021 call_kind = LocationSummary::kNoCall; 3022 } 3023 3024 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(rem, call_kind); 3025 3026 switch (type) { 3027 case Primitive::kPrimInt: { 3028 if (rem->InputAt(1)->IsConstant()) { 3029 locations->SetInAt(0, Location::RequiresRegister()); 3030 locations->SetInAt(1, Location::ConstantLocation(rem->InputAt(1)->AsConstant())); 3031 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3032 int32_t value = rem->InputAt(1)->AsIntConstant()->GetValue(); 3033 if (value == 1 || value == 0 || value == -1) { 3034 // No temp register required. 3035 } else { 3036 locations->AddTemp(Location::RequiresRegister()); 3037 if (!IsPowerOfTwo(AbsOrMin(value))) { 3038 locations->AddTemp(Location::RequiresRegister()); 3039 } 3040 } 3041 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 3042 locations->SetInAt(0, Location::RequiresRegister()); 3043 locations->SetInAt(1, Location::RequiresRegister()); 3044 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3045 locations->AddTemp(Location::RequiresRegister()); 3046 } else { 3047 InvokeRuntimeCallingConvention calling_convention; 3048 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 3049 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 3050 // Note: divrem will compute both the quotient and the remainder as the pair R0 and R1, but 3051 // we only need the latter. 3052 locations->SetOut(Location::RegisterLocation(R1)); 3053 } 3054 break; 3055 } 3056 case Primitive::kPrimLong: { 3057 InvokeRuntimeCallingConvention calling_convention; 3058 locations->SetInAt(0, Location::RegisterPairLocation( 3059 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1))); 3060 locations->SetInAt(1, Location::RegisterPairLocation( 3061 calling_convention.GetRegisterAt(2), calling_convention.GetRegisterAt(3))); 3062 // The runtime helper puts the output in R2,R3. 3063 locations->SetOut(Location::RegisterPairLocation(R2, R3)); 3064 break; 3065 } 3066 case Primitive::kPrimFloat: { 3067 InvokeRuntimeCallingConvention calling_convention; 3068 locations->SetInAt(0, Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(0))); 3069 locations->SetInAt(1, Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(1))); 3070 locations->SetOut(Location::FpuRegisterLocation(S0)); 3071 break; 3072 } 3073 3074 case Primitive::kPrimDouble: { 3075 InvokeRuntimeCallingConvention calling_convention; 3076 locations->SetInAt(0, Location::FpuRegisterPairLocation( 3077 calling_convention.GetFpuRegisterAt(0), calling_convention.GetFpuRegisterAt(1))); 3078 locations->SetInAt(1, Location::FpuRegisterPairLocation( 3079 calling_convention.GetFpuRegisterAt(2), calling_convention.GetFpuRegisterAt(3))); 3080 locations->SetOut(Location::Location::FpuRegisterPairLocation(S0, S1)); 3081 break; 3082 } 3083 3084 default: 3085 LOG(FATAL) << "Unexpected rem type " << type; 3086 } 3087} 3088 3089void InstructionCodeGeneratorARM::VisitRem(HRem* rem) { 3090 LocationSummary* locations = rem->GetLocations(); 3091 Location out = locations->Out(); 3092 Location first = locations->InAt(0); 3093 Location second = locations->InAt(1); 3094 3095 Primitive::Type type = rem->GetResultType(); 3096 switch (type) { 3097 case Primitive::kPrimInt: { 3098 if (second.IsConstant()) { 3099 GenerateDivRemConstantIntegral(rem); 3100 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 3101 Register reg1 = first.AsRegister<Register>(); 3102 Register reg2 = second.AsRegister<Register>(); 3103 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3104 3105 // temp = reg1 / reg2 (integer division) 3106 // dest = reg1 - temp * reg2 3107 __ sdiv(temp, reg1, reg2); 3108 __ mls(out.AsRegister<Register>(), temp, reg2, reg1); 3109 } else { 3110 InvokeRuntimeCallingConvention calling_convention; 3111 DCHECK_EQ(calling_convention.GetRegisterAt(0), first.AsRegister<Register>()); 3112 DCHECK_EQ(calling_convention.GetRegisterAt(1), second.AsRegister<Register>()); 3113 DCHECK_EQ(R1, out.AsRegister<Register>()); 3114 3115 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pIdivmod), rem, rem->GetDexPc(), nullptr); 3116 CheckEntrypointTypes<kQuickIdivmod, int32_t, int32_t, int32_t>(); 3117 } 3118 break; 3119 } 3120 3121 case Primitive::kPrimLong: { 3122 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pLmod), rem, rem->GetDexPc(), nullptr); 3123 CheckEntrypointTypes<kQuickLmod, int64_t, int64_t, int64_t>(); 3124 break; 3125 } 3126 3127 case Primitive::kPrimFloat: { 3128 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pFmodf), rem, rem->GetDexPc(), nullptr); 3129 CheckEntrypointTypes<kQuickFmodf, float, float, float>(); 3130 break; 3131 } 3132 3133 case Primitive::kPrimDouble: { 3134 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pFmod), rem, rem->GetDexPc(), nullptr); 3135 CheckEntrypointTypes<kQuickFmod, double, double, double>(); 3136 break; 3137 } 3138 3139 default: 3140 LOG(FATAL) << "Unexpected rem type " << type; 3141 } 3142} 3143 3144void LocationsBuilderARM::VisitDivZeroCheck(HDivZeroCheck* instruction) { 3145 LocationSummary::CallKind call_kind = instruction->CanThrowIntoCatchBlock() 3146 ? LocationSummary::kCallOnSlowPath 3147 : LocationSummary::kNoCall; 3148 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 3149 locations->SetInAt(0, Location::RegisterOrConstant(instruction->InputAt(0))); 3150 if (instruction->HasUses()) { 3151 locations->SetOut(Location::SameAsFirstInput()); 3152 } 3153} 3154 3155void InstructionCodeGeneratorARM::VisitDivZeroCheck(HDivZeroCheck* instruction) { 3156 SlowPathCode* slow_path = new (GetGraph()->GetArena()) DivZeroCheckSlowPathARM(instruction); 3157 codegen_->AddSlowPath(slow_path); 3158 3159 LocationSummary* locations = instruction->GetLocations(); 3160 Location value = locations->InAt(0); 3161 3162 switch (instruction->GetType()) { 3163 case Primitive::kPrimBoolean: 3164 case Primitive::kPrimByte: 3165 case Primitive::kPrimChar: 3166 case Primitive::kPrimShort: 3167 case Primitive::kPrimInt: { 3168 if (value.IsRegister()) { 3169 __ CompareAndBranchIfZero(value.AsRegister<Register>(), slow_path->GetEntryLabel()); 3170 } else { 3171 DCHECK(value.IsConstant()) << value; 3172 if (value.GetConstant()->AsIntConstant()->GetValue() == 0) { 3173 __ b(slow_path->GetEntryLabel()); 3174 } 3175 } 3176 break; 3177 } 3178 case Primitive::kPrimLong: { 3179 if (value.IsRegisterPair()) { 3180 __ orrs(IP, 3181 value.AsRegisterPairLow<Register>(), 3182 ShifterOperand(value.AsRegisterPairHigh<Register>())); 3183 __ b(slow_path->GetEntryLabel(), EQ); 3184 } else { 3185 DCHECK(value.IsConstant()) << value; 3186 if (value.GetConstant()->AsLongConstant()->GetValue() == 0) { 3187 __ b(slow_path->GetEntryLabel()); 3188 } 3189 } 3190 break; 3191 default: 3192 LOG(FATAL) << "Unexpected type for HDivZeroCheck " << instruction->GetType(); 3193 } 3194 } 3195} 3196 3197void InstructionCodeGeneratorARM::HandleIntegerRotate(LocationSummary* locations) { 3198 Register in = locations->InAt(0).AsRegister<Register>(); 3199 Location rhs = locations->InAt(1); 3200 Register out = locations->Out().AsRegister<Register>(); 3201 3202 if (rhs.IsConstant()) { 3203 // Arm32 and Thumb2 assemblers require a rotation on the interval [1,31], 3204 // so map all rotations to a +ve. equivalent in that range. 3205 // (e.g. left *or* right by -2 bits == 30 bits in the same direction.) 3206 uint32_t rot = CodeGenerator::GetInt32ValueOf(rhs.GetConstant()) & 0x1F; 3207 if (rot) { 3208 // Rotate, mapping left rotations to right equivalents if necessary. 3209 // (e.g. left by 2 bits == right by 30.) 3210 __ Ror(out, in, rot); 3211 } else if (out != in) { 3212 __ Mov(out, in); 3213 } 3214 } else { 3215 __ Ror(out, in, rhs.AsRegister<Register>()); 3216 } 3217} 3218 3219// Gain some speed by mapping all Long rotates onto equivalent pairs of Integer 3220// rotates by swapping input regs (effectively rotating by the first 32-bits of 3221// a larger rotation) or flipping direction (thus treating larger right/left 3222// rotations as sub-word sized rotations in the other direction) as appropriate. 3223void InstructionCodeGeneratorARM::HandleLongRotate(LocationSummary* locations) { 3224 Register in_reg_lo = locations->InAt(0).AsRegisterPairLow<Register>(); 3225 Register in_reg_hi = locations->InAt(0).AsRegisterPairHigh<Register>(); 3226 Location rhs = locations->InAt(1); 3227 Register out_reg_lo = locations->Out().AsRegisterPairLow<Register>(); 3228 Register out_reg_hi = locations->Out().AsRegisterPairHigh<Register>(); 3229 3230 if (rhs.IsConstant()) { 3231 uint64_t rot = CodeGenerator::GetInt64ValueOf(rhs.GetConstant()); 3232 // Map all rotations to +ve. equivalents on the interval [0,63]. 3233 rot &= kMaxLongShiftDistance; 3234 // For rotates over a word in size, 'pre-rotate' by 32-bits to keep rotate 3235 // logic below to a simple pair of binary orr. 3236 // (e.g. 34 bits == in_reg swap + 2 bits right.) 3237 if (rot >= kArmBitsPerWord) { 3238 rot -= kArmBitsPerWord; 3239 std::swap(in_reg_hi, in_reg_lo); 3240 } 3241 // Rotate, or mov to out for zero or word size rotations. 3242 if (rot != 0u) { 3243 __ Lsr(out_reg_hi, in_reg_hi, rot); 3244 __ orr(out_reg_hi, out_reg_hi, ShifterOperand(in_reg_lo, arm::LSL, kArmBitsPerWord - rot)); 3245 __ Lsr(out_reg_lo, in_reg_lo, rot); 3246 __ orr(out_reg_lo, out_reg_lo, ShifterOperand(in_reg_hi, arm::LSL, kArmBitsPerWord - rot)); 3247 } else { 3248 __ Mov(out_reg_lo, in_reg_lo); 3249 __ Mov(out_reg_hi, in_reg_hi); 3250 } 3251 } else { 3252 Register shift_right = locations->GetTemp(0).AsRegister<Register>(); 3253 Register shift_left = locations->GetTemp(1).AsRegister<Register>(); 3254 Label end; 3255 Label shift_by_32_plus_shift_right; 3256 3257 __ and_(shift_right, rhs.AsRegister<Register>(), ShifterOperand(0x1F)); 3258 __ Lsrs(shift_left, rhs.AsRegister<Register>(), 6); 3259 __ rsb(shift_left, shift_right, ShifterOperand(kArmBitsPerWord), AL, kCcKeep); 3260 __ b(&shift_by_32_plus_shift_right, CC); 3261 3262 // out_reg_hi = (reg_hi << shift_left) | (reg_lo >> shift_right). 3263 // out_reg_lo = (reg_lo << shift_left) | (reg_hi >> shift_right). 3264 __ Lsl(out_reg_hi, in_reg_hi, shift_left); 3265 __ Lsr(out_reg_lo, in_reg_lo, shift_right); 3266 __ add(out_reg_hi, out_reg_hi, ShifterOperand(out_reg_lo)); 3267 __ Lsl(out_reg_lo, in_reg_lo, shift_left); 3268 __ Lsr(shift_left, in_reg_hi, shift_right); 3269 __ add(out_reg_lo, out_reg_lo, ShifterOperand(shift_left)); 3270 __ b(&end); 3271 3272 __ Bind(&shift_by_32_plus_shift_right); // Shift by 32+shift_right. 3273 // out_reg_hi = (reg_hi >> shift_right) | (reg_lo << shift_left). 3274 // out_reg_lo = (reg_lo >> shift_right) | (reg_hi << shift_left). 3275 __ Lsr(out_reg_hi, in_reg_hi, shift_right); 3276 __ Lsl(out_reg_lo, in_reg_lo, shift_left); 3277 __ add(out_reg_hi, out_reg_hi, ShifterOperand(out_reg_lo)); 3278 __ Lsr(out_reg_lo, in_reg_lo, shift_right); 3279 __ Lsl(shift_right, in_reg_hi, shift_left); 3280 __ add(out_reg_lo, out_reg_lo, ShifterOperand(shift_right)); 3281 3282 __ Bind(&end); 3283 } 3284} 3285 3286void LocationsBuilderARM::VisitRor(HRor* ror) { 3287 LocationSummary* locations = 3288 new (GetGraph()->GetArena()) LocationSummary(ror, LocationSummary::kNoCall); 3289 switch (ror->GetResultType()) { 3290 case Primitive::kPrimInt: { 3291 locations->SetInAt(0, Location::RequiresRegister()); 3292 locations->SetInAt(1, Location::RegisterOrConstant(ror->InputAt(1))); 3293 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3294 break; 3295 } 3296 case Primitive::kPrimLong: { 3297 locations->SetInAt(0, Location::RequiresRegister()); 3298 if (ror->InputAt(1)->IsConstant()) { 3299 locations->SetInAt(1, Location::ConstantLocation(ror->InputAt(1)->AsConstant())); 3300 } else { 3301 locations->SetInAt(1, Location::RequiresRegister()); 3302 locations->AddTemp(Location::RequiresRegister()); 3303 locations->AddTemp(Location::RequiresRegister()); 3304 } 3305 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3306 break; 3307 } 3308 default: 3309 LOG(FATAL) << "Unexpected operation type " << ror->GetResultType(); 3310 } 3311} 3312 3313void InstructionCodeGeneratorARM::VisitRor(HRor* ror) { 3314 LocationSummary* locations = ror->GetLocations(); 3315 Primitive::Type type = ror->GetResultType(); 3316 switch (type) { 3317 case Primitive::kPrimInt: { 3318 HandleIntegerRotate(locations); 3319 break; 3320 } 3321 case Primitive::kPrimLong: { 3322 HandleLongRotate(locations); 3323 break; 3324 } 3325 default: 3326 LOG(FATAL) << "Unexpected operation type " << type; 3327 UNREACHABLE(); 3328 } 3329} 3330 3331void LocationsBuilderARM::HandleShift(HBinaryOperation* op) { 3332 DCHECK(op->IsShl() || op->IsShr() || op->IsUShr()); 3333 3334 LocationSummary* locations = 3335 new (GetGraph()->GetArena()) LocationSummary(op, LocationSummary::kNoCall); 3336 3337 switch (op->GetResultType()) { 3338 case Primitive::kPrimInt: { 3339 locations->SetInAt(0, Location::RequiresRegister()); 3340 if (op->InputAt(1)->IsConstant()) { 3341 locations->SetInAt(1, Location::ConstantLocation(op->InputAt(1)->AsConstant())); 3342 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3343 } else { 3344 locations->SetInAt(1, Location::RequiresRegister()); 3345 // Make the output overlap, as it will be used to hold the masked 3346 // second input. 3347 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3348 } 3349 break; 3350 } 3351 case Primitive::kPrimLong: { 3352 locations->SetInAt(0, Location::RequiresRegister()); 3353 if (op->InputAt(1)->IsConstant()) { 3354 locations->SetInAt(1, Location::ConstantLocation(op->InputAt(1)->AsConstant())); 3355 // For simplicity, use kOutputOverlap even though we only require that low registers 3356 // don't clash with high registers which the register allocator currently guarantees. 3357 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3358 } else { 3359 locations->SetInAt(1, Location::RequiresRegister()); 3360 locations->AddTemp(Location::RequiresRegister()); 3361 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3362 } 3363 break; 3364 } 3365 default: 3366 LOG(FATAL) << "Unexpected operation type " << op->GetResultType(); 3367 } 3368} 3369 3370void InstructionCodeGeneratorARM::HandleShift(HBinaryOperation* op) { 3371 DCHECK(op->IsShl() || op->IsShr() || op->IsUShr()); 3372 3373 LocationSummary* locations = op->GetLocations(); 3374 Location out = locations->Out(); 3375 Location first = locations->InAt(0); 3376 Location second = locations->InAt(1); 3377 3378 Primitive::Type type = op->GetResultType(); 3379 switch (type) { 3380 case Primitive::kPrimInt: { 3381 Register out_reg = out.AsRegister<Register>(); 3382 Register first_reg = first.AsRegister<Register>(); 3383 if (second.IsRegister()) { 3384 Register second_reg = second.AsRegister<Register>(); 3385 // ARM doesn't mask the shift count so we need to do it ourselves. 3386 __ and_(out_reg, second_reg, ShifterOperand(kMaxIntShiftDistance)); 3387 if (op->IsShl()) { 3388 __ Lsl(out_reg, first_reg, out_reg); 3389 } else if (op->IsShr()) { 3390 __ Asr(out_reg, first_reg, out_reg); 3391 } else { 3392 __ Lsr(out_reg, first_reg, out_reg); 3393 } 3394 } else { 3395 int32_t cst = second.GetConstant()->AsIntConstant()->GetValue(); 3396 uint32_t shift_value = cst & kMaxIntShiftDistance; 3397 if (shift_value == 0) { // ARM does not support shifting with 0 immediate. 3398 __ Mov(out_reg, first_reg); 3399 } else if (op->IsShl()) { 3400 __ Lsl(out_reg, first_reg, shift_value); 3401 } else if (op->IsShr()) { 3402 __ Asr(out_reg, first_reg, shift_value); 3403 } else { 3404 __ Lsr(out_reg, first_reg, shift_value); 3405 } 3406 } 3407 break; 3408 } 3409 case Primitive::kPrimLong: { 3410 Register o_h = out.AsRegisterPairHigh<Register>(); 3411 Register o_l = out.AsRegisterPairLow<Register>(); 3412 3413 Register high = first.AsRegisterPairHigh<Register>(); 3414 Register low = first.AsRegisterPairLow<Register>(); 3415 3416 if (second.IsRegister()) { 3417 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3418 3419 Register second_reg = second.AsRegister<Register>(); 3420 3421 if (op->IsShl()) { 3422 __ and_(o_l, second_reg, ShifterOperand(kMaxLongShiftDistance)); 3423 // Shift the high part 3424 __ Lsl(o_h, high, o_l); 3425 // Shift the low part and `or` what overflew on the high part 3426 __ rsb(temp, o_l, ShifterOperand(kArmBitsPerWord)); 3427 __ Lsr(temp, low, temp); 3428 __ orr(o_h, o_h, ShifterOperand(temp)); 3429 // If the shift is > 32 bits, override the high part 3430 __ subs(temp, o_l, ShifterOperand(kArmBitsPerWord)); 3431 __ it(PL); 3432 __ Lsl(o_h, low, temp, PL); 3433 // Shift the low part 3434 __ Lsl(o_l, low, o_l); 3435 } else if (op->IsShr()) { 3436 __ and_(o_h, second_reg, ShifterOperand(kMaxLongShiftDistance)); 3437 // Shift the low part 3438 __ Lsr(o_l, low, o_h); 3439 // Shift the high part and `or` what underflew on the low part 3440 __ rsb(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3441 __ Lsl(temp, high, temp); 3442 __ orr(o_l, o_l, ShifterOperand(temp)); 3443 // If the shift is > 32 bits, override the low part 3444 __ subs(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3445 __ it(PL); 3446 __ Asr(o_l, high, temp, PL); 3447 // Shift the high part 3448 __ Asr(o_h, high, o_h); 3449 } else { 3450 __ and_(o_h, second_reg, ShifterOperand(kMaxLongShiftDistance)); 3451 // same as Shr except we use `Lsr`s and not `Asr`s 3452 __ Lsr(o_l, low, o_h); 3453 __ rsb(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3454 __ Lsl(temp, high, temp); 3455 __ orr(o_l, o_l, ShifterOperand(temp)); 3456 __ subs(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3457 __ it(PL); 3458 __ Lsr(o_l, high, temp, PL); 3459 __ Lsr(o_h, high, o_h); 3460 } 3461 } else { 3462 // Register allocator doesn't create partial overlap. 3463 DCHECK_NE(o_l, high); 3464 DCHECK_NE(o_h, low); 3465 int32_t cst = second.GetConstant()->AsIntConstant()->GetValue(); 3466 uint32_t shift_value = cst & kMaxLongShiftDistance; 3467 if (shift_value > 32) { 3468 if (op->IsShl()) { 3469 __ Lsl(o_h, low, shift_value - 32); 3470 __ LoadImmediate(o_l, 0); 3471 } else if (op->IsShr()) { 3472 __ Asr(o_l, high, shift_value - 32); 3473 __ Asr(o_h, high, 31); 3474 } else { 3475 __ Lsr(o_l, high, shift_value - 32); 3476 __ LoadImmediate(o_h, 0); 3477 } 3478 } else if (shift_value == 32) { 3479 if (op->IsShl()) { 3480 __ mov(o_h, ShifterOperand(low)); 3481 __ LoadImmediate(o_l, 0); 3482 } else if (op->IsShr()) { 3483 __ mov(o_l, ShifterOperand(high)); 3484 __ Asr(o_h, high, 31); 3485 } else { 3486 __ mov(o_l, ShifterOperand(high)); 3487 __ LoadImmediate(o_h, 0); 3488 } 3489 } else if (shift_value == 1) { 3490 if (op->IsShl()) { 3491 __ Lsls(o_l, low, 1); 3492 __ adc(o_h, high, ShifterOperand(high)); 3493 } else if (op->IsShr()) { 3494 __ Asrs(o_h, high, 1); 3495 __ Rrx(o_l, low); 3496 } else { 3497 __ Lsrs(o_h, high, 1); 3498 __ Rrx(o_l, low); 3499 } 3500 } else { 3501 DCHECK(2 <= shift_value && shift_value < 32) << shift_value; 3502 if (op->IsShl()) { 3503 __ Lsl(o_h, high, shift_value); 3504 __ orr(o_h, o_h, ShifterOperand(low, LSR, 32 - shift_value)); 3505 __ Lsl(o_l, low, shift_value); 3506 } else if (op->IsShr()) { 3507 __ Lsr(o_l, low, shift_value); 3508 __ orr(o_l, o_l, ShifterOperand(high, LSL, 32 - shift_value)); 3509 __ Asr(o_h, high, shift_value); 3510 } else { 3511 __ Lsr(o_l, low, shift_value); 3512 __ orr(o_l, o_l, ShifterOperand(high, LSL, 32 - shift_value)); 3513 __ Lsr(o_h, high, shift_value); 3514 } 3515 } 3516 } 3517 break; 3518 } 3519 default: 3520 LOG(FATAL) << "Unexpected operation type " << type; 3521 UNREACHABLE(); 3522 } 3523} 3524 3525void LocationsBuilderARM::VisitShl(HShl* shl) { 3526 HandleShift(shl); 3527} 3528 3529void InstructionCodeGeneratorARM::VisitShl(HShl* shl) { 3530 HandleShift(shl); 3531} 3532 3533void LocationsBuilderARM::VisitShr(HShr* shr) { 3534 HandleShift(shr); 3535} 3536 3537void InstructionCodeGeneratorARM::VisitShr(HShr* shr) { 3538 HandleShift(shr); 3539} 3540 3541void LocationsBuilderARM::VisitUShr(HUShr* ushr) { 3542 HandleShift(ushr); 3543} 3544 3545void InstructionCodeGeneratorARM::VisitUShr(HUShr* ushr) { 3546 HandleShift(ushr); 3547} 3548 3549void LocationsBuilderARM::VisitNewInstance(HNewInstance* instruction) { 3550 LocationSummary* locations = 3551 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnMainOnly); 3552 if (instruction->IsStringAlloc()) { 3553 locations->AddTemp(Location::RegisterLocation(kMethodRegisterArgument)); 3554 } else { 3555 InvokeRuntimeCallingConvention calling_convention; 3556 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 3557 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 3558 } 3559 locations->SetOut(Location::RegisterLocation(R0)); 3560} 3561 3562void InstructionCodeGeneratorARM::VisitNewInstance(HNewInstance* instruction) { 3563 // Note: if heap poisoning is enabled, the entry point takes cares 3564 // of poisoning the reference. 3565 if (instruction->IsStringAlloc()) { 3566 // String is allocated through StringFactory. Call NewEmptyString entry point. 3567 Register temp = instruction->GetLocations()->GetTemp(0).AsRegister<Register>(); 3568 MemberOffset code_offset = ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmPointerSize); 3569 __ LoadFromOffset(kLoadWord, temp, TR, QUICK_ENTRY_POINT(pNewEmptyString)); 3570 __ LoadFromOffset(kLoadWord, LR, temp, code_offset.Int32Value()); 3571 __ blx(LR); 3572 codegen_->RecordPcInfo(instruction, instruction->GetDexPc()); 3573 } else { 3574 codegen_->InvokeRuntime(instruction->GetEntrypoint(), 3575 instruction, 3576 instruction->GetDexPc(), 3577 nullptr); 3578 CheckEntrypointTypes<kQuickAllocObjectWithAccessCheck, void*, uint32_t, ArtMethod*>(); 3579 } 3580} 3581 3582void LocationsBuilderARM::VisitNewArray(HNewArray* instruction) { 3583 LocationSummary* locations = 3584 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnMainOnly); 3585 InvokeRuntimeCallingConvention calling_convention; 3586 locations->AddTemp(Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 3587 locations->SetOut(Location::RegisterLocation(R0)); 3588 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 3589 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(2))); 3590} 3591 3592void InstructionCodeGeneratorARM::VisitNewArray(HNewArray* instruction) { 3593 InvokeRuntimeCallingConvention calling_convention; 3594 __ LoadImmediate(calling_convention.GetRegisterAt(0), instruction->GetTypeIndex()); 3595 // Note: if heap poisoning is enabled, the entry point takes cares 3596 // of poisoning the reference. 3597 codegen_->InvokeRuntime(instruction->GetEntrypoint(), 3598 instruction, 3599 instruction->GetDexPc(), 3600 nullptr); 3601 CheckEntrypointTypes<kQuickAllocArrayWithAccessCheck, void*, uint32_t, int32_t, ArtMethod*>(); 3602} 3603 3604void LocationsBuilderARM::VisitParameterValue(HParameterValue* instruction) { 3605 LocationSummary* locations = 3606 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 3607 Location location = parameter_visitor_.GetNextLocation(instruction->GetType()); 3608 if (location.IsStackSlot()) { 3609 location = Location::StackSlot(location.GetStackIndex() + codegen_->GetFrameSize()); 3610 } else if (location.IsDoubleStackSlot()) { 3611 location = Location::DoubleStackSlot(location.GetStackIndex() + codegen_->GetFrameSize()); 3612 } 3613 locations->SetOut(location); 3614} 3615 3616void InstructionCodeGeneratorARM::VisitParameterValue( 3617 HParameterValue* instruction ATTRIBUTE_UNUSED) { 3618 // Nothing to do, the parameter is already at its location. 3619} 3620 3621void LocationsBuilderARM::VisitCurrentMethod(HCurrentMethod* instruction) { 3622 LocationSummary* locations = 3623 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 3624 locations->SetOut(Location::RegisterLocation(kMethodRegisterArgument)); 3625} 3626 3627void InstructionCodeGeneratorARM::VisitCurrentMethod(HCurrentMethod* instruction ATTRIBUTE_UNUSED) { 3628 // Nothing to do, the method is already at its location. 3629} 3630 3631void LocationsBuilderARM::VisitNot(HNot* not_) { 3632 LocationSummary* locations = 3633 new (GetGraph()->GetArena()) LocationSummary(not_, LocationSummary::kNoCall); 3634 locations->SetInAt(0, Location::RequiresRegister()); 3635 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3636} 3637 3638void InstructionCodeGeneratorARM::VisitNot(HNot* not_) { 3639 LocationSummary* locations = not_->GetLocations(); 3640 Location out = locations->Out(); 3641 Location in = locations->InAt(0); 3642 switch (not_->GetResultType()) { 3643 case Primitive::kPrimInt: 3644 __ mvn(out.AsRegister<Register>(), ShifterOperand(in.AsRegister<Register>())); 3645 break; 3646 3647 case Primitive::kPrimLong: 3648 __ mvn(out.AsRegisterPairLow<Register>(), 3649 ShifterOperand(in.AsRegisterPairLow<Register>())); 3650 __ mvn(out.AsRegisterPairHigh<Register>(), 3651 ShifterOperand(in.AsRegisterPairHigh<Register>())); 3652 break; 3653 3654 default: 3655 LOG(FATAL) << "Unimplemented type for not operation " << not_->GetResultType(); 3656 } 3657} 3658 3659void LocationsBuilderARM::VisitBooleanNot(HBooleanNot* bool_not) { 3660 LocationSummary* locations = 3661 new (GetGraph()->GetArena()) LocationSummary(bool_not, LocationSummary::kNoCall); 3662 locations->SetInAt(0, Location::RequiresRegister()); 3663 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3664} 3665 3666void InstructionCodeGeneratorARM::VisitBooleanNot(HBooleanNot* bool_not) { 3667 LocationSummary* locations = bool_not->GetLocations(); 3668 Location out = locations->Out(); 3669 Location in = locations->InAt(0); 3670 __ eor(out.AsRegister<Register>(), in.AsRegister<Register>(), ShifterOperand(1)); 3671} 3672 3673void LocationsBuilderARM::VisitCompare(HCompare* compare) { 3674 LocationSummary* locations = 3675 new (GetGraph()->GetArena()) LocationSummary(compare, LocationSummary::kNoCall); 3676 switch (compare->InputAt(0)->GetType()) { 3677 case Primitive::kPrimBoolean: 3678 case Primitive::kPrimByte: 3679 case Primitive::kPrimShort: 3680 case Primitive::kPrimChar: 3681 case Primitive::kPrimInt: 3682 case Primitive::kPrimLong: { 3683 locations->SetInAt(0, Location::RequiresRegister()); 3684 locations->SetInAt(1, Location::RequiresRegister()); 3685 // Output overlaps because it is written before doing the low comparison. 3686 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3687 break; 3688 } 3689 case Primitive::kPrimFloat: 3690 case Primitive::kPrimDouble: { 3691 locations->SetInAt(0, Location::RequiresFpuRegister()); 3692 locations->SetInAt(1, ArithmeticZeroOrFpuRegister(compare->InputAt(1))); 3693 locations->SetOut(Location::RequiresRegister()); 3694 break; 3695 } 3696 default: 3697 LOG(FATAL) << "Unexpected type for compare operation " << compare->InputAt(0)->GetType(); 3698 } 3699} 3700 3701void InstructionCodeGeneratorARM::VisitCompare(HCompare* compare) { 3702 LocationSummary* locations = compare->GetLocations(); 3703 Register out = locations->Out().AsRegister<Register>(); 3704 Location left = locations->InAt(0); 3705 Location right = locations->InAt(1); 3706 3707 Label less, greater, done; 3708 Primitive::Type type = compare->InputAt(0)->GetType(); 3709 Condition less_cond; 3710 switch (type) { 3711 case Primitive::kPrimBoolean: 3712 case Primitive::kPrimByte: 3713 case Primitive::kPrimShort: 3714 case Primitive::kPrimChar: 3715 case Primitive::kPrimInt: { 3716 __ LoadImmediate(out, 0); 3717 __ cmp(left.AsRegister<Register>(), 3718 ShifterOperand(right.AsRegister<Register>())); // Signed compare. 3719 less_cond = LT; 3720 break; 3721 } 3722 case Primitive::kPrimLong: { 3723 __ cmp(left.AsRegisterPairHigh<Register>(), 3724 ShifterOperand(right.AsRegisterPairHigh<Register>())); // Signed compare. 3725 __ b(&less, LT); 3726 __ b(&greater, GT); 3727 // Do LoadImmediate before the last `cmp`, as LoadImmediate might affect the status flags. 3728 __ LoadImmediate(out, 0); 3729 __ cmp(left.AsRegisterPairLow<Register>(), 3730 ShifterOperand(right.AsRegisterPairLow<Register>())); // Unsigned compare. 3731 less_cond = LO; 3732 break; 3733 } 3734 case Primitive::kPrimFloat: 3735 case Primitive::kPrimDouble: { 3736 __ LoadImmediate(out, 0); 3737 GenerateVcmp(compare); 3738 __ vmstat(); // transfer FP status register to ARM APSR. 3739 less_cond = ARMFPCondition(kCondLT, compare->IsGtBias()); 3740 break; 3741 } 3742 default: 3743 LOG(FATAL) << "Unexpected compare type " << type; 3744 UNREACHABLE(); 3745 } 3746 3747 __ b(&done, EQ); 3748 __ b(&less, less_cond); 3749 3750 __ Bind(&greater); 3751 __ LoadImmediate(out, 1); 3752 __ b(&done); 3753 3754 __ Bind(&less); 3755 __ LoadImmediate(out, -1); 3756 3757 __ Bind(&done); 3758} 3759 3760void LocationsBuilderARM::VisitPhi(HPhi* instruction) { 3761 LocationSummary* locations = 3762 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 3763 for (size_t i = 0, e = locations->GetInputCount(); i < e; ++i) { 3764 locations->SetInAt(i, Location::Any()); 3765 } 3766 locations->SetOut(Location::Any()); 3767} 3768 3769void InstructionCodeGeneratorARM::VisitPhi(HPhi* instruction ATTRIBUTE_UNUSED) { 3770 LOG(FATAL) << "Unreachable"; 3771} 3772 3773void CodeGeneratorARM::GenerateMemoryBarrier(MemBarrierKind kind) { 3774 // TODO (ported from quick): revisit ARM barrier kinds. 3775 DmbOptions flavor = DmbOptions::ISH; // Quiet C++ warnings. 3776 switch (kind) { 3777 case MemBarrierKind::kAnyStore: 3778 case MemBarrierKind::kLoadAny: 3779 case MemBarrierKind::kAnyAny: { 3780 flavor = DmbOptions::ISH; 3781 break; 3782 } 3783 case MemBarrierKind::kStoreStore: { 3784 flavor = DmbOptions::ISHST; 3785 break; 3786 } 3787 default: 3788 LOG(FATAL) << "Unexpected memory barrier " << kind; 3789 } 3790 __ dmb(flavor); 3791} 3792 3793void InstructionCodeGeneratorARM::GenerateWideAtomicLoad(Register addr, 3794 uint32_t offset, 3795 Register out_lo, 3796 Register out_hi) { 3797 if (offset != 0) { 3798 // Ensure `out_lo` is different from `addr`, so that loading 3799 // `offset` into `out_lo` does not clutter `addr`. 3800 DCHECK_NE(out_lo, addr); 3801 __ LoadImmediate(out_lo, offset); 3802 __ add(IP, addr, ShifterOperand(out_lo)); 3803 addr = IP; 3804 } 3805 __ ldrexd(out_lo, out_hi, addr); 3806} 3807 3808void InstructionCodeGeneratorARM::GenerateWideAtomicStore(Register addr, 3809 uint32_t offset, 3810 Register value_lo, 3811 Register value_hi, 3812 Register temp1, 3813 Register temp2, 3814 HInstruction* instruction) { 3815 Label fail; 3816 if (offset != 0) { 3817 __ LoadImmediate(temp1, offset); 3818 __ add(IP, addr, ShifterOperand(temp1)); 3819 addr = IP; 3820 } 3821 __ Bind(&fail); 3822 // We need a load followed by store. (The address used in a STREX instruction must 3823 // be the same as the address in the most recently executed LDREX instruction.) 3824 __ ldrexd(temp1, temp2, addr); 3825 codegen_->MaybeRecordImplicitNullCheck(instruction); 3826 __ strexd(temp1, value_lo, value_hi, addr); 3827 __ CompareAndBranchIfNonZero(temp1, &fail); 3828} 3829 3830void LocationsBuilderARM::HandleFieldSet(HInstruction* instruction, const FieldInfo& field_info) { 3831 DCHECK(instruction->IsInstanceFieldSet() || instruction->IsStaticFieldSet()); 3832 3833 LocationSummary* locations = 3834 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 3835 locations->SetInAt(0, Location::RequiresRegister()); 3836 3837 Primitive::Type field_type = field_info.GetFieldType(); 3838 if (Primitive::IsFloatingPointType(field_type)) { 3839 locations->SetInAt(1, Location::RequiresFpuRegister()); 3840 } else { 3841 locations->SetInAt(1, Location::RequiresRegister()); 3842 } 3843 3844 bool is_wide = field_type == Primitive::kPrimLong || field_type == Primitive::kPrimDouble; 3845 bool generate_volatile = field_info.IsVolatile() 3846 && is_wide 3847 && !codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 3848 bool needs_write_barrier = 3849 CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1)); 3850 // Temporary registers for the write barrier. 3851 // TODO: consider renaming StoreNeedsWriteBarrier to StoreNeedsGCMark. 3852 if (needs_write_barrier) { 3853 locations->AddTemp(Location::RequiresRegister()); // Possibly used for reference poisoning too. 3854 locations->AddTemp(Location::RequiresRegister()); 3855 } else if (generate_volatile) { 3856 // ARM encoding have some additional constraints for ldrexd/strexd: 3857 // - registers need to be consecutive 3858 // - the first register should be even but not R14. 3859 // We don't test for ARM yet, and the assertion makes sure that we 3860 // revisit this if we ever enable ARM encoding. 3861 DCHECK_EQ(InstructionSet::kThumb2, codegen_->GetInstructionSet()); 3862 3863 locations->AddTemp(Location::RequiresRegister()); 3864 locations->AddTemp(Location::RequiresRegister()); 3865 if (field_type == Primitive::kPrimDouble) { 3866 // For doubles we need two more registers to copy the value. 3867 locations->AddTemp(Location::RegisterLocation(R2)); 3868 locations->AddTemp(Location::RegisterLocation(R3)); 3869 } 3870 } 3871} 3872 3873void InstructionCodeGeneratorARM::HandleFieldSet(HInstruction* instruction, 3874 const FieldInfo& field_info, 3875 bool value_can_be_null) { 3876 DCHECK(instruction->IsInstanceFieldSet() || instruction->IsStaticFieldSet()); 3877 3878 LocationSummary* locations = instruction->GetLocations(); 3879 Register base = locations->InAt(0).AsRegister<Register>(); 3880 Location value = locations->InAt(1); 3881 3882 bool is_volatile = field_info.IsVolatile(); 3883 bool atomic_ldrd_strd = codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 3884 Primitive::Type field_type = field_info.GetFieldType(); 3885 uint32_t offset = field_info.GetFieldOffset().Uint32Value(); 3886 bool needs_write_barrier = 3887 CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1)); 3888 3889 if (is_volatile) { 3890 codegen_->GenerateMemoryBarrier(MemBarrierKind::kAnyStore); 3891 } 3892 3893 switch (field_type) { 3894 case Primitive::kPrimBoolean: 3895 case Primitive::kPrimByte: { 3896 __ StoreToOffset(kStoreByte, value.AsRegister<Register>(), base, offset); 3897 break; 3898 } 3899 3900 case Primitive::kPrimShort: 3901 case Primitive::kPrimChar: { 3902 __ StoreToOffset(kStoreHalfword, value.AsRegister<Register>(), base, offset); 3903 break; 3904 } 3905 3906 case Primitive::kPrimInt: 3907 case Primitive::kPrimNot: { 3908 if (kPoisonHeapReferences && needs_write_barrier) { 3909 // Note that in the case where `value` is a null reference, 3910 // we do not enter this block, as a null reference does not 3911 // need poisoning. 3912 DCHECK_EQ(field_type, Primitive::kPrimNot); 3913 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3914 __ Mov(temp, value.AsRegister<Register>()); 3915 __ PoisonHeapReference(temp); 3916 __ StoreToOffset(kStoreWord, temp, base, offset); 3917 } else { 3918 __ StoreToOffset(kStoreWord, value.AsRegister<Register>(), base, offset); 3919 } 3920 break; 3921 } 3922 3923 case Primitive::kPrimLong: { 3924 if (is_volatile && !atomic_ldrd_strd) { 3925 GenerateWideAtomicStore(base, offset, 3926 value.AsRegisterPairLow<Register>(), 3927 value.AsRegisterPairHigh<Register>(), 3928 locations->GetTemp(0).AsRegister<Register>(), 3929 locations->GetTemp(1).AsRegister<Register>(), 3930 instruction); 3931 } else { 3932 __ StoreToOffset(kStoreWordPair, value.AsRegisterPairLow<Register>(), base, offset); 3933 codegen_->MaybeRecordImplicitNullCheck(instruction); 3934 } 3935 break; 3936 } 3937 3938 case Primitive::kPrimFloat: { 3939 __ StoreSToOffset(value.AsFpuRegister<SRegister>(), base, offset); 3940 break; 3941 } 3942 3943 case Primitive::kPrimDouble: { 3944 DRegister value_reg = FromLowSToD(value.AsFpuRegisterPairLow<SRegister>()); 3945 if (is_volatile && !atomic_ldrd_strd) { 3946 Register value_reg_lo = locations->GetTemp(0).AsRegister<Register>(); 3947 Register value_reg_hi = locations->GetTemp(1).AsRegister<Register>(); 3948 3949 __ vmovrrd(value_reg_lo, value_reg_hi, value_reg); 3950 3951 GenerateWideAtomicStore(base, offset, 3952 value_reg_lo, 3953 value_reg_hi, 3954 locations->GetTemp(2).AsRegister<Register>(), 3955 locations->GetTemp(3).AsRegister<Register>(), 3956 instruction); 3957 } else { 3958 __ StoreDToOffset(value_reg, base, offset); 3959 codegen_->MaybeRecordImplicitNullCheck(instruction); 3960 } 3961 break; 3962 } 3963 3964 case Primitive::kPrimVoid: 3965 LOG(FATAL) << "Unreachable type " << field_type; 3966 UNREACHABLE(); 3967 } 3968 3969 // Longs and doubles are handled in the switch. 3970 if (field_type != Primitive::kPrimLong && field_type != Primitive::kPrimDouble) { 3971 codegen_->MaybeRecordImplicitNullCheck(instruction); 3972 } 3973 3974 if (CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1))) { 3975 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3976 Register card = locations->GetTemp(1).AsRegister<Register>(); 3977 codegen_->MarkGCCard( 3978 temp, card, base, value.AsRegister<Register>(), value_can_be_null); 3979 } 3980 3981 if (is_volatile) { 3982 codegen_->GenerateMemoryBarrier(MemBarrierKind::kAnyAny); 3983 } 3984} 3985 3986void LocationsBuilderARM::HandleFieldGet(HInstruction* instruction, const FieldInfo& field_info) { 3987 DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet()); 3988 3989 bool object_field_get_with_read_barrier = 3990 kEmitCompilerReadBarrier && (field_info.GetFieldType() == Primitive::kPrimNot); 3991 LocationSummary* locations = 3992 new (GetGraph()->GetArena()) LocationSummary(instruction, 3993 object_field_get_with_read_barrier ? 3994 LocationSummary::kCallOnSlowPath : 3995 LocationSummary::kNoCall); 3996 locations->SetInAt(0, Location::RequiresRegister()); 3997 3998 bool volatile_for_double = field_info.IsVolatile() 3999 && (field_info.GetFieldType() == Primitive::kPrimDouble) 4000 && !codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 4001 // The output overlaps in case of volatile long: we don't want the 4002 // code generated by GenerateWideAtomicLoad to overwrite the 4003 // object's location. Likewise, in the case of an object field get 4004 // with read barriers enabled, we do not want the load to overwrite 4005 // the object's location, as we need it to emit the read barrier. 4006 bool overlap = (field_info.IsVolatile() && (field_info.GetFieldType() == Primitive::kPrimLong)) || 4007 object_field_get_with_read_barrier; 4008 4009 if (Primitive::IsFloatingPointType(instruction->GetType())) { 4010 locations->SetOut(Location::RequiresFpuRegister()); 4011 } else { 4012 locations->SetOut(Location::RequiresRegister(), 4013 (overlap ? Location::kOutputOverlap : Location::kNoOutputOverlap)); 4014 } 4015 if (volatile_for_double) { 4016 // ARM encoding have some additional constraints for ldrexd/strexd: 4017 // - registers need to be consecutive 4018 // - the first register should be even but not R14. 4019 // We don't test for ARM yet, and the assertion makes sure that we 4020 // revisit this if we ever enable ARM encoding. 4021 DCHECK_EQ(InstructionSet::kThumb2, codegen_->GetInstructionSet()); 4022 locations->AddTemp(Location::RequiresRegister()); 4023 locations->AddTemp(Location::RequiresRegister()); 4024 } else if (object_field_get_with_read_barrier && kUseBakerReadBarrier) { 4025 // We need a temporary register for the read barrier marking slow 4026 // path in CodeGeneratorARM::GenerateFieldLoadWithBakerReadBarrier. 4027 locations->AddTemp(Location::RequiresRegister()); 4028 } 4029} 4030 4031Location LocationsBuilderARM::ArithmeticZeroOrFpuRegister(HInstruction* input) { 4032 DCHECK(input->GetType() == Primitive::kPrimDouble || input->GetType() == Primitive::kPrimFloat) 4033 << input->GetType(); 4034 if ((input->IsFloatConstant() && (input->AsFloatConstant()->IsArithmeticZero())) || 4035 (input->IsDoubleConstant() && (input->AsDoubleConstant()->IsArithmeticZero()))) { 4036 return Location::ConstantLocation(input->AsConstant()); 4037 } else { 4038 return Location::RequiresFpuRegister(); 4039 } 4040} 4041 4042Location LocationsBuilderARM::ArmEncodableConstantOrRegister(HInstruction* constant, 4043 Opcode opcode) { 4044 DCHECK(!Primitive::IsFloatingPointType(constant->GetType())); 4045 if (constant->IsConstant() && 4046 CanEncodeConstantAsImmediate(constant->AsConstant(), opcode)) { 4047 return Location::ConstantLocation(constant->AsConstant()); 4048 } 4049 return Location::RequiresRegister(); 4050} 4051 4052bool LocationsBuilderARM::CanEncodeConstantAsImmediate(HConstant* input_cst, 4053 Opcode opcode) { 4054 uint64_t value = static_cast<uint64_t>(Int64FromConstant(input_cst)); 4055 if (Primitive::Is64BitType(input_cst->GetType())) { 4056 Opcode high_opcode = opcode; 4057 SetCc low_set_cc = kCcDontCare; 4058 switch (opcode) { 4059 case SUB: 4060 // Flip the operation to an ADD. 4061 value = -value; 4062 opcode = ADD; 4063 FALLTHROUGH_INTENDED; 4064 case ADD: 4065 if (Low32Bits(value) == 0u) { 4066 return CanEncodeConstantAsImmediate(High32Bits(value), opcode, kCcDontCare); 4067 } 4068 high_opcode = ADC; 4069 low_set_cc = kCcSet; 4070 break; 4071 default: 4072 break; 4073 } 4074 return CanEncodeConstantAsImmediate(Low32Bits(value), opcode, low_set_cc) && 4075 CanEncodeConstantAsImmediate(High32Bits(value), high_opcode, kCcDontCare); 4076 } else { 4077 return CanEncodeConstantAsImmediate(Low32Bits(value), opcode); 4078 } 4079} 4080 4081bool LocationsBuilderARM::CanEncodeConstantAsImmediate(uint32_t value, 4082 Opcode opcode, 4083 SetCc set_cc) { 4084 ShifterOperand so; 4085 ArmAssembler* assembler = codegen_->GetAssembler(); 4086 if (assembler->ShifterOperandCanHold(kNoRegister, kNoRegister, opcode, value, set_cc, &so)) { 4087 return true; 4088 } 4089 Opcode neg_opcode = kNoOperand; 4090 switch (opcode) { 4091 case AND: neg_opcode = BIC; value = ~value; break; 4092 case ORR: neg_opcode = ORN; value = ~value; break; 4093 case ADD: neg_opcode = SUB; value = -value; break; 4094 case ADC: neg_opcode = SBC; value = ~value; break; 4095 case SUB: neg_opcode = ADD; value = -value; break; 4096 case SBC: neg_opcode = ADC; value = ~value; break; 4097 default: 4098 return false; 4099 } 4100 return assembler->ShifterOperandCanHold(kNoRegister, kNoRegister, neg_opcode, value, set_cc, &so); 4101} 4102 4103void InstructionCodeGeneratorARM::HandleFieldGet(HInstruction* instruction, 4104 const FieldInfo& field_info) { 4105 DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet()); 4106 4107 LocationSummary* locations = instruction->GetLocations(); 4108 Location base_loc = locations->InAt(0); 4109 Register base = base_loc.AsRegister<Register>(); 4110 Location out = locations->Out(); 4111 bool is_volatile = field_info.IsVolatile(); 4112 bool atomic_ldrd_strd = codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 4113 Primitive::Type field_type = field_info.GetFieldType(); 4114 uint32_t offset = field_info.GetFieldOffset().Uint32Value(); 4115 4116 switch (field_type) { 4117 case Primitive::kPrimBoolean: 4118 __ LoadFromOffset(kLoadUnsignedByte, out.AsRegister<Register>(), base, offset); 4119 break; 4120 4121 case Primitive::kPrimByte: 4122 __ LoadFromOffset(kLoadSignedByte, out.AsRegister<Register>(), base, offset); 4123 break; 4124 4125 case Primitive::kPrimShort: 4126 __ LoadFromOffset(kLoadSignedHalfword, out.AsRegister<Register>(), base, offset); 4127 break; 4128 4129 case Primitive::kPrimChar: 4130 __ LoadFromOffset(kLoadUnsignedHalfword, out.AsRegister<Register>(), base, offset); 4131 break; 4132 4133 case Primitive::kPrimInt: 4134 __ LoadFromOffset(kLoadWord, out.AsRegister<Register>(), base, offset); 4135 break; 4136 4137 case Primitive::kPrimNot: { 4138 // /* HeapReference<Object> */ out = *(base + offset) 4139 if (kEmitCompilerReadBarrier && kUseBakerReadBarrier) { 4140 Location temp_loc = locations->GetTemp(0); 4141 // Note that a potential implicit null check is handled in this 4142 // CodeGeneratorARM::GenerateFieldLoadWithBakerReadBarrier call. 4143 codegen_->GenerateFieldLoadWithBakerReadBarrier( 4144 instruction, out, base, offset, temp_loc, /* needs_null_check */ true); 4145 if (is_volatile) { 4146 codegen_->GenerateMemoryBarrier(MemBarrierKind::kLoadAny); 4147 } 4148 } else { 4149 __ LoadFromOffset(kLoadWord, out.AsRegister<Register>(), base, offset); 4150 codegen_->MaybeRecordImplicitNullCheck(instruction); 4151 if (is_volatile) { 4152 codegen_->GenerateMemoryBarrier(MemBarrierKind::kLoadAny); 4153 } 4154 // If read barriers are enabled, emit read barriers other than 4155 // Baker's using a slow path (and also unpoison the loaded 4156 // reference, if heap poisoning is enabled). 4157 codegen_->MaybeGenerateReadBarrierSlow(instruction, out, out, base_loc, offset); 4158 } 4159 break; 4160 } 4161 4162 case Primitive::kPrimLong: 4163 if (is_volatile && !atomic_ldrd_strd) { 4164 GenerateWideAtomicLoad(base, offset, 4165 out.AsRegisterPairLow<Register>(), 4166 out.AsRegisterPairHigh<Register>()); 4167 } else { 4168 __ LoadFromOffset(kLoadWordPair, out.AsRegisterPairLow<Register>(), base, offset); 4169 } 4170 break; 4171 4172 case Primitive::kPrimFloat: 4173 __ LoadSFromOffset(out.AsFpuRegister<SRegister>(), base, offset); 4174 break; 4175 4176 case Primitive::kPrimDouble: { 4177 DRegister out_reg = FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()); 4178 if (is_volatile && !atomic_ldrd_strd) { 4179 Register lo = locations->GetTemp(0).AsRegister<Register>(); 4180 Register hi = locations->GetTemp(1).AsRegister<Register>(); 4181 GenerateWideAtomicLoad(base, offset, lo, hi); 4182 codegen_->MaybeRecordImplicitNullCheck(instruction); 4183 __ vmovdrr(out_reg, lo, hi); 4184 } else { 4185 __ LoadDFromOffset(out_reg, base, offset); 4186 codegen_->MaybeRecordImplicitNullCheck(instruction); 4187 } 4188 break; 4189 } 4190 4191 case Primitive::kPrimVoid: 4192 LOG(FATAL) << "Unreachable type " << field_type; 4193 UNREACHABLE(); 4194 } 4195 4196 if (field_type == Primitive::kPrimNot || field_type == Primitive::kPrimDouble) { 4197 // Potential implicit null checks, in the case of reference or 4198 // double fields, are handled in the previous switch statement. 4199 } else { 4200 codegen_->MaybeRecordImplicitNullCheck(instruction); 4201 } 4202 4203 if (is_volatile) { 4204 if (field_type == Primitive::kPrimNot) { 4205 // Memory barriers, in the case of references, are also handled 4206 // in the previous switch statement. 4207 } else { 4208 codegen_->GenerateMemoryBarrier(MemBarrierKind::kLoadAny); 4209 } 4210 } 4211} 4212 4213void LocationsBuilderARM::VisitInstanceFieldSet(HInstanceFieldSet* instruction) { 4214 HandleFieldSet(instruction, instruction->GetFieldInfo()); 4215} 4216 4217void InstructionCodeGeneratorARM::VisitInstanceFieldSet(HInstanceFieldSet* instruction) { 4218 HandleFieldSet(instruction, instruction->GetFieldInfo(), instruction->GetValueCanBeNull()); 4219} 4220 4221void LocationsBuilderARM::VisitInstanceFieldGet(HInstanceFieldGet* instruction) { 4222 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4223} 4224 4225void InstructionCodeGeneratorARM::VisitInstanceFieldGet(HInstanceFieldGet* instruction) { 4226 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4227} 4228 4229void LocationsBuilderARM::VisitStaticFieldGet(HStaticFieldGet* instruction) { 4230 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4231} 4232 4233void InstructionCodeGeneratorARM::VisitStaticFieldGet(HStaticFieldGet* instruction) { 4234 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4235} 4236 4237void LocationsBuilderARM::VisitStaticFieldSet(HStaticFieldSet* instruction) { 4238 HandleFieldSet(instruction, instruction->GetFieldInfo()); 4239} 4240 4241void InstructionCodeGeneratorARM::VisitStaticFieldSet(HStaticFieldSet* instruction) { 4242 HandleFieldSet(instruction, instruction->GetFieldInfo(), instruction->GetValueCanBeNull()); 4243} 4244 4245void LocationsBuilderARM::VisitUnresolvedInstanceFieldGet( 4246 HUnresolvedInstanceFieldGet* instruction) { 4247 FieldAccessCallingConventionARM calling_convention; 4248 codegen_->CreateUnresolvedFieldLocationSummary( 4249 instruction, instruction->GetFieldType(), calling_convention); 4250} 4251 4252void InstructionCodeGeneratorARM::VisitUnresolvedInstanceFieldGet( 4253 HUnresolvedInstanceFieldGet* instruction) { 4254 FieldAccessCallingConventionARM calling_convention; 4255 codegen_->GenerateUnresolvedFieldAccess(instruction, 4256 instruction->GetFieldType(), 4257 instruction->GetFieldIndex(), 4258 instruction->GetDexPc(), 4259 calling_convention); 4260} 4261 4262void LocationsBuilderARM::VisitUnresolvedInstanceFieldSet( 4263 HUnresolvedInstanceFieldSet* instruction) { 4264 FieldAccessCallingConventionARM calling_convention; 4265 codegen_->CreateUnresolvedFieldLocationSummary( 4266 instruction, instruction->GetFieldType(), calling_convention); 4267} 4268 4269void InstructionCodeGeneratorARM::VisitUnresolvedInstanceFieldSet( 4270 HUnresolvedInstanceFieldSet* instruction) { 4271 FieldAccessCallingConventionARM calling_convention; 4272 codegen_->GenerateUnresolvedFieldAccess(instruction, 4273 instruction->GetFieldType(), 4274 instruction->GetFieldIndex(), 4275 instruction->GetDexPc(), 4276 calling_convention); 4277} 4278 4279void LocationsBuilderARM::VisitUnresolvedStaticFieldGet( 4280 HUnresolvedStaticFieldGet* instruction) { 4281 FieldAccessCallingConventionARM calling_convention; 4282 codegen_->CreateUnresolvedFieldLocationSummary( 4283 instruction, instruction->GetFieldType(), calling_convention); 4284} 4285 4286void InstructionCodeGeneratorARM::VisitUnresolvedStaticFieldGet( 4287 HUnresolvedStaticFieldGet* instruction) { 4288 FieldAccessCallingConventionARM calling_convention; 4289 codegen_->GenerateUnresolvedFieldAccess(instruction, 4290 instruction->GetFieldType(), 4291 instruction->GetFieldIndex(), 4292 instruction->GetDexPc(), 4293 calling_convention); 4294} 4295 4296void LocationsBuilderARM::VisitUnresolvedStaticFieldSet( 4297 HUnresolvedStaticFieldSet* instruction) { 4298 FieldAccessCallingConventionARM calling_convention; 4299 codegen_->CreateUnresolvedFieldLocationSummary( 4300 instruction, instruction->GetFieldType(), calling_convention); 4301} 4302 4303void InstructionCodeGeneratorARM::VisitUnresolvedStaticFieldSet( 4304 HUnresolvedStaticFieldSet* instruction) { 4305 FieldAccessCallingConventionARM calling_convention; 4306 codegen_->GenerateUnresolvedFieldAccess(instruction, 4307 instruction->GetFieldType(), 4308 instruction->GetFieldIndex(), 4309 instruction->GetDexPc(), 4310 calling_convention); 4311} 4312 4313void LocationsBuilderARM::VisitNullCheck(HNullCheck* instruction) { 4314 LocationSummary::CallKind call_kind = instruction->CanThrowIntoCatchBlock() 4315 ? LocationSummary::kCallOnSlowPath 4316 : LocationSummary::kNoCall; 4317 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 4318 locations->SetInAt(0, Location::RequiresRegister()); 4319 if (instruction->HasUses()) { 4320 locations->SetOut(Location::SameAsFirstInput()); 4321 } 4322} 4323 4324void CodeGeneratorARM::GenerateImplicitNullCheck(HNullCheck* instruction) { 4325 if (CanMoveNullCheckToUser(instruction)) { 4326 return; 4327 } 4328 Location obj = instruction->GetLocations()->InAt(0); 4329 4330 __ LoadFromOffset(kLoadWord, IP, obj.AsRegister<Register>(), 0); 4331 RecordPcInfo(instruction, instruction->GetDexPc()); 4332} 4333 4334void CodeGeneratorARM::GenerateExplicitNullCheck(HNullCheck* instruction) { 4335 SlowPathCode* slow_path = new (GetGraph()->GetArena()) NullCheckSlowPathARM(instruction); 4336 AddSlowPath(slow_path); 4337 4338 LocationSummary* locations = instruction->GetLocations(); 4339 Location obj = locations->InAt(0); 4340 4341 __ CompareAndBranchIfZero(obj.AsRegister<Register>(), slow_path->GetEntryLabel()); 4342} 4343 4344void InstructionCodeGeneratorARM::VisitNullCheck(HNullCheck* instruction) { 4345 codegen_->GenerateNullCheck(instruction); 4346} 4347 4348static LoadOperandType GetLoadOperandType(Primitive::Type type) { 4349 switch (type) { 4350 case Primitive::kPrimNot: 4351 return kLoadWord; 4352 case Primitive::kPrimBoolean: 4353 return kLoadUnsignedByte; 4354 case Primitive::kPrimByte: 4355 return kLoadSignedByte; 4356 case Primitive::kPrimChar: 4357 return kLoadUnsignedHalfword; 4358 case Primitive::kPrimShort: 4359 return kLoadSignedHalfword; 4360 case Primitive::kPrimInt: 4361 return kLoadWord; 4362 case Primitive::kPrimLong: 4363 return kLoadWordPair; 4364 case Primitive::kPrimFloat: 4365 return kLoadSWord; 4366 case Primitive::kPrimDouble: 4367 return kLoadDWord; 4368 default: 4369 LOG(FATAL) << "Unreachable type " << type; 4370 UNREACHABLE(); 4371 } 4372} 4373 4374static StoreOperandType GetStoreOperandType(Primitive::Type type) { 4375 switch (type) { 4376 case Primitive::kPrimNot: 4377 return kStoreWord; 4378 case Primitive::kPrimBoolean: 4379 case Primitive::kPrimByte: 4380 return kStoreByte; 4381 case Primitive::kPrimChar: 4382 case Primitive::kPrimShort: 4383 return kStoreHalfword; 4384 case Primitive::kPrimInt: 4385 return kStoreWord; 4386 case Primitive::kPrimLong: 4387 return kStoreWordPair; 4388 case Primitive::kPrimFloat: 4389 return kStoreSWord; 4390 case Primitive::kPrimDouble: 4391 return kStoreDWord; 4392 default: 4393 LOG(FATAL) << "Unreachable type " << type; 4394 UNREACHABLE(); 4395 } 4396} 4397 4398void CodeGeneratorARM::LoadFromShiftedRegOffset(Primitive::Type type, 4399 Location out_loc, 4400 Register base, 4401 Register reg_offset, 4402 Condition cond) { 4403 uint32_t shift_count = Primitive::ComponentSizeShift(type); 4404 Address mem_address(base, reg_offset, Shift::LSL, shift_count); 4405 4406 switch (type) { 4407 case Primitive::kPrimByte: 4408 __ ldrsb(out_loc.AsRegister<Register>(), mem_address, cond); 4409 break; 4410 case Primitive::kPrimBoolean: 4411 __ ldrb(out_loc.AsRegister<Register>(), mem_address, cond); 4412 break; 4413 case Primitive::kPrimShort: 4414 __ ldrsh(out_loc.AsRegister<Register>(), mem_address, cond); 4415 break; 4416 case Primitive::kPrimChar: 4417 __ ldrh(out_loc.AsRegister<Register>(), mem_address, cond); 4418 break; 4419 case Primitive::kPrimNot: 4420 case Primitive::kPrimInt: 4421 __ ldr(out_loc.AsRegister<Register>(), mem_address, cond); 4422 break; 4423 // T32 doesn't support LoadFromShiftedRegOffset mem address mode for these types. 4424 case Primitive::kPrimLong: 4425 case Primitive::kPrimFloat: 4426 case Primitive::kPrimDouble: 4427 default: 4428 LOG(FATAL) << "Unreachable type " << type; 4429 UNREACHABLE(); 4430 } 4431} 4432 4433void CodeGeneratorARM::StoreToShiftedRegOffset(Primitive::Type type, 4434 Location loc, 4435 Register base, 4436 Register reg_offset, 4437 Condition cond) { 4438 uint32_t shift_count = Primitive::ComponentSizeShift(type); 4439 Address mem_address(base, reg_offset, Shift::LSL, shift_count); 4440 4441 switch (type) { 4442 case Primitive::kPrimByte: 4443 case Primitive::kPrimBoolean: 4444 __ strb(loc.AsRegister<Register>(), mem_address, cond); 4445 break; 4446 case Primitive::kPrimShort: 4447 case Primitive::kPrimChar: 4448 __ strh(loc.AsRegister<Register>(), mem_address, cond); 4449 break; 4450 case Primitive::kPrimNot: 4451 case Primitive::kPrimInt: 4452 __ str(loc.AsRegister<Register>(), mem_address, cond); 4453 break; 4454 // T32 doesn't support StoreToShiftedRegOffset mem address mode for these types. 4455 case Primitive::kPrimLong: 4456 case Primitive::kPrimFloat: 4457 case Primitive::kPrimDouble: 4458 default: 4459 LOG(FATAL) << "Unreachable type " << type; 4460 UNREACHABLE(); 4461 } 4462} 4463 4464void LocationsBuilderARM::VisitArrayGet(HArrayGet* instruction) { 4465 bool object_array_get_with_read_barrier = 4466 kEmitCompilerReadBarrier && (instruction->GetType() == Primitive::kPrimNot); 4467 LocationSummary* locations = 4468 new (GetGraph()->GetArena()) LocationSummary(instruction, 4469 object_array_get_with_read_barrier ? 4470 LocationSummary::kCallOnSlowPath : 4471 LocationSummary::kNoCall); 4472 locations->SetInAt(0, Location::RequiresRegister()); 4473 locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1))); 4474 if (Primitive::IsFloatingPointType(instruction->GetType())) { 4475 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 4476 } else { 4477 // The output overlaps in the case of an object array get with 4478 // read barriers enabled: we do not want the move to overwrite the 4479 // array's location, as we need it to emit the read barrier. 4480 locations->SetOut( 4481 Location::RequiresRegister(), 4482 object_array_get_with_read_barrier ? Location::kOutputOverlap : Location::kNoOutputOverlap); 4483 } 4484 // We need a temporary register for the read barrier marking slow 4485 // path in CodeGeneratorARM::GenerateArrayLoadWithBakerReadBarrier. 4486 if (object_array_get_with_read_barrier && kUseBakerReadBarrier) { 4487 locations->AddTemp(Location::RequiresRegister()); 4488 } 4489} 4490 4491void InstructionCodeGeneratorARM::VisitArrayGet(HArrayGet* instruction) { 4492 LocationSummary* locations = instruction->GetLocations(); 4493 Location obj_loc = locations->InAt(0); 4494 Register obj = obj_loc.AsRegister<Register>(); 4495 Location index = locations->InAt(1); 4496 Location out_loc = locations->Out(); 4497 uint32_t data_offset = CodeGenerator::GetArrayDataOffset(instruction); 4498 Primitive::Type type = instruction->GetType(); 4499 HInstruction* array_instr = instruction->GetArray(); 4500 bool has_intermediate_address = array_instr->IsIntermediateAddress(); 4501 // The read barrier instrumentation does not support the HIntermediateAddress instruction yet. 4502 DCHECK(!(has_intermediate_address && kEmitCompilerReadBarrier)); 4503 4504 switch (type) { 4505 case Primitive::kPrimBoolean: 4506 case Primitive::kPrimByte: 4507 case Primitive::kPrimShort: 4508 case Primitive::kPrimChar: 4509 case Primitive::kPrimInt: { 4510 if (index.IsConstant()) { 4511 int32_t const_index = index.GetConstant()->AsIntConstant()->GetValue(); 4512 uint32_t full_offset = data_offset + (const_index << Primitive::ComponentSizeShift(type)); 4513 4514 LoadOperandType load_type = GetLoadOperandType(type); 4515 __ LoadFromOffset(load_type, out_loc.AsRegister<Register>(), obj, full_offset); 4516 } else { 4517 Register temp = IP; 4518 4519 if (has_intermediate_address) { 4520 // We do not need to compute the intermediate address from the array: the 4521 // input instruction has done it already. See the comment in 4522 // `TryExtractArrayAccessAddress()`. 4523 if (kIsDebugBuild) { 4524 HIntermediateAddress* tmp = array_instr->AsIntermediateAddress(); 4525 DCHECK_EQ(tmp->GetOffset()->AsIntConstant()->GetValueAsUint64(), data_offset); 4526 } 4527 temp = obj; 4528 } else { 4529 __ add(temp, obj, ShifterOperand(data_offset)); 4530 } 4531 codegen_->LoadFromShiftedRegOffset(type, out_loc, temp, index.AsRegister<Register>()); 4532 } 4533 break; 4534 } 4535 4536 case Primitive::kPrimNot: { 4537 static_assert( 4538 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 4539 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 4540 // /* HeapReference<Object> */ out = 4541 // *(obj + data_offset + index * sizeof(HeapReference<Object>)) 4542 if (kEmitCompilerReadBarrier && kUseBakerReadBarrier) { 4543 Location temp = locations->GetTemp(0); 4544 // Note that a potential implicit null check is handled in this 4545 // CodeGeneratorARM::GenerateArrayLoadWithBakerReadBarrier call. 4546 codegen_->GenerateArrayLoadWithBakerReadBarrier( 4547 instruction, out_loc, obj, data_offset, index, temp, /* needs_null_check */ true); 4548 } else { 4549 Register out = out_loc.AsRegister<Register>(); 4550 if (index.IsConstant()) { 4551 size_t offset = 4552 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4553 __ LoadFromOffset(kLoadWord, out, obj, offset); 4554 codegen_->MaybeRecordImplicitNullCheck(instruction); 4555 // If read barriers are enabled, emit read barriers other than 4556 // Baker's using a slow path (and also unpoison the loaded 4557 // reference, if heap poisoning is enabled). 4558 codegen_->MaybeGenerateReadBarrierSlow(instruction, out_loc, out_loc, obj_loc, offset); 4559 } else { 4560 Register temp = IP; 4561 4562 if (has_intermediate_address) { 4563 // We do not need to compute the intermediate address from the array: the 4564 // input instruction has done it already. See the comment in 4565 // `TryExtractArrayAccessAddress()`. 4566 if (kIsDebugBuild) { 4567 HIntermediateAddress* tmp = array_instr->AsIntermediateAddress(); 4568 DCHECK_EQ(tmp->GetOffset()->AsIntConstant()->GetValueAsUint64(), data_offset); 4569 } 4570 temp = obj; 4571 } else { 4572 __ add(temp, obj, ShifterOperand(data_offset)); 4573 } 4574 codegen_->LoadFromShiftedRegOffset(type, out_loc, temp, index.AsRegister<Register>()); 4575 4576 codegen_->MaybeRecordImplicitNullCheck(instruction); 4577 // If read barriers are enabled, emit read barriers other than 4578 // Baker's using a slow path (and also unpoison the loaded 4579 // reference, if heap poisoning is enabled). 4580 codegen_->MaybeGenerateReadBarrierSlow( 4581 instruction, out_loc, out_loc, obj_loc, data_offset, index); 4582 } 4583 } 4584 break; 4585 } 4586 4587 case Primitive::kPrimLong: { 4588 if (index.IsConstant()) { 4589 size_t offset = 4590 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 4591 __ LoadFromOffset(kLoadWordPair, out_loc.AsRegisterPairLow<Register>(), obj, offset); 4592 } else { 4593 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 4594 __ LoadFromOffset(kLoadWordPair, out_loc.AsRegisterPairLow<Register>(), IP, data_offset); 4595 } 4596 break; 4597 } 4598 4599 case Primitive::kPrimFloat: { 4600 SRegister out = out_loc.AsFpuRegister<SRegister>(); 4601 if (index.IsConstant()) { 4602 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4603 __ LoadSFromOffset(out, obj, offset); 4604 } else { 4605 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 4606 __ LoadSFromOffset(out, IP, data_offset); 4607 } 4608 break; 4609 } 4610 4611 case Primitive::kPrimDouble: { 4612 SRegister out = out_loc.AsFpuRegisterPairLow<SRegister>(); 4613 if (index.IsConstant()) { 4614 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 4615 __ LoadDFromOffset(FromLowSToD(out), obj, offset); 4616 } else { 4617 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 4618 __ LoadDFromOffset(FromLowSToD(out), IP, data_offset); 4619 } 4620 break; 4621 } 4622 4623 case Primitive::kPrimVoid: 4624 LOG(FATAL) << "Unreachable type " << type; 4625 UNREACHABLE(); 4626 } 4627 4628 if (type == Primitive::kPrimNot) { 4629 // Potential implicit null checks, in the case of reference 4630 // arrays, are handled in the previous switch statement. 4631 } else { 4632 codegen_->MaybeRecordImplicitNullCheck(instruction); 4633 } 4634} 4635 4636void LocationsBuilderARM::VisitArraySet(HArraySet* instruction) { 4637 Primitive::Type value_type = instruction->GetComponentType(); 4638 4639 bool needs_write_barrier = 4640 CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue()); 4641 bool may_need_runtime_call_for_type_check = instruction->NeedsTypeCheck(); 4642 4643 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary( 4644 instruction, 4645 may_need_runtime_call_for_type_check ? 4646 LocationSummary::kCallOnSlowPath : 4647 LocationSummary::kNoCall); 4648 4649 locations->SetInAt(0, Location::RequiresRegister()); 4650 locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1))); 4651 if (Primitive::IsFloatingPointType(value_type)) { 4652 locations->SetInAt(2, Location::RequiresFpuRegister()); 4653 } else { 4654 locations->SetInAt(2, Location::RequiresRegister()); 4655 } 4656 if (needs_write_barrier) { 4657 // Temporary registers for the write barrier. 4658 // These registers may be used for Baker read barriers too. 4659 locations->AddTemp(Location::RequiresRegister()); // Possibly used for ref. poisoning too. 4660 locations->AddTemp(Location::RequiresRegister()); 4661 } 4662} 4663 4664void InstructionCodeGeneratorARM::VisitArraySet(HArraySet* instruction) { 4665 LocationSummary* locations = instruction->GetLocations(); 4666 Location array_loc = locations->InAt(0); 4667 Register array = array_loc.AsRegister<Register>(); 4668 Location index = locations->InAt(1); 4669 Primitive::Type value_type = instruction->GetComponentType(); 4670 bool may_need_runtime_call_for_type_check = instruction->NeedsTypeCheck(); 4671 bool needs_write_barrier = 4672 CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue()); 4673 uint32_t data_offset = 4674 mirror::Array::DataOffset(Primitive::ComponentSize(value_type)).Uint32Value(); 4675 Location value_loc = locations->InAt(2); 4676 HInstruction* array_instr = instruction->GetArray(); 4677 bool has_intermediate_address = array_instr->IsIntermediateAddress(); 4678 // The read barrier instrumentation does not support the HIntermediateAddress instruction yet. 4679 DCHECK(!(has_intermediate_address && kEmitCompilerReadBarrier)); 4680 4681 switch (value_type) { 4682 case Primitive::kPrimBoolean: 4683 case Primitive::kPrimByte: 4684 case Primitive::kPrimShort: 4685 case Primitive::kPrimChar: 4686 case Primitive::kPrimInt: { 4687 if (index.IsConstant()) { 4688 int32_t const_index = index.GetConstant()->AsIntConstant()->GetValue(); 4689 uint32_t full_offset = 4690 data_offset + (const_index << Primitive::ComponentSizeShift(value_type)); 4691 StoreOperandType store_type = GetStoreOperandType(value_type); 4692 __ StoreToOffset(store_type, value_loc.AsRegister<Register>(), array, full_offset); 4693 } else { 4694 Register temp = IP; 4695 4696 if (has_intermediate_address) { 4697 // We do not need to compute the intermediate address from the array: the 4698 // input instruction has done it already. See the comment in 4699 // `TryExtractArrayAccessAddress()`. 4700 if (kIsDebugBuild) { 4701 HIntermediateAddress* tmp = array_instr->AsIntermediateAddress(); 4702 DCHECK(tmp->GetOffset()->AsIntConstant()->GetValueAsUint64() == data_offset); 4703 } 4704 temp = array; 4705 } else { 4706 __ add(temp, array, ShifterOperand(data_offset)); 4707 } 4708 codegen_->StoreToShiftedRegOffset(value_type, 4709 value_loc, 4710 temp, 4711 index.AsRegister<Register>()); 4712 } 4713 break; 4714 } 4715 4716 case Primitive::kPrimNot: { 4717 Register value = value_loc.AsRegister<Register>(); 4718 // TryExtractArrayAccessAddress optimization is never applied for non-primitive ArraySet. 4719 // See the comment in instruction_simplifier_shared.cc. 4720 DCHECK(!has_intermediate_address); 4721 4722 if (instruction->InputAt(2)->IsNullConstant()) { 4723 // Just setting null. 4724 if (index.IsConstant()) { 4725 size_t offset = 4726 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4727 __ StoreToOffset(kStoreWord, value, array, offset); 4728 } else { 4729 DCHECK(index.IsRegister()) << index; 4730 __ add(IP, array, ShifterOperand(data_offset)); 4731 codegen_->StoreToShiftedRegOffset(value_type, 4732 value_loc, 4733 IP, 4734 index.AsRegister<Register>()); 4735 } 4736 codegen_->MaybeRecordImplicitNullCheck(instruction); 4737 DCHECK(!needs_write_barrier); 4738 DCHECK(!may_need_runtime_call_for_type_check); 4739 break; 4740 } 4741 4742 DCHECK(needs_write_barrier); 4743 Location temp1_loc = locations->GetTemp(0); 4744 Register temp1 = temp1_loc.AsRegister<Register>(); 4745 Location temp2_loc = locations->GetTemp(1); 4746 Register temp2 = temp2_loc.AsRegister<Register>(); 4747 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 4748 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 4749 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 4750 Label done; 4751 SlowPathCode* slow_path = nullptr; 4752 4753 if (may_need_runtime_call_for_type_check) { 4754 slow_path = new (GetGraph()->GetArena()) ArraySetSlowPathARM(instruction); 4755 codegen_->AddSlowPath(slow_path); 4756 if (instruction->GetValueCanBeNull()) { 4757 Label non_zero; 4758 __ CompareAndBranchIfNonZero(value, &non_zero); 4759 if (index.IsConstant()) { 4760 size_t offset = 4761 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4762 __ StoreToOffset(kStoreWord, value, array, offset); 4763 } else { 4764 DCHECK(index.IsRegister()) << index; 4765 __ add(IP, array, ShifterOperand(data_offset)); 4766 codegen_->StoreToShiftedRegOffset(value_type, 4767 value_loc, 4768 IP, 4769 index.AsRegister<Register>()); 4770 } 4771 codegen_->MaybeRecordImplicitNullCheck(instruction); 4772 __ b(&done); 4773 __ Bind(&non_zero); 4774 } 4775 4776 if (kEmitCompilerReadBarrier) { 4777 if (!kUseBakerReadBarrier) { 4778 // When (non-Baker) read barriers are enabled, the type 4779 // checking instrumentation requires two read barriers 4780 // generated by CodeGeneratorARM::GenerateReadBarrierSlow: 4781 // 4782 // __ Mov(temp2, temp1); 4783 // // /* HeapReference<Class> */ temp1 = temp1->component_type_ 4784 // __ LoadFromOffset(kLoadWord, temp1, temp1, component_offset); 4785 // codegen_->GenerateReadBarrierSlow( 4786 // instruction, temp1_loc, temp1_loc, temp2_loc, component_offset); 4787 // 4788 // // /* HeapReference<Class> */ temp2 = value->klass_ 4789 // __ LoadFromOffset(kLoadWord, temp2, value, class_offset); 4790 // codegen_->GenerateReadBarrierSlow( 4791 // instruction, temp2_loc, temp2_loc, value_loc, class_offset, temp1_loc); 4792 // 4793 // __ cmp(temp1, ShifterOperand(temp2)); 4794 // 4795 // However, the second read barrier may trash `temp`, as it 4796 // is a temporary register, and as such would not be saved 4797 // along with live registers before calling the runtime (nor 4798 // restored afterwards). So in this case, we bail out and 4799 // delegate the work to the array set slow path. 4800 // 4801 // TODO: Extend the register allocator to support a new 4802 // "(locally) live temp" location so as to avoid always 4803 // going into the slow path when read barriers are enabled? 4804 // 4805 // There is no such problem with Baker read barriers (see below). 4806 __ b(slow_path->GetEntryLabel()); 4807 } else { 4808 Register temp3 = IP; 4809 Location temp3_loc = Location::RegisterLocation(temp3); 4810 4811 // Note: `temp3` (scratch register IP) cannot be used as 4812 // `ref` argument of GenerateFieldLoadWithBakerReadBarrier 4813 // calls below (see ReadBarrierMarkSlowPathARM for more 4814 // details). 4815 4816 // /* HeapReference<Class> */ temp1 = array->klass_ 4817 codegen_->GenerateFieldLoadWithBakerReadBarrier(instruction, 4818 temp1_loc, 4819 array, 4820 class_offset, 4821 temp3_loc, 4822 /* needs_null_check */ true); 4823 4824 // /* HeapReference<Class> */ temp1 = temp1->component_type_ 4825 codegen_->GenerateFieldLoadWithBakerReadBarrier(instruction, 4826 temp1_loc, 4827 temp1, 4828 component_offset, 4829 temp3_loc, 4830 /* needs_null_check */ false); 4831 // Register `temp1` is not trashed by the read barrier 4832 // emitted by GenerateFieldLoadWithBakerReadBarrier below, 4833 // as that method produces a call to a ReadBarrierMarkRegX 4834 // entry point, which saves all potentially live registers, 4835 // including temporaries such a `temp1`. 4836 // /* HeapReference<Class> */ temp2 = value->klass_ 4837 codegen_->GenerateFieldLoadWithBakerReadBarrier(instruction, 4838 temp2_loc, 4839 value, 4840 class_offset, 4841 temp3_loc, 4842 /* needs_null_check */ false); 4843 // If heap poisoning is enabled, `temp1` and `temp2` have 4844 // been unpoisoned by the the previous calls to 4845 // CodeGeneratorARM::GenerateFieldLoadWithBakerReadBarrier. 4846 __ cmp(temp1, ShifterOperand(temp2)); 4847 4848 if (instruction->StaticTypeOfArrayIsObjectArray()) { 4849 Label do_put; 4850 __ b(&do_put, EQ); 4851 // We do not need to emit a read barrier for the 4852 // following heap reference load, as `temp1` is only used 4853 // in a comparison with null below, and this reference 4854 // is not kept afterwards. 4855 // /* HeapReference<Class> */ temp1 = temp1->super_class_ 4856 __ LoadFromOffset(kLoadWord, temp1, temp1, super_offset); 4857 // If heap poisoning is enabled, no need to unpoison 4858 // `temp`, as we are comparing against null below. 4859 __ CompareAndBranchIfNonZero(temp1, slow_path->GetEntryLabel()); 4860 __ Bind(&do_put); 4861 } else { 4862 __ b(slow_path->GetEntryLabel(), NE); 4863 } 4864 } 4865 } else { 4866 // Non read barrier code. 4867 4868 // /* HeapReference<Class> */ temp1 = array->klass_ 4869 __ LoadFromOffset(kLoadWord, temp1, array, class_offset); 4870 codegen_->MaybeRecordImplicitNullCheck(instruction); 4871 __ MaybeUnpoisonHeapReference(temp1); 4872 4873 // /* HeapReference<Class> */ temp1 = temp1->component_type_ 4874 __ LoadFromOffset(kLoadWord, temp1, temp1, component_offset); 4875 // /* HeapReference<Class> */ temp2 = value->klass_ 4876 __ LoadFromOffset(kLoadWord, temp2, value, class_offset); 4877 // If heap poisoning is enabled, no need to unpoison `temp1` 4878 // nor `temp2`, as we are comparing two poisoned references. 4879 __ cmp(temp1, ShifterOperand(temp2)); 4880 4881 if (instruction->StaticTypeOfArrayIsObjectArray()) { 4882 Label do_put; 4883 __ b(&do_put, EQ); 4884 // If heap poisoning is enabled, the `temp1` reference has 4885 // not been unpoisoned yet; unpoison it now. 4886 __ MaybeUnpoisonHeapReference(temp1); 4887 4888 // /* HeapReference<Class> */ temp1 = temp1->super_class_ 4889 __ LoadFromOffset(kLoadWord, temp1, temp1, super_offset); 4890 // If heap poisoning is enabled, no need to unpoison 4891 // `temp1`, as we are comparing against null below. 4892 __ CompareAndBranchIfNonZero(temp1, slow_path->GetEntryLabel()); 4893 __ Bind(&do_put); 4894 } else { 4895 __ b(slow_path->GetEntryLabel(), NE); 4896 } 4897 } 4898 } 4899 4900 Register source = value; 4901 if (kPoisonHeapReferences) { 4902 // Note that in the case where `value` is a null reference, 4903 // we do not enter this block, as a null reference does not 4904 // need poisoning. 4905 DCHECK_EQ(value_type, Primitive::kPrimNot); 4906 __ Mov(temp1, value); 4907 __ PoisonHeapReference(temp1); 4908 source = temp1; 4909 } 4910 4911 if (index.IsConstant()) { 4912 size_t offset = 4913 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4914 __ StoreToOffset(kStoreWord, source, array, offset); 4915 } else { 4916 DCHECK(index.IsRegister()) << index; 4917 4918 __ add(IP, array, ShifterOperand(data_offset)); 4919 codegen_->StoreToShiftedRegOffset(value_type, 4920 Location::RegisterLocation(source), 4921 IP, 4922 index.AsRegister<Register>()); 4923 } 4924 4925 if (!may_need_runtime_call_for_type_check) { 4926 codegen_->MaybeRecordImplicitNullCheck(instruction); 4927 } 4928 4929 codegen_->MarkGCCard(temp1, temp2, array, value, instruction->GetValueCanBeNull()); 4930 4931 if (done.IsLinked()) { 4932 __ Bind(&done); 4933 } 4934 4935 if (slow_path != nullptr) { 4936 __ Bind(slow_path->GetExitLabel()); 4937 } 4938 4939 break; 4940 } 4941 4942 case Primitive::kPrimLong: { 4943 Location value = locations->InAt(2); 4944 if (index.IsConstant()) { 4945 size_t offset = 4946 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 4947 __ StoreToOffset(kStoreWordPair, value.AsRegisterPairLow<Register>(), array, offset); 4948 } else { 4949 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 4950 __ StoreToOffset(kStoreWordPair, value.AsRegisterPairLow<Register>(), IP, data_offset); 4951 } 4952 break; 4953 } 4954 4955 case Primitive::kPrimFloat: { 4956 Location value = locations->InAt(2); 4957 DCHECK(value.IsFpuRegister()); 4958 if (index.IsConstant()) { 4959 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4960 __ StoreSToOffset(value.AsFpuRegister<SRegister>(), array, offset); 4961 } else { 4962 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 4963 __ StoreSToOffset(value.AsFpuRegister<SRegister>(), IP, data_offset); 4964 } 4965 break; 4966 } 4967 4968 case Primitive::kPrimDouble: { 4969 Location value = locations->InAt(2); 4970 DCHECK(value.IsFpuRegisterPair()); 4971 if (index.IsConstant()) { 4972 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 4973 __ StoreDToOffset(FromLowSToD(value.AsFpuRegisterPairLow<SRegister>()), array, offset); 4974 } else { 4975 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 4976 __ StoreDToOffset(FromLowSToD(value.AsFpuRegisterPairLow<SRegister>()), IP, data_offset); 4977 } 4978 4979 break; 4980 } 4981 4982 case Primitive::kPrimVoid: 4983 LOG(FATAL) << "Unreachable type " << value_type; 4984 UNREACHABLE(); 4985 } 4986 4987 // Objects are handled in the switch. 4988 if (value_type != Primitive::kPrimNot) { 4989 codegen_->MaybeRecordImplicitNullCheck(instruction); 4990 } 4991} 4992 4993void LocationsBuilderARM::VisitArrayLength(HArrayLength* instruction) { 4994 LocationSummary* locations = 4995 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 4996 locations->SetInAt(0, Location::RequiresRegister()); 4997 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 4998} 4999 5000void InstructionCodeGeneratorARM::VisitArrayLength(HArrayLength* instruction) { 5001 LocationSummary* locations = instruction->GetLocations(); 5002 uint32_t offset = CodeGenerator::GetArrayLengthOffset(instruction); 5003 Register obj = locations->InAt(0).AsRegister<Register>(); 5004 Register out = locations->Out().AsRegister<Register>(); 5005 __ LoadFromOffset(kLoadWord, out, obj, offset); 5006 codegen_->MaybeRecordImplicitNullCheck(instruction); 5007} 5008 5009void LocationsBuilderARM::VisitIntermediateAddress(HIntermediateAddress* instruction) { 5010 // The read barrier instrumentation does not support the HIntermediateAddress instruction yet. 5011 DCHECK(!kEmitCompilerReadBarrier); 5012 LocationSummary* locations = 5013 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 5014 5015 locations->SetInAt(0, Location::RequiresRegister()); 5016 locations->SetInAt(1, Location::RegisterOrConstant(instruction->GetOffset())); 5017 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 5018} 5019 5020void InstructionCodeGeneratorARM::VisitIntermediateAddress(HIntermediateAddress* instruction) { 5021 LocationSummary* locations = instruction->GetLocations(); 5022 Location out = locations->Out(); 5023 Location first = locations->InAt(0); 5024 Location second = locations->InAt(1); 5025 5026 // The read barrier instrumentation does not support the HIntermediateAddress instruction yet. 5027 DCHECK(!kEmitCompilerReadBarrier); 5028 5029 if (second.IsRegister()) { 5030 __ add(out.AsRegister<Register>(), 5031 first.AsRegister<Register>(), 5032 ShifterOperand(second.AsRegister<Register>())); 5033 } else { 5034 __ AddConstant(out.AsRegister<Register>(), 5035 first.AsRegister<Register>(), 5036 second.GetConstant()->AsIntConstant()->GetValue()); 5037 } 5038} 5039 5040void LocationsBuilderARM::VisitBoundsCheck(HBoundsCheck* instruction) { 5041 LocationSummary::CallKind call_kind = instruction->CanThrowIntoCatchBlock() 5042 ? LocationSummary::kCallOnSlowPath 5043 : LocationSummary::kNoCall; 5044 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 5045 locations->SetInAt(0, Location::RequiresRegister()); 5046 locations->SetInAt(1, Location::RequiresRegister()); 5047 if (instruction->HasUses()) { 5048 locations->SetOut(Location::SameAsFirstInput()); 5049 } 5050} 5051 5052void InstructionCodeGeneratorARM::VisitBoundsCheck(HBoundsCheck* instruction) { 5053 LocationSummary* locations = instruction->GetLocations(); 5054 SlowPathCode* slow_path = 5055 new (GetGraph()->GetArena()) BoundsCheckSlowPathARM(instruction); 5056 codegen_->AddSlowPath(slow_path); 5057 5058 Register index = locations->InAt(0).AsRegister<Register>(); 5059 Register length = locations->InAt(1).AsRegister<Register>(); 5060 5061 __ cmp(index, ShifterOperand(length)); 5062 __ b(slow_path->GetEntryLabel(), HS); 5063} 5064 5065void CodeGeneratorARM::MarkGCCard(Register temp, 5066 Register card, 5067 Register object, 5068 Register value, 5069 bool can_be_null) { 5070 Label is_null; 5071 if (can_be_null) { 5072 __ CompareAndBranchIfZero(value, &is_null); 5073 } 5074 __ LoadFromOffset(kLoadWord, card, TR, Thread::CardTableOffset<kArmPointerSize>().Int32Value()); 5075 __ Lsr(temp, object, gc::accounting::CardTable::kCardShift); 5076 __ strb(card, Address(card, temp)); 5077 if (can_be_null) { 5078 __ Bind(&is_null); 5079 } 5080} 5081 5082void LocationsBuilderARM::VisitParallelMove(HParallelMove* instruction ATTRIBUTE_UNUSED) { 5083 LOG(FATAL) << "Unreachable"; 5084} 5085 5086void InstructionCodeGeneratorARM::VisitParallelMove(HParallelMove* instruction) { 5087 codegen_->GetMoveResolver()->EmitNativeCode(instruction); 5088} 5089 5090void LocationsBuilderARM::VisitSuspendCheck(HSuspendCheck* instruction) { 5091 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnSlowPath); 5092} 5093 5094void InstructionCodeGeneratorARM::VisitSuspendCheck(HSuspendCheck* instruction) { 5095 HBasicBlock* block = instruction->GetBlock(); 5096 if (block->GetLoopInformation() != nullptr) { 5097 DCHECK(block->GetLoopInformation()->GetSuspendCheck() == instruction); 5098 // The back edge will generate the suspend check. 5099 return; 5100 } 5101 if (block->IsEntryBlock() && instruction->GetNext()->IsGoto()) { 5102 // The goto will generate the suspend check. 5103 return; 5104 } 5105 GenerateSuspendCheck(instruction, nullptr); 5106} 5107 5108void InstructionCodeGeneratorARM::GenerateSuspendCheck(HSuspendCheck* instruction, 5109 HBasicBlock* successor) { 5110 SuspendCheckSlowPathARM* slow_path = 5111 down_cast<SuspendCheckSlowPathARM*>(instruction->GetSlowPath()); 5112 if (slow_path == nullptr) { 5113 slow_path = new (GetGraph()->GetArena()) SuspendCheckSlowPathARM(instruction, successor); 5114 instruction->SetSlowPath(slow_path); 5115 codegen_->AddSlowPath(slow_path); 5116 if (successor != nullptr) { 5117 DCHECK(successor->IsLoopHeader()); 5118 codegen_->ClearSpillSlotsFromLoopPhisInStackMap(instruction); 5119 } 5120 } else { 5121 DCHECK_EQ(slow_path->GetSuccessor(), successor); 5122 } 5123 5124 __ LoadFromOffset( 5125 kLoadUnsignedHalfword, IP, TR, Thread::ThreadFlagsOffset<kArmPointerSize>().Int32Value()); 5126 if (successor == nullptr) { 5127 __ CompareAndBranchIfNonZero(IP, slow_path->GetEntryLabel()); 5128 __ Bind(slow_path->GetReturnLabel()); 5129 } else { 5130 __ CompareAndBranchIfZero(IP, codegen_->GetLabelOf(successor)); 5131 __ b(slow_path->GetEntryLabel()); 5132 } 5133} 5134 5135ArmAssembler* ParallelMoveResolverARM::GetAssembler() const { 5136 return codegen_->GetAssembler(); 5137} 5138 5139void ParallelMoveResolverARM::EmitMove(size_t index) { 5140 MoveOperands* move = moves_[index]; 5141 Location source = move->GetSource(); 5142 Location destination = move->GetDestination(); 5143 5144 if (source.IsRegister()) { 5145 if (destination.IsRegister()) { 5146 __ Mov(destination.AsRegister<Register>(), source.AsRegister<Register>()); 5147 } else if (destination.IsFpuRegister()) { 5148 __ vmovsr(destination.AsFpuRegister<SRegister>(), source.AsRegister<Register>()); 5149 } else { 5150 DCHECK(destination.IsStackSlot()); 5151 __ StoreToOffset(kStoreWord, source.AsRegister<Register>(), 5152 SP, destination.GetStackIndex()); 5153 } 5154 } else if (source.IsStackSlot()) { 5155 if (destination.IsRegister()) { 5156 __ LoadFromOffset(kLoadWord, destination.AsRegister<Register>(), 5157 SP, source.GetStackIndex()); 5158 } else if (destination.IsFpuRegister()) { 5159 __ LoadSFromOffset(destination.AsFpuRegister<SRegister>(), SP, source.GetStackIndex()); 5160 } else { 5161 DCHECK(destination.IsStackSlot()); 5162 __ LoadFromOffset(kLoadWord, IP, SP, source.GetStackIndex()); 5163 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 5164 } 5165 } else if (source.IsFpuRegister()) { 5166 if (destination.IsRegister()) { 5167 __ vmovrs(destination.AsRegister<Register>(), source.AsFpuRegister<SRegister>()); 5168 } else if (destination.IsFpuRegister()) { 5169 __ vmovs(destination.AsFpuRegister<SRegister>(), source.AsFpuRegister<SRegister>()); 5170 } else { 5171 DCHECK(destination.IsStackSlot()); 5172 __ StoreSToOffset(source.AsFpuRegister<SRegister>(), SP, destination.GetStackIndex()); 5173 } 5174 } else if (source.IsDoubleStackSlot()) { 5175 if (destination.IsDoubleStackSlot()) { 5176 __ LoadDFromOffset(DTMP, SP, source.GetStackIndex()); 5177 __ StoreDToOffset(DTMP, SP, destination.GetStackIndex()); 5178 } else if (destination.IsRegisterPair()) { 5179 DCHECK(ExpectedPairLayout(destination)); 5180 __ LoadFromOffset( 5181 kLoadWordPair, destination.AsRegisterPairLow<Register>(), SP, source.GetStackIndex()); 5182 } else { 5183 DCHECK(destination.IsFpuRegisterPair()) << destination; 5184 __ LoadDFromOffset(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 5185 SP, 5186 source.GetStackIndex()); 5187 } 5188 } else if (source.IsRegisterPair()) { 5189 if (destination.IsRegisterPair()) { 5190 __ Mov(destination.AsRegisterPairLow<Register>(), source.AsRegisterPairLow<Register>()); 5191 __ Mov(destination.AsRegisterPairHigh<Register>(), source.AsRegisterPairHigh<Register>()); 5192 } else if (destination.IsFpuRegisterPair()) { 5193 __ vmovdrr(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 5194 source.AsRegisterPairLow<Register>(), 5195 source.AsRegisterPairHigh<Register>()); 5196 } else { 5197 DCHECK(destination.IsDoubleStackSlot()) << destination; 5198 DCHECK(ExpectedPairLayout(source)); 5199 __ StoreToOffset( 5200 kStoreWordPair, source.AsRegisterPairLow<Register>(), SP, destination.GetStackIndex()); 5201 } 5202 } else if (source.IsFpuRegisterPair()) { 5203 if (destination.IsRegisterPair()) { 5204 __ vmovrrd(destination.AsRegisterPairLow<Register>(), 5205 destination.AsRegisterPairHigh<Register>(), 5206 FromLowSToD(source.AsFpuRegisterPairLow<SRegister>())); 5207 } else if (destination.IsFpuRegisterPair()) { 5208 __ vmovd(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 5209 FromLowSToD(source.AsFpuRegisterPairLow<SRegister>())); 5210 } else { 5211 DCHECK(destination.IsDoubleStackSlot()) << destination; 5212 __ StoreDToOffset(FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()), 5213 SP, 5214 destination.GetStackIndex()); 5215 } 5216 } else { 5217 DCHECK(source.IsConstant()) << source; 5218 HConstant* constant = source.GetConstant(); 5219 if (constant->IsIntConstant() || constant->IsNullConstant()) { 5220 int32_t value = CodeGenerator::GetInt32ValueOf(constant); 5221 if (destination.IsRegister()) { 5222 __ LoadImmediate(destination.AsRegister<Register>(), value); 5223 } else { 5224 DCHECK(destination.IsStackSlot()); 5225 __ LoadImmediate(IP, value); 5226 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 5227 } 5228 } else if (constant->IsLongConstant()) { 5229 int64_t value = constant->AsLongConstant()->GetValue(); 5230 if (destination.IsRegisterPair()) { 5231 __ LoadImmediate(destination.AsRegisterPairLow<Register>(), Low32Bits(value)); 5232 __ LoadImmediate(destination.AsRegisterPairHigh<Register>(), High32Bits(value)); 5233 } else { 5234 DCHECK(destination.IsDoubleStackSlot()) << destination; 5235 __ LoadImmediate(IP, Low32Bits(value)); 5236 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 5237 __ LoadImmediate(IP, High32Bits(value)); 5238 __ StoreToOffset(kStoreWord, IP, SP, destination.GetHighStackIndex(kArmWordSize)); 5239 } 5240 } else if (constant->IsDoubleConstant()) { 5241 double value = constant->AsDoubleConstant()->GetValue(); 5242 if (destination.IsFpuRegisterPair()) { 5243 __ LoadDImmediate(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), value); 5244 } else { 5245 DCHECK(destination.IsDoubleStackSlot()) << destination; 5246 uint64_t int_value = bit_cast<uint64_t, double>(value); 5247 __ LoadImmediate(IP, Low32Bits(int_value)); 5248 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 5249 __ LoadImmediate(IP, High32Bits(int_value)); 5250 __ StoreToOffset(kStoreWord, IP, SP, destination.GetHighStackIndex(kArmWordSize)); 5251 } 5252 } else { 5253 DCHECK(constant->IsFloatConstant()) << constant->DebugName(); 5254 float value = constant->AsFloatConstant()->GetValue(); 5255 if (destination.IsFpuRegister()) { 5256 __ LoadSImmediate(destination.AsFpuRegister<SRegister>(), value); 5257 } else { 5258 DCHECK(destination.IsStackSlot()); 5259 __ LoadImmediate(IP, bit_cast<int32_t, float>(value)); 5260 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 5261 } 5262 } 5263 } 5264} 5265 5266void ParallelMoveResolverARM::Exchange(Register reg, int mem) { 5267 __ Mov(IP, reg); 5268 __ LoadFromOffset(kLoadWord, reg, SP, mem); 5269 __ StoreToOffset(kStoreWord, IP, SP, mem); 5270} 5271 5272void ParallelMoveResolverARM::Exchange(int mem1, int mem2) { 5273 ScratchRegisterScope ensure_scratch(this, IP, R0, codegen_->GetNumberOfCoreRegisters()); 5274 int stack_offset = ensure_scratch.IsSpilled() ? kArmWordSize : 0; 5275 __ LoadFromOffset(kLoadWord, static_cast<Register>(ensure_scratch.GetRegister()), 5276 SP, mem1 + stack_offset); 5277 __ LoadFromOffset(kLoadWord, IP, SP, mem2 + stack_offset); 5278 __ StoreToOffset(kStoreWord, static_cast<Register>(ensure_scratch.GetRegister()), 5279 SP, mem2 + stack_offset); 5280 __ StoreToOffset(kStoreWord, IP, SP, mem1 + stack_offset); 5281} 5282 5283void ParallelMoveResolverARM::EmitSwap(size_t index) { 5284 MoveOperands* move = moves_[index]; 5285 Location source = move->GetSource(); 5286 Location destination = move->GetDestination(); 5287 5288 if (source.IsRegister() && destination.IsRegister()) { 5289 DCHECK_NE(source.AsRegister<Register>(), IP); 5290 DCHECK_NE(destination.AsRegister<Register>(), IP); 5291 __ Mov(IP, source.AsRegister<Register>()); 5292 __ Mov(source.AsRegister<Register>(), destination.AsRegister<Register>()); 5293 __ Mov(destination.AsRegister<Register>(), IP); 5294 } else if (source.IsRegister() && destination.IsStackSlot()) { 5295 Exchange(source.AsRegister<Register>(), destination.GetStackIndex()); 5296 } else if (source.IsStackSlot() && destination.IsRegister()) { 5297 Exchange(destination.AsRegister<Register>(), source.GetStackIndex()); 5298 } else if (source.IsStackSlot() && destination.IsStackSlot()) { 5299 Exchange(source.GetStackIndex(), destination.GetStackIndex()); 5300 } else if (source.IsFpuRegister() && destination.IsFpuRegister()) { 5301 __ vmovrs(IP, source.AsFpuRegister<SRegister>()); 5302 __ vmovs(source.AsFpuRegister<SRegister>(), destination.AsFpuRegister<SRegister>()); 5303 __ vmovsr(destination.AsFpuRegister<SRegister>(), IP); 5304 } else if (source.IsRegisterPair() && destination.IsRegisterPair()) { 5305 __ vmovdrr(DTMP, source.AsRegisterPairLow<Register>(), source.AsRegisterPairHigh<Register>()); 5306 __ Mov(source.AsRegisterPairLow<Register>(), destination.AsRegisterPairLow<Register>()); 5307 __ Mov(source.AsRegisterPairHigh<Register>(), destination.AsRegisterPairHigh<Register>()); 5308 __ vmovrrd(destination.AsRegisterPairLow<Register>(), 5309 destination.AsRegisterPairHigh<Register>(), 5310 DTMP); 5311 } else if (source.IsRegisterPair() || destination.IsRegisterPair()) { 5312 Register low_reg = source.IsRegisterPair() 5313 ? source.AsRegisterPairLow<Register>() 5314 : destination.AsRegisterPairLow<Register>(); 5315 int mem = source.IsRegisterPair() 5316 ? destination.GetStackIndex() 5317 : source.GetStackIndex(); 5318 DCHECK(ExpectedPairLayout(source.IsRegisterPair() ? source : destination)); 5319 __ vmovdrr(DTMP, low_reg, static_cast<Register>(low_reg + 1)); 5320 __ LoadFromOffset(kLoadWordPair, low_reg, SP, mem); 5321 __ StoreDToOffset(DTMP, SP, mem); 5322 } else if (source.IsFpuRegisterPair() && destination.IsFpuRegisterPair()) { 5323 DRegister first = FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()); 5324 DRegister second = FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()); 5325 __ vmovd(DTMP, first); 5326 __ vmovd(first, second); 5327 __ vmovd(second, DTMP); 5328 } else if (source.IsFpuRegisterPair() || destination.IsFpuRegisterPair()) { 5329 DRegister reg = source.IsFpuRegisterPair() 5330 ? FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()) 5331 : FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()); 5332 int mem = source.IsFpuRegisterPair() 5333 ? destination.GetStackIndex() 5334 : source.GetStackIndex(); 5335 __ vmovd(DTMP, reg); 5336 __ LoadDFromOffset(reg, SP, mem); 5337 __ StoreDToOffset(DTMP, SP, mem); 5338 } else if (source.IsFpuRegister() || destination.IsFpuRegister()) { 5339 SRegister reg = source.IsFpuRegister() ? source.AsFpuRegister<SRegister>() 5340 : destination.AsFpuRegister<SRegister>(); 5341 int mem = source.IsFpuRegister() 5342 ? destination.GetStackIndex() 5343 : source.GetStackIndex(); 5344 5345 __ vmovrs(IP, reg); 5346 __ LoadSFromOffset(reg, SP, mem); 5347 __ StoreToOffset(kStoreWord, IP, SP, mem); 5348 } else if (source.IsDoubleStackSlot() && destination.IsDoubleStackSlot()) { 5349 Exchange(source.GetStackIndex(), destination.GetStackIndex()); 5350 Exchange(source.GetHighStackIndex(kArmWordSize), destination.GetHighStackIndex(kArmWordSize)); 5351 } else { 5352 LOG(FATAL) << "Unimplemented" << source << " <-> " << destination; 5353 } 5354} 5355 5356void ParallelMoveResolverARM::SpillScratch(int reg) { 5357 __ Push(static_cast<Register>(reg)); 5358} 5359 5360void ParallelMoveResolverARM::RestoreScratch(int reg) { 5361 __ Pop(static_cast<Register>(reg)); 5362} 5363 5364HLoadClass::LoadKind CodeGeneratorARM::GetSupportedLoadClassKind( 5365 HLoadClass::LoadKind desired_class_load_kind) { 5366 if (kEmitCompilerReadBarrier) { 5367 switch (desired_class_load_kind) { 5368 case HLoadClass::LoadKind::kBootImageLinkTimeAddress: 5369 case HLoadClass::LoadKind::kBootImageLinkTimePcRelative: 5370 case HLoadClass::LoadKind::kBootImageAddress: 5371 // TODO: Implement for read barrier. 5372 return HLoadClass::LoadKind::kDexCacheViaMethod; 5373 default: 5374 break; 5375 } 5376 } 5377 switch (desired_class_load_kind) { 5378 case HLoadClass::LoadKind::kReferrersClass: 5379 break; 5380 case HLoadClass::LoadKind::kBootImageLinkTimeAddress: 5381 DCHECK(!GetCompilerOptions().GetCompilePic()); 5382 break; 5383 case HLoadClass::LoadKind::kBootImageLinkTimePcRelative: 5384 DCHECK(GetCompilerOptions().GetCompilePic()); 5385 break; 5386 case HLoadClass::LoadKind::kBootImageAddress: 5387 break; 5388 case HLoadClass::LoadKind::kDexCacheAddress: 5389 DCHECK(Runtime::Current()->UseJitCompilation()); 5390 break; 5391 case HLoadClass::LoadKind::kDexCachePcRelative: 5392 DCHECK(!Runtime::Current()->UseJitCompilation()); 5393 // We disable pc-relative load when there is an irreducible loop, as the optimization 5394 // is incompatible with it. 5395 // TODO: Create as many ArmDexCacheArraysBase instructions as needed for methods 5396 // with irreducible loops. 5397 if (GetGraph()->HasIrreducibleLoops()) { 5398 return HLoadClass::LoadKind::kDexCacheViaMethod; 5399 } 5400 break; 5401 case HLoadClass::LoadKind::kDexCacheViaMethod: 5402 break; 5403 } 5404 return desired_class_load_kind; 5405} 5406 5407void LocationsBuilderARM::VisitLoadClass(HLoadClass* cls) { 5408 if (cls->NeedsAccessCheck()) { 5409 InvokeRuntimeCallingConvention calling_convention; 5410 CodeGenerator::CreateLoadClassLocationSummary( 5411 cls, 5412 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 5413 Location::RegisterLocation(R0), 5414 /* code_generator_supports_read_barrier */ true); 5415 return; 5416 } 5417 5418 LocationSummary::CallKind call_kind = (cls->NeedsEnvironment() || kEmitCompilerReadBarrier) 5419 ? LocationSummary::kCallOnSlowPath 5420 : LocationSummary::kNoCall; 5421 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(cls, call_kind); 5422 HLoadClass::LoadKind load_kind = cls->GetLoadKind(); 5423 if (load_kind == HLoadClass::LoadKind::kReferrersClass || 5424 load_kind == HLoadClass::LoadKind::kDexCacheViaMethod || 5425 load_kind == HLoadClass::LoadKind::kDexCachePcRelative) { 5426 locations->SetInAt(0, Location::RequiresRegister()); 5427 } 5428 locations->SetOut(Location::RequiresRegister()); 5429} 5430 5431void InstructionCodeGeneratorARM::VisitLoadClass(HLoadClass* cls) { 5432 LocationSummary* locations = cls->GetLocations(); 5433 if (cls->NeedsAccessCheck()) { 5434 codegen_->MoveConstant(locations->GetTemp(0), cls->GetTypeIndex()); 5435 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pInitializeTypeAndVerifyAccess), 5436 cls, 5437 cls->GetDexPc(), 5438 nullptr); 5439 CheckEntrypointTypes<kQuickInitializeTypeAndVerifyAccess, void*, uint32_t>(); 5440 return; 5441 } 5442 5443 Location out_loc = locations->Out(); 5444 Register out = out_loc.AsRegister<Register>(); 5445 5446 bool generate_null_check = false; 5447 switch (cls->GetLoadKind()) { 5448 case HLoadClass::LoadKind::kReferrersClass: { 5449 DCHECK(!cls->CanCallRuntime()); 5450 DCHECK(!cls->MustGenerateClinitCheck()); 5451 // /* GcRoot<mirror::Class> */ out = current_method->declaring_class_ 5452 Register current_method = locations->InAt(0).AsRegister<Register>(); 5453 GenerateGcRootFieldLoad( 5454 cls, out_loc, current_method, ArtMethod::DeclaringClassOffset().Int32Value()); 5455 break; 5456 } 5457 case HLoadClass::LoadKind::kBootImageLinkTimeAddress: { 5458 DCHECK(!kEmitCompilerReadBarrier); 5459 __ LoadLiteral(out, codegen_->DeduplicateBootImageTypeLiteral(cls->GetDexFile(), 5460 cls->GetTypeIndex())); 5461 break; 5462 } 5463 case HLoadClass::LoadKind::kBootImageLinkTimePcRelative: { 5464 DCHECK(!kEmitCompilerReadBarrier); 5465 CodeGeneratorARM::PcRelativePatchInfo* labels = 5466 codegen_->NewPcRelativeTypePatch(cls->GetDexFile(), cls->GetTypeIndex()); 5467 __ BindTrackedLabel(&labels->movw_label); 5468 __ movw(out, /* placeholder */ 0u); 5469 __ BindTrackedLabel(&labels->movt_label); 5470 __ movt(out, /* placeholder */ 0u); 5471 __ BindTrackedLabel(&labels->add_pc_label); 5472 __ add(out, out, ShifterOperand(PC)); 5473 break; 5474 } 5475 case HLoadClass::LoadKind::kBootImageAddress: { 5476 DCHECK(!kEmitCompilerReadBarrier); 5477 DCHECK_NE(cls->GetAddress(), 0u); 5478 uint32_t address = dchecked_integral_cast<uint32_t>(cls->GetAddress()); 5479 __ LoadLiteral(out, codegen_->DeduplicateBootImageAddressLiteral(address)); 5480 break; 5481 } 5482 case HLoadClass::LoadKind::kDexCacheAddress: { 5483 DCHECK_NE(cls->GetAddress(), 0u); 5484 uint32_t address = dchecked_integral_cast<uint32_t>(cls->GetAddress()); 5485 // 16-bit LDR immediate has a 5-bit offset multiplied by the size and that gives 5486 // a 128B range. To try and reduce the number of literals if we load multiple types, 5487 // simply split the dex cache address to a 128B aligned base loaded from a literal 5488 // and the remaining offset embedded in the load. 5489 static_assert(sizeof(GcRoot<mirror::Class>) == 4u, "Expected GC root to be 4 bytes."); 5490 DCHECK_ALIGNED(cls->GetAddress(), 4u); 5491 constexpr size_t offset_bits = /* encoded bits */ 5 + /* scale */ 2; 5492 uint32_t base_address = address & ~MaxInt<uint32_t>(offset_bits); 5493 uint32_t offset = address & MaxInt<uint32_t>(offset_bits); 5494 __ LoadLiteral(out, codegen_->DeduplicateDexCacheAddressLiteral(base_address)); 5495 // /* GcRoot<mirror::Class> */ out = *(base_address + offset) 5496 GenerateGcRootFieldLoad(cls, out_loc, out, offset); 5497 generate_null_check = !cls->IsInDexCache(); 5498 break; 5499 } 5500 case HLoadClass::LoadKind::kDexCachePcRelative: { 5501 Register base_reg = locations->InAt(0).AsRegister<Register>(); 5502 HArmDexCacheArraysBase* base = cls->InputAt(0)->AsArmDexCacheArraysBase(); 5503 int32_t offset = cls->GetDexCacheElementOffset() - base->GetElementOffset(); 5504 // /* GcRoot<mirror::Class> */ out = *(dex_cache_arrays_base + offset) 5505 GenerateGcRootFieldLoad(cls, out_loc, base_reg, offset); 5506 generate_null_check = !cls->IsInDexCache(); 5507 break; 5508 } 5509 case HLoadClass::LoadKind::kDexCacheViaMethod: { 5510 // /* GcRoot<mirror::Class>[] */ out = 5511 // current_method.ptr_sized_fields_->dex_cache_resolved_types_ 5512 Register current_method = locations->InAt(0).AsRegister<Register>(); 5513 __ LoadFromOffset(kLoadWord, 5514 out, 5515 current_method, 5516 ArtMethod::DexCacheResolvedTypesOffset(kArmPointerSize).Int32Value()); 5517 // /* GcRoot<mirror::Class> */ out = out[type_index] 5518 size_t offset = CodeGenerator::GetCacheOffset(cls->GetTypeIndex()); 5519 GenerateGcRootFieldLoad(cls, out_loc, out, offset); 5520 generate_null_check = !cls->IsInDexCache(); 5521 } 5522 } 5523 5524 if (generate_null_check || cls->MustGenerateClinitCheck()) { 5525 DCHECK(cls->CanCallRuntime()); 5526 SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathARM( 5527 cls, cls, cls->GetDexPc(), cls->MustGenerateClinitCheck()); 5528 codegen_->AddSlowPath(slow_path); 5529 if (generate_null_check) { 5530 __ CompareAndBranchIfZero(out, slow_path->GetEntryLabel()); 5531 } 5532 if (cls->MustGenerateClinitCheck()) { 5533 GenerateClassInitializationCheck(slow_path, out); 5534 } else { 5535 __ Bind(slow_path->GetExitLabel()); 5536 } 5537 } 5538} 5539 5540void LocationsBuilderARM::VisitClinitCheck(HClinitCheck* check) { 5541 LocationSummary* locations = 5542 new (GetGraph()->GetArena()) LocationSummary(check, LocationSummary::kCallOnSlowPath); 5543 locations->SetInAt(0, Location::RequiresRegister()); 5544 if (check->HasUses()) { 5545 locations->SetOut(Location::SameAsFirstInput()); 5546 } 5547} 5548 5549void InstructionCodeGeneratorARM::VisitClinitCheck(HClinitCheck* check) { 5550 // We assume the class is not null. 5551 SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathARM( 5552 check->GetLoadClass(), check, check->GetDexPc(), true); 5553 codegen_->AddSlowPath(slow_path); 5554 GenerateClassInitializationCheck(slow_path, 5555 check->GetLocations()->InAt(0).AsRegister<Register>()); 5556} 5557 5558void InstructionCodeGeneratorARM::GenerateClassInitializationCheck( 5559 SlowPathCode* slow_path, Register class_reg) { 5560 __ LoadFromOffset(kLoadWord, IP, class_reg, mirror::Class::StatusOffset().Int32Value()); 5561 __ cmp(IP, ShifterOperand(mirror::Class::kStatusInitialized)); 5562 __ b(slow_path->GetEntryLabel(), LT); 5563 // Even if the initialized flag is set, we may be in a situation where caches are not synced 5564 // properly. Therefore, we do a memory fence. 5565 __ dmb(ISH); 5566 __ Bind(slow_path->GetExitLabel()); 5567} 5568 5569HLoadString::LoadKind CodeGeneratorARM::GetSupportedLoadStringKind( 5570 HLoadString::LoadKind desired_string_load_kind) { 5571 if (kEmitCompilerReadBarrier) { 5572 switch (desired_string_load_kind) { 5573 case HLoadString::LoadKind::kBootImageLinkTimeAddress: 5574 case HLoadString::LoadKind::kBootImageLinkTimePcRelative: 5575 case HLoadString::LoadKind::kBootImageAddress: 5576 // TODO: Implement for read barrier. 5577 return HLoadString::LoadKind::kDexCacheViaMethod; 5578 default: 5579 break; 5580 } 5581 } 5582 switch (desired_string_load_kind) { 5583 case HLoadString::LoadKind::kBootImageLinkTimeAddress: 5584 DCHECK(!GetCompilerOptions().GetCompilePic()); 5585 break; 5586 case HLoadString::LoadKind::kBootImageLinkTimePcRelative: 5587 DCHECK(GetCompilerOptions().GetCompilePic()); 5588 break; 5589 case HLoadString::LoadKind::kBootImageAddress: 5590 break; 5591 case HLoadString::LoadKind::kDexCacheAddress: 5592 DCHECK(Runtime::Current()->UseJitCompilation()); 5593 break; 5594 case HLoadString::LoadKind::kDexCachePcRelative: 5595 DCHECK(!Runtime::Current()->UseJitCompilation()); 5596 // We disable pc-relative load when there is an irreducible loop, as the optimization 5597 // is incompatible with it. 5598 // TODO: Create as many ArmDexCacheArraysBase instructions as needed for methods 5599 // with irreducible loops. 5600 if (GetGraph()->HasIrreducibleLoops()) { 5601 return HLoadString::LoadKind::kDexCacheViaMethod; 5602 } 5603 break; 5604 case HLoadString::LoadKind::kDexCacheViaMethod: 5605 break; 5606 } 5607 return desired_string_load_kind; 5608} 5609 5610void LocationsBuilderARM::VisitLoadString(HLoadString* load) { 5611 LocationSummary::CallKind call_kind = (load->NeedsEnvironment() || kEmitCompilerReadBarrier) 5612 ? LocationSummary::kCallOnSlowPath 5613 : LocationSummary::kNoCall; 5614 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(load, call_kind); 5615 HLoadString::LoadKind load_kind = load->GetLoadKind(); 5616 if (load_kind == HLoadString::LoadKind::kDexCacheViaMethod || 5617 load_kind == HLoadString::LoadKind::kDexCachePcRelative) { 5618 locations->SetInAt(0, Location::RequiresRegister()); 5619 } 5620 locations->SetOut(Location::RequiresRegister()); 5621} 5622 5623void InstructionCodeGeneratorARM::VisitLoadString(HLoadString* load) { 5624 LocationSummary* locations = load->GetLocations(); 5625 Location out_loc = locations->Out(); 5626 Register out = out_loc.AsRegister<Register>(); 5627 5628 switch (load->GetLoadKind()) { 5629 case HLoadString::LoadKind::kBootImageLinkTimeAddress: { 5630 DCHECK(!kEmitCompilerReadBarrier); 5631 __ LoadLiteral(out, codegen_->DeduplicateBootImageStringLiteral(load->GetDexFile(), 5632 load->GetStringIndex())); 5633 return; // No dex cache slow path. 5634 } 5635 case HLoadString::LoadKind::kBootImageLinkTimePcRelative: { 5636 DCHECK(!kEmitCompilerReadBarrier); 5637 CodeGeneratorARM::PcRelativePatchInfo* labels = 5638 codegen_->NewPcRelativeStringPatch(load->GetDexFile(), load->GetStringIndex()); 5639 __ BindTrackedLabel(&labels->movw_label); 5640 __ movw(out, /* placeholder */ 0u); 5641 __ BindTrackedLabel(&labels->movt_label); 5642 __ movt(out, /* placeholder */ 0u); 5643 __ BindTrackedLabel(&labels->add_pc_label); 5644 __ add(out, out, ShifterOperand(PC)); 5645 return; // No dex cache slow path. 5646 } 5647 case HLoadString::LoadKind::kBootImageAddress: { 5648 DCHECK(!kEmitCompilerReadBarrier); 5649 DCHECK_NE(load->GetAddress(), 0u); 5650 uint32_t address = dchecked_integral_cast<uint32_t>(load->GetAddress()); 5651 __ LoadLiteral(out, codegen_->DeduplicateBootImageAddressLiteral(address)); 5652 return; // No dex cache slow path. 5653 } 5654 default: 5655 break; 5656 } 5657 5658 // TODO: Re-add the compiler code to do string dex cache lookup again. 5659 SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadStringSlowPathARM(load); 5660 codegen_->AddSlowPath(slow_path); 5661 __ b(slow_path->GetEntryLabel()); 5662 __ Bind(slow_path->GetExitLabel()); 5663} 5664 5665static int32_t GetExceptionTlsOffset() { 5666 return Thread::ExceptionOffset<kArmPointerSize>().Int32Value(); 5667} 5668 5669void LocationsBuilderARM::VisitLoadException(HLoadException* load) { 5670 LocationSummary* locations = 5671 new (GetGraph()->GetArena()) LocationSummary(load, LocationSummary::kNoCall); 5672 locations->SetOut(Location::RequiresRegister()); 5673} 5674 5675void InstructionCodeGeneratorARM::VisitLoadException(HLoadException* load) { 5676 Register out = load->GetLocations()->Out().AsRegister<Register>(); 5677 __ LoadFromOffset(kLoadWord, out, TR, GetExceptionTlsOffset()); 5678} 5679 5680void LocationsBuilderARM::VisitClearException(HClearException* clear) { 5681 new (GetGraph()->GetArena()) LocationSummary(clear, LocationSummary::kNoCall); 5682} 5683 5684void InstructionCodeGeneratorARM::VisitClearException(HClearException* clear ATTRIBUTE_UNUSED) { 5685 __ LoadImmediate(IP, 0); 5686 __ StoreToOffset(kStoreWord, IP, TR, GetExceptionTlsOffset()); 5687} 5688 5689void LocationsBuilderARM::VisitThrow(HThrow* instruction) { 5690 LocationSummary* locations = 5691 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnMainOnly); 5692 InvokeRuntimeCallingConvention calling_convention; 5693 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 5694} 5695 5696void InstructionCodeGeneratorARM::VisitThrow(HThrow* instruction) { 5697 codegen_->InvokeRuntime( 5698 QUICK_ENTRY_POINT(pDeliverException), instruction, instruction->GetDexPc(), nullptr); 5699 CheckEntrypointTypes<kQuickDeliverException, void, mirror::Object*>(); 5700} 5701 5702static bool TypeCheckNeedsATemporary(TypeCheckKind type_check_kind) { 5703 return kEmitCompilerReadBarrier && 5704 (kUseBakerReadBarrier || 5705 type_check_kind == TypeCheckKind::kAbstractClassCheck || 5706 type_check_kind == TypeCheckKind::kClassHierarchyCheck || 5707 type_check_kind == TypeCheckKind::kArrayObjectCheck); 5708} 5709 5710void LocationsBuilderARM::VisitInstanceOf(HInstanceOf* instruction) { 5711 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 5712 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 5713 switch (type_check_kind) { 5714 case TypeCheckKind::kExactCheck: 5715 case TypeCheckKind::kAbstractClassCheck: 5716 case TypeCheckKind::kClassHierarchyCheck: 5717 case TypeCheckKind::kArrayObjectCheck: 5718 call_kind = 5719 kEmitCompilerReadBarrier ? LocationSummary::kCallOnSlowPath : LocationSummary::kNoCall; 5720 break; 5721 case TypeCheckKind::kArrayCheck: 5722 case TypeCheckKind::kUnresolvedCheck: 5723 case TypeCheckKind::kInterfaceCheck: 5724 call_kind = LocationSummary::kCallOnSlowPath; 5725 break; 5726 } 5727 5728 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 5729 locations->SetInAt(0, Location::RequiresRegister()); 5730 locations->SetInAt(1, Location::RequiresRegister()); 5731 // The "out" register is used as a temporary, so it overlaps with the inputs. 5732 // Note that TypeCheckSlowPathARM uses this register too. 5733 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 5734 // When read barriers are enabled, we need a temporary register for 5735 // some cases. 5736 if (TypeCheckNeedsATemporary(type_check_kind)) { 5737 locations->AddTemp(Location::RequiresRegister()); 5738 } 5739} 5740 5741void InstructionCodeGeneratorARM::VisitInstanceOf(HInstanceOf* instruction) { 5742 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 5743 LocationSummary* locations = instruction->GetLocations(); 5744 Location obj_loc = locations->InAt(0); 5745 Register obj = obj_loc.AsRegister<Register>(); 5746 Register cls = locations->InAt(1).AsRegister<Register>(); 5747 Location out_loc = locations->Out(); 5748 Register out = out_loc.AsRegister<Register>(); 5749 Location maybe_temp_loc = TypeCheckNeedsATemporary(type_check_kind) ? 5750 locations->GetTemp(0) : 5751 Location::NoLocation(); 5752 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 5753 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 5754 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 5755 uint32_t primitive_offset = mirror::Class::PrimitiveTypeOffset().Int32Value(); 5756 Label done, zero; 5757 SlowPathCode* slow_path = nullptr; 5758 5759 // Return 0 if `obj` is null. 5760 // avoid null check if we know obj is not null. 5761 if (instruction->MustDoNullCheck()) { 5762 __ CompareAndBranchIfZero(obj, &zero); 5763 } 5764 5765 // /* HeapReference<Class> */ out = obj->klass_ 5766 GenerateReferenceLoadTwoRegisters(instruction, out_loc, obj_loc, class_offset, maybe_temp_loc); 5767 5768 switch (type_check_kind) { 5769 case TypeCheckKind::kExactCheck: { 5770 __ cmp(out, ShifterOperand(cls)); 5771 // Classes must be equal for the instanceof to succeed. 5772 __ b(&zero, NE); 5773 __ LoadImmediate(out, 1); 5774 __ b(&done); 5775 break; 5776 } 5777 5778 case TypeCheckKind::kAbstractClassCheck: { 5779 // If the class is abstract, we eagerly fetch the super class of the 5780 // object to avoid doing a comparison we know will fail. 5781 Label loop; 5782 __ Bind(&loop); 5783 // /* HeapReference<Class> */ out = out->super_class_ 5784 GenerateReferenceLoadOneRegister(instruction, out_loc, super_offset, maybe_temp_loc); 5785 // If `out` is null, we use it for the result, and jump to `done`. 5786 __ CompareAndBranchIfZero(out, &done); 5787 __ cmp(out, ShifterOperand(cls)); 5788 __ b(&loop, NE); 5789 __ LoadImmediate(out, 1); 5790 if (zero.IsLinked()) { 5791 __ b(&done); 5792 } 5793 break; 5794 } 5795 5796 case TypeCheckKind::kClassHierarchyCheck: { 5797 // Walk over the class hierarchy to find a match. 5798 Label loop, success; 5799 __ Bind(&loop); 5800 __ cmp(out, ShifterOperand(cls)); 5801 __ b(&success, EQ); 5802 // /* HeapReference<Class> */ out = out->super_class_ 5803 GenerateReferenceLoadOneRegister(instruction, out_loc, super_offset, maybe_temp_loc); 5804 __ CompareAndBranchIfNonZero(out, &loop); 5805 // If `out` is null, we use it for the result, and jump to `done`. 5806 __ b(&done); 5807 __ Bind(&success); 5808 __ LoadImmediate(out, 1); 5809 if (zero.IsLinked()) { 5810 __ b(&done); 5811 } 5812 break; 5813 } 5814 5815 case TypeCheckKind::kArrayObjectCheck: { 5816 // Do an exact check. 5817 Label exact_check; 5818 __ cmp(out, ShifterOperand(cls)); 5819 __ b(&exact_check, EQ); 5820 // Otherwise, we need to check that the object's class is a non-primitive array. 5821 // /* HeapReference<Class> */ out = out->component_type_ 5822 GenerateReferenceLoadOneRegister(instruction, out_loc, component_offset, maybe_temp_loc); 5823 // If `out` is null, we use it for the result, and jump to `done`. 5824 __ CompareAndBranchIfZero(out, &done); 5825 __ LoadFromOffset(kLoadUnsignedHalfword, out, out, primitive_offset); 5826 static_assert(Primitive::kPrimNot == 0, "Expected 0 for kPrimNot"); 5827 __ CompareAndBranchIfNonZero(out, &zero); 5828 __ Bind(&exact_check); 5829 __ LoadImmediate(out, 1); 5830 __ b(&done); 5831 break; 5832 } 5833 5834 case TypeCheckKind::kArrayCheck: { 5835 __ cmp(out, ShifterOperand(cls)); 5836 DCHECK(locations->OnlyCallsOnSlowPath()); 5837 slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathARM(instruction, 5838 /* is_fatal */ false); 5839 codegen_->AddSlowPath(slow_path); 5840 __ b(slow_path->GetEntryLabel(), NE); 5841 __ LoadImmediate(out, 1); 5842 if (zero.IsLinked()) { 5843 __ b(&done); 5844 } 5845 break; 5846 } 5847 5848 case TypeCheckKind::kUnresolvedCheck: 5849 case TypeCheckKind::kInterfaceCheck: { 5850 // Note that we indeed only call on slow path, but we always go 5851 // into the slow path for the unresolved and interface check 5852 // cases. 5853 // 5854 // We cannot directly call the InstanceofNonTrivial runtime 5855 // entry point without resorting to a type checking slow path 5856 // here (i.e. by calling InvokeRuntime directly), as it would 5857 // require to assign fixed registers for the inputs of this 5858 // HInstanceOf instruction (following the runtime calling 5859 // convention), which might be cluttered by the potential first 5860 // read barrier emission at the beginning of this method. 5861 // 5862 // TODO: Introduce a new runtime entry point taking the object 5863 // to test (instead of its class) as argument, and let it deal 5864 // with the read barrier issues. This will let us refactor this 5865 // case of the `switch` code as it was previously (with a direct 5866 // call to the runtime not using a type checking slow path). 5867 // This should also be beneficial for the other cases above. 5868 DCHECK(locations->OnlyCallsOnSlowPath()); 5869 slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathARM(instruction, 5870 /* is_fatal */ false); 5871 codegen_->AddSlowPath(slow_path); 5872 __ b(slow_path->GetEntryLabel()); 5873 if (zero.IsLinked()) { 5874 __ b(&done); 5875 } 5876 break; 5877 } 5878 } 5879 5880 if (zero.IsLinked()) { 5881 __ Bind(&zero); 5882 __ LoadImmediate(out, 0); 5883 } 5884 5885 if (done.IsLinked()) { 5886 __ Bind(&done); 5887 } 5888 5889 if (slow_path != nullptr) { 5890 __ Bind(slow_path->GetExitLabel()); 5891 } 5892} 5893 5894void LocationsBuilderARM::VisitCheckCast(HCheckCast* instruction) { 5895 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 5896 bool throws_into_catch = instruction->CanThrowIntoCatchBlock(); 5897 5898 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 5899 switch (type_check_kind) { 5900 case TypeCheckKind::kExactCheck: 5901 case TypeCheckKind::kAbstractClassCheck: 5902 case TypeCheckKind::kClassHierarchyCheck: 5903 case TypeCheckKind::kArrayObjectCheck: 5904 call_kind = (throws_into_catch || kEmitCompilerReadBarrier) ? 5905 LocationSummary::kCallOnSlowPath : 5906 LocationSummary::kNoCall; // In fact, call on a fatal (non-returning) slow path. 5907 break; 5908 case TypeCheckKind::kArrayCheck: 5909 case TypeCheckKind::kUnresolvedCheck: 5910 case TypeCheckKind::kInterfaceCheck: 5911 call_kind = LocationSummary::kCallOnSlowPath; 5912 break; 5913 } 5914 5915 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 5916 locations->SetInAt(0, Location::RequiresRegister()); 5917 locations->SetInAt(1, Location::RequiresRegister()); 5918 // Note that TypeCheckSlowPathARM uses this "temp" register too. 5919 locations->AddTemp(Location::RequiresRegister()); 5920 // When read barriers are enabled, we need an additional temporary 5921 // register for some cases. 5922 if (TypeCheckNeedsATemporary(type_check_kind)) { 5923 locations->AddTemp(Location::RequiresRegister()); 5924 } 5925} 5926 5927void InstructionCodeGeneratorARM::VisitCheckCast(HCheckCast* instruction) { 5928 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 5929 LocationSummary* locations = instruction->GetLocations(); 5930 Location obj_loc = locations->InAt(0); 5931 Register obj = obj_loc.AsRegister<Register>(); 5932 Register cls = locations->InAt(1).AsRegister<Register>(); 5933 Location temp_loc = locations->GetTemp(0); 5934 Register temp = temp_loc.AsRegister<Register>(); 5935 Location maybe_temp2_loc = TypeCheckNeedsATemporary(type_check_kind) ? 5936 locations->GetTemp(1) : 5937 Location::NoLocation(); 5938 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 5939 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 5940 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 5941 uint32_t primitive_offset = mirror::Class::PrimitiveTypeOffset().Int32Value(); 5942 5943 bool is_type_check_slow_path_fatal = 5944 (type_check_kind == TypeCheckKind::kExactCheck || 5945 type_check_kind == TypeCheckKind::kAbstractClassCheck || 5946 type_check_kind == TypeCheckKind::kClassHierarchyCheck || 5947 type_check_kind == TypeCheckKind::kArrayObjectCheck) && 5948 !instruction->CanThrowIntoCatchBlock(); 5949 SlowPathCode* type_check_slow_path = 5950 new (GetGraph()->GetArena()) TypeCheckSlowPathARM(instruction, 5951 is_type_check_slow_path_fatal); 5952 codegen_->AddSlowPath(type_check_slow_path); 5953 5954 Label done; 5955 // Avoid null check if we know obj is not null. 5956 if (instruction->MustDoNullCheck()) { 5957 __ CompareAndBranchIfZero(obj, &done); 5958 } 5959 5960 // /* HeapReference<Class> */ temp = obj->klass_ 5961 GenerateReferenceLoadTwoRegisters(instruction, temp_loc, obj_loc, class_offset, maybe_temp2_loc); 5962 5963 switch (type_check_kind) { 5964 case TypeCheckKind::kExactCheck: 5965 case TypeCheckKind::kArrayCheck: { 5966 __ cmp(temp, ShifterOperand(cls)); 5967 // Jump to slow path for throwing the exception or doing a 5968 // more involved array check. 5969 __ b(type_check_slow_path->GetEntryLabel(), NE); 5970 break; 5971 } 5972 5973 case TypeCheckKind::kAbstractClassCheck: { 5974 // If the class is abstract, we eagerly fetch the super class of the 5975 // object to avoid doing a comparison we know will fail. 5976 Label loop, compare_classes; 5977 __ Bind(&loop); 5978 // /* HeapReference<Class> */ temp = temp->super_class_ 5979 GenerateReferenceLoadOneRegister(instruction, temp_loc, super_offset, maybe_temp2_loc); 5980 5981 // If the class reference currently in `temp` is not null, jump 5982 // to the `compare_classes` label to compare it with the checked 5983 // class. 5984 __ CompareAndBranchIfNonZero(temp, &compare_classes); 5985 // Otherwise, jump to the slow path to throw the exception. 5986 // 5987 // But before, move back the object's class into `temp` before 5988 // going into the slow path, as it has been overwritten in the 5989 // meantime. 5990 // /* HeapReference<Class> */ temp = obj->klass_ 5991 GenerateReferenceLoadTwoRegisters( 5992 instruction, temp_loc, obj_loc, class_offset, maybe_temp2_loc); 5993 __ b(type_check_slow_path->GetEntryLabel()); 5994 5995 __ Bind(&compare_classes); 5996 __ cmp(temp, ShifterOperand(cls)); 5997 __ b(&loop, NE); 5998 break; 5999 } 6000 6001 case TypeCheckKind::kClassHierarchyCheck: { 6002 // Walk over the class hierarchy to find a match. 6003 Label loop; 6004 __ Bind(&loop); 6005 __ cmp(temp, ShifterOperand(cls)); 6006 __ b(&done, EQ); 6007 6008 // /* HeapReference<Class> */ temp = temp->super_class_ 6009 GenerateReferenceLoadOneRegister(instruction, temp_loc, super_offset, maybe_temp2_loc); 6010 6011 // If the class reference currently in `temp` is not null, jump 6012 // back at the beginning of the loop. 6013 __ CompareAndBranchIfNonZero(temp, &loop); 6014 // Otherwise, jump to the slow path to throw the exception. 6015 // 6016 // But before, move back the object's class into `temp` before 6017 // going into the slow path, as it has been overwritten in the 6018 // meantime. 6019 // /* HeapReference<Class> */ temp = obj->klass_ 6020 GenerateReferenceLoadTwoRegisters( 6021 instruction, temp_loc, obj_loc, class_offset, maybe_temp2_loc); 6022 __ b(type_check_slow_path->GetEntryLabel()); 6023 break; 6024 } 6025 6026 case TypeCheckKind::kArrayObjectCheck: { 6027 // Do an exact check. 6028 Label check_non_primitive_component_type; 6029 __ cmp(temp, ShifterOperand(cls)); 6030 __ b(&done, EQ); 6031 6032 // Otherwise, we need to check that the object's class is a non-primitive array. 6033 // /* HeapReference<Class> */ temp = temp->component_type_ 6034 GenerateReferenceLoadOneRegister(instruction, temp_loc, component_offset, maybe_temp2_loc); 6035 6036 // If the component type is not null (i.e. the object is indeed 6037 // an array), jump to label `check_non_primitive_component_type` 6038 // to further check that this component type is not a primitive 6039 // type. 6040 __ CompareAndBranchIfNonZero(temp, &check_non_primitive_component_type); 6041 // Otherwise, jump to the slow path to throw the exception. 6042 // 6043 // But before, move back the object's class into `temp` before 6044 // going into the slow path, as it has been overwritten in the 6045 // meantime. 6046 // /* HeapReference<Class> */ temp = obj->klass_ 6047 GenerateReferenceLoadTwoRegisters( 6048 instruction, temp_loc, obj_loc, class_offset, maybe_temp2_loc); 6049 __ b(type_check_slow_path->GetEntryLabel()); 6050 6051 __ Bind(&check_non_primitive_component_type); 6052 __ LoadFromOffset(kLoadUnsignedHalfword, temp, temp, primitive_offset); 6053 static_assert(Primitive::kPrimNot == 0, "Expected 0 for art::Primitive::kPrimNot"); 6054 __ CompareAndBranchIfZero(temp, &done); 6055 // Same comment as above regarding `temp` and the slow path. 6056 // /* HeapReference<Class> */ temp = obj->klass_ 6057 GenerateReferenceLoadTwoRegisters( 6058 instruction, temp_loc, obj_loc, class_offset, maybe_temp2_loc); 6059 __ b(type_check_slow_path->GetEntryLabel()); 6060 break; 6061 } 6062 6063 case TypeCheckKind::kUnresolvedCheck: 6064 case TypeCheckKind::kInterfaceCheck: 6065 // We always go into the type check slow path for the unresolved 6066 // and interface check cases. 6067 // 6068 // We cannot directly call the CheckCast runtime entry point 6069 // without resorting to a type checking slow path here (i.e. by 6070 // calling InvokeRuntime directly), as it would require to 6071 // assign fixed registers for the inputs of this HInstanceOf 6072 // instruction (following the runtime calling convention), which 6073 // might be cluttered by the potential first read barrier 6074 // emission at the beginning of this method. 6075 // 6076 // TODO: Introduce a new runtime entry point taking the object 6077 // to test (instead of its class) as argument, and let it deal 6078 // with the read barrier issues. This will let us refactor this 6079 // case of the `switch` code as it was previously (with a direct 6080 // call to the runtime not using a type checking slow path). 6081 // This should also be beneficial for the other cases above. 6082 __ b(type_check_slow_path->GetEntryLabel()); 6083 break; 6084 } 6085 __ Bind(&done); 6086 6087 __ Bind(type_check_slow_path->GetExitLabel()); 6088} 6089 6090void LocationsBuilderARM::VisitMonitorOperation(HMonitorOperation* instruction) { 6091 LocationSummary* locations = 6092 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnMainOnly); 6093 InvokeRuntimeCallingConvention calling_convention; 6094 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 6095} 6096 6097void InstructionCodeGeneratorARM::VisitMonitorOperation(HMonitorOperation* instruction) { 6098 codegen_->InvokeRuntime(instruction->IsEnter() 6099 ? QUICK_ENTRY_POINT(pLockObject) : QUICK_ENTRY_POINT(pUnlockObject), 6100 instruction, 6101 instruction->GetDexPc(), 6102 nullptr); 6103 if (instruction->IsEnter()) { 6104 CheckEntrypointTypes<kQuickLockObject, void, mirror::Object*>(); 6105 } else { 6106 CheckEntrypointTypes<kQuickUnlockObject, void, mirror::Object*>(); 6107 } 6108} 6109 6110void LocationsBuilderARM::VisitAnd(HAnd* instruction) { HandleBitwiseOperation(instruction, AND); } 6111void LocationsBuilderARM::VisitOr(HOr* instruction) { HandleBitwiseOperation(instruction, ORR); } 6112void LocationsBuilderARM::VisitXor(HXor* instruction) { HandleBitwiseOperation(instruction, EOR); } 6113 6114void LocationsBuilderARM::HandleBitwiseOperation(HBinaryOperation* instruction, Opcode opcode) { 6115 LocationSummary* locations = 6116 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 6117 DCHECK(instruction->GetResultType() == Primitive::kPrimInt 6118 || instruction->GetResultType() == Primitive::kPrimLong); 6119 // Note: GVN reorders commutative operations to have the constant on the right hand side. 6120 locations->SetInAt(0, Location::RequiresRegister()); 6121 locations->SetInAt(1, ArmEncodableConstantOrRegister(instruction->InputAt(1), opcode)); 6122 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 6123} 6124 6125void InstructionCodeGeneratorARM::VisitAnd(HAnd* instruction) { 6126 HandleBitwiseOperation(instruction); 6127} 6128 6129void InstructionCodeGeneratorARM::VisitOr(HOr* instruction) { 6130 HandleBitwiseOperation(instruction); 6131} 6132 6133void InstructionCodeGeneratorARM::VisitXor(HXor* instruction) { 6134 HandleBitwiseOperation(instruction); 6135} 6136 6137 6138void LocationsBuilderARM::VisitBitwiseNegatedRight(HBitwiseNegatedRight* instruction) { 6139 LocationSummary* locations = 6140 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 6141 DCHECK(instruction->GetResultType() == Primitive::kPrimInt 6142 || instruction->GetResultType() == Primitive::kPrimLong); 6143 6144 locations->SetInAt(0, Location::RequiresRegister()); 6145 locations->SetInAt(1, Location::RequiresRegister()); 6146 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 6147} 6148 6149void InstructionCodeGeneratorARM::VisitBitwiseNegatedRight(HBitwiseNegatedRight* instruction) { 6150 LocationSummary* locations = instruction->GetLocations(); 6151 Location first = locations->InAt(0); 6152 Location second = locations->InAt(1); 6153 Location out = locations->Out(); 6154 6155 if (instruction->GetResultType() == Primitive::kPrimInt) { 6156 Register first_reg = first.AsRegister<Register>(); 6157 ShifterOperand second_reg(second.AsRegister<Register>()); 6158 Register out_reg = out.AsRegister<Register>(); 6159 6160 switch (instruction->GetOpKind()) { 6161 case HInstruction::kAnd: 6162 __ bic(out_reg, first_reg, second_reg); 6163 break; 6164 case HInstruction::kOr: 6165 __ orn(out_reg, first_reg, second_reg); 6166 break; 6167 // There is no EON on arm. 6168 case HInstruction::kXor: 6169 default: 6170 LOG(FATAL) << "Unexpected instruction " << instruction->DebugName(); 6171 UNREACHABLE(); 6172 } 6173 return; 6174 6175 } else { 6176 DCHECK_EQ(instruction->GetResultType(), Primitive::kPrimLong); 6177 Register first_low = first.AsRegisterPairLow<Register>(); 6178 Register first_high = first.AsRegisterPairHigh<Register>(); 6179 ShifterOperand second_low(second.AsRegisterPairLow<Register>()); 6180 ShifterOperand second_high(second.AsRegisterPairHigh<Register>()); 6181 Register out_low = out.AsRegisterPairLow<Register>(); 6182 Register out_high = out.AsRegisterPairHigh<Register>(); 6183 6184 switch (instruction->GetOpKind()) { 6185 case HInstruction::kAnd: 6186 __ bic(out_low, first_low, second_low); 6187 __ bic(out_high, first_high, second_high); 6188 break; 6189 case HInstruction::kOr: 6190 __ orn(out_low, first_low, second_low); 6191 __ orn(out_high, first_high, second_high); 6192 break; 6193 // There is no EON on arm. 6194 case HInstruction::kXor: 6195 default: 6196 LOG(FATAL) << "Unexpected instruction " << instruction->DebugName(); 6197 UNREACHABLE(); 6198 } 6199 } 6200} 6201 6202void InstructionCodeGeneratorARM::GenerateAndConst(Register out, Register first, uint32_t value) { 6203 // Optimize special cases for individual halfs of `and-long` (`and` is simplified earlier). 6204 if (value == 0xffffffffu) { 6205 if (out != first) { 6206 __ mov(out, ShifterOperand(first)); 6207 } 6208 return; 6209 } 6210 if (value == 0u) { 6211 __ mov(out, ShifterOperand(0)); 6212 return; 6213 } 6214 ShifterOperand so; 6215 if (__ ShifterOperandCanHold(kNoRegister, kNoRegister, AND, value, &so)) { 6216 __ and_(out, first, so); 6217 } else { 6218 DCHECK(__ ShifterOperandCanHold(kNoRegister, kNoRegister, BIC, ~value, &so)); 6219 __ bic(out, first, ShifterOperand(~value)); 6220 } 6221} 6222 6223void InstructionCodeGeneratorARM::GenerateOrrConst(Register out, Register first, uint32_t value) { 6224 // Optimize special cases for individual halfs of `or-long` (`or` is simplified earlier). 6225 if (value == 0u) { 6226 if (out != first) { 6227 __ mov(out, ShifterOperand(first)); 6228 } 6229 return; 6230 } 6231 if (value == 0xffffffffu) { 6232 __ mvn(out, ShifterOperand(0)); 6233 return; 6234 } 6235 ShifterOperand so; 6236 if (__ ShifterOperandCanHold(kNoRegister, kNoRegister, ORR, value, &so)) { 6237 __ orr(out, first, so); 6238 } else { 6239 DCHECK(__ ShifterOperandCanHold(kNoRegister, kNoRegister, ORN, ~value, &so)); 6240 __ orn(out, first, ShifterOperand(~value)); 6241 } 6242} 6243 6244void InstructionCodeGeneratorARM::GenerateEorConst(Register out, Register first, uint32_t value) { 6245 // Optimize special case for individual halfs of `xor-long` (`xor` is simplified earlier). 6246 if (value == 0u) { 6247 if (out != first) { 6248 __ mov(out, ShifterOperand(first)); 6249 } 6250 return; 6251 } 6252 __ eor(out, first, ShifterOperand(value)); 6253} 6254 6255void InstructionCodeGeneratorARM::GenerateAddLongConst(Location out, 6256 Location first, 6257 uint64_t value) { 6258 Register out_low = out.AsRegisterPairLow<Register>(); 6259 Register out_high = out.AsRegisterPairHigh<Register>(); 6260 Register first_low = first.AsRegisterPairLow<Register>(); 6261 Register first_high = first.AsRegisterPairHigh<Register>(); 6262 uint32_t value_low = Low32Bits(value); 6263 uint32_t value_high = High32Bits(value); 6264 if (value_low == 0u) { 6265 if (out_low != first_low) { 6266 __ mov(out_low, ShifterOperand(first_low)); 6267 } 6268 __ AddConstant(out_high, first_high, value_high); 6269 return; 6270 } 6271 __ AddConstantSetFlags(out_low, first_low, value_low); 6272 ShifterOperand so; 6273 if (__ ShifterOperandCanHold(out_high, first_high, ADC, value_high, kCcDontCare, &so)) { 6274 __ adc(out_high, first_high, so); 6275 } else if (__ ShifterOperandCanHold(out_low, first_low, SBC, ~value_high, kCcDontCare, &so)) { 6276 __ sbc(out_high, first_high, so); 6277 } else { 6278 LOG(FATAL) << "Unexpected constant " << value_high; 6279 UNREACHABLE(); 6280 } 6281} 6282 6283void InstructionCodeGeneratorARM::HandleBitwiseOperation(HBinaryOperation* instruction) { 6284 LocationSummary* locations = instruction->GetLocations(); 6285 Location first = locations->InAt(0); 6286 Location second = locations->InAt(1); 6287 Location out = locations->Out(); 6288 6289 if (second.IsConstant()) { 6290 uint64_t value = static_cast<uint64_t>(Int64FromConstant(second.GetConstant())); 6291 uint32_t value_low = Low32Bits(value); 6292 if (instruction->GetResultType() == Primitive::kPrimInt) { 6293 Register first_reg = first.AsRegister<Register>(); 6294 Register out_reg = out.AsRegister<Register>(); 6295 if (instruction->IsAnd()) { 6296 GenerateAndConst(out_reg, first_reg, value_low); 6297 } else if (instruction->IsOr()) { 6298 GenerateOrrConst(out_reg, first_reg, value_low); 6299 } else { 6300 DCHECK(instruction->IsXor()); 6301 GenerateEorConst(out_reg, first_reg, value_low); 6302 } 6303 } else { 6304 DCHECK_EQ(instruction->GetResultType(), Primitive::kPrimLong); 6305 uint32_t value_high = High32Bits(value); 6306 Register first_low = first.AsRegisterPairLow<Register>(); 6307 Register first_high = first.AsRegisterPairHigh<Register>(); 6308 Register out_low = out.AsRegisterPairLow<Register>(); 6309 Register out_high = out.AsRegisterPairHigh<Register>(); 6310 if (instruction->IsAnd()) { 6311 GenerateAndConst(out_low, first_low, value_low); 6312 GenerateAndConst(out_high, first_high, value_high); 6313 } else if (instruction->IsOr()) { 6314 GenerateOrrConst(out_low, first_low, value_low); 6315 GenerateOrrConst(out_high, first_high, value_high); 6316 } else { 6317 DCHECK(instruction->IsXor()); 6318 GenerateEorConst(out_low, first_low, value_low); 6319 GenerateEorConst(out_high, first_high, value_high); 6320 } 6321 } 6322 return; 6323 } 6324 6325 if (instruction->GetResultType() == Primitive::kPrimInt) { 6326 Register first_reg = first.AsRegister<Register>(); 6327 ShifterOperand second_reg(second.AsRegister<Register>()); 6328 Register out_reg = out.AsRegister<Register>(); 6329 if (instruction->IsAnd()) { 6330 __ and_(out_reg, first_reg, second_reg); 6331 } else if (instruction->IsOr()) { 6332 __ orr(out_reg, first_reg, second_reg); 6333 } else { 6334 DCHECK(instruction->IsXor()); 6335 __ eor(out_reg, first_reg, second_reg); 6336 } 6337 } else { 6338 DCHECK_EQ(instruction->GetResultType(), Primitive::kPrimLong); 6339 Register first_low = first.AsRegisterPairLow<Register>(); 6340 Register first_high = first.AsRegisterPairHigh<Register>(); 6341 ShifterOperand second_low(second.AsRegisterPairLow<Register>()); 6342 ShifterOperand second_high(second.AsRegisterPairHigh<Register>()); 6343 Register out_low = out.AsRegisterPairLow<Register>(); 6344 Register out_high = out.AsRegisterPairHigh<Register>(); 6345 if (instruction->IsAnd()) { 6346 __ and_(out_low, first_low, second_low); 6347 __ and_(out_high, first_high, second_high); 6348 } else if (instruction->IsOr()) { 6349 __ orr(out_low, first_low, second_low); 6350 __ orr(out_high, first_high, second_high); 6351 } else { 6352 DCHECK(instruction->IsXor()); 6353 __ eor(out_low, first_low, second_low); 6354 __ eor(out_high, first_high, second_high); 6355 } 6356 } 6357} 6358 6359void InstructionCodeGeneratorARM::GenerateReferenceLoadOneRegister(HInstruction* instruction, 6360 Location out, 6361 uint32_t offset, 6362 Location maybe_temp) { 6363 Register out_reg = out.AsRegister<Register>(); 6364 if (kEmitCompilerReadBarrier) { 6365 DCHECK(maybe_temp.IsRegister()) << maybe_temp; 6366 if (kUseBakerReadBarrier) { 6367 // Load with fast path based Baker's read barrier. 6368 // /* HeapReference<Object> */ out = *(out + offset) 6369 codegen_->GenerateFieldLoadWithBakerReadBarrier( 6370 instruction, out, out_reg, offset, maybe_temp, /* needs_null_check */ false); 6371 } else { 6372 // Load with slow path based read barrier. 6373 // Save the value of `out` into `maybe_temp` before overwriting it 6374 // in the following move operation, as we will need it for the 6375 // read barrier below. 6376 __ Mov(maybe_temp.AsRegister<Register>(), out_reg); 6377 // /* HeapReference<Object> */ out = *(out + offset) 6378 __ LoadFromOffset(kLoadWord, out_reg, out_reg, offset); 6379 codegen_->GenerateReadBarrierSlow(instruction, out, out, maybe_temp, offset); 6380 } 6381 } else { 6382 // Plain load with no read barrier. 6383 // /* HeapReference<Object> */ out = *(out + offset) 6384 __ LoadFromOffset(kLoadWord, out_reg, out_reg, offset); 6385 __ MaybeUnpoisonHeapReference(out_reg); 6386 } 6387} 6388 6389void InstructionCodeGeneratorARM::GenerateReferenceLoadTwoRegisters(HInstruction* instruction, 6390 Location out, 6391 Location obj, 6392 uint32_t offset, 6393 Location maybe_temp) { 6394 Register out_reg = out.AsRegister<Register>(); 6395 Register obj_reg = obj.AsRegister<Register>(); 6396 if (kEmitCompilerReadBarrier) { 6397 if (kUseBakerReadBarrier) { 6398 DCHECK(maybe_temp.IsRegister()) << maybe_temp; 6399 // Load with fast path based Baker's read barrier. 6400 // /* HeapReference<Object> */ out = *(obj + offset) 6401 codegen_->GenerateFieldLoadWithBakerReadBarrier( 6402 instruction, out, obj_reg, offset, maybe_temp, /* needs_null_check */ false); 6403 } else { 6404 // Load with slow path based read barrier. 6405 // /* HeapReference<Object> */ out = *(obj + offset) 6406 __ LoadFromOffset(kLoadWord, out_reg, obj_reg, offset); 6407 codegen_->GenerateReadBarrierSlow(instruction, out, out, obj, offset); 6408 } 6409 } else { 6410 // Plain load with no read barrier. 6411 // /* HeapReference<Object> */ out = *(obj + offset) 6412 __ LoadFromOffset(kLoadWord, out_reg, obj_reg, offset); 6413 __ MaybeUnpoisonHeapReference(out_reg); 6414 } 6415} 6416 6417void InstructionCodeGeneratorARM::GenerateGcRootFieldLoad(HInstruction* instruction, 6418 Location root, 6419 Register obj, 6420 uint32_t offset) { 6421 Register root_reg = root.AsRegister<Register>(); 6422 if (kEmitCompilerReadBarrier) { 6423 if (kUseBakerReadBarrier) { 6424 // Fast path implementation of art::ReadBarrier::BarrierForRoot when 6425 // Baker's read barrier are used: 6426 // 6427 // root = obj.field; 6428 // if (Thread::Current()->GetIsGcMarking()) { 6429 // root = ReadBarrier::Mark(root) 6430 // } 6431 6432 // /* GcRoot<mirror::Object> */ root = *(obj + offset) 6433 __ LoadFromOffset(kLoadWord, root_reg, obj, offset); 6434 static_assert( 6435 sizeof(mirror::CompressedReference<mirror::Object>) == sizeof(GcRoot<mirror::Object>), 6436 "art::mirror::CompressedReference<mirror::Object> and art::GcRoot<mirror::Object> " 6437 "have different sizes."); 6438 static_assert(sizeof(mirror::CompressedReference<mirror::Object>) == sizeof(int32_t), 6439 "art::mirror::CompressedReference<mirror::Object> and int32_t " 6440 "have different sizes."); 6441 6442 // Slow path marking the GC root `root`. 6443 SlowPathCode* slow_path = 6444 new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM(instruction, root); 6445 codegen_->AddSlowPath(slow_path); 6446 6447 // IP = Thread::Current()->GetIsGcMarking() 6448 __ LoadFromOffset( 6449 kLoadWord, IP, TR, Thread::IsGcMarkingOffset<kArmPointerSize>().Int32Value()); 6450 __ CompareAndBranchIfNonZero(IP, slow_path->GetEntryLabel()); 6451 __ Bind(slow_path->GetExitLabel()); 6452 } else { 6453 // GC root loaded through a slow path for read barriers other 6454 // than Baker's. 6455 // /* GcRoot<mirror::Object>* */ root = obj + offset 6456 __ AddConstant(root_reg, obj, offset); 6457 // /* mirror::Object* */ root = root->Read() 6458 codegen_->GenerateReadBarrierForRootSlow(instruction, root, root); 6459 } 6460 } else { 6461 // Plain GC root load with no read barrier. 6462 // /* GcRoot<mirror::Object> */ root = *(obj + offset) 6463 __ LoadFromOffset(kLoadWord, root_reg, obj, offset); 6464 // Note that GC roots are not affected by heap poisoning, thus we 6465 // do not have to unpoison `root_reg` here. 6466 } 6467} 6468 6469void CodeGeneratorARM::GenerateFieldLoadWithBakerReadBarrier(HInstruction* instruction, 6470 Location ref, 6471 Register obj, 6472 uint32_t offset, 6473 Location temp, 6474 bool needs_null_check) { 6475 DCHECK(kEmitCompilerReadBarrier); 6476 DCHECK(kUseBakerReadBarrier); 6477 6478 // /* HeapReference<Object> */ ref = *(obj + offset) 6479 Location no_index = Location::NoLocation(); 6480 ScaleFactor no_scale_factor = TIMES_1; 6481 GenerateReferenceLoadWithBakerReadBarrier( 6482 instruction, ref, obj, offset, no_index, no_scale_factor, temp, needs_null_check); 6483} 6484 6485void CodeGeneratorARM::GenerateArrayLoadWithBakerReadBarrier(HInstruction* instruction, 6486 Location ref, 6487 Register obj, 6488 uint32_t data_offset, 6489 Location index, 6490 Location temp, 6491 bool needs_null_check) { 6492 DCHECK(kEmitCompilerReadBarrier); 6493 DCHECK(kUseBakerReadBarrier); 6494 6495 static_assert( 6496 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 6497 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 6498 // /* HeapReference<Object> */ ref = 6499 // *(obj + data_offset + index * sizeof(HeapReference<Object>)) 6500 ScaleFactor scale_factor = TIMES_4; 6501 GenerateReferenceLoadWithBakerReadBarrier( 6502 instruction, ref, obj, data_offset, index, scale_factor, temp, needs_null_check); 6503} 6504 6505void CodeGeneratorARM::GenerateReferenceLoadWithBakerReadBarrier(HInstruction* instruction, 6506 Location ref, 6507 Register obj, 6508 uint32_t offset, 6509 Location index, 6510 ScaleFactor scale_factor, 6511 Location temp, 6512 bool needs_null_check) { 6513 DCHECK(kEmitCompilerReadBarrier); 6514 DCHECK(kUseBakerReadBarrier); 6515 6516 // In slow path based read barriers, the read barrier call is 6517 // inserted after the original load. However, in fast path based 6518 // Baker's read barriers, we need to perform the load of 6519 // mirror::Object::monitor_ *before* the original reference load. 6520 // This load-load ordering is required by the read barrier. 6521 // The fast path/slow path (for Baker's algorithm) should look like: 6522 // 6523 // uint32_t rb_state = Lockword(obj->monitor_).ReadBarrierState(); 6524 // lfence; // Load fence or artificial data dependency to prevent load-load reordering 6525 // HeapReference<Object> ref = *src; // Original reference load. 6526 // bool is_gray = (rb_state == ReadBarrier::gray_ptr_); 6527 // if (is_gray) { 6528 // ref = ReadBarrier::Mark(ref); // Performed by runtime entrypoint slow path. 6529 // } 6530 // 6531 // Note: the original implementation in ReadBarrier::Barrier is 6532 // slightly more complex as it performs additional checks that we do 6533 // not do here for performance reasons. 6534 6535 Register ref_reg = ref.AsRegister<Register>(); 6536 Register temp_reg = temp.AsRegister<Register>(); 6537 uint32_t monitor_offset = mirror::Object::MonitorOffset().Int32Value(); 6538 6539 // /* int32_t */ monitor = obj->monitor_ 6540 __ LoadFromOffset(kLoadWord, temp_reg, obj, monitor_offset); 6541 if (needs_null_check) { 6542 MaybeRecordImplicitNullCheck(instruction); 6543 } 6544 // /* LockWord */ lock_word = LockWord(monitor) 6545 static_assert(sizeof(LockWord) == sizeof(int32_t), 6546 "art::LockWord and int32_t have different sizes."); 6547 6548 // Introduce a dependency on the lock_word including the rb_state, 6549 // which shall prevent load-load reordering without using 6550 // a memory barrier (which would be more expensive). 6551 // `obj` is unchanged by this operation, but its value now depends 6552 // on `temp_reg`. 6553 __ add(obj, obj, ShifterOperand(temp_reg, LSR, 32)); 6554 6555 // The actual reference load. 6556 if (index.IsValid()) { 6557 // Load types involving an "index": ArrayGet and 6558 // UnsafeGetObject/UnsafeGetObjectVolatile intrinsics. 6559 // /* HeapReference<Object> */ ref = *(obj + offset + (index << scale_factor)) 6560 if (index.IsConstant()) { 6561 size_t computed_offset = 6562 (index.GetConstant()->AsIntConstant()->GetValue() << scale_factor) + offset; 6563 __ LoadFromOffset(kLoadWord, ref_reg, obj, computed_offset); 6564 } else { 6565 // Handle the special case of the 6566 // UnsafeGetObject/UnsafeGetObjectVolatile intrinsics, which use 6567 // a register pair as index ("long offset"), of which only the low 6568 // part contains data. 6569 Register index_reg = index.IsRegisterPair() 6570 ? index.AsRegisterPairLow<Register>() 6571 : index.AsRegister<Register>(); 6572 __ add(IP, obj, ShifterOperand(index_reg, LSL, scale_factor)); 6573 __ LoadFromOffset(kLoadWord, ref_reg, IP, offset); 6574 } 6575 } else { 6576 // /* HeapReference<Object> */ ref = *(obj + offset) 6577 __ LoadFromOffset(kLoadWord, ref_reg, obj, offset); 6578 } 6579 6580 // Object* ref = ref_addr->AsMirrorPtr() 6581 __ MaybeUnpoisonHeapReference(ref_reg); 6582 6583 // Slow path marking the object `ref` when it is gray. 6584 SlowPathCode* slow_path = 6585 new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM(instruction, ref); 6586 AddSlowPath(slow_path); 6587 6588 // if (rb_state == ReadBarrier::gray_ptr_) 6589 // ref = ReadBarrier::Mark(ref); 6590 // Given the numeric representation, it's enough to check the low bit of the 6591 // rb_state. We do that by shifting the bit out of the lock word with LSRS 6592 // which can be a 16-bit instruction unlike the TST immediate. 6593 static_assert(ReadBarrier::white_ptr_ == 0, "Expecting white to have value 0"); 6594 static_assert(ReadBarrier::gray_ptr_ == 1, "Expecting gray to have value 1"); 6595 static_assert(ReadBarrier::black_ptr_ == 2, "Expecting black to have value 2"); 6596 __ Lsrs(temp_reg, temp_reg, LockWord::kReadBarrierStateShift + 1); 6597 __ b(slow_path->GetEntryLabel(), CS); // Carry flag is the last bit shifted out by LSRS. 6598 __ Bind(slow_path->GetExitLabel()); 6599} 6600 6601void CodeGeneratorARM::GenerateReadBarrierSlow(HInstruction* instruction, 6602 Location out, 6603 Location ref, 6604 Location obj, 6605 uint32_t offset, 6606 Location index) { 6607 DCHECK(kEmitCompilerReadBarrier); 6608 6609 // Insert a slow path based read barrier *after* the reference load. 6610 // 6611 // If heap poisoning is enabled, the unpoisoning of the loaded 6612 // reference will be carried out by the runtime within the slow 6613 // path. 6614 // 6615 // Note that `ref` currently does not get unpoisoned (when heap 6616 // poisoning is enabled), which is alright as the `ref` argument is 6617 // not used by the artReadBarrierSlow entry point. 6618 // 6619 // TODO: Unpoison `ref` when it is used by artReadBarrierSlow. 6620 SlowPathCode* slow_path = new (GetGraph()->GetArena()) 6621 ReadBarrierForHeapReferenceSlowPathARM(instruction, out, ref, obj, offset, index); 6622 AddSlowPath(slow_path); 6623 6624 __ b(slow_path->GetEntryLabel()); 6625 __ Bind(slow_path->GetExitLabel()); 6626} 6627 6628void CodeGeneratorARM::MaybeGenerateReadBarrierSlow(HInstruction* instruction, 6629 Location out, 6630 Location ref, 6631 Location obj, 6632 uint32_t offset, 6633 Location index) { 6634 if (kEmitCompilerReadBarrier) { 6635 // Baker's read barriers shall be handled by the fast path 6636 // (CodeGeneratorARM::GenerateReferenceLoadWithBakerReadBarrier). 6637 DCHECK(!kUseBakerReadBarrier); 6638 // If heap poisoning is enabled, unpoisoning will be taken care of 6639 // by the runtime within the slow path. 6640 GenerateReadBarrierSlow(instruction, out, ref, obj, offset, index); 6641 } else if (kPoisonHeapReferences) { 6642 __ UnpoisonHeapReference(out.AsRegister<Register>()); 6643 } 6644} 6645 6646void CodeGeneratorARM::GenerateReadBarrierForRootSlow(HInstruction* instruction, 6647 Location out, 6648 Location root) { 6649 DCHECK(kEmitCompilerReadBarrier); 6650 6651 // Insert a slow path based read barrier *after* the GC root load. 6652 // 6653 // Note that GC roots are not affected by heap poisoning, so we do 6654 // not need to do anything special for this here. 6655 SlowPathCode* slow_path = 6656 new (GetGraph()->GetArena()) ReadBarrierForRootSlowPathARM(instruction, out, root); 6657 AddSlowPath(slow_path); 6658 6659 __ b(slow_path->GetEntryLabel()); 6660 __ Bind(slow_path->GetExitLabel()); 6661} 6662 6663HInvokeStaticOrDirect::DispatchInfo CodeGeneratorARM::GetSupportedInvokeStaticOrDirectDispatch( 6664 const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info, 6665 MethodReference target_method) { 6666 HInvokeStaticOrDirect::DispatchInfo dispatch_info = desired_dispatch_info; 6667 // We disable pc-relative load when there is an irreducible loop, as the optimization 6668 // is incompatible with it. 6669 // TODO: Create as many ArmDexCacheArraysBase instructions as needed for methods 6670 // with irreducible loops. 6671 if (GetGraph()->HasIrreducibleLoops() && 6672 (dispatch_info.method_load_kind == 6673 HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative)) { 6674 dispatch_info.method_load_kind = HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod; 6675 } 6676 6677 if (dispatch_info.code_ptr_location == HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative) { 6678 const DexFile& outer_dex_file = GetGraph()->GetDexFile(); 6679 if (&outer_dex_file != target_method.dex_file) { 6680 // Calls across dex files are more likely to exceed the available BL range, 6681 // so use absolute patch with fixup if available and kCallArtMethod otherwise. 6682 HInvokeStaticOrDirect::CodePtrLocation code_ptr_location = 6683 (desired_dispatch_info.method_load_kind == 6684 HInvokeStaticOrDirect::MethodLoadKind::kDirectAddressWithFixup) 6685 ? HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup 6686 : HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod; 6687 return HInvokeStaticOrDirect::DispatchInfo { 6688 dispatch_info.method_load_kind, 6689 code_ptr_location, 6690 dispatch_info.method_load_data, 6691 0u 6692 }; 6693 } 6694 } 6695 return dispatch_info; 6696} 6697 6698Register CodeGeneratorARM::GetInvokeStaticOrDirectExtraParameter(HInvokeStaticOrDirect* invoke, 6699 Register temp) { 6700 DCHECK_EQ(invoke->InputCount(), invoke->GetNumberOfArguments() + 1u); 6701 Location location = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 6702 if (!invoke->GetLocations()->Intrinsified()) { 6703 return location.AsRegister<Register>(); 6704 } 6705 // For intrinsics we allow any location, so it may be on the stack. 6706 if (!location.IsRegister()) { 6707 __ LoadFromOffset(kLoadWord, temp, SP, location.GetStackIndex()); 6708 return temp; 6709 } 6710 // For register locations, check if the register was saved. If so, get it from the stack. 6711 // Note: There is a chance that the register was saved but not overwritten, so we could 6712 // save one load. However, since this is just an intrinsic slow path we prefer this 6713 // simple and more robust approach rather that trying to determine if that's the case. 6714 SlowPathCode* slow_path = GetCurrentSlowPath(); 6715 DCHECK(slow_path != nullptr); // For intrinsified invokes the call is emitted on the slow path. 6716 if (slow_path->IsCoreRegisterSaved(location.AsRegister<Register>())) { 6717 int stack_offset = slow_path->GetStackOffsetOfCoreRegister(location.AsRegister<Register>()); 6718 __ LoadFromOffset(kLoadWord, temp, SP, stack_offset); 6719 return temp; 6720 } 6721 return location.AsRegister<Register>(); 6722} 6723 6724void CodeGeneratorARM::GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) { 6725 // For better instruction scheduling we load the direct code pointer before the method pointer. 6726 switch (invoke->GetCodePtrLocation()) { 6727 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup: 6728 // LR = code address from literal pool with link-time patch. 6729 __ LoadLiteral(LR, DeduplicateMethodCodeLiteral(invoke->GetTargetMethod())); 6730 break; 6731 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect: 6732 // LR = invoke->GetDirectCodePtr(); 6733 __ LoadImmediate(LR, invoke->GetDirectCodePtr()); 6734 break; 6735 default: 6736 break; 6737 } 6738 6739 Location callee_method = temp; // For all kinds except kRecursive, callee will be in temp. 6740 switch (invoke->GetMethodLoadKind()) { 6741 case HInvokeStaticOrDirect::MethodLoadKind::kStringInit: 6742 // temp = thread->string_init_entrypoint 6743 __ LoadFromOffset(kLoadWord, temp.AsRegister<Register>(), TR, invoke->GetStringInitOffset()); 6744 break; 6745 case HInvokeStaticOrDirect::MethodLoadKind::kRecursive: 6746 callee_method = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 6747 break; 6748 case HInvokeStaticOrDirect::MethodLoadKind::kDirectAddress: 6749 __ LoadImmediate(temp.AsRegister<Register>(), invoke->GetMethodAddress()); 6750 break; 6751 case HInvokeStaticOrDirect::MethodLoadKind::kDirectAddressWithFixup: 6752 __ LoadLiteral(temp.AsRegister<Register>(), 6753 DeduplicateMethodAddressLiteral(invoke->GetTargetMethod())); 6754 break; 6755 case HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative: { 6756 HArmDexCacheArraysBase* base = 6757 invoke->InputAt(invoke->GetSpecialInputIndex())->AsArmDexCacheArraysBase(); 6758 Register base_reg = GetInvokeStaticOrDirectExtraParameter(invoke, 6759 temp.AsRegister<Register>()); 6760 int32_t offset = invoke->GetDexCacheArrayOffset() - base->GetElementOffset(); 6761 __ LoadFromOffset(kLoadWord, temp.AsRegister<Register>(), base_reg, offset); 6762 break; 6763 } 6764 case HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod: { 6765 Location current_method = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 6766 Register method_reg; 6767 Register reg = temp.AsRegister<Register>(); 6768 if (current_method.IsRegister()) { 6769 method_reg = current_method.AsRegister<Register>(); 6770 } else { 6771 DCHECK(invoke->GetLocations()->Intrinsified()); 6772 DCHECK(!current_method.IsValid()); 6773 method_reg = reg; 6774 __ LoadFromOffset(kLoadWord, reg, SP, kCurrentMethodStackOffset); 6775 } 6776 // /* ArtMethod*[] */ temp = temp.ptr_sized_fields_->dex_cache_resolved_methods_; 6777 __ LoadFromOffset(kLoadWord, 6778 reg, 6779 method_reg, 6780 ArtMethod::DexCacheResolvedMethodsOffset(kArmPointerSize).Int32Value()); 6781 // temp = temp[index_in_cache]; 6782 // Note: Don't use invoke->GetTargetMethod() as it may point to a different dex file. 6783 uint32_t index_in_cache = invoke->GetDexMethodIndex(); 6784 __ LoadFromOffset(kLoadWord, reg, reg, CodeGenerator::GetCachePointerOffset(index_in_cache)); 6785 break; 6786 } 6787 } 6788 6789 switch (invoke->GetCodePtrLocation()) { 6790 case HInvokeStaticOrDirect::CodePtrLocation::kCallSelf: 6791 __ bl(GetFrameEntryLabel()); 6792 break; 6793 case HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative: 6794 relative_call_patches_.emplace_back(invoke->GetTargetMethod()); 6795 __ BindTrackedLabel(&relative_call_patches_.back().label); 6796 // Arbitrarily branch to the BL itself, override at link time. 6797 __ bl(&relative_call_patches_.back().label); 6798 break; 6799 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup: 6800 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect: 6801 // LR prepared above for better instruction scheduling. 6802 // LR() 6803 __ blx(LR); 6804 break; 6805 case HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod: 6806 // LR = callee_method->entry_point_from_quick_compiled_code_ 6807 __ LoadFromOffset( 6808 kLoadWord, LR, callee_method.AsRegister<Register>(), 6809 ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmPointerSize).Int32Value()); 6810 // LR() 6811 __ blx(LR); 6812 break; 6813 } 6814 6815 DCHECK(!IsLeafMethod()); 6816} 6817 6818void CodeGeneratorARM::GenerateVirtualCall(HInvokeVirtual* invoke, Location temp_location) { 6819 Register temp = temp_location.AsRegister<Register>(); 6820 uint32_t method_offset = mirror::Class::EmbeddedVTableEntryOffset( 6821 invoke->GetVTableIndex(), kArmPointerSize).Uint32Value(); 6822 6823 // Use the calling convention instead of the location of the receiver, as 6824 // intrinsics may have put the receiver in a different register. In the intrinsics 6825 // slow path, the arguments have been moved to the right place, so here we are 6826 // guaranteed that the receiver is the first register of the calling convention. 6827 InvokeDexCallingConvention calling_convention; 6828 Register receiver = calling_convention.GetRegisterAt(0); 6829 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 6830 // /* HeapReference<Class> */ temp = receiver->klass_ 6831 __ LoadFromOffset(kLoadWord, temp, receiver, class_offset); 6832 MaybeRecordImplicitNullCheck(invoke); 6833 // Instead of simply (possibly) unpoisoning `temp` here, we should 6834 // emit a read barrier for the previous class reference load. 6835 // However this is not required in practice, as this is an 6836 // intermediate/temporary reference and because the current 6837 // concurrent copying collector keeps the from-space memory 6838 // intact/accessible until the end of the marking phase (the 6839 // concurrent copying collector may not in the future). 6840 __ MaybeUnpoisonHeapReference(temp); 6841 // temp = temp->GetMethodAt(method_offset); 6842 uint32_t entry_point = ArtMethod::EntryPointFromQuickCompiledCodeOffset( 6843 kArmPointerSize).Int32Value(); 6844 __ LoadFromOffset(kLoadWord, temp, temp, method_offset); 6845 // LR = temp->GetEntryPoint(); 6846 __ LoadFromOffset(kLoadWord, LR, temp, entry_point); 6847 // LR(); 6848 __ blx(LR); 6849} 6850 6851CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewPcRelativeStringPatch( 6852 const DexFile& dex_file, uint32_t string_index) { 6853 return NewPcRelativePatch(dex_file, string_index, &pc_relative_string_patches_); 6854} 6855 6856CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewPcRelativeTypePatch( 6857 const DexFile& dex_file, uint32_t type_index) { 6858 return NewPcRelativePatch(dex_file, type_index, &pc_relative_type_patches_); 6859} 6860 6861CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewPcRelativeDexCacheArrayPatch( 6862 const DexFile& dex_file, uint32_t element_offset) { 6863 return NewPcRelativePatch(dex_file, element_offset, &pc_relative_dex_cache_patches_); 6864} 6865 6866CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewPcRelativePatch( 6867 const DexFile& dex_file, uint32_t offset_or_index, ArenaDeque<PcRelativePatchInfo>* patches) { 6868 patches->emplace_back(dex_file, offset_or_index); 6869 return &patches->back(); 6870} 6871 6872Literal* CodeGeneratorARM::DeduplicateBootImageStringLiteral(const DexFile& dex_file, 6873 uint32_t string_index) { 6874 return boot_image_string_patches_.GetOrCreate( 6875 StringReference(&dex_file, string_index), 6876 [this]() { return __ NewLiteral<uint32_t>(/* placeholder */ 0u); }); 6877} 6878 6879Literal* CodeGeneratorARM::DeduplicateBootImageTypeLiteral(const DexFile& dex_file, 6880 uint32_t type_index) { 6881 return boot_image_type_patches_.GetOrCreate( 6882 TypeReference(&dex_file, type_index), 6883 [this]() { return __ NewLiteral<uint32_t>(/* placeholder */ 0u); }); 6884} 6885 6886Literal* CodeGeneratorARM::DeduplicateBootImageAddressLiteral(uint32_t address) { 6887 bool needs_patch = GetCompilerOptions().GetIncludePatchInformation(); 6888 Uint32ToLiteralMap* map = needs_patch ? &boot_image_address_patches_ : &uint32_literals_; 6889 return DeduplicateUint32Literal(dchecked_integral_cast<uint32_t>(address), map); 6890} 6891 6892Literal* CodeGeneratorARM::DeduplicateDexCacheAddressLiteral(uint32_t address) { 6893 return DeduplicateUint32Literal(address, &uint32_literals_); 6894} 6895 6896void CodeGeneratorARM::EmitLinkerPatches(ArenaVector<LinkerPatch>* linker_patches) { 6897 DCHECK(linker_patches->empty()); 6898 size_t size = 6899 method_patches_.size() + 6900 call_patches_.size() + 6901 relative_call_patches_.size() + 6902 /* MOVW+MOVT for each base */ 2u * pc_relative_dex_cache_patches_.size() + 6903 boot_image_string_patches_.size() + 6904 /* MOVW+MOVT for each base */ 2u * pc_relative_string_patches_.size() + 6905 boot_image_type_patches_.size() + 6906 /* MOVW+MOVT for each base */ 2u * pc_relative_type_patches_.size() + 6907 boot_image_address_patches_.size(); 6908 linker_patches->reserve(size); 6909 for (const auto& entry : method_patches_) { 6910 const MethodReference& target_method = entry.first; 6911 Literal* literal = entry.second; 6912 DCHECK(literal->GetLabel()->IsBound()); 6913 uint32_t literal_offset = literal->GetLabel()->Position(); 6914 linker_patches->push_back(LinkerPatch::MethodPatch(literal_offset, 6915 target_method.dex_file, 6916 target_method.dex_method_index)); 6917 } 6918 for (const auto& entry : call_patches_) { 6919 const MethodReference& target_method = entry.first; 6920 Literal* literal = entry.second; 6921 DCHECK(literal->GetLabel()->IsBound()); 6922 uint32_t literal_offset = literal->GetLabel()->Position(); 6923 linker_patches->push_back(LinkerPatch::CodePatch(literal_offset, 6924 target_method.dex_file, 6925 target_method.dex_method_index)); 6926 } 6927 for (const MethodPatchInfo<Label>& info : relative_call_patches_) { 6928 uint32_t literal_offset = info.label.Position(); 6929 linker_patches->push_back(LinkerPatch::RelativeCodePatch(literal_offset, 6930 info.target_method.dex_file, 6931 info.target_method.dex_method_index)); 6932 } 6933 for (const PcRelativePatchInfo& info : pc_relative_dex_cache_patches_) { 6934 const DexFile& dex_file = info.target_dex_file; 6935 size_t base_element_offset = info.offset_or_index; 6936 DCHECK(info.add_pc_label.IsBound()); 6937 uint32_t add_pc_offset = dchecked_integral_cast<uint32_t>(info.add_pc_label.Position()); 6938 // Add MOVW patch. 6939 DCHECK(info.movw_label.IsBound()); 6940 uint32_t movw_offset = dchecked_integral_cast<uint32_t>(info.movw_label.Position()); 6941 linker_patches->push_back(LinkerPatch::DexCacheArrayPatch(movw_offset, 6942 &dex_file, 6943 add_pc_offset, 6944 base_element_offset)); 6945 // Add MOVT patch. 6946 DCHECK(info.movt_label.IsBound()); 6947 uint32_t movt_offset = dchecked_integral_cast<uint32_t>(info.movt_label.Position()); 6948 linker_patches->push_back(LinkerPatch::DexCacheArrayPatch(movt_offset, 6949 &dex_file, 6950 add_pc_offset, 6951 base_element_offset)); 6952 } 6953 for (const auto& entry : boot_image_string_patches_) { 6954 const StringReference& target_string = entry.first; 6955 Literal* literal = entry.second; 6956 DCHECK(literal->GetLabel()->IsBound()); 6957 uint32_t literal_offset = literal->GetLabel()->Position(); 6958 linker_patches->push_back(LinkerPatch::StringPatch(literal_offset, 6959 target_string.dex_file, 6960 target_string.string_index)); 6961 } 6962 for (const PcRelativePatchInfo& info : pc_relative_string_patches_) { 6963 const DexFile& dex_file = info.target_dex_file; 6964 uint32_t string_index = info.offset_or_index; 6965 DCHECK(info.add_pc_label.IsBound()); 6966 uint32_t add_pc_offset = dchecked_integral_cast<uint32_t>(info.add_pc_label.Position()); 6967 // Add MOVW patch. 6968 DCHECK(info.movw_label.IsBound()); 6969 uint32_t movw_offset = dchecked_integral_cast<uint32_t>(info.movw_label.Position()); 6970 linker_patches->push_back(LinkerPatch::RelativeStringPatch(movw_offset, 6971 &dex_file, 6972 add_pc_offset, 6973 string_index)); 6974 // Add MOVT patch. 6975 DCHECK(info.movt_label.IsBound()); 6976 uint32_t movt_offset = dchecked_integral_cast<uint32_t>(info.movt_label.Position()); 6977 linker_patches->push_back(LinkerPatch::RelativeStringPatch(movt_offset, 6978 &dex_file, 6979 add_pc_offset, 6980 string_index)); 6981 } 6982 for (const auto& entry : boot_image_type_patches_) { 6983 const TypeReference& target_type = entry.first; 6984 Literal* literal = entry.second; 6985 DCHECK(literal->GetLabel()->IsBound()); 6986 uint32_t literal_offset = literal->GetLabel()->Position(); 6987 linker_patches->push_back(LinkerPatch::TypePatch(literal_offset, 6988 target_type.dex_file, 6989 target_type.type_index)); 6990 } 6991 for (const PcRelativePatchInfo& info : pc_relative_type_patches_) { 6992 const DexFile& dex_file = info.target_dex_file; 6993 uint32_t type_index = info.offset_or_index; 6994 DCHECK(info.add_pc_label.IsBound()); 6995 uint32_t add_pc_offset = dchecked_integral_cast<uint32_t>(info.add_pc_label.Position()); 6996 // Add MOVW patch. 6997 DCHECK(info.movw_label.IsBound()); 6998 uint32_t movw_offset = dchecked_integral_cast<uint32_t>(info.movw_label.Position()); 6999 linker_patches->push_back(LinkerPatch::RelativeTypePatch(movw_offset, 7000 &dex_file, 7001 add_pc_offset, 7002 type_index)); 7003 // Add MOVT patch. 7004 DCHECK(info.movt_label.IsBound()); 7005 uint32_t movt_offset = dchecked_integral_cast<uint32_t>(info.movt_label.Position()); 7006 linker_patches->push_back(LinkerPatch::RelativeTypePatch(movt_offset, 7007 &dex_file, 7008 add_pc_offset, 7009 type_index)); 7010 } 7011 for (const auto& entry : boot_image_address_patches_) { 7012 DCHECK(GetCompilerOptions().GetIncludePatchInformation()); 7013 Literal* literal = entry.second; 7014 DCHECK(literal->GetLabel()->IsBound()); 7015 uint32_t literal_offset = literal->GetLabel()->Position(); 7016 linker_patches->push_back(LinkerPatch::RecordPosition(literal_offset)); 7017 } 7018} 7019 7020Literal* CodeGeneratorARM::DeduplicateUint32Literal(uint32_t value, Uint32ToLiteralMap* map) { 7021 return map->GetOrCreate( 7022 value, 7023 [this, value]() { return __ NewLiteral<uint32_t>(value); }); 7024} 7025 7026Literal* CodeGeneratorARM::DeduplicateMethodLiteral(MethodReference target_method, 7027 MethodToLiteralMap* map) { 7028 return map->GetOrCreate( 7029 target_method, 7030 [this]() { return __ NewLiteral<uint32_t>(/* placeholder */ 0u); }); 7031} 7032 7033Literal* CodeGeneratorARM::DeduplicateMethodAddressLiteral(MethodReference target_method) { 7034 return DeduplicateMethodLiteral(target_method, &method_patches_); 7035} 7036 7037Literal* CodeGeneratorARM::DeduplicateMethodCodeLiteral(MethodReference target_method) { 7038 return DeduplicateMethodLiteral(target_method, &call_patches_); 7039} 7040 7041void LocationsBuilderARM::VisitMultiplyAccumulate(HMultiplyAccumulate* instr) { 7042 LocationSummary* locations = 7043 new (GetGraph()->GetArena()) LocationSummary(instr, LocationSummary::kNoCall); 7044 locations->SetInAt(HMultiplyAccumulate::kInputAccumulatorIndex, 7045 Location::RequiresRegister()); 7046 locations->SetInAt(HMultiplyAccumulate::kInputMulLeftIndex, Location::RequiresRegister()); 7047 locations->SetInAt(HMultiplyAccumulate::kInputMulRightIndex, Location::RequiresRegister()); 7048 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 7049} 7050 7051void InstructionCodeGeneratorARM::VisitMultiplyAccumulate(HMultiplyAccumulate* instr) { 7052 LocationSummary* locations = instr->GetLocations(); 7053 Register res = locations->Out().AsRegister<Register>(); 7054 Register accumulator = 7055 locations->InAt(HMultiplyAccumulate::kInputAccumulatorIndex).AsRegister<Register>(); 7056 Register mul_left = 7057 locations->InAt(HMultiplyAccumulate::kInputMulLeftIndex).AsRegister<Register>(); 7058 Register mul_right = 7059 locations->InAt(HMultiplyAccumulate::kInputMulRightIndex).AsRegister<Register>(); 7060 7061 if (instr->GetOpKind() == HInstruction::kAdd) { 7062 __ mla(res, mul_left, mul_right, accumulator); 7063 } else { 7064 __ mls(res, mul_left, mul_right, accumulator); 7065 } 7066} 7067 7068void LocationsBuilderARM::VisitBoundType(HBoundType* instruction ATTRIBUTE_UNUSED) { 7069 // Nothing to do, this should be removed during prepare for register allocator. 7070 LOG(FATAL) << "Unreachable"; 7071} 7072 7073void InstructionCodeGeneratorARM::VisitBoundType(HBoundType* instruction ATTRIBUTE_UNUSED) { 7074 // Nothing to do, this should be removed during prepare for register allocator. 7075 LOG(FATAL) << "Unreachable"; 7076} 7077 7078// Simple implementation of packed switch - generate cascaded compare/jumps. 7079void LocationsBuilderARM::VisitPackedSwitch(HPackedSwitch* switch_instr) { 7080 LocationSummary* locations = 7081 new (GetGraph()->GetArena()) LocationSummary(switch_instr, LocationSummary::kNoCall); 7082 locations->SetInAt(0, Location::RequiresRegister()); 7083 if (switch_instr->GetNumEntries() > kPackedSwitchCompareJumpThreshold && 7084 codegen_->GetAssembler()->IsThumb()) { 7085 locations->AddTemp(Location::RequiresRegister()); // We need a temp for the table base. 7086 if (switch_instr->GetStartValue() != 0) { 7087 locations->AddTemp(Location::RequiresRegister()); // We need a temp for the bias. 7088 } 7089 } 7090} 7091 7092void InstructionCodeGeneratorARM::VisitPackedSwitch(HPackedSwitch* switch_instr) { 7093 int32_t lower_bound = switch_instr->GetStartValue(); 7094 uint32_t num_entries = switch_instr->GetNumEntries(); 7095 LocationSummary* locations = switch_instr->GetLocations(); 7096 Register value_reg = locations->InAt(0).AsRegister<Register>(); 7097 HBasicBlock* default_block = switch_instr->GetDefaultBlock(); 7098 7099 if (num_entries <= kPackedSwitchCompareJumpThreshold || !codegen_->GetAssembler()->IsThumb()) { 7100 // Create a series of compare/jumps. 7101 Register temp_reg = IP; 7102 // Note: It is fine for the below AddConstantSetFlags() using IP register to temporarily store 7103 // the immediate, because IP is used as the destination register. For the other 7104 // AddConstantSetFlags() and GenerateCompareWithImmediate(), the immediate values are constant, 7105 // and they can be encoded in the instruction without making use of IP register. 7106 __ AddConstantSetFlags(temp_reg, value_reg, -lower_bound); 7107 7108 const ArenaVector<HBasicBlock*>& successors = switch_instr->GetBlock()->GetSuccessors(); 7109 // Jump to successors[0] if value == lower_bound. 7110 __ b(codegen_->GetLabelOf(successors[0]), EQ); 7111 int32_t last_index = 0; 7112 for (; num_entries - last_index > 2; last_index += 2) { 7113 __ AddConstantSetFlags(temp_reg, temp_reg, -2); 7114 // Jump to successors[last_index + 1] if value < case_value[last_index + 2]. 7115 __ b(codegen_->GetLabelOf(successors[last_index + 1]), LO); 7116 // Jump to successors[last_index + 2] if value == case_value[last_index + 2]. 7117 __ b(codegen_->GetLabelOf(successors[last_index + 2]), EQ); 7118 } 7119 if (num_entries - last_index == 2) { 7120 // The last missing case_value. 7121 __ CmpConstant(temp_reg, 1); 7122 __ b(codegen_->GetLabelOf(successors[last_index + 1]), EQ); 7123 } 7124 7125 // And the default for any other value. 7126 if (!codegen_->GoesToNextBlock(switch_instr->GetBlock(), default_block)) { 7127 __ b(codegen_->GetLabelOf(default_block)); 7128 } 7129 } else { 7130 // Create a table lookup. 7131 Register temp_reg = locations->GetTemp(0).AsRegister<Register>(); 7132 7133 // Materialize a pointer to the switch table 7134 std::vector<Label*> labels(num_entries); 7135 const ArenaVector<HBasicBlock*>& successors = switch_instr->GetBlock()->GetSuccessors(); 7136 for (uint32_t i = 0; i < num_entries; i++) { 7137 labels[i] = codegen_->GetLabelOf(successors[i]); 7138 } 7139 JumpTable* table = __ CreateJumpTable(std::move(labels), temp_reg); 7140 7141 // Remove the bias. 7142 Register key_reg; 7143 if (lower_bound != 0) { 7144 key_reg = locations->GetTemp(1).AsRegister<Register>(); 7145 __ AddConstant(key_reg, value_reg, -lower_bound); 7146 } else { 7147 key_reg = value_reg; 7148 } 7149 7150 // Check whether the value is in the table, jump to default block if not. 7151 __ CmpConstant(key_reg, num_entries - 1); 7152 __ b(codegen_->GetLabelOf(default_block), Condition::HI); 7153 7154 // Load the displacement from the table. 7155 __ ldr(temp_reg, Address(temp_reg, key_reg, Shift::LSL, 2)); 7156 7157 // Dispatch is a direct add to the PC (for Thumb2). 7158 __ EmitJumpTableDispatch(table, temp_reg); 7159 } 7160} 7161 7162void LocationsBuilderARM::VisitArmDexCacheArraysBase(HArmDexCacheArraysBase* base) { 7163 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(base); 7164 locations->SetOut(Location::RequiresRegister()); 7165} 7166 7167void InstructionCodeGeneratorARM::VisitArmDexCacheArraysBase(HArmDexCacheArraysBase* base) { 7168 Register base_reg = base->GetLocations()->Out().AsRegister<Register>(); 7169 CodeGeneratorARM::PcRelativePatchInfo* labels = 7170 codegen_->NewPcRelativeDexCacheArrayPatch(base->GetDexFile(), base->GetElementOffset()); 7171 __ BindTrackedLabel(&labels->movw_label); 7172 __ movw(base_reg, /* placeholder */ 0u); 7173 __ BindTrackedLabel(&labels->movt_label); 7174 __ movt(base_reg, /* placeholder */ 0u); 7175 __ BindTrackedLabel(&labels->add_pc_label); 7176 __ add(base_reg, base_reg, ShifterOperand(PC)); 7177} 7178 7179void CodeGeneratorARM::MoveFromReturnRegister(Location trg, Primitive::Type type) { 7180 if (!trg.IsValid()) { 7181 DCHECK_EQ(type, Primitive::kPrimVoid); 7182 return; 7183 } 7184 7185 DCHECK_NE(type, Primitive::kPrimVoid); 7186 7187 Location return_loc = InvokeDexCallingConventionVisitorARM().GetReturnLocation(type); 7188 if (return_loc.Equals(trg)) { 7189 return; 7190 } 7191 7192 // TODO: Consider pairs in the parallel move resolver, then this could be nicely merged 7193 // with the last branch. 7194 if (type == Primitive::kPrimLong) { 7195 HParallelMove parallel_move(GetGraph()->GetArena()); 7196 parallel_move.AddMove(return_loc.ToLow(), trg.ToLow(), Primitive::kPrimInt, nullptr); 7197 parallel_move.AddMove(return_loc.ToHigh(), trg.ToHigh(), Primitive::kPrimInt, nullptr); 7198 GetMoveResolver()->EmitNativeCode(¶llel_move); 7199 } else if (type == Primitive::kPrimDouble) { 7200 HParallelMove parallel_move(GetGraph()->GetArena()); 7201 parallel_move.AddMove(return_loc.ToLow(), trg.ToLow(), Primitive::kPrimFloat, nullptr); 7202 parallel_move.AddMove(return_loc.ToHigh(), trg.ToHigh(), Primitive::kPrimFloat, nullptr); 7203 GetMoveResolver()->EmitNativeCode(¶llel_move); 7204 } else { 7205 // Let the parallel move resolver take care of all of this. 7206 HParallelMove parallel_move(GetGraph()->GetArena()); 7207 parallel_move.AddMove(return_loc, trg, type, nullptr); 7208 GetMoveResolver()->EmitNativeCode(¶llel_move); 7209 } 7210} 7211 7212void LocationsBuilderARM::VisitClassTableGet(HClassTableGet* instruction) { 7213 LocationSummary* locations = 7214 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 7215 locations->SetInAt(0, Location::RequiresRegister()); 7216 locations->SetOut(Location::RequiresRegister()); 7217} 7218 7219void InstructionCodeGeneratorARM::VisitClassTableGet(HClassTableGet* instruction) { 7220 LocationSummary* locations = instruction->GetLocations(); 7221 if (instruction->GetTableKind() == HClassTableGet::TableKind::kVTable) { 7222 uint32_t method_offset = mirror::Class::EmbeddedVTableEntryOffset( 7223 instruction->GetIndex(), kArmPointerSize).SizeValue(); 7224 __ LoadFromOffset(kLoadWord, 7225 locations->Out().AsRegister<Register>(), 7226 locations->InAt(0).AsRegister<Register>(), 7227 method_offset); 7228 } else { 7229 uint32_t method_offset = static_cast<uint32_t>(ImTable::OffsetOfElement( 7230 instruction->GetIndex(), kArmPointerSize)); 7231 __ LoadFromOffset(kLoadWord, 7232 locations->Out().AsRegister<Register>(), 7233 locations->InAt(0).AsRegister<Register>(), 7234 mirror::Class::ImtPtrOffset(kArmPointerSize).Uint32Value()); 7235 __ LoadFromOffset(kLoadWord, 7236 locations->Out().AsRegister<Register>(), 7237 locations->Out().AsRegister<Register>(), 7238 method_offset); 7239 } 7240} 7241 7242#undef __ 7243#undef QUICK_ENTRY_POINT 7244 7245} // namespace arm 7246} // namespace art 7247