code_generator_arm.cc revision 07c919feccdf47f997842a131a802aa6b891e34a
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 66static constexpr int kRegListThreshold = 4; 67 68// SaveLiveRegisters and RestoreLiveRegisters from SlowPathCodeARM operate on sets of S registers, 69// for each live D registers they treat two corresponding S registers as live ones. 70// 71// Two following functions (SaveContiguousSRegisterList, RestoreContiguousSRegisterList) build 72// from a list of contiguous S registers a list of contiguous D registers (processing first/last 73// S registers corner cases) and save/restore this new list treating them as D registers. 74// - decreasing code size 75// - avoiding hazards on Cortex-A57, when a pair of S registers for an actual live D register is 76// restored and then used in regular non SlowPath code as D register. 77// 78// For the following example (v means the S register is live): 79// D names: | D0 | D1 | D2 | D4 | ... 80// S names: | S0 | S1 | S2 | S3 | S4 | S5 | S6 | S7 | ... 81// Live? | | v | v | v | v | v | v | | ... 82// 83// S1 and S6 will be saved/restored independently; D registers list (D1, D2) will be processed 84// as D registers. 85static size_t SaveContiguousSRegisterList(size_t first, 86 size_t last, 87 CodeGenerator* codegen, 88 size_t stack_offset) { 89 DCHECK_LE(first, last); 90 if ((first == last) && (first == 0)) { 91 stack_offset += codegen->SaveFloatingPointRegister(stack_offset, first); 92 return stack_offset; 93 } 94 if (first % 2 == 1) { 95 stack_offset += codegen->SaveFloatingPointRegister(stack_offset, first++); 96 } 97 98 bool save_last = false; 99 if (last % 2 == 0) { 100 save_last = true; 101 --last; 102 } 103 104 if (first < last) { 105 DRegister d_reg = static_cast<DRegister>(first / 2); 106 DCHECK_EQ((last - first + 1) % 2, 0u); 107 size_t number_of_d_regs = (last - first + 1) / 2; 108 109 if (number_of_d_regs == 1) { 110 __ StoreDToOffset(d_reg, SP, stack_offset); 111 } else if (number_of_d_regs > 1) { 112 __ add(IP, SP, ShifterOperand(stack_offset)); 113 __ vstmiad(IP, d_reg, number_of_d_regs); 114 } 115 stack_offset += number_of_d_regs * kArmWordSize * 2; 116 } 117 118 if (save_last) { 119 stack_offset += codegen->SaveFloatingPointRegister(stack_offset, last + 1); 120 } 121 122 return stack_offset; 123} 124 125static size_t RestoreContiguousSRegisterList(size_t first, 126 size_t last, 127 CodeGenerator* codegen, 128 size_t stack_offset) { 129 DCHECK_LE(first, last); 130 if ((first == last) && (first == 0)) { 131 stack_offset += codegen->RestoreFloatingPointRegister(stack_offset, first); 132 return stack_offset; 133 } 134 if (first % 2 == 1) { 135 stack_offset += codegen->RestoreFloatingPointRegister(stack_offset, first++); 136 } 137 138 bool restore_last = false; 139 if (last % 2 == 0) { 140 restore_last = true; 141 --last; 142 } 143 144 if (first < last) { 145 DRegister d_reg = static_cast<DRegister>(first / 2); 146 DCHECK_EQ((last - first + 1) % 2, 0u); 147 size_t number_of_d_regs = (last - first + 1) / 2; 148 if (number_of_d_regs == 1) { 149 __ LoadDFromOffset(d_reg, SP, stack_offset); 150 } else if (number_of_d_regs > 1) { 151 __ add(IP, SP, ShifterOperand(stack_offset)); 152 __ vldmiad(IP, d_reg, number_of_d_regs); 153 } 154 stack_offset += number_of_d_regs * kArmWordSize * 2; 155 } 156 157 if (restore_last) { 158 stack_offset += codegen->RestoreFloatingPointRegister(stack_offset, last + 1); 159 } 160 161 return stack_offset; 162} 163 164void SlowPathCodeARM::SaveLiveRegisters(CodeGenerator* codegen, LocationSummary* locations) { 165 size_t stack_offset = codegen->GetFirstRegisterSlotInSlowPath(); 166 size_t orig_offset = stack_offset; 167 168 const uint32_t core_spills = codegen->GetSlowPathSpills(locations, /* core_registers */ true); 169 for (uint32_t i : LowToHighBits(core_spills)) { 170 // If the register holds an object, update the stack mask. 171 if (locations->RegisterContainsObject(i)) { 172 locations->SetStackBit(stack_offset / kVRegSize); 173 } 174 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 175 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 176 saved_core_stack_offsets_[i] = stack_offset; 177 stack_offset += kArmWordSize; 178 } 179 180 int reg_num = POPCOUNT(core_spills); 181 if (reg_num != 0) { 182 if (reg_num > kRegListThreshold) { 183 __ StoreList(RegList(core_spills), orig_offset); 184 } else { 185 stack_offset = orig_offset; 186 for (uint32_t i : LowToHighBits(core_spills)) { 187 stack_offset += codegen->SaveCoreRegister(stack_offset, i); 188 } 189 } 190 } 191 192 uint32_t fp_spills = codegen->GetSlowPathSpills(locations, /* core_registers */ false); 193 orig_offset = stack_offset; 194 for (uint32_t i : LowToHighBits(fp_spills)) { 195 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 196 saved_fpu_stack_offsets_[i] = stack_offset; 197 stack_offset += kArmWordSize; 198 } 199 200 stack_offset = orig_offset; 201 while (fp_spills != 0u) { 202 uint32_t begin = CTZ(fp_spills); 203 uint32_t tmp = fp_spills + (1u << begin); 204 fp_spills &= tmp; // Clear the contiguous range of 1s. 205 uint32_t end = (tmp == 0u) ? 32u : CTZ(tmp); // CTZ(0) is undefined. 206 stack_offset = SaveContiguousSRegisterList(begin, end - 1, codegen, stack_offset); 207 } 208 DCHECK_LE(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 209} 210 211void SlowPathCodeARM::RestoreLiveRegisters(CodeGenerator* codegen, LocationSummary* locations) { 212 size_t stack_offset = codegen->GetFirstRegisterSlotInSlowPath(); 213 size_t orig_offset = stack_offset; 214 215 const uint32_t core_spills = codegen->GetSlowPathSpills(locations, /* core_registers */ true); 216 for (uint32_t i : LowToHighBits(core_spills)) { 217 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 218 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 219 stack_offset += kArmWordSize; 220 } 221 222 int reg_num = POPCOUNT(core_spills); 223 if (reg_num != 0) { 224 if (reg_num > kRegListThreshold) { 225 __ LoadList(RegList(core_spills), orig_offset); 226 } else { 227 stack_offset = orig_offset; 228 for (uint32_t i : LowToHighBits(core_spills)) { 229 stack_offset += codegen->RestoreCoreRegister(stack_offset, i); 230 } 231 } 232 } 233 234 uint32_t fp_spills = codegen->GetSlowPathSpills(locations, /* core_registers */ false); 235 while (fp_spills != 0u) { 236 uint32_t begin = CTZ(fp_spills); 237 uint32_t tmp = fp_spills + (1u << begin); 238 fp_spills &= tmp; // Clear the contiguous range of 1s. 239 uint32_t end = (tmp == 0u) ? 32u : CTZ(tmp); // CTZ(0) is undefined. 240 stack_offset = RestoreContiguousSRegisterList(begin, end - 1, codegen, stack_offset); 241 } 242 DCHECK_LE(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 243} 244 245class NullCheckSlowPathARM : public SlowPathCodeARM { 246 public: 247 explicit NullCheckSlowPathARM(HNullCheck* instruction) : SlowPathCodeARM(instruction) {} 248 249 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 250 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 251 __ Bind(GetEntryLabel()); 252 if (instruction_->CanThrowIntoCatchBlock()) { 253 // Live registers will be restored in the catch block if caught. 254 SaveLiveRegisters(codegen, instruction_->GetLocations()); 255 } 256 arm_codegen->InvokeRuntime(kQuickThrowNullPointer, 257 instruction_, 258 instruction_->GetDexPc(), 259 this); 260 CheckEntrypointTypes<kQuickThrowNullPointer, void, void>(); 261 } 262 263 bool IsFatal() const OVERRIDE { return true; } 264 265 const char* GetDescription() const OVERRIDE { return "NullCheckSlowPathARM"; } 266 267 private: 268 DISALLOW_COPY_AND_ASSIGN(NullCheckSlowPathARM); 269}; 270 271class DivZeroCheckSlowPathARM : public SlowPathCodeARM { 272 public: 273 explicit DivZeroCheckSlowPathARM(HDivZeroCheck* instruction) : SlowPathCodeARM(instruction) {} 274 275 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 276 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 277 __ Bind(GetEntryLabel()); 278 arm_codegen->InvokeRuntime(kQuickThrowDivZero, instruction_, instruction_->GetDexPc(), this); 279 CheckEntrypointTypes<kQuickThrowDivZero, void, void>(); 280 } 281 282 bool IsFatal() const OVERRIDE { return true; } 283 284 const char* GetDescription() const OVERRIDE { return "DivZeroCheckSlowPathARM"; } 285 286 private: 287 DISALLOW_COPY_AND_ASSIGN(DivZeroCheckSlowPathARM); 288}; 289 290class SuspendCheckSlowPathARM : public SlowPathCodeARM { 291 public: 292 SuspendCheckSlowPathARM(HSuspendCheck* instruction, HBasicBlock* successor) 293 : SlowPathCodeARM(instruction), successor_(successor) {} 294 295 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 296 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 297 __ Bind(GetEntryLabel()); 298 arm_codegen->InvokeRuntime(kQuickTestSuspend, instruction_, instruction_->GetDexPc(), this); 299 CheckEntrypointTypes<kQuickTestSuspend, void, void>(); 300 if (successor_ == nullptr) { 301 __ b(GetReturnLabel()); 302 } else { 303 __ b(arm_codegen->GetLabelOf(successor_)); 304 } 305 } 306 307 Label* GetReturnLabel() { 308 DCHECK(successor_ == nullptr); 309 return &return_label_; 310 } 311 312 HBasicBlock* GetSuccessor() const { 313 return successor_; 314 } 315 316 const char* GetDescription() const OVERRIDE { return "SuspendCheckSlowPathARM"; } 317 318 private: 319 // If not null, the block to branch to after the suspend check. 320 HBasicBlock* const successor_; 321 322 // If `successor_` is null, the label to branch to after the suspend check. 323 Label return_label_; 324 325 DISALLOW_COPY_AND_ASSIGN(SuspendCheckSlowPathARM); 326}; 327 328class BoundsCheckSlowPathARM : public SlowPathCodeARM { 329 public: 330 explicit BoundsCheckSlowPathARM(HBoundsCheck* instruction) 331 : SlowPathCodeARM(instruction) {} 332 333 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 334 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 335 LocationSummary* locations = instruction_->GetLocations(); 336 337 __ Bind(GetEntryLabel()); 338 if (instruction_->CanThrowIntoCatchBlock()) { 339 // Live registers will be restored in the catch block if caught. 340 SaveLiveRegisters(codegen, instruction_->GetLocations()); 341 } 342 // We're moving two locations to locations that could overlap, so we need a parallel 343 // move resolver. 344 InvokeRuntimeCallingConvention calling_convention; 345 codegen->EmitParallelMoves( 346 locations->InAt(0), 347 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 348 Primitive::kPrimInt, 349 locations->InAt(1), 350 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 351 Primitive::kPrimInt); 352 QuickEntrypointEnum entrypoint = instruction_->AsBoundsCheck()->IsStringCharAt() 353 ? kQuickThrowStringBounds 354 : kQuickThrowArrayBounds; 355 arm_codegen->InvokeRuntime(entrypoint, instruction_, instruction_->GetDexPc(), this); 356 CheckEntrypointTypes<kQuickThrowStringBounds, void, int32_t, int32_t>(); 357 CheckEntrypointTypes<kQuickThrowArrayBounds, void, int32_t, int32_t>(); 358 } 359 360 bool IsFatal() const OVERRIDE { return true; } 361 362 const char* GetDescription() const OVERRIDE { return "BoundsCheckSlowPathARM"; } 363 364 private: 365 DISALLOW_COPY_AND_ASSIGN(BoundsCheckSlowPathARM); 366}; 367 368class LoadClassSlowPathARM : public SlowPathCodeARM { 369 public: 370 LoadClassSlowPathARM(HLoadClass* cls, 371 HInstruction* at, 372 uint32_t dex_pc, 373 bool do_clinit) 374 : SlowPathCodeARM(at), cls_(cls), at_(at), dex_pc_(dex_pc), do_clinit_(do_clinit) { 375 DCHECK(at->IsLoadClass() || at->IsClinitCheck()); 376 } 377 378 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 379 LocationSummary* locations = at_->GetLocations(); 380 381 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 382 __ Bind(GetEntryLabel()); 383 SaveLiveRegisters(codegen, locations); 384 385 InvokeRuntimeCallingConvention calling_convention; 386 __ LoadImmediate(calling_convention.GetRegisterAt(0), cls_->GetTypeIndex()); 387 QuickEntrypointEnum entrypoint = do_clinit_ ? kQuickInitializeStaticStorage 388 : kQuickInitializeType; 389 arm_codegen->InvokeRuntime(entrypoint, at_, dex_pc_, this); 390 if (do_clinit_) { 391 CheckEntrypointTypes<kQuickInitializeStaticStorage, void*, uint32_t>(); 392 } else { 393 CheckEntrypointTypes<kQuickInitializeType, void*, uint32_t>(); 394 } 395 396 // Move the class to the desired location. 397 Location out = locations->Out(); 398 if (out.IsValid()) { 399 DCHECK(out.IsRegister() && !locations->GetLiveRegisters()->ContainsCoreRegister(out.reg())); 400 arm_codegen->Move32(locations->Out(), Location::RegisterLocation(R0)); 401 } 402 RestoreLiveRegisters(codegen, locations); 403 __ b(GetExitLabel()); 404 } 405 406 const char* GetDescription() const OVERRIDE { return "LoadClassSlowPathARM"; } 407 408 private: 409 // The class this slow path will load. 410 HLoadClass* const cls_; 411 412 // The instruction where this slow path is happening. 413 // (Might be the load class or an initialization check). 414 HInstruction* const at_; 415 416 // The dex PC of `at_`. 417 const uint32_t dex_pc_; 418 419 // Whether to initialize the class. 420 const bool do_clinit_; 421 422 DISALLOW_COPY_AND_ASSIGN(LoadClassSlowPathARM); 423}; 424 425class LoadStringSlowPathARM : public SlowPathCodeARM { 426 public: 427 explicit LoadStringSlowPathARM(HLoadString* instruction) : SlowPathCodeARM(instruction) {} 428 429 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 430 LocationSummary* locations = instruction_->GetLocations(); 431 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg())); 432 HLoadString* load = instruction_->AsLoadString(); 433 const uint32_t string_index = load->GetStringIndex(); 434 Register out = locations->Out().AsRegister<Register>(); 435 Register temp = locations->GetTemp(0).AsRegister<Register>(); 436 constexpr bool call_saves_everything_except_r0 = (!kUseReadBarrier || kUseBakerReadBarrier); 437 438 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 439 __ Bind(GetEntryLabel()); 440 SaveLiveRegisters(codegen, locations); 441 442 InvokeRuntimeCallingConvention calling_convention; 443 // In the unlucky case that the `temp` is R0, we preserve the address in `out` across 444 // the kSaveEverything call (or use `out` for the address after non-kSaveEverything call). 445 bool temp_is_r0 = (temp == calling_convention.GetRegisterAt(0)); 446 Register entry_address = temp_is_r0 ? out : temp; 447 DCHECK_NE(entry_address, calling_convention.GetRegisterAt(0)); 448 if (call_saves_everything_except_r0 && temp_is_r0) { 449 __ mov(entry_address, ShifterOperand(temp)); 450 } 451 452 __ LoadImmediate(calling_convention.GetRegisterAt(0), string_index); 453 arm_codegen->InvokeRuntime(kQuickResolveString, instruction_, instruction_->GetDexPc(), this); 454 CheckEntrypointTypes<kQuickResolveString, void*, uint32_t>(); 455 456 // Store the resolved String to the .bss entry. 457 if (call_saves_everything_except_r0) { 458 // The string entry address was preserved in `entry_address` thanks to kSaveEverything. 459 __ str(R0, Address(entry_address)); 460 } else { 461 // For non-Baker read barrier, we need to re-calculate the address of the string entry. 462 CodeGeneratorARM::PcRelativePatchInfo* labels = 463 arm_codegen->NewPcRelativeStringPatch(load->GetDexFile(), string_index); 464 __ BindTrackedLabel(&labels->movw_label); 465 __ movw(entry_address, /* placeholder */ 0u); 466 __ BindTrackedLabel(&labels->movt_label); 467 __ movt(entry_address, /* placeholder */ 0u); 468 __ BindTrackedLabel(&labels->add_pc_label); 469 __ add(entry_address, entry_address, ShifterOperand(PC)); 470 __ str(R0, Address(entry_address)); 471 } 472 473 arm_codegen->Move32(locations->Out(), Location::RegisterLocation(R0)); 474 RestoreLiveRegisters(codegen, locations); 475 476 __ b(GetExitLabel()); 477 } 478 479 const char* GetDescription() const OVERRIDE { return "LoadStringSlowPathARM"; } 480 481 private: 482 DISALLOW_COPY_AND_ASSIGN(LoadStringSlowPathARM); 483}; 484 485class TypeCheckSlowPathARM : public SlowPathCodeARM { 486 public: 487 TypeCheckSlowPathARM(HInstruction* instruction, bool is_fatal) 488 : SlowPathCodeARM(instruction), is_fatal_(is_fatal) {} 489 490 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 491 LocationSummary* locations = instruction_->GetLocations(); 492 Location arg0, arg1; 493 if (instruction_->IsInstanceOf()) { 494 arg0 = locations->InAt(1); 495 arg1 = locations->Out(); 496 } else { 497 arg0 = locations->InAt(0); 498 arg1 = locations->InAt(1); 499 } 500 DCHECK(instruction_->IsCheckCast() 501 || !locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg())); 502 503 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 504 __ Bind(GetEntryLabel()); 505 506 if (!is_fatal_) { 507 SaveLiveRegisters(codegen, locations); 508 } 509 510 // We're moving two locations to locations that could overlap, so we need a parallel 511 // move resolver. 512 InvokeRuntimeCallingConvention calling_convention; 513 codegen->EmitParallelMoves(arg0, 514 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 515 Primitive::kPrimNot, 516 arg1, 517 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 518 Primitive::kPrimNot); 519 if (instruction_->IsInstanceOf()) { 520 arm_codegen->InvokeRuntime(kQuickInstanceofNonTrivial, 521 instruction_, 522 instruction_->GetDexPc(), 523 this); 524 CheckEntrypointTypes<kQuickInstanceofNonTrivial, size_t, mirror::Class*, mirror::Class*>(); 525 arm_codegen->Move32(locations->Out(), Location::RegisterLocation(R0)); 526 } else { 527 DCHECK(instruction_->IsCheckCast()); 528 arm_codegen->InvokeRuntime(kQuickCheckInstanceOf, 529 instruction_, 530 instruction_->GetDexPc(), 531 this); 532 CheckEntrypointTypes<kQuickCheckInstanceOf, void, mirror::Object*, mirror::Class*>(); 533 } 534 535 if (!is_fatal_) { 536 RestoreLiveRegisters(codegen, locations); 537 __ b(GetExitLabel()); 538 } 539 } 540 541 const char* GetDescription() const OVERRIDE { return "TypeCheckSlowPathARM"; } 542 543 bool IsFatal() const OVERRIDE { return is_fatal_; } 544 545 private: 546 const bool is_fatal_; 547 548 DISALLOW_COPY_AND_ASSIGN(TypeCheckSlowPathARM); 549}; 550 551class DeoptimizationSlowPathARM : public SlowPathCodeARM { 552 public: 553 explicit DeoptimizationSlowPathARM(HDeoptimize* instruction) 554 : SlowPathCodeARM(instruction) {} 555 556 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 557 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 558 __ Bind(GetEntryLabel()); 559 arm_codegen->InvokeRuntime(kQuickDeoptimize, instruction_, instruction_->GetDexPc(), this); 560 CheckEntrypointTypes<kQuickDeoptimize, void, void>(); 561 } 562 563 const char* GetDescription() const OVERRIDE { return "DeoptimizationSlowPathARM"; } 564 565 private: 566 DISALLOW_COPY_AND_ASSIGN(DeoptimizationSlowPathARM); 567}; 568 569class ArraySetSlowPathARM : public SlowPathCodeARM { 570 public: 571 explicit ArraySetSlowPathARM(HInstruction* instruction) : SlowPathCodeARM(instruction) {} 572 573 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 574 LocationSummary* locations = instruction_->GetLocations(); 575 __ Bind(GetEntryLabel()); 576 SaveLiveRegisters(codegen, locations); 577 578 InvokeRuntimeCallingConvention calling_convention; 579 HParallelMove parallel_move(codegen->GetGraph()->GetArena()); 580 parallel_move.AddMove( 581 locations->InAt(0), 582 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 583 Primitive::kPrimNot, 584 nullptr); 585 parallel_move.AddMove( 586 locations->InAt(1), 587 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 588 Primitive::kPrimInt, 589 nullptr); 590 parallel_move.AddMove( 591 locations->InAt(2), 592 Location::RegisterLocation(calling_convention.GetRegisterAt(2)), 593 Primitive::kPrimNot, 594 nullptr); 595 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 596 597 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 598 arm_codegen->InvokeRuntime(kQuickAputObject, instruction_, instruction_->GetDexPc(), this); 599 CheckEntrypointTypes<kQuickAputObject, void, mirror::Array*, int32_t, mirror::Object*>(); 600 RestoreLiveRegisters(codegen, locations); 601 __ b(GetExitLabel()); 602 } 603 604 const char* GetDescription() const OVERRIDE { return "ArraySetSlowPathARM"; } 605 606 private: 607 DISALLOW_COPY_AND_ASSIGN(ArraySetSlowPathARM); 608}; 609 610// Slow path marking an object reference `ref` during a read 611// barrier. The field `obj.field` in the object `obj` holding this 612// reference does not get updated by this slow path after marking (see 613// ReadBarrierMarkAndUpdateFieldSlowPathARM below for that). 614// 615// This means that after the execution of this slow path, `ref` will 616// always be up-to-date, but `obj.field` may not; i.e., after the 617// flip, `ref` will be a to-space reference, but `obj.field` will 618// probably still be a from-space reference (unless it gets updated by 619// another thread, or if another thread installed another object 620// reference (different from `ref`) in `obj.field`). 621class ReadBarrierMarkSlowPathARM : public SlowPathCodeARM { 622 public: 623 ReadBarrierMarkSlowPathARM(HInstruction* instruction, Location ref) 624 : SlowPathCodeARM(instruction), ref_(ref) { 625 DCHECK(kEmitCompilerReadBarrier); 626 } 627 628 const char* GetDescription() const OVERRIDE { return "ReadBarrierMarkSlowPathARM"; } 629 630 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 631 LocationSummary* locations = instruction_->GetLocations(); 632 Register ref_reg = ref_.AsRegister<Register>(); 633 DCHECK(locations->CanCall()); 634 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(ref_reg)) << ref_reg; 635 DCHECK(instruction_->IsInstanceFieldGet() || 636 instruction_->IsStaticFieldGet() || 637 instruction_->IsArrayGet() || 638 instruction_->IsArraySet() || 639 instruction_->IsLoadClass() || 640 instruction_->IsLoadString() || 641 instruction_->IsInstanceOf() || 642 instruction_->IsCheckCast() || 643 (instruction_->IsInvokeVirtual() && instruction_->GetLocations()->Intrinsified()) || 644 (instruction_->IsInvokeStaticOrDirect() && instruction_->GetLocations()->Intrinsified())) 645 << "Unexpected instruction in read barrier marking slow path: " 646 << instruction_->DebugName(); 647 // The read barrier instrumentation of object ArrayGet 648 // instructions does not support the HIntermediateAddress 649 // instruction. 650 DCHECK(!(instruction_->IsArrayGet() && 651 instruction_->AsArrayGet()->GetArray()->IsIntermediateAddress())); 652 653 __ Bind(GetEntryLabel()); 654 // No need to save live registers; it's taken care of by the 655 // entrypoint. Also, there is no need to update the stack mask, 656 // as this runtime call will not trigger a garbage collection. 657 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 658 DCHECK_NE(ref_reg, SP); 659 DCHECK_NE(ref_reg, LR); 660 DCHECK_NE(ref_reg, PC); 661 // IP is used internally by the ReadBarrierMarkRegX entry point 662 // as a temporary, it cannot be the entry point's input/output. 663 DCHECK_NE(ref_reg, IP); 664 DCHECK(0 <= ref_reg && ref_reg < kNumberOfCoreRegisters) << ref_reg; 665 // "Compact" slow path, saving two moves. 666 // 667 // Instead of using the standard runtime calling convention (input 668 // and output in R0): 669 // 670 // R0 <- ref 671 // R0 <- ReadBarrierMark(R0) 672 // ref <- R0 673 // 674 // we just use rX (the register containing `ref`) as input and output 675 // of a dedicated entrypoint: 676 // 677 // rX <- ReadBarrierMarkRegX(rX) 678 // 679 int32_t entry_point_offset = 680 CodeGenerator::GetReadBarrierMarkEntryPointsOffset<kArmPointerSize>(ref_reg); 681 // This runtime call does not require a stack map. 682 arm_codegen->InvokeRuntimeWithoutRecordingPcInfo(entry_point_offset, instruction_, this); 683 __ b(GetExitLabel()); 684 } 685 686 private: 687 // The location (register) of the marked object reference. 688 const Location ref_; 689 690 DISALLOW_COPY_AND_ASSIGN(ReadBarrierMarkSlowPathARM); 691}; 692 693// Slow path marking an object reference `ref` during a read barrier, 694// and if needed, atomically updating the field `obj.field` in the 695// object `obj` holding this reference after marking (contrary to 696// ReadBarrierMarkSlowPathARM above, which never tries to update 697// `obj.field`). 698// 699// This means that after the execution of this slow path, both `ref` 700// and `obj.field` will be up-to-date; i.e., after the flip, both will 701// hold the same to-space reference (unless another thread installed 702// another object reference (different from `ref`) in `obj.field`). 703class ReadBarrierMarkAndUpdateFieldSlowPathARM : public SlowPathCodeARM { 704 public: 705 ReadBarrierMarkAndUpdateFieldSlowPathARM(HInstruction* instruction, 706 Location ref, 707 Register obj, 708 Location field_offset, 709 Register temp1, 710 Register temp2) 711 : SlowPathCodeARM(instruction), 712 ref_(ref), 713 obj_(obj), 714 field_offset_(field_offset), 715 temp1_(temp1), 716 temp2_(temp2) { 717 DCHECK(kEmitCompilerReadBarrier); 718 } 719 720 const char* GetDescription() const OVERRIDE { return "ReadBarrierMarkAndUpdateFieldSlowPathARM"; } 721 722 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 723 LocationSummary* locations = instruction_->GetLocations(); 724 Register ref_reg = ref_.AsRegister<Register>(); 725 DCHECK(locations->CanCall()); 726 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(ref_reg)) << ref_reg; 727 // This slow path is only used by the UnsafeCASObject intrinsic. 728 DCHECK((instruction_->IsInvokeVirtual() && instruction_->GetLocations()->Intrinsified())) 729 << "Unexpected instruction in read barrier marking and field updating slow path: " 730 << instruction_->DebugName(); 731 DCHECK(instruction_->GetLocations()->Intrinsified()); 732 DCHECK_EQ(instruction_->AsInvoke()->GetIntrinsic(), Intrinsics::kUnsafeCASObject); 733 DCHECK(field_offset_.IsRegisterPair()) << field_offset_; 734 735 __ Bind(GetEntryLabel()); 736 737 // Save the old reference. 738 // Note that we cannot use IP to save the old reference, as IP is 739 // used internally by the ReadBarrierMarkRegX entry point, and we 740 // need the old reference after the call to that entry point. 741 DCHECK_NE(temp1_, IP); 742 __ Mov(temp1_, ref_reg); 743 744 // No need to save live registers; it's taken care of by the 745 // entrypoint. Also, there is no need to update the stack mask, 746 // as this runtime call will not trigger a garbage collection. 747 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 748 DCHECK_NE(ref_reg, SP); 749 DCHECK_NE(ref_reg, LR); 750 DCHECK_NE(ref_reg, PC); 751 // IP is used internally by the ReadBarrierMarkRegX entry point 752 // as a temporary, it cannot be the entry point's input/output. 753 DCHECK_NE(ref_reg, IP); 754 DCHECK(0 <= ref_reg && ref_reg < kNumberOfCoreRegisters) << ref_reg; 755 // "Compact" slow path, saving two moves. 756 // 757 // Instead of using the standard runtime calling convention (input 758 // and output in R0): 759 // 760 // R0 <- ref 761 // R0 <- ReadBarrierMark(R0) 762 // ref <- R0 763 // 764 // we just use rX (the register containing `ref`) as input and output 765 // of a dedicated entrypoint: 766 // 767 // rX <- ReadBarrierMarkRegX(rX) 768 // 769 int32_t entry_point_offset = 770 CodeGenerator::GetReadBarrierMarkEntryPointsOffset<kArmPointerSize>(ref_reg); 771 // This runtime call does not require a stack map. 772 arm_codegen->InvokeRuntimeWithoutRecordingPcInfo(entry_point_offset, instruction_, this); 773 774 // If the new reference is different from the old reference, 775 // update the field in the holder (`*(obj_ + field_offset_)`). 776 // 777 // Note that this field could also hold a different object, if 778 // another thread had concurrently changed it. In that case, the 779 // LDREX/SUBS/ITNE sequence of instructions in the compare-and-set 780 // (CAS) operation below would abort the CAS, leaving the field 781 // as-is. 782 Label done; 783 __ cmp(temp1_, ShifterOperand(ref_reg)); 784 __ b(&done, EQ); 785 786 // Update the the holder's field atomically. This may fail if 787 // mutator updates before us, but it's OK. This is achieved 788 // using a strong compare-and-set (CAS) operation with relaxed 789 // memory synchronization ordering, where the expected value is 790 // the old reference and the desired value is the new reference. 791 792 // Convenience aliases. 793 Register base = obj_; 794 // The UnsafeCASObject intrinsic uses a register pair as field 795 // offset ("long offset"), of which only the low part contains 796 // data. 797 Register offset = field_offset_.AsRegisterPairLow<Register>(); 798 Register expected = temp1_; 799 Register value = ref_reg; 800 Register tmp_ptr = IP; // Pointer to actual memory. 801 Register tmp = temp2_; // Value in memory. 802 803 __ add(tmp_ptr, base, ShifterOperand(offset)); 804 805 if (kPoisonHeapReferences) { 806 __ PoisonHeapReference(expected); 807 if (value == expected) { 808 // Do not poison `value`, as it is the same register as 809 // `expected`, which has just been poisoned. 810 } else { 811 __ PoisonHeapReference(value); 812 } 813 } 814 815 // do { 816 // tmp = [r_ptr] - expected; 817 // } while (tmp == 0 && failure([r_ptr] <- r_new_value)); 818 819 Label loop_head, exit_loop; 820 __ Bind(&loop_head); 821 822 __ ldrex(tmp, tmp_ptr); 823 824 __ subs(tmp, tmp, ShifterOperand(expected)); 825 826 __ it(NE); 827 __ clrex(NE); 828 829 __ b(&exit_loop, NE); 830 831 __ strex(tmp, value, tmp_ptr); 832 __ cmp(tmp, ShifterOperand(1)); 833 __ b(&loop_head, EQ); 834 835 __ Bind(&exit_loop); 836 837 if (kPoisonHeapReferences) { 838 __ UnpoisonHeapReference(expected); 839 if (value == expected) { 840 // Do not unpoison `value`, as it is the same register as 841 // `expected`, which has just been unpoisoned. 842 } else { 843 __ UnpoisonHeapReference(value); 844 } 845 } 846 847 __ Bind(&done); 848 __ b(GetExitLabel()); 849 } 850 851 private: 852 // The location (register) of the marked object reference. 853 const Location ref_; 854 // The register containing the object holding the marked object reference field. 855 const Register obj_; 856 // The location of the offset of the marked reference field within `obj_`. 857 Location field_offset_; 858 859 const Register temp1_; 860 const Register temp2_; 861 862 DISALLOW_COPY_AND_ASSIGN(ReadBarrierMarkAndUpdateFieldSlowPathARM); 863}; 864 865// Slow path generating a read barrier for a heap reference. 866class ReadBarrierForHeapReferenceSlowPathARM : public SlowPathCodeARM { 867 public: 868 ReadBarrierForHeapReferenceSlowPathARM(HInstruction* instruction, 869 Location out, 870 Location ref, 871 Location obj, 872 uint32_t offset, 873 Location index) 874 : SlowPathCodeARM(instruction), 875 out_(out), 876 ref_(ref), 877 obj_(obj), 878 offset_(offset), 879 index_(index) { 880 DCHECK(kEmitCompilerReadBarrier); 881 // If `obj` is equal to `out` or `ref`, it means the initial object 882 // has been overwritten by (or after) the heap object reference load 883 // to be instrumented, e.g.: 884 // 885 // __ LoadFromOffset(kLoadWord, out, out, offset); 886 // codegen_->GenerateReadBarrierSlow(instruction, out_loc, out_loc, out_loc, offset); 887 // 888 // In that case, we have lost the information about the original 889 // object, and the emitted read barrier cannot work properly. 890 DCHECK(!obj.Equals(out)) << "obj=" << obj << " out=" << out; 891 DCHECK(!obj.Equals(ref)) << "obj=" << obj << " ref=" << ref; 892 } 893 894 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 895 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 896 LocationSummary* locations = instruction_->GetLocations(); 897 Register reg_out = out_.AsRegister<Register>(); 898 DCHECK(locations->CanCall()); 899 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg_out)); 900 DCHECK(instruction_->IsInstanceFieldGet() || 901 instruction_->IsStaticFieldGet() || 902 instruction_->IsArrayGet() || 903 instruction_->IsInstanceOf() || 904 instruction_->IsCheckCast() || 905 (instruction_->IsInvokeVirtual()) && instruction_->GetLocations()->Intrinsified()) 906 << "Unexpected instruction in read barrier for heap reference slow path: " 907 << instruction_->DebugName(); 908 // The read barrier instrumentation of object ArrayGet 909 // instructions does not support the HIntermediateAddress 910 // instruction. 911 DCHECK(!(instruction_->IsArrayGet() && 912 instruction_->AsArrayGet()->GetArray()->IsIntermediateAddress())); 913 914 __ Bind(GetEntryLabel()); 915 SaveLiveRegisters(codegen, locations); 916 917 // We may have to change the index's value, but as `index_` is a 918 // constant member (like other "inputs" of this slow path), 919 // introduce a copy of it, `index`. 920 Location index = index_; 921 if (index_.IsValid()) { 922 // Handle `index_` for HArrayGet and UnsafeGetObject/UnsafeGetObjectVolatile intrinsics. 923 if (instruction_->IsArrayGet()) { 924 // Compute the actual memory offset and store it in `index`. 925 Register index_reg = index_.AsRegister<Register>(); 926 DCHECK(locations->GetLiveRegisters()->ContainsCoreRegister(index_reg)); 927 if (codegen->IsCoreCalleeSaveRegister(index_reg)) { 928 // We are about to change the value of `index_reg` (see the 929 // calls to art::arm::Thumb2Assembler::Lsl and 930 // art::arm::Thumb2Assembler::AddConstant below), but it has 931 // not been saved by the previous call to 932 // art::SlowPathCode::SaveLiveRegisters, as it is a 933 // callee-save register -- 934 // art::SlowPathCode::SaveLiveRegisters does not consider 935 // callee-save registers, as it has been designed with the 936 // assumption that callee-save registers are supposed to be 937 // handled by the called function. So, as a callee-save 938 // register, `index_reg` _would_ eventually be saved onto 939 // the stack, but it would be too late: we would have 940 // changed its value earlier. Therefore, we manually save 941 // it here into another freely available register, 942 // `free_reg`, chosen of course among the caller-save 943 // registers (as a callee-save `free_reg` register would 944 // exhibit the same problem). 945 // 946 // Note we could have requested a temporary register from 947 // the register allocator instead; but we prefer not to, as 948 // this is a slow path, and we know we can find a 949 // caller-save register that is available. 950 Register free_reg = FindAvailableCallerSaveRegister(codegen); 951 __ Mov(free_reg, index_reg); 952 index_reg = free_reg; 953 index = Location::RegisterLocation(index_reg); 954 } else { 955 // The initial register stored in `index_` has already been 956 // saved in the call to art::SlowPathCode::SaveLiveRegisters 957 // (as it is not a callee-save register), so we can freely 958 // use it. 959 } 960 // Shifting the index value contained in `index_reg` by the scale 961 // factor (2) cannot overflow in practice, as the runtime is 962 // unable to allocate object arrays with a size larger than 963 // 2^26 - 1 (that is, 2^28 - 4 bytes). 964 __ Lsl(index_reg, index_reg, TIMES_4); 965 static_assert( 966 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 967 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 968 __ AddConstant(index_reg, index_reg, offset_); 969 } else { 970 // In the case of the UnsafeGetObject/UnsafeGetObjectVolatile 971 // intrinsics, `index_` is not shifted by a scale factor of 2 972 // (as in the case of ArrayGet), as it is actually an offset 973 // to an object field within an object. 974 DCHECK(instruction_->IsInvoke()) << instruction_->DebugName(); 975 DCHECK(instruction_->GetLocations()->Intrinsified()); 976 DCHECK((instruction_->AsInvoke()->GetIntrinsic() == Intrinsics::kUnsafeGetObject) || 977 (instruction_->AsInvoke()->GetIntrinsic() == Intrinsics::kUnsafeGetObjectVolatile)) 978 << instruction_->AsInvoke()->GetIntrinsic(); 979 DCHECK_EQ(offset_, 0U); 980 DCHECK(index_.IsRegisterPair()); 981 // UnsafeGet's offset location is a register pair, the low 982 // part contains the correct offset. 983 index = index_.ToLow(); 984 } 985 } 986 987 // We're moving two or three locations to locations that could 988 // overlap, so we need a parallel move resolver. 989 InvokeRuntimeCallingConvention calling_convention; 990 HParallelMove parallel_move(codegen->GetGraph()->GetArena()); 991 parallel_move.AddMove(ref_, 992 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 993 Primitive::kPrimNot, 994 nullptr); 995 parallel_move.AddMove(obj_, 996 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 997 Primitive::kPrimNot, 998 nullptr); 999 if (index.IsValid()) { 1000 parallel_move.AddMove(index, 1001 Location::RegisterLocation(calling_convention.GetRegisterAt(2)), 1002 Primitive::kPrimInt, 1003 nullptr); 1004 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 1005 } else { 1006 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 1007 __ LoadImmediate(calling_convention.GetRegisterAt(2), offset_); 1008 } 1009 arm_codegen->InvokeRuntime(kQuickReadBarrierSlow, instruction_, instruction_->GetDexPc(), this); 1010 CheckEntrypointTypes< 1011 kQuickReadBarrierSlow, mirror::Object*, mirror::Object*, mirror::Object*, uint32_t>(); 1012 arm_codegen->Move32(out_, Location::RegisterLocation(R0)); 1013 1014 RestoreLiveRegisters(codegen, locations); 1015 __ b(GetExitLabel()); 1016 } 1017 1018 const char* GetDescription() const OVERRIDE { return "ReadBarrierForHeapReferenceSlowPathARM"; } 1019 1020 private: 1021 Register FindAvailableCallerSaveRegister(CodeGenerator* codegen) { 1022 size_t ref = static_cast<int>(ref_.AsRegister<Register>()); 1023 size_t obj = static_cast<int>(obj_.AsRegister<Register>()); 1024 for (size_t i = 0, e = codegen->GetNumberOfCoreRegisters(); i < e; ++i) { 1025 if (i != ref && i != obj && !codegen->IsCoreCalleeSaveRegister(i)) { 1026 return static_cast<Register>(i); 1027 } 1028 } 1029 // We shall never fail to find a free caller-save register, as 1030 // there are more than two core caller-save registers on ARM 1031 // (meaning it is possible to find one which is different from 1032 // `ref` and `obj`). 1033 DCHECK_GT(codegen->GetNumberOfCoreCallerSaveRegisters(), 2u); 1034 LOG(FATAL) << "Could not find a free caller-save register"; 1035 UNREACHABLE(); 1036 } 1037 1038 const Location out_; 1039 const Location ref_; 1040 const Location obj_; 1041 const uint32_t offset_; 1042 // An additional location containing an index to an array. 1043 // Only used for HArrayGet and the UnsafeGetObject & 1044 // UnsafeGetObjectVolatile intrinsics. 1045 const Location index_; 1046 1047 DISALLOW_COPY_AND_ASSIGN(ReadBarrierForHeapReferenceSlowPathARM); 1048}; 1049 1050// Slow path generating a read barrier for a GC root. 1051class ReadBarrierForRootSlowPathARM : public SlowPathCodeARM { 1052 public: 1053 ReadBarrierForRootSlowPathARM(HInstruction* instruction, Location out, Location root) 1054 : SlowPathCodeARM(instruction), out_(out), root_(root) { 1055 DCHECK(kEmitCompilerReadBarrier); 1056 } 1057 1058 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 1059 LocationSummary* locations = instruction_->GetLocations(); 1060 Register reg_out = out_.AsRegister<Register>(); 1061 DCHECK(locations->CanCall()); 1062 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg_out)); 1063 DCHECK(instruction_->IsLoadClass() || instruction_->IsLoadString()) 1064 << "Unexpected instruction in read barrier for GC root slow path: " 1065 << instruction_->DebugName(); 1066 1067 __ Bind(GetEntryLabel()); 1068 SaveLiveRegisters(codegen, locations); 1069 1070 InvokeRuntimeCallingConvention calling_convention; 1071 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 1072 arm_codegen->Move32(Location::RegisterLocation(calling_convention.GetRegisterAt(0)), root_); 1073 arm_codegen->InvokeRuntime(kQuickReadBarrierForRootSlow, 1074 instruction_, 1075 instruction_->GetDexPc(), 1076 this); 1077 CheckEntrypointTypes<kQuickReadBarrierForRootSlow, mirror::Object*, GcRoot<mirror::Object>*>(); 1078 arm_codegen->Move32(out_, Location::RegisterLocation(R0)); 1079 1080 RestoreLiveRegisters(codegen, locations); 1081 __ b(GetExitLabel()); 1082 } 1083 1084 const char* GetDescription() const OVERRIDE { return "ReadBarrierForRootSlowPathARM"; } 1085 1086 private: 1087 const Location out_; 1088 const Location root_; 1089 1090 DISALLOW_COPY_AND_ASSIGN(ReadBarrierForRootSlowPathARM); 1091}; 1092 1093#undef __ 1094// NOLINT on __ macro to suppress wrong warning/fix (misc-macro-parentheses) from clang-tidy. 1095#define __ down_cast<ArmAssembler*>(GetAssembler())-> // NOLINT 1096 1097inline Condition ARMCondition(IfCondition cond) { 1098 switch (cond) { 1099 case kCondEQ: return EQ; 1100 case kCondNE: return NE; 1101 case kCondLT: return LT; 1102 case kCondLE: return LE; 1103 case kCondGT: return GT; 1104 case kCondGE: return GE; 1105 case kCondB: return LO; 1106 case kCondBE: return LS; 1107 case kCondA: return HI; 1108 case kCondAE: return HS; 1109 } 1110 LOG(FATAL) << "Unreachable"; 1111 UNREACHABLE(); 1112} 1113 1114// Maps signed condition to unsigned condition. 1115inline Condition ARMUnsignedCondition(IfCondition cond) { 1116 switch (cond) { 1117 case kCondEQ: return EQ; 1118 case kCondNE: return NE; 1119 // Signed to unsigned. 1120 case kCondLT: return LO; 1121 case kCondLE: return LS; 1122 case kCondGT: return HI; 1123 case kCondGE: return HS; 1124 // Unsigned remain unchanged. 1125 case kCondB: return LO; 1126 case kCondBE: return LS; 1127 case kCondA: return HI; 1128 case kCondAE: return HS; 1129 } 1130 LOG(FATAL) << "Unreachable"; 1131 UNREACHABLE(); 1132} 1133 1134inline Condition ARMFPCondition(IfCondition cond, bool gt_bias) { 1135 // The ARM condition codes can express all the necessary branches, see the 1136 // "Meaning (floating-point)" column in the table A8-1 of the ARMv7 reference manual. 1137 // There is no dex instruction or HIR that would need the missing conditions 1138 // "equal or unordered" or "not equal". 1139 switch (cond) { 1140 case kCondEQ: return EQ; 1141 case kCondNE: return NE /* unordered */; 1142 case kCondLT: return gt_bias ? CC : LT /* unordered */; 1143 case kCondLE: return gt_bias ? LS : LE /* unordered */; 1144 case kCondGT: return gt_bias ? HI /* unordered */ : GT; 1145 case kCondGE: return gt_bias ? CS /* unordered */ : GE; 1146 default: 1147 LOG(FATAL) << "UNREACHABLE"; 1148 UNREACHABLE(); 1149 } 1150} 1151 1152void CodeGeneratorARM::DumpCoreRegister(std::ostream& stream, int reg) const { 1153 stream << Register(reg); 1154} 1155 1156void CodeGeneratorARM::DumpFloatingPointRegister(std::ostream& stream, int reg) const { 1157 stream << SRegister(reg); 1158} 1159 1160size_t CodeGeneratorARM::SaveCoreRegister(size_t stack_index, uint32_t reg_id) { 1161 __ StoreToOffset(kStoreWord, static_cast<Register>(reg_id), SP, stack_index); 1162 return kArmWordSize; 1163} 1164 1165size_t CodeGeneratorARM::RestoreCoreRegister(size_t stack_index, uint32_t reg_id) { 1166 __ LoadFromOffset(kLoadWord, static_cast<Register>(reg_id), SP, stack_index); 1167 return kArmWordSize; 1168} 1169 1170size_t CodeGeneratorARM::SaveFloatingPointRegister(size_t stack_index, uint32_t reg_id) { 1171 __ StoreSToOffset(static_cast<SRegister>(reg_id), SP, stack_index); 1172 return kArmWordSize; 1173} 1174 1175size_t CodeGeneratorARM::RestoreFloatingPointRegister(size_t stack_index, uint32_t reg_id) { 1176 __ LoadSFromOffset(static_cast<SRegister>(reg_id), SP, stack_index); 1177 return kArmWordSize; 1178} 1179 1180CodeGeneratorARM::CodeGeneratorARM(HGraph* graph, 1181 const ArmInstructionSetFeatures& isa_features, 1182 const CompilerOptions& compiler_options, 1183 OptimizingCompilerStats* stats) 1184 : CodeGenerator(graph, 1185 kNumberOfCoreRegisters, 1186 kNumberOfSRegisters, 1187 kNumberOfRegisterPairs, 1188 ComputeRegisterMask(reinterpret_cast<const int*>(kCoreCalleeSaves), 1189 arraysize(kCoreCalleeSaves)), 1190 ComputeRegisterMask(reinterpret_cast<const int*>(kFpuCalleeSaves), 1191 arraysize(kFpuCalleeSaves)), 1192 compiler_options, 1193 stats), 1194 block_labels_(nullptr), 1195 location_builder_(graph, this), 1196 instruction_visitor_(graph, this), 1197 move_resolver_(graph->GetArena(), this), 1198 assembler_(graph->GetArena()), 1199 isa_features_(isa_features), 1200 uint32_literals_(std::less<uint32_t>(), 1201 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 1202 method_patches_(MethodReferenceComparator(), 1203 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 1204 call_patches_(MethodReferenceComparator(), 1205 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 1206 relative_call_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 1207 pc_relative_dex_cache_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 1208 boot_image_string_patches_(StringReferenceValueComparator(), 1209 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 1210 pc_relative_string_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 1211 boot_image_type_patches_(TypeReferenceValueComparator(), 1212 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 1213 pc_relative_type_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 1214 boot_image_address_patches_(std::less<uint32_t>(), 1215 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)) { 1216 // Always save the LR register to mimic Quick. 1217 AddAllocatedRegister(Location::RegisterLocation(LR)); 1218} 1219 1220void CodeGeneratorARM::Finalize(CodeAllocator* allocator) { 1221 // Ensure that we fix up branches and literal loads and emit the literal pool. 1222 __ FinalizeCode(); 1223 1224 // Adjust native pc offsets in stack maps. 1225 for (size_t i = 0, num = stack_map_stream_.GetNumberOfStackMaps(); i != num; ++i) { 1226 uint32_t old_position = stack_map_stream_.GetStackMap(i).native_pc_offset; 1227 uint32_t new_position = __ GetAdjustedPosition(old_position); 1228 stack_map_stream_.SetStackMapNativePcOffset(i, new_position); 1229 } 1230 // Adjust pc offsets for the disassembly information. 1231 if (disasm_info_ != nullptr) { 1232 GeneratedCodeInterval* frame_entry_interval = disasm_info_->GetFrameEntryInterval(); 1233 frame_entry_interval->start = __ GetAdjustedPosition(frame_entry_interval->start); 1234 frame_entry_interval->end = __ GetAdjustedPosition(frame_entry_interval->end); 1235 for (auto& it : *disasm_info_->GetInstructionIntervals()) { 1236 it.second.start = __ GetAdjustedPosition(it.second.start); 1237 it.second.end = __ GetAdjustedPosition(it.second.end); 1238 } 1239 for (auto& it : *disasm_info_->GetSlowPathIntervals()) { 1240 it.code_interval.start = __ GetAdjustedPosition(it.code_interval.start); 1241 it.code_interval.end = __ GetAdjustedPosition(it.code_interval.end); 1242 } 1243 } 1244 1245 CodeGenerator::Finalize(allocator); 1246} 1247 1248void CodeGeneratorARM::SetupBlockedRegisters() const { 1249 // Stack register, LR and PC are always reserved. 1250 blocked_core_registers_[SP] = true; 1251 blocked_core_registers_[LR] = true; 1252 blocked_core_registers_[PC] = true; 1253 1254 // Reserve thread register. 1255 blocked_core_registers_[TR] = true; 1256 1257 // Reserve temp register. 1258 blocked_core_registers_[IP] = true; 1259 1260 if (GetGraph()->IsDebuggable()) { 1261 // Stubs do not save callee-save floating point registers. If the graph 1262 // is debuggable, we need to deal with these registers differently. For 1263 // now, just block them. 1264 for (size_t i = 0; i < arraysize(kFpuCalleeSaves); ++i) { 1265 blocked_fpu_registers_[kFpuCalleeSaves[i]] = true; 1266 } 1267 } 1268} 1269 1270InstructionCodeGeneratorARM::InstructionCodeGeneratorARM(HGraph* graph, CodeGeneratorARM* codegen) 1271 : InstructionCodeGenerator(graph, codegen), 1272 assembler_(codegen->GetAssembler()), 1273 codegen_(codegen) {} 1274 1275void CodeGeneratorARM::ComputeSpillMask() { 1276 core_spill_mask_ = allocated_registers_.GetCoreRegisters() & core_callee_save_mask_; 1277 DCHECK_NE(core_spill_mask_, 0u) << "At least the return address register must be saved"; 1278 // There is no easy instruction to restore just the PC on thumb2. We spill and 1279 // restore another arbitrary register. 1280 core_spill_mask_ |= (1 << kCoreAlwaysSpillRegister); 1281 fpu_spill_mask_ = allocated_registers_.GetFloatingPointRegisters() & fpu_callee_save_mask_; 1282 // We use vpush and vpop for saving and restoring floating point registers, which take 1283 // a SRegister and the number of registers to save/restore after that SRegister. We 1284 // therefore update the `fpu_spill_mask_` to also contain those registers not allocated, 1285 // but in the range. 1286 if (fpu_spill_mask_ != 0) { 1287 uint32_t least_significant_bit = LeastSignificantBit(fpu_spill_mask_); 1288 uint32_t most_significant_bit = MostSignificantBit(fpu_spill_mask_); 1289 for (uint32_t i = least_significant_bit + 1 ; i < most_significant_bit; ++i) { 1290 fpu_spill_mask_ |= (1 << i); 1291 } 1292 } 1293} 1294 1295static dwarf::Reg DWARFReg(Register reg) { 1296 return dwarf::Reg::ArmCore(static_cast<int>(reg)); 1297} 1298 1299static dwarf::Reg DWARFReg(SRegister reg) { 1300 return dwarf::Reg::ArmFp(static_cast<int>(reg)); 1301} 1302 1303void CodeGeneratorARM::GenerateFrameEntry() { 1304 bool skip_overflow_check = 1305 IsLeafMethod() && !FrameNeedsStackCheck(GetFrameSize(), InstructionSet::kArm); 1306 DCHECK(GetCompilerOptions().GetImplicitStackOverflowChecks()); 1307 __ Bind(&frame_entry_label_); 1308 1309 if (HasEmptyFrame()) { 1310 return; 1311 } 1312 1313 if (!skip_overflow_check) { 1314 __ AddConstant(IP, SP, -static_cast<int32_t>(GetStackOverflowReservedBytes(kArm))); 1315 __ LoadFromOffset(kLoadWord, IP, IP, 0); 1316 RecordPcInfo(nullptr, 0); 1317 } 1318 1319 __ PushList(core_spill_mask_); 1320 __ cfi().AdjustCFAOffset(kArmWordSize * POPCOUNT(core_spill_mask_)); 1321 __ cfi().RelOffsetForMany(DWARFReg(kMethodRegisterArgument), 0, core_spill_mask_, kArmWordSize); 1322 if (fpu_spill_mask_ != 0) { 1323 SRegister start_register = SRegister(LeastSignificantBit(fpu_spill_mask_)); 1324 __ vpushs(start_register, POPCOUNT(fpu_spill_mask_)); 1325 __ cfi().AdjustCFAOffset(kArmWordSize * POPCOUNT(fpu_spill_mask_)); 1326 __ cfi().RelOffsetForMany(DWARFReg(S0), 0, fpu_spill_mask_, kArmWordSize); 1327 } 1328 int adjust = GetFrameSize() - FrameEntrySpillSize(); 1329 __ AddConstant(SP, -adjust); 1330 __ cfi().AdjustCFAOffset(adjust); 1331 1332 // Save the current method if we need it. Note that we do not 1333 // do this in HCurrentMethod, as the instruction might have been removed 1334 // in the SSA graph. 1335 if (RequiresCurrentMethod()) { 1336 __ StoreToOffset(kStoreWord, kMethodRegisterArgument, SP, 0); 1337 } 1338} 1339 1340void CodeGeneratorARM::GenerateFrameExit() { 1341 if (HasEmptyFrame()) { 1342 __ bx(LR); 1343 return; 1344 } 1345 __ cfi().RememberState(); 1346 int adjust = GetFrameSize() - FrameEntrySpillSize(); 1347 __ AddConstant(SP, adjust); 1348 __ cfi().AdjustCFAOffset(-adjust); 1349 if (fpu_spill_mask_ != 0) { 1350 SRegister start_register = SRegister(LeastSignificantBit(fpu_spill_mask_)); 1351 __ vpops(start_register, POPCOUNT(fpu_spill_mask_)); 1352 __ cfi().AdjustCFAOffset(-static_cast<int>(kArmPointerSize) * POPCOUNT(fpu_spill_mask_)); 1353 __ cfi().RestoreMany(DWARFReg(SRegister(0)), fpu_spill_mask_); 1354 } 1355 // Pop LR into PC to return. 1356 DCHECK_NE(core_spill_mask_ & (1 << LR), 0U); 1357 uint32_t pop_mask = (core_spill_mask_ & (~(1 << LR))) | 1 << PC; 1358 __ PopList(pop_mask); 1359 __ cfi().RestoreState(); 1360 __ cfi().DefCFAOffset(GetFrameSize()); 1361} 1362 1363void CodeGeneratorARM::Bind(HBasicBlock* block) { 1364 Label* label = GetLabelOf(block); 1365 __ BindTrackedLabel(label); 1366} 1367 1368Location InvokeDexCallingConventionVisitorARM::GetNextLocation(Primitive::Type type) { 1369 switch (type) { 1370 case Primitive::kPrimBoolean: 1371 case Primitive::kPrimByte: 1372 case Primitive::kPrimChar: 1373 case Primitive::kPrimShort: 1374 case Primitive::kPrimInt: 1375 case Primitive::kPrimNot: { 1376 uint32_t index = gp_index_++; 1377 uint32_t stack_index = stack_index_++; 1378 if (index < calling_convention.GetNumberOfRegisters()) { 1379 return Location::RegisterLocation(calling_convention.GetRegisterAt(index)); 1380 } else { 1381 return Location::StackSlot(calling_convention.GetStackOffsetOf(stack_index)); 1382 } 1383 } 1384 1385 case Primitive::kPrimLong: { 1386 uint32_t index = gp_index_; 1387 uint32_t stack_index = stack_index_; 1388 gp_index_ += 2; 1389 stack_index_ += 2; 1390 if (index + 1 < calling_convention.GetNumberOfRegisters()) { 1391 if (calling_convention.GetRegisterAt(index) == R1) { 1392 // Skip R1, and use R2_R3 instead. 1393 gp_index_++; 1394 index++; 1395 } 1396 } 1397 if (index + 1 < calling_convention.GetNumberOfRegisters()) { 1398 DCHECK_EQ(calling_convention.GetRegisterAt(index) + 1, 1399 calling_convention.GetRegisterAt(index + 1)); 1400 1401 return Location::RegisterPairLocation(calling_convention.GetRegisterAt(index), 1402 calling_convention.GetRegisterAt(index + 1)); 1403 } else { 1404 return Location::DoubleStackSlot(calling_convention.GetStackOffsetOf(stack_index)); 1405 } 1406 } 1407 1408 case Primitive::kPrimFloat: { 1409 uint32_t stack_index = stack_index_++; 1410 if (float_index_ % 2 == 0) { 1411 float_index_ = std::max(double_index_, float_index_); 1412 } 1413 if (float_index_ < calling_convention.GetNumberOfFpuRegisters()) { 1414 return Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(float_index_++)); 1415 } else { 1416 return Location::StackSlot(calling_convention.GetStackOffsetOf(stack_index)); 1417 } 1418 } 1419 1420 case Primitive::kPrimDouble: { 1421 double_index_ = std::max(double_index_, RoundUp(float_index_, 2)); 1422 uint32_t stack_index = stack_index_; 1423 stack_index_ += 2; 1424 if (double_index_ + 1 < calling_convention.GetNumberOfFpuRegisters()) { 1425 uint32_t index = double_index_; 1426 double_index_ += 2; 1427 Location result = Location::FpuRegisterPairLocation( 1428 calling_convention.GetFpuRegisterAt(index), 1429 calling_convention.GetFpuRegisterAt(index + 1)); 1430 DCHECK(ExpectedPairLayout(result)); 1431 return result; 1432 } else { 1433 return Location::DoubleStackSlot(calling_convention.GetStackOffsetOf(stack_index)); 1434 } 1435 } 1436 1437 case Primitive::kPrimVoid: 1438 LOG(FATAL) << "Unexpected parameter type " << type; 1439 break; 1440 } 1441 return Location::NoLocation(); 1442} 1443 1444Location InvokeDexCallingConventionVisitorARM::GetReturnLocation(Primitive::Type type) const { 1445 switch (type) { 1446 case Primitive::kPrimBoolean: 1447 case Primitive::kPrimByte: 1448 case Primitive::kPrimChar: 1449 case Primitive::kPrimShort: 1450 case Primitive::kPrimInt: 1451 case Primitive::kPrimNot: { 1452 return Location::RegisterLocation(R0); 1453 } 1454 1455 case Primitive::kPrimFloat: { 1456 return Location::FpuRegisterLocation(S0); 1457 } 1458 1459 case Primitive::kPrimLong: { 1460 return Location::RegisterPairLocation(R0, R1); 1461 } 1462 1463 case Primitive::kPrimDouble: { 1464 return Location::FpuRegisterPairLocation(S0, S1); 1465 } 1466 1467 case Primitive::kPrimVoid: 1468 return Location::NoLocation(); 1469 } 1470 1471 UNREACHABLE(); 1472} 1473 1474Location InvokeDexCallingConventionVisitorARM::GetMethodLocation() const { 1475 return Location::RegisterLocation(kMethodRegisterArgument); 1476} 1477 1478void CodeGeneratorARM::Move32(Location destination, Location source) { 1479 if (source.Equals(destination)) { 1480 return; 1481 } 1482 if (destination.IsRegister()) { 1483 if (source.IsRegister()) { 1484 __ Mov(destination.AsRegister<Register>(), source.AsRegister<Register>()); 1485 } else if (source.IsFpuRegister()) { 1486 __ vmovrs(destination.AsRegister<Register>(), source.AsFpuRegister<SRegister>()); 1487 } else { 1488 __ LoadFromOffset(kLoadWord, destination.AsRegister<Register>(), SP, source.GetStackIndex()); 1489 } 1490 } else if (destination.IsFpuRegister()) { 1491 if (source.IsRegister()) { 1492 __ vmovsr(destination.AsFpuRegister<SRegister>(), source.AsRegister<Register>()); 1493 } else if (source.IsFpuRegister()) { 1494 __ vmovs(destination.AsFpuRegister<SRegister>(), source.AsFpuRegister<SRegister>()); 1495 } else { 1496 __ LoadSFromOffset(destination.AsFpuRegister<SRegister>(), SP, source.GetStackIndex()); 1497 } 1498 } else { 1499 DCHECK(destination.IsStackSlot()) << destination; 1500 if (source.IsRegister()) { 1501 __ StoreToOffset(kStoreWord, source.AsRegister<Register>(), SP, destination.GetStackIndex()); 1502 } else if (source.IsFpuRegister()) { 1503 __ StoreSToOffset(source.AsFpuRegister<SRegister>(), SP, destination.GetStackIndex()); 1504 } else { 1505 DCHECK(source.IsStackSlot()) << source; 1506 __ LoadFromOffset(kLoadWord, IP, SP, source.GetStackIndex()); 1507 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 1508 } 1509 } 1510} 1511 1512void CodeGeneratorARM::Move64(Location destination, Location source) { 1513 if (source.Equals(destination)) { 1514 return; 1515 } 1516 if (destination.IsRegisterPair()) { 1517 if (source.IsRegisterPair()) { 1518 EmitParallelMoves( 1519 Location::RegisterLocation(source.AsRegisterPairHigh<Register>()), 1520 Location::RegisterLocation(destination.AsRegisterPairHigh<Register>()), 1521 Primitive::kPrimInt, 1522 Location::RegisterLocation(source.AsRegisterPairLow<Register>()), 1523 Location::RegisterLocation(destination.AsRegisterPairLow<Register>()), 1524 Primitive::kPrimInt); 1525 } else if (source.IsFpuRegister()) { 1526 UNIMPLEMENTED(FATAL); 1527 } else if (source.IsFpuRegisterPair()) { 1528 __ vmovrrd(destination.AsRegisterPairLow<Register>(), 1529 destination.AsRegisterPairHigh<Register>(), 1530 FromLowSToD(source.AsFpuRegisterPairLow<SRegister>())); 1531 } else { 1532 DCHECK(source.IsDoubleStackSlot()); 1533 DCHECK(ExpectedPairLayout(destination)); 1534 __ LoadFromOffset(kLoadWordPair, destination.AsRegisterPairLow<Register>(), 1535 SP, source.GetStackIndex()); 1536 } 1537 } else if (destination.IsFpuRegisterPair()) { 1538 if (source.IsDoubleStackSlot()) { 1539 __ LoadDFromOffset(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 1540 SP, 1541 source.GetStackIndex()); 1542 } else if (source.IsRegisterPair()) { 1543 __ vmovdrr(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 1544 source.AsRegisterPairLow<Register>(), 1545 source.AsRegisterPairHigh<Register>()); 1546 } else { 1547 UNIMPLEMENTED(FATAL); 1548 } 1549 } else { 1550 DCHECK(destination.IsDoubleStackSlot()); 1551 if (source.IsRegisterPair()) { 1552 // No conflict possible, so just do the moves. 1553 if (source.AsRegisterPairLow<Register>() == R1) { 1554 DCHECK_EQ(source.AsRegisterPairHigh<Register>(), R2); 1555 __ StoreToOffset(kStoreWord, R1, SP, destination.GetStackIndex()); 1556 __ StoreToOffset(kStoreWord, R2, SP, destination.GetHighStackIndex(kArmWordSize)); 1557 } else { 1558 __ StoreToOffset(kStoreWordPair, source.AsRegisterPairLow<Register>(), 1559 SP, destination.GetStackIndex()); 1560 } 1561 } else if (source.IsFpuRegisterPair()) { 1562 __ StoreDToOffset(FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()), 1563 SP, 1564 destination.GetStackIndex()); 1565 } else { 1566 DCHECK(source.IsDoubleStackSlot()); 1567 EmitParallelMoves( 1568 Location::StackSlot(source.GetStackIndex()), 1569 Location::StackSlot(destination.GetStackIndex()), 1570 Primitive::kPrimInt, 1571 Location::StackSlot(source.GetHighStackIndex(kArmWordSize)), 1572 Location::StackSlot(destination.GetHighStackIndex(kArmWordSize)), 1573 Primitive::kPrimInt); 1574 } 1575 } 1576} 1577 1578void CodeGeneratorARM::MoveConstant(Location location, int32_t value) { 1579 DCHECK(location.IsRegister()); 1580 __ LoadImmediate(location.AsRegister<Register>(), value); 1581} 1582 1583void CodeGeneratorARM::MoveLocation(Location dst, Location src, Primitive::Type dst_type) { 1584 HParallelMove move(GetGraph()->GetArena()); 1585 move.AddMove(src, dst, dst_type, nullptr); 1586 GetMoveResolver()->EmitNativeCode(&move); 1587} 1588 1589void CodeGeneratorARM::AddLocationAsTemp(Location location, LocationSummary* locations) { 1590 if (location.IsRegister()) { 1591 locations->AddTemp(location); 1592 } else if (location.IsRegisterPair()) { 1593 locations->AddTemp(Location::RegisterLocation(location.AsRegisterPairLow<Register>())); 1594 locations->AddTemp(Location::RegisterLocation(location.AsRegisterPairHigh<Register>())); 1595 } else { 1596 UNIMPLEMENTED(FATAL) << "AddLocationAsTemp not implemented for location " << location; 1597 } 1598} 1599 1600void CodeGeneratorARM::InvokeRuntime(QuickEntrypointEnum entrypoint, 1601 HInstruction* instruction, 1602 uint32_t dex_pc, 1603 SlowPathCode* slow_path) { 1604 ValidateInvokeRuntime(entrypoint, instruction, slow_path); 1605 GenerateInvokeRuntime(GetThreadOffset<kArmPointerSize>(entrypoint).Int32Value()); 1606 if (EntrypointRequiresStackMap(entrypoint)) { 1607 RecordPcInfo(instruction, dex_pc, slow_path); 1608 } 1609} 1610 1611void CodeGeneratorARM::InvokeRuntimeWithoutRecordingPcInfo(int32_t entry_point_offset, 1612 HInstruction* instruction, 1613 SlowPathCode* slow_path) { 1614 ValidateInvokeRuntimeWithoutRecordingPcInfo(instruction, slow_path); 1615 GenerateInvokeRuntime(entry_point_offset); 1616} 1617 1618void CodeGeneratorARM::GenerateInvokeRuntime(int32_t entry_point_offset) { 1619 __ LoadFromOffset(kLoadWord, LR, TR, entry_point_offset); 1620 __ blx(LR); 1621} 1622 1623void InstructionCodeGeneratorARM::HandleGoto(HInstruction* got, HBasicBlock* successor) { 1624 DCHECK(!successor->IsExitBlock()); 1625 1626 HBasicBlock* block = got->GetBlock(); 1627 HInstruction* previous = got->GetPrevious(); 1628 1629 HLoopInformation* info = block->GetLoopInformation(); 1630 if (info != nullptr && info->IsBackEdge(*block) && info->HasSuspendCheck()) { 1631 codegen_->ClearSpillSlotsFromLoopPhisInStackMap(info->GetSuspendCheck()); 1632 GenerateSuspendCheck(info->GetSuspendCheck(), successor); 1633 return; 1634 } 1635 1636 if (block->IsEntryBlock() && (previous != nullptr) && previous->IsSuspendCheck()) { 1637 GenerateSuspendCheck(previous->AsSuspendCheck(), nullptr); 1638 } 1639 if (!codegen_->GoesToNextBlock(got->GetBlock(), successor)) { 1640 __ b(codegen_->GetLabelOf(successor)); 1641 } 1642} 1643 1644void LocationsBuilderARM::VisitGoto(HGoto* got) { 1645 got->SetLocations(nullptr); 1646} 1647 1648void InstructionCodeGeneratorARM::VisitGoto(HGoto* got) { 1649 HandleGoto(got, got->GetSuccessor()); 1650} 1651 1652void LocationsBuilderARM::VisitTryBoundary(HTryBoundary* try_boundary) { 1653 try_boundary->SetLocations(nullptr); 1654} 1655 1656void InstructionCodeGeneratorARM::VisitTryBoundary(HTryBoundary* try_boundary) { 1657 HBasicBlock* successor = try_boundary->GetNormalFlowSuccessor(); 1658 if (!successor->IsExitBlock()) { 1659 HandleGoto(try_boundary, successor); 1660 } 1661} 1662 1663void LocationsBuilderARM::VisitExit(HExit* exit) { 1664 exit->SetLocations(nullptr); 1665} 1666 1667void InstructionCodeGeneratorARM::VisitExit(HExit* exit ATTRIBUTE_UNUSED) { 1668} 1669 1670void InstructionCodeGeneratorARM::GenerateVcmp(HInstruction* instruction) { 1671 Primitive::Type type = instruction->InputAt(0)->GetType(); 1672 Location lhs_loc = instruction->GetLocations()->InAt(0); 1673 Location rhs_loc = instruction->GetLocations()->InAt(1); 1674 if (rhs_loc.IsConstant()) { 1675 // 0.0 is the only immediate that can be encoded directly in 1676 // a VCMP instruction. 1677 // 1678 // Both the JLS (section 15.20.1) and the JVMS (section 6.5) 1679 // specify that in a floating-point comparison, positive zero 1680 // and negative zero are considered equal, so we can use the 1681 // literal 0.0 for both cases here. 1682 // 1683 // Note however that some methods (Float.equal, Float.compare, 1684 // Float.compareTo, Double.equal, Double.compare, 1685 // Double.compareTo, Math.max, Math.min, StrictMath.max, 1686 // StrictMath.min) consider 0.0 to be (strictly) greater than 1687 // -0.0. So if we ever translate calls to these methods into a 1688 // HCompare instruction, we must handle the -0.0 case with 1689 // care here. 1690 DCHECK(rhs_loc.GetConstant()->IsArithmeticZero()); 1691 if (type == Primitive::kPrimFloat) { 1692 __ vcmpsz(lhs_loc.AsFpuRegister<SRegister>()); 1693 } else { 1694 DCHECK_EQ(type, Primitive::kPrimDouble); 1695 __ vcmpdz(FromLowSToD(lhs_loc.AsFpuRegisterPairLow<SRegister>())); 1696 } 1697 } else { 1698 if (type == Primitive::kPrimFloat) { 1699 __ vcmps(lhs_loc.AsFpuRegister<SRegister>(), rhs_loc.AsFpuRegister<SRegister>()); 1700 } else { 1701 DCHECK_EQ(type, Primitive::kPrimDouble); 1702 __ vcmpd(FromLowSToD(lhs_loc.AsFpuRegisterPairLow<SRegister>()), 1703 FromLowSToD(rhs_loc.AsFpuRegisterPairLow<SRegister>())); 1704 } 1705 } 1706} 1707 1708void InstructionCodeGeneratorARM::GenerateFPJumps(HCondition* cond, 1709 Label* true_label, 1710 Label* false_label ATTRIBUTE_UNUSED) { 1711 __ vmstat(); // transfer FP status register to ARM APSR. 1712 __ b(true_label, ARMFPCondition(cond->GetCondition(), cond->IsGtBias())); 1713} 1714 1715void InstructionCodeGeneratorARM::GenerateLongComparesAndJumps(HCondition* cond, 1716 Label* true_label, 1717 Label* false_label) { 1718 LocationSummary* locations = cond->GetLocations(); 1719 Location left = locations->InAt(0); 1720 Location right = locations->InAt(1); 1721 IfCondition if_cond = cond->GetCondition(); 1722 1723 Register left_high = left.AsRegisterPairHigh<Register>(); 1724 Register left_low = left.AsRegisterPairLow<Register>(); 1725 IfCondition true_high_cond = if_cond; 1726 IfCondition false_high_cond = cond->GetOppositeCondition(); 1727 Condition final_condition = ARMUnsignedCondition(if_cond); // unsigned on lower part 1728 1729 // Set the conditions for the test, remembering that == needs to be 1730 // decided using the low words. 1731 // TODO: consider avoiding jumps with temporary and CMP low+SBC high 1732 switch (if_cond) { 1733 case kCondEQ: 1734 case kCondNE: 1735 // Nothing to do. 1736 break; 1737 case kCondLT: 1738 false_high_cond = kCondGT; 1739 break; 1740 case kCondLE: 1741 true_high_cond = kCondLT; 1742 break; 1743 case kCondGT: 1744 false_high_cond = kCondLT; 1745 break; 1746 case kCondGE: 1747 true_high_cond = kCondGT; 1748 break; 1749 case kCondB: 1750 false_high_cond = kCondA; 1751 break; 1752 case kCondBE: 1753 true_high_cond = kCondB; 1754 break; 1755 case kCondA: 1756 false_high_cond = kCondB; 1757 break; 1758 case kCondAE: 1759 true_high_cond = kCondA; 1760 break; 1761 } 1762 if (right.IsConstant()) { 1763 int64_t value = right.GetConstant()->AsLongConstant()->GetValue(); 1764 int32_t val_low = Low32Bits(value); 1765 int32_t val_high = High32Bits(value); 1766 1767 __ CmpConstant(left_high, val_high); 1768 if (if_cond == kCondNE) { 1769 __ b(true_label, ARMCondition(true_high_cond)); 1770 } else if (if_cond == kCondEQ) { 1771 __ b(false_label, ARMCondition(false_high_cond)); 1772 } else { 1773 __ b(true_label, ARMCondition(true_high_cond)); 1774 __ b(false_label, ARMCondition(false_high_cond)); 1775 } 1776 // Must be equal high, so compare the lows. 1777 __ CmpConstant(left_low, val_low); 1778 } else { 1779 Register right_high = right.AsRegisterPairHigh<Register>(); 1780 Register right_low = right.AsRegisterPairLow<Register>(); 1781 1782 __ cmp(left_high, ShifterOperand(right_high)); 1783 if (if_cond == kCondNE) { 1784 __ b(true_label, ARMCondition(true_high_cond)); 1785 } else if (if_cond == kCondEQ) { 1786 __ b(false_label, ARMCondition(false_high_cond)); 1787 } else { 1788 __ b(true_label, ARMCondition(true_high_cond)); 1789 __ b(false_label, ARMCondition(false_high_cond)); 1790 } 1791 // Must be equal high, so compare the lows. 1792 __ cmp(left_low, ShifterOperand(right_low)); 1793 } 1794 // The last comparison might be unsigned. 1795 // TODO: optimize cases where this is always true/false 1796 __ b(true_label, final_condition); 1797} 1798 1799void InstructionCodeGeneratorARM::GenerateCompareTestAndBranch(HCondition* condition, 1800 Label* true_target_in, 1801 Label* false_target_in) { 1802 // Generated branching requires both targets to be explicit. If either of the 1803 // targets is nullptr (fallthrough) use and bind `fallthrough_target` instead. 1804 Label fallthrough_target; 1805 Label* true_target = true_target_in == nullptr ? &fallthrough_target : true_target_in; 1806 Label* false_target = false_target_in == nullptr ? &fallthrough_target : false_target_in; 1807 1808 Primitive::Type type = condition->InputAt(0)->GetType(); 1809 switch (type) { 1810 case Primitive::kPrimLong: 1811 GenerateLongComparesAndJumps(condition, true_target, false_target); 1812 break; 1813 case Primitive::kPrimFloat: 1814 case Primitive::kPrimDouble: 1815 GenerateVcmp(condition); 1816 GenerateFPJumps(condition, true_target, false_target); 1817 break; 1818 default: 1819 LOG(FATAL) << "Unexpected compare type " << type; 1820 } 1821 1822 if (false_target != &fallthrough_target) { 1823 __ b(false_target); 1824 } 1825 1826 if (fallthrough_target.IsLinked()) { 1827 __ Bind(&fallthrough_target); 1828 } 1829} 1830 1831void InstructionCodeGeneratorARM::GenerateTestAndBranch(HInstruction* instruction, 1832 size_t condition_input_index, 1833 Label* true_target, 1834 Label* false_target) { 1835 HInstruction* cond = instruction->InputAt(condition_input_index); 1836 1837 if (true_target == nullptr && false_target == nullptr) { 1838 // Nothing to do. The code always falls through. 1839 return; 1840 } else if (cond->IsIntConstant()) { 1841 // Constant condition, statically compared against "true" (integer value 1). 1842 if (cond->AsIntConstant()->IsTrue()) { 1843 if (true_target != nullptr) { 1844 __ b(true_target); 1845 } 1846 } else { 1847 DCHECK(cond->AsIntConstant()->IsFalse()) << cond->AsIntConstant()->GetValue(); 1848 if (false_target != nullptr) { 1849 __ b(false_target); 1850 } 1851 } 1852 return; 1853 } 1854 1855 // The following code generates these patterns: 1856 // (1) true_target == nullptr && false_target != nullptr 1857 // - opposite condition true => branch to false_target 1858 // (2) true_target != nullptr && false_target == nullptr 1859 // - condition true => branch to true_target 1860 // (3) true_target != nullptr && false_target != nullptr 1861 // - condition true => branch to true_target 1862 // - branch to false_target 1863 if (IsBooleanValueOrMaterializedCondition(cond)) { 1864 // Condition has been materialized, compare the output to 0. 1865 Location cond_val = instruction->GetLocations()->InAt(condition_input_index); 1866 DCHECK(cond_val.IsRegister()); 1867 if (true_target == nullptr) { 1868 __ CompareAndBranchIfZero(cond_val.AsRegister<Register>(), false_target); 1869 } else { 1870 __ CompareAndBranchIfNonZero(cond_val.AsRegister<Register>(), true_target); 1871 } 1872 } else { 1873 // Condition has not been materialized. Use its inputs as the comparison and 1874 // its condition as the branch condition. 1875 HCondition* condition = cond->AsCondition(); 1876 1877 // If this is a long or FP comparison that has been folded into 1878 // the HCondition, generate the comparison directly. 1879 Primitive::Type type = condition->InputAt(0)->GetType(); 1880 if (type == Primitive::kPrimLong || Primitive::IsFloatingPointType(type)) { 1881 GenerateCompareTestAndBranch(condition, true_target, false_target); 1882 return; 1883 } 1884 1885 LocationSummary* locations = cond->GetLocations(); 1886 DCHECK(locations->InAt(0).IsRegister()); 1887 Register left = locations->InAt(0).AsRegister<Register>(); 1888 Location right = locations->InAt(1); 1889 if (right.IsRegister()) { 1890 __ cmp(left, ShifterOperand(right.AsRegister<Register>())); 1891 } else { 1892 DCHECK(right.IsConstant()); 1893 __ CmpConstant(left, CodeGenerator::GetInt32ValueOf(right.GetConstant())); 1894 } 1895 if (true_target == nullptr) { 1896 __ b(false_target, ARMCondition(condition->GetOppositeCondition())); 1897 } else { 1898 __ b(true_target, ARMCondition(condition->GetCondition())); 1899 } 1900 } 1901 1902 // If neither branch falls through (case 3), the conditional branch to `true_target` 1903 // was already emitted (case 2) and we need to emit a jump to `false_target`. 1904 if (true_target != nullptr && false_target != nullptr) { 1905 __ b(false_target); 1906 } 1907} 1908 1909void LocationsBuilderARM::VisitIf(HIf* if_instr) { 1910 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(if_instr); 1911 if (IsBooleanValueOrMaterializedCondition(if_instr->InputAt(0))) { 1912 locations->SetInAt(0, Location::RequiresRegister()); 1913 } 1914} 1915 1916void InstructionCodeGeneratorARM::VisitIf(HIf* if_instr) { 1917 HBasicBlock* true_successor = if_instr->IfTrueSuccessor(); 1918 HBasicBlock* false_successor = if_instr->IfFalseSuccessor(); 1919 Label* true_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), true_successor) ? 1920 nullptr : codegen_->GetLabelOf(true_successor); 1921 Label* false_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), false_successor) ? 1922 nullptr : codegen_->GetLabelOf(false_successor); 1923 GenerateTestAndBranch(if_instr, /* condition_input_index */ 0, true_target, false_target); 1924} 1925 1926void LocationsBuilderARM::VisitDeoptimize(HDeoptimize* deoptimize) { 1927 LocationSummary* locations = new (GetGraph()->GetArena()) 1928 LocationSummary(deoptimize, LocationSummary::kCallOnSlowPath); 1929 locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty()); // No caller-save registers. 1930 if (IsBooleanValueOrMaterializedCondition(deoptimize->InputAt(0))) { 1931 locations->SetInAt(0, Location::RequiresRegister()); 1932 } 1933} 1934 1935void InstructionCodeGeneratorARM::VisitDeoptimize(HDeoptimize* deoptimize) { 1936 SlowPathCodeARM* slow_path = deopt_slow_paths_.NewSlowPath<DeoptimizationSlowPathARM>(deoptimize); 1937 GenerateTestAndBranch(deoptimize, 1938 /* condition_input_index */ 0, 1939 slow_path->GetEntryLabel(), 1940 /* false_target */ nullptr); 1941} 1942 1943void LocationsBuilderARM::VisitSelect(HSelect* select) { 1944 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(select); 1945 if (Primitive::IsFloatingPointType(select->GetType())) { 1946 locations->SetInAt(0, Location::RequiresFpuRegister()); 1947 locations->SetInAt(1, Location::RequiresFpuRegister()); 1948 } else { 1949 locations->SetInAt(0, Location::RequiresRegister()); 1950 locations->SetInAt(1, Location::RequiresRegister()); 1951 } 1952 if (IsBooleanValueOrMaterializedCondition(select->GetCondition())) { 1953 locations->SetInAt(2, Location::RequiresRegister()); 1954 } 1955 locations->SetOut(Location::SameAsFirstInput()); 1956} 1957 1958void InstructionCodeGeneratorARM::VisitSelect(HSelect* select) { 1959 LocationSummary* locations = select->GetLocations(); 1960 Label false_target; 1961 GenerateTestAndBranch(select, 1962 /* condition_input_index */ 2, 1963 /* true_target */ nullptr, 1964 &false_target); 1965 codegen_->MoveLocation(locations->Out(), locations->InAt(1), select->GetType()); 1966 __ Bind(&false_target); 1967} 1968 1969void LocationsBuilderARM::VisitNativeDebugInfo(HNativeDebugInfo* info) { 1970 new (GetGraph()->GetArena()) LocationSummary(info); 1971} 1972 1973void InstructionCodeGeneratorARM::VisitNativeDebugInfo(HNativeDebugInfo*) { 1974 // MaybeRecordNativeDebugInfo is already called implicitly in CodeGenerator::Compile. 1975} 1976 1977void CodeGeneratorARM::GenerateNop() { 1978 __ nop(); 1979} 1980 1981void LocationsBuilderARM::HandleCondition(HCondition* cond) { 1982 LocationSummary* locations = 1983 new (GetGraph()->GetArena()) LocationSummary(cond, LocationSummary::kNoCall); 1984 // Handle the long/FP comparisons made in instruction simplification. 1985 switch (cond->InputAt(0)->GetType()) { 1986 case Primitive::kPrimLong: 1987 locations->SetInAt(0, Location::RequiresRegister()); 1988 locations->SetInAt(1, Location::RegisterOrConstant(cond->InputAt(1))); 1989 if (!cond->IsEmittedAtUseSite()) { 1990 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 1991 } 1992 break; 1993 1994 case Primitive::kPrimFloat: 1995 case Primitive::kPrimDouble: 1996 locations->SetInAt(0, Location::RequiresFpuRegister()); 1997 locations->SetInAt(1, ArithmeticZeroOrFpuRegister(cond->InputAt(1))); 1998 if (!cond->IsEmittedAtUseSite()) { 1999 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2000 } 2001 break; 2002 2003 default: 2004 locations->SetInAt(0, Location::RequiresRegister()); 2005 locations->SetInAt(1, Location::RegisterOrConstant(cond->InputAt(1))); 2006 if (!cond->IsEmittedAtUseSite()) { 2007 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2008 } 2009 } 2010} 2011 2012void InstructionCodeGeneratorARM::HandleCondition(HCondition* cond) { 2013 if (cond->IsEmittedAtUseSite()) { 2014 return; 2015 } 2016 2017 LocationSummary* locations = cond->GetLocations(); 2018 Location left = locations->InAt(0); 2019 Location right = locations->InAt(1); 2020 Register out = locations->Out().AsRegister<Register>(); 2021 Label true_label, false_label; 2022 2023 switch (cond->InputAt(0)->GetType()) { 2024 default: { 2025 // Integer case. 2026 if (right.IsRegister()) { 2027 __ cmp(left.AsRegister<Register>(), ShifterOperand(right.AsRegister<Register>())); 2028 } else { 2029 DCHECK(right.IsConstant()); 2030 __ CmpConstant(left.AsRegister<Register>(), 2031 CodeGenerator::GetInt32ValueOf(right.GetConstant())); 2032 } 2033 __ it(ARMCondition(cond->GetCondition()), kItElse); 2034 __ mov(locations->Out().AsRegister<Register>(), ShifterOperand(1), 2035 ARMCondition(cond->GetCondition())); 2036 __ mov(locations->Out().AsRegister<Register>(), ShifterOperand(0), 2037 ARMCondition(cond->GetOppositeCondition())); 2038 return; 2039 } 2040 case Primitive::kPrimLong: 2041 GenerateLongComparesAndJumps(cond, &true_label, &false_label); 2042 break; 2043 case Primitive::kPrimFloat: 2044 case Primitive::kPrimDouble: 2045 GenerateVcmp(cond); 2046 GenerateFPJumps(cond, &true_label, &false_label); 2047 break; 2048 } 2049 2050 // Convert the jumps into the result. 2051 Label done_label; 2052 2053 // False case: result = 0. 2054 __ Bind(&false_label); 2055 __ LoadImmediate(out, 0); 2056 __ b(&done_label); 2057 2058 // True case: result = 1. 2059 __ Bind(&true_label); 2060 __ LoadImmediate(out, 1); 2061 __ Bind(&done_label); 2062} 2063 2064void LocationsBuilderARM::VisitEqual(HEqual* comp) { 2065 HandleCondition(comp); 2066} 2067 2068void InstructionCodeGeneratorARM::VisitEqual(HEqual* comp) { 2069 HandleCondition(comp); 2070} 2071 2072void LocationsBuilderARM::VisitNotEqual(HNotEqual* comp) { 2073 HandleCondition(comp); 2074} 2075 2076void InstructionCodeGeneratorARM::VisitNotEqual(HNotEqual* comp) { 2077 HandleCondition(comp); 2078} 2079 2080void LocationsBuilderARM::VisitLessThan(HLessThan* comp) { 2081 HandleCondition(comp); 2082} 2083 2084void InstructionCodeGeneratorARM::VisitLessThan(HLessThan* comp) { 2085 HandleCondition(comp); 2086} 2087 2088void LocationsBuilderARM::VisitLessThanOrEqual(HLessThanOrEqual* comp) { 2089 HandleCondition(comp); 2090} 2091 2092void InstructionCodeGeneratorARM::VisitLessThanOrEqual(HLessThanOrEqual* comp) { 2093 HandleCondition(comp); 2094} 2095 2096void LocationsBuilderARM::VisitGreaterThan(HGreaterThan* comp) { 2097 HandleCondition(comp); 2098} 2099 2100void InstructionCodeGeneratorARM::VisitGreaterThan(HGreaterThan* comp) { 2101 HandleCondition(comp); 2102} 2103 2104void LocationsBuilderARM::VisitGreaterThanOrEqual(HGreaterThanOrEqual* comp) { 2105 HandleCondition(comp); 2106} 2107 2108void InstructionCodeGeneratorARM::VisitGreaterThanOrEqual(HGreaterThanOrEqual* comp) { 2109 HandleCondition(comp); 2110} 2111 2112void LocationsBuilderARM::VisitBelow(HBelow* comp) { 2113 HandleCondition(comp); 2114} 2115 2116void InstructionCodeGeneratorARM::VisitBelow(HBelow* comp) { 2117 HandleCondition(comp); 2118} 2119 2120void LocationsBuilderARM::VisitBelowOrEqual(HBelowOrEqual* comp) { 2121 HandleCondition(comp); 2122} 2123 2124void InstructionCodeGeneratorARM::VisitBelowOrEqual(HBelowOrEqual* comp) { 2125 HandleCondition(comp); 2126} 2127 2128void LocationsBuilderARM::VisitAbove(HAbove* comp) { 2129 HandleCondition(comp); 2130} 2131 2132void InstructionCodeGeneratorARM::VisitAbove(HAbove* comp) { 2133 HandleCondition(comp); 2134} 2135 2136void LocationsBuilderARM::VisitAboveOrEqual(HAboveOrEqual* comp) { 2137 HandleCondition(comp); 2138} 2139 2140void InstructionCodeGeneratorARM::VisitAboveOrEqual(HAboveOrEqual* comp) { 2141 HandleCondition(comp); 2142} 2143 2144void LocationsBuilderARM::VisitIntConstant(HIntConstant* constant) { 2145 LocationSummary* locations = 2146 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 2147 locations->SetOut(Location::ConstantLocation(constant)); 2148} 2149 2150void InstructionCodeGeneratorARM::VisitIntConstant(HIntConstant* constant ATTRIBUTE_UNUSED) { 2151 // Will be generated at use site. 2152} 2153 2154void LocationsBuilderARM::VisitNullConstant(HNullConstant* constant) { 2155 LocationSummary* locations = 2156 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 2157 locations->SetOut(Location::ConstantLocation(constant)); 2158} 2159 2160void InstructionCodeGeneratorARM::VisitNullConstant(HNullConstant* constant ATTRIBUTE_UNUSED) { 2161 // Will be generated at use site. 2162} 2163 2164void LocationsBuilderARM::VisitLongConstant(HLongConstant* constant) { 2165 LocationSummary* locations = 2166 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 2167 locations->SetOut(Location::ConstantLocation(constant)); 2168} 2169 2170void InstructionCodeGeneratorARM::VisitLongConstant(HLongConstant* constant ATTRIBUTE_UNUSED) { 2171 // Will be generated at use site. 2172} 2173 2174void LocationsBuilderARM::VisitFloatConstant(HFloatConstant* constant) { 2175 LocationSummary* locations = 2176 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 2177 locations->SetOut(Location::ConstantLocation(constant)); 2178} 2179 2180void InstructionCodeGeneratorARM::VisitFloatConstant(HFloatConstant* constant ATTRIBUTE_UNUSED) { 2181 // Will be generated at use site. 2182} 2183 2184void LocationsBuilderARM::VisitDoubleConstant(HDoubleConstant* constant) { 2185 LocationSummary* locations = 2186 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 2187 locations->SetOut(Location::ConstantLocation(constant)); 2188} 2189 2190void InstructionCodeGeneratorARM::VisitDoubleConstant(HDoubleConstant* constant ATTRIBUTE_UNUSED) { 2191 // Will be generated at use site. 2192} 2193 2194void LocationsBuilderARM::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) { 2195 memory_barrier->SetLocations(nullptr); 2196} 2197 2198void InstructionCodeGeneratorARM::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) { 2199 codegen_->GenerateMemoryBarrier(memory_barrier->GetBarrierKind()); 2200} 2201 2202void LocationsBuilderARM::VisitReturnVoid(HReturnVoid* ret) { 2203 ret->SetLocations(nullptr); 2204} 2205 2206void InstructionCodeGeneratorARM::VisitReturnVoid(HReturnVoid* ret ATTRIBUTE_UNUSED) { 2207 codegen_->GenerateFrameExit(); 2208} 2209 2210void LocationsBuilderARM::VisitReturn(HReturn* ret) { 2211 LocationSummary* locations = 2212 new (GetGraph()->GetArena()) LocationSummary(ret, LocationSummary::kNoCall); 2213 locations->SetInAt(0, parameter_visitor_.GetReturnLocation(ret->InputAt(0)->GetType())); 2214} 2215 2216void InstructionCodeGeneratorARM::VisitReturn(HReturn* ret ATTRIBUTE_UNUSED) { 2217 codegen_->GenerateFrameExit(); 2218} 2219 2220void LocationsBuilderARM::VisitInvokeUnresolved(HInvokeUnresolved* invoke) { 2221 // The trampoline uses the same calling convention as dex calling conventions, 2222 // except instead of loading arg0/r0 with the target Method*, arg0/r0 will contain 2223 // the method_idx. 2224 HandleInvoke(invoke); 2225} 2226 2227void InstructionCodeGeneratorARM::VisitInvokeUnresolved(HInvokeUnresolved* invoke) { 2228 codegen_->GenerateInvokeUnresolvedRuntimeCall(invoke); 2229} 2230 2231void LocationsBuilderARM::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) { 2232 // Explicit clinit checks triggered by static invokes must have been pruned by 2233 // art::PrepareForRegisterAllocation. 2234 DCHECK(!invoke->IsStaticWithExplicitClinitCheck()); 2235 2236 IntrinsicLocationsBuilderARM intrinsic(codegen_); 2237 if (intrinsic.TryDispatch(invoke)) { 2238 if (invoke->GetLocations()->CanCall() && invoke->HasPcRelativeDexCache()) { 2239 invoke->GetLocations()->SetInAt(invoke->GetSpecialInputIndex(), Location::Any()); 2240 } 2241 return; 2242 } 2243 2244 HandleInvoke(invoke); 2245 2246 // For PC-relative dex cache the invoke has an extra input, the PC-relative address base. 2247 if (invoke->HasPcRelativeDexCache()) { 2248 invoke->GetLocations()->SetInAt(invoke->GetSpecialInputIndex(), Location::RequiresRegister()); 2249 } 2250} 2251 2252static bool TryGenerateIntrinsicCode(HInvoke* invoke, CodeGeneratorARM* codegen) { 2253 if (invoke->GetLocations()->Intrinsified()) { 2254 IntrinsicCodeGeneratorARM intrinsic(codegen); 2255 intrinsic.Dispatch(invoke); 2256 return true; 2257 } 2258 return false; 2259} 2260 2261void InstructionCodeGeneratorARM::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) { 2262 // Explicit clinit checks triggered by static invokes must have been pruned by 2263 // art::PrepareForRegisterAllocation. 2264 DCHECK(!invoke->IsStaticWithExplicitClinitCheck()); 2265 2266 if (TryGenerateIntrinsicCode(invoke, codegen_)) { 2267 return; 2268 } 2269 2270 LocationSummary* locations = invoke->GetLocations(); 2271 codegen_->GenerateStaticOrDirectCall( 2272 invoke, locations->HasTemps() ? locations->GetTemp(0) : Location::NoLocation()); 2273 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 2274} 2275 2276void LocationsBuilderARM::HandleInvoke(HInvoke* invoke) { 2277 InvokeDexCallingConventionVisitorARM calling_convention_visitor; 2278 CodeGenerator::CreateCommonInvokeLocationSummary(invoke, &calling_convention_visitor); 2279} 2280 2281void LocationsBuilderARM::VisitInvokeVirtual(HInvokeVirtual* invoke) { 2282 IntrinsicLocationsBuilderARM intrinsic(codegen_); 2283 if (intrinsic.TryDispatch(invoke)) { 2284 return; 2285 } 2286 2287 HandleInvoke(invoke); 2288} 2289 2290void InstructionCodeGeneratorARM::VisitInvokeVirtual(HInvokeVirtual* invoke) { 2291 if (TryGenerateIntrinsicCode(invoke, codegen_)) { 2292 return; 2293 } 2294 2295 codegen_->GenerateVirtualCall(invoke, invoke->GetLocations()->GetTemp(0)); 2296 DCHECK(!codegen_->IsLeafMethod()); 2297 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 2298} 2299 2300void LocationsBuilderARM::VisitInvokeInterface(HInvokeInterface* invoke) { 2301 HandleInvoke(invoke); 2302 // Add the hidden argument. 2303 invoke->GetLocations()->AddTemp(Location::RegisterLocation(R12)); 2304} 2305 2306void InstructionCodeGeneratorARM::VisitInvokeInterface(HInvokeInterface* invoke) { 2307 // TODO: b/18116999, our IMTs can miss an IncompatibleClassChangeError. 2308 LocationSummary* locations = invoke->GetLocations(); 2309 Register temp = locations->GetTemp(0).AsRegister<Register>(); 2310 Register hidden_reg = locations->GetTemp(1).AsRegister<Register>(); 2311 Location receiver = locations->InAt(0); 2312 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 2313 2314 // Set the hidden argument. This is safe to do this here, as R12 2315 // won't be modified thereafter, before the `blx` (call) instruction. 2316 DCHECK_EQ(R12, hidden_reg); 2317 __ LoadImmediate(hidden_reg, invoke->GetDexMethodIndex()); 2318 2319 if (receiver.IsStackSlot()) { 2320 __ LoadFromOffset(kLoadWord, temp, SP, receiver.GetStackIndex()); 2321 // /* HeapReference<Class> */ temp = temp->klass_ 2322 __ LoadFromOffset(kLoadWord, temp, temp, class_offset); 2323 } else { 2324 // /* HeapReference<Class> */ temp = receiver->klass_ 2325 __ LoadFromOffset(kLoadWord, temp, receiver.AsRegister<Register>(), class_offset); 2326 } 2327 codegen_->MaybeRecordImplicitNullCheck(invoke); 2328 // Instead of simply (possibly) unpoisoning `temp` here, we should 2329 // emit a read barrier for the previous class reference load. 2330 // However this is not required in practice, as this is an 2331 // intermediate/temporary reference and because the current 2332 // concurrent copying collector keeps the from-space memory 2333 // intact/accessible until the end of the marking phase (the 2334 // concurrent copying collector may not in the future). 2335 __ MaybeUnpoisonHeapReference(temp); 2336 __ LoadFromOffset(kLoadWord, temp, temp, 2337 mirror::Class::ImtPtrOffset(kArmPointerSize).Uint32Value()); 2338 uint32_t method_offset = static_cast<uint32_t>(ImTable::OffsetOfElement( 2339 invoke->GetImtIndex(), kArmPointerSize)); 2340 // temp = temp->GetImtEntryAt(method_offset); 2341 __ LoadFromOffset(kLoadWord, temp, temp, method_offset); 2342 uint32_t entry_point = 2343 ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmPointerSize).Int32Value(); 2344 // LR = temp->GetEntryPoint(); 2345 __ LoadFromOffset(kLoadWord, LR, temp, entry_point); 2346 // LR(); 2347 __ blx(LR); 2348 DCHECK(!codegen_->IsLeafMethod()); 2349 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 2350} 2351 2352void LocationsBuilderARM::VisitNeg(HNeg* neg) { 2353 LocationSummary* locations = 2354 new (GetGraph()->GetArena()) LocationSummary(neg, LocationSummary::kNoCall); 2355 switch (neg->GetResultType()) { 2356 case Primitive::kPrimInt: { 2357 locations->SetInAt(0, Location::RequiresRegister()); 2358 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2359 break; 2360 } 2361 case Primitive::kPrimLong: { 2362 locations->SetInAt(0, Location::RequiresRegister()); 2363 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 2364 break; 2365 } 2366 2367 case Primitive::kPrimFloat: 2368 case Primitive::kPrimDouble: 2369 locations->SetInAt(0, Location::RequiresFpuRegister()); 2370 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2371 break; 2372 2373 default: 2374 LOG(FATAL) << "Unexpected neg type " << neg->GetResultType(); 2375 } 2376} 2377 2378void InstructionCodeGeneratorARM::VisitNeg(HNeg* neg) { 2379 LocationSummary* locations = neg->GetLocations(); 2380 Location out = locations->Out(); 2381 Location in = locations->InAt(0); 2382 switch (neg->GetResultType()) { 2383 case Primitive::kPrimInt: 2384 DCHECK(in.IsRegister()); 2385 __ rsb(out.AsRegister<Register>(), in.AsRegister<Register>(), ShifterOperand(0)); 2386 break; 2387 2388 case Primitive::kPrimLong: 2389 DCHECK(in.IsRegisterPair()); 2390 // out.lo = 0 - in.lo (and update the carry/borrow (C) flag) 2391 __ rsbs(out.AsRegisterPairLow<Register>(), 2392 in.AsRegisterPairLow<Register>(), 2393 ShifterOperand(0)); 2394 // We cannot emit an RSC (Reverse Subtract with Carry) 2395 // instruction here, as it does not exist in the Thumb-2 2396 // instruction set. We use the following approach 2397 // using SBC and SUB instead. 2398 // 2399 // out.hi = -C 2400 __ sbc(out.AsRegisterPairHigh<Register>(), 2401 out.AsRegisterPairHigh<Register>(), 2402 ShifterOperand(out.AsRegisterPairHigh<Register>())); 2403 // out.hi = out.hi - in.hi 2404 __ sub(out.AsRegisterPairHigh<Register>(), 2405 out.AsRegisterPairHigh<Register>(), 2406 ShifterOperand(in.AsRegisterPairHigh<Register>())); 2407 break; 2408 2409 case Primitive::kPrimFloat: 2410 DCHECK(in.IsFpuRegister()); 2411 __ vnegs(out.AsFpuRegister<SRegister>(), in.AsFpuRegister<SRegister>()); 2412 break; 2413 2414 case Primitive::kPrimDouble: 2415 DCHECK(in.IsFpuRegisterPair()); 2416 __ vnegd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2417 FromLowSToD(in.AsFpuRegisterPairLow<SRegister>())); 2418 break; 2419 2420 default: 2421 LOG(FATAL) << "Unexpected neg type " << neg->GetResultType(); 2422 } 2423} 2424 2425void LocationsBuilderARM::VisitTypeConversion(HTypeConversion* conversion) { 2426 Primitive::Type result_type = conversion->GetResultType(); 2427 Primitive::Type input_type = conversion->GetInputType(); 2428 DCHECK_NE(result_type, input_type); 2429 2430 // The float-to-long, double-to-long and long-to-float type conversions 2431 // rely on a call to the runtime. 2432 LocationSummary::CallKind call_kind = 2433 (((input_type == Primitive::kPrimFloat || input_type == Primitive::kPrimDouble) 2434 && result_type == Primitive::kPrimLong) 2435 || (input_type == Primitive::kPrimLong && result_type == Primitive::kPrimFloat)) 2436 ? LocationSummary::kCallOnMainOnly 2437 : LocationSummary::kNoCall; 2438 LocationSummary* locations = 2439 new (GetGraph()->GetArena()) LocationSummary(conversion, call_kind); 2440 2441 // The Java language does not allow treating boolean as an integral type but 2442 // our bit representation makes it safe. 2443 2444 switch (result_type) { 2445 case Primitive::kPrimByte: 2446 switch (input_type) { 2447 case Primitive::kPrimLong: 2448 // Type conversion from long to byte is a result of code transformations. 2449 case Primitive::kPrimBoolean: 2450 // Boolean input is a result of code transformations. 2451 case Primitive::kPrimShort: 2452 case Primitive::kPrimInt: 2453 case Primitive::kPrimChar: 2454 // Processing a Dex `int-to-byte' instruction. 2455 locations->SetInAt(0, Location::RequiresRegister()); 2456 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2457 break; 2458 2459 default: 2460 LOG(FATAL) << "Unexpected type conversion from " << input_type 2461 << " to " << result_type; 2462 } 2463 break; 2464 2465 case Primitive::kPrimShort: 2466 switch (input_type) { 2467 case Primitive::kPrimLong: 2468 // Type conversion from long to short is a result of code transformations. 2469 case Primitive::kPrimBoolean: 2470 // Boolean input is a result of code transformations. 2471 case Primitive::kPrimByte: 2472 case Primitive::kPrimInt: 2473 case Primitive::kPrimChar: 2474 // Processing a Dex `int-to-short' instruction. 2475 locations->SetInAt(0, Location::RequiresRegister()); 2476 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2477 break; 2478 2479 default: 2480 LOG(FATAL) << "Unexpected type conversion from " << input_type 2481 << " to " << result_type; 2482 } 2483 break; 2484 2485 case Primitive::kPrimInt: 2486 switch (input_type) { 2487 case Primitive::kPrimLong: 2488 // Processing a Dex `long-to-int' instruction. 2489 locations->SetInAt(0, Location::Any()); 2490 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2491 break; 2492 2493 case Primitive::kPrimFloat: 2494 // Processing a Dex `float-to-int' instruction. 2495 locations->SetInAt(0, Location::RequiresFpuRegister()); 2496 locations->SetOut(Location::RequiresRegister()); 2497 locations->AddTemp(Location::RequiresFpuRegister()); 2498 break; 2499 2500 case Primitive::kPrimDouble: 2501 // Processing a Dex `double-to-int' instruction. 2502 locations->SetInAt(0, Location::RequiresFpuRegister()); 2503 locations->SetOut(Location::RequiresRegister()); 2504 locations->AddTemp(Location::RequiresFpuRegister()); 2505 break; 2506 2507 default: 2508 LOG(FATAL) << "Unexpected type conversion from " << input_type 2509 << " to " << result_type; 2510 } 2511 break; 2512 2513 case Primitive::kPrimLong: 2514 switch (input_type) { 2515 case Primitive::kPrimBoolean: 2516 // Boolean input is a result of code transformations. 2517 case Primitive::kPrimByte: 2518 case Primitive::kPrimShort: 2519 case Primitive::kPrimInt: 2520 case Primitive::kPrimChar: 2521 // Processing a Dex `int-to-long' instruction. 2522 locations->SetInAt(0, Location::RequiresRegister()); 2523 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2524 break; 2525 2526 case Primitive::kPrimFloat: { 2527 // Processing a Dex `float-to-long' instruction. 2528 InvokeRuntimeCallingConvention calling_convention; 2529 locations->SetInAt(0, Location::FpuRegisterLocation( 2530 calling_convention.GetFpuRegisterAt(0))); 2531 locations->SetOut(Location::RegisterPairLocation(R0, R1)); 2532 break; 2533 } 2534 2535 case Primitive::kPrimDouble: { 2536 // Processing a Dex `double-to-long' instruction. 2537 InvokeRuntimeCallingConvention calling_convention; 2538 locations->SetInAt(0, Location::FpuRegisterPairLocation( 2539 calling_convention.GetFpuRegisterAt(0), 2540 calling_convention.GetFpuRegisterAt(1))); 2541 locations->SetOut(Location::RegisterPairLocation(R0, R1)); 2542 break; 2543 } 2544 2545 default: 2546 LOG(FATAL) << "Unexpected type conversion from " << input_type 2547 << " to " << result_type; 2548 } 2549 break; 2550 2551 case Primitive::kPrimChar: 2552 switch (input_type) { 2553 case Primitive::kPrimLong: 2554 // Type conversion from long to char is a result of code transformations. 2555 case Primitive::kPrimBoolean: 2556 // Boolean input is a result of code transformations. 2557 case Primitive::kPrimByte: 2558 case Primitive::kPrimShort: 2559 case Primitive::kPrimInt: 2560 // Processing a Dex `int-to-char' instruction. 2561 locations->SetInAt(0, Location::RequiresRegister()); 2562 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2563 break; 2564 2565 default: 2566 LOG(FATAL) << "Unexpected type conversion from " << input_type 2567 << " to " << result_type; 2568 } 2569 break; 2570 2571 case Primitive::kPrimFloat: 2572 switch (input_type) { 2573 case Primitive::kPrimBoolean: 2574 // Boolean input is a result of code transformations. 2575 case Primitive::kPrimByte: 2576 case Primitive::kPrimShort: 2577 case Primitive::kPrimInt: 2578 case Primitive::kPrimChar: 2579 // Processing a Dex `int-to-float' instruction. 2580 locations->SetInAt(0, Location::RequiresRegister()); 2581 locations->SetOut(Location::RequiresFpuRegister()); 2582 break; 2583 2584 case Primitive::kPrimLong: { 2585 // Processing a Dex `long-to-float' instruction. 2586 InvokeRuntimeCallingConvention calling_convention; 2587 locations->SetInAt(0, Location::RegisterPairLocation( 2588 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1))); 2589 locations->SetOut(Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(0))); 2590 break; 2591 } 2592 2593 case Primitive::kPrimDouble: 2594 // Processing a Dex `double-to-float' instruction. 2595 locations->SetInAt(0, Location::RequiresFpuRegister()); 2596 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2597 break; 2598 2599 default: 2600 LOG(FATAL) << "Unexpected type conversion from " << input_type 2601 << " to " << result_type; 2602 }; 2603 break; 2604 2605 case Primitive::kPrimDouble: 2606 switch (input_type) { 2607 case Primitive::kPrimBoolean: 2608 // Boolean input is a result of code transformations. 2609 case Primitive::kPrimByte: 2610 case Primitive::kPrimShort: 2611 case Primitive::kPrimInt: 2612 case Primitive::kPrimChar: 2613 // Processing a Dex `int-to-double' instruction. 2614 locations->SetInAt(0, Location::RequiresRegister()); 2615 locations->SetOut(Location::RequiresFpuRegister()); 2616 break; 2617 2618 case Primitive::kPrimLong: 2619 // Processing a Dex `long-to-double' instruction. 2620 locations->SetInAt(0, Location::RequiresRegister()); 2621 locations->SetOut(Location::RequiresFpuRegister()); 2622 locations->AddTemp(Location::RequiresFpuRegister()); 2623 locations->AddTemp(Location::RequiresFpuRegister()); 2624 break; 2625 2626 case Primitive::kPrimFloat: 2627 // Processing a Dex `float-to-double' instruction. 2628 locations->SetInAt(0, Location::RequiresFpuRegister()); 2629 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2630 break; 2631 2632 default: 2633 LOG(FATAL) << "Unexpected type conversion from " << input_type 2634 << " to " << result_type; 2635 }; 2636 break; 2637 2638 default: 2639 LOG(FATAL) << "Unexpected type conversion from " << input_type 2640 << " to " << result_type; 2641 } 2642} 2643 2644void InstructionCodeGeneratorARM::VisitTypeConversion(HTypeConversion* conversion) { 2645 LocationSummary* locations = conversion->GetLocations(); 2646 Location out = locations->Out(); 2647 Location in = locations->InAt(0); 2648 Primitive::Type result_type = conversion->GetResultType(); 2649 Primitive::Type input_type = conversion->GetInputType(); 2650 DCHECK_NE(result_type, input_type); 2651 switch (result_type) { 2652 case Primitive::kPrimByte: 2653 switch (input_type) { 2654 case Primitive::kPrimLong: 2655 // Type conversion from long to byte is a result of code transformations. 2656 __ sbfx(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>(), 0, 8); 2657 break; 2658 case Primitive::kPrimBoolean: 2659 // Boolean input is a result of code transformations. 2660 case Primitive::kPrimShort: 2661 case Primitive::kPrimInt: 2662 case Primitive::kPrimChar: 2663 // Processing a Dex `int-to-byte' instruction. 2664 __ sbfx(out.AsRegister<Register>(), in.AsRegister<Register>(), 0, 8); 2665 break; 2666 2667 default: 2668 LOG(FATAL) << "Unexpected type conversion from " << input_type 2669 << " to " << result_type; 2670 } 2671 break; 2672 2673 case Primitive::kPrimShort: 2674 switch (input_type) { 2675 case Primitive::kPrimLong: 2676 // Type conversion from long to short is a result of code transformations. 2677 __ sbfx(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>(), 0, 16); 2678 break; 2679 case Primitive::kPrimBoolean: 2680 // Boolean input is a result of code transformations. 2681 case Primitive::kPrimByte: 2682 case Primitive::kPrimInt: 2683 case Primitive::kPrimChar: 2684 // Processing a Dex `int-to-short' instruction. 2685 __ sbfx(out.AsRegister<Register>(), in.AsRegister<Register>(), 0, 16); 2686 break; 2687 2688 default: 2689 LOG(FATAL) << "Unexpected type conversion from " << input_type 2690 << " to " << result_type; 2691 } 2692 break; 2693 2694 case Primitive::kPrimInt: 2695 switch (input_type) { 2696 case Primitive::kPrimLong: 2697 // Processing a Dex `long-to-int' instruction. 2698 DCHECK(out.IsRegister()); 2699 if (in.IsRegisterPair()) { 2700 __ Mov(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>()); 2701 } else if (in.IsDoubleStackSlot()) { 2702 __ LoadFromOffset(kLoadWord, out.AsRegister<Register>(), SP, in.GetStackIndex()); 2703 } else { 2704 DCHECK(in.IsConstant()); 2705 DCHECK(in.GetConstant()->IsLongConstant()); 2706 int64_t value = in.GetConstant()->AsLongConstant()->GetValue(); 2707 __ LoadImmediate(out.AsRegister<Register>(), static_cast<int32_t>(value)); 2708 } 2709 break; 2710 2711 case Primitive::kPrimFloat: { 2712 // Processing a Dex `float-to-int' instruction. 2713 SRegister temp = locations->GetTemp(0).AsFpuRegisterPairLow<SRegister>(); 2714 __ vcvtis(temp, in.AsFpuRegister<SRegister>()); 2715 __ vmovrs(out.AsRegister<Register>(), temp); 2716 break; 2717 } 2718 2719 case Primitive::kPrimDouble: { 2720 // Processing a Dex `double-to-int' instruction. 2721 SRegister temp_s = locations->GetTemp(0).AsFpuRegisterPairLow<SRegister>(); 2722 __ vcvtid(temp_s, FromLowSToD(in.AsFpuRegisterPairLow<SRegister>())); 2723 __ vmovrs(out.AsRegister<Register>(), temp_s); 2724 break; 2725 } 2726 2727 default: 2728 LOG(FATAL) << "Unexpected type conversion from " << input_type 2729 << " to " << result_type; 2730 } 2731 break; 2732 2733 case Primitive::kPrimLong: 2734 switch (input_type) { 2735 case Primitive::kPrimBoolean: 2736 // Boolean input is a result of code transformations. 2737 case Primitive::kPrimByte: 2738 case Primitive::kPrimShort: 2739 case Primitive::kPrimInt: 2740 case Primitive::kPrimChar: 2741 // Processing a Dex `int-to-long' instruction. 2742 DCHECK(out.IsRegisterPair()); 2743 DCHECK(in.IsRegister()); 2744 __ Mov(out.AsRegisterPairLow<Register>(), in.AsRegister<Register>()); 2745 // Sign extension. 2746 __ Asr(out.AsRegisterPairHigh<Register>(), 2747 out.AsRegisterPairLow<Register>(), 2748 31); 2749 break; 2750 2751 case Primitive::kPrimFloat: 2752 // Processing a Dex `float-to-long' instruction. 2753 codegen_->InvokeRuntime(kQuickF2l, conversion, conversion->GetDexPc()); 2754 CheckEntrypointTypes<kQuickF2l, int64_t, float>(); 2755 break; 2756 2757 case Primitive::kPrimDouble: 2758 // Processing a Dex `double-to-long' instruction. 2759 codegen_->InvokeRuntime(kQuickD2l, conversion, conversion->GetDexPc()); 2760 CheckEntrypointTypes<kQuickD2l, int64_t, double>(); 2761 break; 2762 2763 default: 2764 LOG(FATAL) << "Unexpected type conversion from " << input_type 2765 << " to " << result_type; 2766 } 2767 break; 2768 2769 case Primitive::kPrimChar: 2770 switch (input_type) { 2771 case Primitive::kPrimLong: 2772 // Type conversion from long to char is a result of code transformations. 2773 __ ubfx(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>(), 0, 16); 2774 break; 2775 case Primitive::kPrimBoolean: 2776 // Boolean input is a result of code transformations. 2777 case Primitive::kPrimByte: 2778 case Primitive::kPrimShort: 2779 case Primitive::kPrimInt: 2780 // Processing a Dex `int-to-char' instruction. 2781 __ ubfx(out.AsRegister<Register>(), in.AsRegister<Register>(), 0, 16); 2782 break; 2783 2784 default: 2785 LOG(FATAL) << "Unexpected type conversion from " << input_type 2786 << " to " << result_type; 2787 } 2788 break; 2789 2790 case Primitive::kPrimFloat: 2791 switch (input_type) { 2792 case Primitive::kPrimBoolean: 2793 // Boolean input is a result of code transformations. 2794 case Primitive::kPrimByte: 2795 case Primitive::kPrimShort: 2796 case Primitive::kPrimInt: 2797 case Primitive::kPrimChar: { 2798 // Processing a Dex `int-to-float' instruction. 2799 __ vmovsr(out.AsFpuRegister<SRegister>(), in.AsRegister<Register>()); 2800 __ vcvtsi(out.AsFpuRegister<SRegister>(), out.AsFpuRegister<SRegister>()); 2801 break; 2802 } 2803 2804 case Primitive::kPrimLong: 2805 // Processing a Dex `long-to-float' instruction. 2806 codegen_->InvokeRuntime(kQuickL2f, conversion, conversion->GetDexPc()); 2807 CheckEntrypointTypes<kQuickL2f, float, int64_t>(); 2808 break; 2809 2810 case Primitive::kPrimDouble: 2811 // Processing a Dex `double-to-float' instruction. 2812 __ vcvtsd(out.AsFpuRegister<SRegister>(), 2813 FromLowSToD(in.AsFpuRegisterPairLow<SRegister>())); 2814 break; 2815 2816 default: 2817 LOG(FATAL) << "Unexpected type conversion from " << input_type 2818 << " to " << result_type; 2819 }; 2820 break; 2821 2822 case Primitive::kPrimDouble: 2823 switch (input_type) { 2824 case Primitive::kPrimBoolean: 2825 // Boolean input is a result of code transformations. 2826 case Primitive::kPrimByte: 2827 case Primitive::kPrimShort: 2828 case Primitive::kPrimInt: 2829 case Primitive::kPrimChar: { 2830 // Processing a Dex `int-to-double' instruction. 2831 __ vmovsr(out.AsFpuRegisterPairLow<SRegister>(), in.AsRegister<Register>()); 2832 __ vcvtdi(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2833 out.AsFpuRegisterPairLow<SRegister>()); 2834 break; 2835 } 2836 2837 case Primitive::kPrimLong: { 2838 // Processing a Dex `long-to-double' instruction. 2839 Register low = in.AsRegisterPairLow<Register>(); 2840 Register high = in.AsRegisterPairHigh<Register>(); 2841 SRegister out_s = out.AsFpuRegisterPairLow<SRegister>(); 2842 DRegister out_d = FromLowSToD(out_s); 2843 SRegister temp_s = locations->GetTemp(0).AsFpuRegisterPairLow<SRegister>(); 2844 DRegister temp_d = FromLowSToD(temp_s); 2845 SRegister constant_s = locations->GetTemp(1).AsFpuRegisterPairLow<SRegister>(); 2846 DRegister constant_d = FromLowSToD(constant_s); 2847 2848 // temp_d = int-to-double(high) 2849 __ vmovsr(temp_s, high); 2850 __ vcvtdi(temp_d, temp_s); 2851 // constant_d = k2Pow32EncodingForDouble 2852 __ LoadDImmediate(constant_d, bit_cast<double, int64_t>(k2Pow32EncodingForDouble)); 2853 // out_d = unsigned-to-double(low) 2854 __ vmovsr(out_s, low); 2855 __ vcvtdu(out_d, out_s); 2856 // out_d += temp_d * constant_d 2857 __ vmlad(out_d, temp_d, constant_d); 2858 break; 2859 } 2860 2861 case Primitive::kPrimFloat: 2862 // Processing a Dex `float-to-double' instruction. 2863 __ vcvtds(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2864 in.AsFpuRegister<SRegister>()); 2865 break; 2866 2867 default: 2868 LOG(FATAL) << "Unexpected type conversion from " << input_type 2869 << " to " << result_type; 2870 }; 2871 break; 2872 2873 default: 2874 LOG(FATAL) << "Unexpected type conversion from " << input_type 2875 << " to " << result_type; 2876 } 2877} 2878 2879void LocationsBuilderARM::VisitAdd(HAdd* add) { 2880 LocationSummary* locations = 2881 new (GetGraph()->GetArena()) LocationSummary(add, LocationSummary::kNoCall); 2882 switch (add->GetResultType()) { 2883 case Primitive::kPrimInt: { 2884 locations->SetInAt(0, Location::RequiresRegister()); 2885 locations->SetInAt(1, Location::RegisterOrConstant(add->InputAt(1))); 2886 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2887 break; 2888 } 2889 2890 case Primitive::kPrimLong: { 2891 locations->SetInAt(0, Location::RequiresRegister()); 2892 locations->SetInAt(1, ArmEncodableConstantOrRegister(add->InputAt(1), ADD)); 2893 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2894 break; 2895 } 2896 2897 case Primitive::kPrimFloat: 2898 case Primitive::kPrimDouble: { 2899 locations->SetInAt(0, Location::RequiresFpuRegister()); 2900 locations->SetInAt(1, Location::RequiresFpuRegister()); 2901 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2902 break; 2903 } 2904 2905 default: 2906 LOG(FATAL) << "Unexpected add type " << add->GetResultType(); 2907 } 2908} 2909 2910void InstructionCodeGeneratorARM::VisitAdd(HAdd* add) { 2911 LocationSummary* locations = add->GetLocations(); 2912 Location out = locations->Out(); 2913 Location first = locations->InAt(0); 2914 Location second = locations->InAt(1); 2915 switch (add->GetResultType()) { 2916 case Primitive::kPrimInt: 2917 if (second.IsRegister()) { 2918 __ add(out.AsRegister<Register>(), 2919 first.AsRegister<Register>(), 2920 ShifterOperand(second.AsRegister<Register>())); 2921 } else { 2922 __ AddConstant(out.AsRegister<Register>(), 2923 first.AsRegister<Register>(), 2924 second.GetConstant()->AsIntConstant()->GetValue()); 2925 } 2926 break; 2927 2928 case Primitive::kPrimLong: { 2929 if (second.IsConstant()) { 2930 uint64_t value = static_cast<uint64_t>(Int64FromConstant(second.GetConstant())); 2931 GenerateAddLongConst(out, first, value); 2932 } else { 2933 DCHECK(second.IsRegisterPair()); 2934 __ adds(out.AsRegisterPairLow<Register>(), 2935 first.AsRegisterPairLow<Register>(), 2936 ShifterOperand(second.AsRegisterPairLow<Register>())); 2937 __ adc(out.AsRegisterPairHigh<Register>(), 2938 first.AsRegisterPairHigh<Register>(), 2939 ShifterOperand(second.AsRegisterPairHigh<Register>())); 2940 } 2941 break; 2942 } 2943 2944 case Primitive::kPrimFloat: 2945 __ vadds(out.AsFpuRegister<SRegister>(), 2946 first.AsFpuRegister<SRegister>(), 2947 second.AsFpuRegister<SRegister>()); 2948 break; 2949 2950 case Primitive::kPrimDouble: 2951 __ vaddd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2952 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 2953 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 2954 break; 2955 2956 default: 2957 LOG(FATAL) << "Unexpected add type " << add->GetResultType(); 2958 } 2959} 2960 2961void LocationsBuilderARM::VisitSub(HSub* sub) { 2962 LocationSummary* locations = 2963 new (GetGraph()->GetArena()) LocationSummary(sub, LocationSummary::kNoCall); 2964 switch (sub->GetResultType()) { 2965 case Primitive::kPrimInt: { 2966 locations->SetInAt(0, Location::RequiresRegister()); 2967 locations->SetInAt(1, Location::RegisterOrConstant(sub->InputAt(1))); 2968 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2969 break; 2970 } 2971 2972 case Primitive::kPrimLong: { 2973 locations->SetInAt(0, Location::RequiresRegister()); 2974 locations->SetInAt(1, ArmEncodableConstantOrRegister(sub->InputAt(1), SUB)); 2975 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2976 break; 2977 } 2978 case Primitive::kPrimFloat: 2979 case Primitive::kPrimDouble: { 2980 locations->SetInAt(0, Location::RequiresFpuRegister()); 2981 locations->SetInAt(1, Location::RequiresFpuRegister()); 2982 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2983 break; 2984 } 2985 default: 2986 LOG(FATAL) << "Unexpected sub type " << sub->GetResultType(); 2987 } 2988} 2989 2990void InstructionCodeGeneratorARM::VisitSub(HSub* sub) { 2991 LocationSummary* locations = sub->GetLocations(); 2992 Location out = locations->Out(); 2993 Location first = locations->InAt(0); 2994 Location second = locations->InAt(1); 2995 switch (sub->GetResultType()) { 2996 case Primitive::kPrimInt: { 2997 if (second.IsRegister()) { 2998 __ sub(out.AsRegister<Register>(), 2999 first.AsRegister<Register>(), 3000 ShifterOperand(second.AsRegister<Register>())); 3001 } else { 3002 __ AddConstant(out.AsRegister<Register>(), 3003 first.AsRegister<Register>(), 3004 -second.GetConstant()->AsIntConstant()->GetValue()); 3005 } 3006 break; 3007 } 3008 3009 case Primitive::kPrimLong: { 3010 if (second.IsConstant()) { 3011 uint64_t value = static_cast<uint64_t>(Int64FromConstant(second.GetConstant())); 3012 GenerateAddLongConst(out, first, -value); 3013 } else { 3014 DCHECK(second.IsRegisterPair()); 3015 __ subs(out.AsRegisterPairLow<Register>(), 3016 first.AsRegisterPairLow<Register>(), 3017 ShifterOperand(second.AsRegisterPairLow<Register>())); 3018 __ sbc(out.AsRegisterPairHigh<Register>(), 3019 first.AsRegisterPairHigh<Register>(), 3020 ShifterOperand(second.AsRegisterPairHigh<Register>())); 3021 } 3022 break; 3023 } 3024 3025 case Primitive::kPrimFloat: { 3026 __ vsubs(out.AsFpuRegister<SRegister>(), 3027 first.AsFpuRegister<SRegister>(), 3028 second.AsFpuRegister<SRegister>()); 3029 break; 3030 } 3031 3032 case Primitive::kPrimDouble: { 3033 __ vsubd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 3034 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 3035 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 3036 break; 3037 } 3038 3039 3040 default: 3041 LOG(FATAL) << "Unexpected sub type " << sub->GetResultType(); 3042 } 3043} 3044 3045void LocationsBuilderARM::VisitMul(HMul* mul) { 3046 LocationSummary* locations = 3047 new (GetGraph()->GetArena()) LocationSummary(mul, LocationSummary::kNoCall); 3048 switch (mul->GetResultType()) { 3049 case Primitive::kPrimInt: 3050 case Primitive::kPrimLong: { 3051 locations->SetInAt(0, Location::RequiresRegister()); 3052 locations->SetInAt(1, Location::RequiresRegister()); 3053 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3054 break; 3055 } 3056 3057 case Primitive::kPrimFloat: 3058 case Primitive::kPrimDouble: { 3059 locations->SetInAt(0, Location::RequiresFpuRegister()); 3060 locations->SetInAt(1, Location::RequiresFpuRegister()); 3061 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 3062 break; 3063 } 3064 3065 default: 3066 LOG(FATAL) << "Unexpected mul type " << mul->GetResultType(); 3067 } 3068} 3069 3070void InstructionCodeGeneratorARM::VisitMul(HMul* mul) { 3071 LocationSummary* locations = mul->GetLocations(); 3072 Location out = locations->Out(); 3073 Location first = locations->InAt(0); 3074 Location second = locations->InAt(1); 3075 switch (mul->GetResultType()) { 3076 case Primitive::kPrimInt: { 3077 __ mul(out.AsRegister<Register>(), 3078 first.AsRegister<Register>(), 3079 second.AsRegister<Register>()); 3080 break; 3081 } 3082 case Primitive::kPrimLong: { 3083 Register out_hi = out.AsRegisterPairHigh<Register>(); 3084 Register out_lo = out.AsRegisterPairLow<Register>(); 3085 Register in1_hi = first.AsRegisterPairHigh<Register>(); 3086 Register in1_lo = first.AsRegisterPairLow<Register>(); 3087 Register in2_hi = second.AsRegisterPairHigh<Register>(); 3088 Register in2_lo = second.AsRegisterPairLow<Register>(); 3089 3090 // Extra checks to protect caused by the existence of R1_R2. 3091 // The algorithm is wrong if out.hi is either in1.lo or in2.lo: 3092 // (e.g. in1=r0_r1, in2=r2_r3 and out=r1_r2); 3093 DCHECK_NE(out_hi, in1_lo); 3094 DCHECK_NE(out_hi, in2_lo); 3095 3096 // input: in1 - 64 bits, in2 - 64 bits 3097 // output: out 3098 // formula: out.hi : out.lo = (in1.lo * in2.hi + in1.hi * in2.lo)* 2^32 + in1.lo * in2.lo 3099 // parts: out.hi = in1.lo * in2.hi + in1.hi * in2.lo + (in1.lo * in2.lo)[63:32] 3100 // parts: out.lo = (in1.lo * in2.lo)[31:0] 3101 3102 // IP <- in1.lo * in2.hi 3103 __ mul(IP, in1_lo, in2_hi); 3104 // out.hi <- in1.lo * in2.hi + in1.hi * in2.lo 3105 __ mla(out_hi, in1_hi, in2_lo, IP); 3106 // out.lo <- (in1.lo * in2.lo)[31:0]; 3107 __ umull(out_lo, IP, in1_lo, in2_lo); 3108 // out.hi <- in2.hi * in1.lo + in2.lo * in1.hi + (in1.lo * in2.lo)[63:32] 3109 __ add(out_hi, out_hi, ShifterOperand(IP)); 3110 break; 3111 } 3112 3113 case Primitive::kPrimFloat: { 3114 __ vmuls(out.AsFpuRegister<SRegister>(), 3115 first.AsFpuRegister<SRegister>(), 3116 second.AsFpuRegister<SRegister>()); 3117 break; 3118 } 3119 3120 case Primitive::kPrimDouble: { 3121 __ vmuld(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 3122 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 3123 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 3124 break; 3125 } 3126 3127 default: 3128 LOG(FATAL) << "Unexpected mul type " << mul->GetResultType(); 3129 } 3130} 3131 3132void InstructionCodeGeneratorARM::DivRemOneOrMinusOne(HBinaryOperation* instruction) { 3133 DCHECK(instruction->IsDiv() || instruction->IsRem()); 3134 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 3135 3136 LocationSummary* locations = instruction->GetLocations(); 3137 Location second = locations->InAt(1); 3138 DCHECK(second.IsConstant()); 3139 3140 Register out = locations->Out().AsRegister<Register>(); 3141 Register dividend = locations->InAt(0).AsRegister<Register>(); 3142 int32_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 3143 DCHECK(imm == 1 || imm == -1); 3144 3145 if (instruction->IsRem()) { 3146 __ LoadImmediate(out, 0); 3147 } else { 3148 if (imm == 1) { 3149 __ Mov(out, dividend); 3150 } else { 3151 __ rsb(out, dividend, ShifterOperand(0)); 3152 } 3153 } 3154} 3155 3156void InstructionCodeGeneratorARM::DivRemByPowerOfTwo(HBinaryOperation* instruction) { 3157 DCHECK(instruction->IsDiv() || instruction->IsRem()); 3158 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 3159 3160 LocationSummary* locations = instruction->GetLocations(); 3161 Location second = locations->InAt(1); 3162 DCHECK(second.IsConstant()); 3163 3164 Register out = locations->Out().AsRegister<Register>(); 3165 Register dividend = locations->InAt(0).AsRegister<Register>(); 3166 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3167 int32_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 3168 uint32_t abs_imm = static_cast<uint32_t>(AbsOrMin(imm)); 3169 int ctz_imm = CTZ(abs_imm); 3170 3171 if (ctz_imm == 1) { 3172 __ Lsr(temp, dividend, 32 - ctz_imm); 3173 } else { 3174 __ Asr(temp, dividend, 31); 3175 __ Lsr(temp, temp, 32 - ctz_imm); 3176 } 3177 __ add(out, temp, ShifterOperand(dividend)); 3178 3179 if (instruction->IsDiv()) { 3180 __ Asr(out, out, ctz_imm); 3181 if (imm < 0) { 3182 __ rsb(out, out, ShifterOperand(0)); 3183 } 3184 } else { 3185 __ ubfx(out, out, 0, ctz_imm); 3186 __ sub(out, out, ShifterOperand(temp)); 3187 } 3188} 3189 3190void InstructionCodeGeneratorARM::GenerateDivRemWithAnyConstant(HBinaryOperation* instruction) { 3191 DCHECK(instruction->IsDiv() || instruction->IsRem()); 3192 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 3193 3194 LocationSummary* locations = instruction->GetLocations(); 3195 Location second = locations->InAt(1); 3196 DCHECK(second.IsConstant()); 3197 3198 Register out = locations->Out().AsRegister<Register>(); 3199 Register dividend = locations->InAt(0).AsRegister<Register>(); 3200 Register temp1 = locations->GetTemp(0).AsRegister<Register>(); 3201 Register temp2 = locations->GetTemp(1).AsRegister<Register>(); 3202 int64_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 3203 3204 int64_t magic; 3205 int shift; 3206 CalculateMagicAndShiftForDivRem(imm, false /* is_long */, &magic, &shift); 3207 3208 __ LoadImmediate(temp1, magic); 3209 __ smull(temp2, temp1, dividend, temp1); 3210 3211 if (imm > 0 && magic < 0) { 3212 __ add(temp1, temp1, ShifterOperand(dividend)); 3213 } else if (imm < 0 && magic > 0) { 3214 __ sub(temp1, temp1, ShifterOperand(dividend)); 3215 } 3216 3217 if (shift != 0) { 3218 __ Asr(temp1, temp1, shift); 3219 } 3220 3221 if (instruction->IsDiv()) { 3222 __ sub(out, temp1, ShifterOperand(temp1, ASR, 31)); 3223 } else { 3224 __ sub(temp1, temp1, ShifterOperand(temp1, ASR, 31)); 3225 // TODO: Strength reduction for mls. 3226 __ LoadImmediate(temp2, imm); 3227 __ mls(out, temp1, temp2, dividend); 3228 } 3229} 3230 3231void InstructionCodeGeneratorARM::GenerateDivRemConstantIntegral(HBinaryOperation* instruction) { 3232 DCHECK(instruction->IsDiv() || instruction->IsRem()); 3233 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 3234 3235 LocationSummary* locations = instruction->GetLocations(); 3236 Location second = locations->InAt(1); 3237 DCHECK(second.IsConstant()); 3238 3239 int32_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 3240 if (imm == 0) { 3241 // Do not generate anything. DivZeroCheck would prevent any code to be executed. 3242 } else if (imm == 1 || imm == -1) { 3243 DivRemOneOrMinusOne(instruction); 3244 } else if (IsPowerOfTwo(AbsOrMin(imm))) { 3245 DivRemByPowerOfTwo(instruction); 3246 } else { 3247 DCHECK(imm <= -2 || imm >= 2); 3248 GenerateDivRemWithAnyConstant(instruction); 3249 } 3250} 3251 3252void LocationsBuilderARM::VisitDiv(HDiv* div) { 3253 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 3254 if (div->GetResultType() == Primitive::kPrimLong) { 3255 // pLdiv runtime call. 3256 call_kind = LocationSummary::kCallOnMainOnly; 3257 } else if (div->GetResultType() == Primitive::kPrimInt && div->InputAt(1)->IsConstant()) { 3258 // sdiv will be replaced by other instruction sequence. 3259 } else if (div->GetResultType() == Primitive::kPrimInt && 3260 !codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 3261 // pIdivmod runtime call. 3262 call_kind = LocationSummary::kCallOnMainOnly; 3263 } 3264 3265 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(div, call_kind); 3266 3267 switch (div->GetResultType()) { 3268 case Primitive::kPrimInt: { 3269 if (div->InputAt(1)->IsConstant()) { 3270 locations->SetInAt(0, Location::RequiresRegister()); 3271 locations->SetInAt(1, Location::ConstantLocation(div->InputAt(1)->AsConstant())); 3272 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3273 int32_t value = div->InputAt(1)->AsIntConstant()->GetValue(); 3274 if (value == 1 || value == 0 || value == -1) { 3275 // No temp register required. 3276 } else { 3277 locations->AddTemp(Location::RequiresRegister()); 3278 if (!IsPowerOfTwo(AbsOrMin(value))) { 3279 locations->AddTemp(Location::RequiresRegister()); 3280 } 3281 } 3282 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 3283 locations->SetInAt(0, Location::RequiresRegister()); 3284 locations->SetInAt(1, Location::RequiresRegister()); 3285 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3286 } else { 3287 InvokeRuntimeCallingConvention calling_convention; 3288 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 3289 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 3290 // Note: divrem will compute both the quotient and the remainder as the pair R0 and R1, but 3291 // we only need the former. 3292 locations->SetOut(Location::RegisterLocation(R0)); 3293 } 3294 break; 3295 } 3296 case Primitive::kPrimLong: { 3297 InvokeRuntimeCallingConvention calling_convention; 3298 locations->SetInAt(0, Location::RegisterPairLocation( 3299 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1))); 3300 locations->SetInAt(1, Location::RegisterPairLocation( 3301 calling_convention.GetRegisterAt(2), calling_convention.GetRegisterAt(3))); 3302 locations->SetOut(Location::RegisterPairLocation(R0, R1)); 3303 break; 3304 } 3305 case Primitive::kPrimFloat: 3306 case Primitive::kPrimDouble: { 3307 locations->SetInAt(0, Location::RequiresFpuRegister()); 3308 locations->SetInAt(1, Location::RequiresFpuRegister()); 3309 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 3310 break; 3311 } 3312 3313 default: 3314 LOG(FATAL) << "Unexpected div type " << div->GetResultType(); 3315 } 3316} 3317 3318void InstructionCodeGeneratorARM::VisitDiv(HDiv* div) { 3319 LocationSummary* locations = div->GetLocations(); 3320 Location out = locations->Out(); 3321 Location first = locations->InAt(0); 3322 Location second = locations->InAt(1); 3323 3324 switch (div->GetResultType()) { 3325 case Primitive::kPrimInt: { 3326 if (second.IsConstant()) { 3327 GenerateDivRemConstantIntegral(div); 3328 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 3329 __ sdiv(out.AsRegister<Register>(), 3330 first.AsRegister<Register>(), 3331 second.AsRegister<Register>()); 3332 } else { 3333 InvokeRuntimeCallingConvention calling_convention; 3334 DCHECK_EQ(calling_convention.GetRegisterAt(0), first.AsRegister<Register>()); 3335 DCHECK_EQ(calling_convention.GetRegisterAt(1), second.AsRegister<Register>()); 3336 DCHECK_EQ(R0, out.AsRegister<Register>()); 3337 3338 codegen_->InvokeRuntime(kQuickIdivmod, div, div->GetDexPc()); 3339 CheckEntrypointTypes<kQuickIdivmod, int32_t, int32_t, int32_t>(); 3340 } 3341 break; 3342 } 3343 3344 case Primitive::kPrimLong: { 3345 InvokeRuntimeCallingConvention calling_convention; 3346 DCHECK_EQ(calling_convention.GetRegisterAt(0), first.AsRegisterPairLow<Register>()); 3347 DCHECK_EQ(calling_convention.GetRegisterAt(1), first.AsRegisterPairHigh<Register>()); 3348 DCHECK_EQ(calling_convention.GetRegisterAt(2), second.AsRegisterPairLow<Register>()); 3349 DCHECK_EQ(calling_convention.GetRegisterAt(3), second.AsRegisterPairHigh<Register>()); 3350 DCHECK_EQ(R0, out.AsRegisterPairLow<Register>()); 3351 DCHECK_EQ(R1, out.AsRegisterPairHigh<Register>()); 3352 3353 codegen_->InvokeRuntime(kQuickLdiv, div, div->GetDexPc()); 3354 CheckEntrypointTypes<kQuickLdiv, int64_t, int64_t, int64_t>(); 3355 break; 3356 } 3357 3358 case Primitive::kPrimFloat: { 3359 __ vdivs(out.AsFpuRegister<SRegister>(), 3360 first.AsFpuRegister<SRegister>(), 3361 second.AsFpuRegister<SRegister>()); 3362 break; 3363 } 3364 3365 case Primitive::kPrimDouble: { 3366 __ vdivd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 3367 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 3368 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 3369 break; 3370 } 3371 3372 default: 3373 LOG(FATAL) << "Unexpected div type " << div->GetResultType(); 3374 } 3375} 3376 3377void LocationsBuilderARM::VisitRem(HRem* rem) { 3378 Primitive::Type type = rem->GetResultType(); 3379 3380 // Most remainders are implemented in the runtime. 3381 LocationSummary::CallKind call_kind = LocationSummary::kCallOnMainOnly; 3382 if (rem->GetResultType() == Primitive::kPrimInt && rem->InputAt(1)->IsConstant()) { 3383 // sdiv will be replaced by other instruction sequence. 3384 call_kind = LocationSummary::kNoCall; 3385 } else if ((rem->GetResultType() == Primitive::kPrimInt) 3386 && codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 3387 // Have hardware divide instruction for int, do it with three instructions. 3388 call_kind = LocationSummary::kNoCall; 3389 } 3390 3391 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(rem, call_kind); 3392 3393 switch (type) { 3394 case Primitive::kPrimInt: { 3395 if (rem->InputAt(1)->IsConstant()) { 3396 locations->SetInAt(0, Location::RequiresRegister()); 3397 locations->SetInAt(1, Location::ConstantLocation(rem->InputAt(1)->AsConstant())); 3398 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3399 int32_t value = rem->InputAt(1)->AsIntConstant()->GetValue(); 3400 if (value == 1 || value == 0 || value == -1) { 3401 // No temp register required. 3402 } else { 3403 locations->AddTemp(Location::RequiresRegister()); 3404 if (!IsPowerOfTwo(AbsOrMin(value))) { 3405 locations->AddTemp(Location::RequiresRegister()); 3406 } 3407 } 3408 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 3409 locations->SetInAt(0, Location::RequiresRegister()); 3410 locations->SetInAt(1, Location::RequiresRegister()); 3411 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3412 locations->AddTemp(Location::RequiresRegister()); 3413 } else { 3414 InvokeRuntimeCallingConvention calling_convention; 3415 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 3416 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 3417 // Note: divrem will compute both the quotient and the remainder as the pair R0 and R1, but 3418 // we only need the latter. 3419 locations->SetOut(Location::RegisterLocation(R1)); 3420 } 3421 break; 3422 } 3423 case Primitive::kPrimLong: { 3424 InvokeRuntimeCallingConvention calling_convention; 3425 locations->SetInAt(0, Location::RegisterPairLocation( 3426 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1))); 3427 locations->SetInAt(1, Location::RegisterPairLocation( 3428 calling_convention.GetRegisterAt(2), calling_convention.GetRegisterAt(3))); 3429 // The runtime helper puts the output in R2,R3. 3430 locations->SetOut(Location::RegisterPairLocation(R2, R3)); 3431 break; 3432 } 3433 case Primitive::kPrimFloat: { 3434 InvokeRuntimeCallingConvention calling_convention; 3435 locations->SetInAt(0, Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(0))); 3436 locations->SetInAt(1, Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(1))); 3437 locations->SetOut(Location::FpuRegisterLocation(S0)); 3438 break; 3439 } 3440 3441 case Primitive::kPrimDouble: { 3442 InvokeRuntimeCallingConvention calling_convention; 3443 locations->SetInAt(0, Location::FpuRegisterPairLocation( 3444 calling_convention.GetFpuRegisterAt(0), calling_convention.GetFpuRegisterAt(1))); 3445 locations->SetInAt(1, Location::FpuRegisterPairLocation( 3446 calling_convention.GetFpuRegisterAt(2), calling_convention.GetFpuRegisterAt(3))); 3447 locations->SetOut(Location::Location::FpuRegisterPairLocation(S0, S1)); 3448 break; 3449 } 3450 3451 default: 3452 LOG(FATAL) << "Unexpected rem type " << type; 3453 } 3454} 3455 3456void InstructionCodeGeneratorARM::VisitRem(HRem* rem) { 3457 LocationSummary* locations = rem->GetLocations(); 3458 Location out = locations->Out(); 3459 Location first = locations->InAt(0); 3460 Location second = locations->InAt(1); 3461 3462 Primitive::Type type = rem->GetResultType(); 3463 switch (type) { 3464 case Primitive::kPrimInt: { 3465 if (second.IsConstant()) { 3466 GenerateDivRemConstantIntegral(rem); 3467 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 3468 Register reg1 = first.AsRegister<Register>(); 3469 Register reg2 = second.AsRegister<Register>(); 3470 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3471 3472 // temp = reg1 / reg2 (integer division) 3473 // dest = reg1 - temp * reg2 3474 __ sdiv(temp, reg1, reg2); 3475 __ mls(out.AsRegister<Register>(), temp, reg2, reg1); 3476 } else { 3477 InvokeRuntimeCallingConvention calling_convention; 3478 DCHECK_EQ(calling_convention.GetRegisterAt(0), first.AsRegister<Register>()); 3479 DCHECK_EQ(calling_convention.GetRegisterAt(1), second.AsRegister<Register>()); 3480 DCHECK_EQ(R1, out.AsRegister<Register>()); 3481 3482 codegen_->InvokeRuntime(kQuickIdivmod, rem, rem->GetDexPc()); 3483 CheckEntrypointTypes<kQuickIdivmod, int32_t, int32_t, int32_t>(); 3484 } 3485 break; 3486 } 3487 3488 case Primitive::kPrimLong: { 3489 codegen_->InvokeRuntime(kQuickLmod, rem, rem->GetDexPc()); 3490 CheckEntrypointTypes<kQuickLmod, int64_t, int64_t, int64_t>(); 3491 break; 3492 } 3493 3494 case Primitive::kPrimFloat: { 3495 codegen_->InvokeRuntime(kQuickFmodf, rem, rem->GetDexPc()); 3496 CheckEntrypointTypes<kQuickFmodf, float, float, float>(); 3497 break; 3498 } 3499 3500 case Primitive::kPrimDouble: { 3501 codegen_->InvokeRuntime(kQuickFmod, rem, rem->GetDexPc()); 3502 CheckEntrypointTypes<kQuickFmod, double, double, double>(); 3503 break; 3504 } 3505 3506 default: 3507 LOG(FATAL) << "Unexpected rem type " << type; 3508 } 3509} 3510 3511void LocationsBuilderARM::VisitDivZeroCheck(HDivZeroCheck* instruction) { 3512 LocationSummary* locations = codegen_->CreateThrowingSlowPathLocations(instruction); 3513 locations->SetInAt(0, Location::RegisterOrConstant(instruction->InputAt(0))); 3514} 3515 3516void InstructionCodeGeneratorARM::VisitDivZeroCheck(HDivZeroCheck* instruction) { 3517 SlowPathCodeARM* slow_path = new (GetGraph()->GetArena()) DivZeroCheckSlowPathARM(instruction); 3518 codegen_->AddSlowPath(slow_path); 3519 3520 LocationSummary* locations = instruction->GetLocations(); 3521 Location value = locations->InAt(0); 3522 3523 switch (instruction->GetType()) { 3524 case Primitive::kPrimBoolean: 3525 case Primitive::kPrimByte: 3526 case Primitive::kPrimChar: 3527 case Primitive::kPrimShort: 3528 case Primitive::kPrimInt: { 3529 if (value.IsRegister()) { 3530 __ CompareAndBranchIfZero(value.AsRegister<Register>(), slow_path->GetEntryLabel()); 3531 } else { 3532 DCHECK(value.IsConstant()) << value; 3533 if (value.GetConstant()->AsIntConstant()->GetValue() == 0) { 3534 __ b(slow_path->GetEntryLabel()); 3535 } 3536 } 3537 break; 3538 } 3539 case Primitive::kPrimLong: { 3540 if (value.IsRegisterPair()) { 3541 __ orrs(IP, 3542 value.AsRegisterPairLow<Register>(), 3543 ShifterOperand(value.AsRegisterPairHigh<Register>())); 3544 __ b(slow_path->GetEntryLabel(), EQ); 3545 } else { 3546 DCHECK(value.IsConstant()) << value; 3547 if (value.GetConstant()->AsLongConstant()->GetValue() == 0) { 3548 __ b(slow_path->GetEntryLabel()); 3549 } 3550 } 3551 break; 3552 default: 3553 LOG(FATAL) << "Unexpected type for HDivZeroCheck " << instruction->GetType(); 3554 } 3555 } 3556} 3557 3558void InstructionCodeGeneratorARM::HandleIntegerRotate(LocationSummary* locations) { 3559 Register in = locations->InAt(0).AsRegister<Register>(); 3560 Location rhs = locations->InAt(1); 3561 Register out = locations->Out().AsRegister<Register>(); 3562 3563 if (rhs.IsConstant()) { 3564 // Arm32 and Thumb2 assemblers require a rotation on the interval [1,31], 3565 // so map all rotations to a +ve. equivalent in that range. 3566 // (e.g. left *or* right by -2 bits == 30 bits in the same direction.) 3567 uint32_t rot = CodeGenerator::GetInt32ValueOf(rhs.GetConstant()) & 0x1F; 3568 if (rot) { 3569 // Rotate, mapping left rotations to right equivalents if necessary. 3570 // (e.g. left by 2 bits == right by 30.) 3571 __ Ror(out, in, rot); 3572 } else if (out != in) { 3573 __ Mov(out, in); 3574 } 3575 } else { 3576 __ Ror(out, in, rhs.AsRegister<Register>()); 3577 } 3578} 3579 3580// Gain some speed by mapping all Long rotates onto equivalent pairs of Integer 3581// rotates by swapping input regs (effectively rotating by the first 32-bits of 3582// a larger rotation) or flipping direction (thus treating larger right/left 3583// rotations as sub-word sized rotations in the other direction) as appropriate. 3584void InstructionCodeGeneratorARM::HandleLongRotate(LocationSummary* locations) { 3585 Register in_reg_lo = locations->InAt(0).AsRegisterPairLow<Register>(); 3586 Register in_reg_hi = locations->InAt(0).AsRegisterPairHigh<Register>(); 3587 Location rhs = locations->InAt(1); 3588 Register out_reg_lo = locations->Out().AsRegisterPairLow<Register>(); 3589 Register out_reg_hi = locations->Out().AsRegisterPairHigh<Register>(); 3590 3591 if (rhs.IsConstant()) { 3592 uint64_t rot = CodeGenerator::GetInt64ValueOf(rhs.GetConstant()); 3593 // Map all rotations to +ve. equivalents on the interval [0,63]. 3594 rot &= kMaxLongShiftDistance; 3595 // For rotates over a word in size, 'pre-rotate' by 32-bits to keep rotate 3596 // logic below to a simple pair of binary orr. 3597 // (e.g. 34 bits == in_reg swap + 2 bits right.) 3598 if (rot >= kArmBitsPerWord) { 3599 rot -= kArmBitsPerWord; 3600 std::swap(in_reg_hi, in_reg_lo); 3601 } 3602 // Rotate, or mov to out for zero or word size rotations. 3603 if (rot != 0u) { 3604 __ Lsr(out_reg_hi, in_reg_hi, rot); 3605 __ orr(out_reg_hi, out_reg_hi, ShifterOperand(in_reg_lo, arm::LSL, kArmBitsPerWord - rot)); 3606 __ Lsr(out_reg_lo, in_reg_lo, rot); 3607 __ orr(out_reg_lo, out_reg_lo, ShifterOperand(in_reg_hi, arm::LSL, kArmBitsPerWord - rot)); 3608 } else { 3609 __ Mov(out_reg_lo, in_reg_lo); 3610 __ Mov(out_reg_hi, in_reg_hi); 3611 } 3612 } else { 3613 Register shift_right = locations->GetTemp(0).AsRegister<Register>(); 3614 Register shift_left = locations->GetTemp(1).AsRegister<Register>(); 3615 Label end; 3616 Label shift_by_32_plus_shift_right; 3617 3618 __ and_(shift_right, rhs.AsRegister<Register>(), ShifterOperand(0x1F)); 3619 __ Lsrs(shift_left, rhs.AsRegister<Register>(), 6); 3620 __ rsb(shift_left, shift_right, ShifterOperand(kArmBitsPerWord), AL, kCcKeep); 3621 __ b(&shift_by_32_plus_shift_right, CC); 3622 3623 // out_reg_hi = (reg_hi << shift_left) | (reg_lo >> shift_right). 3624 // out_reg_lo = (reg_lo << shift_left) | (reg_hi >> shift_right). 3625 __ Lsl(out_reg_hi, in_reg_hi, shift_left); 3626 __ Lsr(out_reg_lo, in_reg_lo, shift_right); 3627 __ add(out_reg_hi, out_reg_hi, ShifterOperand(out_reg_lo)); 3628 __ Lsl(out_reg_lo, in_reg_lo, shift_left); 3629 __ Lsr(shift_left, in_reg_hi, shift_right); 3630 __ add(out_reg_lo, out_reg_lo, ShifterOperand(shift_left)); 3631 __ b(&end); 3632 3633 __ Bind(&shift_by_32_plus_shift_right); // Shift by 32+shift_right. 3634 // out_reg_hi = (reg_hi >> shift_right) | (reg_lo << shift_left). 3635 // out_reg_lo = (reg_lo >> shift_right) | (reg_hi << shift_left). 3636 __ Lsr(out_reg_hi, in_reg_hi, shift_right); 3637 __ Lsl(out_reg_lo, in_reg_lo, shift_left); 3638 __ add(out_reg_hi, out_reg_hi, ShifterOperand(out_reg_lo)); 3639 __ Lsr(out_reg_lo, in_reg_lo, shift_right); 3640 __ Lsl(shift_right, in_reg_hi, shift_left); 3641 __ add(out_reg_lo, out_reg_lo, ShifterOperand(shift_right)); 3642 3643 __ Bind(&end); 3644 } 3645} 3646 3647void LocationsBuilderARM::VisitRor(HRor* ror) { 3648 LocationSummary* locations = 3649 new (GetGraph()->GetArena()) LocationSummary(ror, LocationSummary::kNoCall); 3650 switch (ror->GetResultType()) { 3651 case Primitive::kPrimInt: { 3652 locations->SetInAt(0, Location::RequiresRegister()); 3653 locations->SetInAt(1, Location::RegisterOrConstant(ror->InputAt(1))); 3654 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3655 break; 3656 } 3657 case Primitive::kPrimLong: { 3658 locations->SetInAt(0, Location::RequiresRegister()); 3659 if (ror->InputAt(1)->IsConstant()) { 3660 locations->SetInAt(1, Location::ConstantLocation(ror->InputAt(1)->AsConstant())); 3661 } else { 3662 locations->SetInAt(1, Location::RequiresRegister()); 3663 locations->AddTemp(Location::RequiresRegister()); 3664 locations->AddTemp(Location::RequiresRegister()); 3665 } 3666 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3667 break; 3668 } 3669 default: 3670 LOG(FATAL) << "Unexpected operation type " << ror->GetResultType(); 3671 } 3672} 3673 3674void InstructionCodeGeneratorARM::VisitRor(HRor* ror) { 3675 LocationSummary* locations = ror->GetLocations(); 3676 Primitive::Type type = ror->GetResultType(); 3677 switch (type) { 3678 case Primitive::kPrimInt: { 3679 HandleIntegerRotate(locations); 3680 break; 3681 } 3682 case Primitive::kPrimLong: { 3683 HandleLongRotate(locations); 3684 break; 3685 } 3686 default: 3687 LOG(FATAL) << "Unexpected operation type " << type; 3688 UNREACHABLE(); 3689 } 3690} 3691 3692void LocationsBuilderARM::HandleShift(HBinaryOperation* op) { 3693 DCHECK(op->IsShl() || op->IsShr() || op->IsUShr()); 3694 3695 LocationSummary* locations = 3696 new (GetGraph()->GetArena()) LocationSummary(op, LocationSummary::kNoCall); 3697 3698 switch (op->GetResultType()) { 3699 case Primitive::kPrimInt: { 3700 locations->SetInAt(0, Location::RequiresRegister()); 3701 if (op->InputAt(1)->IsConstant()) { 3702 locations->SetInAt(1, Location::ConstantLocation(op->InputAt(1)->AsConstant())); 3703 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3704 } else { 3705 locations->SetInAt(1, Location::RequiresRegister()); 3706 // Make the output overlap, as it will be used to hold the masked 3707 // second input. 3708 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3709 } 3710 break; 3711 } 3712 case Primitive::kPrimLong: { 3713 locations->SetInAt(0, Location::RequiresRegister()); 3714 if (op->InputAt(1)->IsConstant()) { 3715 locations->SetInAt(1, Location::ConstantLocation(op->InputAt(1)->AsConstant())); 3716 // For simplicity, use kOutputOverlap even though we only require that low registers 3717 // don't clash with high registers which the register allocator currently guarantees. 3718 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3719 } else { 3720 locations->SetInAt(1, Location::RequiresRegister()); 3721 locations->AddTemp(Location::RequiresRegister()); 3722 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3723 } 3724 break; 3725 } 3726 default: 3727 LOG(FATAL) << "Unexpected operation type " << op->GetResultType(); 3728 } 3729} 3730 3731void InstructionCodeGeneratorARM::HandleShift(HBinaryOperation* op) { 3732 DCHECK(op->IsShl() || op->IsShr() || op->IsUShr()); 3733 3734 LocationSummary* locations = op->GetLocations(); 3735 Location out = locations->Out(); 3736 Location first = locations->InAt(0); 3737 Location second = locations->InAt(1); 3738 3739 Primitive::Type type = op->GetResultType(); 3740 switch (type) { 3741 case Primitive::kPrimInt: { 3742 Register out_reg = out.AsRegister<Register>(); 3743 Register first_reg = first.AsRegister<Register>(); 3744 if (second.IsRegister()) { 3745 Register second_reg = second.AsRegister<Register>(); 3746 // ARM doesn't mask the shift count so we need to do it ourselves. 3747 __ and_(out_reg, second_reg, ShifterOperand(kMaxIntShiftDistance)); 3748 if (op->IsShl()) { 3749 __ Lsl(out_reg, first_reg, out_reg); 3750 } else if (op->IsShr()) { 3751 __ Asr(out_reg, first_reg, out_reg); 3752 } else { 3753 __ Lsr(out_reg, first_reg, out_reg); 3754 } 3755 } else { 3756 int32_t cst = second.GetConstant()->AsIntConstant()->GetValue(); 3757 uint32_t shift_value = cst & kMaxIntShiftDistance; 3758 if (shift_value == 0) { // ARM does not support shifting with 0 immediate. 3759 __ Mov(out_reg, first_reg); 3760 } else if (op->IsShl()) { 3761 __ Lsl(out_reg, first_reg, shift_value); 3762 } else if (op->IsShr()) { 3763 __ Asr(out_reg, first_reg, shift_value); 3764 } else { 3765 __ Lsr(out_reg, first_reg, shift_value); 3766 } 3767 } 3768 break; 3769 } 3770 case Primitive::kPrimLong: { 3771 Register o_h = out.AsRegisterPairHigh<Register>(); 3772 Register o_l = out.AsRegisterPairLow<Register>(); 3773 3774 Register high = first.AsRegisterPairHigh<Register>(); 3775 Register low = first.AsRegisterPairLow<Register>(); 3776 3777 if (second.IsRegister()) { 3778 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3779 3780 Register second_reg = second.AsRegister<Register>(); 3781 3782 if (op->IsShl()) { 3783 __ and_(o_l, second_reg, ShifterOperand(kMaxLongShiftDistance)); 3784 // Shift the high part 3785 __ Lsl(o_h, high, o_l); 3786 // Shift the low part and `or` what overflew on the high part 3787 __ rsb(temp, o_l, ShifterOperand(kArmBitsPerWord)); 3788 __ Lsr(temp, low, temp); 3789 __ orr(o_h, o_h, ShifterOperand(temp)); 3790 // If the shift is > 32 bits, override the high part 3791 __ subs(temp, o_l, ShifterOperand(kArmBitsPerWord)); 3792 __ it(PL); 3793 __ Lsl(o_h, low, temp, PL); 3794 // Shift the low part 3795 __ Lsl(o_l, low, o_l); 3796 } else if (op->IsShr()) { 3797 __ and_(o_h, second_reg, ShifterOperand(kMaxLongShiftDistance)); 3798 // Shift the low part 3799 __ Lsr(o_l, low, o_h); 3800 // Shift the high part and `or` what underflew on the low part 3801 __ rsb(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3802 __ Lsl(temp, high, temp); 3803 __ orr(o_l, o_l, ShifterOperand(temp)); 3804 // If the shift is > 32 bits, override the low part 3805 __ subs(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3806 __ it(PL); 3807 __ Asr(o_l, high, temp, PL); 3808 // Shift the high part 3809 __ Asr(o_h, high, o_h); 3810 } else { 3811 __ and_(o_h, second_reg, ShifterOperand(kMaxLongShiftDistance)); 3812 // same as Shr except we use `Lsr`s and not `Asr`s 3813 __ Lsr(o_l, low, o_h); 3814 __ rsb(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3815 __ Lsl(temp, high, temp); 3816 __ orr(o_l, o_l, ShifterOperand(temp)); 3817 __ subs(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3818 __ it(PL); 3819 __ Lsr(o_l, high, temp, PL); 3820 __ Lsr(o_h, high, o_h); 3821 } 3822 } else { 3823 // Register allocator doesn't create partial overlap. 3824 DCHECK_NE(o_l, high); 3825 DCHECK_NE(o_h, low); 3826 int32_t cst = second.GetConstant()->AsIntConstant()->GetValue(); 3827 uint32_t shift_value = cst & kMaxLongShiftDistance; 3828 if (shift_value > 32) { 3829 if (op->IsShl()) { 3830 __ Lsl(o_h, low, shift_value - 32); 3831 __ LoadImmediate(o_l, 0); 3832 } else if (op->IsShr()) { 3833 __ Asr(o_l, high, shift_value - 32); 3834 __ Asr(o_h, high, 31); 3835 } else { 3836 __ Lsr(o_l, high, shift_value - 32); 3837 __ LoadImmediate(o_h, 0); 3838 } 3839 } else if (shift_value == 32) { 3840 if (op->IsShl()) { 3841 __ mov(o_h, ShifterOperand(low)); 3842 __ LoadImmediate(o_l, 0); 3843 } else if (op->IsShr()) { 3844 __ mov(o_l, ShifterOperand(high)); 3845 __ Asr(o_h, high, 31); 3846 } else { 3847 __ mov(o_l, ShifterOperand(high)); 3848 __ LoadImmediate(o_h, 0); 3849 } 3850 } else if (shift_value == 1) { 3851 if (op->IsShl()) { 3852 __ Lsls(o_l, low, 1); 3853 __ adc(o_h, high, ShifterOperand(high)); 3854 } else if (op->IsShr()) { 3855 __ Asrs(o_h, high, 1); 3856 __ Rrx(o_l, low); 3857 } else { 3858 __ Lsrs(o_h, high, 1); 3859 __ Rrx(o_l, low); 3860 } 3861 } else { 3862 DCHECK(2 <= shift_value && shift_value < 32) << shift_value; 3863 if (op->IsShl()) { 3864 __ Lsl(o_h, high, shift_value); 3865 __ orr(o_h, o_h, ShifterOperand(low, LSR, 32 - shift_value)); 3866 __ Lsl(o_l, low, shift_value); 3867 } else if (op->IsShr()) { 3868 __ Lsr(o_l, low, shift_value); 3869 __ orr(o_l, o_l, ShifterOperand(high, LSL, 32 - shift_value)); 3870 __ Asr(o_h, high, shift_value); 3871 } else { 3872 __ Lsr(o_l, low, shift_value); 3873 __ orr(o_l, o_l, ShifterOperand(high, LSL, 32 - shift_value)); 3874 __ Lsr(o_h, high, shift_value); 3875 } 3876 } 3877 } 3878 break; 3879 } 3880 default: 3881 LOG(FATAL) << "Unexpected operation type " << type; 3882 UNREACHABLE(); 3883 } 3884} 3885 3886void LocationsBuilderARM::VisitShl(HShl* shl) { 3887 HandleShift(shl); 3888} 3889 3890void InstructionCodeGeneratorARM::VisitShl(HShl* shl) { 3891 HandleShift(shl); 3892} 3893 3894void LocationsBuilderARM::VisitShr(HShr* shr) { 3895 HandleShift(shr); 3896} 3897 3898void InstructionCodeGeneratorARM::VisitShr(HShr* shr) { 3899 HandleShift(shr); 3900} 3901 3902void LocationsBuilderARM::VisitUShr(HUShr* ushr) { 3903 HandleShift(ushr); 3904} 3905 3906void InstructionCodeGeneratorARM::VisitUShr(HUShr* ushr) { 3907 HandleShift(ushr); 3908} 3909 3910void LocationsBuilderARM::VisitNewInstance(HNewInstance* instruction) { 3911 LocationSummary* locations = 3912 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnMainOnly); 3913 if (instruction->IsStringAlloc()) { 3914 locations->AddTemp(Location::RegisterLocation(kMethodRegisterArgument)); 3915 } else { 3916 InvokeRuntimeCallingConvention calling_convention; 3917 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 3918 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 3919 } 3920 locations->SetOut(Location::RegisterLocation(R0)); 3921} 3922 3923void InstructionCodeGeneratorARM::VisitNewInstance(HNewInstance* instruction) { 3924 // Note: if heap poisoning is enabled, the entry point takes cares 3925 // of poisoning the reference. 3926 if (instruction->IsStringAlloc()) { 3927 // String is allocated through StringFactory. Call NewEmptyString entry point. 3928 Register temp = instruction->GetLocations()->GetTemp(0).AsRegister<Register>(); 3929 MemberOffset code_offset = ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmPointerSize); 3930 __ LoadFromOffset(kLoadWord, temp, TR, QUICK_ENTRY_POINT(pNewEmptyString)); 3931 __ LoadFromOffset(kLoadWord, LR, temp, code_offset.Int32Value()); 3932 __ blx(LR); 3933 codegen_->RecordPcInfo(instruction, instruction->GetDexPc()); 3934 } else { 3935 codegen_->InvokeRuntime(instruction->GetEntrypoint(), instruction, instruction->GetDexPc()); 3936 CheckEntrypointTypes<kQuickAllocObjectWithAccessCheck, void*, uint32_t, ArtMethod*>(); 3937 } 3938} 3939 3940void LocationsBuilderARM::VisitNewArray(HNewArray* instruction) { 3941 LocationSummary* locations = 3942 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnMainOnly); 3943 InvokeRuntimeCallingConvention calling_convention; 3944 locations->AddTemp(Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 3945 locations->SetOut(Location::RegisterLocation(R0)); 3946 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 3947 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(2))); 3948} 3949 3950void InstructionCodeGeneratorARM::VisitNewArray(HNewArray* instruction) { 3951 InvokeRuntimeCallingConvention calling_convention; 3952 __ LoadImmediate(calling_convention.GetRegisterAt(0), instruction->GetTypeIndex()); 3953 // Note: if heap poisoning is enabled, the entry point takes cares 3954 // of poisoning the reference. 3955 codegen_->InvokeRuntime(instruction->GetEntrypoint(), instruction, instruction->GetDexPc()); 3956 CheckEntrypointTypes<kQuickAllocArrayWithAccessCheck, void*, uint32_t, int32_t, ArtMethod*>(); 3957} 3958 3959void LocationsBuilderARM::VisitParameterValue(HParameterValue* instruction) { 3960 LocationSummary* locations = 3961 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 3962 Location location = parameter_visitor_.GetNextLocation(instruction->GetType()); 3963 if (location.IsStackSlot()) { 3964 location = Location::StackSlot(location.GetStackIndex() + codegen_->GetFrameSize()); 3965 } else if (location.IsDoubleStackSlot()) { 3966 location = Location::DoubleStackSlot(location.GetStackIndex() + codegen_->GetFrameSize()); 3967 } 3968 locations->SetOut(location); 3969} 3970 3971void InstructionCodeGeneratorARM::VisitParameterValue( 3972 HParameterValue* instruction ATTRIBUTE_UNUSED) { 3973 // Nothing to do, the parameter is already at its location. 3974} 3975 3976void LocationsBuilderARM::VisitCurrentMethod(HCurrentMethod* instruction) { 3977 LocationSummary* locations = 3978 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 3979 locations->SetOut(Location::RegisterLocation(kMethodRegisterArgument)); 3980} 3981 3982void InstructionCodeGeneratorARM::VisitCurrentMethod(HCurrentMethod* instruction ATTRIBUTE_UNUSED) { 3983 // Nothing to do, the method is already at its location. 3984} 3985 3986void LocationsBuilderARM::VisitNot(HNot* not_) { 3987 LocationSummary* locations = 3988 new (GetGraph()->GetArena()) LocationSummary(not_, LocationSummary::kNoCall); 3989 locations->SetInAt(0, Location::RequiresRegister()); 3990 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3991} 3992 3993void InstructionCodeGeneratorARM::VisitNot(HNot* not_) { 3994 LocationSummary* locations = not_->GetLocations(); 3995 Location out = locations->Out(); 3996 Location in = locations->InAt(0); 3997 switch (not_->GetResultType()) { 3998 case Primitive::kPrimInt: 3999 __ mvn(out.AsRegister<Register>(), ShifterOperand(in.AsRegister<Register>())); 4000 break; 4001 4002 case Primitive::kPrimLong: 4003 __ mvn(out.AsRegisterPairLow<Register>(), 4004 ShifterOperand(in.AsRegisterPairLow<Register>())); 4005 __ mvn(out.AsRegisterPairHigh<Register>(), 4006 ShifterOperand(in.AsRegisterPairHigh<Register>())); 4007 break; 4008 4009 default: 4010 LOG(FATAL) << "Unimplemented type for not operation " << not_->GetResultType(); 4011 } 4012} 4013 4014void LocationsBuilderARM::VisitBooleanNot(HBooleanNot* bool_not) { 4015 LocationSummary* locations = 4016 new (GetGraph()->GetArena()) LocationSummary(bool_not, LocationSummary::kNoCall); 4017 locations->SetInAt(0, Location::RequiresRegister()); 4018 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 4019} 4020 4021void InstructionCodeGeneratorARM::VisitBooleanNot(HBooleanNot* bool_not) { 4022 LocationSummary* locations = bool_not->GetLocations(); 4023 Location out = locations->Out(); 4024 Location in = locations->InAt(0); 4025 __ eor(out.AsRegister<Register>(), in.AsRegister<Register>(), ShifterOperand(1)); 4026} 4027 4028void LocationsBuilderARM::VisitCompare(HCompare* compare) { 4029 LocationSummary* locations = 4030 new (GetGraph()->GetArena()) LocationSummary(compare, LocationSummary::kNoCall); 4031 switch (compare->InputAt(0)->GetType()) { 4032 case Primitive::kPrimBoolean: 4033 case Primitive::kPrimByte: 4034 case Primitive::kPrimShort: 4035 case Primitive::kPrimChar: 4036 case Primitive::kPrimInt: 4037 case Primitive::kPrimLong: { 4038 locations->SetInAt(0, Location::RequiresRegister()); 4039 locations->SetInAt(1, Location::RequiresRegister()); 4040 // Output overlaps because it is written before doing the low comparison. 4041 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 4042 break; 4043 } 4044 case Primitive::kPrimFloat: 4045 case Primitive::kPrimDouble: { 4046 locations->SetInAt(0, Location::RequiresFpuRegister()); 4047 locations->SetInAt(1, ArithmeticZeroOrFpuRegister(compare->InputAt(1))); 4048 locations->SetOut(Location::RequiresRegister()); 4049 break; 4050 } 4051 default: 4052 LOG(FATAL) << "Unexpected type for compare operation " << compare->InputAt(0)->GetType(); 4053 } 4054} 4055 4056void InstructionCodeGeneratorARM::VisitCompare(HCompare* compare) { 4057 LocationSummary* locations = compare->GetLocations(); 4058 Register out = locations->Out().AsRegister<Register>(); 4059 Location left = locations->InAt(0); 4060 Location right = locations->InAt(1); 4061 4062 Label less, greater, done; 4063 Primitive::Type type = compare->InputAt(0)->GetType(); 4064 Condition less_cond; 4065 switch (type) { 4066 case Primitive::kPrimBoolean: 4067 case Primitive::kPrimByte: 4068 case Primitive::kPrimShort: 4069 case Primitive::kPrimChar: 4070 case Primitive::kPrimInt: { 4071 __ LoadImmediate(out, 0); 4072 __ cmp(left.AsRegister<Register>(), 4073 ShifterOperand(right.AsRegister<Register>())); // Signed compare. 4074 less_cond = LT; 4075 break; 4076 } 4077 case Primitive::kPrimLong: { 4078 __ cmp(left.AsRegisterPairHigh<Register>(), 4079 ShifterOperand(right.AsRegisterPairHigh<Register>())); // Signed compare. 4080 __ b(&less, LT); 4081 __ b(&greater, GT); 4082 // Do LoadImmediate before the last `cmp`, as LoadImmediate might affect the status flags. 4083 __ LoadImmediate(out, 0); 4084 __ cmp(left.AsRegisterPairLow<Register>(), 4085 ShifterOperand(right.AsRegisterPairLow<Register>())); // Unsigned compare. 4086 less_cond = LO; 4087 break; 4088 } 4089 case Primitive::kPrimFloat: 4090 case Primitive::kPrimDouble: { 4091 __ LoadImmediate(out, 0); 4092 GenerateVcmp(compare); 4093 __ vmstat(); // transfer FP status register to ARM APSR. 4094 less_cond = ARMFPCondition(kCondLT, compare->IsGtBias()); 4095 break; 4096 } 4097 default: 4098 LOG(FATAL) << "Unexpected compare type " << type; 4099 UNREACHABLE(); 4100 } 4101 4102 __ b(&done, EQ); 4103 __ b(&less, less_cond); 4104 4105 __ Bind(&greater); 4106 __ LoadImmediate(out, 1); 4107 __ b(&done); 4108 4109 __ Bind(&less); 4110 __ LoadImmediate(out, -1); 4111 4112 __ Bind(&done); 4113} 4114 4115void LocationsBuilderARM::VisitPhi(HPhi* instruction) { 4116 LocationSummary* locations = 4117 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 4118 for (size_t i = 0, e = locations->GetInputCount(); i < e; ++i) { 4119 locations->SetInAt(i, Location::Any()); 4120 } 4121 locations->SetOut(Location::Any()); 4122} 4123 4124void InstructionCodeGeneratorARM::VisitPhi(HPhi* instruction ATTRIBUTE_UNUSED) { 4125 LOG(FATAL) << "Unreachable"; 4126} 4127 4128void CodeGeneratorARM::GenerateMemoryBarrier(MemBarrierKind kind) { 4129 // TODO (ported from quick): revisit ARM barrier kinds. 4130 DmbOptions flavor = DmbOptions::ISH; // Quiet C++ warnings. 4131 switch (kind) { 4132 case MemBarrierKind::kAnyStore: 4133 case MemBarrierKind::kLoadAny: 4134 case MemBarrierKind::kAnyAny: { 4135 flavor = DmbOptions::ISH; 4136 break; 4137 } 4138 case MemBarrierKind::kStoreStore: { 4139 flavor = DmbOptions::ISHST; 4140 break; 4141 } 4142 default: 4143 LOG(FATAL) << "Unexpected memory barrier " << kind; 4144 } 4145 __ dmb(flavor); 4146} 4147 4148void InstructionCodeGeneratorARM::GenerateWideAtomicLoad(Register addr, 4149 uint32_t offset, 4150 Register out_lo, 4151 Register out_hi) { 4152 if (offset != 0) { 4153 // Ensure `out_lo` is different from `addr`, so that loading 4154 // `offset` into `out_lo` does not clutter `addr`. 4155 DCHECK_NE(out_lo, addr); 4156 __ LoadImmediate(out_lo, offset); 4157 __ add(IP, addr, ShifterOperand(out_lo)); 4158 addr = IP; 4159 } 4160 __ ldrexd(out_lo, out_hi, addr); 4161} 4162 4163void InstructionCodeGeneratorARM::GenerateWideAtomicStore(Register addr, 4164 uint32_t offset, 4165 Register value_lo, 4166 Register value_hi, 4167 Register temp1, 4168 Register temp2, 4169 HInstruction* instruction) { 4170 Label fail; 4171 if (offset != 0) { 4172 __ LoadImmediate(temp1, offset); 4173 __ add(IP, addr, ShifterOperand(temp1)); 4174 addr = IP; 4175 } 4176 __ Bind(&fail); 4177 // We need a load followed by store. (The address used in a STREX instruction must 4178 // be the same as the address in the most recently executed LDREX instruction.) 4179 __ ldrexd(temp1, temp2, addr); 4180 codegen_->MaybeRecordImplicitNullCheck(instruction); 4181 __ strexd(temp1, value_lo, value_hi, addr); 4182 __ CompareAndBranchIfNonZero(temp1, &fail); 4183} 4184 4185void LocationsBuilderARM::HandleFieldSet(HInstruction* instruction, const FieldInfo& field_info) { 4186 DCHECK(instruction->IsInstanceFieldSet() || instruction->IsStaticFieldSet()); 4187 4188 LocationSummary* locations = 4189 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 4190 locations->SetInAt(0, Location::RequiresRegister()); 4191 4192 Primitive::Type field_type = field_info.GetFieldType(); 4193 if (Primitive::IsFloatingPointType(field_type)) { 4194 locations->SetInAt(1, Location::RequiresFpuRegister()); 4195 } else { 4196 locations->SetInAt(1, Location::RequiresRegister()); 4197 } 4198 4199 bool is_wide = field_type == Primitive::kPrimLong || field_type == Primitive::kPrimDouble; 4200 bool generate_volatile = field_info.IsVolatile() 4201 && is_wide 4202 && !codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 4203 bool needs_write_barrier = 4204 CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1)); 4205 // Temporary registers for the write barrier. 4206 // TODO: consider renaming StoreNeedsWriteBarrier to StoreNeedsGCMark. 4207 if (needs_write_barrier) { 4208 locations->AddTemp(Location::RequiresRegister()); // Possibly used for reference poisoning too. 4209 locations->AddTemp(Location::RequiresRegister()); 4210 } else if (generate_volatile) { 4211 // ARM encoding have some additional constraints for ldrexd/strexd: 4212 // - registers need to be consecutive 4213 // - the first register should be even but not R14. 4214 // We don't test for ARM yet, and the assertion makes sure that we 4215 // revisit this if we ever enable ARM encoding. 4216 DCHECK_EQ(InstructionSet::kThumb2, codegen_->GetInstructionSet()); 4217 4218 locations->AddTemp(Location::RequiresRegister()); 4219 locations->AddTemp(Location::RequiresRegister()); 4220 if (field_type == Primitive::kPrimDouble) { 4221 // For doubles we need two more registers to copy the value. 4222 locations->AddTemp(Location::RegisterLocation(R2)); 4223 locations->AddTemp(Location::RegisterLocation(R3)); 4224 } 4225 } 4226} 4227 4228void InstructionCodeGeneratorARM::HandleFieldSet(HInstruction* instruction, 4229 const FieldInfo& field_info, 4230 bool value_can_be_null) { 4231 DCHECK(instruction->IsInstanceFieldSet() || instruction->IsStaticFieldSet()); 4232 4233 LocationSummary* locations = instruction->GetLocations(); 4234 Register base = locations->InAt(0).AsRegister<Register>(); 4235 Location value = locations->InAt(1); 4236 4237 bool is_volatile = field_info.IsVolatile(); 4238 bool atomic_ldrd_strd = codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 4239 Primitive::Type field_type = field_info.GetFieldType(); 4240 uint32_t offset = field_info.GetFieldOffset().Uint32Value(); 4241 bool needs_write_barrier = 4242 CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1)); 4243 4244 if (is_volatile) { 4245 codegen_->GenerateMemoryBarrier(MemBarrierKind::kAnyStore); 4246 } 4247 4248 switch (field_type) { 4249 case Primitive::kPrimBoolean: 4250 case Primitive::kPrimByte: { 4251 __ StoreToOffset(kStoreByte, value.AsRegister<Register>(), base, offset); 4252 break; 4253 } 4254 4255 case Primitive::kPrimShort: 4256 case Primitive::kPrimChar: { 4257 __ StoreToOffset(kStoreHalfword, value.AsRegister<Register>(), base, offset); 4258 break; 4259 } 4260 4261 case Primitive::kPrimInt: 4262 case Primitive::kPrimNot: { 4263 if (kPoisonHeapReferences && needs_write_barrier) { 4264 // Note that in the case where `value` is a null reference, 4265 // we do not enter this block, as a null reference does not 4266 // need poisoning. 4267 DCHECK_EQ(field_type, Primitive::kPrimNot); 4268 Register temp = locations->GetTemp(0).AsRegister<Register>(); 4269 __ Mov(temp, value.AsRegister<Register>()); 4270 __ PoisonHeapReference(temp); 4271 __ StoreToOffset(kStoreWord, temp, base, offset); 4272 } else { 4273 __ StoreToOffset(kStoreWord, value.AsRegister<Register>(), base, offset); 4274 } 4275 break; 4276 } 4277 4278 case Primitive::kPrimLong: { 4279 if (is_volatile && !atomic_ldrd_strd) { 4280 GenerateWideAtomicStore(base, offset, 4281 value.AsRegisterPairLow<Register>(), 4282 value.AsRegisterPairHigh<Register>(), 4283 locations->GetTemp(0).AsRegister<Register>(), 4284 locations->GetTemp(1).AsRegister<Register>(), 4285 instruction); 4286 } else { 4287 __ StoreToOffset(kStoreWordPair, value.AsRegisterPairLow<Register>(), base, offset); 4288 codegen_->MaybeRecordImplicitNullCheck(instruction); 4289 } 4290 break; 4291 } 4292 4293 case Primitive::kPrimFloat: { 4294 __ StoreSToOffset(value.AsFpuRegister<SRegister>(), base, offset); 4295 break; 4296 } 4297 4298 case Primitive::kPrimDouble: { 4299 DRegister value_reg = FromLowSToD(value.AsFpuRegisterPairLow<SRegister>()); 4300 if (is_volatile && !atomic_ldrd_strd) { 4301 Register value_reg_lo = locations->GetTemp(0).AsRegister<Register>(); 4302 Register value_reg_hi = locations->GetTemp(1).AsRegister<Register>(); 4303 4304 __ vmovrrd(value_reg_lo, value_reg_hi, value_reg); 4305 4306 GenerateWideAtomicStore(base, offset, 4307 value_reg_lo, 4308 value_reg_hi, 4309 locations->GetTemp(2).AsRegister<Register>(), 4310 locations->GetTemp(3).AsRegister<Register>(), 4311 instruction); 4312 } else { 4313 __ StoreDToOffset(value_reg, base, offset); 4314 codegen_->MaybeRecordImplicitNullCheck(instruction); 4315 } 4316 break; 4317 } 4318 4319 case Primitive::kPrimVoid: 4320 LOG(FATAL) << "Unreachable type " << field_type; 4321 UNREACHABLE(); 4322 } 4323 4324 // Longs and doubles are handled in the switch. 4325 if (field_type != Primitive::kPrimLong && field_type != Primitive::kPrimDouble) { 4326 codegen_->MaybeRecordImplicitNullCheck(instruction); 4327 } 4328 4329 if (CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1))) { 4330 Register temp = locations->GetTemp(0).AsRegister<Register>(); 4331 Register card = locations->GetTemp(1).AsRegister<Register>(); 4332 codegen_->MarkGCCard( 4333 temp, card, base, value.AsRegister<Register>(), value_can_be_null); 4334 } 4335 4336 if (is_volatile) { 4337 codegen_->GenerateMemoryBarrier(MemBarrierKind::kAnyAny); 4338 } 4339} 4340 4341void LocationsBuilderARM::HandleFieldGet(HInstruction* instruction, const FieldInfo& field_info) { 4342 DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet()); 4343 4344 bool object_field_get_with_read_barrier = 4345 kEmitCompilerReadBarrier && (field_info.GetFieldType() == Primitive::kPrimNot); 4346 LocationSummary* locations = 4347 new (GetGraph()->GetArena()) LocationSummary(instruction, 4348 object_field_get_with_read_barrier ? 4349 LocationSummary::kCallOnSlowPath : 4350 LocationSummary::kNoCall); 4351 if (object_field_get_with_read_barrier && kUseBakerReadBarrier) { 4352 locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty()); // No caller-save registers. 4353 } 4354 locations->SetInAt(0, Location::RequiresRegister()); 4355 4356 bool volatile_for_double = field_info.IsVolatile() 4357 && (field_info.GetFieldType() == Primitive::kPrimDouble) 4358 && !codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 4359 // The output overlaps in case of volatile long: we don't want the 4360 // code generated by GenerateWideAtomicLoad to overwrite the 4361 // object's location. Likewise, in the case of an object field get 4362 // with read barriers enabled, we do not want the load to overwrite 4363 // the object's location, as we need it to emit the read barrier. 4364 bool overlap = (field_info.IsVolatile() && (field_info.GetFieldType() == Primitive::kPrimLong)) || 4365 object_field_get_with_read_barrier; 4366 4367 if (Primitive::IsFloatingPointType(instruction->GetType())) { 4368 locations->SetOut(Location::RequiresFpuRegister()); 4369 } else { 4370 locations->SetOut(Location::RequiresRegister(), 4371 (overlap ? Location::kOutputOverlap : Location::kNoOutputOverlap)); 4372 } 4373 if (volatile_for_double) { 4374 // ARM encoding have some additional constraints for ldrexd/strexd: 4375 // - registers need to be consecutive 4376 // - the first register should be even but not R14. 4377 // We don't test for ARM yet, and the assertion makes sure that we 4378 // revisit this if we ever enable ARM encoding. 4379 DCHECK_EQ(InstructionSet::kThumb2, codegen_->GetInstructionSet()); 4380 locations->AddTemp(Location::RequiresRegister()); 4381 locations->AddTemp(Location::RequiresRegister()); 4382 } else if (object_field_get_with_read_barrier && kUseBakerReadBarrier) { 4383 // We need a temporary register for the read barrier marking slow 4384 // path in CodeGeneratorARM::GenerateFieldLoadWithBakerReadBarrier. 4385 locations->AddTemp(Location::RequiresRegister()); 4386 } 4387} 4388 4389Location LocationsBuilderARM::ArithmeticZeroOrFpuRegister(HInstruction* input) { 4390 DCHECK(input->GetType() == Primitive::kPrimDouble || input->GetType() == Primitive::kPrimFloat) 4391 << input->GetType(); 4392 if ((input->IsFloatConstant() && (input->AsFloatConstant()->IsArithmeticZero())) || 4393 (input->IsDoubleConstant() && (input->AsDoubleConstant()->IsArithmeticZero()))) { 4394 return Location::ConstantLocation(input->AsConstant()); 4395 } else { 4396 return Location::RequiresFpuRegister(); 4397 } 4398} 4399 4400Location LocationsBuilderARM::ArmEncodableConstantOrRegister(HInstruction* constant, 4401 Opcode opcode) { 4402 DCHECK(!Primitive::IsFloatingPointType(constant->GetType())); 4403 if (constant->IsConstant() && 4404 CanEncodeConstantAsImmediate(constant->AsConstant(), opcode)) { 4405 return Location::ConstantLocation(constant->AsConstant()); 4406 } 4407 return Location::RequiresRegister(); 4408} 4409 4410bool LocationsBuilderARM::CanEncodeConstantAsImmediate(HConstant* input_cst, 4411 Opcode opcode) { 4412 uint64_t value = static_cast<uint64_t>(Int64FromConstant(input_cst)); 4413 if (Primitive::Is64BitType(input_cst->GetType())) { 4414 Opcode high_opcode = opcode; 4415 SetCc low_set_cc = kCcDontCare; 4416 switch (opcode) { 4417 case SUB: 4418 // Flip the operation to an ADD. 4419 value = -value; 4420 opcode = ADD; 4421 FALLTHROUGH_INTENDED; 4422 case ADD: 4423 if (Low32Bits(value) == 0u) { 4424 return CanEncodeConstantAsImmediate(High32Bits(value), opcode, kCcDontCare); 4425 } 4426 high_opcode = ADC; 4427 low_set_cc = kCcSet; 4428 break; 4429 default: 4430 break; 4431 } 4432 return CanEncodeConstantAsImmediate(Low32Bits(value), opcode, low_set_cc) && 4433 CanEncodeConstantAsImmediate(High32Bits(value), high_opcode, kCcDontCare); 4434 } else { 4435 return CanEncodeConstantAsImmediate(Low32Bits(value), opcode); 4436 } 4437} 4438 4439bool LocationsBuilderARM::CanEncodeConstantAsImmediate(uint32_t value, 4440 Opcode opcode, 4441 SetCc set_cc) { 4442 ShifterOperand so; 4443 ArmAssembler* assembler = codegen_->GetAssembler(); 4444 if (assembler->ShifterOperandCanHold(kNoRegister, kNoRegister, opcode, value, set_cc, &so)) { 4445 return true; 4446 } 4447 Opcode neg_opcode = kNoOperand; 4448 switch (opcode) { 4449 case AND: neg_opcode = BIC; value = ~value; break; 4450 case ORR: neg_opcode = ORN; value = ~value; break; 4451 case ADD: neg_opcode = SUB; value = -value; break; 4452 case ADC: neg_opcode = SBC; value = ~value; break; 4453 case SUB: neg_opcode = ADD; value = -value; break; 4454 case SBC: neg_opcode = ADC; value = ~value; break; 4455 default: 4456 return false; 4457 } 4458 return assembler->ShifterOperandCanHold(kNoRegister, kNoRegister, neg_opcode, value, set_cc, &so); 4459} 4460 4461void InstructionCodeGeneratorARM::HandleFieldGet(HInstruction* instruction, 4462 const FieldInfo& field_info) { 4463 DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet()); 4464 4465 LocationSummary* locations = instruction->GetLocations(); 4466 Location base_loc = locations->InAt(0); 4467 Register base = base_loc.AsRegister<Register>(); 4468 Location out = locations->Out(); 4469 bool is_volatile = field_info.IsVolatile(); 4470 bool atomic_ldrd_strd = codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 4471 Primitive::Type field_type = field_info.GetFieldType(); 4472 uint32_t offset = field_info.GetFieldOffset().Uint32Value(); 4473 4474 switch (field_type) { 4475 case Primitive::kPrimBoolean: 4476 __ LoadFromOffset(kLoadUnsignedByte, out.AsRegister<Register>(), base, offset); 4477 break; 4478 4479 case Primitive::kPrimByte: 4480 __ LoadFromOffset(kLoadSignedByte, out.AsRegister<Register>(), base, offset); 4481 break; 4482 4483 case Primitive::kPrimShort: 4484 __ LoadFromOffset(kLoadSignedHalfword, out.AsRegister<Register>(), base, offset); 4485 break; 4486 4487 case Primitive::kPrimChar: 4488 __ LoadFromOffset(kLoadUnsignedHalfword, out.AsRegister<Register>(), base, offset); 4489 break; 4490 4491 case Primitive::kPrimInt: 4492 __ LoadFromOffset(kLoadWord, out.AsRegister<Register>(), base, offset); 4493 break; 4494 4495 case Primitive::kPrimNot: { 4496 // /* HeapReference<Object> */ out = *(base + offset) 4497 if (kEmitCompilerReadBarrier && kUseBakerReadBarrier) { 4498 Location temp_loc = locations->GetTemp(0); 4499 // Note that a potential implicit null check is handled in this 4500 // CodeGeneratorARM::GenerateFieldLoadWithBakerReadBarrier call. 4501 codegen_->GenerateFieldLoadWithBakerReadBarrier( 4502 instruction, out, base, offset, temp_loc, /* needs_null_check */ true); 4503 if (is_volatile) { 4504 codegen_->GenerateMemoryBarrier(MemBarrierKind::kLoadAny); 4505 } 4506 } else { 4507 __ LoadFromOffset(kLoadWord, out.AsRegister<Register>(), base, offset); 4508 codegen_->MaybeRecordImplicitNullCheck(instruction); 4509 if (is_volatile) { 4510 codegen_->GenerateMemoryBarrier(MemBarrierKind::kLoadAny); 4511 } 4512 // If read barriers are enabled, emit read barriers other than 4513 // Baker's using a slow path (and also unpoison the loaded 4514 // reference, if heap poisoning is enabled). 4515 codegen_->MaybeGenerateReadBarrierSlow(instruction, out, out, base_loc, offset); 4516 } 4517 break; 4518 } 4519 4520 case Primitive::kPrimLong: 4521 if (is_volatile && !atomic_ldrd_strd) { 4522 GenerateWideAtomicLoad(base, offset, 4523 out.AsRegisterPairLow<Register>(), 4524 out.AsRegisterPairHigh<Register>()); 4525 } else { 4526 __ LoadFromOffset(kLoadWordPair, out.AsRegisterPairLow<Register>(), base, offset); 4527 } 4528 break; 4529 4530 case Primitive::kPrimFloat: 4531 __ LoadSFromOffset(out.AsFpuRegister<SRegister>(), base, offset); 4532 break; 4533 4534 case Primitive::kPrimDouble: { 4535 DRegister out_reg = FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()); 4536 if (is_volatile && !atomic_ldrd_strd) { 4537 Register lo = locations->GetTemp(0).AsRegister<Register>(); 4538 Register hi = locations->GetTemp(1).AsRegister<Register>(); 4539 GenerateWideAtomicLoad(base, offset, lo, hi); 4540 codegen_->MaybeRecordImplicitNullCheck(instruction); 4541 __ vmovdrr(out_reg, lo, hi); 4542 } else { 4543 __ LoadDFromOffset(out_reg, base, offset); 4544 codegen_->MaybeRecordImplicitNullCheck(instruction); 4545 } 4546 break; 4547 } 4548 4549 case Primitive::kPrimVoid: 4550 LOG(FATAL) << "Unreachable type " << field_type; 4551 UNREACHABLE(); 4552 } 4553 4554 if (field_type == Primitive::kPrimNot || field_type == Primitive::kPrimDouble) { 4555 // Potential implicit null checks, in the case of reference or 4556 // double fields, are handled in the previous switch statement. 4557 } else { 4558 codegen_->MaybeRecordImplicitNullCheck(instruction); 4559 } 4560 4561 if (is_volatile) { 4562 if (field_type == Primitive::kPrimNot) { 4563 // Memory barriers, in the case of references, are also handled 4564 // in the previous switch statement. 4565 } else { 4566 codegen_->GenerateMemoryBarrier(MemBarrierKind::kLoadAny); 4567 } 4568 } 4569} 4570 4571void LocationsBuilderARM::VisitInstanceFieldSet(HInstanceFieldSet* instruction) { 4572 HandleFieldSet(instruction, instruction->GetFieldInfo()); 4573} 4574 4575void InstructionCodeGeneratorARM::VisitInstanceFieldSet(HInstanceFieldSet* instruction) { 4576 HandleFieldSet(instruction, instruction->GetFieldInfo(), instruction->GetValueCanBeNull()); 4577} 4578 4579void LocationsBuilderARM::VisitInstanceFieldGet(HInstanceFieldGet* instruction) { 4580 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4581} 4582 4583void InstructionCodeGeneratorARM::VisitInstanceFieldGet(HInstanceFieldGet* instruction) { 4584 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4585} 4586 4587void LocationsBuilderARM::VisitStaticFieldGet(HStaticFieldGet* instruction) { 4588 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4589} 4590 4591void InstructionCodeGeneratorARM::VisitStaticFieldGet(HStaticFieldGet* instruction) { 4592 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4593} 4594 4595void LocationsBuilderARM::VisitStaticFieldSet(HStaticFieldSet* instruction) { 4596 HandleFieldSet(instruction, instruction->GetFieldInfo()); 4597} 4598 4599void InstructionCodeGeneratorARM::VisitStaticFieldSet(HStaticFieldSet* instruction) { 4600 HandleFieldSet(instruction, instruction->GetFieldInfo(), instruction->GetValueCanBeNull()); 4601} 4602 4603void LocationsBuilderARM::VisitUnresolvedInstanceFieldGet( 4604 HUnresolvedInstanceFieldGet* instruction) { 4605 FieldAccessCallingConventionARM calling_convention; 4606 codegen_->CreateUnresolvedFieldLocationSummary( 4607 instruction, instruction->GetFieldType(), calling_convention); 4608} 4609 4610void InstructionCodeGeneratorARM::VisitUnresolvedInstanceFieldGet( 4611 HUnresolvedInstanceFieldGet* instruction) { 4612 FieldAccessCallingConventionARM calling_convention; 4613 codegen_->GenerateUnresolvedFieldAccess(instruction, 4614 instruction->GetFieldType(), 4615 instruction->GetFieldIndex(), 4616 instruction->GetDexPc(), 4617 calling_convention); 4618} 4619 4620void LocationsBuilderARM::VisitUnresolvedInstanceFieldSet( 4621 HUnresolvedInstanceFieldSet* instruction) { 4622 FieldAccessCallingConventionARM calling_convention; 4623 codegen_->CreateUnresolvedFieldLocationSummary( 4624 instruction, instruction->GetFieldType(), calling_convention); 4625} 4626 4627void InstructionCodeGeneratorARM::VisitUnresolvedInstanceFieldSet( 4628 HUnresolvedInstanceFieldSet* instruction) { 4629 FieldAccessCallingConventionARM calling_convention; 4630 codegen_->GenerateUnresolvedFieldAccess(instruction, 4631 instruction->GetFieldType(), 4632 instruction->GetFieldIndex(), 4633 instruction->GetDexPc(), 4634 calling_convention); 4635} 4636 4637void LocationsBuilderARM::VisitUnresolvedStaticFieldGet( 4638 HUnresolvedStaticFieldGet* instruction) { 4639 FieldAccessCallingConventionARM calling_convention; 4640 codegen_->CreateUnresolvedFieldLocationSummary( 4641 instruction, instruction->GetFieldType(), calling_convention); 4642} 4643 4644void InstructionCodeGeneratorARM::VisitUnresolvedStaticFieldGet( 4645 HUnresolvedStaticFieldGet* instruction) { 4646 FieldAccessCallingConventionARM calling_convention; 4647 codegen_->GenerateUnresolvedFieldAccess(instruction, 4648 instruction->GetFieldType(), 4649 instruction->GetFieldIndex(), 4650 instruction->GetDexPc(), 4651 calling_convention); 4652} 4653 4654void LocationsBuilderARM::VisitUnresolvedStaticFieldSet( 4655 HUnresolvedStaticFieldSet* instruction) { 4656 FieldAccessCallingConventionARM calling_convention; 4657 codegen_->CreateUnresolvedFieldLocationSummary( 4658 instruction, instruction->GetFieldType(), calling_convention); 4659} 4660 4661void InstructionCodeGeneratorARM::VisitUnresolvedStaticFieldSet( 4662 HUnresolvedStaticFieldSet* instruction) { 4663 FieldAccessCallingConventionARM calling_convention; 4664 codegen_->GenerateUnresolvedFieldAccess(instruction, 4665 instruction->GetFieldType(), 4666 instruction->GetFieldIndex(), 4667 instruction->GetDexPc(), 4668 calling_convention); 4669} 4670 4671void LocationsBuilderARM::VisitNullCheck(HNullCheck* instruction) { 4672 LocationSummary* locations = codegen_->CreateThrowingSlowPathLocations(instruction); 4673 locations->SetInAt(0, Location::RequiresRegister()); 4674} 4675 4676void CodeGeneratorARM::GenerateImplicitNullCheck(HNullCheck* instruction) { 4677 if (CanMoveNullCheckToUser(instruction)) { 4678 return; 4679 } 4680 Location obj = instruction->GetLocations()->InAt(0); 4681 4682 __ LoadFromOffset(kLoadWord, IP, obj.AsRegister<Register>(), 0); 4683 RecordPcInfo(instruction, instruction->GetDexPc()); 4684} 4685 4686void CodeGeneratorARM::GenerateExplicitNullCheck(HNullCheck* instruction) { 4687 SlowPathCodeARM* slow_path = new (GetGraph()->GetArena()) NullCheckSlowPathARM(instruction); 4688 AddSlowPath(slow_path); 4689 4690 LocationSummary* locations = instruction->GetLocations(); 4691 Location obj = locations->InAt(0); 4692 4693 __ CompareAndBranchIfZero(obj.AsRegister<Register>(), slow_path->GetEntryLabel()); 4694} 4695 4696void InstructionCodeGeneratorARM::VisitNullCheck(HNullCheck* instruction) { 4697 codegen_->GenerateNullCheck(instruction); 4698} 4699 4700static LoadOperandType GetLoadOperandType(Primitive::Type type) { 4701 switch (type) { 4702 case Primitive::kPrimNot: 4703 return kLoadWord; 4704 case Primitive::kPrimBoolean: 4705 return kLoadUnsignedByte; 4706 case Primitive::kPrimByte: 4707 return kLoadSignedByte; 4708 case Primitive::kPrimChar: 4709 return kLoadUnsignedHalfword; 4710 case Primitive::kPrimShort: 4711 return kLoadSignedHalfword; 4712 case Primitive::kPrimInt: 4713 return kLoadWord; 4714 case Primitive::kPrimLong: 4715 return kLoadWordPair; 4716 case Primitive::kPrimFloat: 4717 return kLoadSWord; 4718 case Primitive::kPrimDouble: 4719 return kLoadDWord; 4720 default: 4721 LOG(FATAL) << "Unreachable type " << type; 4722 UNREACHABLE(); 4723 } 4724} 4725 4726static StoreOperandType GetStoreOperandType(Primitive::Type type) { 4727 switch (type) { 4728 case Primitive::kPrimNot: 4729 return kStoreWord; 4730 case Primitive::kPrimBoolean: 4731 case Primitive::kPrimByte: 4732 return kStoreByte; 4733 case Primitive::kPrimChar: 4734 case Primitive::kPrimShort: 4735 return kStoreHalfword; 4736 case Primitive::kPrimInt: 4737 return kStoreWord; 4738 case Primitive::kPrimLong: 4739 return kStoreWordPair; 4740 case Primitive::kPrimFloat: 4741 return kStoreSWord; 4742 case Primitive::kPrimDouble: 4743 return kStoreDWord; 4744 default: 4745 LOG(FATAL) << "Unreachable type " << type; 4746 UNREACHABLE(); 4747 } 4748} 4749 4750void CodeGeneratorARM::LoadFromShiftedRegOffset(Primitive::Type type, 4751 Location out_loc, 4752 Register base, 4753 Register reg_offset, 4754 Condition cond) { 4755 uint32_t shift_count = Primitive::ComponentSizeShift(type); 4756 Address mem_address(base, reg_offset, Shift::LSL, shift_count); 4757 4758 switch (type) { 4759 case Primitive::kPrimByte: 4760 __ ldrsb(out_loc.AsRegister<Register>(), mem_address, cond); 4761 break; 4762 case Primitive::kPrimBoolean: 4763 __ ldrb(out_loc.AsRegister<Register>(), mem_address, cond); 4764 break; 4765 case Primitive::kPrimShort: 4766 __ ldrsh(out_loc.AsRegister<Register>(), mem_address, cond); 4767 break; 4768 case Primitive::kPrimChar: 4769 __ ldrh(out_loc.AsRegister<Register>(), mem_address, cond); 4770 break; 4771 case Primitive::kPrimNot: 4772 case Primitive::kPrimInt: 4773 __ ldr(out_loc.AsRegister<Register>(), mem_address, cond); 4774 break; 4775 // T32 doesn't support LoadFromShiftedRegOffset mem address mode for these types. 4776 case Primitive::kPrimLong: 4777 case Primitive::kPrimFloat: 4778 case Primitive::kPrimDouble: 4779 default: 4780 LOG(FATAL) << "Unreachable type " << type; 4781 UNREACHABLE(); 4782 } 4783} 4784 4785void CodeGeneratorARM::StoreToShiftedRegOffset(Primitive::Type type, 4786 Location loc, 4787 Register base, 4788 Register reg_offset, 4789 Condition cond) { 4790 uint32_t shift_count = Primitive::ComponentSizeShift(type); 4791 Address mem_address(base, reg_offset, Shift::LSL, shift_count); 4792 4793 switch (type) { 4794 case Primitive::kPrimByte: 4795 case Primitive::kPrimBoolean: 4796 __ strb(loc.AsRegister<Register>(), mem_address, cond); 4797 break; 4798 case Primitive::kPrimShort: 4799 case Primitive::kPrimChar: 4800 __ strh(loc.AsRegister<Register>(), mem_address, cond); 4801 break; 4802 case Primitive::kPrimNot: 4803 case Primitive::kPrimInt: 4804 __ str(loc.AsRegister<Register>(), mem_address, cond); 4805 break; 4806 // T32 doesn't support StoreToShiftedRegOffset mem address mode for these types. 4807 case Primitive::kPrimLong: 4808 case Primitive::kPrimFloat: 4809 case Primitive::kPrimDouble: 4810 default: 4811 LOG(FATAL) << "Unreachable type " << type; 4812 UNREACHABLE(); 4813 } 4814} 4815 4816void LocationsBuilderARM::VisitArrayGet(HArrayGet* instruction) { 4817 bool object_array_get_with_read_barrier = 4818 kEmitCompilerReadBarrier && (instruction->GetType() == Primitive::kPrimNot); 4819 LocationSummary* locations = 4820 new (GetGraph()->GetArena()) LocationSummary(instruction, 4821 object_array_get_with_read_barrier ? 4822 LocationSummary::kCallOnSlowPath : 4823 LocationSummary::kNoCall); 4824 if (object_array_get_with_read_barrier && kUseBakerReadBarrier) { 4825 locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty()); // No caller-save registers. 4826 } 4827 locations->SetInAt(0, Location::RequiresRegister()); 4828 locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1))); 4829 if (Primitive::IsFloatingPointType(instruction->GetType())) { 4830 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 4831 } else { 4832 // The output overlaps in the case of an object array get with 4833 // read barriers enabled: we do not want the move to overwrite the 4834 // array's location, as we need it to emit the read barrier. 4835 locations->SetOut( 4836 Location::RequiresRegister(), 4837 object_array_get_with_read_barrier ? Location::kOutputOverlap : Location::kNoOutputOverlap); 4838 } 4839 // We need a temporary register for the read barrier marking slow 4840 // path in CodeGeneratorARM::GenerateArrayLoadWithBakerReadBarrier. 4841 // Also need for String compression feature. 4842 if ((object_array_get_with_read_barrier && kUseBakerReadBarrier) 4843 || (mirror::kUseStringCompression && instruction->IsStringCharAt())) { 4844 locations->AddTemp(Location::RequiresRegister()); 4845 } 4846} 4847 4848void InstructionCodeGeneratorARM::VisitArrayGet(HArrayGet* instruction) { 4849 LocationSummary* locations = instruction->GetLocations(); 4850 Location obj_loc = locations->InAt(0); 4851 Register obj = obj_loc.AsRegister<Register>(); 4852 Location index = locations->InAt(1); 4853 Location out_loc = locations->Out(); 4854 uint32_t data_offset = CodeGenerator::GetArrayDataOffset(instruction); 4855 Primitive::Type type = instruction->GetType(); 4856 const bool maybe_compressed_char_at = mirror::kUseStringCompression && 4857 instruction->IsStringCharAt(); 4858 HInstruction* array_instr = instruction->GetArray(); 4859 bool has_intermediate_address = array_instr->IsIntermediateAddress(); 4860 4861 switch (type) { 4862 case Primitive::kPrimBoolean: 4863 case Primitive::kPrimByte: 4864 case Primitive::kPrimShort: 4865 case Primitive::kPrimChar: 4866 case Primitive::kPrimInt: { 4867 if (index.IsConstant()) { 4868 int32_t const_index = index.GetConstant()->AsIntConstant()->GetValue(); 4869 if (maybe_compressed_char_at) { 4870 Register length = IP; 4871 Label uncompressed_load, done; 4872 uint32_t count_offset = mirror::String::CountOffset().Uint32Value(); 4873 __ LoadFromOffset(kLoadWord, length, obj, count_offset); 4874 codegen_->MaybeRecordImplicitNullCheck(instruction); 4875 __ cmp(length, ShifterOperand(0)); 4876 __ b(&uncompressed_load, GE); 4877 __ LoadFromOffset(kLoadUnsignedByte, 4878 out_loc.AsRegister<Register>(), 4879 obj, 4880 data_offset + const_index); 4881 __ b(&done); 4882 __ Bind(&uncompressed_load); 4883 __ LoadFromOffset(GetLoadOperandType(Primitive::kPrimChar), 4884 out_loc.AsRegister<Register>(), 4885 obj, 4886 data_offset + (const_index << 1)); 4887 __ Bind(&done); 4888 } else { 4889 uint32_t full_offset = data_offset + (const_index << Primitive::ComponentSizeShift(type)); 4890 4891 LoadOperandType load_type = GetLoadOperandType(type); 4892 __ LoadFromOffset(load_type, out_loc.AsRegister<Register>(), obj, full_offset); 4893 } 4894 } else { 4895 Register temp = IP; 4896 4897 if (has_intermediate_address) { 4898 // We do not need to compute the intermediate address from the array: the 4899 // input instruction has done it already. See the comment in 4900 // `TryExtractArrayAccessAddress()`. 4901 if (kIsDebugBuild) { 4902 HIntermediateAddress* tmp = array_instr->AsIntermediateAddress(); 4903 DCHECK_EQ(tmp->GetOffset()->AsIntConstant()->GetValueAsUint64(), data_offset); 4904 } 4905 temp = obj; 4906 } else { 4907 __ add(temp, obj, ShifterOperand(data_offset)); 4908 } 4909 if (maybe_compressed_char_at) { 4910 Label uncompressed_load, done; 4911 uint32_t count_offset = mirror::String::CountOffset().Uint32Value(); 4912 Register length = locations->GetTemp(0).AsRegister<Register>(); 4913 __ LoadFromOffset(kLoadWord, length, obj, count_offset); 4914 codegen_->MaybeRecordImplicitNullCheck(instruction); 4915 __ cmp(length, ShifterOperand(0)); 4916 __ b(&uncompressed_load, GE); 4917 __ ldrb(out_loc.AsRegister<Register>(), 4918 Address(temp, index.AsRegister<Register>(), Shift::LSL, 0)); 4919 __ b(&done); 4920 __ Bind(&uncompressed_load); 4921 __ ldrh(out_loc.AsRegister<Register>(), 4922 Address(temp, index.AsRegister<Register>(), Shift::LSL, 1)); 4923 __ Bind(&done); 4924 } else { 4925 codegen_->LoadFromShiftedRegOffset(type, out_loc, temp, index.AsRegister<Register>()); 4926 } 4927 } 4928 break; 4929 } 4930 4931 case Primitive::kPrimNot: { 4932 // The read barrier instrumentation of object ArrayGet 4933 // instructions does not support the HIntermediateAddress 4934 // instruction. 4935 DCHECK(!(has_intermediate_address && kEmitCompilerReadBarrier)); 4936 4937 static_assert( 4938 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 4939 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 4940 // /* HeapReference<Object> */ out = 4941 // *(obj + data_offset + index * sizeof(HeapReference<Object>)) 4942 if (kEmitCompilerReadBarrier && kUseBakerReadBarrier) { 4943 Location temp = locations->GetTemp(0); 4944 // Note that a potential implicit null check is handled in this 4945 // CodeGeneratorARM::GenerateArrayLoadWithBakerReadBarrier call. 4946 codegen_->GenerateArrayLoadWithBakerReadBarrier( 4947 instruction, out_loc, obj, data_offset, index, temp, /* needs_null_check */ true); 4948 } else { 4949 Register out = out_loc.AsRegister<Register>(); 4950 if (index.IsConstant()) { 4951 size_t offset = 4952 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4953 __ LoadFromOffset(kLoadWord, out, obj, offset); 4954 codegen_->MaybeRecordImplicitNullCheck(instruction); 4955 // If read barriers are enabled, emit read barriers other than 4956 // Baker's using a slow path (and also unpoison the loaded 4957 // reference, if heap poisoning is enabled). 4958 codegen_->MaybeGenerateReadBarrierSlow(instruction, out_loc, out_loc, obj_loc, offset); 4959 } else { 4960 Register temp = IP; 4961 4962 if (has_intermediate_address) { 4963 // We do not need to compute the intermediate address from the array: the 4964 // input instruction has done it already. See the comment in 4965 // `TryExtractArrayAccessAddress()`. 4966 if (kIsDebugBuild) { 4967 HIntermediateAddress* tmp = array_instr->AsIntermediateAddress(); 4968 DCHECK_EQ(tmp->GetOffset()->AsIntConstant()->GetValueAsUint64(), data_offset); 4969 } 4970 temp = obj; 4971 } else { 4972 __ add(temp, obj, ShifterOperand(data_offset)); 4973 } 4974 codegen_->LoadFromShiftedRegOffset(type, out_loc, temp, index.AsRegister<Register>()); 4975 4976 codegen_->MaybeRecordImplicitNullCheck(instruction); 4977 // If read barriers are enabled, emit read barriers other than 4978 // Baker's using a slow path (and also unpoison the loaded 4979 // reference, if heap poisoning is enabled). 4980 codegen_->MaybeGenerateReadBarrierSlow( 4981 instruction, out_loc, out_loc, obj_loc, data_offset, index); 4982 } 4983 } 4984 break; 4985 } 4986 4987 case Primitive::kPrimLong: { 4988 if (index.IsConstant()) { 4989 size_t offset = 4990 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 4991 __ LoadFromOffset(kLoadWordPair, out_loc.AsRegisterPairLow<Register>(), obj, offset); 4992 } else { 4993 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 4994 __ LoadFromOffset(kLoadWordPair, out_loc.AsRegisterPairLow<Register>(), IP, data_offset); 4995 } 4996 break; 4997 } 4998 4999 case Primitive::kPrimFloat: { 5000 SRegister out = out_loc.AsFpuRegister<SRegister>(); 5001 if (index.IsConstant()) { 5002 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 5003 __ LoadSFromOffset(out, obj, offset); 5004 } else { 5005 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 5006 __ LoadSFromOffset(out, IP, data_offset); 5007 } 5008 break; 5009 } 5010 5011 case Primitive::kPrimDouble: { 5012 SRegister out = out_loc.AsFpuRegisterPairLow<SRegister>(); 5013 if (index.IsConstant()) { 5014 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 5015 __ LoadDFromOffset(FromLowSToD(out), obj, offset); 5016 } else { 5017 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 5018 __ LoadDFromOffset(FromLowSToD(out), IP, data_offset); 5019 } 5020 break; 5021 } 5022 5023 case Primitive::kPrimVoid: 5024 LOG(FATAL) << "Unreachable type " << type; 5025 UNREACHABLE(); 5026 } 5027 5028 if (type == Primitive::kPrimNot) { 5029 // Potential implicit null checks, in the case of reference 5030 // arrays, are handled in the previous switch statement. 5031 } else if (!maybe_compressed_char_at) { 5032 codegen_->MaybeRecordImplicitNullCheck(instruction); 5033 } 5034} 5035 5036void LocationsBuilderARM::VisitArraySet(HArraySet* instruction) { 5037 Primitive::Type value_type = instruction->GetComponentType(); 5038 5039 bool needs_write_barrier = 5040 CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue()); 5041 bool may_need_runtime_call_for_type_check = instruction->NeedsTypeCheck(); 5042 5043 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary( 5044 instruction, 5045 may_need_runtime_call_for_type_check ? 5046 LocationSummary::kCallOnSlowPath : 5047 LocationSummary::kNoCall); 5048 5049 locations->SetInAt(0, Location::RequiresRegister()); 5050 locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1))); 5051 if (Primitive::IsFloatingPointType(value_type)) { 5052 locations->SetInAt(2, Location::RequiresFpuRegister()); 5053 } else { 5054 locations->SetInAt(2, Location::RequiresRegister()); 5055 } 5056 if (needs_write_barrier) { 5057 // Temporary registers for the write barrier. 5058 locations->AddTemp(Location::RequiresRegister()); // Possibly used for ref. poisoning too. 5059 locations->AddTemp(Location::RequiresRegister()); 5060 } 5061} 5062 5063void InstructionCodeGeneratorARM::VisitArraySet(HArraySet* instruction) { 5064 LocationSummary* locations = instruction->GetLocations(); 5065 Location array_loc = locations->InAt(0); 5066 Register array = array_loc.AsRegister<Register>(); 5067 Location index = locations->InAt(1); 5068 Primitive::Type value_type = instruction->GetComponentType(); 5069 bool may_need_runtime_call_for_type_check = instruction->NeedsTypeCheck(); 5070 bool needs_write_barrier = 5071 CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue()); 5072 uint32_t data_offset = 5073 mirror::Array::DataOffset(Primitive::ComponentSize(value_type)).Uint32Value(); 5074 Location value_loc = locations->InAt(2); 5075 HInstruction* array_instr = instruction->GetArray(); 5076 bool has_intermediate_address = array_instr->IsIntermediateAddress(); 5077 5078 switch (value_type) { 5079 case Primitive::kPrimBoolean: 5080 case Primitive::kPrimByte: 5081 case Primitive::kPrimShort: 5082 case Primitive::kPrimChar: 5083 case Primitive::kPrimInt: { 5084 if (index.IsConstant()) { 5085 int32_t const_index = index.GetConstant()->AsIntConstant()->GetValue(); 5086 uint32_t full_offset = 5087 data_offset + (const_index << Primitive::ComponentSizeShift(value_type)); 5088 StoreOperandType store_type = GetStoreOperandType(value_type); 5089 __ StoreToOffset(store_type, value_loc.AsRegister<Register>(), array, full_offset); 5090 } else { 5091 Register temp = IP; 5092 5093 if (has_intermediate_address) { 5094 // We do not need to compute the intermediate address from the array: the 5095 // input instruction has done it already. See the comment in 5096 // `TryExtractArrayAccessAddress()`. 5097 if (kIsDebugBuild) { 5098 HIntermediateAddress* tmp = array_instr->AsIntermediateAddress(); 5099 DCHECK(tmp->GetOffset()->AsIntConstant()->GetValueAsUint64() == data_offset); 5100 } 5101 temp = array; 5102 } else { 5103 __ add(temp, array, ShifterOperand(data_offset)); 5104 } 5105 codegen_->StoreToShiftedRegOffset(value_type, 5106 value_loc, 5107 temp, 5108 index.AsRegister<Register>()); 5109 } 5110 break; 5111 } 5112 5113 case Primitive::kPrimNot: { 5114 Register value = value_loc.AsRegister<Register>(); 5115 // TryExtractArrayAccessAddress optimization is never applied for non-primitive ArraySet. 5116 // See the comment in instruction_simplifier_shared.cc. 5117 DCHECK(!has_intermediate_address); 5118 5119 if (instruction->InputAt(2)->IsNullConstant()) { 5120 // Just setting null. 5121 if (index.IsConstant()) { 5122 size_t offset = 5123 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 5124 __ StoreToOffset(kStoreWord, value, array, offset); 5125 } else { 5126 DCHECK(index.IsRegister()) << index; 5127 __ add(IP, array, ShifterOperand(data_offset)); 5128 codegen_->StoreToShiftedRegOffset(value_type, 5129 value_loc, 5130 IP, 5131 index.AsRegister<Register>()); 5132 } 5133 codegen_->MaybeRecordImplicitNullCheck(instruction); 5134 DCHECK(!needs_write_barrier); 5135 DCHECK(!may_need_runtime_call_for_type_check); 5136 break; 5137 } 5138 5139 DCHECK(needs_write_barrier); 5140 Location temp1_loc = locations->GetTemp(0); 5141 Register temp1 = temp1_loc.AsRegister<Register>(); 5142 Location temp2_loc = locations->GetTemp(1); 5143 Register temp2 = temp2_loc.AsRegister<Register>(); 5144 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 5145 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 5146 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 5147 Label done; 5148 SlowPathCodeARM* slow_path = nullptr; 5149 5150 if (may_need_runtime_call_for_type_check) { 5151 slow_path = new (GetGraph()->GetArena()) ArraySetSlowPathARM(instruction); 5152 codegen_->AddSlowPath(slow_path); 5153 if (instruction->GetValueCanBeNull()) { 5154 Label non_zero; 5155 __ CompareAndBranchIfNonZero(value, &non_zero); 5156 if (index.IsConstant()) { 5157 size_t offset = 5158 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 5159 __ StoreToOffset(kStoreWord, value, array, offset); 5160 } else { 5161 DCHECK(index.IsRegister()) << index; 5162 __ add(IP, array, ShifterOperand(data_offset)); 5163 codegen_->StoreToShiftedRegOffset(value_type, 5164 value_loc, 5165 IP, 5166 index.AsRegister<Register>()); 5167 } 5168 codegen_->MaybeRecordImplicitNullCheck(instruction); 5169 __ b(&done); 5170 __ Bind(&non_zero); 5171 } 5172 5173 // Note that when read barriers are enabled, the type checks 5174 // are performed without read barriers. This is fine, even in 5175 // the case where a class object is in the from-space after 5176 // the flip, as a comparison involving such a type would not 5177 // produce a false positive; it may of course produce a false 5178 // negative, in which case we would take the ArraySet slow 5179 // path. 5180 5181 // /* HeapReference<Class> */ temp1 = array->klass_ 5182 __ LoadFromOffset(kLoadWord, temp1, array, class_offset); 5183 codegen_->MaybeRecordImplicitNullCheck(instruction); 5184 __ MaybeUnpoisonHeapReference(temp1); 5185 5186 // /* HeapReference<Class> */ temp1 = temp1->component_type_ 5187 __ LoadFromOffset(kLoadWord, temp1, temp1, component_offset); 5188 // /* HeapReference<Class> */ temp2 = value->klass_ 5189 __ LoadFromOffset(kLoadWord, temp2, value, class_offset); 5190 // If heap poisoning is enabled, no need to unpoison `temp1` 5191 // nor `temp2`, as we are comparing two poisoned references. 5192 __ cmp(temp1, ShifterOperand(temp2)); 5193 5194 if (instruction->StaticTypeOfArrayIsObjectArray()) { 5195 Label do_put; 5196 __ b(&do_put, EQ); 5197 // If heap poisoning is enabled, the `temp1` reference has 5198 // not been unpoisoned yet; unpoison it now. 5199 __ MaybeUnpoisonHeapReference(temp1); 5200 5201 // /* HeapReference<Class> */ temp1 = temp1->super_class_ 5202 __ LoadFromOffset(kLoadWord, temp1, temp1, super_offset); 5203 // If heap poisoning is enabled, no need to unpoison 5204 // `temp1`, as we are comparing against null below. 5205 __ CompareAndBranchIfNonZero(temp1, slow_path->GetEntryLabel()); 5206 __ Bind(&do_put); 5207 } else { 5208 __ b(slow_path->GetEntryLabel(), NE); 5209 } 5210 } 5211 5212 Register source = value; 5213 if (kPoisonHeapReferences) { 5214 // Note that in the case where `value` is a null reference, 5215 // we do not enter this block, as a null reference does not 5216 // need poisoning. 5217 DCHECK_EQ(value_type, Primitive::kPrimNot); 5218 __ Mov(temp1, value); 5219 __ PoisonHeapReference(temp1); 5220 source = temp1; 5221 } 5222 5223 if (index.IsConstant()) { 5224 size_t offset = 5225 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 5226 __ StoreToOffset(kStoreWord, source, array, offset); 5227 } else { 5228 DCHECK(index.IsRegister()) << index; 5229 5230 __ add(IP, array, ShifterOperand(data_offset)); 5231 codegen_->StoreToShiftedRegOffset(value_type, 5232 Location::RegisterLocation(source), 5233 IP, 5234 index.AsRegister<Register>()); 5235 } 5236 5237 if (!may_need_runtime_call_for_type_check) { 5238 codegen_->MaybeRecordImplicitNullCheck(instruction); 5239 } 5240 5241 codegen_->MarkGCCard(temp1, temp2, array, value, instruction->GetValueCanBeNull()); 5242 5243 if (done.IsLinked()) { 5244 __ Bind(&done); 5245 } 5246 5247 if (slow_path != nullptr) { 5248 __ Bind(slow_path->GetExitLabel()); 5249 } 5250 5251 break; 5252 } 5253 5254 case Primitive::kPrimLong: { 5255 Location value = locations->InAt(2); 5256 if (index.IsConstant()) { 5257 size_t offset = 5258 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 5259 __ StoreToOffset(kStoreWordPair, value.AsRegisterPairLow<Register>(), array, offset); 5260 } else { 5261 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 5262 __ StoreToOffset(kStoreWordPair, value.AsRegisterPairLow<Register>(), IP, data_offset); 5263 } 5264 break; 5265 } 5266 5267 case Primitive::kPrimFloat: { 5268 Location value = locations->InAt(2); 5269 DCHECK(value.IsFpuRegister()); 5270 if (index.IsConstant()) { 5271 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 5272 __ StoreSToOffset(value.AsFpuRegister<SRegister>(), array, offset); 5273 } else { 5274 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 5275 __ StoreSToOffset(value.AsFpuRegister<SRegister>(), IP, data_offset); 5276 } 5277 break; 5278 } 5279 5280 case Primitive::kPrimDouble: { 5281 Location value = locations->InAt(2); 5282 DCHECK(value.IsFpuRegisterPair()); 5283 if (index.IsConstant()) { 5284 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 5285 __ StoreDToOffset(FromLowSToD(value.AsFpuRegisterPairLow<SRegister>()), array, offset); 5286 } else { 5287 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 5288 __ StoreDToOffset(FromLowSToD(value.AsFpuRegisterPairLow<SRegister>()), IP, data_offset); 5289 } 5290 5291 break; 5292 } 5293 5294 case Primitive::kPrimVoid: 5295 LOG(FATAL) << "Unreachable type " << value_type; 5296 UNREACHABLE(); 5297 } 5298 5299 // Objects are handled in the switch. 5300 if (value_type != Primitive::kPrimNot) { 5301 codegen_->MaybeRecordImplicitNullCheck(instruction); 5302 } 5303} 5304 5305void LocationsBuilderARM::VisitArrayLength(HArrayLength* instruction) { 5306 LocationSummary* locations = 5307 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 5308 locations->SetInAt(0, Location::RequiresRegister()); 5309 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 5310} 5311 5312void InstructionCodeGeneratorARM::VisitArrayLength(HArrayLength* instruction) { 5313 LocationSummary* locations = instruction->GetLocations(); 5314 uint32_t offset = CodeGenerator::GetArrayLengthOffset(instruction); 5315 Register obj = locations->InAt(0).AsRegister<Register>(); 5316 Register out = locations->Out().AsRegister<Register>(); 5317 __ LoadFromOffset(kLoadWord, out, obj, offset); 5318 codegen_->MaybeRecordImplicitNullCheck(instruction); 5319 // Mask out compression flag from String's array length. 5320 if (mirror::kUseStringCompression && instruction->IsStringLength()) { 5321 __ bic(out, out, ShifterOperand(1u << 31)); 5322 } 5323} 5324 5325void LocationsBuilderARM::VisitIntermediateAddress(HIntermediateAddress* instruction) { 5326 LocationSummary* locations = 5327 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 5328 5329 locations->SetInAt(0, Location::RequiresRegister()); 5330 locations->SetInAt(1, Location::RegisterOrConstant(instruction->GetOffset())); 5331 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 5332} 5333 5334void InstructionCodeGeneratorARM::VisitIntermediateAddress(HIntermediateAddress* instruction) { 5335 LocationSummary* locations = instruction->GetLocations(); 5336 Location out = locations->Out(); 5337 Location first = locations->InAt(0); 5338 Location second = locations->InAt(1); 5339 5340 if (second.IsRegister()) { 5341 __ add(out.AsRegister<Register>(), 5342 first.AsRegister<Register>(), 5343 ShifterOperand(second.AsRegister<Register>())); 5344 } else { 5345 __ AddConstant(out.AsRegister<Register>(), 5346 first.AsRegister<Register>(), 5347 second.GetConstant()->AsIntConstant()->GetValue()); 5348 } 5349} 5350 5351void LocationsBuilderARM::VisitBoundsCheck(HBoundsCheck* instruction) { 5352 RegisterSet caller_saves = RegisterSet::Empty(); 5353 InvokeRuntimeCallingConvention calling_convention; 5354 caller_saves.Add(Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 5355 caller_saves.Add(Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 5356 LocationSummary* locations = codegen_->CreateThrowingSlowPathLocations(instruction, caller_saves); 5357 locations->SetInAt(0, Location::RequiresRegister()); 5358 locations->SetInAt(1, Location::RequiresRegister()); 5359} 5360 5361void InstructionCodeGeneratorARM::VisitBoundsCheck(HBoundsCheck* instruction) { 5362 LocationSummary* locations = instruction->GetLocations(); 5363 SlowPathCodeARM* slow_path = 5364 new (GetGraph()->GetArena()) BoundsCheckSlowPathARM(instruction); 5365 codegen_->AddSlowPath(slow_path); 5366 5367 Register index = locations->InAt(0).AsRegister<Register>(); 5368 Register length = locations->InAt(1).AsRegister<Register>(); 5369 5370 __ cmp(index, ShifterOperand(length)); 5371 __ b(slow_path->GetEntryLabel(), HS); 5372} 5373 5374void CodeGeneratorARM::MarkGCCard(Register temp, 5375 Register card, 5376 Register object, 5377 Register value, 5378 bool can_be_null) { 5379 Label is_null; 5380 if (can_be_null) { 5381 __ CompareAndBranchIfZero(value, &is_null); 5382 } 5383 __ LoadFromOffset(kLoadWord, card, TR, Thread::CardTableOffset<kArmPointerSize>().Int32Value()); 5384 __ Lsr(temp, object, gc::accounting::CardTable::kCardShift); 5385 __ strb(card, Address(card, temp)); 5386 if (can_be_null) { 5387 __ Bind(&is_null); 5388 } 5389} 5390 5391void LocationsBuilderARM::VisitParallelMove(HParallelMove* instruction ATTRIBUTE_UNUSED) { 5392 LOG(FATAL) << "Unreachable"; 5393} 5394 5395void InstructionCodeGeneratorARM::VisitParallelMove(HParallelMove* instruction) { 5396 codegen_->GetMoveResolver()->EmitNativeCode(instruction); 5397} 5398 5399void LocationsBuilderARM::VisitSuspendCheck(HSuspendCheck* instruction) { 5400 LocationSummary* locations = 5401 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnSlowPath); 5402 locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty()); // No caller-save registers. 5403} 5404 5405void InstructionCodeGeneratorARM::VisitSuspendCheck(HSuspendCheck* instruction) { 5406 HBasicBlock* block = instruction->GetBlock(); 5407 if (block->GetLoopInformation() != nullptr) { 5408 DCHECK(block->GetLoopInformation()->GetSuspendCheck() == instruction); 5409 // The back edge will generate the suspend check. 5410 return; 5411 } 5412 if (block->IsEntryBlock() && instruction->GetNext()->IsGoto()) { 5413 // The goto will generate the suspend check. 5414 return; 5415 } 5416 GenerateSuspendCheck(instruction, nullptr); 5417} 5418 5419void InstructionCodeGeneratorARM::GenerateSuspendCheck(HSuspendCheck* instruction, 5420 HBasicBlock* successor) { 5421 SuspendCheckSlowPathARM* slow_path = 5422 down_cast<SuspendCheckSlowPathARM*>(instruction->GetSlowPath()); 5423 if (slow_path == nullptr) { 5424 slow_path = new (GetGraph()->GetArena()) SuspendCheckSlowPathARM(instruction, successor); 5425 instruction->SetSlowPath(slow_path); 5426 codegen_->AddSlowPath(slow_path); 5427 if (successor != nullptr) { 5428 DCHECK(successor->IsLoopHeader()); 5429 codegen_->ClearSpillSlotsFromLoopPhisInStackMap(instruction); 5430 } 5431 } else { 5432 DCHECK_EQ(slow_path->GetSuccessor(), successor); 5433 } 5434 5435 __ LoadFromOffset( 5436 kLoadUnsignedHalfword, IP, TR, Thread::ThreadFlagsOffset<kArmPointerSize>().Int32Value()); 5437 if (successor == nullptr) { 5438 __ CompareAndBranchIfNonZero(IP, slow_path->GetEntryLabel()); 5439 __ Bind(slow_path->GetReturnLabel()); 5440 } else { 5441 __ CompareAndBranchIfZero(IP, codegen_->GetLabelOf(successor)); 5442 __ b(slow_path->GetEntryLabel()); 5443 } 5444} 5445 5446ArmAssembler* ParallelMoveResolverARM::GetAssembler() const { 5447 return codegen_->GetAssembler(); 5448} 5449 5450void ParallelMoveResolverARM::EmitMove(size_t index) { 5451 MoveOperands* move = moves_[index]; 5452 Location source = move->GetSource(); 5453 Location destination = move->GetDestination(); 5454 5455 if (source.IsRegister()) { 5456 if (destination.IsRegister()) { 5457 __ Mov(destination.AsRegister<Register>(), source.AsRegister<Register>()); 5458 } else if (destination.IsFpuRegister()) { 5459 __ vmovsr(destination.AsFpuRegister<SRegister>(), source.AsRegister<Register>()); 5460 } else { 5461 DCHECK(destination.IsStackSlot()); 5462 __ StoreToOffset(kStoreWord, source.AsRegister<Register>(), 5463 SP, destination.GetStackIndex()); 5464 } 5465 } else if (source.IsStackSlot()) { 5466 if (destination.IsRegister()) { 5467 __ LoadFromOffset(kLoadWord, destination.AsRegister<Register>(), 5468 SP, source.GetStackIndex()); 5469 } else if (destination.IsFpuRegister()) { 5470 __ LoadSFromOffset(destination.AsFpuRegister<SRegister>(), SP, source.GetStackIndex()); 5471 } else { 5472 DCHECK(destination.IsStackSlot()); 5473 __ LoadFromOffset(kLoadWord, IP, SP, source.GetStackIndex()); 5474 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 5475 } 5476 } else if (source.IsFpuRegister()) { 5477 if (destination.IsRegister()) { 5478 __ vmovrs(destination.AsRegister<Register>(), source.AsFpuRegister<SRegister>()); 5479 } else if (destination.IsFpuRegister()) { 5480 __ vmovs(destination.AsFpuRegister<SRegister>(), source.AsFpuRegister<SRegister>()); 5481 } else { 5482 DCHECK(destination.IsStackSlot()); 5483 __ StoreSToOffset(source.AsFpuRegister<SRegister>(), SP, destination.GetStackIndex()); 5484 } 5485 } else if (source.IsDoubleStackSlot()) { 5486 if (destination.IsDoubleStackSlot()) { 5487 __ LoadDFromOffset(DTMP, SP, source.GetStackIndex()); 5488 __ StoreDToOffset(DTMP, SP, destination.GetStackIndex()); 5489 } else if (destination.IsRegisterPair()) { 5490 DCHECK(ExpectedPairLayout(destination)); 5491 __ LoadFromOffset( 5492 kLoadWordPair, destination.AsRegisterPairLow<Register>(), SP, source.GetStackIndex()); 5493 } else { 5494 DCHECK(destination.IsFpuRegisterPair()) << destination; 5495 __ LoadDFromOffset(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 5496 SP, 5497 source.GetStackIndex()); 5498 } 5499 } else if (source.IsRegisterPair()) { 5500 if (destination.IsRegisterPair()) { 5501 __ Mov(destination.AsRegisterPairLow<Register>(), source.AsRegisterPairLow<Register>()); 5502 __ Mov(destination.AsRegisterPairHigh<Register>(), source.AsRegisterPairHigh<Register>()); 5503 } else if (destination.IsFpuRegisterPair()) { 5504 __ vmovdrr(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 5505 source.AsRegisterPairLow<Register>(), 5506 source.AsRegisterPairHigh<Register>()); 5507 } else { 5508 DCHECK(destination.IsDoubleStackSlot()) << destination; 5509 DCHECK(ExpectedPairLayout(source)); 5510 __ StoreToOffset( 5511 kStoreWordPair, source.AsRegisterPairLow<Register>(), SP, destination.GetStackIndex()); 5512 } 5513 } else if (source.IsFpuRegisterPair()) { 5514 if (destination.IsRegisterPair()) { 5515 __ vmovrrd(destination.AsRegisterPairLow<Register>(), 5516 destination.AsRegisterPairHigh<Register>(), 5517 FromLowSToD(source.AsFpuRegisterPairLow<SRegister>())); 5518 } else if (destination.IsFpuRegisterPair()) { 5519 __ vmovd(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 5520 FromLowSToD(source.AsFpuRegisterPairLow<SRegister>())); 5521 } else { 5522 DCHECK(destination.IsDoubleStackSlot()) << destination; 5523 __ StoreDToOffset(FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()), 5524 SP, 5525 destination.GetStackIndex()); 5526 } 5527 } else { 5528 DCHECK(source.IsConstant()) << source; 5529 HConstant* constant = source.GetConstant(); 5530 if (constant->IsIntConstant() || constant->IsNullConstant()) { 5531 int32_t value = CodeGenerator::GetInt32ValueOf(constant); 5532 if (destination.IsRegister()) { 5533 __ LoadImmediate(destination.AsRegister<Register>(), value); 5534 } else { 5535 DCHECK(destination.IsStackSlot()); 5536 __ LoadImmediate(IP, value); 5537 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 5538 } 5539 } else if (constant->IsLongConstant()) { 5540 int64_t value = constant->AsLongConstant()->GetValue(); 5541 if (destination.IsRegisterPair()) { 5542 __ LoadImmediate(destination.AsRegisterPairLow<Register>(), Low32Bits(value)); 5543 __ LoadImmediate(destination.AsRegisterPairHigh<Register>(), High32Bits(value)); 5544 } else { 5545 DCHECK(destination.IsDoubleStackSlot()) << destination; 5546 __ LoadImmediate(IP, Low32Bits(value)); 5547 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 5548 __ LoadImmediate(IP, High32Bits(value)); 5549 __ StoreToOffset(kStoreWord, IP, SP, destination.GetHighStackIndex(kArmWordSize)); 5550 } 5551 } else if (constant->IsDoubleConstant()) { 5552 double value = constant->AsDoubleConstant()->GetValue(); 5553 if (destination.IsFpuRegisterPair()) { 5554 __ LoadDImmediate(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), value); 5555 } else { 5556 DCHECK(destination.IsDoubleStackSlot()) << destination; 5557 uint64_t int_value = bit_cast<uint64_t, double>(value); 5558 __ LoadImmediate(IP, Low32Bits(int_value)); 5559 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 5560 __ LoadImmediate(IP, High32Bits(int_value)); 5561 __ StoreToOffset(kStoreWord, IP, SP, destination.GetHighStackIndex(kArmWordSize)); 5562 } 5563 } else { 5564 DCHECK(constant->IsFloatConstant()) << constant->DebugName(); 5565 float value = constant->AsFloatConstant()->GetValue(); 5566 if (destination.IsFpuRegister()) { 5567 __ LoadSImmediate(destination.AsFpuRegister<SRegister>(), value); 5568 } else { 5569 DCHECK(destination.IsStackSlot()); 5570 __ LoadImmediate(IP, bit_cast<int32_t, float>(value)); 5571 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 5572 } 5573 } 5574 } 5575} 5576 5577void ParallelMoveResolverARM::Exchange(Register reg, int mem) { 5578 __ Mov(IP, reg); 5579 __ LoadFromOffset(kLoadWord, reg, SP, mem); 5580 __ StoreToOffset(kStoreWord, IP, SP, mem); 5581} 5582 5583void ParallelMoveResolverARM::Exchange(int mem1, int mem2) { 5584 ScratchRegisterScope ensure_scratch(this, IP, R0, codegen_->GetNumberOfCoreRegisters()); 5585 int stack_offset = ensure_scratch.IsSpilled() ? kArmWordSize : 0; 5586 __ LoadFromOffset(kLoadWord, static_cast<Register>(ensure_scratch.GetRegister()), 5587 SP, mem1 + stack_offset); 5588 __ LoadFromOffset(kLoadWord, IP, SP, mem2 + stack_offset); 5589 __ StoreToOffset(kStoreWord, static_cast<Register>(ensure_scratch.GetRegister()), 5590 SP, mem2 + stack_offset); 5591 __ StoreToOffset(kStoreWord, IP, SP, mem1 + stack_offset); 5592} 5593 5594void ParallelMoveResolverARM::EmitSwap(size_t index) { 5595 MoveOperands* move = moves_[index]; 5596 Location source = move->GetSource(); 5597 Location destination = move->GetDestination(); 5598 5599 if (source.IsRegister() && destination.IsRegister()) { 5600 DCHECK_NE(source.AsRegister<Register>(), IP); 5601 DCHECK_NE(destination.AsRegister<Register>(), IP); 5602 __ Mov(IP, source.AsRegister<Register>()); 5603 __ Mov(source.AsRegister<Register>(), destination.AsRegister<Register>()); 5604 __ Mov(destination.AsRegister<Register>(), IP); 5605 } else if (source.IsRegister() && destination.IsStackSlot()) { 5606 Exchange(source.AsRegister<Register>(), destination.GetStackIndex()); 5607 } else if (source.IsStackSlot() && destination.IsRegister()) { 5608 Exchange(destination.AsRegister<Register>(), source.GetStackIndex()); 5609 } else if (source.IsStackSlot() && destination.IsStackSlot()) { 5610 Exchange(source.GetStackIndex(), destination.GetStackIndex()); 5611 } else if (source.IsFpuRegister() && destination.IsFpuRegister()) { 5612 __ vmovrs(IP, source.AsFpuRegister<SRegister>()); 5613 __ vmovs(source.AsFpuRegister<SRegister>(), destination.AsFpuRegister<SRegister>()); 5614 __ vmovsr(destination.AsFpuRegister<SRegister>(), IP); 5615 } else if (source.IsRegisterPair() && destination.IsRegisterPair()) { 5616 __ vmovdrr(DTMP, source.AsRegisterPairLow<Register>(), source.AsRegisterPairHigh<Register>()); 5617 __ Mov(source.AsRegisterPairLow<Register>(), destination.AsRegisterPairLow<Register>()); 5618 __ Mov(source.AsRegisterPairHigh<Register>(), destination.AsRegisterPairHigh<Register>()); 5619 __ vmovrrd(destination.AsRegisterPairLow<Register>(), 5620 destination.AsRegisterPairHigh<Register>(), 5621 DTMP); 5622 } else if (source.IsRegisterPair() || destination.IsRegisterPair()) { 5623 Register low_reg = source.IsRegisterPair() 5624 ? source.AsRegisterPairLow<Register>() 5625 : destination.AsRegisterPairLow<Register>(); 5626 int mem = source.IsRegisterPair() 5627 ? destination.GetStackIndex() 5628 : source.GetStackIndex(); 5629 DCHECK(ExpectedPairLayout(source.IsRegisterPair() ? source : destination)); 5630 __ vmovdrr(DTMP, low_reg, static_cast<Register>(low_reg + 1)); 5631 __ LoadFromOffset(kLoadWordPair, low_reg, SP, mem); 5632 __ StoreDToOffset(DTMP, SP, mem); 5633 } else if (source.IsFpuRegisterPair() && destination.IsFpuRegisterPair()) { 5634 DRegister first = FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()); 5635 DRegister second = FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()); 5636 __ vmovd(DTMP, first); 5637 __ vmovd(first, second); 5638 __ vmovd(second, DTMP); 5639 } else if (source.IsFpuRegisterPair() || destination.IsFpuRegisterPair()) { 5640 DRegister reg = source.IsFpuRegisterPair() 5641 ? FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()) 5642 : FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()); 5643 int mem = source.IsFpuRegisterPair() 5644 ? destination.GetStackIndex() 5645 : source.GetStackIndex(); 5646 __ vmovd(DTMP, reg); 5647 __ LoadDFromOffset(reg, SP, mem); 5648 __ StoreDToOffset(DTMP, SP, mem); 5649 } else if (source.IsFpuRegister() || destination.IsFpuRegister()) { 5650 SRegister reg = source.IsFpuRegister() ? source.AsFpuRegister<SRegister>() 5651 : destination.AsFpuRegister<SRegister>(); 5652 int mem = source.IsFpuRegister() 5653 ? destination.GetStackIndex() 5654 : source.GetStackIndex(); 5655 5656 __ vmovrs(IP, reg); 5657 __ LoadSFromOffset(reg, SP, mem); 5658 __ StoreToOffset(kStoreWord, IP, SP, mem); 5659 } else if (source.IsDoubleStackSlot() && destination.IsDoubleStackSlot()) { 5660 Exchange(source.GetStackIndex(), destination.GetStackIndex()); 5661 Exchange(source.GetHighStackIndex(kArmWordSize), destination.GetHighStackIndex(kArmWordSize)); 5662 } else { 5663 LOG(FATAL) << "Unimplemented" << source << " <-> " << destination; 5664 } 5665} 5666 5667void ParallelMoveResolverARM::SpillScratch(int reg) { 5668 __ Push(static_cast<Register>(reg)); 5669} 5670 5671void ParallelMoveResolverARM::RestoreScratch(int reg) { 5672 __ Pop(static_cast<Register>(reg)); 5673} 5674 5675HLoadClass::LoadKind CodeGeneratorARM::GetSupportedLoadClassKind( 5676 HLoadClass::LoadKind desired_class_load_kind) { 5677 switch (desired_class_load_kind) { 5678 case HLoadClass::LoadKind::kReferrersClass: 5679 break; 5680 case HLoadClass::LoadKind::kBootImageLinkTimeAddress: 5681 DCHECK(!GetCompilerOptions().GetCompilePic()); 5682 break; 5683 case HLoadClass::LoadKind::kBootImageLinkTimePcRelative: 5684 DCHECK(GetCompilerOptions().GetCompilePic()); 5685 break; 5686 case HLoadClass::LoadKind::kBootImageAddress: 5687 break; 5688 case HLoadClass::LoadKind::kDexCacheAddress: 5689 DCHECK(Runtime::Current()->UseJitCompilation()); 5690 break; 5691 case HLoadClass::LoadKind::kDexCachePcRelative: 5692 DCHECK(!Runtime::Current()->UseJitCompilation()); 5693 // We disable pc-relative load when there is an irreducible loop, as the optimization 5694 // is incompatible with it. 5695 // TODO: Create as many ArmDexCacheArraysBase instructions as needed for methods 5696 // with irreducible loops. 5697 if (GetGraph()->HasIrreducibleLoops()) { 5698 return HLoadClass::LoadKind::kDexCacheViaMethod; 5699 } 5700 break; 5701 case HLoadClass::LoadKind::kDexCacheViaMethod: 5702 break; 5703 } 5704 return desired_class_load_kind; 5705} 5706 5707void LocationsBuilderARM::VisitLoadClass(HLoadClass* cls) { 5708 if (cls->NeedsAccessCheck()) { 5709 InvokeRuntimeCallingConvention calling_convention; 5710 CodeGenerator::CreateLoadClassLocationSummary( 5711 cls, 5712 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 5713 Location::RegisterLocation(R0), 5714 /* code_generator_supports_read_barrier */ true); 5715 return; 5716 } 5717 5718 const bool requires_read_barrier = kEmitCompilerReadBarrier && !cls->IsInBootImage(); 5719 LocationSummary::CallKind call_kind = (cls->NeedsEnvironment() || requires_read_barrier) 5720 ? LocationSummary::kCallOnSlowPath 5721 : LocationSummary::kNoCall; 5722 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(cls, call_kind); 5723 if (kUseBakerReadBarrier && requires_read_barrier && !cls->NeedsEnvironment()) { 5724 locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty()); // No caller-save registers. 5725 } 5726 5727 HLoadClass::LoadKind load_kind = cls->GetLoadKind(); 5728 if (load_kind == HLoadClass::LoadKind::kReferrersClass || 5729 load_kind == HLoadClass::LoadKind::kDexCacheViaMethod || 5730 load_kind == HLoadClass::LoadKind::kDexCachePcRelative) { 5731 locations->SetInAt(0, Location::RequiresRegister()); 5732 } 5733 locations->SetOut(Location::RequiresRegister()); 5734} 5735 5736void InstructionCodeGeneratorARM::VisitLoadClass(HLoadClass* cls) { 5737 LocationSummary* locations = cls->GetLocations(); 5738 if (cls->NeedsAccessCheck()) { 5739 codegen_->MoveConstant(locations->GetTemp(0), cls->GetTypeIndex()); 5740 codegen_->InvokeRuntime(kQuickInitializeTypeAndVerifyAccess, cls, cls->GetDexPc()); 5741 CheckEntrypointTypes<kQuickInitializeTypeAndVerifyAccess, void*, uint32_t>(); 5742 return; 5743 } 5744 5745 Location out_loc = locations->Out(); 5746 Register out = out_loc.AsRegister<Register>(); 5747 5748 const bool requires_read_barrier = kEmitCompilerReadBarrier && !cls->IsInBootImage(); 5749 bool generate_null_check = false; 5750 switch (cls->GetLoadKind()) { 5751 case HLoadClass::LoadKind::kReferrersClass: { 5752 DCHECK(!cls->CanCallRuntime()); 5753 DCHECK(!cls->MustGenerateClinitCheck()); 5754 // /* GcRoot<mirror::Class> */ out = current_method->declaring_class_ 5755 Register current_method = locations->InAt(0).AsRegister<Register>(); 5756 GenerateGcRootFieldLoad(cls, 5757 out_loc, 5758 current_method, 5759 ArtMethod::DeclaringClassOffset().Int32Value(), 5760 requires_read_barrier); 5761 break; 5762 } 5763 case HLoadClass::LoadKind::kBootImageLinkTimeAddress: { 5764 DCHECK(!requires_read_barrier); 5765 __ LoadLiteral(out, codegen_->DeduplicateBootImageTypeLiteral(cls->GetDexFile(), 5766 cls->GetTypeIndex())); 5767 break; 5768 } 5769 case HLoadClass::LoadKind::kBootImageLinkTimePcRelative: { 5770 DCHECK(!requires_read_barrier); 5771 CodeGeneratorARM::PcRelativePatchInfo* labels = 5772 codegen_->NewPcRelativeTypePatch(cls->GetDexFile(), cls->GetTypeIndex()); 5773 __ BindTrackedLabel(&labels->movw_label); 5774 __ movw(out, /* placeholder */ 0u); 5775 __ BindTrackedLabel(&labels->movt_label); 5776 __ movt(out, /* placeholder */ 0u); 5777 __ BindTrackedLabel(&labels->add_pc_label); 5778 __ add(out, out, ShifterOperand(PC)); 5779 break; 5780 } 5781 case HLoadClass::LoadKind::kBootImageAddress: { 5782 DCHECK(!requires_read_barrier); 5783 DCHECK_NE(cls->GetAddress(), 0u); 5784 uint32_t address = dchecked_integral_cast<uint32_t>(cls->GetAddress()); 5785 __ LoadLiteral(out, codegen_->DeduplicateBootImageAddressLiteral(address)); 5786 break; 5787 } 5788 case HLoadClass::LoadKind::kDexCacheAddress: { 5789 DCHECK_NE(cls->GetAddress(), 0u); 5790 uint32_t address = dchecked_integral_cast<uint32_t>(cls->GetAddress()); 5791 // 16-bit LDR immediate has a 5-bit offset multiplied by the size and that gives 5792 // a 128B range. To try and reduce the number of literals if we load multiple types, 5793 // simply split the dex cache address to a 128B aligned base loaded from a literal 5794 // and the remaining offset embedded in the load. 5795 static_assert(sizeof(GcRoot<mirror::Class>) == 4u, "Expected GC root to be 4 bytes."); 5796 DCHECK_ALIGNED(cls->GetAddress(), 4u); 5797 constexpr size_t offset_bits = /* encoded bits */ 5 + /* scale */ 2; 5798 uint32_t base_address = address & ~MaxInt<uint32_t>(offset_bits); 5799 uint32_t offset = address & MaxInt<uint32_t>(offset_bits); 5800 __ LoadLiteral(out, codegen_->DeduplicateDexCacheAddressLiteral(base_address)); 5801 // /* GcRoot<mirror::Class> */ out = *(base_address + offset) 5802 GenerateGcRootFieldLoad(cls, out_loc, out, offset, requires_read_barrier); 5803 generate_null_check = !cls->IsInDexCache(); 5804 break; 5805 } 5806 case HLoadClass::LoadKind::kDexCachePcRelative: { 5807 Register base_reg = locations->InAt(0).AsRegister<Register>(); 5808 HArmDexCacheArraysBase* base = cls->InputAt(0)->AsArmDexCacheArraysBase(); 5809 int32_t offset = cls->GetDexCacheElementOffset() - base->GetElementOffset(); 5810 // /* GcRoot<mirror::Class> */ out = *(dex_cache_arrays_base + offset) 5811 GenerateGcRootFieldLoad(cls, out_loc, base_reg, offset, requires_read_barrier); 5812 generate_null_check = !cls->IsInDexCache(); 5813 break; 5814 } 5815 case HLoadClass::LoadKind::kDexCacheViaMethod: { 5816 // /* GcRoot<mirror::Class>[] */ out = 5817 // current_method.ptr_sized_fields_->dex_cache_resolved_types_ 5818 Register current_method = locations->InAt(0).AsRegister<Register>(); 5819 __ LoadFromOffset(kLoadWord, 5820 out, 5821 current_method, 5822 ArtMethod::DexCacheResolvedTypesOffset(kArmPointerSize).Int32Value()); 5823 // /* GcRoot<mirror::Class> */ out = out[type_index] 5824 size_t offset = CodeGenerator::GetCacheOffset(cls->GetTypeIndex()); 5825 GenerateGcRootFieldLoad(cls, out_loc, out, offset, requires_read_barrier); 5826 generate_null_check = !cls->IsInDexCache(); 5827 } 5828 } 5829 5830 if (generate_null_check || cls->MustGenerateClinitCheck()) { 5831 DCHECK(cls->CanCallRuntime()); 5832 SlowPathCodeARM* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathARM( 5833 cls, cls, cls->GetDexPc(), cls->MustGenerateClinitCheck()); 5834 codegen_->AddSlowPath(slow_path); 5835 if (generate_null_check) { 5836 __ CompareAndBranchIfZero(out, slow_path->GetEntryLabel()); 5837 } 5838 if (cls->MustGenerateClinitCheck()) { 5839 GenerateClassInitializationCheck(slow_path, out); 5840 } else { 5841 __ Bind(slow_path->GetExitLabel()); 5842 } 5843 } 5844} 5845 5846void LocationsBuilderARM::VisitClinitCheck(HClinitCheck* check) { 5847 LocationSummary* locations = 5848 new (GetGraph()->GetArena()) LocationSummary(check, LocationSummary::kCallOnSlowPath); 5849 locations->SetInAt(0, Location::RequiresRegister()); 5850 if (check->HasUses()) { 5851 locations->SetOut(Location::SameAsFirstInput()); 5852 } 5853} 5854 5855void InstructionCodeGeneratorARM::VisitClinitCheck(HClinitCheck* check) { 5856 // We assume the class is not null. 5857 SlowPathCodeARM* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathARM( 5858 check->GetLoadClass(), check, check->GetDexPc(), true); 5859 codegen_->AddSlowPath(slow_path); 5860 GenerateClassInitializationCheck(slow_path, 5861 check->GetLocations()->InAt(0).AsRegister<Register>()); 5862} 5863 5864void InstructionCodeGeneratorARM::GenerateClassInitializationCheck( 5865 SlowPathCodeARM* slow_path, Register class_reg) { 5866 __ LoadFromOffset(kLoadWord, IP, class_reg, mirror::Class::StatusOffset().Int32Value()); 5867 __ cmp(IP, ShifterOperand(mirror::Class::kStatusInitialized)); 5868 __ b(slow_path->GetEntryLabel(), LT); 5869 // Even if the initialized flag is set, we may be in a situation where caches are not synced 5870 // properly. Therefore, we do a memory fence. 5871 __ dmb(ISH); 5872 __ Bind(slow_path->GetExitLabel()); 5873} 5874 5875HLoadString::LoadKind CodeGeneratorARM::GetSupportedLoadStringKind( 5876 HLoadString::LoadKind desired_string_load_kind) { 5877 switch (desired_string_load_kind) { 5878 case HLoadString::LoadKind::kBootImageLinkTimeAddress: 5879 DCHECK(!GetCompilerOptions().GetCompilePic()); 5880 break; 5881 case HLoadString::LoadKind::kBootImageLinkTimePcRelative: 5882 DCHECK(GetCompilerOptions().GetCompilePic()); 5883 break; 5884 case HLoadString::LoadKind::kBootImageAddress: 5885 break; 5886 case HLoadString::LoadKind::kDexCacheAddress: 5887 DCHECK(Runtime::Current()->UseJitCompilation()); 5888 break; 5889 case HLoadString::LoadKind::kBssEntry: 5890 DCHECK(!Runtime::Current()->UseJitCompilation()); 5891 break; 5892 case HLoadString::LoadKind::kJitTableAddress: 5893 DCHECK(Runtime::Current()->UseJitCompilation()); 5894 return HLoadString::LoadKind::kDexCacheViaMethod; 5895 case HLoadString::LoadKind::kDexCacheViaMethod: 5896 break; 5897 } 5898 return desired_string_load_kind; 5899} 5900 5901void LocationsBuilderARM::VisitLoadString(HLoadString* load) { 5902 LocationSummary::CallKind call_kind = load->NeedsEnvironment() 5903 ? ((load->GetLoadKind() == HLoadString::LoadKind::kDexCacheViaMethod) 5904 ? LocationSummary::kCallOnMainOnly 5905 : LocationSummary::kCallOnSlowPath) 5906 : LocationSummary::kNoCall; 5907 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(load, call_kind); 5908 5909 HLoadString::LoadKind load_kind = load->GetLoadKind(); 5910 if (load_kind == HLoadString::LoadKind::kDexCacheViaMethod) { 5911 locations->SetOut(Location::RegisterLocation(R0)); 5912 } else { 5913 locations->SetOut(Location::RequiresRegister()); 5914 if (load_kind == HLoadString::LoadKind::kBssEntry) { 5915 if (!kUseReadBarrier || kUseBakerReadBarrier) { 5916 // Rely on the pResolveString and/or marking to save everything, including temps. 5917 // Note that IP may theoretically be clobbered by saving/restoring the live register 5918 // (only one thanks to the custom calling convention), so we request a different temp. 5919 locations->AddTemp(Location::RequiresRegister()); 5920 RegisterSet caller_saves = RegisterSet::Empty(); 5921 InvokeRuntimeCallingConvention calling_convention; 5922 caller_saves.Add(Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 5923 // TODO: Add GetReturnLocation() to the calling convention so that we can DCHECK() 5924 // that the the kPrimNot result register is the same as the first argument register. 5925 locations->SetCustomSlowPathCallerSaves(caller_saves); 5926 } else { 5927 // For non-Baker read barrier we have a temp-clobbering call. 5928 } 5929 } 5930 } 5931} 5932 5933void InstructionCodeGeneratorARM::VisitLoadString(HLoadString* load) { 5934 LocationSummary* locations = load->GetLocations(); 5935 Location out_loc = locations->Out(); 5936 Register out = out_loc.AsRegister<Register>(); 5937 HLoadString::LoadKind load_kind = load->GetLoadKind(); 5938 5939 switch (load_kind) { 5940 case HLoadString::LoadKind::kBootImageLinkTimeAddress: { 5941 __ LoadLiteral(out, codegen_->DeduplicateBootImageStringLiteral(load->GetDexFile(), 5942 load->GetStringIndex())); 5943 return; // No dex cache slow path. 5944 } 5945 case HLoadString::LoadKind::kBootImageLinkTimePcRelative: { 5946 DCHECK(codegen_->GetCompilerOptions().IsBootImage()); 5947 CodeGeneratorARM::PcRelativePatchInfo* labels = 5948 codegen_->NewPcRelativeStringPatch(load->GetDexFile(), load->GetStringIndex()); 5949 __ BindTrackedLabel(&labels->movw_label); 5950 __ movw(out, /* placeholder */ 0u); 5951 __ BindTrackedLabel(&labels->movt_label); 5952 __ movt(out, /* placeholder */ 0u); 5953 __ BindTrackedLabel(&labels->add_pc_label); 5954 __ add(out, out, ShifterOperand(PC)); 5955 return; // No dex cache slow path. 5956 } 5957 case HLoadString::LoadKind::kBootImageAddress: { 5958 DCHECK_NE(load->GetAddress(), 0u); 5959 uint32_t address = dchecked_integral_cast<uint32_t>(load->GetAddress()); 5960 __ LoadLiteral(out, codegen_->DeduplicateBootImageAddressLiteral(address)); 5961 return; // No dex cache slow path. 5962 } 5963 case HLoadString::LoadKind::kBssEntry: { 5964 DCHECK(!codegen_->GetCompilerOptions().IsBootImage()); 5965 Register temp = locations->GetTemp(0).AsRegister<Register>(); 5966 CodeGeneratorARM::PcRelativePatchInfo* labels = 5967 codegen_->NewPcRelativeStringPatch(load->GetDexFile(), load->GetStringIndex()); 5968 __ BindTrackedLabel(&labels->movw_label); 5969 __ movw(temp, /* placeholder */ 0u); 5970 __ BindTrackedLabel(&labels->movt_label); 5971 __ movt(temp, /* placeholder */ 0u); 5972 __ BindTrackedLabel(&labels->add_pc_label); 5973 __ add(temp, temp, ShifterOperand(PC)); 5974 GenerateGcRootFieldLoad(load, out_loc, temp, /* offset */ 0, kEmitCompilerReadBarrier); 5975 SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadStringSlowPathARM(load); 5976 codegen_->AddSlowPath(slow_path); 5977 __ CompareAndBranchIfZero(out, slow_path->GetEntryLabel()); 5978 __ Bind(slow_path->GetExitLabel()); 5979 return; 5980 } 5981 default: 5982 break; 5983 } 5984 5985 // TODO: Consider re-adding the compiler code to do string dex cache lookup again. 5986 DCHECK(load_kind == HLoadString::LoadKind::kDexCacheViaMethod); 5987 InvokeRuntimeCallingConvention calling_convention; 5988 DCHECK_EQ(calling_convention.GetRegisterAt(0), out); 5989 __ LoadImmediate(calling_convention.GetRegisterAt(0), load->GetStringIndex()); 5990 codegen_->InvokeRuntime(kQuickResolveString, load, load->GetDexPc()); 5991 CheckEntrypointTypes<kQuickResolveString, void*, uint32_t>(); 5992} 5993 5994static int32_t GetExceptionTlsOffset() { 5995 return Thread::ExceptionOffset<kArmPointerSize>().Int32Value(); 5996} 5997 5998void LocationsBuilderARM::VisitLoadException(HLoadException* load) { 5999 LocationSummary* locations = 6000 new (GetGraph()->GetArena()) LocationSummary(load, LocationSummary::kNoCall); 6001 locations->SetOut(Location::RequiresRegister()); 6002} 6003 6004void InstructionCodeGeneratorARM::VisitLoadException(HLoadException* load) { 6005 Register out = load->GetLocations()->Out().AsRegister<Register>(); 6006 __ LoadFromOffset(kLoadWord, out, TR, GetExceptionTlsOffset()); 6007} 6008 6009void LocationsBuilderARM::VisitClearException(HClearException* clear) { 6010 new (GetGraph()->GetArena()) LocationSummary(clear, LocationSummary::kNoCall); 6011} 6012 6013void InstructionCodeGeneratorARM::VisitClearException(HClearException* clear ATTRIBUTE_UNUSED) { 6014 __ LoadImmediate(IP, 0); 6015 __ StoreToOffset(kStoreWord, IP, TR, GetExceptionTlsOffset()); 6016} 6017 6018void LocationsBuilderARM::VisitThrow(HThrow* instruction) { 6019 LocationSummary* locations = 6020 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnMainOnly); 6021 InvokeRuntimeCallingConvention calling_convention; 6022 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 6023} 6024 6025void InstructionCodeGeneratorARM::VisitThrow(HThrow* instruction) { 6026 codegen_->InvokeRuntime(kQuickDeliverException, instruction, instruction->GetDexPc()); 6027 CheckEntrypointTypes<kQuickDeliverException, void, mirror::Object*>(); 6028} 6029 6030static bool TypeCheckNeedsATemporary(TypeCheckKind type_check_kind) { 6031 return kEmitCompilerReadBarrier && 6032 (kUseBakerReadBarrier || 6033 type_check_kind == TypeCheckKind::kAbstractClassCheck || 6034 type_check_kind == TypeCheckKind::kClassHierarchyCheck || 6035 type_check_kind == TypeCheckKind::kArrayObjectCheck); 6036} 6037 6038void LocationsBuilderARM::VisitInstanceOf(HInstanceOf* instruction) { 6039 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 6040 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 6041 bool baker_read_barrier_slow_path = false; 6042 switch (type_check_kind) { 6043 case TypeCheckKind::kExactCheck: 6044 case TypeCheckKind::kAbstractClassCheck: 6045 case TypeCheckKind::kClassHierarchyCheck: 6046 case TypeCheckKind::kArrayObjectCheck: 6047 call_kind = 6048 kEmitCompilerReadBarrier ? LocationSummary::kCallOnSlowPath : LocationSummary::kNoCall; 6049 baker_read_barrier_slow_path = kUseBakerReadBarrier; 6050 break; 6051 case TypeCheckKind::kArrayCheck: 6052 case TypeCheckKind::kUnresolvedCheck: 6053 case TypeCheckKind::kInterfaceCheck: 6054 call_kind = LocationSummary::kCallOnSlowPath; 6055 break; 6056 } 6057 6058 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 6059 if (baker_read_barrier_slow_path) { 6060 locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty()); // No caller-save registers. 6061 } 6062 locations->SetInAt(0, Location::RequiresRegister()); 6063 locations->SetInAt(1, Location::RequiresRegister()); 6064 // The "out" register is used as a temporary, so it overlaps with the inputs. 6065 // Note that TypeCheckSlowPathARM uses this register too. 6066 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 6067 // When read barriers are enabled, we need a temporary register for 6068 // some cases. 6069 if (TypeCheckNeedsATemporary(type_check_kind)) { 6070 locations->AddTemp(Location::RequiresRegister()); 6071 } 6072} 6073 6074void InstructionCodeGeneratorARM::VisitInstanceOf(HInstanceOf* instruction) { 6075 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 6076 LocationSummary* locations = instruction->GetLocations(); 6077 Location obj_loc = locations->InAt(0); 6078 Register obj = obj_loc.AsRegister<Register>(); 6079 Register cls = locations->InAt(1).AsRegister<Register>(); 6080 Location out_loc = locations->Out(); 6081 Register out = out_loc.AsRegister<Register>(); 6082 Location maybe_temp_loc = TypeCheckNeedsATemporary(type_check_kind) ? 6083 locations->GetTemp(0) : 6084 Location::NoLocation(); 6085 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 6086 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 6087 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 6088 uint32_t primitive_offset = mirror::Class::PrimitiveTypeOffset().Int32Value(); 6089 Label done, zero; 6090 SlowPathCodeARM* slow_path = nullptr; 6091 6092 // Return 0 if `obj` is null. 6093 // avoid null check if we know obj is not null. 6094 if (instruction->MustDoNullCheck()) { 6095 __ CompareAndBranchIfZero(obj, &zero); 6096 } 6097 6098 // /* HeapReference<Class> */ out = obj->klass_ 6099 GenerateReferenceLoadTwoRegisters(instruction, out_loc, obj_loc, class_offset, maybe_temp_loc); 6100 6101 switch (type_check_kind) { 6102 case TypeCheckKind::kExactCheck: { 6103 __ cmp(out, ShifterOperand(cls)); 6104 // Classes must be equal for the instanceof to succeed. 6105 __ b(&zero, NE); 6106 __ LoadImmediate(out, 1); 6107 __ b(&done); 6108 break; 6109 } 6110 6111 case TypeCheckKind::kAbstractClassCheck: { 6112 // If the class is abstract, we eagerly fetch the super class of the 6113 // object to avoid doing a comparison we know will fail. 6114 Label loop; 6115 __ Bind(&loop); 6116 // /* HeapReference<Class> */ out = out->super_class_ 6117 GenerateReferenceLoadOneRegister(instruction, out_loc, super_offset, maybe_temp_loc); 6118 // If `out` is null, we use it for the result, and jump to `done`. 6119 __ CompareAndBranchIfZero(out, &done); 6120 __ cmp(out, ShifterOperand(cls)); 6121 __ b(&loop, NE); 6122 __ LoadImmediate(out, 1); 6123 if (zero.IsLinked()) { 6124 __ b(&done); 6125 } 6126 break; 6127 } 6128 6129 case TypeCheckKind::kClassHierarchyCheck: { 6130 // Walk over the class hierarchy to find a match. 6131 Label loop, success; 6132 __ Bind(&loop); 6133 __ cmp(out, ShifterOperand(cls)); 6134 __ b(&success, EQ); 6135 // /* HeapReference<Class> */ out = out->super_class_ 6136 GenerateReferenceLoadOneRegister(instruction, out_loc, super_offset, maybe_temp_loc); 6137 __ CompareAndBranchIfNonZero(out, &loop); 6138 // If `out` is null, we use it for the result, and jump to `done`. 6139 __ b(&done); 6140 __ Bind(&success); 6141 __ LoadImmediate(out, 1); 6142 if (zero.IsLinked()) { 6143 __ b(&done); 6144 } 6145 break; 6146 } 6147 6148 case TypeCheckKind::kArrayObjectCheck: { 6149 // Do an exact check. 6150 Label exact_check; 6151 __ cmp(out, ShifterOperand(cls)); 6152 __ b(&exact_check, EQ); 6153 // Otherwise, we need to check that the object's class is a non-primitive array. 6154 // /* HeapReference<Class> */ out = out->component_type_ 6155 GenerateReferenceLoadOneRegister(instruction, out_loc, component_offset, maybe_temp_loc); 6156 // If `out` is null, we use it for the result, and jump to `done`. 6157 __ CompareAndBranchIfZero(out, &done); 6158 __ LoadFromOffset(kLoadUnsignedHalfword, out, out, primitive_offset); 6159 static_assert(Primitive::kPrimNot == 0, "Expected 0 for kPrimNot"); 6160 __ CompareAndBranchIfNonZero(out, &zero); 6161 __ Bind(&exact_check); 6162 __ LoadImmediate(out, 1); 6163 __ b(&done); 6164 break; 6165 } 6166 6167 case TypeCheckKind::kArrayCheck: { 6168 __ cmp(out, ShifterOperand(cls)); 6169 DCHECK(locations->OnlyCallsOnSlowPath()); 6170 slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathARM(instruction, 6171 /* is_fatal */ false); 6172 codegen_->AddSlowPath(slow_path); 6173 __ b(slow_path->GetEntryLabel(), NE); 6174 __ LoadImmediate(out, 1); 6175 if (zero.IsLinked()) { 6176 __ b(&done); 6177 } 6178 break; 6179 } 6180 6181 case TypeCheckKind::kUnresolvedCheck: 6182 case TypeCheckKind::kInterfaceCheck: { 6183 // Note that we indeed only call on slow path, but we always go 6184 // into the slow path for the unresolved and interface check 6185 // cases. 6186 // 6187 // We cannot directly call the InstanceofNonTrivial runtime 6188 // entry point without resorting to a type checking slow path 6189 // here (i.e. by calling InvokeRuntime directly), as it would 6190 // require to assign fixed registers for the inputs of this 6191 // HInstanceOf instruction (following the runtime calling 6192 // convention), which might be cluttered by the potential first 6193 // read barrier emission at the beginning of this method. 6194 // 6195 // TODO: Introduce a new runtime entry point taking the object 6196 // to test (instead of its class) as argument, and let it deal 6197 // with the read barrier issues. This will let us refactor this 6198 // case of the `switch` code as it was previously (with a direct 6199 // call to the runtime not using a type checking slow path). 6200 // This should also be beneficial for the other cases above. 6201 DCHECK(locations->OnlyCallsOnSlowPath()); 6202 slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathARM(instruction, 6203 /* is_fatal */ false); 6204 codegen_->AddSlowPath(slow_path); 6205 __ b(slow_path->GetEntryLabel()); 6206 if (zero.IsLinked()) { 6207 __ b(&done); 6208 } 6209 break; 6210 } 6211 } 6212 6213 if (zero.IsLinked()) { 6214 __ Bind(&zero); 6215 __ LoadImmediate(out, 0); 6216 } 6217 6218 if (done.IsLinked()) { 6219 __ Bind(&done); 6220 } 6221 6222 if (slow_path != nullptr) { 6223 __ Bind(slow_path->GetExitLabel()); 6224 } 6225} 6226 6227void LocationsBuilderARM::VisitCheckCast(HCheckCast* instruction) { 6228 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 6229 bool throws_into_catch = instruction->CanThrowIntoCatchBlock(); 6230 6231 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 6232 switch (type_check_kind) { 6233 case TypeCheckKind::kExactCheck: 6234 case TypeCheckKind::kAbstractClassCheck: 6235 case TypeCheckKind::kClassHierarchyCheck: 6236 case TypeCheckKind::kArrayObjectCheck: 6237 call_kind = (throws_into_catch || kEmitCompilerReadBarrier) ? 6238 LocationSummary::kCallOnSlowPath : 6239 LocationSummary::kNoCall; // In fact, call on a fatal (non-returning) slow path. 6240 break; 6241 case TypeCheckKind::kArrayCheck: 6242 case TypeCheckKind::kUnresolvedCheck: 6243 case TypeCheckKind::kInterfaceCheck: 6244 call_kind = LocationSummary::kCallOnSlowPath; 6245 break; 6246 } 6247 6248 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 6249 locations->SetInAt(0, Location::RequiresRegister()); 6250 locations->SetInAt(1, Location::RequiresRegister()); 6251 // Note that TypeCheckSlowPathARM uses this "temp" register too. 6252 locations->AddTemp(Location::RequiresRegister()); 6253 // When read barriers are enabled, we need an additional temporary 6254 // register for some cases. 6255 if (TypeCheckNeedsATemporary(type_check_kind)) { 6256 locations->AddTemp(Location::RequiresRegister()); 6257 } 6258} 6259 6260void InstructionCodeGeneratorARM::VisitCheckCast(HCheckCast* instruction) { 6261 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 6262 LocationSummary* locations = instruction->GetLocations(); 6263 Location obj_loc = locations->InAt(0); 6264 Register obj = obj_loc.AsRegister<Register>(); 6265 Register cls = locations->InAt(1).AsRegister<Register>(); 6266 Location temp_loc = locations->GetTemp(0); 6267 Register temp = temp_loc.AsRegister<Register>(); 6268 Location maybe_temp2_loc = TypeCheckNeedsATemporary(type_check_kind) ? 6269 locations->GetTemp(1) : 6270 Location::NoLocation(); 6271 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 6272 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 6273 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 6274 uint32_t primitive_offset = mirror::Class::PrimitiveTypeOffset().Int32Value(); 6275 6276 bool is_type_check_slow_path_fatal = 6277 (type_check_kind == TypeCheckKind::kExactCheck || 6278 type_check_kind == TypeCheckKind::kAbstractClassCheck || 6279 type_check_kind == TypeCheckKind::kClassHierarchyCheck || 6280 type_check_kind == TypeCheckKind::kArrayObjectCheck) && 6281 !instruction->CanThrowIntoCatchBlock(); 6282 SlowPathCodeARM* type_check_slow_path = 6283 new (GetGraph()->GetArena()) TypeCheckSlowPathARM(instruction, 6284 is_type_check_slow_path_fatal); 6285 codegen_->AddSlowPath(type_check_slow_path); 6286 6287 Label done; 6288 // Avoid null check if we know obj is not null. 6289 if (instruction->MustDoNullCheck()) { 6290 __ CompareAndBranchIfZero(obj, &done); 6291 } 6292 6293 // /* HeapReference<Class> */ temp = obj->klass_ 6294 GenerateReferenceLoadTwoRegisters(instruction, temp_loc, obj_loc, class_offset, maybe_temp2_loc); 6295 6296 switch (type_check_kind) { 6297 case TypeCheckKind::kExactCheck: 6298 case TypeCheckKind::kArrayCheck: { 6299 __ cmp(temp, ShifterOperand(cls)); 6300 // Jump to slow path for throwing the exception or doing a 6301 // more involved array check. 6302 __ b(type_check_slow_path->GetEntryLabel(), NE); 6303 break; 6304 } 6305 6306 case TypeCheckKind::kAbstractClassCheck: { 6307 // If the class is abstract, we eagerly fetch the super class of the 6308 // object to avoid doing a comparison we know will fail. 6309 Label loop; 6310 __ Bind(&loop); 6311 // /* HeapReference<Class> */ temp = temp->super_class_ 6312 GenerateReferenceLoadOneRegister(instruction, temp_loc, super_offset, maybe_temp2_loc); 6313 6314 // If the class reference currently in `temp` is null, jump to the slow path to throw the 6315 // exception. 6316 __ CompareAndBranchIfZero(temp, type_check_slow_path->GetEntryLabel()); 6317 6318 // Otherwise, compare the classes. 6319 __ cmp(temp, ShifterOperand(cls)); 6320 __ b(&loop, NE); 6321 break; 6322 } 6323 6324 case TypeCheckKind::kClassHierarchyCheck: { 6325 // Walk over the class hierarchy to find a match. 6326 Label loop; 6327 __ Bind(&loop); 6328 __ cmp(temp, ShifterOperand(cls)); 6329 __ b(&done, EQ); 6330 6331 // /* HeapReference<Class> */ temp = temp->super_class_ 6332 GenerateReferenceLoadOneRegister(instruction, temp_loc, super_offset, maybe_temp2_loc); 6333 6334 // If the class reference currently in `temp` is null, jump to the slow path to throw the 6335 // exception. 6336 __ CompareAndBranchIfZero(temp, type_check_slow_path->GetEntryLabel()); 6337 // Otherwise, jump to the beginning of the loop. 6338 __ b(&loop); 6339 break; 6340 } 6341 6342 case TypeCheckKind::kArrayObjectCheck: { 6343 // Do an exact check. 6344 __ cmp(temp, ShifterOperand(cls)); 6345 __ b(&done, EQ); 6346 6347 // Otherwise, we need to check that the object's class is a non-primitive array. 6348 // /* HeapReference<Class> */ temp = temp->component_type_ 6349 GenerateReferenceLoadOneRegister(instruction, temp_loc, component_offset, maybe_temp2_loc); 6350 // If the component type is null, jump to the slow path to throw the exception. 6351 __ CompareAndBranchIfZero(temp, type_check_slow_path->GetEntryLabel()); 6352 // Otherwise,the object is indeed an array, jump to label `check_non_primitive_component_type` 6353 // to further check that this component type is not a primitive type. 6354 __ LoadFromOffset(kLoadUnsignedHalfword, temp, temp, primitive_offset); 6355 static_assert(Primitive::kPrimNot == 0, "Expected 0 for art::Primitive::kPrimNot"); 6356 __ CompareAndBranchIfNonZero(temp, type_check_slow_path->GetEntryLabel()); 6357 break; 6358 } 6359 6360 case TypeCheckKind::kUnresolvedCheck: 6361 case TypeCheckKind::kInterfaceCheck: 6362 // We always go into the type check slow path for the unresolved 6363 // and interface check cases. 6364 // 6365 // We cannot directly call the CheckCast runtime entry point 6366 // without resorting to a type checking slow path here (i.e. by 6367 // calling InvokeRuntime directly), as it would require to 6368 // assign fixed registers for the inputs of this HInstanceOf 6369 // instruction (following the runtime calling convention), which 6370 // might be cluttered by the potential first read barrier 6371 // emission at the beginning of this method. 6372 __ b(type_check_slow_path->GetEntryLabel()); 6373 break; 6374 } 6375 __ Bind(&done); 6376 6377 __ Bind(type_check_slow_path->GetExitLabel()); 6378} 6379 6380void LocationsBuilderARM::VisitMonitorOperation(HMonitorOperation* instruction) { 6381 LocationSummary* locations = 6382 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnMainOnly); 6383 InvokeRuntimeCallingConvention calling_convention; 6384 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 6385} 6386 6387void InstructionCodeGeneratorARM::VisitMonitorOperation(HMonitorOperation* instruction) { 6388 codegen_->InvokeRuntime(instruction->IsEnter() ? kQuickLockObject : kQuickUnlockObject, 6389 instruction, 6390 instruction->GetDexPc()); 6391 if (instruction->IsEnter()) { 6392 CheckEntrypointTypes<kQuickLockObject, void, mirror::Object*>(); 6393 } else { 6394 CheckEntrypointTypes<kQuickUnlockObject, void, mirror::Object*>(); 6395 } 6396} 6397 6398void LocationsBuilderARM::VisitAnd(HAnd* instruction) { HandleBitwiseOperation(instruction, AND); } 6399void LocationsBuilderARM::VisitOr(HOr* instruction) { HandleBitwiseOperation(instruction, ORR); } 6400void LocationsBuilderARM::VisitXor(HXor* instruction) { HandleBitwiseOperation(instruction, EOR); } 6401 6402void LocationsBuilderARM::HandleBitwiseOperation(HBinaryOperation* instruction, Opcode opcode) { 6403 LocationSummary* locations = 6404 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 6405 DCHECK(instruction->GetResultType() == Primitive::kPrimInt 6406 || instruction->GetResultType() == Primitive::kPrimLong); 6407 // Note: GVN reorders commutative operations to have the constant on the right hand side. 6408 locations->SetInAt(0, Location::RequiresRegister()); 6409 locations->SetInAt(1, ArmEncodableConstantOrRegister(instruction->InputAt(1), opcode)); 6410 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 6411} 6412 6413void InstructionCodeGeneratorARM::VisitAnd(HAnd* instruction) { 6414 HandleBitwiseOperation(instruction); 6415} 6416 6417void InstructionCodeGeneratorARM::VisitOr(HOr* instruction) { 6418 HandleBitwiseOperation(instruction); 6419} 6420 6421void InstructionCodeGeneratorARM::VisitXor(HXor* instruction) { 6422 HandleBitwiseOperation(instruction); 6423} 6424 6425 6426void LocationsBuilderARM::VisitBitwiseNegatedRight(HBitwiseNegatedRight* instruction) { 6427 LocationSummary* locations = 6428 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 6429 DCHECK(instruction->GetResultType() == Primitive::kPrimInt 6430 || instruction->GetResultType() == Primitive::kPrimLong); 6431 6432 locations->SetInAt(0, Location::RequiresRegister()); 6433 locations->SetInAt(1, Location::RequiresRegister()); 6434 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 6435} 6436 6437void InstructionCodeGeneratorARM::VisitBitwiseNegatedRight(HBitwiseNegatedRight* instruction) { 6438 LocationSummary* locations = instruction->GetLocations(); 6439 Location first = locations->InAt(0); 6440 Location second = locations->InAt(1); 6441 Location out = locations->Out(); 6442 6443 if (instruction->GetResultType() == Primitive::kPrimInt) { 6444 Register first_reg = first.AsRegister<Register>(); 6445 ShifterOperand second_reg(second.AsRegister<Register>()); 6446 Register out_reg = out.AsRegister<Register>(); 6447 6448 switch (instruction->GetOpKind()) { 6449 case HInstruction::kAnd: 6450 __ bic(out_reg, first_reg, second_reg); 6451 break; 6452 case HInstruction::kOr: 6453 __ orn(out_reg, first_reg, second_reg); 6454 break; 6455 // There is no EON on arm. 6456 case HInstruction::kXor: 6457 default: 6458 LOG(FATAL) << "Unexpected instruction " << instruction->DebugName(); 6459 UNREACHABLE(); 6460 } 6461 return; 6462 6463 } else { 6464 DCHECK_EQ(instruction->GetResultType(), Primitive::kPrimLong); 6465 Register first_low = first.AsRegisterPairLow<Register>(); 6466 Register first_high = first.AsRegisterPairHigh<Register>(); 6467 ShifterOperand second_low(second.AsRegisterPairLow<Register>()); 6468 ShifterOperand second_high(second.AsRegisterPairHigh<Register>()); 6469 Register out_low = out.AsRegisterPairLow<Register>(); 6470 Register out_high = out.AsRegisterPairHigh<Register>(); 6471 6472 switch (instruction->GetOpKind()) { 6473 case HInstruction::kAnd: 6474 __ bic(out_low, first_low, second_low); 6475 __ bic(out_high, first_high, second_high); 6476 break; 6477 case HInstruction::kOr: 6478 __ orn(out_low, first_low, second_low); 6479 __ orn(out_high, first_high, second_high); 6480 break; 6481 // There is no EON on arm. 6482 case HInstruction::kXor: 6483 default: 6484 LOG(FATAL) << "Unexpected instruction " << instruction->DebugName(); 6485 UNREACHABLE(); 6486 } 6487 } 6488} 6489 6490void InstructionCodeGeneratorARM::GenerateAndConst(Register out, Register first, uint32_t value) { 6491 // Optimize special cases for individual halfs of `and-long` (`and` is simplified earlier). 6492 if (value == 0xffffffffu) { 6493 if (out != first) { 6494 __ mov(out, ShifterOperand(first)); 6495 } 6496 return; 6497 } 6498 if (value == 0u) { 6499 __ mov(out, ShifterOperand(0)); 6500 return; 6501 } 6502 ShifterOperand so; 6503 if (__ ShifterOperandCanHold(kNoRegister, kNoRegister, AND, value, &so)) { 6504 __ and_(out, first, so); 6505 } else { 6506 DCHECK(__ ShifterOperandCanHold(kNoRegister, kNoRegister, BIC, ~value, &so)); 6507 __ bic(out, first, ShifterOperand(~value)); 6508 } 6509} 6510 6511void InstructionCodeGeneratorARM::GenerateOrrConst(Register out, Register first, uint32_t value) { 6512 // Optimize special cases for individual halfs of `or-long` (`or` is simplified earlier). 6513 if (value == 0u) { 6514 if (out != first) { 6515 __ mov(out, ShifterOperand(first)); 6516 } 6517 return; 6518 } 6519 if (value == 0xffffffffu) { 6520 __ mvn(out, ShifterOperand(0)); 6521 return; 6522 } 6523 ShifterOperand so; 6524 if (__ ShifterOperandCanHold(kNoRegister, kNoRegister, ORR, value, &so)) { 6525 __ orr(out, first, so); 6526 } else { 6527 DCHECK(__ ShifterOperandCanHold(kNoRegister, kNoRegister, ORN, ~value, &so)); 6528 __ orn(out, first, ShifterOperand(~value)); 6529 } 6530} 6531 6532void InstructionCodeGeneratorARM::GenerateEorConst(Register out, Register first, uint32_t value) { 6533 // Optimize special case for individual halfs of `xor-long` (`xor` is simplified earlier). 6534 if (value == 0u) { 6535 if (out != first) { 6536 __ mov(out, ShifterOperand(first)); 6537 } 6538 return; 6539 } 6540 __ eor(out, first, ShifterOperand(value)); 6541} 6542 6543void InstructionCodeGeneratorARM::GenerateAddLongConst(Location out, 6544 Location first, 6545 uint64_t value) { 6546 Register out_low = out.AsRegisterPairLow<Register>(); 6547 Register out_high = out.AsRegisterPairHigh<Register>(); 6548 Register first_low = first.AsRegisterPairLow<Register>(); 6549 Register first_high = first.AsRegisterPairHigh<Register>(); 6550 uint32_t value_low = Low32Bits(value); 6551 uint32_t value_high = High32Bits(value); 6552 if (value_low == 0u) { 6553 if (out_low != first_low) { 6554 __ mov(out_low, ShifterOperand(first_low)); 6555 } 6556 __ AddConstant(out_high, first_high, value_high); 6557 return; 6558 } 6559 __ AddConstantSetFlags(out_low, first_low, value_low); 6560 ShifterOperand so; 6561 if (__ ShifterOperandCanHold(out_high, first_high, ADC, value_high, kCcDontCare, &so)) { 6562 __ adc(out_high, first_high, so); 6563 } else if (__ ShifterOperandCanHold(out_low, first_low, SBC, ~value_high, kCcDontCare, &so)) { 6564 __ sbc(out_high, first_high, so); 6565 } else { 6566 LOG(FATAL) << "Unexpected constant " << value_high; 6567 UNREACHABLE(); 6568 } 6569} 6570 6571void InstructionCodeGeneratorARM::HandleBitwiseOperation(HBinaryOperation* instruction) { 6572 LocationSummary* locations = instruction->GetLocations(); 6573 Location first = locations->InAt(0); 6574 Location second = locations->InAt(1); 6575 Location out = locations->Out(); 6576 6577 if (second.IsConstant()) { 6578 uint64_t value = static_cast<uint64_t>(Int64FromConstant(second.GetConstant())); 6579 uint32_t value_low = Low32Bits(value); 6580 if (instruction->GetResultType() == Primitive::kPrimInt) { 6581 Register first_reg = first.AsRegister<Register>(); 6582 Register out_reg = out.AsRegister<Register>(); 6583 if (instruction->IsAnd()) { 6584 GenerateAndConst(out_reg, first_reg, value_low); 6585 } else if (instruction->IsOr()) { 6586 GenerateOrrConst(out_reg, first_reg, value_low); 6587 } else { 6588 DCHECK(instruction->IsXor()); 6589 GenerateEorConst(out_reg, first_reg, value_low); 6590 } 6591 } else { 6592 DCHECK_EQ(instruction->GetResultType(), Primitive::kPrimLong); 6593 uint32_t value_high = High32Bits(value); 6594 Register first_low = first.AsRegisterPairLow<Register>(); 6595 Register first_high = first.AsRegisterPairHigh<Register>(); 6596 Register out_low = out.AsRegisterPairLow<Register>(); 6597 Register out_high = out.AsRegisterPairHigh<Register>(); 6598 if (instruction->IsAnd()) { 6599 GenerateAndConst(out_low, first_low, value_low); 6600 GenerateAndConst(out_high, first_high, value_high); 6601 } else if (instruction->IsOr()) { 6602 GenerateOrrConst(out_low, first_low, value_low); 6603 GenerateOrrConst(out_high, first_high, value_high); 6604 } else { 6605 DCHECK(instruction->IsXor()); 6606 GenerateEorConst(out_low, first_low, value_low); 6607 GenerateEorConst(out_high, first_high, value_high); 6608 } 6609 } 6610 return; 6611 } 6612 6613 if (instruction->GetResultType() == Primitive::kPrimInt) { 6614 Register first_reg = first.AsRegister<Register>(); 6615 ShifterOperand second_reg(second.AsRegister<Register>()); 6616 Register out_reg = out.AsRegister<Register>(); 6617 if (instruction->IsAnd()) { 6618 __ and_(out_reg, first_reg, second_reg); 6619 } else if (instruction->IsOr()) { 6620 __ orr(out_reg, first_reg, second_reg); 6621 } else { 6622 DCHECK(instruction->IsXor()); 6623 __ eor(out_reg, first_reg, second_reg); 6624 } 6625 } else { 6626 DCHECK_EQ(instruction->GetResultType(), Primitive::kPrimLong); 6627 Register first_low = first.AsRegisterPairLow<Register>(); 6628 Register first_high = first.AsRegisterPairHigh<Register>(); 6629 ShifterOperand second_low(second.AsRegisterPairLow<Register>()); 6630 ShifterOperand second_high(second.AsRegisterPairHigh<Register>()); 6631 Register out_low = out.AsRegisterPairLow<Register>(); 6632 Register out_high = out.AsRegisterPairHigh<Register>(); 6633 if (instruction->IsAnd()) { 6634 __ and_(out_low, first_low, second_low); 6635 __ and_(out_high, first_high, second_high); 6636 } else if (instruction->IsOr()) { 6637 __ orr(out_low, first_low, second_low); 6638 __ orr(out_high, first_high, second_high); 6639 } else { 6640 DCHECK(instruction->IsXor()); 6641 __ eor(out_low, first_low, second_low); 6642 __ eor(out_high, first_high, second_high); 6643 } 6644 } 6645} 6646 6647void InstructionCodeGeneratorARM::GenerateReferenceLoadOneRegister(HInstruction* instruction, 6648 Location out, 6649 uint32_t offset, 6650 Location maybe_temp) { 6651 Register out_reg = out.AsRegister<Register>(); 6652 if (kEmitCompilerReadBarrier) { 6653 DCHECK(maybe_temp.IsRegister()) << maybe_temp; 6654 if (kUseBakerReadBarrier) { 6655 // Load with fast path based Baker's read barrier. 6656 // /* HeapReference<Object> */ out = *(out + offset) 6657 codegen_->GenerateFieldLoadWithBakerReadBarrier( 6658 instruction, out, out_reg, offset, maybe_temp, /* needs_null_check */ false); 6659 } else { 6660 // Load with slow path based read barrier. 6661 // Save the value of `out` into `maybe_temp` before overwriting it 6662 // in the following move operation, as we will need it for the 6663 // read barrier below. 6664 __ Mov(maybe_temp.AsRegister<Register>(), out_reg); 6665 // /* HeapReference<Object> */ out = *(out + offset) 6666 __ LoadFromOffset(kLoadWord, out_reg, out_reg, offset); 6667 codegen_->GenerateReadBarrierSlow(instruction, out, out, maybe_temp, offset); 6668 } 6669 } else { 6670 // Plain load with no read barrier. 6671 // /* HeapReference<Object> */ out = *(out + offset) 6672 __ LoadFromOffset(kLoadWord, out_reg, out_reg, offset); 6673 __ MaybeUnpoisonHeapReference(out_reg); 6674 } 6675} 6676 6677void InstructionCodeGeneratorARM::GenerateReferenceLoadTwoRegisters(HInstruction* instruction, 6678 Location out, 6679 Location obj, 6680 uint32_t offset, 6681 Location maybe_temp) { 6682 Register out_reg = out.AsRegister<Register>(); 6683 Register obj_reg = obj.AsRegister<Register>(); 6684 if (kEmitCompilerReadBarrier) { 6685 if (kUseBakerReadBarrier) { 6686 DCHECK(maybe_temp.IsRegister()) << maybe_temp; 6687 // Load with fast path based Baker's read barrier. 6688 // /* HeapReference<Object> */ out = *(obj + offset) 6689 codegen_->GenerateFieldLoadWithBakerReadBarrier( 6690 instruction, out, obj_reg, offset, maybe_temp, /* needs_null_check */ false); 6691 } else { 6692 // Load with slow path based read barrier. 6693 // /* HeapReference<Object> */ out = *(obj + offset) 6694 __ LoadFromOffset(kLoadWord, out_reg, obj_reg, offset); 6695 codegen_->GenerateReadBarrierSlow(instruction, out, out, obj, offset); 6696 } 6697 } else { 6698 // Plain load with no read barrier. 6699 // /* HeapReference<Object> */ out = *(obj + offset) 6700 __ LoadFromOffset(kLoadWord, out_reg, obj_reg, offset); 6701 __ MaybeUnpoisonHeapReference(out_reg); 6702 } 6703} 6704 6705void InstructionCodeGeneratorARM::GenerateGcRootFieldLoad(HInstruction* instruction, 6706 Location root, 6707 Register obj, 6708 uint32_t offset, 6709 bool requires_read_barrier) { 6710 Register root_reg = root.AsRegister<Register>(); 6711 if (requires_read_barrier) { 6712 DCHECK(kEmitCompilerReadBarrier); 6713 if (kUseBakerReadBarrier) { 6714 // Fast path implementation of art::ReadBarrier::BarrierForRoot when 6715 // Baker's read barrier are used: 6716 // 6717 // root = obj.field; 6718 // if (Thread::Current()->GetIsGcMarking()) { 6719 // root = ReadBarrier::Mark(root) 6720 // } 6721 6722 // /* GcRoot<mirror::Object> */ root = *(obj + offset) 6723 __ LoadFromOffset(kLoadWord, root_reg, obj, offset); 6724 static_assert( 6725 sizeof(mirror::CompressedReference<mirror::Object>) == sizeof(GcRoot<mirror::Object>), 6726 "art::mirror::CompressedReference<mirror::Object> and art::GcRoot<mirror::Object> " 6727 "have different sizes."); 6728 static_assert(sizeof(mirror::CompressedReference<mirror::Object>) == sizeof(int32_t), 6729 "art::mirror::CompressedReference<mirror::Object> and int32_t " 6730 "have different sizes."); 6731 6732 // Slow path marking the GC root `root`. 6733 SlowPathCodeARM* slow_path = 6734 new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM(instruction, root); 6735 codegen_->AddSlowPath(slow_path); 6736 6737 // IP = Thread::Current()->GetIsGcMarking() 6738 __ LoadFromOffset( 6739 kLoadWord, IP, TR, Thread::IsGcMarkingOffset<kArmPointerSize>().Int32Value()); 6740 __ CompareAndBranchIfNonZero(IP, slow_path->GetEntryLabel()); 6741 __ Bind(slow_path->GetExitLabel()); 6742 } else { 6743 // GC root loaded through a slow path for read barriers other 6744 // than Baker's. 6745 // /* GcRoot<mirror::Object>* */ root = obj + offset 6746 __ AddConstant(root_reg, obj, offset); 6747 // /* mirror::Object* */ root = root->Read() 6748 codegen_->GenerateReadBarrierForRootSlow(instruction, root, root); 6749 } 6750 } else { 6751 // Plain GC root load with no read barrier. 6752 // /* GcRoot<mirror::Object> */ root = *(obj + offset) 6753 __ LoadFromOffset(kLoadWord, root_reg, obj, offset); 6754 // Note that GC roots are not affected by heap poisoning, thus we 6755 // do not have to unpoison `root_reg` here. 6756 } 6757} 6758 6759void CodeGeneratorARM::GenerateFieldLoadWithBakerReadBarrier(HInstruction* instruction, 6760 Location ref, 6761 Register obj, 6762 uint32_t offset, 6763 Location temp, 6764 bool needs_null_check) { 6765 DCHECK(kEmitCompilerReadBarrier); 6766 DCHECK(kUseBakerReadBarrier); 6767 6768 // /* HeapReference<Object> */ ref = *(obj + offset) 6769 Location no_index = Location::NoLocation(); 6770 ScaleFactor no_scale_factor = TIMES_1; 6771 GenerateReferenceLoadWithBakerReadBarrier( 6772 instruction, ref, obj, offset, no_index, no_scale_factor, temp, needs_null_check); 6773} 6774 6775void CodeGeneratorARM::GenerateArrayLoadWithBakerReadBarrier(HInstruction* instruction, 6776 Location ref, 6777 Register obj, 6778 uint32_t data_offset, 6779 Location index, 6780 Location temp, 6781 bool needs_null_check) { 6782 DCHECK(kEmitCompilerReadBarrier); 6783 DCHECK(kUseBakerReadBarrier); 6784 6785 static_assert( 6786 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 6787 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 6788 // /* HeapReference<Object> */ ref = 6789 // *(obj + data_offset + index * sizeof(HeapReference<Object>)) 6790 ScaleFactor scale_factor = TIMES_4; 6791 GenerateReferenceLoadWithBakerReadBarrier( 6792 instruction, ref, obj, data_offset, index, scale_factor, temp, needs_null_check); 6793} 6794 6795void CodeGeneratorARM::GenerateReferenceLoadWithBakerReadBarrier(HInstruction* instruction, 6796 Location ref, 6797 Register obj, 6798 uint32_t offset, 6799 Location index, 6800 ScaleFactor scale_factor, 6801 Location temp, 6802 bool needs_null_check, 6803 bool always_update_field, 6804 Register* temp2) { 6805 DCHECK(kEmitCompilerReadBarrier); 6806 DCHECK(kUseBakerReadBarrier); 6807 6808 // In slow path based read barriers, the read barrier call is 6809 // inserted after the original load. However, in fast path based 6810 // Baker's read barriers, we need to perform the load of 6811 // mirror::Object::monitor_ *before* the original reference load. 6812 // This load-load ordering is required by the read barrier. 6813 // The fast path/slow path (for Baker's algorithm) should look like: 6814 // 6815 // uint32_t rb_state = Lockword(obj->monitor_).ReadBarrierState(); 6816 // lfence; // Load fence or artificial data dependency to prevent load-load reordering 6817 // HeapReference<Object> ref = *src; // Original reference load. 6818 // bool is_gray = (rb_state == ReadBarrier::GrayState()); 6819 // if (is_gray) { 6820 // ref = ReadBarrier::Mark(ref); // Performed by runtime entrypoint slow path. 6821 // } 6822 // 6823 // Note: the original implementation in ReadBarrier::Barrier is 6824 // slightly more complex as it performs additional checks that we do 6825 // not do here for performance reasons. 6826 6827 Register ref_reg = ref.AsRegister<Register>(); 6828 Register temp_reg = temp.AsRegister<Register>(); 6829 uint32_t monitor_offset = mirror::Object::MonitorOffset().Int32Value(); 6830 6831 // /* int32_t */ monitor = obj->monitor_ 6832 __ LoadFromOffset(kLoadWord, temp_reg, obj, monitor_offset); 6833 if (needs_null_check) { 6834 MaybeRecordImplicitNullCheck(instruction); 6835 } 6836 // /* LockWord */ lock_word = LockWord(monitor) 6837 static_assert(sizeof(LockWord) == sizeof(int32_t), 6838 "art::LockWord and int32_t have different sizes."); 6839 6840 // Introduce a dependency on the lock_word including the rb_state, 6841 // which shall prevent load-load reordering without using 6842 // a memory barrier (which would be more expensive). 6843 // `obj` is unchanged by this operation, but its value now depends 6844 // on `temp_reg`. 6845 __ add(obj, obj, ShifterOperand(temp_reg, LSR, 32)); 6846 6847 // The actual reference load. 6848 if (index.IsValid()) { 6849 // Load types involving an "index": ArrayGet, 6850 // UnsafeGetObject/UnsafeGetObjectVolatile and UnsafeCASObject 6851 // intrinsics. 6852 // /* HeapReference<Object> */ ref = *(obj + offset + (index << scale_factor)) 6853 if (index.IsConstant()) { 6854 size_t computed_offset = 6855 (index.GetConstant()->AsIntConstant()->GetValue() << scale_factor) + offset; 6856 __ LoadFromOffset(kLoadWord, ref_reg, obj, computed_offset); 6857 } else { 6858 // Handle the special case of the 6859 // UnsafeGetObject/UnsafeGetObjectVolatile and UnsafeCASObject 6860 // intrinsics, which use a register pair as index ("long 6861 // offset"), of which only the low part contains data. 6862 Register index_reg = index.IsRegisterPair() 6863 ? index.AsRegisterPairLow<Register>() 6864 : index.AsRegister<Register>(); 6865 __ add(IP, obj, ShifterOperand(index_reg, LSL, scale_factor)); 6866 __ LoadFromOffset(kLoadWord, ref_reg, IP, offset); 6867 } 6868 } else { 6869 // /* HeapReference<Object> */ ref = *(obj + offset) 6870 __ LoadFromOffset(kLoadWord, ref_reg, obj, offset); 6871 } 6872 6873 // Object* ref = ref_addr->AsMirrorPtr() 6874 __ MaybeUnpoisonHeapReference(ref_reg); 6875 6876 // Slow path marking the object `ref` when it is gray. 6877 SlowPathCodeARM* slow_path; 6878 if (always_update_field) { 6879 DCHECK(temp2 != nullptr); 6880 // ReadBarrierMarkAndUpdateFieldSlowPathARM only supports address 6881 // of the form `obj + field_offset`, where `obj` is a register and 6882 // `field_offset` is a register pair (of which only the lower half 6883 // is used). Thus `offset` and `scale_factor` above are expected 6884 // to be null in this code path. 6885 DCHECK_EQ(offset, 0u); 6886 DCHECK_EQ(scale_factor, ScaleFactor::TIMES_1); 6887 slow_path = new (GetGraph()->GetArena()) ReadBarrierMarkAndUpdateFieldSlowPathARM( 6888 instruction, ref, obj, /* field_offset */ index, temp_reg, *temp2); 6889 } else { 6890 slow_path = new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM(instruction, ref); 6891 } 6892 AddSlowPath(slow_path); 6893 6894 // if (rb_state == ReadBarrier::GrayState()) 6895 // ref = ReadBarrier::Mark(ref); 6896 // Given the numeric representation, it's enough to check the low bit of the 6897 // rb_state. We do that by shifting the bit out of the lock word with LSRS 6898 // which can be a 16-bit instruction unlike the TST immediate. 6899 static_assert(ReadBarrier::WhiteState() == 0, "Expecting white to have value 0"); 6900 static_assert(ReadBarrier::GrayState() == 1, "Expecting gray to have value 1"); 6901 __ Lsrs(temp_reg, temp_reg, LockWord::kReadBarrierStateShift + 1); 6902 __ b(slow_path->GetEntryLabel(), CS); // Carry flag is the last bit shifted out by LSRS. 6903 __ Bind(slow_path->GetExitLabel()); 6904} 6905 6906void CodeGeneratorARM::GenerateReadBarrierSlow(HInstruction* instruction, 6907 Location out, 6908 Location ref, 6909 Location obj, 6910 uint32_t offset, 6911 Location index) { 6912 DCHECK(kEmitCompilerReadBarrier); 6913 6914 // Insert a slow path based read barrier *after* the reference load. 6915 // 6916 // If heap poisoning is enabled, the unpoisoning of the loaded 6917 // reference will be carried out by the runtime within the slow 6918 // path. 6919 // 6920 // Note that `ref` currently does not get unpoisoned (when heap 6921 // poisoning is enabled), which is alright as the `ref` argument is 6922 // not used by the artReadBarrierSlow entry point. 6923 // 6924 // TODO: Unpoison `ref` when it is used by artReadBarrierSlow. 6925 SlowPathCodeARM* slow_path = new (GetGraph()->GetArena()) 6926 ReadBarrierForHeapReferenceSlowPathARM(instruction, out, ref, obj, offset, index); 6927 AddSlowPath(slow_path); 6928 6929 __ b(slow_path->GetEntryLabel()); 6930 __ Bind(slow_path->GetExitLabel()); 6931} 6932 6933void CodeGeneratorARM::MaybeGenerateReadBarrierSlow(HInstruction* instruction, 6934 Location out, 6935 Location ref, 6936 Location obj, 6937 uint32_t offset, 6938 Location index) { 6939 if (kEmitCompilerReadBarrier) { 6940 // Baker's read barriers shall be handled by the fast path 6941 // (CodeGeneratorARM::GenerateReferenceLoadWithBakerReadBarrier). 6942 DCHECK(!kUseBakerReadBarrier); 6943 // If heap poisoning is enabled, unpoisoning will be taken care of 6944 // by the runtime within the slow path. 6945 GenerateReadBarrierSlow(instruction, out, ref, obj, offset, index); 6946 } else if (kPoisonHeapReferences) { 6947 __ UnpoisonHeapReference(out.AsRegister<Register>()); 6948 } 6949} 6950 6951void CodeGeneratorARM::GenerateReadBarrierForRootSlow(HInstruction* instruction, 6952 Location out, 6953 Location root) { 6954 DCHECK(kEmitCompilerReadBarrier); 6955 6956 // Insert a slow path based read barrier *after* the GC root load. 6957 // 6958 // Note that GC roots are not affected by heap poisoning, so we do 6959 // not need to do anything special for this here. 6960 SlowPathCodeARM* slow_path = 6961 new (GetGraph()->GetArena()) ReadBarrierForRootSlowPathARM(instruction, out, root); 6962 AddSlowPath(slow_path); 6963 6964 __ b(slow_path->GetEntryLabel()); 6965 __ Bind(slow_path->GetExitLabel()); 6966} 6967 6968HInvokeStaticOrDirect::DispatchInfo CodeGeneratorARM::GetSupportedInvokeStaticOrDirectDispatch( 6969 const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info, 6970 HInvokeStaticOrDirect* invoke) { 6971 HInvokeStaticOrDirect::DispatchInfo dispatch_info = desired_dispatch_info; 6972 // We disable pc-relative load when there is an irreducible loop, as the optimization 6973 // is incompatible with it. 6974 // TODO: Create as many ArmDexCacheArraysBase instructions as needed for methods 6975 // with irreducible loops. 6976 if (GetGraph()->HasIrreducibleLoops() && 6977 (dispatch_info.method_load_kind == 6978 HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative)) { 6979 dispatch_info.method_load_kind = HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod; 6980 } 6981 6982 if (dispatch_info.code_ptr_location == HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative) { 6983 const DexFile& outer_dex_file = GetGraph()->GetDexFile(); 6984 if (&outer_dex_file != invoke->GetTargetMethod().dex_file) { 6985 // Calls across dex files are more likely to exceed the available BL range, 6986 // so use absolute patch with fixup if available and kCallArtMethod otherwise. 6987 HInvokeStaticOrDirect::CodePtrLocation code_ptr_location = 6988 (desired_dispatch_info.method_load_kind == 6989 HInvokeStaticOrDirect::MethodLoadKind::kDirectAddressWithFixup) 6990 ? HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup 6991 : HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod; 6992 return HInvokeStaticOrDirect::DispatchInfo { 6993 dispatch_info.method_load_kind, 6994 code_ptr_location, 6995 dispatch_info.method_load_data, 6996 0u 6997 }; 6998 } 6999 } 7000 return dispatch_info; 7001} 7002 7003Register CodeGeneratorARM::GetInvokeStaticOrDirectExtraParameter(HInvokeStaticOrDirect* invoke, 7004 Register temp) { 7005 DCHECK_EQ(invoke->InputCount(), invoke->GetNumberOfArguments() + 1u); 7006 Location location = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 7007 if (!invoke->GetLocations()->Intrinsified()) { 7008 return location.AsRegister<Register>(); 7009 } 7010 // For intrinsics we allow any location, so it may be on the stack. 7011 if (!location.IsRegister()) { 7012 __ LoadFromOffset(kLoadWord, temp, SP, location.GetStackIndex()); 7013 return temp; 7014 } 7015 // For register locations, check if the register was saved. If so, get it from the stack. 7016 // Note: There is a chance that the register was saved but not overwritten, so we could 7017 // save one load. However, since this is just an intrinsic slow path we prefer this 7018 // simple and more robust approach rather that trying to determine if that's the case. 7019 SlowPathCode* slow_path = GetCurrentSlowPath(); 7020 DCHECK(slow_path != nullptr); // For intrinsified invokes the call is emitted on the slow path. 7021 if (slow_path->IsCoreRegisterSaved(location.AsRegister<Register>())) { 7022 int stack_offset = slow_path->GetStackOffsetOfCoreRegister(location.AsRegister<Register>()); 7023 __ LoadFromOffset(kLoadWord, temp, SP, stack_offset); 7024 return temp; 7025 } 7026 return location.AsRegister<Register>(); 7027} 7028 7029void CodeGeneratorARM::GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) { 7030 // For better instruction scheduling we load the direct code pointer before the method pointer. 7031 switch (invoke->GetCodePtrLocation()) { 7032 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup: 7033 // LR = code address from literal pool with link-time patch. 7034 __ LoadLiteral(LR, DeduplicateMethodCodeLiteral(invoke->GetTargetMethod())); 7035 break; 7036 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect: 7037 // LR = invoke->GetDirectCodePtr(); 7038 __ LoadImmediate(LR, invoke->GetDirectCodePtr()); 7039 break; 7040 default: 7041 break; 7042 } 7043 7044 Location callee_method = temp; // For all kinds except kRecursive, callee will be in temp. 7045 switch (invoke->GetMethodLoadKind()) { 7046 case HInvokeStaticOrDirect::MethodLoadKind::kStringInit: { 7047 uint32_t offset = 7048 GetThreadOffset<kArmPointerSize>(invoke->GetStringInitEntryPoint()).Int32Value(); 7049 // temp = thread->string_init_entrypoint 7050 __ LoadFromOffset(kLoadWord, temp.AsRegister<Register>(), TR, offset); 7051 break; 7052 } 7053 case HInvokeStaticOrDirect::MethodLoadKind::kRecursive: 7054 callee_method = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 7055 break; 7056 case HInvokeStaticOrDirect::MethodLoadKind::kDirectAddress: 7057 __ LoadImmediate(temp.AsRegister<Register>(), invoke->GetMethodAddress()); 7058 break; 7059 case HInvokeStaticOrDirect::MethodLoadKind::kDirectAddressWithFixup: 7060 __ LoadLiteral(temp.AsRegister<Register>(), 7061 DeduplicateMethodAddressLiteral(invoke->GetTargetMethod())); 7062 break; 7063 case HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative: { 7064 HArmDexCacheArraysBase* base = 7065 invoke->InputAt(invoke->GetSpecialInputIndex())->AsArmDexCacheArraysBase(); 7066 Register base_reg = GetInvokeStaticOrDirectExtraParameter(invoke, 7067 temp.AsRegister<Register>()); 7068 int32_t offset = invoke->GetDexCacheArrayOffset() - base->GetElementOffset(); 7069 __ LoadFromOffset(kLoadWord, temp.AsRegister<Register>(), base_reg, offset); 7070 break; 7071 } 7072 case HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod: { 7073 Location current_method = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 7074 Register method_reg; 7075 Register reg = temp.AsRegister<Register>(); 7076 if (current_method.IsRegister()) { 7077 method_reg = current_method.AsRegister<Register>(); 7078 } else { 7079 DCHECK(invoke->GetLocations()->Intrinsified()); 7080 DCHECK(!current_method.IsValid()); 7081 method_reg = reg; 7082 __ LoadFromOffset(kLoadWord, reg, SP, kCurrentMethodStackOffset); 7083 } 7084 // /* ArtMethod*[] */ temp = temp.ptr_sized_fields_->dex_cache_resolved_methods_; 7085 __ LoadFromOffset(kLoadWord, 7086 reg, 7087 method_reg, 7088 ArtMethod::DexCacheResolvedMethodsOffset(kArmPointerSize).Int32Value()); 7089 // temp = temp[index_in_cache]; 7090 // Note: Don't use invoke->GetTargetMethod() as it may point to a different dex file. 7091 uint32_t index_in_cache = invoke->GetDexMethodIndex(); 7092 __ LoadFromOffset(kLoadWord, reg, reg, CodeGenerator::GetCachePointerOffset(index_in_cache)); 7093 break; 7094 } 7095 } 7096 7097 switch (invoke->GetCodePtrLocation()) { 7098 case HInvokeStaticOrDirect::CodePtrLocation::kCallSelf: 7099 __ bl(GetFrameEntryLabel()); 7100 break; 7101 case HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative: 7102 relative_call_patches_.emplace_back(*invoke->GetTargetMethod().dex_file, 7103 invoke->GetTargetMethod().dex_method_index); 7104 __ BindTrackedLabel(&relative_call_patches_.back().label); 7105 // Arbitrarily branch to the BL itself, override at link time. 7106 __ bl(&relative_call_patches_.back().label); 7107 break; 7108 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup: 7109 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect: 7110 // LR prepared above for better instruction scheduling. 7111 // LR() 7112 __ blx(LR); 7113 break; 7114 case HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod: 7115 // LR = callee_method->entry_point_from_quick_compiled_code_ 7116 __ LoadFromOffset( 7117 kLoadWord, LR, callee_method.AsRegister<Register>(), 7118 ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmPointerSize).Int32Value()); 7119 // LR() 7120 __ blx(LR); 7121 break; 7122 } 7123 7124 DCHECK(!IsLeafMethod()); 7125} 7126 7127void CodeGeneratorARM::GenerateVirtualCall(HInvokeVirtual* invoke, Location temp_location) { 7128 Register temp = temp_location.AsRegister<Register>(); 7129 uint32_t method_offset = mirror::Class::EmbeddedVTableEntryOffset( 7130 invoke->GetVTableIndex(), kArmPointerSize).Uint32Value(); 7131 7132 // Use the calling convention instead of the location of the receiver, as 7133 // intrinsics may have put the receiver in a different register. In the intrinsics 7134 // slow path, the arguments have been moved to the right place, so here we are 7135 // guaranteed that the receiver is the first register of the calling convention. 7136 InvokeDexCallingConvention calling_convention; 7137 Register receiver = calling_convention.GetRegisterAt(0); 7138 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 7139 // /* HeapReference<Class> */ temp = receiver->klass_ 7140 __ LoadFromOffset(kLoadWord, temp, receiver, class_offset); 7141 MaybeRecordImplicitNullCheck(invoke); 7142 // Instead of simply (possibly) unpoisoning `temp` here, we should 7143 // emit a read barrier for the previous class reference load. 7144 // However this is not required in practice, as this is an 7145 // intermediate/temporary reference and because the current 7146 // concurrent copying collector keeps the from-space memory 7147 // intact/accessible until the end of the marking phase (the 7148 // concurrent copying collector may not in the future). 7149 __ MaybeUnpoisonHeapReference(temp); 7150 // temp = temp->GetMethodAt(method_offset); 7151 uint32_t entry_point = ArtMethod::EntryPointFromQuickCompiledCodeOffset( 7152 kArmPointerSize).Int32Value(); 7153 __ LoadFromOffset(kLoadWord, temp, temp, method_offset); 7154 // LR = temp->GetEntryPoint(); 7155 __ LoadFromOffset(kLoadWord, LR, temp, entry_point); 7156 // LR(); 7157 __ blx(LR); 7158} 7159 7160CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewPcRelativeStringPatch( 7161 const DexFile& dex_file, uint32_t string_index) { 7162 return NewPcRelativePatch(dex_file, string_index, &pc_relative_string_patches_); 7163} 7164 7165CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewPcRelativeTypePatch( 7166 const DexFile& dex_file, uint32_t type_index) { 7167 return NewPcRelativePatch(dex_file, type_index, &pc_relative_type_patches_); 7168} 7169 7170CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewPcRelativeDexCacheArrayPatch( 7171 const DexFile& dex_file, uint32_t element_offset) { 7172 return NewPcRelativePatch(dex_file, element_offset, &pc_relative_dex_cache_patches_); 7173} 7174 7175CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewPcRelativePatch( 7176 const DexFile& dex_file, uint32_t offset_or_index, ArenaDeque<PcRelativePatchInfo>* patches) { 7177 patches->emplace_back(dex_file, offset_or_index); 7178 return &patches->back(); 7179} 7180 7181Literal* CodeGeneratorARM::DeduplicateBootImageStringLiteral(const DexFile& dex_file, 7182 uint32_t string_index) { 7183 return boot_image_string_patches_.GetOrCreate( 7184 StringReference(&dex_file, string_index), 7185 [this]() { return __ NewLiteral<uint32_t>(/* placeholder */ 0u); }); 7186} 7187 7188Literal* CodeGeneratorARM::DeduplicateBootImageTypeLiteral(const DexFile& dex_file, 7189 uint32_t type_index) { 7190 return boot_image_type_patches_.GetOrCreate( 7191 TypeReference(&dex_file, type_index), 7192 [this]() { return __ NewLiteral<uint32_t>(/* placeholder */ 0u); }); 7193} 7194 7195Literal* CodeGeneratorARM::DeduplicateBootImageAddressLiteral(uint32_t address) { 7196 bool needs_patch = GetCompilerOptions().GetIncludePatchInformation(); 7197 Uint32ToLiteralMap* map = needs_patch ? &boot_image_address_patches_ : &uint32_literals_; 7198 return DeduplicateUint32Literal(dchecked_integral_cast<uint32_t>(address), map); 7199} 7200 7201Literal* CodeGeneratorARM::DeduplicateDexCacheAddressLiteral(uint32_t address) { 7202 return DeduplicateUint32Literal(address, &uint32_literals_); 7203} 7204 7205template <LinkerPatch (*Factory)(size_t, const DexFile*, uint32_t, uint32_t)> 7206inline void CodeGeneratorARM::EmitPcRelativeLinkerPatches( 7207 const ArenaDeque<PcRelativePatchInfo>& infos, 7208 ArenaVector<LinkerPatch>* linker_patches) { 7209 for (const PcRelativePatchInfo& info : infos) { 7210 const DexFile& dex_file = info.target_dex_file; 7211 size_t offset_or_index = info.offset_or_index; 7212 DCHECK(info.add_pc_label.IsBound()); 7213 uint32_t add_pc_offset = dchecked_integral_cast<uint32_t>(info.add_pc_label.Position()); 7214 // Add MOVW patch. 7215 DCHECK(info.movw_label.IsBound()); 7216 uint32_t movw_offset = dchecked_integral_cast<uint32_t>(info.movw_label.Position()); 7217 linker_patches->push_back(Factory(movw_offset, &dex_file, add_pc_offset, offset_or_index)); 7218 // Add MOVT patch. 7219 DCHECK(info.movt_label.IsBound()); 7220 uint32_t movt_offset = dchecked_integral_cast<uint32_t>(info.movt_label.Position()); 7221 linker_patches->push_back(Factory(movt_offset, &dex_file, add_pc_offset, offset_or_index)); 7222 } 7223} 7224 7225void CodeGeneratorARM::EmitLinkerPatches(ArenaVector<LinkerPatch>* linker_patches) { 7226 DCHECK(linker_patches->empty()); 7227 size_t size = 7228 method_patches_.size() + 7229 call_patches_.size() + 7230 relative_call_patches_.size() + 7231 /* MOVW+MOVT for each entry */ 2u * pc_relative_dex_cache_patches_.size() + 7232 boot_image_string_patches_.size() + 7233 /* MOVW+MOVT for each entry */ 2u * pc_relative_string_patches_.size() + 7234 boot_image_type_patches_.size() + 7235 /* MOVW+MOVT for each entry */ 2u * pc_relative_type_patches_.size() + 7236 boot_image_address_patches_.size(); 7237 linker_patches->reserve(size); 7238 for (const auto& entry : method_patches_) { 7239 const MethodReference& target_method = entry.first; 7240 Literal* literal = entry.second; 7241 DCHECK(literal->GetLabel()->IsBound()); 7242 uint32_t literal_offset = literal->GetLabel()->Position(); 7243 linker_patches->push_back(LinkerPatch::MethodPatch(literal_offset, 7244 target_method.dex_file, 7245 target_method.dex_method_index)); 7246 } 7247 for (const auto& entry : call_patches_) { 7248 const MethodReference& target_method = entry.first; 7249 Literal* literal = entry.second; 7250 DCHECK(literal->GetLabel()->IsBound()); 7251 uint32_t literal_offset = literal->GetLabel()->Position(); 7252 linker_patches->push_back(LinkerPatch::CodePatch(literal_offset, 7253 target_method.dex_file, 7254 target_method.dex_method_index)); 7255 } 7256 for (const PatchInfo<Label>& info : relative_call_patches_) { 7257 uint32_t literal_offset = info.label.Position(); 7258 linker_patches->push_back( 7259 LinkerPatch::RelativeCodePatch(literal_offset, &info.dex_file, info.index)); 7260 } 7261 EmitPcRelativeLinkerPatches<LinkerPatch::DexCacheArrayPatch>(pc_relative_dex_cache_patches_, 7262 linker_patches); 7263 for (const auto& entry : boot_image_string_patches_) { 7264 const StringReference& target_string = entry.first; 7265 Literal* literal = entry.second; 7266 DCHECK(literal->GetLabel()->IsBound()); 7267 uint32_t literal_offset = literal->GetLabel()->Position(); 7268 linker_patches->push_back(LinkerPatch::StringPatch(literal_offset, 7269 target_string.dex_file, 7270 target_string.string_index)); 7271 } 7272 if (!GetCompilerOptions().IsBootImage()) { 7273 EmitPcRelativeLinkerPatches<LinkerPatch::StringBssEntryPatch>(pc_relative_string_patches_, 7274 linker_patches); 7275 } else { 7276 EmitPcRelativeLinkerPatches<LinkerPatch::RelativeStringPatch>(pc_relative_string_patches_, 7277 linker_patches); 7278 } 7279 for (const auto& entry : boot_image_type_patches_) { 7280 const TypeReference& target_type = entry.first; 7281 Literal* literal = entry.second; 7282 DCHECK(literal->GetLabel()->IsBound()); 7283 uint32_t literal_offset = literal->GetLabel()->Position(); 7284 linker_patches->push_back(LinkerPatch::TypePatch(literal_offset, 7285 target_type.dex_file, 7286 target_type.type_index)); 7287 } 7288 EmitPcRelativeLinkerPatches<LinkerPatch::RelativeTypePatch>(pc_relative_type_patches_, 7289 linker_patches); 7290 for (const auto& entry : boot_image_address_patches_) { 7291 DCHECK(GetCompilerOptions().GetIncludePatchInformation()); 7292 Literal* literal = entry.second; 7293 DCHECK(literal->GetLabel()->IsBound()); 7294 uint32_t literal_offset = literal->GetLabel()->Position(); 7295 linker_patches->push_back(LinkerPatch::RecordPosition(literal_offset)); 7296 } 7297} 7298 7299Literal* CodeGeneratorARM::DeduplicateUint32Literal(uint32_t value, Uint32ToLiteralMap* map) { 7300 return map->GetOrCreate( 7301 value, 7302 [this, value]() { return __ NewLiteral<uint32_t>(value); }); 7303} 7304 7305Literal* CodeGeneratorARM::DeduplicateMethodLiteral(MethodReference target_method, 7306 MethodToLiteralMap* map) { 7307 return map->GetOrCreate( 7308 target_method, 7309 [this]() { return __ NewLiteral<uint32_t>(/* placeholder */ 0u); }); 7310} 7311 7312Literal* CodeGeneratorARM::DeduplicateMethodAddressLiteral(MethodReference target_method) { 7313 return DeduplicateMethodLiteral(target_method, &method_patches_); 7314} 7315 7316Literal* CodeGeneratorARM::DeduplicateMethodCodeLiteral(MethodReference target_method) { 7317 return DeduplicateMethodLiteral(target_method, &call_patches_); 7318} 7319 7320void LocationsBuilderARM::VisitMultiplyAccumulate(HMultiplyAccumulate* instr) { 7321 LocationSummary* locations = 7322 new (GetGraph()->GetArena()) LocationSummary(instr, LocationSummary::kNoCall); 7323 locations->SetInAt(HMultiplyAccumulate::kInputAccumulatorIndex, 7324 Location::RequiresRegister()); 7325 locations->SetInAt(HMultiplyAccumulate::kInputMulLeftIndex, Location::RequiresRegister()); 7326 locations->SetInAt(HMultiplyAccumulate::kInputMulRightIndex, Location::RequiresRegister()); 7327 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 7328} 7329 7330void InstructionCodeGeneratorARM::VisitMultiplyAccumulate(HMultiplyAccumulate* instr) { 7331 LocationSummary* locations = instr->GetLocations(); 7332 Register res = locations->Out().AsRegister<Register>(); 7333 Register accumulator = 7334 locations->InAt(HMultiplyAccumulate::kInputAccumulatorIndex).AsRegister<Register>(); 7335 Register mul_left = 7336 locations->InAt(HMultiplyAccumulate::kInputMulLeftIndex).AsRegister<Register>(); 7337 Register mul_right = 7338 locations->InAt(HMultiplyAccumulate::kInputMulRightIndex).AsRegister<Register>(); 7339 7340 if (instr->GetOpKind() == HInstruction::kAdd) { 7341 __ mla(res, mul_left, mul_right, accumulator); 7342 } else { 7343 __ mls(res, mul_left, mul_right, accumulator); 7344 } 7345} 7346 7347void LocationsBuilderARM::VisitBoundType(HBoundType* instruction ATTRIBUTE_UNUSED) { 7348 // Nothing to do, this should be removed during prepare for register allocator. 7349 LOG(FATAL) << "Unreachable"; 7350} 7351 7352void InstructionCodeGeneratorARM::VisitBoundType(HBoundType* instruction ATTRIBUTE_UNUSED) { 7353 // Nothing to do, this should be removed during prepare for register allocator. 7354 LOG(FATAL) << "Unreachable"; 7355} 7356 7357// Simple implementation of packed switch - generate cascaded compare/jumps. 7358void LocationsBuilderARM::VisitPackedSwitch(HPackedSwitch* switch_instr) { 7359 LocationSummary* locations = 7360 new (GetGraph()->GetArena()) LocationSummary(switch_instr, LocationSummary::kNoCall); 7361 locations->SetInAt(0, Location::RequiresRegister()); 7362 if (switch_instr->GetNumEntries() > kPackedSwitchCompareJumpThreshold && 7363 codegen_->GetAssembler()->IsThumb()) { 7364 locations->AddTemp(Location::RequiresRegister()); // We need a temp for the table base. 7365 if (switch_instr->GetStartValue() != 0) { 7366 locations->AddTemp(Location::RequiresRegister()); // We need a temp for the bias. 7367 } 7368 } 7369} 7370 7371void InstructionCodeGeneratorARM::VisitPackedSwitch(HPackedSwitch* switch_instr) { 7372 int32_t lower_bound = switch_instr->GetStartValue(); 7373 uint32_t num_entries = switch_instr->GetNumEntries(); 7374 LocationSummary* locations = switch_instr->GetLocations(); 7375 Register value_reg = locations->InAt(0).AsRegister<Register>(); 7376 HBasicBlock* default_block = switch_instr->GetDefaultBlock(); 7377 7378 if (num_entries <= kPackedSwitchCompareJumpThreshold || !codegen_->GetAssembler()->IsThumb()) { 7379 // Create a series of compare/jumps. 7380 Register temp_reg = IP; 7381 // Note: It is fine for the below AddConstantSetFlags() using IP register to temporarily store 7382 // the immediate, because IP is used as the destination register. For the other 7383 // AddConstantSetFlags() and GenerateCompareWithImmediate(), the immediate values are constant, 7384 // and they can be encoded in the instruction without making use of IP register. 7385 __ AddConstantSetFlags(temp_reg, value_reg, -lower_bound); 7386 7387 const ArenaVector<HBasicBlock*>& successors = switch_instr->GetBlock()->GetSuccessors(); 7388 // Jump to successors[0] if value == lower_bound. 7389 __ b(codegen_->GetLabelOf(successors[0]), EQ); 7390 int32_t last_index = 0; 7391 for (; num_entries - last_index > 2; last_index += 2) { 7392 __ AddConstantSetFlags(temp_reg, temp_reg, -2); 7393 // Jump to successors[last_index + 1] if value < case_value[last_index + 2]. 7394 __ b(codegen_->GetLabelOf(successors[last_index + 1]), LO); 7395 // Jump to successors[last_index + 2] if value == case_value[last_index + 2]. 7396 __ b(codegen_->GetLabelOf(successors[last_index + 2]), EQ); 7397 } 7398 if (num_entries - last_index == 2) { 7399 // The last missing case_value. 7400 __ CmpConstant(temp_reg, 1); 7401 __ b(codegen_->GetLabelOf(successors[last_index + 1]), EQ); 7402 } 7403 7404 // And the default for any other value. 7405 if (!codegen_->GoesToNextBlock(switch_instr->GetBlock(), default_block)) { 7406 __ b(codegen_->GetLabelOf(default_block)); 7407 } 7408 } else { 7409 // Create a table lookup. 7410 Register temp_reg = locations->GetTemp(0).AsRegister<Register>(); 7411 7412 // Materialize a pointer to the switch table 7413 std::vector<Label*> labels(num_entries); 7414 const ArenaVector<HBasicBlock*>& successors = switch_instr->GetBlock()->GetSuccessors(); 7415 for (uint32_t i = 0; i < num_entries; i++) { 7416 labels[i] = codegen_->GetLabelOf(successors[i]); 7417 } 7418 JumpTable* table = __ CreateJumpTable(std::move(labels), temp_reg); 7419 7420 // Remove the bias. 7421 Register key_reg; 7422 if (lower_bound != 0) { 7423 key_reg = locations->GetTemp(1).AsRegister<Register>(); 7424 __ AddConstant(key_reg, value_reg, -lower_bound); 7425 } else { 7426 key_reg = value_reg; 7427 } 7428 7429 // Check whether the value is in the table, jump to default block if not. 7430 __ CmpConstant(key_reg, num_entries - 1); 7431 __ b(codegen_->GetLabelOf(default_block), Condition::HI); 7432 7433 // Load the displacement from the table. 7434 __ ldr(temp_reg, Address(temp_reg, key_reg, Shift::LSL, 2)); 7435 7436 // Dispatch is a direct add to the PC (for Thumb2). 7437 __ EmitJumpTableDispatch(table, temp_reg); 7438 } 7439} 7440 7441void LocationsBuilderARM::VisitArmDexCacheArraysBase(HArmDexCacheArraysBase* base) { 7442 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(base); 7443 locations->SetOut(Location::RequiresRegister()); 7444} 7445 7446void InstructionCodeGeneratorARM::VisitArmDexCacheArraysBase(HArmDexCacheArraysBase* base) { 7447 Register base_reg = base->GetLocations()->Out().AsRegister<Register>(); 7448 CodeGeneratorARM::PcRelativePatchInfo* labels = 7449 codegen_->NewPcRelativeDexCacheArrayPatch(base->GetDexFile(), base->GetElementOffset()); 7450 __ BindTrackedLabel(&labels->movw_label); 7451 __ movw(base_reg, /* placeholder */ 0u); 7452 __ BindTrackedLabel(&labels->movt_label); 7453 __ movt(base_reg, /* placeholder */ 0u); 7454 __ BindTrackedLabel(&labels->add_pc_label); 7455 __ add(base_reg, base_reg, ShifterOperand(PC)); 7456} 7457 7458void CodeGeneratorARM::MoveFromReturnRegister(Location trg, Primitive::Type type) { 7459 if (!trg.IsValid()) { 7460 DCHECK_EQ(type, Primitive::kPrimVoid); 7461 return; 7462 } 7463 7464 DCHECK_NE(type, Primitive::kPrimVoid); 7465 7466 Location return_loc = InvokeDexCallingConventionVisitorARM().GetReturnLocation(type); 7467 if (return_loc.Equals(trg)) { 7468 return; 7469 } 7470 7471 // TODO: Consider pairs in the parallel move resolver, then this could be nicely merged 7472 // with the last branch. 7473 if (type == Primitive::kPrimLong) { 7474 HParallelMove parallel_move(GetGraph()->GetArena()); 7475 parallel_move.AddMove(return_loc.ToLow(), trg.ToLow(), Primitive::kPrimInt, nullptr); 7476 parallel_move.AddMove(return_loc.ToHigh(), trg.ToHigh(), Primitive::kPrimInt, nullptr); 7477 GetMoveResolver()->EmitNativeCode(¶llel_move); 7478 } else if (type == Primitive::kPrimDouble) { 7479 HParallelMove parallel_move(GetGraph()->GetArena()); 7480 parallel_move.AddMove(return_loc.ToLow(), trg.ToLow(), Primitive::kPrimFloat, nullptr); 7481 parallel_move.AddMove(return_loc.ToHigh(), trg.ToHigh(), Primitive::kPrimFloat, nullptr); 7482 GetMoveResolver()->EmitNativeCode(¶llel_move); 7483 } else { 7484 // Let the parallel move resolver take care of all of this. 7485 HParallelMove parallel_move(GetGraph()->GetArena()); 7486 parallel_move.AddMove(return_loc, trg, type, nullptr); 7487 GetMoveResolver()->EmitNativeCode(¶llel_move); 7488 } 7489} 7490 7491void LocationsBuilderARM::VisitClassTableGet(HClassTableGet* instruction) { 7492 LocationSummary* locations = 7493 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 7494 locations->SetInAt(0, Location::RequiresRegister()); 7495 locations->SetOut(Location::RequiresRegister()); 7496} 7497 7498void InstructionCodeGeneratorARM::VisitClassTableGet(HClassTableGet* instruction) { 7499 LocationSummary* locations = instruction->GetLocations(); 7500 if (instruction->GetTableKind() == HClassTableGet::TableKind::kVTable) { 7501 uint32_t method_offset = mirror::Class::EmbeddedVTableEntryOffset( 7502 instruction->GetIndex(), kArmPointerSize).SizeValue(); 7503 __ LoadFromOffset(kLoadWord, 7504 locations->Out().AsRegister<Register>(), 7505 locations->InAt(0).AsRegister<Register>(), 7506 method_offset); 7507 } else { 7508 uint32_t method_offset = static_cast<uint32_t>(ImTable::OffsetOfElement( 7509 instruction->GetIndex(), kArmPointerSize)); 7510 __ LoadFromOffset(kLoadWord, 7511 locations->Out().AsRegister<Register>(), 7512 locations->InAt(0).AsRegister<Register>(), 7513 mirror::Class::ImtPtrOffset(kArmPointerSize).Uint32Value()); 7514 __ LoadFromOffset(kLoadWord, 7515 locations->Out().AsRegister<Register>(), 7516 locations->Out().AsRegister<Register>(), 7517 method_offset); 7518 } 7519} 7520 7521#undef __ 7522#undef QUICK_ENTRY_POINT 7523 7524} // namespace arm 7525} // namespace art 7526