code_generator_arm.cc revision e807ff725159dabab3a3028bbb76f83ebcfa40de
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), 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 = instruction_->GetLocations(); 380 381 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 382 __ Bind(GetEntryLabel()); 383 SaveLiveRegisters(codegen, locations); 384 385 InvokeRuntimeCallingConvention calling_convention; 386 dex::TypeIndex type_index = cls_->GetTypeIndex(); 387 __ LoadImmediate(calling_convention.GetRegisterAt(0), type_index.index_); 388 QuickEntrypointEnum entrypoint = do_clinit_ ? kQuickInitializeStaticStorage 389 : kQuickInitializeType; 390 arm_codegen->InvokeRuntime(entrypoint, instruction_, dex_pc_, this); 391 if (do_clinit_) { 392 CheckEntrypointTypes<kQuickInitializeStaticStorage, void*, uint32_t>(); 393 } else { 394 CheckEntrypointTypes<kQuickInitializeType, void*, uint32_t>(); 395 } 396 397 // Move the class to the desired location. 398 Location out = locations->Out(); 399 if (out.IsValid()) { 400 DCHECK(out.IsRegister() && !locations->GetLiveRegisters()->ContainsCoreRegister(out.reg())); 401 arm_codegen->Move32(locations->Out(), Location::RegisterLocation(R0)); 402 } 403 RestoreLiveRegisters(codegen, locations); 404 // For HLoadClass/kBssEntry, store the resolved Class to the BSS entry. 405 DCHECK_EQ(instruction_->IsLoadClass(), cls_ == instruction_); 406 if (cls_ == instruction_ && cls_->GetLoadKind() == HLoadClass::LoadKind::kBssEntry) { 407 DCHECK(out.IsValid()); 408 // TODO: Change art_quick_initialize_type/art_quick_initialize_static_storage to 409 // kSaveEverything and use a temporary for the .bss entry address in the fast path, 410 // so that we can avoid another calculation here. 411 CodeGeneratorARM::PcRelativePatchInfo* labels = 412 arm_codegen->NewTypeBssEntryPatch(cls_->GetDexFile(), type_index); 413 __ BindTrackedLabel(&labels->movw_label); 414 __ movw(IP, /* placeholder */ 0u); 415 __ BindTrackedLabel(&labels->movt_label); 416 __ movt(IP, /* placeholder */ 0u); 417 __ BindTrackedLabel(&labels->add_pc_label); 418 __ add(IP, IP, ShifterOperand(PC)); 419 __ str(locations->Out().AsRegister<Register>(), Address(IP)); 420 } 421 __ b(GetExitLabel()); 422 } 423 424 const char* GetDescription() const OVERRIDE { return "LoadClassSlowPathARM"; } 425 426 private: 427 // The class this slow path will load. 428 HLoadClass* const cls_; 429 430 // The dex PC of `at_`. 431 const uint32_t dex_pc_; 432 433 // Whether to initialize the class. 434 const bool do_clinit_; 435 436 DISALLOW_COPY_AND_ASSIGN(LoadClassSlowPathARM); 437}; 438 439class LoadStringSlowPathARM : public SlowPathCodeARM { 440 public: 441 explicit LoadStringSlowPathARM(HLoadString* instruction) : SlowPathCodeARM(instruction) {} 442 443 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 444 LocationSummary* locations = instruction_->GetLocations(); 445 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg())); 446 HLoadString* load = instruction_->AsLoadString(); 447 const dex::StringIndex string_index = load->GetStringIndex(); 448 Register out = locations->Out().AsRegister<Register>(); 449 Register temp = locations->GetTemp(0).AsRegister<Register>(); 450 constexpr bool call_saves_everything_except_r0 = (!kUseReadBarrier || kUseBakerReadBarrier); 451 452 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 453 __ Bind(GetEntryLabel()); 454 SaveLiveRegisters(codegen, locations); 455 456 InvokeRuntimeCallingConvention calling_convention; 457 // In the unlucky case that the `temp` is R0, we preserve the address in `out` across 458 // the kSaveEverything call (or use `out` for the address after non-kSaveEverything call). 459 bool temp_is_r0 = (temp == calling_convention.GetRegisterAt(0)); 460 Register entry_address = temp_is_r0 ? out : temp; 461 DCHECK_NE(entry_address, calling_convention.GetRegisterAt(0)); 462 if (call_saves_everything_except_r0 && temp_is_r0) { 463 __ mov(entry_address, ShifterOperand(temp)); 464 } 465 466 __ LoadImmediate(calling_convention.GetRegisterAt(0), string_index.index_); 467 arm_codegen->InvokeRuntime(kQuickResolveString, instruction_, instruction_->GetDexPc(), this); 468 CheckEntrypointTypes<kQuickResolveString, void*, uint32_t>(); 469 470 // Store the resolved String to the .bss entry. 471 if (call_saves_everything_except_r0) { 472 // The string entry address was preserved in `entry_address` thanks to kSaveEverything. 473 __ str(R0, Address(entry_address)); 474 } else { 475 // For non-Baker read barrier, we need to re-calculate the address of the string entry. 476 CodeGeneratorARM::PcRelativePatchInfo* labels = 477 arm_codegen->NewPcRelativeStringPatch(load->GetDexFile(), string_index); 478 __ BindTrackedLabel(&labels->movw_label); 479 __ movw(entry_address, /* placeholder */ 0u); 480 __ BindTrackedLabel(&labels->movt_label); 481 __ movt(entry_address, /* placeholder */ 0u); 482 __ BindTrackedLabel(&labels->add_pc_label); 483 __ add(entry_address, entry_address, ShifterOperand(PC)); 484 __ str(R0, Address(entry_address)); 485 } 486 487 arm_codegen->Move32(locations->Out(), Location::RegisterLocation(R0)); 488 RestoreLiveRegisters(codegen, locations); 489 490 __ b(GetExitLabel()); 491 } 492 493 const char* GetDescription() const OVERRIDE { return "LoadStringSlowPathARM"; } 494 495 private: 496 DISALLOW_COPY_AND_ASSIGN(LoadStringSlowPathARM); 497}; 498 499class TypeCheckSlowPathARM : public SlowPathCodeARM { 500 public: 501 TypeCheckSlowPathARM(HInstruction* instruction, bool is_fatal) 502 : SlowPathCodeARM(instruction), is_fatal_(is_fatal) {} 503 504 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 505 LocationSummary* locations = instruction_->GetLocations(); 506 DCHECK(instruction_->IsCheckCast() 507 || !locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg())); 508 509 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 510 __ Bind(GetEntryLabel()); 511 512 if (!is_fatal_) { 513 SaveLiveRegisters(codegen, locations); 514 } 515 516 // We're moving two locations to locations that could overlap, so we need a parallel 517 // move resolver. 518 InvokeRuntimeCallingConvention calling_convention; 519 codegen->EmitParallelMoves(locations->InAt(0), 520 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 521 Primitive::kPrimNot, 522 locations->InAt(1), 523 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 524 Primitive::kPrimNot); 525 if (instruction_->IsInstanceOf()) { 526 arm_codegen->InvokeRuntime(kQuickInstanceofNonTrivial, 527 instruction_, 528 instruction_->GetDexPc(), 529 this); 530 CheckEntrypointTypes<kQuickInstanceofNonTrivial, size_t, mirror::Object*, mirror::Class*>(); 531 arm_codegen->Move32(locations->Out(), Location::RegisterLocation(R0)); 532 } else { 533 DCHECK(instruction_->IsCheckCast()); 534 arm_codegen->InvokeRuntime(kQuickCheckInstanceOf, 535 instruction_, 536 instruction_->GetDexPc(), 537 this); 538 CheckEntrypointTypes<kQuickCheckInstanceOf, void, mirror::Object*, mirror::Class*>(); 539 } 540 541 if (!is_fatal_) { 542 RestoreLiveRegisters(codegen, locations); 543 __ b(GetExitLabel()); 544 } 545 } 546 547 const char* GetDescription() const OVERRIDE { return "TypeCheckSlowPathARM"; } 548 549 bool IsFatal() const OVERRIDE { return is_fatal_; } 550 551 private: 552 const bool is_fatal_; 553 554 DISALLOW_COPY_AND_ASSIGN(TypeCheckSlowPathARM); 555}; 556 557class DeoptimizationSlowPathARM : public SlowPathCodeARM { 558 public: 559 explicit DeoptimizationSlowPathARM(HDeoptimize* instruction) 560 : SlowPathCodeARM(instruction) {} 561 562 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 563 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 564 __ Bind(GetEntryLabel()); 565 arm_codegen->InvokeRuntime(kQuickDeoptimize, instruction_, instruction_->GetDexPc(), this); 566 CheckEntrypointTypes<kQuickDeoptimize, void, void>(); 567 } 568 569 const char* GetDescription() const OVERRIDE { return "DeoptimizationSlowPathARM"; } 570 571 private: 572 DISALLOW_COPY_AND_ASSIGN(DeoptimizationSlowPathARM); 573}; 574 575class ArraySetSlowPathARM : public SlowPathCodeARM { 576 public: 577 explicit ArraySetSlowPathARM(HInstruction* instruction) : SlowPathCodeARM(instruction) {} 578 579 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 580 LocationSummary* locations = instruction_->GetLocations(); 581 __ Bind(GetEntryLabel()); 582 SaveLiveRegisters(codegen, locations); 583 584 InvokeRuntimeCallingConvention calling_convention; 585 HParallelMove parallel_move(codegen->GetGraph()->GetArena()); 586 parallel_move.AddMove( 587 locations->InAt(0), 588 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 589 Primitive::kPrimNot, 590 nullptr); 591 parallel_move.AddMove( 592 locations->InAt(1), 593 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 594 Primitive::kPrimInt, 595 nullptr); 596 parallel_move.AddMove( 597 locations->InAt(2), 598 Location::RegisterLocation(calling_convention.GetRegisterAt(2)), 599 Primitive::kPrimNot, 600 nullptr); 601 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 602 603 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 604 arm_codegen->InvokeRuntime(kQuickAputObject, instruction_, instruction_->GetDexPc(), this); 605 CheckEntrypointTypes<kQuickAputObject, void, mirror::Array*, int32_t, mirror::Object*>(); 606 RestoreLiveRegisters(codegen, locations); 607 __ b(GetExitLabel()); 608 } 609 610 const char* GetDescription() const OVERRIDE { return "ArraySetSlowPathARM"; } 611 612 private: 613 DISALLOW_COPY_AND_ASSIGN(ArraySetSlowPathARM); 614}; 615 616// Slow path marking an object reference `ref` during a read 617// barrier. The field `obj.field` in the object `obj` holding this 618// reference does not get updated by this slow path after marking (see 619// ReadBarrierMarkAndUpdateFieldSlowPathARM below for that). 620// 621// This means that after the execution of this slow path, `ref` will 622// always be up-to-date, but `obj.field` may not; i.e., after the 623// flip, `ref` will be a to-space reference, but `obj.field` will 624// probably still be a from-space reference (unless it gets updated by 625// another thread, or if another thread installed another object 626// reference (different from `ref`) in `obj.field`). 627class ReadBarrierMarkSlowPathARM : public SlowPathCodeARM { 628 public: 629 ReadBarrierMarkSlowPathARM(HInstruction* instruction, 630 Location ref, 631 Location entrypoint = Location::NoLocation()) 632 : SlowPathCodeARM(instruction), ref_(ref), entrypoint_(entrypoint) { 633 DCHECK(kEmitCompilerReadBarrier); 634 } 635 636 const char* GetDescription() const OVERRIDE { return "ReadBarrierMarkSlowPathARM"; } 637 638 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 639 LocationSummary* locations = instruction_->GetLocations(); 640 Register ref_reg = ref_.AsRegister<Register>(); 641 DCHECK(locations->CanCall()); 642 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(ref_reg)) << ref_reg; 643 DCHECK(instruction_->IsInstanceFieldGet() || 644 instruction_->IsStaticFieldGet() || 645 instruction_->IsArrayGet() || 646 instruction_->IsArraySet() || 647 instruction_->IsLoadClass() || 648 instruction_->IsLoadString() || 649 instruction_->IsInstanceOf() || 650 instruction_->IsCheckCast() || 651 (instruction_->IsInvokeVirtual() && instruction_->GetLocations()->Intrinsified()) || 652 (instruction_->IsInvokeStaticOrDirect() && instruction_->GetLocations()->Intrinsified())) 653 << "Unexpected instruction in read barrier marking slow path: " 654 << instruction_->DebugName(); 655 // The read barrier instrumentation of object ArrayGet 656 // instructions does not support the HIntermediateAddress 657 // instruction. 658 DCHECK(!(instruction_->IsArrayGet() && 659 instruction_->AsArrayGet()->GetArray()->IsIntermediateAddress())); 660 661 __ Bind(GetEntryLabel()); 662 // No need to save live registers; it's taken care of by the 663 // entrypoint. Also, there is no need to update the stack mask, 664 // as this runtime call will not trigger a garbage collection. 665 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 666 DCHECK_NE(ref_reg, SP); 667 DCHECK_NE(ref_reg, LR); 668 DCHECK_NE(ref_reg, PC); 669 // IP is used internally by the ReadBarrierMarkRegX entry point 670 // as a temporary, it cannot be the entry point's input/output. 671 DCHECK_NE(ref_reg, IP); 672 DCHECK(0 <= ref_reg && ref_reg < kNumberOfCoreRegisters) << ref_reg; 673 // "Compact" slow path, saving two moves. 674 // 675 // Instead of using the standard runtime calling convention (input 676 // and output in R0): 677 // 678 // R0 <- ref 679 // R0 <- ReadBarrierMark(R0) 680 // ref <- R0 681 // 682 // we just use rX (the register containing `ref`) as input and output 683 // of a dedicated entrypoint: 684 // 685 // rX <- ReadBarrierMarkRegX(rX) 686 // 687 if (entrypoint_.IsValid()) { 688 arm_codegen->ValidateInvokeRuntimeWithoutRecordingPcInfo(instruction_, this); 689 __ blx(entrypoint_.AsRegister<Register>()); 690 } else { 691 int32_t entry_point_offset = 692 CodeGenerator::GetReadBarrierMarkEntryPointsOffset<kArmPointerSize>(ref_reg); 693 // This runtime call does not require a stack map. 694 arm_codegen->InvokeRuntimeWithoutRecordingPcInfo(entry_point_offset, instruction_, this); 695 } 696 __ b(GetExitLabel()); 697 } 698 699 private: 700 // The location (register) of the marked object reference. 701 const Location ref_; 702 703 // The location of the entrypoint if already loaded. 704 const Location entrypoint_; 705 706 DISALLOW_COPY_AND_ASSIGN(ReadBarrierMarkSlowPathARM); 707}; 708 709// Slow path marking an object reference `ref` during a read barrier, 710// and if needed, atomically updating the field `obj.field` in the 711// object `obj` holding this reference after marking (contrary to 712// ReadBarrierMarkSlowPathARM above, which never tries to update 713// `obj.field`). 714// 715// This means that after the execution of this slow path, both `ref` 716// and `obj.field` will be up-to-date; i.e., after the flip, both will 717// hold the same to-space reference (unless another thread installed 718// another object reference (different from `ref`) in `obj.field`). 719class ReadBarrierMarkAndUpdateFieldSlowPathARM : public SlowPathCodeARM { 720 public: 721 ReadBarrierMarkAndUpdateFieldSlowPathARM(HInstruction* instruction, 722 Location ref, 723 Register obj, 724 Location field_offset, 725 Register temp1, 726 Register temp2) 727 : SlowPathCodeARM(instruction), 728 ref_(ref), 729 obj_(obj), 730 field_offset_(field_offset), 731 temp1_(temp1), 732 temp2_(temp2) { 733 DCHECK(kEmitCompilerReadBarrier); 734 } 735 736 const char* GetDescription() const OVERRIDE { return "ReadBarrierMarkAndUpdateFieldSlowPathARM"; } 737 738 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 739 LocationSummary* locations = instruction_->GetLocations(); 740 Register ref_reg = ref_.AsRegister<Register>(); 741 DCHECK(locations->CanCall()); 742 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(ref_reg)) << ref_reg; 743 // This slow path is only used by the UnsafeCASObject intrinsic. 744 DCHECK((instruction_->IsInvokeVirtual() && instruction_->GetLocations()->Intrinsified())) 745 << "Unexpected instruction in read barrier marking and field updating slow path: " 746 << instruction_->DebugName(); 747 DCHECK(instruction_->GetLocations()->Intrinsified()); 748 DCHECK_EQ(instruction_->AsInvoke()->GetIntrinsic(), Intrinsics::kUnsafeCASObject); 749 DCHECK(field_offset_.IsRegisterPair()) << field_offset_; 750 751 __ Bind(GetEntryLabel()); 752 753 // Save the old reference. 754 // Note that we cannot use IP to save the old reference, as IP is 755 // used internally by the ReadBarrierMarkRegX entry point, and we 756 // need the old reference after the call to that entry point. 757 DCHECK_NE(temp1_, IP); 758 __ Mov(temp1_, ref_reg); 759 760 // No need to save live registers; it's taken care of by the 761 // entrypoint. Also, there is no need to update the stack mask, 762 // as this runtime call will not trigger a garbage collection. 763 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 764 DCHECK_NE(ref_reg, SP); 765 DCHECK_NE(ref_reg, LR); 766 DCHECK_NE(ref_reg, PC); 767 // IP is used internally by the ReadBarrierMarkRegX entry point 768 // as a temporary, it cannot be the entry point's input/output. 769 DCHECK_NE(ref_reg, IP); 770 DCHECK(0 <= ref_reg && ref_reg < kNumberOfCoreRegisters) << ref_reg; 771 // "Compact" slow path, saving two moves. 772 // 773 // Instead of using the standard runtime calling convention (input 774 // and output in R0): 775 // 776 // R0 <- ref 777 // R0 <- ReadBarrierMark(R0) 778 // ref <- R0 779 // 780 // we just use rX (the register containing `ref`) as input and output 781 // of a dedicated entrypoint: 782 // 783 // rX <- ReadBarrierMarkRegX(rX) 784 // 785 int32_t entry_point_offset = 786 CodeGenerator::GetReadBarrierMarkEntryPointsOffset<kArmPointerSize>(ref_reg); 787 // This runtime call does not require a stack map. 788 arm_codegen->InvokeRuntimeWithoutRecordingPcInfo(entry_point_offset, instruction_, this); 789 790 // If the new reference is different from the old reference, 791 // update the field in the holder (`*(obj_ + field_offset_)`). 792 // 793 // Note that this field could also hold a different object, if 794 // another thread had concurrently changed it. In that case, the 795 // LDREX/SUBS/ITNE sequence of instructions in the compare-and-set 796 // (CAS) operation below would abort the CAS, leaving the field 797 // as-is. 798 Label done; 799 __ cmp(temp1_, ShifterOperand(ref_reg)); 800 __ b(&done, EQ); 801 802 // Update the the holder's field atomically. This may fail if 803 // mutator updates before us, but it's OK. This is achieved 804 // using a strong compare-and-set (CAS) operation with relaxed 805 // memory synchronization ordering, where the expected value is 806 // the old reference and the desired value is the new reference. 807 808 // Convenience aliases. 809 Register base = obj_; 810 // The UnsafeCASObject intrinsic uses a register pair as field 811 // offset ("long offset"), of which only the low part contains 812 // data. 813 Register offset = field_offset_.AsRegisterPairLow<Register>(); 814 Register expected = temp1_; 815 Register value = ref_reg; 816 Register tmp_ptr = IP; // Pointer to actual memory. 817 Register tmp = temp2_; // Value in memory. 818 819 __ add(tmp_ptr, base, ShifterOperand(offset)); 820 821 if (kPoisonHeapReferences) { 822 __ PoisonHeapReference(expected); 823 if (value == expected) { 824 // Do not poison `value`, as it is the same register as 825 // `expected`, which has just been poisoned. 826 } else { 827 __ PoisonHeapReference(value); 828 } 829 } 830 831 // do { 832 // tmp = [r_ptr] - expected; 833 // } while (tmp == 0 && failure([r_ptr] <- r_new_value)); 834 835 Label loop_head, exit_loop; 836 __ Bind(&loop_head); 837 838 __ ldrex(tmp, tmp_ptr); 839 840 __ subs(tmp, tmp, ShifterOperand(expected)); 841 842 __ it(NE); 843 __ clrex(NE); 844 845 __ b(&exit_loop, NE); 846 847 __ strex(tmp, value, tmp_ptr); 848 __ cmp(tmp, ShifterOperand(1)); 849 __ b(&loop_head, EQ); 850 851 __ Bind(&exit_loop); 852 853 if (kPoisonHeapReferences) { 854 __ UnpoisonHeapReference(expected); 855 if (value == expected) { 856 // Do not unpoison `value`, as it is the same register as 857 // `expected`, which has just been unpoisoned. 858 } else { 859 __ UnpoisonHeapReference(value); 860 } 861 } 862 863 __ Bind(&done); 864 __ b(GetExitLabel()); 865 } 866 867 private: 868 // The location (register) of the marked object reference. 869 const Location ref_; 870 // The register containing the object holding the marked object reference field. 871 const Register obj_; 872 // The location of the offset of the marked reference field within `obj_`. 873 Location field_offset_; 874 875 const Register temp1_; 876 const Register temp2_; 877 878 DISALLOW_COPY_AND_ASSIGN(ReadBarrierMarkAndUpdateFieldSlowPathARM); 879}; 880 881// Slow path generating a read barrier for a heap reference. 882class ReadBarrierForHeapReferenceSlowPathARM : public SlowPathCodeARM { 883 public: 884 ReadBarrierForHeapReferenceSlowPathARM(HInstruction* instruction, 885 Location out, 886 Location ref, 887 Location obj, 888 uint32_t offset, 889 Location index) 890 : SlowPathCodeARM(instruction), 891 out_(out), 892 ref_(ref), 893 obj_(obj), 894 offset_(offset), 895 index_(index) { 896 DCHECK(kEmitCompilerReadBarrier); 897 // If `obj` is equal to `out` or `ref`, it means the initial object 898 // has been overwritten by (or after) the heap object reference load 899 // to be instrumented, e.g.: 900 // 901 // __ LoadFromOffset(kLoadWord, out, out, offset); 902 // codegen_->GenerateReadBarrierSlow(instruction, out_loc, out_loc, out_loc, offset); 903 // 904 // In that case, we have lost the information about the original 905 // object, and the emitted read barrier cannot work properly. 906 DCHECK(!obj.Equals(out)) << "obj=" << obj << " out=" << out; 907 DCHECK(!obj.Equals(ref)) << "obj=" << obj << " ref=" << ref; 908 } 909 910 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 911 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 912 LocationSummary* locations = instruction_->GetLocations(); 913 Register reg_out = out_.AsRegister<Register>(); 914 DCHECK(locations->CanCall()); 915 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg_out)); 916 DCHECK(instruction_->IsInstanceFieldGet() || 917 instruction_->IsStaticFieldGet() || 918 instruction_->IsArrayGet() || 919 instruction_->IsInstanceOf() || 920 instruction_->IsCheckCast() || 921 (instruction_->IsInvokeVirtual()) && instruction_->GetLocations()->Intrinsified()) 922 << "Unexpected instruction in read barrier for heap reference slow path: " 923 << instruction_->DebugName(); 924 // The read barrier instrumentation of object ArrayGet 925 // instructions does not support the HIntermediateAddress 926 // instruction. 927 DCHECK(!(instruction_->IsArrayGet() && 928 instruction_->AsArrayGet()->GetArray()->IsIntermediateAddress())); 929 930 __ Bind(GetEntryLabel()); 931 SaveLiveRegisters(codegen, locations); 932 933 // We may have to change the index's value, but as `index_` is a 934 // constant member (like other "inputs" of this slow path), 935 // introduce a copy of it, `index`. 936 Location index = index_; 937 if (index_.IsValid()) { 938 // Handle `index_` for HArrayGet and UnsafeGetObject/UnsafeGetObjectVolatile intrinsics. 939 if (instruction_->IsArrayGet()) { 940 // Compute the actual memory offset and store it in `index`. 941 Register index_reg = index_.AsRegister<Register>(); 942 DCHECK(locations->GetLiveRegisters()->ContainsCoreRegister(index_reg)); 943 if (codegen->IsCoreCalleeSaveRegister(index_reg)) { 944 // We are about to change the value of `index_reg` (see the 945 // calls to art::arm::Thumb2Assembler::Lsl and 946 // art::arm::Thumb2Assembler::AddConstant below), but it has 947 // not been saved by the previous call to 948 // art::SlowPathCode::SaveLiveRegisters, as it is a 949 // callee-save register -- 950 // art::SlowPathCode::SaveLiveRegisters does not consider 951 // callee-save registers, as it has been designed with the 952 // assumption that callee-save registers are supposed to be 953 // handled by the called function. So, as a callee-save 954 // register, `index_reg` _would_ eventually be saved onto 955 // the stack, but it would be too late: we would have 956 // changed its value earlier. Therefore, we manually save 957 // it here into another freely available register, 958 // `free_reg`, chosen of course among the caller-save 959 // registers (as a callee-save `free_reg` register would 960 // exhibit the same problem). 961 // 962 // Note we could have requested a temporary register from 963 // the register allocator instead; but we prefer not to, as 964 // this is a slow path, and we know we can find a 965 // caller-save register that is available. 966 Register free_reg = FindAvailableCallerSaveRegister(codegen); 967 __ Mov(free_reg, index_reg); 968 index_reg = free_reg; 969 index = Location::RegisterLocation(index_reg); 970 } else { 971 // The initial register stored in `index_` has already been 972 // saved in the call to art::SlowPathCode::SaveLiveRegisters 973 // (as it is not a callee-save register), so we can freely 974 // use it. 975 } 976 // Shifting the index value contained in `index_reg` by the scale 977 // factor (2) cannot overflow in practice, as the runtime is 978 // unable to allocate object arrays with a size larger than 979 // 2^26 - 1 (that is, 2^28 - 4 bytes). 980 __ Lsl(index_reg, index_reg, TIMES_4); 981 static_assert( 982 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 983 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 984 __ AddConstant(index_reg, index_reg, offset_); 985 } else { 986 // In the case of the UnsafeGetObject/UnsafeGetObjectVolatile 987 // intrinsics, `index_` is not shifted by a scale factor of 2 988 // (as in the case of ArrayGet), as it is actually an offset 989 // to an object field within an object. 990 DCHECK(instruction_->IsInvoke()) << instruction_->DebugName(); 991 DCHECK(instruction_->GetLocations()->Intrinsified()); 992 DCHECK((instruction_->AsInvoke()->GetIntrinsic() == Intrinsics::kUnsafeGetObject) || 993 (instruction_->AsInvoke()->GetIntrinsic() == Intrinsics::kUnsafeGetObjectVolatile)) 994 << instruction_->AsInvoke()->GetIntrinsic(); 995 DCHECK_EQ(offset_, 0U); 996 DCHECK(index_.IsRegisterPair()); 997 // UnsafeGet's offset location is a register pair, the low 998 // part contains the correct offset. 999 index = index_.ToLow(); 1000 } 1001 } 1002 1003 // We're moving two or three locations to locations that could 1004 // overlap, so we need a parallel move resolver. 1005 InvokeRuntimeCallingConvention calling_convention; 1006 HParallelMove parallel_move(codegen->GetGraph()->GetArena()); 1007 parallel_move.AddMove(ref_, 1008 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 1009 Primitive::kPrimNot, 1010 nullptr); 1011 parallel_move.AddMove(obj_, 1012 Location::RegisterLocation(calling_convention.GetRegisterAt(1)), 1013 Primitive::kPrimNot, 1014 nullptr); 1015 if (index.IsValid()) { 1016 parallel_move.AddMove(index, 1017 Location::RegisterLocation(calling_convention.GetRegisterAt(2)), 1018 Primitive::kPrimInt, 1019 nullptr); 1020 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 1021 } else { 1022 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 1023 __ LoadImmediate(calling_convention.GetRegisterAt(2), offset_); 1024 } 1025 arm_codegen->InvokeRuntime(kQuickReadBarrierSlow, instruction_, instruction_->GetDexPc(), this); 1026 CheckEntrypointTypes< 1027 kQuickReadBarrierSlow, mirror::Object*, mirror::Object*, mirror::Object*, uint32_t>(); 1028 arm_codegen->Move32(out_, Location::RegisterLocation(R0)); 1029 1030 RestoreLiveRegisters(codegen, locations); 1031 __ b(GetExitLabel()); 1032 } 1033 1034 const char* GetDescription() const OVERRIDE { return "ReadBarrierForHeapReferenceSlowPathARM"; } 1035 1036 private: 1037 Register FindAvailableCallerSaveRegister(CodeGenerator* codegen) { 1038 size_t ref = static_cast<int>(ref_.AsRegister<Register>()); 1039 size_t obj = static_cast<int>(obj_.AsRegister<Register>()); 1040 for (size_t i = 0, e = codegen->GetNumberOfCoreRegisters(); i < e; ++i) { 1041 if (i != ref && i != obj && !codegen->IsCoreCalleeSaveRegister(i)) { 1042 return static_cast<Register>(i); 1043 } 1044 } 1045 // We shall never fail to find a free caller-save register, as 1046 // there are more than two core caller-save registers on ARM 1047 // (meaning it is possible to find one which is different from 1048 // `ref` and `obj`). 1049 DCHECK_GT(codegen->GetNumberOfCoreCallerSaveRegisters(), 2u); 1050 LOG(FATAL) << "Could not find a free caller-save register"; 1051 UNREACHABLE(); 1052 } 1053 1054 const Location out_; 1055 const Location ref_; 1056 const Location obj_; 1057 const uint32_t offset_; 1058 // An additional location containing an index to an array. 1059 // Only used for HArrayGet and the UnsafeGetObject & 1060 // UnsafeGetObjectVolatile intrinsics. 1061 const Location index_; 1062 1063 DISALLOW_COPY_AND_ASSIGN(ReadBarrierForHeapReferenceSlowPathARM); 1064}; 1065 1066// Slow path generating a read barrier for a GC root. 1067class ReadBarrierForRootSlowPathARM : public SlowPathCodeARM { 1068 public: 1069 ReadBarrierForRootSlowPathARM(HInstruction* instruction, Location out, Location root) 1070 : SlowPathCodeARM(instruction), out_(out), root_(root) { 1071 DCHECK(kEmitCompilerReadBarrier); 1072 } 1073 1074 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 1075 LocationSummary* locations = instruction_->GetLocations(); 1076 Register reg_out = out_.AsRegister<Register>(); 1077 DCHECK(locations->CanCall()); 1078 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(reg_out)); 1079 DCHECK(instruction_->IsLoadClass() || instruction_->IsLoadString()) 1080 << "Unexpected instruction in read barrier for GC root slow path: " 1081 << instruction_->DebugName(); 1082 1083 __ Bind(GetEntryLabel()); 1084 SaveLiveRegisters(codegen, locations); 1085 1086 InvokeRuntimeCallingConvention calling_convention; 1087 CodeGeneratorARM* arm_codegen = down_cast<CodeGeneratorARM*>(codegen); 1088 arm_codegen->Move32(Location::RegisterLocation(calling_convention.GetRegisterAt(0)), root_); 1089 arm_codegen->InvokeRuntime(kQuickReadBarrierForRootSlow, 1090 instruction_, 1091 instruction_->GetDexPc(), 1092 this); 1093 CheckEntrypointTypes<kQuickReadBarrierForRootSlow, mirror::Object*, GcRoot<mirror::Object>*>(); 1094 arm_codegen->Move32(out_, Location::RegisterLocation(R0)); 1095 1096 RestoreLiveRegisters(codegen, locations); 1097 __ b(GetExitLabel()); 1098 } 1099 1100 const char* GetDescription() const OVERRIDE { return "ReadBarrierForRootSlowPathARM"; } 1101 1102 private: 1103 const Location out_; 1104 const Location root_; 1105 1106 DISALLOW_COPY_AND_ASSIGN(ReadBarrierForRootSlowPathARM); 1107}; 1108 1109#undef __ 1110// NOLINT on __ macro to suppress wrong warning/fix (misc-macro-parentheses) from clang-tidy. 1111#define __ down_cast<ArmAssembler*>(GetAssembler())-> // NOLINT 1112 1113inline Condition ARMCondition(IfCondition cond) { 1114 switch (cond) { 1115 case kCondEQ: return EQ; 1116 case kCondNE: return NE; 1117 case kCondLT: return LT; 1118 case kCondLE: return LE; 1119 case kCondGT: return GT; 1120 case kCondGE: return GE; 1121 case kCondB: return LO; 1122 case kCondBE: return LS; 1123 case kCondA: return HI; 1124 case kCondAE: return HS; 1125 } 1126 LOG(FATAL) << "Unreachable"; 1127 UNREACHABLE(); 1128} 1129 1130// Maps signed condition to unsigned condition. 1131inline Condition ARMUnsignedCondition(IfCondition cond) { 1132 switch (cond) { 1133 case kCondEQ: return EQ; 1134 case kCondNE: return NE; 1135 // Signed to unsigned. 1136 case kCondLT: return LO; 1137 case kCondLE: return LS; 1138 case kCondGT: return HI; 1139 case kCondGE: return HS; 1140 // Unsigned remain unchanged. 1141 case kCondB: return LO; 1142 case kCondBE: return LS; 1143 case kCondA: return HI; 1144 case kCondAE: return HS; 1145 } 1146 LOG(FATAL) << "Unreachable"; 1147 UNREACHABLE(); 1148} 1149 1150inline Condition ARMFPCondition(IfCondition cond, bool gt_bias) { 1151 // The ARM condition codes can express all the necessary branches, see the 1152 // "Meaning (floating-point)" column in the table A8-1 of the ARMv7 reference manual. 1153 // There is no dex instruction or HIR that would need the missing conditions 1154 // "equal or unordered" or "not equal". 1155 switch (cond) { 1156 case kCondEQ: return EQ; 1157 case kCondNE: return NE /* unordered */; 1158 case kCondLT: return gt_bias ? CC : LT /* unordered */; 1159 case kCondLE: return gt_bias ? LS : LE /* unordered */; 1160 case kCondGT: return gt_bias ? HI /* unordered */ : GT; 1161 case kCondGE: return gt_bias ? CS /* unordered */ : GE; 1162 default: 1163 LOG(FATAL) << "UNREACHABLE"; 1164 UNREACHABLE(); 1165 } 1166} 1167 1168void CodeGeneratorARM::DumpCoreRegister(std::ostream& stream, int reg) const { 1169 stream << Register(reg); 1170} 1171 1172void CodeGeneratorARM::DumpFloatingPointRegister(std::ostream& stream, int reg) const { 1173 stream << SRegister(reg); 1174} 1175 1176size_t CodeGeneratorARM::SaveCoreRegister(size_t stack_index, uint32_t reg_id) { 1177 __ StoreToOffset(kStoreWord, static_cast<Register>(reg_id), SP, stack_index); 1178 return kArmWordSize; 1179} 1180 1181size_t CodeGeneratorARM::RestoreCoreRegister(size_t stack_index, uint32_t reg_id) { 1182 __ LoadFromOffset(kLoadWord, static_cast<Register>(reg_id), SP, stack_index); 1183 return kArmWordSize; 1184} 1185 1186size_t CodeGeneratorARM::SaveFloatingPointRegister(size_t stack_index, uint32_t reg_id) { 1187 __ StoreSToOffset(static_cast<SRegister>(reg_id), SP, stack_index); 1188 return kArmWordSize; 1189} 1190 1191size_t CodeGeneratorARM::RestoreFloatingPointRegister(size_t stack_index, uint32_t reg_id) { 1192 __ LoadSFromOffset(static_cast<SRegister>(reg_id), SP, stack_index); 1193 return kArmWordSize; 1194} 1195 1196CodeGeneratorARM::CodeGeneratorARM(HGraph* graph, 1197 const ArmInstructionSetFeatures& isa_features, 1198 const CompilerOptions& compiler_options, 1199 OptimizingCompilerStats* stats) 1200 : CodeGenerator(graph, 1201 kNumberOfCoreRegisters, 1202 kNumberOfSRegisters, 1203 kNumberOfRegisterPairs, 1204 ComputeRegisterMask(reinterpret_cast<const int*>(kCoreCalleeSaves), 1205 arraysize(kCoreCalleeSaves)), 1206 ComputeRegisterMask(reinterpret_cast<const int*>(kFpuCalleeSaves), 1207 arraysize(kFpuCalleeSaves)), 1208 compiler_options, 1209 stats), 1210 block_labels_(nullptr), 1211 location_builder_(graph, this), 1212 instruction_visitor_(graph, this), 1213 move_resolver_(graph->GetArena(), this), 1214 assembler_(graph->GetArena()), 1215 isa_features_(isa_features), 1216 uint32_literals_(std::less<uint32_t>(), 1217 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 1218 pc_relative_dex_cache_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 1219 boot_image_string_patches_(StringReferenceValueComparator(), 1220 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 1221 pc_relative_string_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 1222 boot_image_type_patches_(TypeReferenceValueComparator(), 1223 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 1224 pc_relative_type_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 1225 type_bss_entry_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 1226 boot_image_address_patches_(std::less<uint32_t>(), 1227 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 1228 jit_string_patches_(StringReferenceValueComparator(), 1229 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 1230 jit_class_patches_(TypeReferenceValueComparator(), 1231 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)) { 1232 // Always save the LR register to mimic Quick. 1233 AddAllocatedRegister(Location::RegisterLocation(LR)); 1234} 1235 1236void CodeGeneratorARM::Finalize(CodeAllocator* allocator) { 1237 // Ensure that we fix up branches and literal loads and emit the literal pool. 1238 __ FinalizeCode(); 1239 1240 // Adjust native pc offsets in stack maps. 1241 for (size_t i = 0, num = stack_map_stream_.GetNumberOfStackMaps(); i != num; ++i) { 1242 uint32_t old_position = 1243 stack_map_stream_.GetStackMap(i).native_pc_code_offset.Uint32Value(kThumb2); 1244 uint32_t new_position = __ GetAdjustedPosition(old_position); 1245 stack_map_stream_.SetStackMapNativePcOffset(i, new_position); 1246 } 1247 // Adjust pc offsets for the disassembly information. 1248 if (disasm_info_ != nullptr) { 1249 GeneratedCodeInterval* frame_entry_interval = disasm_info_->GetFrameEntryInterval(); 1250 frame_entry_interval->start = __ GetAdjustedPosition(frame_entry_interval->start); 1251 frame_entry_interval->end = __ GetAdjustedPosition(frame_entry_interval->end); 1252 for (auto& it : *disasm_info_->GetInstructionIntervals()) { 1253 it.second.start = __ GetAdjustedPosition(it.second.start); 1254 it.second.end = __ GetAdjustedPosition(it.second.end); 1255 } 1256 for (auto& it : *disasm_info_->GetSlowPathIntervals()) { 1257 it.code_interval.start = __ GetAdjustedPosition(it.code_interval.start); 1258 it.code_interval.end = __ GetAdjustedPosition(it.code_interval.end); 1259 } 1260 } 1261 1262 CodeGenerator::Finalize(allocator); 1263} 1264 1265void CodeGeneratorARM::SetupBlockedRegisters() const { 1266 // Stack register, LR and PC are always reserved. 1267 blocked_core_registers_[SP] = true; 1268 blocked_core_registers_[LR] = true; 1269 blocked_core_registers_[PC] = true; 1270 1271 // Reserve thread register. 1272 blocked_core_registers_[TR] = true; 1273 1274 // Reserve temp register. 1275 blocked_core_registers_[IP] = true; 1276 1277 if (GetGraph()->IsDebuggable()) { 1278 // Stubs do not save callee-save floating point registers. If the graph 1279 // is debuggable, we need to deal with these registers differently. For 1280 // now, just block them. 1281 for (size_t i = 0; i < arraysize(kFpuCalleeSaves); ++i) { 1282 blocked_fpu_registers_[kFpuCalleeSaves[i]] = true; 1283 } 1284 } 1285} 1286 1287InstructionCodeGeneratorARM::InstructionCodeGeneratorARM(HGraph* graph, CodeGeneratorARM* codegen) 1288 : InstructionCodeGenerator(graph, codegen), 1289 assembler_(codegen->GetAssembler()), 1290 codegen_(codegen) {} 1291 1292void CodeGeneratorARM::ComputeSpillMask() { 1293 core_spill_mask_ = allocated_registers_.GetCoreRegisters() & core_callee_save_mask_; 1294 DCHECK_NE(core_spill_mask_, 0u) << "At least the return address register must be saved"; 1295 // There is no easy instruction to restore just the PC on thumb2. We spill and 1296 // restore another arbitrary register. 1297 core_spill_mask_ |= (1 << kCoreAlwaysSpillRegister); 1298 fpu_spill_mask_ = allocated_registers_.GetFloatingPointRegisters() & fpu_callee_save_mask_; 1299 // We use vpush and vpop for saving and restoring floating point registers, which take 1300 // a SRegister and the number of registers to save/restore after that SRegister. We 1301 // therefore update the `fpu_spill_mask_` to also contain those registers not allocated, 1302 // but in the range. 1303 if (fpu_spill_mask_ != 0) { 1304 uint32_t least_significant_bit = LeastSignificantBit(fpu_spill_mask_); 1305 uint32_t most_significant_bit = MostSignificantBit(fpu_spill_mask_); 1306 for (uint32_t i = least_significant_bit + 1 ; i < most_significant_bit; ++i) { 1307 fpu_spill_mask_ |= (1 << i); 1308 } 1309 } 1310} 1311 1312static dwarf::Reg DWARFReg(Register reg) { 1313 return dwarf::Reg::ArmCore(static_cast<int>(reg)); 1314} 1315 1316static dwarf::Reg DWARFReg(SRegister reg) { 1317 return dwarf::Reg::ArmFp(static_cast<int>(reg)); 1318} 1319 1320void CodeGeneratorARM::GenerateFrameEntry() { 1321 bool skip_overflow_check = 1322 IsLeafMethod() && !FrameNeedsStackCheck(GetFrameSize(), InstructionSet::kArm); 1323 DCHECK(GetCompilerOptions().GetImplicitStackOverflowChecks()); 1324 __ Bind(&frame_entry_label_); 1325 1326 if (HasEmptyFrame()) { 1327 return; 1328 } 1329 1330 if (!skip_overflow_check) { 1331 __ AddConstant(IP, SP, -static_cast<int32_t>(GetStackOverflowReservedBytes(kArm))); 1332 __ LoadFromOffset(kLoadWord, IP, IP, 0); 1333 RecordPcInfo(nullptr, 0); 1334 } 1335 1336 __ PushList(core_spill_mask_); 1337 __ cfi().AdjustCFAOffset(kArmWordSize * POPCOUNT(core_spill_mask_)); 1338 __ cfi().RelOffsetForMany(DWARFReg(kMethodRegisterArgument), 0, core_spill_mask_, kArmWordSize); 1339 if (fpu_spill_mask_ != 0) { 1340 SRegister start_register = SRegister(LeastSignificantBit(fpu_spill_mask_)); 1341 __ vpushs(start_register, POPCOUNT(fpu_spill_mask_)); 1342 __ cfi().AdjustCFAOffset(kArmWordSize * POPCOUNT(fpu_spill_mask_)); 1343 __ cfi().RelOffsetForMany(DWARFReg(S0), 0, fpu_spill_mask_, kArmWordSize); 1344 } 1345 1346 if (GetGraph()->HasShouldDeoptimizeFlag()) { 1347 // Initialize should_deoptimize flag to 0. 1348 __ mov(IP, ShifterOperand(0)); 1349 __ StoreToOffset(kStoreWord, IP, SP, -kShouldDeoptimizeFlagSize); 1350 } 1351 1352 int adjust = GetFrameSize() - FrameEntrySpillSize(); 1353 __ AddConstant(SP, -adjust); 1354 __ cfi().AdjustCFAOffset(adjust); 1355 1356 // Save the current method if we need it. Note that we do not 1357 // do this in HCurrentMethod, as the instruction might have been removed 1358 // in the SSA graph. 1359 if (RequiresCurrentMethod()) { 1360 __ StoreToOffset(kStoreWord, kMethodRegisterArgument, SP, 0); 1361 } 1362} 1363 1364void CodeGeneratorARM::GenerateFrameExit() { 1365 if (HasEmptyFrame()) { 1366 __ bx(LR); 1367 return; 1368 } 1369 __ cfi().RememberState(); 1370 int adjust = GetFrameSize() - FrameEntrySpillSize(); 1371 __ AddConstant(SP, adjust); 1372 __ cfi().AdjustCFAOffset(-adjust); 1373 if (fpu_spill_mask_ != 0) { 1374 SRegister start_register = SRegister(LeastSignificantBit(fpu_spill_mask_)); 1375 __ vpops(start_register, POPCOUNT(fpu_spill_mask_)); 1376 __ cfi().AdjustCFAOffset(-static_cast<int>(kArmPointerSize) * POPCOUNT(fpu_spill_mask_)); 1377 __ cfi().RestoreMany(DWARFReg(SRegister(0)), fpu_spill_mask_); 1378 } 1379 // Pop LR into PC to return. 1380 DCHECK_NE(core_spill_mask_ & (1 << LR), 0U); 1381 uint32_t pop_mask = (core_spill_mask_ & (~(1 << LR))) | 1 << PC; 1382 __ PopList(pop_mask); 1383 __ cfi().RestoreState(); 1384 __ cfi().DefCFAOffset(GetFrameSize()); 1385} 1386 1387void CodeGeneratorARM::Bind(HBasicBlock* block) { 1388 Label* label = GetLabelOf(block); 1389 __ BindTrackedLabel(label); 1390} 1391 1392Location InvokeDexCallingConventionVisitorARM::GetNextLocation(Primitive::Type type) { 1393 switch (type) { 1394 case Primitive::kPrimBoolean: 1395 case Primitive::kPrimByte: 1396 case Primitive::kPrimChar: 1397 case Primitive::kPrimShort: 1398 case Primitive::kPrimInt: 1399 case Primitive::kPrimNot: { 1400 uint32_t index = gp_index_++; 1401 uint32_t stack_index = stack_index_++; 1402 if (index < calling_convention.GetNumberOfRegisters()) { 1403 return Location::RegisterLocation(calling_convention.GetRegisterAt(index)); 1404 } else { 1405 return Location::StackSlot(calling_convention.GetStackOffsetOf(stack_index)); 1406 } 1407 } 1408 1409 case Primitive::kPrimLong: { 1410 uint32_t index = gp_index_; 1411 uint32_t stack_index = stack_index_; 1412 gp_index_ += 2; 1413 stack_index_ += 2; 1414 if (index + 1 < calling_convention.GetNumberOfRegisters()) { 1415 if (calling_convention.GetRegisterAt(index) == R1) { 1416 // Skip R1, and use R2_R3 instead. 1417 gp_index_++; 1418 index++; 1419 } 1420 } 1421 if (index + 1 < calling_convention.GetNumberOfRegisters()) { 1422 DCHECK_EQ(calling_convention.GetRegisterAt(index) + 1, 1423 calling_convention.GetRegisterAt(index + 1)); 1424 1425 return Location::RegisterPairLocation(calling_convention.GetRegisterAt(index), 1426 calling_convention.GetRegisterAt(index + 1)); 1427 } else { 1428 return Location::DoubleStackSlot(calling_convention.GetStackOffsetOf(stack_index)); 1429 } 1430 } 1431 1432 case Primitive::kPrimFloat: { 1433 uint32_t stack_index = stack_index_++; 1434 if (float_index_ % 2 == 0) { 1435 float_index_ = std::max(double_index_, float_index_); 1436 } 1437 if (float_index_ < calling_convention.GetNumberOfFpuRegisters()) { 1438 return Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(float_index_++)); 1439 } else { 1440 return Location::StackSlot(calling_convention.GetStackOffsetOf(stack_index)); 1441 } 1442 } 1443 1444 case Primitive::kPrimDouble: { 1445 double_index_ = std::max(double_index_, RoundUp(float_index_, 2)); 1446 uint32_t stack_index = stack_index_; 1447 stack_index_ += 2; 1448 if (double_index_ + 1 < calling_convention.GetNumberOfFpuRegisters()) { 1449 uint32_t index = double_index_; 1450 double_index_ += 2; 1451 Location result = Location::FpuRegisterPairLocation( 1452 calling_convention.GetFpuRegisterAt(index), 1453 calling_convention.GetFpuRegisterAt(index + 1)); 1454 DCHECK(ExpectedPairLayout(result)); 1455 return result; 1456 } else { 1457 return Location::DoubleStackSlot(calling_convention.GetStackOffsetOf(stack_index)); 1458 } 1459 } 1460 1461 case Primitive::kPrimVoid: 1462 LOG(FATAL) << "Unexpected parameter type " << type; 1463 break; 1464 } 1465 return Location::NoLocation(); 1466} 1467 1468Location InvokeDexCallingConventionVisitorARM::GetReturnLocation(Primitive::Type type) const { 1469 switch (type) { 1470 case Primitive::kPrimBoolean: 1471 case Primitive::kPrimByte: 1472 case Primitive::kPrimChar: 1473 case Primitive::kPrimShort: 1474 case Primitive::kPrimInt: 1475 case Primitive::kPrimNot: { 1476 return Location::RegisterLocation(R0); 1477 } 1478 1479 case Primitive::kPrimFloat: { 1480 return Location::FpuRegisterLocation(S0); 1481 } 1482 1483 case Primitive::kPrimLong: { 1484 return Location::RegisterPairLocation(R0, R1); 1485 } 1486 1487 case Primitive::kPrimDouble: { 1488 return Location::FpuRegisterPairLocation(S0, S1); 1489 } 1490 1491 case Primitive::kPrimVoid: 1492 return Location::NoLocation(); 1493 } 1494 1495 UNREACHABLE(); 1496} 1497 1498Location InvokeDexCallingConventionVisitorARM::GetMethodLocation() const { 1499 return Location::RegisterLocation(kMethodRegisterArgument); 1500} 1501 1502void CodeGeneratorARM::Move32(Location destination, Location source) { 1503 if (source.Equals(destination)) { 1504 return; 1505 } 1506 if (destination.IsRegister()) { 1507 if (source.IsRegister()) { 1508 __ Mov(destination.AsRegister<Register>(), source.AsRegister<Register>()); 1509 } else if (source.IsFpuRegister()) { 1510 __ vmovrs(destination.AsRegister<Register>(), source.AsFpuRegister<SRegister>()); 1511 } else { 1512 __ LoadFromOffset(kLoadWord, destination.AsRegister<Register>(), SP, source.GetStackIndex()); 1513 } 1514 } else if (destination.IsFpuRegister()) { 1515 if (source.IsRegister()) { 1516 __ vmovsr(destination.AsFpuRegister<SRegister>(), source.AsRegister<Register>()); 1517 } else if (source.IsFpuRegister()) { 1518 __ vmovs(destination.AsFpuRegister<SRegister>(), source.AsFpuRegister<SRegister>()); 1519 } else { 1520 __ LoadSFromOffset(destination.AsFpuRegister<SRegister>(), SP, source.GetStackIndex()); 1521 } 1522 } else { 1523 DCHECK(destination.IsStackSlot()) << destination; 1524 if (source.IsRegister()) { 1525 __ StoreToOffset(kStoreWord, source.AsRegister<Register>(), SP, destination.GetStackIndex()); 1526 } else if (source.IsFpuRegister()) { 1527 __ StoreSToOffset(source.AsFpuRegister<SRegister>(), SP, destination.GetStackIndex()); 1528 } else { 1529 DCHECK(source.IsStackSlot()) << source; 1530 __ LoadFromOffset(kLoadWord, IP, SP, source.GetStackIndex()); 1531 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 1532 } 1533 } 1534} 1535 1536void CodeGeneratorARM::Move64(Location destination, Location source) { 1537 if (source.Equals(destination)) { 1538 return; 1539 } 1540 if (destination.IsRegisterPair()) { 1541 if (source.IsRegisterPair()) { 1542 EmitParallelMoves( 1543 Location::RegisterLocation(source.AsRegisterPairHigh<Register>()), 1544 Location::RegisterLocation(destination.AsRegisterPairHigh<Register>()), 1545 Primitive::kPrimInt, 1546 Location::RegisterLocation(source.AsRegisterPairLow<Register>()), 1547 Location::RegisterLocation(destination.AsRegisterPairLow<Register>()), 1548 Primitive::kPrimInt); 1549 } else if (source.IsFpuRegister()) { 1550 UNIMPLEMENTED(FATAL); 1551 } else if (source.IsFpuRegisterPair()) { 1552 __ vmovrrd(destination.AsRegisterPairLow<Register>(), 1553 destination.AsRegisterPairHigh<Register>(), 1554 FromLowSToD(source.AsFpuRegisterPairLow<SRegister>())); 1555 } else { 1556 DCHECK(source.IsDoubleStackSlot()); 1557 DCHECK(ExpectedPairLayout(destination)); 1558 __ LoadFromOffset(kLoadWordPair, destination.AsRegisterPairLow<Register>(), 1559 SP, source.GetStackIndex()); 1560 } 1561 } else if (destination.IsFpuRegisterPair()) { 1562 if (source.IsDoubleStackSlot()) { 1563 __ LoadDFromOffset(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 1564 SP, 1565 source.GetStackIndex()); 1566 } else if (source.IsRegisterPair()) { 1567 __ vmovdrr(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 1568 source.AsRegisterPairLow<Register>(), 1569 source.AsRegisterPairHigh<Register>()); 1570 } else { 1571 UNIMPLEMENTED(FATAL); 1572 } 1573 } else { 1574 DCHECK(destination.IsDoubleStackSlot()); 1575 if (source.IsRegisterPair()) { 1576 // No conflict possible, so just do the moves. 1577 if (source.AsRegisterPairLow<Register>() == R1) { 1578 DCHECK_EQ(source.AsRegisterPairHigh<Register>(), R2); 1579 __ StoreToOffset(kStoreWord, R1, SP, destination.GetStackIndex()); 1580 __ StoreToOffset(kStoreWord, R2, SP, destination.GetHighStackIndex(kArmWordSize)); 1581 } else { 1582 __ StoreToOffset(kStoreWordPair, source.AsRegisterPairLow<Register>(), 1583 SP, destination.GetStackIndex()); 1584 } 1585 } else if (source.IsFpuRegisterPair()) { 1586 __ StoreDToOffset(FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()), 1587 SP, 1588 destination.GetStackIndex()); 1589 } else { 1590 DCHECK(source.IsDoubleStackSlot()); 1591 EmitParallelMoves( 1592 Location::StackSlot(source.GetStackIndex()), 1593 Location::StackSlot(destination.GetStackIndex()), 1594 Primitive::kPrimInt, 1595 Location::StackSlot(source.GetHighStackIndex(kArmWordSize)), 1596 Location::StackSlot(destination.GetHighStackIndex(kArmWordSize)), 1597 Primitive::kPrimInt); 1598 } 1599 } 1600} 1601 1602void CodeGeneratorARM::MoveConstant(Location location, int32_t value) { 1603 DCHECK(location.IsRegister()); 1604 __ LoadImmediate(location.AsRegister<Register>(), value); 1605} 1606 1607void CodeGeneratorARM::MoveLocation(Location dst, Location src, Primitive::Type dst_type) { 1608 HParallelMove move(GetGraph()->GetArena()); 1609 move.AddMove(src, dst, dst_type, nullptr); 1610 GetMoveResolver()->EmitNativeCode(&move); 1611} 1612 1613void CodeGeneratorARM::AddLocationAsTemp(Location location, LocationSummary* locations) { 1614 if (location.IsRegister()) { 1615 locations->AddTemp(location); 1616 } else if (location.IsRegisterPair()) { 1617 locations->AddTemp(Location::RegisterLocation(location.AsRegisterPairLow<Register>())); 1618 locations->AddTemp(Location::RegisterLocation(location.AsRegisterPairHigh<Register>())); 1619 } else { 1620 UNIMPLEMENTED(FATAL) << "AddLocationAsTemp not implemented for location " << location; 1621 } 1622} 1623 1624void CodeGeneratorARM::InvokeRuntime(QuickEntrypointEnum entrypoint, 1625 HInstruction* instruction, 1626 uint32_t dex_pc, 1627 SlowPathCode* slow_path) { 1628 ValidateInvokeRuntime(entrypoint, instruction, slow_path); 1629 GenerateInvokeRuntime(GetThreadOffset<kArmPointerSize>(entrypoint).Int32Value()); 1630 if (EntrypointRequiresStackMap(entrypoint)) { 1631 RecordPcInfo(instruction, dex_pc, slow_path); 1632 } 1633} 1634 1635void CodeGeneratorARM::InvokeRuntimeWithoutRecordingPcInfo(int32_t entry_point_offset, 1636 HInstruction* instruction, 1637 SlowPathCode* slow_path) { 1638 ValidateInvokeRuntimeWithoutRecordingPcInfo(instruction, slow_path); 1639 GenerateInvokeRuntime(entry_point_offset); 1640} 1641 1642void CodeGeneratorARM::GenerateInvokeRuntime(int32_t entry_point_offset) { 1643 __ LoadFromOffset(kLoadWord, LR, TR, entry_point_offset); 1644 __ blx(LR); 1645} 1646 1647void InstructionCodeGeneratorARM::HandleGoto(HInstruction* got, HBasicBlock* successor) { 1648 DCHECK(!successor->IsExitBlock()); 1649 1650 HBasicBlock* block = got->GetBlock(); 1651 HInstruction* previous = got->GetPrevious(); 1652 1653 HLoopInformation* info = block->GetLoopInformation(); 1654 if (info != nullptr && info->IsBackEdge(*block) && info->HasSuspendCheck()) { 1655 codegen_->ClearSpillSlotsFromLoopPhisInStackMap(info->GetSuspendCheck()); 1656 GenerateSuspendCheck(info->GetSuspendCheck(), successor); 1657 return; 1658 } 1659 1660 if (block->IsEntryBlock() && (previous != nullptr) && previous->IsSuspendCheck()) { 1661 GenerateSuspendCheck(previous->AsSuspendCheck(), nullptr); 1662 } 1663 if (!codegen_->GoesToNextBlock(got->GetBlock(), successor)) { 1664 __ b(codegen_->GetLabelOf(successor)); 1665 } 1666} 1667 1668void LocationsBuilderARM::VisitGoto(HGoto* got) { 1669 got->SetLocations(nullptr); 1670} 1671 1672void InstructionCodeGeneratorARM::VisitGoto(HGoto* got) { 1673 HandleGoto(got, got->GetSuccessor()); 1674} 1675 1676void LocationsBuilderARM::VisitTryBoundary(HTryBoundary* try_boundary) { 1677 try_boundary->SetLocations(nullptr); 1678} 1679 1680void InstructionCodeGeneratorARM::VisitTryBoundary(HTryBoundary* try_boundary) { 1681 HBasicBlock* successor = try_boundary->GetNormalFlowSuccessor(); 1682 if (!successor->IsExitBlock()) { 1683 HandleGoto(try_boundary, successor); 1684 } 1685} 1686 1687void LocationsBuilderARM::VisitExit(HExit* exit) { 1688 exit->SetLocations(nullptr); 1689} 1690 1691void InstructionCodeGeneratorARM::VisitExit(HExit* exit ATTRIBUTE_UNUSED) { 1692} 1693 1694void InstructionCodeGeneratorARM::GenerateVcmp(HInstruction* instruction) { 1695 Primitive::Type type = instruction->InputAt(0)->GetType(); 1696 Location lhs_loc = instruction->GetLocations()->InAt(0); 1697 Location rhs_loc = instruction->GetLocations()->InAt(1); 1698 if (rhs_loc.IsConstant()) { 1699 // 0.0 is the only immediate that can be encoded directly in 1700 // a VCMP instruction. 1701 // 1702 // Both the JLS (section 15.20.1) and the JVMS (section 6.5) 1703 // specify that in a floating-point comparison, positive zero 1704 // and negative zero are considered equal, so we can use the 1705 // literal 0.0 for both cases here. 1706 // 1707 // Note however that some methods (Float.equal, Float.compare, 1708 // Float.compareTo, Double.equal, Double.compare, 1709 // Double.compareTo, Math.max, Math.min, StrictMath.max, 1710 // StrictMath.min) consider 0.0 to be (strictly) greater than 1711 // -0.0. So if we ever translate calls to these methods into a 1712 // HCompare instruction, we must handle the -0.0 case with 1713 // care here. 1714 DCHECK(rhs_loc.GetConstant()->IsArithmeticZero()); 1715 if (type == Primitive::kPrimFloat) { 1716 __ vcmpsz(lhs_loc.AsFpuRegister<SRegister>()); 1717 } else { 1718 DCHECK_EQ(type, Primitive::kPrimDouble); 1719 __ vcmpdz(FromLowSToD(lhs_loc.AsFpuRegisterPairLow<SRegister>())); 1720 } 1721 } else { 1722 if (type == Primitive::kPrimFloat) { 1723 __ vcmps(lhs_loc.AsFpuRegister<SRegister>(), rhs_loc.AsFpuRegister<SRegister>()); 1724 } else { 1725 DCHECK_EQ(type, Primitive::kPrimDouble); 1726 __ vcmpd(FromLowSToD(lhs_loc.AsFpuRegisterPairLow<SRegister>()), 1727 FromLowSToD(rhs_loc.AsFpuRegisterPairLow<SRegister>())); 1728 } 1729 } 1730} 1731 1732void InstructionCodeGeneratorARM::GenerateFPJumps(HCondition* cond, 1733 Label* true_label, 1734 Label* false_label ATTRIBUTE_UNUSED) { 1735 __ vmstat(); // transfer FP status register to ARM APSR. 1736 __ b(true_label, ARMFPCondition(cond->GetCondition(), cond->IsGtBias())); 1737} 1738 1739void InstructionCodeGeneratorARM::GenerateLongComparesAndJumps(HCondition* cond, 1740 Label* true_label, 1741 Label* false_label) { 1742 LocationSummary* locations = cond->GetLocations(); 1743 Location left = locations->InAt(0); 1744 Location right = locations->InAt(1); 1745 IfCondition if_cond = cond->GetCondition(); 1746 1747 Register left_high = left.AsRegisterPairHigh<Register>(); 1748 Register left_low = left.AsRegisterPairLow<Register>(); 1749 IfCondition true_high_cond = if_cond; 1750 IfCondition false_high_cond = cond->GetOppositeCondition(); 1751 Condition final_condition = ARMUnsignedCondition(if_cond); // unsigned on lower part 1752 1753 // Set the conditions for the test, remembering that == needs to be 1754 // decided using the low words. 1755 // TODO: consider avoiding jumps with temporary and CMP low+SBC high 1756 switch (if_cond) { 1757 case kCondEQ: 1758 case kCondNE: 1759 // Nothing to do. 1760 break; 1761 case kCondLT: 1762 false_high_cond = kCondGT; 1763 break; 1764 case kCondLE: 1765 true_high_cond = kCondLT; 1766 break; 1767 case kCondGT: 1768 false_high_cond = kCondLT; 1769 break; 1770 case kCondGE: 1771 true_high_cond = kCondGT; 1772 break; 1773 case kCondB: 1774 false_high_cond = kCondA; 1775 break; 1776 case kCondBE: 1777 true_high_cond = kCondB; 1778 break; 1779 case kCondA: 1780 false_high_cond = kCondB; 1781 break; 1782 case kCondAE: 1783 true_high_cond = kCondA; 1784 break; 1785 } 1786 if (right.IsConstant()) { 1787 int64_t value = right.GetConstant()->AsLongConstant()->GetValue(); 1788 int32_t val_low = Low32Bits(value); 1789 int32_t val_high = High32Bits(value); 1790 1791 __ CmpConstant(left_high, val_high); 1792 if (if_cond == kCondNE) { 1793 __ b(true_label, ARMCondition(true_high_cond)); 1794 } else if (if_cond == kCondEQ) { 1795 __ b(false_label, ARMCondition(false_high_cond)); 1796 } else { 1797 __ b(true_label, ARMCondition(true_high_cond)); 1798 __ b(false_label, ARMCondition(false_high_cond)); 1799 } 1800 // Must be equal high, so compare the lows. 1801 __ CmpConstant(left_low, val_low); 1802 } else { 1803 Register right_high = right.AsRegisterPairHigh<Register>(); 1804 Register right_low = right.AsRegisterPairLow<Register>(); 1805 1806 __ cmp(left_high, ShifterOperand(right_high)); 1807 if (if_cond == kCondNE) { 1808 __ b(true_label, ARMCondition(true_high_cond)); 1809 } else if (if_cond == kCondEQ) { 1810 __ b(false_label, ARMCondition(false_high_cond)); 1811 } else { 1812 __ b(true_label, ARMCondition(true_high_cond)); 1813 __ b(false_label, ARMCondition(false_high_cond)); 1814 } 1815 // Must be equal high, so compare the lows. 1816 __ cmp(left_low, ShifterOperand(right_low)); 1817 } 1818 // The last comparison might be unsigned. 1819 // TODO: optimize cases where this is always true/false 1820 __ b(true_label, final_condition); 1821} 1822 1823void InstructionCodeGeneratorARM::GenerateCompareTestAndBranch(HCondition* condition, 1824 Label* true_target_in, 1825 Label* false_target_in) { 1826 // Generated branching requires both targets to be explicit. If either of the 1827 // targets is nullptr (fallthrough) use and bind `fallthrough_target` instead. 1828 Label fallthrough_target; 1829 Label* true_target = true_target_in == nullptr ? &fallthrough_target : true_target_in; 1830 Label* false_target = false_target_in == nullptr ? &fallthrough_target : false_target_in; 1831 1832 Primitive::Type type = condition->InputAt(0)->GetType(); 1833 switch (type) { 1834 case Primitive::kPrimLong: 1835 GenerateLongComparesAndJumps(condition, true_target, false_target); 1836 break; 1837 case Primitive::kPrimFloat: 1838 case Primitive::kPrimDouble: 1839 GenerateVcmp(condition); 1840 GenerateFPJumps(condition, true_target, false_target); 1841 break; 1842 default: 1843 LOG(FATAL) << "Unexpected compare type " << type; 1844 } 1845 1846 if (false_target != &fallthrough_target) { 1847 __ b(false_target); 1848 } 1849 1850 if (fallthrough_target.IsLinked()) { 1851 __ Bind(&fallthrough_target); 1852 } 1853} 1854 1855void InstructionCodeGeneratorARM::GenerateTestAndBranch(HInstruction* instruction, 1856 size_t condition_input_index, 1857 Label* true_target, 1858 Label* false_target) { 1859 HInstruction* cond = instruction->InputAt(condition_input_index); 1860 1861 if (true_target == nullptr && false_target == nullptr) { 1862 // Nothing to do. The code always falls through. 1863 return; 1864 } else if (cond->IsIntConstant()) { 1865 // Constant condition, statically compared against "true" (integer value 1). 1866 if (cond->AsIntConstant()->IsTrue()) { 1867 if (true_target != nullptr) { 1868 __ b(true_target); 1869 } 1870 } else { 1871 DCHECK(cond->AsIntConstant()->IsFalse()) << cond->AsIntConstant()->GetValue(); 1872 if (false_target != nullptr) { 1873 __ b(false_target); 1874 } 1875 } 1876 return; 1877 } 1878 1879 // The following code generates these patterns: 1880 // (1) true_target == nullptr && false_target != nullptr 1881 // - opposite condition true => branch to false_target 1882 // (2) true_target != nullptr && false_target == nullptr 1883 // - condition true => branch to true_target 1884 // (3) true_target != nullptr && false_target != nullptr 1885 // - condition true => branch to true_target 1886 // - branch to false_target 1887 if (IsBooleanValueOrMaterializedCondition(cond)) { 1888 // Condition has been materialized, compare the output to 0. 1889 Location cond_val = instruction->GetLocations()->InAt(condition_input_index); 1890 DCHECK(cond_val.IsRegister()); 1891 if (true_target == nullptr) { 1892 __ CompareAndBranchIfZero(cond_val.AsRegister<Register>(), false_target); 1893 } else { 1894 __ CompareAndBranchIfNonZero(cond_val.AsRegister<Register>(), true_target); 1895 } 1896 } else { 1897 // Condition has not been materialized. Use its inputs as the comparison and 1898 // its condition as the branch condition. 1899 HCondition* condition = cond->AsCondition(); 1900 1901 // If this is a long or FP comparison that has been folded into 1902 // the HCondition, generate the comparison directly. 1903 Primitive::Type type = condition->InputAt(0)->GetType(); 1904 if (type == Primitive::kPrimLong || Primitive::IsFloatingPointType(type)) { 1905 GenerateCompareTestAndBranch(condition, true_target, false_target); 1906 return; 1907 } 1908 1909 LocationSummary* locations = cond->GetLocations(); 1910 DCHECK(locations->InAt(0).IsRegister()); 1911 Register left = locations->InAt(0).AsRegister<Register>(); 1912 Location right = locations->InAt(1); 1913 if (right.IsRegister()) { 1914 __ cmp(left, ShifterOperand(right.AsRegister<Register>())); 1915 } else { 1916 DCHECK(right.IsConstant()); 1917 __ CmpConstant(left, CodeGenerator::GetInt32ValueOf(right.GetConstant())); 1918 } 1919 if (true_target == nullptr) { 1920 __ b(false_target, ARMCondition(condition->GetOppositeCondition())); 1921 } else { 1922 __ b(true_target, ARMCondition(condition->GetCondition())); 1923 } 1924 } 1925 1926 // If neither branch falls through (case 3), the conditional branch to `true_target` 1927 // was already emitted (case 2) and we need to emit a jump to `false_target`. 1928 if (true_target != nullptr && false_target != nullptr) { 1929 __ b(false_target); 1930 } 1931} 1932 1933void LocationsBuilderARM::VisitIf(HIf* if_instr) { 1934 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(if_instr); 1935 if (IsBooleanValueOrMaterializedCondition(if_instr->InputAt(0))) { 1936 locations->SetInAt(0, Location::RequiresRegister()); 1937 } 1938} 1939 1940void InstructionCodeGeneratorARM::VisitIf(HIf* if_instr) { 1941 HBasicBlock* true_successor = if_instr->IfTrueSuccessor(); 1942 HBasicBlock* false_successor = if_instr->IfFalseSuccessor(); 1943 Label* true_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), true_successor) ? 1944 nullptr : codegen_->GetLabelOf(true_successor); 1945 Label* false_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), false_successor) ? 1946 nullptr : codegen_->GetLabelOf(false_successor); 1947 GenerateTestAndBranch(if_instr, /* condition_input_index */ 0, true_target, false_target); 1948} 1949 1950void LocationsBuilderARM::VisitDeoptimize(HDeoptimize* deoptimize) { 1951 LocationSummary* locations = new (GetGraph()->GetArena()) 1952 LocationSummary(deoptimize, LocationSummary::kCallOnSlowPath); 1953 locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty()); // No caller-save registers. 1954 if (IsBooleanValueOrMaterializedCondition(deoptimize->InputAt(0))) { 1955 locations->SetInAt(0, Location::RequiresRegister()); 1956 } 1957} 1958 1959void InstructionCodeGeneratorARM::VisitDeoptimize(HDeoptimize* deoptimize) { 1960 SlowPathCodeARM* slow_path = deopt_slow_paths_.NewSlowPath<DeoptimizationSlowPathARM>(deoptimize); 1961 GenerateTestAndBranch(deoptimize, 1962 /* condition_input_index */ 0, 1963 slow_path->GetEntryLabel(), 1964 /* false_target */ nullptr); 1965} 1966 1967void LocationsBuilderARM::VisitShouldDeoptimizeFlag(HShouldDeoptimizeFlag* flag) { 1968 LocationSummary* locations = new (GetGraph()->GetArena()) 1969 LocationSummary(flag, LocationSummary::kNoCall); 1970 locations->SetOut(Location::RequiresRegister()); 1971} 1972 1973void InstructionCodeGeneratorARM::VisitShouldDeoptimizeFlag(HShouldDeoptimizeFlag* flag) { 1974 __ LoadFromOffset(kLoadWord, 1975 flag->GetLocations()->Out().AsRegister<Register>(), 1976 SP, 1977 codegen_->GetStackOffsetOfShouldDeoptimizeFlag()); 1978} 1979 1980void LocationsBuilderARM::VisitSelect(HSelect* select) { 1981 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(select); 1982 if (Primitive::IsFloatingPointType(select->GetType())) { 1983 locations->SetInAt(0, Location::RequiresFpuRegister()); 1984 locations->SetInAt(1, Location::RequiresFpuRegister()); 1985 } else { 1986 locations->SetInAt(0, Location::RequiresRegister()); 1987 locations->SetInAt(1, Location::RequiresRegister()); 1988 } 1989 if (IsBooleanValueOrMaterializedCondition(select->GetCondition())) { 1990 locations->SetInAt(2, Location::RequiresRegister()); 1991 } 1992 locations->SetOut(Location::SameAsFirstInput()); 1993} 1994 1995void InstructionCodeGeneratorARM::VisitSelect(HSelect* select) { 1996 LocationSummary* locations = select->GetLocations(); 1997 Label false_target; 1998 GenerateTestAndBranch(select, 1999 /* condition_input_index */ 2, 2000 /* true_target */ nullptr, 2001 &false_target); 2002 codegen_->MoveLocation(locations->Out(), locations->InAt(1), select->GetType()); 2003 __ Bind(&false_target); 2004} 2005 2006void LocationsBuilderARM::VisitNativeDebugInfo(HNativeDebugInfo* info) { 2007 new (GetGraph()->GetArena()) LocationSummary(info); 2008} 2009 2010void InstructionCodeGeneratorARM::VisitNativeDebugInfo(HNativeDebugInfo*) { 2011 // MaybeRecordNativeDebugInfo is already called implicitly in CodeGenerator::Compile. 2012} 2013 2014void CodeGeneratorARM::GenerateNop() { 2015 __ nop(); 2016} 2017 2018void LocationsBuilderARM::HandleCondition(HCondition* cond) { 2019 LocationSummary* locations = 2020 new (GetGraph()->GetArena()) LocationSummary(cond, LocationSummary::kNoCall); 2021 // Handle the long/FP comparisons made in instruction simplification. 2022 switch (cond->InputAt(0)->GetType()) { 2023 case Primitive::kPrimLong: 2024 locations->SetInAt(0, Location::RequiresRegister()); 2025 locations->SetInAt(1, Location::RegisterOrConstant(cond->InputAt(1))); 2026 if (!cond->IsEmittedAtUseSite()) { 2027 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 2028 } 2029 break; 2030 2031 case Primitive::kPrimFloat: 2032 case Primitive::kPrimDouble: 2033 locations->SetInAt(0, Location::RequiresFpuRegister()); 2034 locations->SetInAt(1, ArithmeticZeroOrFpuRegister(cond->InputAt(1))); 2035 if (!cond->IsEmittedAtUseSite()) { 2036 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2037 } 2038 break; 2039 2040 default: 2041 locations->SetInAt(0, Location::RequiresRegister()); 2042 locations->SetInAt(1, Location::RegisterOrConstant(cond->InputAt(1))); 2043 if (!cond->IsEmittedAtUseSite()) { 2044 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2045 } 2046 } 2047} 2048 2049void InstructionCodeGeneratorARM::HandleCondition(HCondition* cond) { 2050 if (cond->IsEmittedAtUseSite()) { 2051 return; 2052 } 2053 2054 LocationSummary* locations = cond->GetLocations(); 2055 Location left = locations->InAt(0); 2056 Location right = locations->InAt(1); 2057 Register out = locations->Out().AsRegister<Register>(); 2058 Label true_label, false_label; 2059 2060 switch (cond->InputAt(0)->GetType()) { 2061 default: { 2062 // Integer case. 2063 if (right.IsRegister()) { 2064 __ cmp(left.AsRegister<Register>(), ShifterOperand(right.AsRegister<Register>())); 2065 } else { 2066 DCHECK(right.IsConstant()); 2067 __ CmpConstant(left.AsRegister<Register>(), 2068 CodeGenerator::GetInt32ValueOf(right.GetConstant())); 2069 } 2070 __ it(ARMCondition(cond->GetCondition()), kItElse); 2071 __ mov(locations->Out().AsRegister<Register>(), ShifterOperand(1), 2072 ARMCondition(cond->GetCondition())); 2073 __ mov(locations->Out().AsRegister<Register>(), ShifterOperand(0), 2074 ARMCondition(cond->GetOppositeCondition())); 2075 return; 2076 } 2077 case Primitive::kPrimLong: 2078 GenerateLongComparesAndJumps(cond, &true_label, &false_label); 2079 break; 2080 case Primitive::kPrimFloat: 2081 case Primitive::kPrimDouble: 2082 GenerateVcmp(cond); 2083 GenerateFPJumps(cond, &true_label, &false_label); 2084 break; 2085 } 2086 2087 // Convert the jumps into the result. 2088 Label done_label; 2089 2090 // False case: result = 0. 2091 __ Bind(&false_label); 2092 __ LoadImmediate(out, 0); 2093 __ b(&done_label); 2094 2095 // True case: result = 1. 2096 __ Bind(&true_label); 2097 __ LoadImmediate(out, 1); 2098 __ Bind(&done_label); 2099} 2100 2101void LocationsBuilderARM::VisitEqual(HEqual* comp) { 2102 HandleCondition(comp); 2103} 2104 2105void InstructionCodeGeneratorARM::VisitEqual(HEqual* comp) { 2106 HandleCondition(comp); 2107} 2108 2109void LocationsBuilderARM::VisitNotEqual(HNotEqual* comp) { 2110 HandleCondition(comp); 2111} 2112 2113void InstructionCodeGeneratorARM::VisitNotEqual(HNotEqual* comp) { 2114 HandleCondition(comp); 2115} 2116 2117void LocationsBuilderARM::VisitLessThan(HLessThan* comp) { 2118 HandleCondition(comp); 2119} 2120 2121void InstructionCodeGeneratorARM::VisitLessThan(HLessThan* comp) { 2122 HandleCondition(comp); 2123} 2124 2125void LocationsBuilderARM::VisitLessThanOrEqual(HLessThanOrEqual* comp) { 2126 HandleCondition(comp); 2127} 2128 2129void InstructionCodeGeneratorARM::VisitLessThanOrEqual(HLessThanOrEqual* comp) { 2130 HandleCondition(comp); 2131} 2132 2133void LocationsBuilderARM::VisitGreaterThan(HGreaterThan* comp) { 2134 HandleCondition(comp); 2135} 2136 2137void InstructionCodeGeneratorARM::VisitGreaterThan(HGreaterThan* comp) { 2138 HandleCondition(comp); 2139} 2140 2141void LocationsBuilderARM::VisitGreaterThanOrEqual(HGreaterThanOrEqual* comp) { 2142 HandleCondition(comp); 2143} 2144 2145void InstructionCodeGeneratorARM::VisitGreaterThanOrEqual(HGreaterThanOrEqual* comp) { 2146 HandleCondition(comp); 2147} 2148 2149void LocationsBuilderARM::VisitBelow(HBelow* comp) { 2150 HandleCondition(comp); 2151} 2152 2153void InstructionCodeGeneratorARM::VisitBelow(HBelow* comp) { 2154 HandleCondition(comp); 2155} 2156 2157void LocationsBuilderARM::VisitBelowOrEqual(HBelowOrEqual* comp) { 2158 HandleCondition(comp); 2159} 2160 2161void InstructionCodeGeneratorARM::VisitBelowOrEqual(HBelowOrEqual* comp) { 2162 HandleCondition(comp); 2163} 2164 2165void LocationsBuilderARM::VisitAbove(HAbove* comp) { 2166 HandleCondition(comp); 2167} 2168 2169void InstructionCodeGeneratorARM::VisitAbove(HAbove* comp) { 2170 HandleCondition(comp); 2171} 2172 2173void LocationsBuilderARM::VisitAboveOrEqual(HAboveOrEqual* comp) { 2174 HandleCondition(comp); 2175} 2176 2177void InstructionCodeGeneratorARM::VisitAboveOrEqual(HAboveOrEqual* comp) { 2178 HandleCondition(comp); 2179} 2180 2181void LocationsBuilderARM::VisitIntConstant(HIntConstant* constant) { 2182 LocationSummary* locations = 2183 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 2184 locations->SetOut(Location::ConstantLocation(constant)); 2185} 2186 2187void InstructionCodeGeneratorARM::VisitIntConstant(HIntConstant* constant ATTRIBUTE_UNUSED) { 2188 // Will be generated at use site. 2189} 2190 2191void LocationsBuilderARM::VisitNullConstant(HNullConstant* constant) { 2192 LocationSummary* locations = 2193 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 2194 locations->SetOut(Location::ConstantLocation(constant)); 2195} 2196 2197void InstructionCodeGeneratorARM::VisitNullConstant(HNullConstant* constant ATTRIBUTE_UNUSED) { 2198 // Will be generated at use site. 2199} 2200 2201void LocationsBuilderARM::VisitLongConstant(HLongConstant* constant) { 2202 LocationSummary* locations = 2203 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 2204 locations->SetOut(Location::ConstantLocation(constant)); 2205} 2206 2207void InstructionCodeGeneratorARM::VisitLongConstant(HLongConstant* constant ATTRIBUTE_UNUSED) { 2208 // Will be generated at use site. 2209} 2210 2211void LocationsBuilderARM::VisitFloatConstant(HFloatConstant* constant) { 2212 LocationSummary* locations = 2213 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 2214 locations->SetOut(Location::ConstantLocation(constant)); 2215} 2216 2217void InstructionCodeGeneratorARM::VisitFloatConstant(HFloatConstant* constant ATTRIBUTE_UNUSED) { 2218 // Will be generated at use site. 2219} 2220 2221void LocationsBuilderARM::VisitDoubleConstant(HDoubleConstant* constant) { 2222 LocationSummary* locations = 2223 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 2224 locations->SetOut(Location::ConstantLocation(constant)); 2225} 2226 2227void InstructionCodeGeneratorARM::VisitDoubleConstant(HDoubleConstant* constant ATTRIBUTE_UNUSED) { 2228 // Will be generated at use site. 2229} 2230 2231void LocationsBuilderARM::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) { 2232 memory_barrier->SetLocations(nullptr); 2233} 2234 2235void InstructionCodeGeneratorARM::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) { 2236 codegen_->GenerateMemoryBarrier(memory_barrier->GetBarrierKind()); 2237} 2238 2239void LocationsBuilderARM::VisitReturnVoid(HReturnVoid* ret) { 2240 ret->SetLocations(nullptr); 2241} 2242 2243void InstructionCodeGeneratorARM::VisitReturnVoid(HReturnVoid* ret ATTRIBUTE_UNUSED) { 2244 codegen_->GenerateFrameExit(); 2245} 2246 2247void LocationsBuilderARM::VisitReturn(HReturn* ret) { 2248 LocationSummary* locations = 2249 new (GetGraph()->GetArena()) LocationSummary(ret, LocationSummary::kNoCall); 2250 locations->SetInAt(0, parameter_visitor_.GetReturnLocation(ret->InputAt(0)->GetType())); 2251} 2252 2253void InstructionCodeGeneratorARM::VisitReturn(HReturn* ret ATTRIBUTE_UNUSED) { 2254 codegen_->GenerateFrameExit(); 2255} 2256 2257void LocationsBuilderARM::VisitInvokeUnresolved(HInvokeUnresolved* invoke) { 2258 // The trampoline uses the same calling convention as dex calling conventions, 2259 // except instead of loading arg0/r0 with the target Method*, arg0/r0 will contain 2260 // the method_idx. 2261 HandleInvoke(invoke); 2262} 2263 2264void InstructionCodeGeneratorARM::VisitInvokeUnresolved(HInvokeUnresolved* invoke) { 2265 codegen_->GenerateInvokeUnresolvedRuntimeCall(invoke); 2266} 2267 2268void LocationsBuilderARM::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) { 2269 // Explicit clinit checks triggered by static invokes must have been pruned by 2270 // art::PrepareForRegisterAllocation. 2271 DCHECK(!invoke->IsStaticWithExplicitClinitCheck()); 2272 2273 IntrinsicLocationsBuilderARM intrinsic(codegen_); 2274 if (intrinsic.TryDispatch(invoke)) { 2275 if (invoke->GetLocations()->CanCall() && invoke->HasPcRelativeDexCache()) { 2276 invoke->GetLocations()->SetInAt(invoke->GetSpecialInputIndex(), Location::Any()); 2277 } 2278 return; 2279 } 2280 2281 HandleInvoke(invoke); 2282 2283 // For PC-relative dex cache the invoke has an extra input, the PC-relative address base. 2284 if (invoke->HasPcRelativeDexCache()) { 2285 invoke->GetLocations()->SetInAt(invoke->GetSpecialInputIndex(), Location::RequiresRegister()); 2286 } 2287} 2288 2289static bool TryGenerateIntrinsicCode(HInvoke* invoke, CodeGeneratorARM* codegen) { 2290 if (invoke->GetLocations()->Intrinsified()) { 2291 IntrinsicCodeGeneratorARM intrinsic(codegen); 2292 intrinsic.Dispatch(invoke); 2293 return true; 2294 } 2295 return false; 2296} 2297 2298void InstructionCodeGeneratorARM::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) { 2299 // Explicit clinit checks triggered by static invokes must have been pruned by 2300 // art::PrepareForRegisterAllocation. 2301 DCHECK(!invoke->IsStaticWithExplicitClinitCheck()); 2302 2303 if (TryGenerateIntrinsicCode(invoke, codegen_)) { 2304 return; 2305 } 2306 2307 LocationSummary* locations = invoke->GetLocations(); 2308 codegen_->GenerateStaticOrDirectCall( 2309 invoke, locations->HasTemps() ? locations->GetTemp(0) : Location::NoLocation()); 2310 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 2311} 2312 2313void LocationsBuilderARM::HandleInvoke(HInvoke* invoke) { 2314 InvokeDexCallingConventionVisitorARM calling_convention_visitor; 2315 CodeGenerator::CreateCommonInvokeLocationSummary(invoke, &calling_convention_visitor); 2316} 2317 2318void LocationsBuilderARM::VisitInvokeVirtual(HInvokeVirtual* invoke) { 2319 IntrinsicLocationsBuilderARM intrinsic(codegen_); 2320 if (intrinsic.TryDispatch(invoke)) { 2321 return; 2322 } 2323 2324 HandleInvoke(invoke); 2325} 2326 2327void InstructionCodeGeneratorARM::VisitInvokeVirtual(HInvokeVirtual* invoke) { 2328 if (TryGenerateIntrinsicCode(invoke, codegen_)) { 2329 return; 2330 } 2331 2332 codegen_->GenerateVirtualCall(invoke, invoke->GetLocations()->GetTemp(0)); 2333 DCHECK(!codegen_->IsLeafMethod()); 2334 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 2335} 2336 2337void LocationsBuilderARM::VisitInvokeInterface(HInvokeInterface* invoke) { 2338 HandleInvoke(invoke); 2339 // Add the hidden argument. 2340 invoke->GetLocations()->AddTemp(Location::RegisterLocation(R12)); 2341} 2342 2343void InstructionCodeGeneratorARM::VisitInvokeInterface(HInvokeInterface* invoke) { 2344 // TODO: b/18116999, our IMTs can miss an IncompatibleClassChangeError. 2345 LocationSummary* locations = invoke->GetLocations(); 2346 Register temp = locations->GetTemp(0).AsRegister<Register>(); 2347 Register hidden_reg = locations->GetTemp(1).AsRegister<Register>(); 2348 Location receiver = locations->InAt(0); 2349 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 2350 2351 // Set the hidden argument. This is safe to do this here, as R12 2352 // won't be modified thereafter, before the `blx` (call) instruction. 2353 DCHECK_EQ(R12, hidden_reg); 2354 __ LoadImmediate(hidden_reg, invoke->GetDexMethodIndex()); 2355 2356 if (receiver.IsStackSlot()) { 2357 __ LoadFromOffset(kLoadWord, temp, SP, receiver.GetStackIndex()); 2358 // /* HeapReference<Class> */ temp = temp->klass_ 2359 __ LoadFromOffset(kLoadWord, temp, temp, class_offset); 2360 } else { 2361 // /* HeapReference<Class> */ temp = receiver->klass_ 2362 __ LoadFromOffset(kLoadWord, temp, receiver.AsRegister<Register>(), class_offset); 2363 } 2364 codegen_->MaybeRecordImplicitNullCheck(invoke); 2365 // Instead of simply (possibly) unpoisoning `temp` here, we should 2366 // emit a read barrier for the previous class reference load. 2367 // However this is not required in practice, as this is an 2368 // intermediate/temporary reference and because the current 2369 // concurrent copying collector keeps the from-space memory 2370 // intact/accessible until the end of the marking phase (the 2371 // concurrent copying collector may not in the future). 2372 __ MaybeUnpoisonHeapReference(temp); 2373 __ LoadFromOffset(kLoadWord, temp, temp, 2374 mirror::Class::ImtPtrOffset(kArmPointerSize).Uint32Value()); 2375 uint32_t method_offset = static_cast<uint32_t>(ImTable::OffsetOfElement( 2376 invoke->GetImtIndex(), kArmPointerSize)); 2377 // temp = temp->GetImtEntryAt(method_offset); 2378 __ LoadFromOffset(kLoadWord, temp, temp, method_offset); 2379 uint32_t entry_point = 2380 ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmPointerSize).Int32Value(); 2381 // LR = temp->GetEntryPoint(); 2382 __ LoadFromOffset(kLoadWord, LR, temp, entry_point); 2383 // LR(); 2384 __ blx(LR); 2385 DCHECK(!codegen_->IsLeafMethod()); 2386 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 2387} 2388 2389void LocationsBuilderARM::VisitInvokePolymorphic(HInvokePolymorphic* invoke) { 2390 HandleInvoke(invoke); 2391} 2392 2393void InstructionCodeGeneratorARM::VisitInvokePolymorphic(HInvokePolymorphic* invoke) { 2394 codegen_->GenerateInvokePolymorphicCall(invoke); 2395} 2396 2397void LocationsBuilderARM::VisitNeg(HNeg* neg) { 2398 LocationSummary* locations = 2399 new (GetGraph()->GetArena()) LocationSummary(neg, LocationSummary::kNoCall); 2400 switch (neg->GetResultType()) { 2401 case Primitive::kPrimInt: { 2402 locations->SetInAt(0, Location::RequiresRegister()); 2403 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2404 break; 2405 } 2406 case Primitive::kPrimLong: { 2407 locations->SetInAt(0, Location::RequiresRegister()); 2408 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 2409 break; 2410 } 2411 2412 case Primitive::kPrimFloat: 2413 case Primitive::kPrimDouble: 2414 locations->SetInAt(0, Location::RequiresFpuRegister()); 2415 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2416 break; 2417 2418 default: 2419 LOG(FATAL) << "Unexpected neg type " << neg->GetResultType(); 2420 } 2421} 2422 2423void InstructionCodeGeneratorARM::VisitNeg(HNeg* neg) { 2424 LocationSummary* locations = neg->GetLocations(); 2425 Location out = locations->Out(); 2426 Location in = locations->InAt(0); 2427 switch (neg->GetResultType()) { 2428 case Primitive::kPrimInt: 2429 DCHECK(in.IsRegister()); 2430 __ rsb(out.AsRegister<Register>(), in.AsRegister<Register>(), ShifterOperand(0)); 2431 break; 2432 2433 case Primitive::kPrimLong: 2434 DCHECK(in.IsRegisterPair()); 2435 // out.lo = 0 - in.lo (and update the carry/borrow (C) flag) 2436 __ rsbs(out.AsRegisterPairLow<Register>(), 2437 in.AsRegisterPairLow<Register>(), 2438 ShifterOperand(0)); 2439 // We cannot emit an RSC (Reverse Subtract with Carry) 2440 // instruction here, as it does not exist in the Thumb-2 2441 // instruction set. We use the following approach 2442 // using SBC and SUB instead. 2443 // 2444 // out.hi = -C 2445 __ sbc(out.AsRegisterPairHigh<Register>(), 2446 out.AsRegisterPairHigh<Register>(), 2447 ShifterOperand(out.AsRegisterPairHigh<Register>())); 2448 // out.hi = out.hi - in.hi 2449 __ sub(out.AsRegisterPairHigh<Register>(), 2450 out.AsRegisterPairHigh<Register>(), 2451 ShifterOperand(in.AsRegisterPairHigh<Register>())); 2452 break; 2453 2454 case Primitive::kPrimFloat: 2455 DCHECK(in.IsFpuRegister()); 2456 __ vnegs(out.AsFpuRegister<SRegister>(), in.AsFpuRegister<SRegister>()); 2457 break; 2458 2459 case Primitive::kPrimDouble: 2460 DCHECK(in.IsFpuRegisterPair()); 2461 __ vnegd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2462 FromLowSToD(in.AsFpuRegisterPairLow<SRegister>())); 2463 break; 2464 2465 default: 2466 LOG(FATAL) << "Unexpected neg type " << neg->GetResultType(); 2467 } 2468} 2469 2470void LocationsBuilderARM::VisitTypeConversion(HTypeConversion* conversion) { 2471 Primitive::Type result_type = conversion->GetResultType(); 2472 Primitive::Type input_type = conversion->GetInputType(); 2473 DCHECK_NE(result_type, input_type); 2474 2475 // The float-to-long, double-to-long and long-to-float type conversions 2476 // rely on a call to the runtime. 2477 LocationSummary::CallKind call_kind = 2478 (((input_type == Primitive::kPrimFloat || input_type == Primitive::kPrimDouble) 2479 && result_type == Primitive::kPrimLong) 2480 || (input_type == Primitive::kPrimLong && result_type == Primitive::kPrimFloat)) 2481 ? LocationSummary::kCallOnMainOnly 2482 : LocationSummary::kNoCall; 2483 LocationSummary* locations = 2484 new (GetGraph()->GetArena()) LocationSummary(conversion, call_kind); 2485 2486 // The Java language does not allow treating boolean as an integral type but 2487 // our bit representation makes it safe. 2488 2489 switch (result_type) { 2490 case Primitive::kPrimByte: 2491 switch (input_type) { 2492 case Primitive::kPrimLong: 2493 // Type conversion from long to byte is a result of code transformations. 2494 case Primitive::kPrimBoolean: 2495 // Boolean input is a result of code transformations. 2496 case Primitive::kPrimShort: 2497 case Primitive::kPrimInt: 2498 case Primitive::kPrimChar: 2499 // Processing a Dex `int-to-byte' instruction. 2500 locations->SetInAt(0, Location::RequiresRegister()); 2501 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2502 break; 2503 2504 default: 2505 LOG(FATAL) << "Unexpected type conversion from " << input_type 2506 << " to " << result_type; 2507 } 2508 break; 2509 2510 case Primitive::kPrimShort: 2511 switch (input_type) { 2512 case Primitive::kPrimLong: 2513 // Type conversion from long to short is a result of code transformations. 2514 case Primitive::kPrimBoolean: 2515 // Boolean input is a result of code transformations. 2516 case Primitive::kPrimByte: 2517 case Primitive::kPrimInt: 2518 case Primitive::kPrimChar: 2519 // Processing a Dex `int-to-short' instruction. 2520 locations->SetInAt(0, Location::RequiresRegister()); 2521 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2522 break; 2523 2524 default: 2525 LOG(FATAL) << "Unexpected type conversion from " << input_type 2526 << " to " << result_type; 2527 } 2528 break; 2529 2530 case Primitive::kPrimInt: 2531 switch (input_type) { 2532 case Primitive::kPrimLong: 2533 // Processing a Dex `long-to-int' instruction. 2534 locations->SetInAt(0, Location::Any()); 2535 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2536 break; 2537 2538 case Primitive::kPrimFloat: 2539 // Processing a Dex `float-to-int' instruction. 2540 locations->SetInAt(0, Location::RequiresFpuRegister()); 2541 locations->SetOut(Location::RequiresRegister()); 2542 locations->AddTemp(Location::RequiresFpuRegister()); 2543 break; 2544 2545 case Primitive::kPrimDouble: 2546 // Processing a Dex `double-to-int' instruction. 2547 locations->SetInAt(0, Location::RequiresFpuRegister()); 2548 locations->SetOut(Location::RequiresRegister()); 2549 locations->AddTemp(Location::RequiresFpuRegister()); 2550 break; 2551 2552 default: 2553 LOG(FATAL) << "Unexpected type conversion from " << input_type 2554 << " to " << result_type; 2555 } 2556 break; 2557 2558 case Primitive::kPrimLong: 2559 switch (input_type) { 2560 case Primitive::kPrimBoolean: 2561 // Boolean input is a result of code transformations. 2562 case Primitive::kPrimByte: 2563 case Primitive::kPrimShort: 2564 case Primitive::kPrimInt: 2565 case Primitive::kPrimChar: 2566 // Processing a Dex `int-to-long' instruction. 2567 locations->SetInAt(0, Location::RequiresRegister()); 2568 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2569 break; 2570 2571 case Primitive::kPrimFloat: { 2572 // Processing a Dex `float-to-long' instruction. 2573 InvokeRuntimeCallingConvention calling_convention; 2574 locations->SetInAt(0, Location::FpuRegisterLocation( 2575 calling_convention.GetFpuRegisterAt(0))); 2576 locations->SetOut(Location::RegisterPairLocation(R0, R1)); 2577 break; 2578 } 2579 2580 case Primitive::kPrimDouble: { 2581 // Processing a Dex `double-to-long' instruction. 2582 InvokeRuntimeCallingConvention calling_convention; 2583 locations->SetInAt(0, Location::FpuRegisterPairLocation( 2584 calling_convention.GetFpuRegisterAt(0), 2585 calling_convention.GetFpuRegisterAt(1))); 2586 locations->SetOut(Location::RegisterPairLocation(R0, R1)); 2587 break; 2588 } 2589 2590 default: 2591 LOG(FATAL) << "Unexpected type conversion from " << input_type 2592 << " to " << result_type; 2593 } 2594 break; 2595 2596 case Primitive::kPrimChar: 2597 switch (input_type) { 2598 case Primitive::kPrimLong: 2599 // Type conversion from long to char is a result of code transformations. 2600 case Primitive::kPrimBoolean: 2601 // Boolean input is a result of code transformations. 2602 case Primitive::kPrimByte: 2603 case Primitive::kPrimShort: 2604 case Primitive::kPrimInt: 2605 // Processing a Dex `int-to-char' instruction. 2606 locations->SetInAt(0, Location::RequiresRegister()); 2607 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2608 break; 2609 2610 default: 2611 LOG(FATAL) << "Unexpected type conversion from " << input_type 2612 << " to " << result_type; 2613 } 2614 break; 2615 2616 case Primitive::kPrimFloat: 2617 switch (input_type) { 2618 case Primitive::kPrimBoolean: 2619 // Boolean input is a result of code transformations. 2620 case Primitive::kPrimByte: 2621 case Primitive::kPrimShort: 2622 case Primitive::kPrimInt: 2623 case Primitive::kPrimChar: 2624 // Processing a Dex `int-to-float' instruction. 2625 locations->SetInAt(0, Location::RequiresRegister()); 2626 locations->SetOut(Location::RequiresFpuRegister()); 2627 break; 2628 2629 case Primitive::kPrimLong: { 2630 // Processing a Dex `long-to-float' instruction. 2631 InvokeRuntimeCallingConvention calling_convention; 2632 locations->SetInAt(0, Location::RegisterPairLocation( 2633 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1))); 2634 locations->SetOut(Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(0))); 2635 break; 2636 } 2637 2638 case Primitive::kPrimDouble: 2639 // Processing a Dex `double-to-float' instruction. 2640 locations->SetInAt(0, Location::RequiresFpuRegister()); 2641 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2642 break; 2643 2644 default: 2645 LOG(FATAL) << "Unexpected type conversion from " << input_type 2646 << " to " << result_type; 2647 }; 2648 break; 2649 2650 case Primitive::kPrimDouble: 2651 switch (input_type) { 2652 case Primitive::kPrimBoolean: 2653 // Boolean input is a result of code transformations. 2654 case Primitive::kPrimByte: 2655 case Primitive::kPrimShort: 2656 case Primitive::kPrimInt: 2657 case Primitive::kPrimChar: 2658 // Processing a Dex `int-to-double' instruction. 2659 locations->SetInAt(0, Location::RequiresRegister()); 2660 locations->SetOut(Location::RequiresFpuRegister()); 2661 break; 2662 2663 case Primitive::kPrimLong: 2664 // Processing a Dex `long-to-double' instruction. 2665 locations->SetInAt(0, Location::RequiresRegister()); 2666 locations->SetOut(Location::RequiresFpuRegister()); 2667 locations->AddTemp(Location::RequiresFpuRegister()); 2668 locations->AddTemp(Location::RequiresFpuRegister()); 2669 break; 2670 2671 case Primitive::kPrimFloat: 2672 // Processing a Dex `float-to-double' instruction. 2673 locations->SetInAt(0, Location::RequiresFpuRegister()); 2674 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2675 break; 2676 2677 default: 2678 LOG(FATAL) << "Unexpected type conversion from " << input_type 2679 << " to " << result_type; 2680 }; 2681 break; 2682 2683 default: 2684 LOG(FATAL) << "Unexpected type conversion from " << input_type 2685 << " to " << result_type; 2686 } 2687} 2688 2689void InstructionCodeGeneratorARM::VisitTypeConversion(HTypeConversion* conversion) { 2690 LocationSummary* locations = conversion->GetLocations(); 2691 Location out = locations->Out(); 2692 Location in = locations->InAt(0); 2693 Primitive::Type result_type = conversion->GetResultType(); 2694 Primitive::Type input_type = conversion->GetInputType(); 2695 DCHECK_NE(result_type, input_type); 2696 switch (result_type) { 2697 case Primitive::kPrimByte: 2698 switch (input_type) { 2699 case Primitive::kPrimLong: 2700 // Type conversion from long to byte is a result of code transformations. 2701 __ sbfx(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>(), 0, 8); 2702 break; 2703 case Primitive::kPrimBoolean: 2704 // Boolean input is a result of code transformations. 2705 case Primitive::kPrimShort: 2706 case Primitive::kPrimInt: 2707 case Primitive::kPrimChar: 2708 // Processing a Dex `int-to-byte' instruction. 2709 __ sbfx(out.AsRegister<Register>(), in.AsRegister<Register>(), 0, 8); 2710 break; 2711 2712 default: 2713 LOG(FATAL) << "Unexpected type conversion from " << input_type 2714 << " to " << result_type; 2715 } 2716 break; 2717 2718 case Primitive::kPrimShort: 2719 switch (input_type) { 2720 case Primitive::kPrimLong: 2721 // Type conversion from long to short is a result of code transformations. 2722 __ sbfx(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>(), 0, 16); 2723 break; 2724 case Primitive::kPrimBoolean: 2725 // Boolean input is a result of code transformations. 2726 case Primitive::kPrimByte: 2727 case Primitive::kPrimInt: 2728 case Primitive::kPrimChar: 2729 // Processing a Dex `int-to-short' instruction. 2730 __ sbfx(out.AsRegister<Register>(), in.AsRegister<Register>(), 0, 16); 2731 break; 2732 2733 default: 2734 LOG(FATAL) << "Unexpected type conversion from " << input_type 2735 << " to " << result_type; 2736 } 2737 break; 2738 2739 case Primitive::kPrimInt: 2740 switch (input_type) { 2741 case Primitive::kPrimLong: 2742 // Processing a Dex `long-to-int' instruction. 2743 DCHECK(out.IsRegister()); 2744 if (in.IsRegisterPair()) { 2745 __ Mov(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>()); 2746 } else if (in.IsDoubleStackSlot()) { 2747 __ LoadFromOffset(kLoadWord, out.AsRegister<Register>(), SP, in.GetStackIndex()); 2748 } else { 2749 DCHECK(in.IsConstant()); 2750 DCHECK(in.GetConstant()->IsLongConstant()); 2751 int64_t value = in.GetConstant()->AsLongConstant()->GetValue(); 2752 __ LoadImmediate(out.AsRegister<Register>(), static_cast<int32_t>(value)); 2753 } 2754 break; 2755 2756 case Primitive::kPrimFloat: { 2757 // Processing a Dex `float-to-int' instruction. 2758 SRegister temp = locations->GetTemp(0).AsFpuRegisterPairLow<SRegister>(); 2759 __ vcvtis(temp, in.AsFpuRegister<SRegister>()); 2760 __ vmovrs(out.AsRegister<Register>(), temp); 2761 break; 2762 } 2763 2764 case Primitive::kPrimDouble: { 2765 // Processing a Dex `double-to-int' instruction. 2766 SRegister temp_s = locations->GetTemp(0).AsFpuRegisterPairLow<SRegister>(); 2767 __ vcvtid(temp_s, FromLowSToD(in.AsFpuRegisterPairLow<SRegister>())); 2768 __ vmovrs(out.AsRegister<Register>(), temp_s); 2769 break; 2770 } 2771 2772 default: 2773 LOG(FATAL) << "Unexpected type conversion from " << input_type 2774 << " to " << result_type; 2775 } 2776 break; 2777 2778 case Primitive::kPrimLong: 2779 switch (input_type) { 2780 case Primitive::kPrimBoolean: 2781 // Boolean input is a result of code transformations. 2782 case Primitive::kPrimByte: 2783 case Primitive::kPrimShort: 2784 case Primitive::kPrimInt: 2785 case Primitive::kPrimChar: 2786 // Processing a Dex `int-to-long' instruction. 2787 DCHECK(out.IsRegisterPair()); 2788 DCHECK(in.IsRegister()); 2789 __ Mov(out.AsRegisterPairLow<Register>(), in.AsRegister<Register>()); 2790 // Sign extension. 2791 __ Asr(out.AsRegisterPairHigh<Register>(), 2792 out.AsRegisterPairLow<Register>(), 2793 31); 2794 break; 2795 2796 case Primitive::kPrimFloat: 2797 // Processing a Dex `float-to-long' instruction. 2798 codegen_->InvokeRuntime(kQuickF2l, conversion, conversion->GetDexPc()); 2799 CheckEntrypointTypes<kQuickF2l, int64_t, float>(); 2800 break; 2801 2802 case Primitive::kPrimDouble: 2803 // Processing a Dex `double-to-long' instruction. 2804 codegen_->InvokeRuntime(kQuickD2l, conversion, conversion->GetDexPc()); 2805 CheckEntrypointTypes<kQuickD2l, int64_t, double>(); 2806 break; 2807 2808 default: 2809 LOG(FATAL) << "Unexpected type conversion from " << input_type 2810 << " to " << result_type; 2811 } 2812 break; 2813 2814 case Primitive::kPrimChar: 2815 switch (input_type) { 2816 case Primitive::kPrimLong: 2817 // Type conversion from long to char is a result of code transformations. 2818 __ ubfx(out.AsRegister<Register>(), in.AsRegisterPairLow<Register>(), 0, 16); 2819 break; 2820 case Primitive::kPrimBoolean: 2821 // Boolean input is a result of code transformations. 2822 case Primitive::kPrimByte: 2823 case Primitive::kPrimShort: 2824 case Primitive::kPrimInt: 2825 // Processing a Dex `int-to-char' instruction. 2826 __ ubfx(out.AsRegister<Register>(), in.AsRegister<Register>(), 0, 16); 2827 break; 2828 2829 default: 2830 LOG(FATAL) << "Unexpected type conversion from " << input_type 2831 << " to " << result_type; 2832 } 2833 break; 2834 2835 case Primitive::kPrimFloat: 2836 switch (input_type) { 2837 case Primitive::kPrimBoolean: 2838 // Boolean input is a result of code transformations. 2839 case Primitive::kPrimByte: 2840 case Primitive::kPrimShort: 2841 case Primitive::kPrimInt: 2842 case Primitive::kPrimChar: { 2843 // Processing a Dex `int-to-float' instruction. 2844 __ vmovsr(out.AsFpuRegister<SRegister>(), in.AsRegister<Register>()); 2845 __ vcvtsi(out.AsFpuRegister<SRegister>(), out.AsFpuRegister<SRegister>()); 2846 break; 2847 } 2848 2849 case Primitive::kPrimLong: 2850 // Processing a Dex `long-to-float' instruction. 2851 codegen_->InvokeRuntime(kQuickL2f, conversion, conversion->GetDexPc()); 2852 CheckEntrypointTypes<kQuickL2f, float, int64_t>(); 2853 break; 2854 2855 case Primitive::kPrimDouble: 2856 // Processing a Dex `double-to-float' instruction. 2857 __ vcvtsd(out.AsFpuRegister<SRegister>(), 2858 FromLowSToD(in.AsFpuRegisterPairLow<SRegister>())); 2859 break; 2860 2861 default: 2862 LOG(FATAL) << "Unexpected type conversion from " << input_type 2863 << " to " << result_type; 2864 }; 2865 break; 2866 2867 case Primitive::kPrimDouble: 2868 switch (input_type) { 2869 case Primitive::kPrimBoolean: 2870 // Boolean input is a result of code transformations. 2871 case Primitive::kPrimByte: 2872 case Primitive::kPrimShort: 2873 case Primitive::kPrimInt: 2874 case Primitive::kPrimChar: { 2875 // Processing a Dex `int-to-double' instruction. 2876 __ vmovsr(out.AsFpuRegisterPairLow<SRegister>(), in.AsRegister<Register>()); 2877 __ vcvtdi(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2878 out.AsFpuRegisterPairLow<SRegister>()); 2879 break; 2880 } 2881 2882 case Primitive::kPrimLong: { 2883 // Processing a Dex `long-to-double' instruction. 2884 Register low = in.AsRegisterPairLow<Register>(); 2885 Register high = in.AsRegisterPairHigh<Register>(); 2886 SRegister out_s = out.AsFpuRegisterPairLow<SRegister>(); 2887 DRegister out_d = FromLowSToD(out_s); 2888 SRegister temp_s = locations->GetTemp(0).AsFpuRegisterPairLow<SRegister>(); 2889 DRegister temp_d = FromLowSToD(temp_s); 2890 SRegister constant_s = locations->GetTemp(1).AsFpuRegisterPairLow<SRegister>(); 2891 DRegister constant_d = FromLowSToD(constant_s); 2892 2893 // temp_d = int-to-double(high) 2894 __ vmovsr(temp_s, high); 2895 __ vcvtdi(temp_d, temp_s); 2896 // constant_d = k2Pow32EncodingForDouble 2897 __ LoadDImmediate(constant_d, bit_cast<double, int64_t>(k2Pow32EncodingForDouble)); 2898 // out_d = unsigned-to-double(low) 2899 __ vmovsr(out_s, low); 2900 __ vcvtdu(out_d, out_s); 2901 // out_d += temp_d * constant_d 2902 __ vmlad(out_d, temp_d, constant_d); 2903 break; 2904 } 2905 2906 case Primitive::kPrimFloat: 2907 // Processing a Dex `float-to-double' instruction. 2908 __ vcvtds(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2909 in.AsFpuRegister<SRegister>()); 2910 break; 2911 2912 default: 2913 LOG(FATAL) << "Unexpected type conversion from " << input_type 2914 << " to " << result_type; 2915 }; 2916 break; 2917 2918 default: 2919 LOG(FATAL) << "Unexpected type conversion from " << input_type 2920 << " to " << result_type; 2921 } 2922} 2923 2924void LocationsBuilderARM::VisitAdd(HAdd* add) { 2925 LocationSummary* locations = 2926 new (GetGraph()->GetArena()) LocationSummary(add, LocationSummary::kNoCall); 2927 switch (add->GetResultType()) { 2928 case Primitive::kPrimInt: { 2929 locations->SetInAt(0, Location::RequiresRegister()); 2930 locations->SetInAt(1, Location::RegisterOrConstant(add->InputAt(1))); 2931 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2932 break; 2933 } 2934 2935 case Primitive::kPrimLong: { 2936 locations->SetInAt(0, Location::RequiresRegister()); 2937 locations->SetInAt(1, ArmEncodableConstantOrRegister(add->InputAt(1), ADD)); 2938 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2939 break; 2940 } 2941 2942 case Primitive::kPrimFloat: 2943 case Primitive::kPrimDouble: { 2944 locations->SetInAt(0, Location::RequiresFpuRegister()); 2945 locations->SetInAt(1, Location::RequiresFpuRegister()); 2946 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2947 break; 2948 } 2949 2950 default: 2951 LOG(FATAL) << "Unexpected add type " << add->GetResultType(); 2952 } 2953} 2954 2955void InstructionCodeGeneratorARM::VisitAdd(HAdd* add) { 2956 LocationSummary* locations = add->GetLocations(); 2957 Location out = locations->Out(); 2958 Location first = locations->InAt(0); 2959 Location second = locations->InAt(1); 2960 switch (add->GetResultType()) { 2961 case Primitive::kPrimInt: 2962 if (second.IsRegister()) { 2963 __ add(out.AsRegister<Register>(), 2964 first.AsRegister<Register>(), 2965 ShifterOperand(second.AsRegister<Register>())); 2966 } else { 2967 __ AddConstant(out.AsRegister<Register>(), 2968 first.AsRegister<Register>(), 2969 second.GetConstant()->AsIntConstant()->GetValue()); 2970 } 2971 break; 2972 2973 case Primitive::kPrimLong: { 2974 if (second.IsConstant()) { 2975 uint64_t value = static_cast<uint64_t>(Int64FromConstant(second.GetConstant())); 2976 GenerateAddLongConst(out, first, value); 2977 } else { 2978 DCHECK(second.IsRegisterPair()); 2979 __ adds(out.AsRegisterPairLow<Register>(), 2980 first.AsRegisterPairLow<Register>(), 2981 ShifterOperand(second.AsRegisterPairLow<Register>())); 2982 __ adc(out.AsRegisterPairHigh<Register>(), 2983 first.AsRegisterPairHigh<Register>(), 2984 ShifterOperand(second.AsRegisterPairHigh<Register>())); 2985 } 2986 break; 2987 } 2988 2989 case Primitive::kPrimFloat: 2990 __ vadds(out.AsFpuRegister<SRegister>(), 2991 first.AsFpuRegister<SRegister>(), 2992 second.AsFpuRegister<SRegister>()); 2993 break; 2994 2995 case Primitive::kPrimDouble: 2996 __ vaddd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 2997 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 2998 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 2999 break; 3000 3001 default: 3002 LOG(FATAL) << "Unexpected add type " << add->GetResultType(); 3003 } 3004} 3005 3006void LocationsBuilderARM::VisitSub(HSub* sub) { 3007 LocationSummary* locations = 3008 new (GetGraph()->GetArena()) LocationSummary(sub, LocationSummary::kNoCall); 3009 switch (sub->GetResultType()) { 3010 case Primitive::kPrimInt: { 3011 locations->SetInAt(0, Location::RequiresRegister()); 3012 locations->SetInAt(1, Location::RegisterOrConstant(sub->InputAt(1))); 3013 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3014 break; 3015 } 3016 3017 case Primitive::kPrimLong: { 3018 locations->SetInAt(0, Location::RequiresRegister()); 3019 locations->SetInAt(1, ArmEncodableConstantOrRegister(sub->InputAt(1), SUB)); 3020 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3021 break; 3022 } 3023 case Primitive::kPrimFloat: 3024 case Primitive::kPrimDouble: { 3025 locations->SetInAt(0, Location::RequiresFpuRegister()); 3026 locations->SetInAt(1, Location::RequiresFpuRegister()); 3027 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 3028 break; 3029 } 3030 default: 3031 LOG(FATAL) << "Unexpected sub type " << sub->GetResultType(); 3032 } 3033} 3034 3035void InstructionCodeGeneratorARM::VisitSub(HSub* sub) { 3036 LocationSummary* locations = sub->GetLocations(); 3037 Location out = locations->Out(); 3038 Location first = locations->InAt(0); 3039 Location second = locations->InAt(1); 3040 switch (sub->GetResultType()) { 3041 case Primitive::kPrimInt: { 3042 if (second.IsRegister()) { 3043 __ sub(out.AsRegister<Register>(), 3044 first.AsRegister<Register>(), 3045 ShifterOperand(second.AsRegister<Register>())); 3046 } else { 3047 __ AddConstant(out.AsRegister<Register>(), 3048 first.AsRegister<Register>(), 3049 -second.GetConstant()->AsIntConstant()->GetValue()); 3050 } 3051 break; 3052 } 3053 3054 case Primitive::kPrimLong: { 3055 if (second.IsConstant()) { 3056 uint64_t value = static_cast<uint64_t>(Int64FromConstant(second.GetConstant())); 3057 GenerateAddLongConst(out, first, -value); 3058 } else { 3059 DCHECK(second.IsRegisterPair()); 3060 __ subs(out.AsRegisterPairLow<Register>(), 3061 first.AsRegisterPairLow<Register>(), 3062 ShifterOperand(second.AsRegisterPairLow<Register>())); 3063 __ sbc(out.AsRegisterPairHigh<Register>(), 3064 first.AsRegisterPairHigh<Register>(), 3065 ShifterOperand(second.AsRegisterPairHigh<Register>())); 3066 } 3067 break; 3068 } 3069 3070 case Primitive::kPrimFloat: { 3071 __ vsubs(out.AsFpuRegister<SRegister>(), 3072 first.AsFpuRegister<SRegister>(), 3073 second.AsFpuRegister<SRegister>()); 3074 break; 3075 } 3076 3077 case Primitive::kPrimDouble: { 3078 __ vsubd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 3079 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 3080 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 3081 break; 3082 } 3083 3084 3085 default: 3086 LOG(FATAL) << "Unexpected sub type " << sub->GetResultType(); 3087 } 3088} 3089 3090void LocationsBuilderARM::VisitMul(HMul* mul) { 3091 LocationSummary* locations = 3092 new (GetGraph()->GetArena()) LocationSummary(mul, LocationSummary::kNoCall); 3093 switch (mul->GetResultType()) { 3094 case Primitive::kPrimInt: 3095 case Primitive::kPrimLong: { 3096 locations->SetInAt(0, Location::RequiresRegister()); 3097 locations->SetInAt(1, Location::RequiresRegister()); 3098 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3099 break; 3100 } 3101 3102 case Primitive::kPrimFloat: 3103 case Primitive::kPrimDouble: { 3104 locations->SetInAt(0, Location::RequiresFpuRegister()); 3105 locations->SetInAt(1, Location::RequiresFpuRegister()); 3106 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 3107 break; 3108 } 3109 3110 default: 3111 LOG(FATAL) << "Unexpected mul type " << mul->GetResultType(); 3112 } 3113} 3114 3115void InstructionCodeGeneratorARM::VisitMul(HMul* mul) { 3116 LocationSummary* locations = mul->GetLocations(); 3117 Location out = locations->Out(); 3118 Location first = locations->InAt(0); 3119 Location second = locations->InAt(1); 3120 switch (mul->GetResultType()) { 3121 case Primitive::kPrimInt: { 3122 __ mul(out.AsRegister<Register>(), 3123 first.AsRegister<Register>(), 3124 second.AsRegister<Register>()); 3125 break; 3126 } 3127 case Primitive::kPrimLong: { 3128 Register out_hi = out.AsRegisterPairHigh<Register>(); 3129 Register out_lo = out.AsRegisterPairLow<Register>(); 3130 Register in1_hi = first.AsRegisterPairHigh<Register>(); 3131 Register in1_lo = first.AsRegisterPairLow<Register>(); 3132 Register in2_hi = second.AsRegisterPairHigh<Register>(); 3133 Register in2_lo = second.AsRegisterPairLow<Register>(); 3134 3135 // Extra checks to protect caused by the existence of R1_R2. 3136 // The algorithm is wrong if out.hi is either in1.lo or in2.lo: 3137 // (e.g. in1=r0_r1, in2=r2_r3 and out=r1_r2); 3138 DCHECK_NE(out_hi, in1_lo); 3139 DCHECK_NE(out_hi, in2_lo); 3140 3141 // input: in1 - 64 bits, in2 - 64 bits 3142 // output: out 3143 // formula: out.hi : out.lo = (in1.lo * in2.hi + in1.hi * in2.lo)* 2^32 + in1.lo * in2.lo 3144 // parts: out.hi = in1.lo * in2.hi + in1.hi * in2.lo + (in1.lo * in2.lo)[63:32] 3145 // parts: out.lo = (in1.lo * in2.lo)[31:0] 3146 3147 // IP <- in1.lo * in2.hi 3148 __ mul(IP, in1_lo, in2_hi); 3149 // out.hi <- in1.lo * in2.hi + in1.hi * in2.lo 3150 __ mla(out_hi, in1_hi, in2_lo, IP); 3151 // out.lo <- (in1.lo * in2.lo)[31:0]; 3152 __ umull(out_lo, IP, in1_lo, in2_lo); 3153 // out.hi <- in2.hi * in1.lo + in2.lo * in1.hi + (in1.lo * in2.lo)[63:32] 3154 __ add(out_hi, out_hi, ShifterOperand(IP)); 3155 break; 3156 } 3157 3158 case Primitive::kPrimFloat: { 3159 __ vmuls(out.AsFpuRegister<SRegister>(), 3160 first.AsFpuRegister<SRegister>(), 3161 second.AsFpuRegister<SRegister>()); 3162 break; 3163 } 3164 3165 case Primitive::kPrimDouble: { 3166 __ vmuld(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 3167 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 3168 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 3169 break; 3170 } 3171 3172 default: 3173 LOG(FATAL) << "Unexpected mul type " << mul->GetResultType(); 3174 } 3175} 3176 3177void InstructionCodeGeneratorARM::DivRemOneOrMinusOne(HBinaryOperation* instruction) { 3178 DCHECK(instruction->IsDiv() || instruction->IsRem()); 3179 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 3180 3181 LocationSummary* locations = instruction->GetLocations(); 3182 Location second = locations->InAt(1); 3183 DCHECK(second.IsConstant()); 3184 3185 Register out = locations->Out().AsRegister<Register>(); 3186 Register dividend = locations->InAt(0).AsRegister<Register>(); 3187 int32_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 3188 DCHECK(imm == 1 || imm == -1); 3189 3190 if (instruction->IsRem()) { 3191 __ LoadImmediate(out, 0); 3192 } else { 3193 if (imm == 1) { 3194 __ Mov(out, dividend); 3195 } else { 3196 __ rsb(out, dividend, ShifterOperand(0)); 3197 } 3198 } 3199} 3200 3201void InstructionCodeGeneratorARM::DivRemByPowerOfTwo(HBinaryOperation* instruction) { 3202 DCHECK(instruction->IsDiv() || instruction->IsRem()); 3203 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 3204 3205 LocationSummary* locations = instruction->GetLocations(); 3206 Location second = locations->InAt(1); 3207 DCHECK(second.IsConstant()); 3208 3209 Register out = locations->Out().AsRegister<Register>(); 3210 Register dividend = locations->InAt(0).AsRegister<Register>(); 3211 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3212 int32_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 3213 uint32_t abs_imm = static_cast<uint32_t>(AbsOrMin(imm)); 3214 int ctz_imm = CTZ(abs_imm); 3215 3216 if (ctz_imm == 1) { 3217 __ Lsr(temp, dividend, 32 - ctz_imm); 3218 } else { 3219 __ Asr(temp, dividend, 31); 3220 __ Lsr(temp, temp, 32 - ctz_imm); 3221 } 3222 __ add(out, temp, ShifterOperand(dividend)); 3223 3224 if (instruction->IsDiv()) { 3225 __ Asr(out, out, ctz_imm); 3226 if (imm < 0) { 3227 __ rsb(out, out, ShifterOperand(0)); 3228 } 3229 } else { 3230 __ ubfx(out, out, 0, ctz_imm); 3231 __ sub(out, out, ShifterOperand(temp)); 3232 } 3233} 3234 3235void InstructionCodeGeneratorARM::GenerateDivRemWithAnyConstant(HBinaryOperation* instruction) { 3236 DCHECK(instruction->IsDiv() || instruction->IsRem()); 3237 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 3238 3239 LocationSummary* locations = instruction->GetLocations(); 3240 Location second = locations->InAt(1); 3241 DCHECK(second.IsConstant()); 3242 3243 Register out = locations->Out().AsRegister<Register>(); 3244 Register dividend = locations->InAt(0).AsRegister<Register>(); 3245 Register temp1 = locations->GetTemp(0).AsRegister<Register>(); 3246 Register temp2 = locations->GetTemp(1).AsRegister<Register>(); 3247 int64_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 3248 3249 int64_t magic; 3250 int shift; 3251 CalculateMagicAndShiftForDivRem(imm, false /* is_long */, &magic, &shift); 3252 3253 __ LoadImmediate(temp1, magic); 3254 __ smull(temp2, temp1, dividend, temp1); 3255 3256 if (imm > 0 && magic < 0) { 3257 __ add(temp1, temp1, ShifterOperand(dividend)); 3258 } else if (imm < 0 && magic > 0) { 3259 __ sub(temp1, temp1, ShifterOperand(dividend)); 3260 } 3261 3262 if (shift != 0) { 3263 __ Asr(temp1, temp1, shift); 3264 } 3265 3266 if (instruction->IsDiv()) { 3267 __ sub(out, temp1, ShifterOperand(temp1, ASR, 31)); 3268 } else { 3269 __ sub(temp1, temp1, ShifterOperand(temp1, ASR, 31)); 3270 // TODO: Strength reduction for mls. 3271 __ LoadImmediate(temp2, imm); 3272 __ mls(out, temp1, temp2, dividend); 3273 } 3274} 3275 3276void InstructionCodeGeneratorARM::GenerateDivRemConstantIntegral(HBinaryOperation* instruction) { 3277 DCHECK(instruction->IsDiv() || instruction->IsRem()); 3278 DCHECK(instruction->GetResultType() == Primitive::kPrimInt); 3279 3280 LocationSummary* locations = instruction->GetLocations(); 3281 Location second = locations->InAt(1); 3282 DCHECK(second.IsConstant()); 3283 3284 int32_t imm = second.GetConstant()->AsIntConstant()->GetValue(); 3285 if (imm == 0) { 3286 // Do not generate anything. DivZeroCheck would prevent any code to be executed. 3287 } else if (imm == 1 || imm == -1) { 3288 DivRemOneOrMinusOne(instruction); 3289 } else if (IsPowerOfTwo(AbsOrMin(imm))) { 3290 DivRemByPowerOfTwo(instruction); 3291 } else { 3292 DCHECK(imm <= -2 || imm >= 2); 3293 GenerateDivRemWithAnyConstant(instruction); 3294 } 3295} 3296 3297void LocationsBuilderARM::VisitDiv(HDiv* div) { 3298 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 3299 if (div->GetResultType() == Primitive::kPrimLong) { 3300 // pLdiv runtime call. 3301 call_kind = LocationSummary::kCallOnMainOnly; 3302 } else if (div->GetResultType() == Primitive::kPrimInt && div->InputAt(1)->IsConstant()) { 3303 // sdiv will be replaced by other instruction sequence. 3304 } else if (div->GetResultType() == Primitive::kPrimInt && 3305 !codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 3306 // pIdivmod runtime call. 3307 call_kind = LocationSummary::kCallOnMainOnly; 3308 } 3309 3310 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(div, call_kind); 3311 3312 switch (div->GetResultType()) { 3313 case Primitive::kPrimInt: { 3314 if (div->InputAt(1)->IsConstant()) { 3315 locations->SetInAt(0, Location::RequiresRegister()); 3316 locations->SetInAt(1, Location::ConstantLocation(div->InputAt(1)->AsConstant())); 3317 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3318 int32_t value = div->InputAt(1)->AsIntConstant()->GetValue(); 3319 if (value == 1 || value == 0 || value == -1) { 3320 // No temp register required. 3321 } else { 3322 locations->AddTemp(Location::RequiresRegister()); 3323 if (!IsPowerOfTwo(AbsOrMin(value))) { 3324 locations->AddTemp(Location::RequiresRegister()); 3325 } 3326 } 3327 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 3328 locations->SetInAt(0, Location::RequiresRegister()); 3329 locations->SetInAt(1, Location::RequiresRegister()); 3330 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3331 } else { 3332 InvokeRuntimeCallingConvention calling_convention; 3333 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 3334 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 3335 // Note: divrem will compute both the quotient and the remainder as the pair R0 and R1, but 3336 // we only need the former. 3337 locations->SetOut(Location::RegisterLocation(R0)); 3338 } 3339 break; 3340 } 3341 case Primitive::kPrimLong: { 3342 InvokeRuntimeCallingConvention calling_convention; 3343 locations->SetInAt(0, Location::RegisterPairLocation( 3344 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1))); 3345 locations->SetInAt(1, Location::RegisterPairLocation( 3346 calling_convention.GetRegisterAt(2), calling_convention.GetRegisterAt(3))); 3347 locations->SetOut(Location::RegisterPairLocation(R0, R1)); 3348 break; 3349 } 3350 case Primitive::kPrimFloat: 3351 case Primitive::kPrimDouble: { 3352 locations->SetInAt(0, Location::RequiresFpuRegister()); 3353 locations->SetInAt(1, Location::RequiresFpuRegister()); 3354 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 3355 break; 3356 } 3357 3358 default: 3359 LOG(FATAL) << "Unexpected div type " << div->GetResultType(); 3360 } 3361} 3362 3363void InstructionCodeGeneratorARM::VisitDiv(HDiv* div) { 3364 LocationSummary* locations = div->GetLocations(); 3365 Location out = locations->Out(); 3366 Location first = locations->InAt(0); 3367 Location second = locations->InAt(1); 3368 3369 switch (div->GetResultType()) { 3370 case Primitive::kPrimInt: { 3371 if (second.IsConstant()) { 3372 GenerateDivRemConstantIntegral(div); 3373 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 3374 __ sdiv(out.AsRegister<Register>(), 3375 first.AsRegister<Register>(), 3376 second.AsRegister<Register>()); 3377 } else { 3378 InvokeRuntimeCallingConvention calling_convention; 3379 DCHECK_EQ(calling_convention.GetRegisterAt(0), first.AsRegister<Register>()); 3380 DCHECK_EQ(calling_convention.GetRegisterAt(1), second.AsRegister<Register>()); 3381 DCHECK_EQ(R0, out.AsRegister<Register>()); 3382 3383 codegen_->InvokeRuntime(kQuickIdivmod, div, div->GetDexPc()); 3384 CheckEntrypointTypes<kQuickIdivmod, int32_t, int32_t, int32_t>(); 3385 } 3386 break; 3387 } 3388 3389 case Primitive::kPrimLong: { 3390 InvokeRuntimeCallingConvention calling_convention; 3391 DCHECK_EQ(calling_convention.GetRegisterAt(0), first.AsRegisterPairLow<Register>()); 3392 DCHECK_EQ(calling_convention.GetRegisterAt(1), first.AsRegisterPairHigh<Register>()); 3393 DCHECK_EQ(calling_convention.GetRegisterAt(2), second.AsRegisterPairLow<Register>()); 3394 DCHECK_EQ(calling_convention.GetRegisterAt(3), second.AsRegisterPairHigh<Register>()); 3395 DCHECK_EQ(R0, out.AsRegisterPairLow<Register>()); 3396 DCHECK_EQ(R1, out.AsRegisterPairHigh<Register>()); 3397 3398 codegen_->InvokeRuntime(kQuickLdiv, div, div->GetDexPc()); 3399 CheckEntrypointTypes<kQuickLdiv, int64_t, int64_t, int64_t>(); 3400 break; 3401 } 3402 3403 case Primitive::kPrimFloat: { 3404 __ vdivs(out.AsFpuRegister<SRegister>(), 3405 first.AsFpuRegister<SRegister>(), 3406 second.AsFpuRegister<SRegister>()); 3407 break; 3408 } 3409 3410 case Primitive::kPrimDouble: { 3411 __ vdivd(FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()), 3412 FromLowSToD(first.AsFpuRegisterPairLow<SRegister>()), 3413 FromLowSToD(second.AsFpuRegisterPairLow<SRegister>())); 3414 break; 3415 } 3416 3417 default: 3418 LOG(FATAL) << "Unexpected div type " << div->GetResultType(); 3419 } 3420} 3421 3422void LocationsBuilderARM::VisitRem(HRem* rem) { 3423 Primitive::Type type = rem->GetResultType(); 3424 3425 // Most remainders are implemented in the runtime. 3426 LocationSummary::CallKind call_kind = LocationSummary::kCallOnMainOnly; 3427 if (rem->GetResultType() == Primitive::kPrimInt && rem->InputAt(1)->IsConstant()) { 3428 // sdiv will be replaced by other instruction sequence. 3429 call_kind = LocationSummary::kNoCall; 3430 } else if ((rem->GetResultType() == Primitive::kPrimInt) 3431 && codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 3432 // Have hardware divide instruction for int, do it with three instructions. 3433 call_kind = LocationSummary::kNoCall; 3434 } 3435 3436 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(rem, call_kind); 3437 3438 switch (type) { 3439 case Primitive::kPrimInt: { 3440 if (rem->InputAt(1)->IsConstant()) { 3441 locations->SetInAt(0, Location::RequiresRegister()); 3442 locations->SetInAt(1, Location::ConstantLocation(rem->InputAt(1)->AsConstant())); 3443 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3444 int32_t value = rem->InputAt(1)->AsIntConstant()->GetValue(); 3445 if (value == 1 || value == 0 || value == -1) { 3446 // No temp register required. 3447 } else { 3448 locations->AddTemp(Location::RequiresRegister()); 3449 if (!IsPowerOfTwo(AbsOrMin(value))) { 3450 locations->AddTemp(Location::RequiresRegister()); 3451 } 3452 } 3453 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 3454 locations->SetInAt(0, Location::RequiresRegister()); 3455 locations->SetInAt(1, Location::RequiresRegister()); 3456 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3457 locations->AddTemp(Location::RequiresRegister()); 3458 } else { 3459 InvokeRuntimeCallingConvention calling_convention; 3460 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 3461 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 3462 // Note: divrem will compute both the quotient and the remainder as the pair R0 and R1, but 3463 // we only need the latter. 3464 locations->SetOut(Location::RegisterLocation(R1)); 3465 } 3466 break; 3467 } 3468 case Primitive::kPrimLong: { 3469 InvokeRuntimeCallingConvention calling_convention; 3470 locations->SetInAt(0, Location::RegisterPairLocation( 3471 calling_convention.GetRegisterAt(0), calling_convention.GetRegisterAt(1))); 3472 locations->SetInAt(1, Location::RegisterPairLocation( 3473 calling_convention.GetRegisterAt(2), calling_convention.GetRegisterAt(3))); 3474 // The runtime helper puts the output in R2,R3. 3475 locations->SetOut(Location::RegisterPairLocation(R2, R3)); 3476 break; 3477 } 3478 case Primitive::kPrimFloat: { 3479 InvokeRuntimeCallingConvention calling_convention; 3480 locations->SetInAt(0, Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(0))); 3481 locations->SetInAt(1, Location::FpuRegisterLocation(calling_convention.GetFpuRegisterAt(1))); 3482 locations->SetOut(Location::FpuRegisterLocation(S0)); 3483 break; 3484 } 3485 3486 case Primitive::kPrimDouble: { 3487 InvokeRuntimeCallingConvention calling_convention; 3488 locations->SetInAt(0, Location::FpuRegisterPairLocation( 3489 calling_convention.GetFpuRegisterAt(0), calling_convention.GetFpuRegisterAt(1))); 3490 locations->SetInAt(1, Location::FpuRegisterPairLocation( 3491 calling_convention.GetFpuRegisterAt(2), calling_convention.GetFpuRegisterAt(3))); 3492 locations->SetOut(Location::Location::FpuRegisterPairLocation(S0, S1)); 3493 break; 3494 } 3495 3496 default: 3497 LOG(FATAL) << "Unexpected rem type " << type; 3498 } 3499} 3500 3501void InstructionCodeGeneratorARM::VisitRem(HRem* rem) { 3502 LocationSummary* locations = rem->GetLocations(); 3503 Location out = locations->Out(); 3504 Location first = locations->InAt(0); 3505 Location second = locations->InAt(1); 3506 3507 Primitive::Type type = rem->GetResultType(); 3508 switch (type) { 3509 case Primitive::kPrimInt: { 3510 if (second.IsConstant()) { 3511 GenerateDivRemConstantIntegral(rem); 3512 } else if (codegen_->GetInstructionSetFeatures().HasDivideInstruction()) { 3513 Register reg1 = first.AsRegister<Register>(); 3514 Register reg2 = second.AsRegister<Register>(); 3515 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3516 3517 // temp = reg1 / reg2 (integer division) 3518 // dest = reg1 - temp * reg2 3519 __ sdiv(temp, reg1, reg2); 3520 __ mls(out.AsRegister<Register>(), temp, reg2, reg1); 3521 } else { 3522 InvokeRuntimeCallingConvention calling_convention; 3523 DCHECK_EQ(calling_convention.GetRegisterAt(0), first.AsRegister<Register>()); 3524 DCHECK_EQ(calling_convention.GetRegisterAt(1), second.AsRegister<Register>()); 3525 DCHECK_EQ(R1, out.AsRegister<Register>()); 3526 3527 codegen_->InvokeRuntime(kQuickIdivmod, rem, rem->GetDexPc()); 3528 CheckEntrypointTypes<kQuickIdivmod, int32_t, int32_t, int32_t>(); 3529 } 3530 break; 3531 } 3532 3533 case Primitive::kPrimLong: { 3534 codegen_->InvokeRuntime(kQuickLmod, rem, rem->GetDexPc()); 3535 CheckEntrypointTypes<kQuickLmod, int64_t, int64_t, int64_t>(); 3536 break; 3537 } 3538 3539 case Primitive::kPrimFloat: { 3540 codegen_->InvokeRuntime(kQuickFmodf, rem, rem->GetDexPc()); 3541 CheckEntrypointTypes<kQuickFmodf, float, float, float>(); 3542 break; 3543 } 3544 3545 case Primitive::kPrimDouble: { 3546 codegen_->InvokeRuntime(kQuickFmod, rem, rem->GetDexPc()); 3547 CheckEntrypointTypes<kQuickFmod, double, double, double>(); 3548 break; 3549 } 3550 3551 default: 3552 LOG(FATAL) << "Unexpected rem type " << type; 3553 } 3554} 3555 3556void LocationsBuilderARM::VisitDivZeroCheck(HDivZeroCheck* instruction) { 3557 LocationSummary* locations = codegen_->CreateThrowingSlowPathLocations(instruction); 3558 locations->SetInAt(0, Location::RegisterOrConstant(instruction->InputAt(0))); 3559} 3560 3561void InstructionCodeGeneratorARM::VisitDivZeroCheck(HDivZeroCheck* instruction) { 3562 SlowPathCodeARM* slow_path = new (GetGraph()->GetArena()) DivZeroCheckSlowPathARM(instruction); 3563 codegen_->AddSlowPath(slow_path); 3564 3565 LocationSummary* locations = instruction->GetLocations(); 3566 Location value = locations->InAt(0); 3567 3568 switch (instruction->GetType()) { 3569 case Primitive::kPrimBoolean: 3570 case Primitive::kPrimByte: 3571 case Primitive::kPrimChar: 3572 case Primitive::kPrimShort: 3573 case Primitive::kPrimInt: { 3574 if (value.IsRegister()) { 3575 __ CompareAndBranchIfZero(value.AsRegister<Register>(), slow_path->GetEntryLabel()); 3576 } else { 3577 DCHECK(value.IsConstant()) << value; 3578 if (value.GetConstant()->AsIntConstant()->GetValue() == 0) { 3579 __ b(slow_path->GetEntryLabel()); 3580 } 3581 } 3582 break; 3583 } 3584 case Primitive::kPrimLong: { 3585 if (value.IsRegisterPair()) { 3586 __ orrs(IP, 3587 value.AsRegisterPairLow<Register>(), 3588 ShifterOperand(value.AsRegisterPairHigh<Register>())); 3589 __ b(slow_path->GetEntryLabel(), EQ); 3590 } else { 3591 DCHECK(value.IsConstant()) << value; 3592 if (value.GetConstant()->AsLongConstant()->GetValue() == 0) { 3593 __ b(slow_path->GetEntryLabel()); 3594 } 3595 } 3596 break; 3597 default: 3598 LOG(FATAL) << "Unexpected type for HDivZeroCheck " << instruction->GetType(); 3599 } 3600 } 3601} 3602 3603void InstructionCodeGeneratorARM::HandleIntegerRotate(LocationSummary* locations) { 3604 Register in = locations->InAt(0).AsRegister<Register>(); 3605 Location rhs = locations->InAt(1); 3606 Register out = locations->Out().AsRegister<Register>(); 3607 3608 if (rhs.IsConstant()) { 3609 // Arm32 and Thumb2 assemblers require a rotation on the interval [1,31], 3610 // so map all rotations to a +ve. equivalent in that range. 3611 // (e.g. left *or* right by -2 bits == 30 bits in the same direction.) 3612 uint32_t rot = CodeGenerator::GetInt32ValueOf(rhs.GetConstant()) & 0x1F; 3613 if (rot) { 3614 // Rotate, mapping left rotations to right equivalents if necessary. 3615 // (e.g. left by 2 bits == right by 30.) 3616 __ Ror(out, in, rot); 3617 } else if (out != in) { 3618 __ Mov(out, in); 3619 } 3620 } else { 3621 __ Ror(out, in, rhs.AsRegister<Register>()); 3622 } 3623} 3624 3625// Gain some speed by mapping all Long rotates onto equivalent pairs of Integer 3626// rotates by swapping input regs (effectively rotating by the first 32-bits of 3627// a larger rotation) or flipping direction (thus treating larger right/left 3628// rotations as sub-word sized rotations in the other direction) as appropriate. 3629void InstructionCodeGeneratorARM::HandleLongRotate(LocationSummary* locations) { 3630 Register in_reg_lo = locations->InAt(0).AsRegisterPairLow<Register>(); 3631 Register in_reg_hi = locations->InAt(0).AsRegisterPairHigh<Register>(); 3632 Location rhs = locations->InAt(1); 3633 Register out_reg_lo = locations->Out().AsRegisterPairLow<Register>(); 3634 Register out_reg_hi = locations->Out().AsRegisterPairHigh<Register>(); 3635 3636 if (rhs.IsConstant()) { 3637 uint64_t rot = CodeGenerator::GetInt64ValueOf(rhs.GetConstant()); 3638 // Map all rotations to +ve. equivalents on the interval [0,63]. 3639 rot &= kMaxLongShiftDistance; 3640 // For rotates over a word in size, 'pre-rotate' by 32-bits to keep rotate 3641 // logic below to a simple pair of binary orr. 3642 // (e.g. 34 bits == in_reg swap + 2 bits right.) 3643 if (rot >= kArmBitsPerWord) { 3644 rot -= kArmBitsPerWord; 3645 std::swap(in_reg_hi, in_reg_lo); 3646 } 3647 // Rotate, or mov to out for zero or word size rotations. 3648 if (rot != 0u) { 3649 __ Lsr(out_reg_hi, in_reg_hi, rot); 3650 __ orr(out_reg_hi, out_reg_hi, ShifterOperand(in_reg_lo, arm::LSL, kArmBitsPerWord - rot)); 3651 __ Lsr(out_reg_lo, in_reg_lo, rot); 3652 __ orr(out_reg_lo, out_reg_lo, ShifterOperand(in_reg_hi, arm::LSL, kArmBitsPerWord - rot)); 3653 } else { 3654 __ Mov(out_reg_lo, in_reg_lo); 3655 __ Mov(out_reg_hi, in_reg_hi); 3656 } 3657 } else { 3658 Register shift_right = locations->GetTemp(0).AsRegister<Register>(); 3659 Register shift_left = locations->GetTemp(1).AsRegister<Register>(); 3660 Label end; 3661 Label shift_by_32_plus_shift_right; 3662 3663 __ and_(shift_right, rhs.AsRegister<Register>(), ShifterOperand(0x1F)); 3664 __ Lsrs(shift_left, rhs.AsRegister<Register>(), 6); 3665 __ rsb(shift_left, shift_right, ShifterOperand(kArmBitsPerWord), AL, kCcKeep); 3666 __ b(&shift_by_32_plus_shift_right, CC); 3667 3668 // out_reg_hi = (reg_hi << shift_left) | (reg_lo >> shift_right). 3669 // out_reg_lo = (reg_lo << shift_left) | (reg_hi >> shift_right). 3670 __ Lsl(out_reg_hi, in_reg_hi, shift_left); 3671 __ Lsr(out_reg_lo, in_reg_lo, shift_right); 3672 __ add(out_reg_hi, out_reg_hi, ShifterOperand(out_reg_lo)); 3673 __ Lsl(out_reg_lo, in_reg_lo, shift_left); 3674 __ Lsr(shift_left, in_reg_hi, shift_right); 3675 __ add(out_reg_lo, out_reg_lo, ShifterOperand(shift_left)); 3676 __ b(&end); 3677 3678 __ Bind(&shift_by_32_plus_shift_right); // Shift by 32+shift_right. 3679 // out_reg_hi = (reg_hi >> shift_right) | (reg_lo << shift_left). 3680 // out_reg_lo = (reg_lo >> shift_right) | (reg_hi << shift_left). 3681 __ Lsr(out_reg_hi, in_reg_hi, shift_right); 3682 __ Lsl(out_reg_lo, in_reg_lo, shift_left); 3683 __ add(out_reg_hi, out_reg_hi, ShifterOperand(out_reg_lo)); 3684 __ Lsr(out_reg_lo, in_reg_lo, shift_right); 3685 __ Lsl(shift_right, in_reg_hi, shift_left); 3686 __ add(out_reg_lo, out_reg_lo, ShifterOperand(shift_right)); 3687 3688 __ Bind(&end); 3689 } 3690} 3691 3692void LocationsBuilderARM::VisitRor(HRor* ror) { 3693 LocationSummary* locations = 3694 new (GetGraph()->GetArena()) LocationSummary(ror, LocationSummary::kNoCall); 3695 switch (ror->GetResultType()) { 3696 case Primitive::kPrimInt: { 3697 locations->SetInAt(0, Location::RequiresRegister()); 3698 locations->SetInAt(1, Location::RegisterOrConstant(ror->InputAt(1))); 3699 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3700 break; 3701 } 3702 case Primitive::kPrimLong: { 3703 locations->SetInAt(0, Location::RequiresRegister()); 3704 if (ror->InputAt(1)->IsConstant()) { 3705 locations->SetInAt(1, Location::ConstantLocation(ror->InputAt(1)->AsConstant())); 3706 } else { 3707 locations->SetInAt(1, Location::RequiresRegister()); 3708 locations->AddTemp(Location::RequiresRegister()); 3709 locations->AddTemp(Location::RequiresRegister()); 3710 } 3711 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3712 break; 3713 } 3714 default: 3715 LOG(FATAL) << "Unexpected operation type " << ror->GetResultType(); 3716 } 3717} 3718 3719void InstructionCodeGeneratorARM::VisitRor(HRor* ror) { 3720 LocationSummary* locations = ror->GetLocations(); 3721 Primitive::Type type = ror->GetResultType(); 3722 switch (type) { 3723 case Primitive::kPrimInt: { 3724 HandleIntegerRotate(locations); 3725 break; 3726 } 3727 case Primitive::kPrimLong: { 3728 HandleLongRotate(locations); 3729 break; 3730 } 3731 default: 3732 LOG(FATAL) << "Unexpected operation type " << type; 3733 UNREACHABLE(); 3734 } 3735} 3736 3737void LocationsBuilderARM::HandleShift(HBinaryOperation* op) { 3738 DCHECK(op->IsShl() || op->IsShr() || op->IsUShr()); 3739 3740 LocationSummary* locations = 3741 new (GetGraph()->GetArena()) LocationSummary(op, LocationSummary::kNoCall); 3742 3743 switch (op->GetResultType()) { 3744 case Primitive::kPrimInt: { 3745 locations->SetInAt(0, Location::RequiresRegister()); 3746 if (op->InputAt(1)->IsConstant()) { 3747 locations->SetInAt(1, Location::ConstantLocation(op->InputAt(1)->AsConstant())); 3748 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3749 } else { 3750 locations->SetInAt(1, Location::RequiresRegister()); 3751 // Make the output overlap, as it will be used to hold the masked 3752 // second input. 3753 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3754 } 3755 break; 3756 } 3757 case Primitive::kPrimLong: { 3758 locations->SetInAt(0, Location::RequiresRegister()); 3759 if (op->InputAt(1)->IsConstant()) { 3760 locations->SetInAt(1, Location::ConstantLocation(op->InputAt(1)->AsConstant())); 3761 // For simplicity, use kOutputOverlap even though we only require that low registers 3762 // don't clash with high registers which the register allocator currently guarantees. 3763 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3764 } else { 3765 locations->SetInAt(1, Location::RequiresRegister()); 3766 locations->AddTemp(Location::RequiresRegister()); 3767 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3768 } 3769 break; 3770 } 3771 default: 3772 LOG(FATAL) << "Unexpected operation type " << op->GetResultType(); 3773 } 3774} 3775 3776void InstructionCodeGeneratorARM::HandleShift(HBinaryOperation* op) { 3777 DCHECK(op->IsShl() || op->IsShr() || op->IsUShr()); 3778 3779 LocationSummary* locations = op->GetLocations(); 3780 Location out = locations->Out(); 3781 Location first = locations->InAt(0); 3782 Location second = locations->InAt(1); 3783 3784 Primitive::Type type = op->GetResultType(); 3785 switch (type) { 3786 case Primitive::kPrimInt: { 3787 Register out_reg = out.AsRegister<Register>(); 3788 Register first_reg = first.AsRegister<Register>(); 3789 if (second.IsRegister()) { 3790 Register second_reg = second.AsRegister<Register>(); 3791 // ARM doesn't mask the shift count so we need to do it ourselves. 3792 __ and_(out_reg, second_reg, ShifterOperand(kMaxIntShiftDistance)); 3793 if (op->IsShl()) { 3794 __ Lsl(out_reg, first_reg, out_reg); 3795 } else if (op->IsShr()) { 3796 __ Asr(out_reg, first_reg, out_reg); 3797 } else { 3798 __ Lsr(out_reg, first_reg, out_reg); 3799 } 3800 } else { 3801 int32_t cst = second.GetConstant()->AsIntConstant()->GetValue(); 3802 uint32_t shift_value = cst & kMaxIntShiftDistance; 3803 if (shift_value == 0) { // ARM does not support shifting with 0 immediate. 3804 __ Mov(out_reg, first_reg); 3805 } else if (op->IsShl()) { 3806 __ Lsl(out_reg, first_reg, shift_value); 3807 } else if (op->IsShr()) { 3808 __ Asr(out_reg, first_reg, shift_value); 3809 } else { 3810 __ Lsr(out_reg, first_reg, shift_value); 3811 } 3812 } 3813 break; 3814 } 3815 case Primitive::kPrimLong: { 3816 Register o_h = out.AsRegisterPairHigh<Register>(); 3817 Register o_l = out.AsRegisterPairLow<Register>(); 3818 3819 Register high = first.AsRegisterPairHigh<Register>(); 3820 Register low = first.AsRegisterPairLow<Register>(); 3821 3822 if (second.IsRegister()) { 3823 Register temp = locations->GetTemp(0).AsRegister<Register>(); 3824 3825 Register second_reg = second.AsRegister<Register>(); 3826 3827 if (op->IsShl()) { 3828 __ and_(o_l, second_reg, ShifterOperand(kMaxLongShiftDistance)); 3829 // Shift the high part 3830 __ Lsl(o_h, high, o_l); 3831 // Shift the low part and `or` what overflew on the high part 3832 __ rsb(temp, o_l, ShifterOperand(kArmBitsPerWord)); 3833 __ Lsr(temp, low, temp); 3834 __ orr(o_h, o_h, ShifterOperand(temp)); 3835 // If the shift is > 32 bits, override the high part 3836 __ subs(temp, o_l, ShifterOperand(kArmBitsPerWord)); 3837 __ it(PL); 3838 __ Lsl(o_h, low, temp, PL); 3839 // Shift the low part 3840 __ Lsl(o_l, low, o_l); 3841 } else if (op->IsShr()) { 3842 __ and_(o_h, second_reg, ShifterOperand(kMaxLongShiftDistance)); 3843 // Shift the low part 3844 __ Lsr(o_l, low, o_h); 3845 // Shift the high part and `or` what underflew on the low part 3846 __ rsb(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3847 __ Lsl(temp, high, temp); 3848 __ orr(o_l, o_l, ShifterOperand(temp)); 3849 // If the shift is > 32 bits, override the low part 3850 __ subs(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3851 __ it(PL); 3852 __ Asr(o_l, high, temp, PL); 3853 // Shift the high part 3854 __ Asr(o_h, high, o_h); 3855 } else { 3856 __ and_(o_h, second_reg, ShifterOperand(kMaxLongShiftDistance)); 3857 // same as Shr except we use `Lsr`s and not `Asr`s 3858 __ Lsr(o_l, low, o_h); 3859 __ rsb(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3860 __ Lsl(temp, high, temp); 3861 __ orr(o_l, o_l, ShifterOperand(temp)); 3862 __ subs(temp, o_h, ShifterOperand(kArmBitsPerWord)); 3863 __ it(PL); 3864 __ Lsr(o_l, high, temp, PL); 3865 __ Lsr(o_h, high, o_h); 3866 } 3867 } else { 3868 // Register allocator doesn't create partial overlap. 3869 DCHECK_NE(o_l, high); 3870 DCHECK_NE(o_h, low); 3871 int32_t cst = second.GetConstant()->AsIntConstant()->GetValue(); 3872 uint32_t shift_value = cst & kMaxLongShiftDistance; 3873 if (shift_value > 32) { 3874 if (op->IsShl()) { 3875 __ Lsl(o_h, low, shift_value - 32); 3876 __ LoadImmediate(o_l, 0); 3877 } else if (op->IsShr()) { 3878 __ Asr(o_l, high, shift_value - 32); 3879 __ Asr(o_h, high, 31); 3880 } else { 3881 __ Lsr(o_l, high, shift_value - 32); 3882 __ LoadImmediate(o_h, 0); 3883 } 3884 } else if (shift_value == 32) { 3885 if (op->IsShl()) { 3886 __ mov(o_h, ShifterOperand(low)); 3887 __ LoadImmediate(o_l, 0); 3888 } else if (op->IsShr()) { 3889 __ mov(o_l, ShifterOperand(high)); 3890 __ Asr(o_h, high, 31); 3891 } else { 3892 __ mov(o_l, ShifterOperand(high)); 3893 __ LoadImmediate(o_h, 0); 3894 } 3895 } else if (shift_value == 1) { 3896 if (op->IsShl()) { 3897 __ Lsls(o_l, low, 1); 3898 __ adc(o_h, high, ShifterOperand(high)); 3899 } else if (op->IsShr()) { 3900 __ Asrs(o_h, high, 1); 3901 __ Rrx(o_l, low); 3902 } else { 3903 __ Lsrs(o_h, high, 1); 3904 __ Rrx(o_l, low); 3905 } 3906 } else { 3907 DCHECK(2 <= shift_value && shift_value < 32) << shift_value; 3908 if (op->IsShl()) { 3909 __ Lsl(o_h, high, shift_value); 3910 __ orr(o_h, o_h, ShifterOperand(low, LSR, 32 - shift_value)); 3911 __ Lsl(o_l, low, shift_value); 3912 } else if (op->IsShr()) { 3913 __ Lsr(o_l, low, shift_value); 3914 __ orr(o_l, o_l, ShifterOperand(high, LSL, 32 - shift_value)); 3915 __ Asr(o_h, high, shift_value); 3916 } else { 3917 __ Lsr(o_l, low, shift_value); 3918 __ orr(o_l, o_l, ShifterOperand(high, LSL, 32 - shift_value)); 3919 __ Lsr(o_h, high, shift_value); 3920 } 3921 } 3922 } 3923 break; 3924 } 3925 default: 3926 LOG(FATAL) << "Unexpected operation type " << type; 3927 UNREACHABLE(); 3928 } 3929} 3930 3931void LocationsBuilderARM::VisitShl(HShl* shl) { 3932 HandleShift(shl); 3933} 3934 3935void InstructionCodeGeneratorARM::VisitShl(HShl* shl) { 3936 HandleShift(shl); 3937} 3938 3939void LocationsBuilderARM::VisitShr(HShr* shr) { 3940 HandleShift(shr); 3941} 3942 3943void InstructionCodeGeneratorARM::VisitShr(HShr* shr) { 3944 HandleShift(shr); 3945} 3946 3947void LocationsBuilderARM::VisitUShr(HUShr* ushr) { 3948 HandleShift(ushr); 3949} 3950 3951void InstructionCodeGeneratorARM::VisitUShr(HUShr* ushr) { 3952 HandleShift(ushr); 3953} 3954 3955void LocationsBuilderARM::VisitNewInstance(HNewInstance* instruction) { 3956 LocationSummary* locations = 3957 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnMainOnly); 3958 if (instruction->IsStringAlloc()) { 3959 locations->AddTemp(Location::RegisterLocation(kMethodRegisterArgument)); 3960 } else { 3961 InvokeRuntimeCallingConvention calling_convention; 3962 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 3963 } 3964 locations->SetOut(Location::RegisterLocation(R0)); 3965} 3966 3967void InstructionCodeGeneratorARM::VisitNewInstance(HNewInstance* instruction) { 3968 // Note: if heap poisoning is enabled, the entry point takes cares 3969 // of poisoning the reference. 3970 if (instruction->IsStringAlloc()) { 3971 // String is allocated through StringFactory. Call NewEmptyString entry point. 3972 Register temp = instruction->GetLocations()->GetTemp(0).AsRegister<Register>(); 3973 MemberOffset code_offset = ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmPointerSize); 3974 __ LoadFromOffset(kLoadWord, temp, TR, QUICK_ENTRY_POINT(pNewEmptyString)); 3975 __ LoadFromOffset(kLoadWord, LR, temp, code_offset.Int32Value()); 3976 __ blx(LR); 3977 codegen_->RecordPcInfo(instruction, instruction->GetDexPc()); 3978 } else { 3979 codegen_->InvokeRuntime(instruction->GetEntrypoint(), instruction, instruction->GetDexPc()); 3980 CheckEntrypointTypes<kQuickAllocObjectWithChecks, void*, mirror::Class*>(); 3981 } 3982} 3983 3984void LocationsBuilderARM::VisitNewArray(HNewArray* instruction) { 3985 LocationSummary* locations = 3986 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnMainOnly); 3987 InvokeRuntimeCallingConvention calling_convention; 3988 locations->SetOut(Location::RegisterLocation(R0)); 3989 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 3990 locations->SetInAt(1, Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 3991} 3992 3993void InstructionCodeGeneratorARM::VisitNewArray(HNewArray* instruction) { 3994 // Note: if heap poisoning is enabled, the entry point takes cares 3995 // of poisoning the reference. 3996 codegen_->InvokeRuntime(kQuickAllocArrayResolved, instruction, instruction->GetDexPc()); 3997 CheckEntrypointTypes<kQuickAllocArrayResolved, void*, mirror::Class*, int32_t>(); 3998} 3999 4000void LocationsBuilderARM::VisitParameterValue(HParameterValue* instruction) { 4001 LocationSummary* locations = 4002 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 4003 Location location = parameter_visitor_.GetNextLocation(instruction->GetType()); 4004 if (location.IsStackSlot()) { 4005 location = Location::StackSlot(location.GetStackIndex() + codegen_->GetFrameSize()); 4006 } else if (location.IsDoubleStackSlot()) { 4007 location = Location::DoubleStackSlot(location.GetStackIndex() + codegen_->GetFrameSize()); 4008 } 4009 locations->SetOut(location); 4010} 4011 4012void InstructionCodeGeneratorARM::VisitParameterValue( 4013 HParameterValue* instruction ATTRIBUTE_UNUSED) { 4014 // Nothing to do, the parameter is already at its location. 4015} 4016 4017void LocationsBuilderARM::VisitCurrentMethod(HCurrentMethod* instruction) { 4018 LocationSummary* locations = 4019 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 4020 locations->SetOut(Location::RegisterLocation(kMethodRegisterArgument)); 4021} 4022 4023void InstructionCodeGeneratorARM::VisitCurrentMethod(HCurrentMethod* instruction ATTRIBUTE_UNUSED) { 4024 // Nothing to do, the method is already at its location. 4025} 4026 4027void LocationsBuilderARM::VisitNot(HNot* not_) { 4028 LocationSummary* locations = 4029 new (GetGraph()->GetArena()) LocationSummary(not_, LocationSummary::kNoCall); 4030 locations->SetInAt(0, Location::RequiresRegister()); 4031 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 4032} 4033 4034void InstructionCodeGeneratorARM::VisitNot(HNot* not_) { 4035 LocationSummary* locations = not_->GetLocations(); 4036 Location out = locations->Out(); 4037 Location in = locations->InAt(0); 4038 switch (not_->GetResultType()) { 4039 case Primitive::kPrimInt: 4040 __ mvn(out.AsRegister<Register>(), ShifterOperand(in.AsRegister<Register>())); 4041 break; 4042 4043 case Primitive::kPrimLong: 4044 __ mvn(out.AsRegisterPairLow<Register>(), 4045 ShifterOperand(in.AsRegisterPairLow<Register>())); 4046 __ mvn(out.AsRegisterPairHigh<Register>(), 4047 ShifterOperand(in.AsRegisterPairHigh<Register>())); 4048 break; 4049 4050 default: 4051 LOG(FATAL) << "Unimplemented type for not operation " << not_->GetResultType(); 4052 } 4053} 4054 4055void LocationsBuilderARM::VisitBooleanNot(HBooleanNot* bool_not) { 4056 LocationSummary* locations = 4057 new (GetGraph()->GetArena()) LocationSummary(bool_not, LocationSummary::kNoCall); 4058 locations->SetInAt(0, Location::RequiresRegister()); 4059 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 4060} 4061 4062void InstructionCodeGeneratorARM::VisitBooleanNot(HBooleanNot* bool_not) { 4063 LocationSummary* locations = bool_not->GetLocations(); 4064 Location out = locations->Out(); 4065 Location in = locations->InAt(0); 4066 __ eor(out.AsRegister<Register>(), in.AsRegister<Register>(), ShifterOperand(1)); 4067} 4068 4069void LocationsBuilderARM::VisitCompare(HCompare* compare) { 4070 LocationSummary* locations = 4071 new (GetGraph()->GetArena()) LocationSummary(compare, LocationSummary::kNoCall); 4072 switch (compare->InputAt(0)->GetType()) { 4073 case Primitive::kPrimBoolean: 4074 case Primitive::kPrimByte: 4075 case Primitive::kPrimShort: 4076 case Primitive::kPrimChar: 4077 case Primitive::kPrimInt: 4078 case Primitive::kPrimLong: { 4079 locations->SetInAt(0, Location::RequiresRegister()); 4080 locations->SetInAt(1, Location::RequiresRegister()); 4081 // Output overlaps because it is written before doing the low comparison. 4082 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 4083 break; 4084 } 4085 case Primitive::kPrimFloat: 4086 case Primitive::kPrimDouble: { 4087 locations->SetInAt(0, Location::RequiresFpuRegister()); 4088 locations->SetInAt(1, ArithmeticZeroOrFpuRegister(compare->InputAt(1))); 4089 locations->SetOut(Location::RequiresRegister()); 4090 break; 4091 } 4092 default: 4093 LOG(FATAL) << "Unexpected type for compare operation " << compare->InputAt(0)->GetType(); 4094 } 4095} 4096 4097void InstructionCodeGeneratorARM::VisitCompare(HCompare* compare) { 4098 LocationSummary* locations = compare->GetLocations(); 4099 Register out = locations->Out().AsRegister<Register>(); 4100 Location left = locations->InAt(0); 4101 Location right = locations->InAt(1); 4102 4103 Label less, greater, done; 4104 Primitive::Type type = compare->InputAt(0)->GetType(); 4105 Condition less_cond; 4106 switch (type) { 4107 case Primitive::kPrimBoolean: 4108 case Primitive::kPrimByte: 4109 case Primitive::kPrimShort: 4110 case Primitive::kPrimChar: 4111 case Primitive::kPrimInt: { 4112 __ LoadImmediate(out, 0); 4113 __ cmp(left.AsRegister<Register>(), 4114 ShifterOperand(right.AsRegister<Register>())); // Signed compare. 4115 less_cond = LT; 4116 break; 4117 } 4118 case Primitive::kPrimLong: { 4119 __ cmp(left.AsRegisterPairHigh<Register>(), 4120 ShifterOperand(right.AsRegisterPairHigh<Register>())); // Signed compare. 4121 __ b(&less, LT); 4122 __ b(&greater, GT); 4123 // Do LoadImmediate before the last `cmp`, as LoadImmediate might affect the status flags. 4124 __ LoadImmediate(out, 0); 4125 __ cmp(left.AsRegisterPairLow<Register>(), 4126 ShifterOperand(right.AsRegisterPairLow<Register>())); // Unsigned compare. 4127 less_cond = LO; 4128 break; 4129 } 4130 case Primitive::kPrimFloat: 4131 case Primitive::kPrimDouble: { 4132 __ LoadImmediate(out, 0); 4133 GenerateVcmp(compare); 4134 __ vmstat(); // transfer FP status register to ARM APSR. 4135 less_cond = ARMFPCondition(kCondLT, compare->IsGtBias()); 4136 break; 4137 } 4138 default: 4139 LOG(FATAL) << "Unexpected compare type " << type; 4140 UNREACHABLE(); 4141 } 4142 4143 __ b(&done, EQ); 4144 __ b(&less, less_cond); 4145 4146 __ Bind(&greater); 4147 __ LoadImmediate(out, 1); 4148 __ b(&done); 4149 4150 __ Bind(&less); 4151 __ LoadImmediate(out, -1); 4152 4153 __ Bind(&done); 4154} 4155 4156void LocationsBuilderARM::VisitPhi(HPhi* instruction) { 4157 LocationSummary* locations = 4158 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 4159 for (size_t i = 0, e = locations->GetInputCount(); i < e; ++i) { 4160 locations->SetInAt(i, Location::Any()); 4161 } 4162 locations->SetOut(Location::Any()); 4163} 4164 4165void InstructionCodeGeneratorARM::VisitPhi(HPhi* instruction ATTRIBUTE_UNUSED) { 4166 LOG(FATAL) << "Unreachable"; 4167} 4168 4169void CodeGeneratorARM::GenerateMemoryBarrier(MemBarrierKind kind) { 4170 // TODO (ported from quick): revisit ARM barrier kinds. 4171 DmbOptions flavor = DmbOptions::ISH; // Quiet C++ warnings. 4172 switch (kind) { 4173 case MemBarrierKind::kAnyStore: 4174 case MemBarrierKind::kLoadAny: 4175 case MemBarrierKind::kAnyAny: { 4176 flavor = DmbOptions::ISH; 4177 break; 4178 } 4179 case MemBarrierKind::kStoreStore: { 4180 flavor = DmbOptions::ISHST; 4181 break; 4182 } 4183 default: 4184 LOG(FATAL) << "Unexpected memory barrier " << kind; 4185 } 4186 __ dmb(flavor); 4187} 4188 4189void InstructionCodeGeneratorARM::GenerateWideAtomicLoad(Register addr, 4190 uint32_t offset, 4191 Register out_lo, 4192 Register out_hi) { 4193 if (offset != 0) { 4194 // Ensure `out_lo` is different from `addr`, so that loading 4195 // `offset` into `out_lo` does not clutter `addr`. 4196 DCHECK_NE(out_lo, addr); 4197 __ LoadImmediate(out_lo, offset); 4198 __ add(IP, addr, ShifterOperand(out_lo)); 4199 addr = IP; 4200 } 4201 __ ldrexd(out_lo, out_hi, addr); 4202} 4203 4204void InstructionCodeGeneratorARM::GenerateWideAtomicStore(Register addr, 4205 uint32_t offset, 4206 Register value_lo, 4207 Register value_hi, 4208 Register temp1, 4209 Register temp2, 4210 HInstruction* instruction) { 4211 Label fail; 4212 if (offset != 0) { 4213 __ LoadImmediate(temp1, offset); 4214 __ add(IP, addr, ShifterOperand(temp1)); 4215 addr = IP; 4216 } 4217 __ Bind(&fail); 4218 // We need a load followed by store. (The address used in a STREX instruction must 4219 // be the same as the address in the most recently executed LDREX instruction.) 4220 __ ldrexd(temp1, temp2, addr); 4221 codegen_->MaybeRecordImplicitNullCheck(instruction); 4222 __ strexd(temp1, value_lo, value_hi, addr); 4223 __ CompareAndBranchIfNonZero(temp1, &fail); 4224} 4225 4226void LocationsBuilderARM::HandleFieldSet(HInstruction* instruction, const FieldInfo& field_info) { 4227 DCHECK(instruction->IsInstanceFieldSet() || instruction->IsStaticFieldSet()); 4228 4229 LocationSummary* locations = 4230 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 4231 locations->SetInAt(0, Location::RequiresRegister()); 4232 4233 Primitive::Type field_type = field_info.GetFieldType(); 4234 if (Primitive::IsFloatingPointType(field_type)) { 4235 locations->SetInAt(1, Location::RequiresFpuRegister()); 4236 } else { 4237 locations->SetInAt(1, Location::RequiresRegister()); 4238 } 4239 4240 bool is_wide = field_type == Primitive::kPrimLong || field_type == Primitive::kPrimDouble; 4241 bool generate_volatile = field_info.IsVolatile() 4242 && is_wide 4243 && !codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 4244 bool needs_write_barrier = 4245 CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1)); 4246 // Temporary registers for the write barrier. 4247 // TODO: consider renaming StoreNeedsWriteBarrier to StoreNeedsGCMark. 4248 if (needs_write_barrier) { 4249 locations->AddTemp(Location::RequiresRegister()); // Possibly used for reference poisoning too. 4250 locations->AddTemp(Location::RequiresRegister()); 4251 } else if (generate_volatile) { 4252 // ARM encoding have some additional constraints for ldrexd/strexd: 4253 // - registers need to be consecutive 4254 // - the first register should be even but not R14. 4255 // We don't test for ARM yet, and the assertion makes sure that we 4256 // revisit this if we ever enable ARM encoding. 4257 DCHECK_EQ(InstructionSet::kThumb2, codegen_->GetInstructionSet()); 4258 4259 locations->AddTemp(Location::RequiresRegister()); 4260 locations->AddTemp(Location::RequiresRegister()); 4261 if (field_type == Primitive::kPrimDouble) { 4262 // For doubles we need two more registers to copy the value. 4263 locations->AddTemp(Location::RegisterLocation(R2)); 4264 locations->AddTemp(Location::RegisterLocation(R3)); 4265 } 4266 } 4267} 4268 4269void InstructionCodeGeneratorARM::HandleFieldSet(HInstruction* instruction, 4270 const FieldInfo& field_info, 4271 bool value_can_be_null) { 4272 DCHECK(instruction->IsInstanceFieldSet() || instruction->IsStaticFieldSet()); 4273 4274 LocationSummary* locations = instruction->GetLocations(); 4275 Register base = locations->InAt(0).AsRegister<Register>(); 4276 Location value = locations->InAt(1); 4277 4278 bool is_volatile = field_info.IsVolatile(); 4279 bool atomic_ldrd_strd = codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 4280 Primitive::Type field_type = field_info.GetFieldType(); 4281 uint32_t offset = field_info.GetFieldOffset().Uint32Value(); 4282 bool needs_write_barrier = 4283 CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1)); 4284 4285 if (is_volatile) { 4286 codegen_->GenerateMemoryBarrier(MemBarrierKind::kAnyStore); 4287 } 4288 4289 switch (field_type) { 4290 case Primitive::kPrimBoolean: 4291 case Primitive::kPrimByte: { 4292 __ StoreToOffset(kStoreByte, value.AsRegister<Register>(), base, offset); 4293 break; 4294 } 4295 4296 case Primitive::kPrimShort: 4297 case Primitive::kPrimChar: { 4298 __ StoreToOffset(kStoreHalfword, value.AsRegister<Register>(), base, offset); 4299 break; 4300 } 4301 4302 case Primitive::kPrimInt: 4303 case Primitive::kPrimNot: { 4304 if (kPoisonHeapReferences && needs_write_barrier) { 4305 // Note that in the case where `value` is a null reference, 4306 // we do not enter this block, as a null reference does not 4307 // need poisoning. 4308 DCHECK_EQ(field_type, Primitive::kPrimNot); 4309 Register temp = locations->GetTemp(0).AsRegister<Register>(); 4310 __ Mov(temp, value.AsRegister<Register>()); 4311 __ PoisonHeapReference(temp); 4312 __ StoreToOffset(kStoreWord, temp, base, offset); 4313 } else { 4314 __ StoreToOffset(kStoreWord, value.AsRegister<Register>(), base, offset); 4315 } 4316 break; 4317 } 4318 4319 case Primitive::kPrimLong: { 4320 if (is_volatile && !atomic_ldrd_strd) { 4321 GenerateWideAtomicStore(base, offset, 4322 value.AsRegisterPairLow<Register>(), 4323 value.AsRegisterPairHigh<Register>(), 4324 locations->GetTemp(0).AsRegister<Register>(), 4325 locations->GetTemp(1).AsRegister<Register>(), 4326 instruction); 4327 } else { 4328 __ StoreToOffset(kStoreWordPair, value.AsRegisterPairLow<Register>(), base, offset); 4329 codegen_->MaybeRecordImplicitNullCheck(instruction); 4330 } 4331 break; 4332 } 4333 4334 case Primitive::kPrimFloat: { 4335 __ StoreSToOffset(value.AsFpuRegister<SRegister>(), base, offset); 4336 break; 4337 } 4338 4339 case Primitive::kPrimDouble: { 4340 DRegister value_reg = FromLowSToD(value.AsFpuRegisterPairLow<SRegister>()); 4341 if (is_volatile && !atomic_ldrd_strd) { 4342 Register value_reg_lo = locations->GetTemp(0).AsRegister<Register>(); 4343 Register value_reg_hi = locations->GetTemp(1).AsRegister<Register>(); 4344 4345 __ vmovrrd(value_reg_lo, value_reg_hi, value_reg); 4346 4347 GenerateWideAtomicStore(base, offset, 4348 value_reg_lo, 4349 value_reg_hi, 4350 locations->GetTemp(2).AsRegister<Register>(), 4351 locations->GetTemp(3).AsRegister<Register>(), 4352 instruction); 4353 } else { 4354 __ StoreDToOffset(value_reg, base, offset); 4355 codegen_->MaybeRecordImplicitNullCheck(instruction); 4356 } 4357 break; 4358 } 4359 4360 case Primitive::kPrimVoid: 4361 LOG(FATAL) << "Unreachable type " << field_type; 4362 UNREACHABLE(); 4363 } 4364 4365 // Longs and doubles are handled in the switch. 4366 if (field_type != Primitive::kPrimLong && field_type != Primitive::kPrimDouble) { 4367 codegen_->MaybeRecordImplicitNullCheck(instruction); 4368 } 4369 4370 if (CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1))) { 4371 Register temp = locations->GetTemp(0).AsRegister<Register>(); 4372 Register card = locations->GetTemp(1).AsRegister<Register>(); 4373 codegen_->MarkGCCard( 4374 temp, card, base, value.AsRegister<Register>(), value_can_be_null); 4375 } 4376 4377 if (is_volatile) { 4378 codegen_->GenerateMemoryBarrier(MemBarrierKind::kAnyAny); 4379 } 4380} 4381 4382void LocationsBuilderARM::HandleFieldGet(HInstruction* instruction, const FieldInfo& field_info) { 4383 DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet()); 4384 4385 bool object_field_get_with_read_barrier = 4386 kEmitCompilerReadBarrier && (field_info.GetFieldType() == Primitive::kPrimNot); 4387 LocationSummary* locations = 4388 new (GetGraph()->GetArena()) LocationSummary(instruction, 4389 object_field_get_with_read_barrier ? 4390 LocationSummary::kCallOnSlowPath : 4391 LocationSummary::kNoCall); 4392 if (object_field_get_with_read_barrier && kUseBakerReadBarrier) { 4393 locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty()); // No caller-save registers. 4394 } 4395 locations->SetInAt(0, Location::RequiresRegister()); 4396 4397 bool volatile_for_double = field_info.IsVolatile() 4398 && (field_info.GetFieldType() == Primitive::kPrimDouble) 4399 && !codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 4400 // The output overlaps in case of volatile long: we don't want the 4401 // code generated by GenerateWideAtomicLoad to overwrite the 4402 // object's location. Likewise, in the case of an object field get 4403 // with read barriers enabled, we do not want the load to overwrite 4404 // the object's location, as we need it to emit the read barrier. 4405 bool overlap = (field_info.IsVolatile() && (field_info.GetFieldType() == Primitive::kPrimLong)) || 4406 object_field_get_with_read_barrier; 4407 4408 if (Primitive::IsFloatingPointType(instruction->GetType())) { 4409 locations->SetOut(Location::RequiresFpuRegister()); 4410 } else { 4411 locations->SetOut(Location::RequiresRegister(), 4412 (overlap ? Location::kOutputOverlap : Location::kNoOutputOverlap)); 4413 } 4414 if (volatile_for_double) { 4415 // ARM encoding have some additional constraints for ldrexd/strexd: 4416 // - registers need to be consecutive 4417 // - the first register should be even but not R14. 4418 // We don't test for ARM yet, and the assertion makes sure that we 4419 // revisit this if we ever enable ARM encoding. 4420 DCHECK_EQ(InstructionSet::kThumb2, codegen_->GetInstructionSet()); 4421 locations->AddTemp(Location::RequiresRegister()); 4422 locations->AddTemp(Location::RequiresRegister()); 4423 } else if (object_field_get_with_read_barrier && kUseBakerReadBarrier) { 4424 // We need a temporary register for the read barrier marking slow 4425 // path in CodeGeneratorARM::GenerateFieldLoadWithBakerReadBarrier. 4426 locations->AddTemp(Location::RequiresRegister()); 4427 } 4428} 4429 4430Location LocationsBuilderARM::ArithmeticZeroOrFpuRegister(HInstruction* input) { 4431 DCHECK(input->GetType() == Primitive::kPrimDouble || input->GetType() == Primitive::kPrimFloat) 4432 << input->GetType(); 4433 if ((input->IsFloatConstant() && (input->AsFloatConstant()->IsArithmeticZero())) || 4434 (input->IsDoubleConstant() && (input->AsDoubleConstant()->IsArithmeticZero()))) { 4435 return Location::ConstantLocation(input->AsConstant()); 4436 } else { 4437 return Location::RequiresFpuRegister(); 4438 } 4439} 4440 4441Location LocationsBuilderARM::ArmEncodableConstantOrRegister(HInstruction* constant, 4442 Opcode opcode) { 4443 DCHECK(!Primitive::IsFloatingPointType(constant->GetType())); 4444 if (constant->IsConstant() && 4445 CanEncodeConstantAsImmediate(constant->AsConstant(), opcode)) { 4446 return Location::ConstantLocation(constant->AsConstant()); 4447 } 4448 return Location::RequiresRegister(); 4449} 4450 4451bool LocationsBuilderARM::CanEncodeConstantAsImmediate(HConstant* input_cst, 4452 Opcode opcode) { 4453 uint64_t value = static_cast<uint64_t>(Int64FromConstant(input_cst)); 4454 if (Primitive::Is64BitType(input_cst->GetType())) { 4455 Opcode high_opcode = opcode; 4456 SetCc low_set_cc = kCcDontCare; 4457 switch (opcode) { 4458 case SUB: 4459 // Flip the operation to an ADD. 4460 value = -value; 4461 opcode = ADD; 4462 FALLTHROUGH_INTENDED; 4463 case ADD: 4464 if (Low32Bits(value) == 0u) { 4465 return CanEncodeConstantAsImmediate(High32Bits(value), opcode, kCcDontCare); 4466 } 4467 high_opcode = ADC; 4468 low_set_cc = kCcSet; 4469 break; 4470 default: 4471 break; 4472 } 4473 return CanEncodeConstantAsImmediate(Low32Bits(value), opcode, low_set_cc) && 4474 CanEncodeConstantAsImmediate(High32Bits(value), high_opcode, kCcDontCare); 4475 } else { 4476 return CanEncodeConstantAsImmediate(Low32Bits(value), opcode); 4477 } 4478} 4479 4480bool LocationsBuilderARM::CanEncodeConstantAsImmediate(uint32_t value, 4481 Opcode opcode, 4482 SetCc set_cc) { 4483 ShifterOperand so; 4484 ArmAssembler* assembler = codegen_->GetAssembler(); 4485 if (assembler->ShifterOperandCanHold(kNoRegister, kNoRegister, opcode, value, set_cc, &so)) { 4486 return true; 4487 } 4488 Opcode neg_opcode = kNoOperand; 4489 switch (opcode) { 4490 case AND: neg_opcode = BIC; value = ~value; break; 4491 case ORR: neg_opcode = ORN; value = ~value; break; 4492 case ADD: neg_opcode = SUB; value = -value; break; 4493 case ADC: neg_opcode = SBC; value = ~value; break; 4494 case SUB: neg_opcode = ADD; value = -value; break; 4495 case SBC: neg_opcode = ADC; value = ~value; break; 4496 default: 4497 return false; 4498 } 4499 return assembler->ShifterOperandCanHold(kNoRegister, kNoRegister, neg_opcode, value, set_cc, &so); 4500} 4501 4502void InstructionCodeGeneratorARM::HandleFieldGet(HInstruction* instruction, 4503 const FieldInfo& field_info) { 4504 DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet()); 4505 4506 LocationSummary* locations = instruction->GetLocations(); 4507 Location base_loc = locations->InAt(0); 4508 Register base = base_loc.AsRegister<Register>(); 4509 Location out = locations->Out(); 4510 bool is_volatile = field_info.IsVolatile(); 4511 bool atomic_ldrd_strd = codegen_->GetInstructionSetFeatures().HasAtomicLdrdAndStrd(); 4512 Primitive::Type field_type = field_info.GetFieldType(); 4513 uint32_t offset = field_info.GetFieldOffset().Uint32Value(); 4514 4515 switch (field_type) { 4516 case Primitive::kPrimBoolean: 4517 __ LoadFromOffset(kLoadUnsignedByte, out.AsRegister<Register>(), base, offset); 4518 break; 4519 4520 case Primitive::kPrimByte: 4521 __ LoadFromOffset(kLoadSignedByte, out.AsRegister<Register>(), base, offset); 4522 break; 4523 4524 case Primitive::kPrimShort: 4525 __ LoadFromOffset(kLoadSignedHalfword, out.AsRegister<Register>(), base, offset); 4526 break; 4527 4528 case Primitive::kPrimChar: 4529 __ LoadFromOffset(kLoadUnsignedHalfword, out.AsRegister<Register>(), base, offset); 4530 break; 4531 4532 case Primitive::kPrimInt: 4533 __ LoadFromOffset(kLoadWord, out.AsRegister<Register>(), base, offset); 4534 break; 4535 4536 case Primitive::kPrimNot: { 4537 // /* HeapReference<Object> */ out = *(base + offset) 4538 if (kEmitCompilerReadBarrier && kUseBakerReadBarrier) { 4539 Location temp_loc = locations->GetTemp(0); 4540 // Note that a potential implicit null check is handled in this 4541 // CodeGeneratorARM::GenerateFieldLoadWithBakerReadBarrier call. 4542 codegen_->GenerateFieldLoadWithBakerReadBarrier( 4543 instruction, out, base, offset, temp_loc, /* needs_null_check */ true); 4544 if (is_volatile) { 4545 codegen_->GenerateMemoryBarrier(MemBarrierKind::kLoadAny); 4546 } 4547 } else { 4548 __ LoadFromOffset(kLoadWord, out.AsRegister<Register>(), base, offset); 4549 codegen_->MaybeRecordImplicitNullCheck(instruction); 4550 if (is_volatile) { 4551 codegen_->GenerateMemoryBarrier(MemBarrierKind::kLoadAny); 4552 } 4553 // If read barriers are enabled, emit read barriers other than 4554 // Baker's using a slow path (and also unpoison the loaded 4555 // reference, if heap poisoning is enabled). 4556 codegen_->MaybeGenerateReadBarrierSlow(instruction, out, out, base_loc, offset); 4557 } 4558 break; 4559 } 4560 4561 case Primitive::kPrimLong: 4562 if (is_volatile && !atomic_ldrd_strd) { 4563 GenerateWideAtomicLoad(base, offset, 4564 out.AsRegisterPairLow<Register>(), 4565 out.AsRegisterPairHigh<Register>()); 4566 } else { 4567 __ LoadFromOffset(kLoadWordPair, out.AsRegisterPairLow<Register>(), base, offset); 4568 } 4569 break; 4570 4571 case Primitive::kPrimFloat: 4572 __ LoadSFromOffset(out.AsFpuRegister<SRegister>(), base, offset); 4573 break; 4574 4575 case Primitive::kPrimDouble: { 4576 DRegister out_reg = FromLowSToD(out.AsFpuRegisterPairLow<SRegister>()); 4577 if (is_volatile && !atomic_ldrd_strd) { 4578 Register lo = locations->GetTemp(0).AsRegister<Register>(); 4579 Register hi = locations->GetTemp(1).AsRegister<Register>(); 4580 GenerateWideAtomicLoad(base, offset, lo, hi); 4581 codegen_->MaybeRecordImplicitNullCheck(instruction); 4582 __ vmovdrr(out_reg, lo, hi); 4583 } else { 4584 __ LoadDFromOffset(out_reg, base, offset); 4585 codegen_->MaybeRecordImplicitNullCheck(instruction); 4586 } 4587 break; 4588 } 4589 4590 case Primitive::kPrimVoid: 4591 LOG(FATAL) << "Unreachable type " << field_type; 4592 UNREACHABLE(); 4593 } 4594 4595 if (field_type == Primitive::kPrimNot || field_type == Primitive::kPrimDouble) { 4596 // Potential implicit null checks, in the case of reference or 4597 // double fields, are handled in the previous switch statement. 4598 } else { 4599 codegen_->MaybeRecordImplicitNullCheck(instruction); 4600 } 4601 4602 if (is_volatile) { 4603 if (field_type == Primitive::kPrimNot) { 4604 // Memory barriers, in the case of references, are also handled 4605 // in the previous switch statement. 4606 } else { 4607 codegen_->GenerateMemoryBarrier(MemBarrierKind::kLoadAny); 4608 } 4609 } 4610} 4611 4612void LocationsBuilderARM::VisitInstanceFieldSet(HInstanceFieldSet* instruction) { 4613 HandleFieldSet(instruction, instruction->GetFieldInfo()); 4614} 4615 4616void InstructionCodeGeneratorARM::VisitInstanceFieldSet(HInstanceFieldSet* instruction) { 4617 HandleFieldSet(instruction, instruction->GetFieldInfo(), instruction->GetValueCanBeNull()); 4618} 4619 4620void LocationsBuilderARM::VisitInstanceFieldGet(HInstanceFieldGet* instruction) { 4621 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4622} 4623 4624void InstructionCodeGeneratorARM::VisitInstanceFieldGet(HInstanceFieldGet* instruction) { 4625 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4626} 4627 4628void LocationsBuilderARM::VisitStaticFieldGet(HStaticFieldGet* instruction) { 4629 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4630} 4631 4632void InstructionCodeGeneratorARM::VisitStaticFieldGet(HStaticFieldGet* instruction) { 4633 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4634} 4635 4636void LocationsBuilderARM::VisitStaticFieldSet(HStaticFieldSet* instruction) { 4637 HandleFieldSet(instruction, instruction->GetFieldInfo()); 4638} 4639 4640void InstructionCodeGeneratorARM::VisitStaticFieldSet(HStaticFieldSet* instruction) { 4641 HandleFieldSet(instruction, instruction->GetFieldInfo(), instruction->GetValueCanBeNull()); 4642} 4643 4644void LocationsBuilderARM::VisitUnresolvedInstanceFieldGet( 4645 HUnresolvedInstanceFieldGet* instruction) { 4646 FieldAccessCallingConventionARM calling_convention; 4647 codegen_->CreateUnresolvedFieldLocationSummary( 4648 instruction, instruction->GetFieldType(), calling_convention); 4649} 4650 4651void InstructionCodeGeneratorARM::VisitUnresolvedInstanceFieldGet( 4652 HUnresolvedInstanceFieldGet* instruction) { 4653 FieldAccessCallingConventionARM calling_convention; 4654 codegen_->GenerateUnresolvedFieldAccess(instruction, 4655 instruction->GetFieldType(), 4656 instruction->GetFieldIndex(), 4657 instruction->GetDexPc(), 4658 calling_convention); 4659} 4660 4661void LocationsBuilderARM::VisitUnresolvedInstanceFieldSet( 4662 HUnresolvedInstanceFieldSet* instruction) { 4663 FieldAccessCallingConventionARM calling_convention; 4664 codegen_->CreateUnresolvedFieldLocationSummary( 4665 instruction, instruction->GetFieldType(), calling_convention); 4666} 4667 4668void InstructionCodeGeneratorARM::VisitUnresolvedInstanceFieldSet( 4669 HUnresolvedInstanceFieldSet* instruction) { 4670 FieldAccessCallingConventionARM calling_convention; 4671 codegen_->GenerateUnresolvedFieldAccess(instruction, 4672 instruction->GetFieldType(), 4673 instruction->GetFieldIndex(), 4674 instruction->GetDexPc(), 4675 calling_convention); 4676} 4677 4678void LocationsBuilderARM::VisitUnresolvedStaticFieldGet( 4679 HUnresolvedStaticFieldGet* instruction) { 4680 FieldAccessCallingConventionARM calling_convention; 4681 codegen_->CreateUnresolvedFieldLocationSummary( 4682 instruction, instruction->GetFieldType(), calling_convention); 4683} 4684 4685void InstructionCodeGeneratorARM::VisitUnresolvedStaticFieldGet( 4686 HUnresolvedStaticFieldGet* instruction) { 4687 FieldAccessCallingConventionARM calling_convention; 4688 codegen_->GenerateUnresolvedFieldAccess(instruction, 4689 instruction->GetFieldType(), 4690 instruction->GetFieldIndex(), 4691 instruction->GetDexPc(), 4692 calling_convention); 4693} 4694 4695void LocationsBuilderARM::VisitUnresolvedStaticFieldSet( 4696 HUnresolvedStaticFieldSet* instruction) { 4697 FieldAccessCallingConventionARM calling_convention; 4698 codegen_->CreateUnresolvedFieldLocationSummary( 4699 instruction, instruction->GetFieldType(), calling_convention); 4700} 4701 4702void InstructionCodeGeneratorARM::VisitUnresolvedStaticFieldSet( 4703 HUnresolvedStaticFieldSet* instruction) { 4704 FieldAccessCallingConventionARM calling_convention; 4705 codegen_->GenerateUnresolvedFieldAccess(instruction, 4706 instruction->GetFieldType(), 4707 instruction->GetFieldIndex(), 4708 instruction->GetDexPc(), 4709 calling_convention); 4710} 4711 4712void LocationsBuilderARM::VisitNullCheck(HNullCheck* instruction) { 4713 LocationSummary* locations = codegen_->CreateThrowingSlowPathLocations(instruction); 4714 locations->SetInAt(0, Location::RequiresRegister()); 4715} 4716 4717void CodeGeneratorARM::GenerateImplicitNullCheck(HNullCheck* instruction) { 4718 if (CanMoveNullCheckToUser(instruction)) { 4719 return; 4720 } 4721 Location obj = instruction->GetLocations()->InAt(0); 4722 4723 __ LoadFromOffset(kLoadWord, IP, obj.AsRegister<Register>(), 0); 4724 RecordPcInfo(instruction, instruction->GetDexPc()); 4725} 4726 4727void CodeGeneratorARM::GenerateExplicitNullCheck(HNullCheck* instruction) { 4728 SlowPathCodeARM* slow_path = new (GetGraph()->GetArena()) NullCheckSlowPathARM(instruction); 4729 AddSlowPath(slow_path); 4730 4731 LocationSummary* locations = instruction->GetLocations(); 4732 Location obj = locations->InAt(0); 4733 4734 __ CompareAndBranchIfZero(obj.AsRegister<Register>(), slow_path->GetEntryLabel()); 4735} 4736 4737void InstructionCodeGeneratorARM::VisitNullCheck(HNullCheck* instruction) { 4738 codegen_->GenerateNullCheck(instruction); 4739} 4740 4741static LoadOperandType GetLoadOperandType(Primitive::Type type) { 4742 switch (type) { 4743 case Primitive::kPrimNot: 4744 return kLoadWord; 4745 case Primitive::kPrimBoolean: 4746 return kLoadUnsignedByte; 4747 case Primitive::kPrimByte: 4748 return kLoadSignedByte; 4749 case Primitive::kPrimChar: 4750 return kLoadUnsignedHalfword; 4751 case Primitive::kPrimShort: 4752 return kLoadSignedHalfword; 4753 case Primitive::kPrimInt: 4754 return kLoadWord; 4755 case Primitive::kPrimLong: 4756 return kLoadWordPair; 4757 case Primitive::kPrimFloat: 4758 return kLoadSWord; 4759 case Primitive::kPrimDouble: 4760 return kLoadDWord; 4761 default: 4762 LOG(FATAL) << "Unreachable type " << type; 4763 UNREACHABLE(); 4764 } 4765} 4766 4767static StoreOperandType GetStoreOperandType(Primitive::Type type) { 4768 switch (type) { 4769 case Primitive::kPrimNot: 4770 return kStoreWord; 4771 case Primitive::kPrimBoolean: 4772 case Primitive::kPrimByte: 4773 return kStoreByte; 4774 case Primitive::kPrimChar: 4775 case Primitive::kPrimShort: 4776 return kStoreHalfword; 4777 case Primitive::kPrimInt: 4778 return kStoreWord; 4779 case Primitive::kPrimLong: 4780 return kStoreWordPair; 4781 case Primitive::kPrimFloat: 4782 return kStoreSWord; 4783 case Primitive::kPrimDouble: 4784 return kStoreDWord; 4785 default: 4786 LOG(FATAL) << "Unreachable type " << type; 4787 UNREACHABLE(); 4788 } 4789} 4790 4791void CodeGeneratorARM::LoadFromShiftedRegOffset(Primitive::Type type, 4792 Location out_loc, 4793 Register base, 4794 Register reg_offset, 4795 Condition cond) { 4796 uint32_t shift_count = Primitive::ComponentSizeShift(type); 4797 Address mem_address(base, reg_offset, Shift::LSL, shift_count); 4798 4799 switch (type) { 4800 case Primitive::kPrimByte: 4801 __ ldrsb(out_loc.AsRegister<Register>(), mem_address, cond); 4802 break; 4803 case Primitive::kPrimBoolean: 4804 __ ldrb(out_loc.AsRegister<Register>(), mem_address, cond); 4805 break; 4806 case Primitive::kPrimShort: 4807 __ ldrsh(out_loc.AsRegister<Register>(), mem_address, cond); 4808 break; 4809 case Primitive::kPrimChar: 4810 __ ldrh(out_loc.AsRegister<Register>(), mem_address, cond); 4811 break; 4812 case Primitive::kPrimNot: 4813 case Primitive::kPrimInt: 4814 __ ldr(out_loc.AsRegister<Register>(), mem_address, cond); 4815 break; 4816 // T32 doesn't support LoadFromShiftedRegOffset mem address mode for these types. 4817 case Primitive::kPrimLong: 4818 case Primitive::kPrimFloat: 4819 case Primitive::kPrimDouble: 4820 default: 4821 LOG(FATAL) << "Unreachable type " << type; 4822 UNREACHABLE(); 4823 } 4824} 4825 4826void CodeGeneratorARM::StoreToShiftedRegOffset(Primitive::Type type, 4827 Location loc, 4828 Register base, 4829 Register reg_offset, 4830 Condition cond) { 4831 uint32_t shift_count = Primitive::ComponentSizeShift(type); 4832 Address mem_address(base, reg_offset, Shift::LSL, shift_count); 4833 4834 switch (type) { 4835 case Primitive::kPrimByte: 4836 case Primitive::kPrimBoolean: 4837 __ strb(loc.AsRegister<Register>(), mem_address, cond); 4838 break; 4839 case Primitive::kPrimShort: 4840 case Primitive::kPrimChar: 4841 __ strh(loc.AsRegister<Register>(), mem_address, cond); 4842 break; 4843 case Primitive::kPrimNot: 4844 case Primitive::kPrimInt: 4845 __ str(loc.AsRegister<Register>(), mem_address, cond); 4846 break; 4847 // T32 doesn't support StoreToShiftedRegOffset mem address mode for these types. 4848 case Primitive::kPrimLong: 4849 case Primitive::kPrimFloat: 4850 case Primitive::kPrimDouble: 4851 default: 4852 LOG(FATAL) << "Unreachable type " << type; 4853 UNREACHABLE(); 4854 } 4855} 4856 4857void LocationsBuilderARM::VisitArrayGet(HArrayGet* instruction) { 4858 bool object_array_get_with_read_barrier = 4859 kEmitCompilerReadBarrier && (instruction->GetType() == Primitive::kPrimNot); 4860 LocationSummary* locations = 4861 new (GetGraph()->GetArena()) LocationSummary(instruction, 4862 object_array_get_with_read_barrier ? 4863 LocationSummary::kCallOnSlowPath : 4864 LocationSummary::kNoCall); 4865 if (object_array_get_with_read_barrier && kUseBakerReadBarrier) { 4866 locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty()); // No caller-save registers. 4867 } 4868 locations->SetInAt(0, Location::RequiresRegister()); 4869 locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1))); 4870 if (Primitive::IsFloatingPointType(instruction->GetType())) { 4871 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 4872 } else { 4873 // The output overlaps in the case of an object array get with 4874 // read barriers enabled: we do not want the move to overwrite the 4875 // array's location, as we need it to emit the read barrier. 4876 locations->SetOut( 4877 Location::RequiresRegister(), 4878 object_array_get_with_read_barrier ? Location::kOutputOverlap : Location::kNoOutputOverlap); 4879 } 4880 // We need a temporary register for the read barrier marking slow 4881 // path in CodeGeneratorARM::GenerateArrayLoadWithBakerReadBarrier. 4882 // Also need for String compression feature. 4883 if ((object_array_get_with_read_barrier && kUseBakerReadBarrier) 4884 || (mirror::kUseStringCompression && instruction->IsStringCharAt())) { 4885 locations->AddTemp(Location::RequiresRegister()); 4886 } 4887} 4888 4889void InstructionCodeGeneratorARM::VisitArrayGet(HArrayGet* instruction) { 4890 LocationSummary* locations = instruction->GetLocations(); 4891 Location obj_loc = locations->InAt(0); 4892 Register obj = obj_loc.AsRegister<Register>(); 4893 Location index = locations->InAt(1); 4894 Location out_loc = locations->Out(); 4895 uint32_t data_offset = CodeGenerator::GetArrayDataOffset(instruction); 4896 Primitive::Type type = instruction->GetType(); 4897 const bool maybe_compressed_char_at = mirror::kUseStringCompression && 4898 instruction->IsStringCharAt(); 4899 HInstruction* array_instr = instruction->GetArray(); 4900 bool has_intermediate_address = array_instr->IsIntermediateAddress(); 4901 4902 switch (type) { 4903 case Primitive::kPrimBoolean: 4904 case Primitive::kPrimByte: 4905 case Primitive::kPrimShort: 4906 case Primitive::kPrimChar: 4907 case Primitive::kPrimInt: { 4908 Register length; 4909 if (maybe_compressed_char_at) { 4910 length = locations->GetTemp(0).AsRegister<Register>(); 4911 uint32_t count_offset = mirror::String::CountOffset().Uint32Value(); 4912 __ LoadFromOffset(kLoadWord, length, obj, count_offset); 4913 codegen_->MaybeRecordImplicitNullCheck(instruction); 4914 } 4915 if (index.IsConstant()) { 4916 int32_t const_index = index.GetConstant()->AsIntConstant()->GetValue(); 4917 if (maybe_compressed_char_at) { 4918 Label uncompressed_load, done; 4919 __ Lsrs(length, length, 1u); // LSRS has a 16-bit encoding, TST (immediate) does not. 4920 static_assert(static_cast<uint32_t>(mirror::StringCompressionFlag::kCompressed) == 0u, 4921 "Expecting 0=compressed, 1=uncompressed"); 4922 __ b(&uncompressed_load, CS); 4923 __ LoadFromOffset(kLoadUnsignedByte, 4924 out_loc.AsRegister<Register>(), 4925 obj, 4926 data_offset + const_index); 4927 __ b(&done); 4928 __ Bind(&uncompressed_load); 4929 __ LoadFromOffset(GetLoadOperandType(Primitive::kPrimChar), 4930 out_loc.AsRegister<Register>(), 4931 obj, 4932 data_offset + (const_index << 1)); 4933 __ Bind(&done); 4934 } else { 4935 uint32_t full_offset = data_offset + (const_index << Primitive::ComponentSizeShift(type)); 4936 4937 LoadOperandType load_type = GetLoadOperandType(type); 4938 __ LoadFromOffset(load_type, out_loc.AsRegister<Register>(), obj, full_offset); 4939 } 4940 } else { 4941 Register temp = IP; 4942 4943 if (has_intermediate_address) { 4944 // We do not need to compute the intermediate address from the array: the 4945 // input instruction has done it already. See the comment in 4946 // `TryExtractArrayAccessAddress()`. 4947 if (kIsDebugBuild) { 4948 HIntermediateAddress* tmp = array_instr->AsIntermediateAddress(); 4949 DCHECK_EQ(tmp->GetOffset()->AsIntConstant()->GetValueAsUint64(), data_offset); 4950 } 4951 temp = obj; 4952 } else { 4953 __ add(temp, obj, ShifterOperand(data_offset)); 4954 } 4955 if (maybe_compressed_char_at) { 4956 Label uncompressed_load, done; 4957 __ Lsrs(length, length, 1u); // LSRS has a 16-bit encoding, TST (immediate) does not. 4958 static_assert(static_cast<uint32_t>(mirror::StringCompressionFlag::kCompressed) == 0u, 4959 "Expecting 0=compressed, 1=uncompressed"); 4960 __ b(&uncompressed_load, CS); 4961 __ ldrb(out_loc.AsRegister<Register>(), 4962 Address(temp, index.AsRegister<Register>(), Shift::LSL, 0)); 4963 __ b(&done); 4964 __ Bind(&uncompressed_load); 4965 __ ldrh(out_loc.AsRegister<Register>(), 4966 Address(temp, index.AsRegister<Register>(), Shift::LSL, 1)); 4967 __ Bind(&done); 4968 } else { 4969 codegen_->LoadFromShiftedRegOffset(type, out_loc, temp, index.AsRegister<Register>()); 4970 } 4971 } 4972 break; 4973 } 4974 4975 case Primitive::kPrimNot: { 4976 // The read barrier instrumentation of object ArrayGet 4977 // instructions does not support the HIntermediateAddress 4978 // instruction. 4979 DCHECK(!(has_intermediate_address && kEmitCompilerReadBarrier)); 4980 4981 static_assert( 4982 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 4983 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 4984 // /* HeapReference<Object> */ out = 4985 // *(obj + data_offset + index * sizeof(HeapReference<Object>)) 4986 if (kEmitCompilerReadBarrier && kUseBakerReadBarrier) { 4987 Location temp = locations->GetTemp(0); 4988 // Note that a potential implicit null check is handled in this 4989 // CodeGeneratorARM::GenerateArrayLoadWithBakerReadBarrier call. 4990 codegen_->GenerateArrayLoadWithBakerReadBarrier( 4991 instruction, out_loc, obj, data_offset, index, temp, /* needs_null_check */ true); 4992 } else { 4993 Register out = out_loc.AsRegister<Register>(); 4994 if (index.IsConstant()) { 4995 size_t offset = 4996 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 4997 __ LoadFromOffset(kLoadWord, out, obj, offset); 4998 codegen_->MaybeRecordImplicitNullCheck(instruction); 4999 // If read barriers are enabled, emit read barriers other than 5000 // Baker's using a slow path (and also unpoison the loaded 5001 // reference, if heap poisoning is enabled). 5002 codegen_->MaybeGenerateReadBarrierSlow(instruction, out_loc, out_loc, obj_loc, offset); 5003 } else { 5004 Register temp = IP; 5005 5006 if (has_intermediate_address) { 5007 // We do not need to compute the intermediate address from the array: the 5008 // input instruction has done it already. See the comment in 5009 // `TryExtractArrayAccessAddress()`. 5010 if (kIsDebugBuild) { 5011 HIntermediateAddress* tmp = array_instr->AsIntermediateAddress(); 5012 DCHECK_EQ(tmp->GetOffset()->AsIntConstant()->GetValueAsUint64(), data_offset); 5013 } 5014 temp = obj; 5015 } else { 5016 __ add(temp, obj, ShifterOperand(data_offset)); 5017 } 5018 codegen_->LoadFromShiftedRegOffset(type, out_loc, temp, index.AsRegister<Register>()); 5019 5020 codegen_->MaybeRecordImplicitNullCheck(instruction); 5021 // If read barriers are enabled, emit read barriers other than 5022 // Baker's using a slow path (and also unpoison the loaded 5023 // reference, if heap poisoning is enabled). 5024 codegen_->MaybeGenerateReadBarrierSlow( 5025 instruction, out_loc, out_loc, obj_loc, data_offset, index); 5026 } 5027 } 5028 break; 5029 } 5030 5031 case Primitive::kPrimLong: { 5032 if (index.IsConstant()) { 5033 size_t offset = 5034 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 5035 __ LoadFromOffset(kLoadWordPair, out_loc.AsRegisterPairLow<Register>(), obj, offset); 5036 } else { 5037 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 5038 __ LoadFromOffset(kLoadWordPair, out_loc.AsRegisterPairLow<Register>(), IP, data_offset); 5039 } 5040 break; 5041 } 5042 5043 case Primitive::kPrimFloat: { 5044 SRegister out = out_loc.AsFpuRegister<SRegister>(); 5045 if (index.IsConstant()) { 5046 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 5047 __ LoadSFromOffset(out, obj, offset); 5048 } else { 5049 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 5050 __ LoadSFromOffset(out, IP, data_offset); 5051 } 5052 break; 5053 } 5054 5055 case Primitive::kPrimDouble: { 5056 SRegister out = out_loc.AsFpuRegisterPairLow<SRegister>(); 5057 if (index.IsConstant()) { 5058 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 5059 __ LoadDFromOffset(FromLowSToD(out), obj, offset); 5060 } else { 5061 __ add(IP, obj, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 5062 __ LoadDFromOffset(FromLowSToD(out), IP, data_offset); 5063 } 5064 break; 5065 } 5066 5067 case Primitive::kPrimVoid: 5068 LOG(FATAL) << "Unreachable type " << type; 5069 UNREACHABLE(); 5070 } 5071 5072 if (type == Primitive::kPrimNot) { 5073 // Potential implicit null checks, in the case of reference 5074 // arrays, are handled in the previous switch statement. 5075 } else if (!maybe_compressed_char_at) { 5076 codegen_->MaybeRecordImplicitNullCheck(instruction); 5077 } 5078} 5079 5080void LocationsBuilderARM::VisitArraySet(HArraySet* instruction) { 5081 Primitive::Type value_type = instruction->GetComponentType(); 5082 5083 bool needs_write_barrier = 5084 CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue()); 5085 bool may_need_runtime_call_for_type_check = instruction->NeedsTypeCheck(); 5086 5087 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary( 5088 instruction, 5089 may_need_runtime_call_for_type_check ? 5090 LocationSummary::kCallOnSlowPath : 5091 LocationSummary::kNoCall); 5092 5093 locations->SetInAt(0, Location::RequiresRegister()); 5094 locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1))); 5095 if (Primitive::IsFloatingPointType(value_type)) { 5096 locations->SetInAt(2, Location::RequiresFpuRegister()); 5097 } else { 5098 locations->SetInAt(2, Location::RequiresRegister()); 5099 } 5100 if (needs_write_barrier) { 5101 // Temporary registers for the write barrier. 5102 locations->AddTemp(Location::RequiresRegister()); // Possibly used for ref. poisoning too. 5103 locations->AddTemp(Location::RequiresRegister()); 5104 } 5105} 5106 5107void InstructionCodeGeneratorARM::VisitArraySet(HArraySet* instruction) { 5108 LocationSummary* locations = instruction->GetLocations(); 5109 Location array_loc = locations->InAt(0); 5110 Register array = array_loc.AsRegister<Register>(); 5111 Location index = locations->InAt(1); 5112 Primitive::Type value_type = instruction->GetComponentType(); 5113 bool may_need_runtime_call_for_type_check = instruction->NeedsTypeCheck(); 5114 bool needs_write_barrier = 5115 CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue()); 5116 uint32_t data_offset = 5117 mirror::Array::DataOffset(Primitive::ComponentSize(value_type)).Uint32Value(); 5118 Location value_loc = locations->InAt(2); 5119 HInstruction* array_instr = instruction->GetArray(); 5120 bool has_intermediate_address = array_instr->IsIntermediateAddress(); 5121 5122 switch (value_type) { 5123 case Primitive::kPrimBoolean: 5124 case Primitive::kPrimByte: 5125 case Primitive::kPrimShort: 5126 case Primitive::kPrimChar: 5127 case Primitive::kPrimInt: { 5128 if (index.IsConstant()) { 5129 int32_t const_index = index.GetConstant()->AsIntConstant()->GetValue(); 5130 uint32_t full_offset = 5131 data_offset + (const_index << Primitive::ComponentSizeShift(value_type)); 5132 StoreOperandType store_type = GetStoreOperandType(value_type); 5133 __ StoreToOffset(store_type, value_loc.AsRegister<Register>(), array, full_offset); 5134 } else { 5135 Register temp = IP; 5136 5137 if (has_intermediate_address) { 5138 // We do not need to compute the intermediate address from the array: the 5139 // input instruction has done it already. See the comment in 5140 // `TryExtractArrayAccessAddress()`. 5141 if (kIsDebugBuild) { 5142 HIntermediateAddress* tmp = array_instr->AsIntermediateAddress(); 5143 DCHECK(tmp->GetOffset()->AsIntConstant()->GetValueAsUint64() == data_offset); 5144 } 5145 temp = array; 5146 } else { 5147 __ add(temp, array, ShifterOperand(data_offset)); 5148 } 5149 codegen_->StoreToShiftedRegOffset(value_type, 5150 value_loc, 5151 temp, 5152 index.AsRegister<Register>()); 5153 } 5154 break; 5155 } 5156 5157 case Primitive::kPrimNot: { 5158 Register value = value_loc.AsRegister<Register>(); 5159 // TryExtractArrayAccessAddress optimization is never applied for non-primitive ArraySet. 5160 // See the comment in instruction_simplifier_shared.cc. 5161 DCHECK(!has_intermediate_address); 5162 5163 if (instruction->InputAt(2)->IsNullConstant()) { 5164 // Just setting null. 5165 if (index.IsConstant()) { 5166 size_t offset = 5167 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 5168 __ StoreToOffset(kStoreWord, value, array, offset); 5169 } else { 5170 DCHECK(index.IsRegister()) << index; 5171 __ add(IP, array, ShifterOperand(data_offset)); 5172 codegen_->StoreToShiftedRegOffset(value_type, 5173 value_loc, 5174 IP, 5175 index.AsRegister<Register>()); 5176 } 5177 codegen_->MaybeRecordImplicitNullCheck(instruction); 5178 DCHECK(!needs_write_barrier); 5179 DCHECK(!may_need_runtime_call_for_type_check); 5180 break; 5181 } 5182 5183 DCHECK(needs_write_barrier); 5184 Location temp1_loc = locations->GetTemp(0); 5185 Register temp1 = temp1_loc.AsRegister<Register>(); 5186 Location temp2_loc = locations->GetTemp(1); 5187 Register temp2 = temp2_loc.AsRegister<Register>(); 5188 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 5189 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 5190 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 5191 Label done; 5192 SlowPathCodeARM* slow_path = nullptr; 5193 5194 if (may_need_runtime_call_for_type_check) { 5195 slow_path = new (GetGraph()->GetArena()) ArraySetSlowPathARM(instruction); 5196 codegen_->AddSlowPath(slow_path); 5197 if (instruction->GetValueCanBeNull()) { 5198 Label non_zero; 5199 __ CompareAndBranchIfNonZero(value, &non_zero); 5200 if (index.IsConstant()) { 5201 size_t offset = 5202 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 5203 __ StoreToOffset(kStoreWord, value, array, offset); 5204 } else { 5205 DCHECK(index.IsRegister()) << index; 5206 __ add(IP, array, ShifterOperand(data_offset)); 5207 codegen_->StoreToShiftedRegOffset(value_type, 5208 value_loc, 5209 IP, 5210 index.AsRegister<Register>()); 5211 } 5212 codegen_->MaybeRecordImplicitNullCheck(instruction); 5213 __ b(&done); 5214 __ Bind(&non_zero); 5215 } 5216 5217 // Note that when read barriers are enabled, the type checks 5218 // are performed without read barriers. This is fine, even in 5219 // the case where a class object is in the from-space after 5220 // the flip, as a comparison involving such a type would not 5221 // produce a false positive; it may of course produce a false 5222 // negative, in which case we would take the ArraySet slow 5223 // path. 5224 5225 // /* HeapReference<Class> */ temp1 = array->klass_ 5226 __ LoadFromOffset(kLoadWord, temp1, array, class_offset); 5227 codegen_->MaybeRecordImplicitNullCheck(instruction); 5228 __ MaybeUnpoisonHeapReference(temp1); 5229 5230 // /* HeapReference<Class> */ temp1 = temp1->component_type_ 5231 __ LoadFromOffset(kLoadWord, temp1, temp1, component_offset); 5232 // /* HeapReference<Class> */ temp2 = value->klass_ 5233 __ LoadFromOffset(kLoadWord, temp2, value, class_offset); 5234 // If heap poisoning is enabled, no need to unpoison `temp1` 5235 // nor `temp2`, as we are comparing two poisoned references. 5236 __ cmp(temp1, ShifterOperand(temp2)); 5237 5238 if (instruction->StaticTypeOfArrayIsObjectArray()) { 5239 Label do_put; 5240 __ b(&do_put, EQ); 5241 // If heap poisoning is enabled, the `temp1` reference has 5242 // not been unpoisoned yet; unpoison it now. 5243 __ MaybeUnpoisonHeapReference(temp1); 5244 5245 // /* HeapReference<Class> */ temp1 = temp1->super_class_ 5246 __ LoadFromOffset(kLoadWord, temp1, temp1, super_offset); 5247 // If heap poisoning is enabled, no need to unpoison 5248 // `temp1`, as we are comparing against null below. 5249 __ CompareAndBranchIfNonZero(temp1, slow_path->GetEntryLabel()); 5250 __ Bind(&do_put); 5251 } else { 5252 __ b(slow_path->GetEntryLabel(), NE); 5253 } 5254 } 5255 5256 Register source = value; 5257 if (kPoisonHeapReferences) { 5258 // Note that in the case where `value` is a null reference, 5259 // we do not enter this block, as a null reference does not 5260 // need poisoning. 5261 DCHECK_EQ(value_type, Primitive::kPrimNot); 5262 __ Mov(temp1, value); 5263 __ PoisonHeapReference(temp1); 5264 source = temp1; 5265 } 5266 5267 if (index.IsConstant()) { 5268 size_t offset = 5269 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 5270 __ StoreToOffset(kStoreWord, source, array, offset); 5271 } else { 5272 DCHECK(index.IsRegister()) << index; 5273 5274 __ add(IP, array, ShifterOperand(data_offset)); 5275 codegen_->StoreToShiftedRegOffset(value_type, 5276 Location::RegisterLocation(source), 5277 IP, 5278 index.AsRegister<Register>()); 5279 } 5280 5281 if (!may_need_runtime_call_for_type_check) { 5282 codegen_->MaybeRecordImplicitNullCheck(instruction); 5283 } 5284 5285 codegen_->MarkGCCard(temp1, temp2, array, value, instruction->GetValueCanBeNull()); 5286 5287 if (done.IsLinked()) { 5288 __ Bind(&done); 5289 } 5290 5291 if (slow_path != nullptr) { 5292 __ Bind(slow_path->GetExitLabel()); 5293 } 5294 5295 break; 5296 } 5297 5298 case Primitive::kPrimLong: { 5299 Location value = locations->InAt(2); 5300 if (index.IsConstant()) { 5301 size_t offset = 5302 (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 5303 __ StoreToOffset(kStoreWordPair, value.AsRegisterPairLow<Register>(), array, offset); 5304 } else { 5305 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 5306 __ StoreToOffset(kStoreWordPair, value.AsRegisterPairLow<Register>(), IP, data_offset); 5307 } 5308 break; 5309 } 5310 5311 case Primitive::kPrimFloat: { 5312 Location value = locations->InAt(2); 5313 DCHECK(value.IsFpuRegister()); 5314 if (index.IsConstant()) { 5315 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_4) + data_offset; 5316 __ StoreSToOffset(value.AsFpuRegister<SRegister>(), array, offset); 5317 } else { 5318 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_4)); 5319 __ StoreSToOffset(value.AsFpuRegister<SRegister>(), IP, data_offset); 5320 } 5321 break; 5322 } 5323 5324 case Primitive::kPrimDouble: { 5325 Location value = locations->InAt(2); 5326 DCHECK(value.IsFpuRegisterPair()); 5327 if (index.IsConstant()) { 5328 size_t offset = (index.GetConstant()->AsIntConstant()->GetValue() << TIMES_8) + data_offset; 5329 __ StoreDToOffset(FromLowSToD(value.AsFpuRegisterPairLow<SRegister>()), array, offset); 5330 } else { 5331 __ add(IP, array, ShifterOperand(index.AsRegister<Register>(), LSL, TIMES_8)); 5332 __ StoreDToOffset(FromLowSToD(value.AsFpuRegisterPairLow<SRegister>()), IP, data_offset); 5333 } 5334 5335 break; 5336 } 5337 5338 case Primitive::kPrimVoid: 5339 LOG(FATAL) << "Unreachable type " << value_type; 5340 UNREACHABLE(); 5341 } 5342 5343 // Objects are handled in the switch. 5344 if (value_type != Primitive::kPrimNot) { 5345 codegen_->MaybeRecordImplicitNullCheck(instruction); 5346 } 5347} 5348 5349void LocationsBuilderARM::VisitArrayLength(HArrayLength* instruction) { 5350 LocationSummary* locations = 5351 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 5352 locations->SetInAt(0, Location::RequiresRegister()); 5353 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 5354} 5355 5356void InstructionCodeGeneratorARM::VisitArrayLength(HArrayLength* instruction) { 5357 LocationSummary* locations = instruction->GetLocations(); 5358 uint32_t offset = CodeGenerator::GetArrayLengthOffset(instruction); 5359 Register obj = locations->InAt(0).AsRegister<Register>(); 5360 Register out = locations->Out().AsRegister<Register>(); 5361 __ LoadFromOffset(kLoadWord, out, obj, offset); 5362 codegen_->MaybeRecordImplicitNullCheck(instruction); 5363 // Mask out compression flag from String's array length. 5364 if (mirror::kUseStringCompression && instruction->IsStringLength()) { 5365 __ Lsr(out, out, 1u); 5366 } 5367} 5368 5369void LocationsBuilderARM::VisitIntermediateAddress(HIntermediateAddress* instruction) { 5370 LocationSummary* locations = 5371 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 5372 5373 locations->SetInAt(0, Location::RequiresRegister()); 5374 locations->SetInAt(1, Location::RegisterOrConstant(instruction->GetOffset())); 5375 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 5376} 5377 5378void InstructionCodeGeneratorARM::VisitIntermediateAddress(HIntermediateAddress* instruction) { 5379 LocationSummary* locations = instruction->GetLocations(); 5380 Location out = locations->Out(); 5381 Location first = locations->InAt(0); 5382 Location second = locations->InAt(1); 5383 5384 if (second.IsRegister()) { 5385 __ add(out.AsRegister<Register>(), 5386 first.AsRegister<Register>(), 5387 ShifterOperand(second.AsRegister<Register>())); 5388 } else { 5389 __ AddConstant(out.AsRegister<Register>(), 5390 first.AsRegister<Register>(), 5391 second.GetConstant()->AsIntConstant()->GetValue()); 5392 } 5393} 5394 5395void LocationsBuilderARM::VisitBoundsCheck(HBoundsCheck* instruction) { 5396 RegisterSet caller_saves = RegisterSet::Empty(); 5397 InvokeRuntimeCallingConvention calling_convention; 5398 caller_saves.Add(Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 5399 caller_saves.Add(Location::RegisterLocation(calling_convention.GetRegisterAt(1))); 5400 LocationSummary* locations = codegen_->CreateThrowingSlowPathLocations(instruction, caller_saves); 5401 locations->SetInAt(0, Location::RequiresRegister()); 5402 locations->SetInAt(1, Location::RequiresRegister()); 5403} 5404 5405void InstructionCodeGeneratorARM::VisitBoundsCheck(HBoundsCheck* instruction) { 5406 LocationSummary* locations = instruction->GetLocations(); 5407 SlowPathCodeARM* slow_path = 5408 new (GetGraph()->GetArena()) BoundsCheckSlowPathARM(instruction); 5409 codegen_->AddSlowPath(slow_path); 5410 5411 Register index = locations->InAt(0).AsRegister<Register>(); 5412 Register length = locations->InAt(1).AsRegister<Register>(); 5413 5414 __ cmp(index, ShifterOperand(length)); 5415 __ b(slow_path->GetEntryLabel(), HS); 5416} 5417 5418void CodeGeneratorARM::MarkGCCard(Register temp, 5419 Register card, 5420 Register object, 5421 Register value, 5422 bool can_be_null) { 5423 Label is_null; 5424 if (can_be_null) { 5425 __ CompareAndBranchIfZero(value, &is_null); 5426 } 5427 __ LoadFromOffset(kLoadWord, card, TR, Thread::CardTableOffset<kArmPointerSize>().Int32Value()); 5428 __ Lsr(temp, object, gc::accounting::CardTable::kCardShift); 5429 __ strb(card, Address(card, temp)); 5430 if (can_be_null) { 5431 __ Bind(&is_null); 5432 } 5433} 5434 5435void LocationsBuilderARM::VisitParallelMove(HParallelMove* instruction ATTRIBUTE_UNUSED) { 5436 LOG(FATAL) << "Unreachable"; 5437} 5438 5439void InstructionCodeGeneratorARM::VisitParallelMove(HParallelMove* instruction) { 5440 codegen_->GetMoveResolver()->EmitNativeCode(instruction); 5441} 5442 5443void LocationsBuilderARM::VisitSuspendCheck(HSuspendCheck* instruction) { 5444 LocationSummary* locations = 5445 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnSlowPath); 5446 locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty()); // No caller-save registers. 5447} 5448 5449void InstructionCodeGeneratorARM::VisitSuspendCheck(HSuspendCheck* instruction) { 5450 HBasicBlock* block = instruction->GetBlock(); 5451 if (block->GetLoopInformation() != nullptr) { 5452 DCHECK(block->GetLoopInformation()->GetSuspendCheck() == instruction); 5453 // The back edge will generate the suspend check. 5454 return; 5455 } 5456 if (block->IsEntryBlock() && instruction->GetNext()->IsGoto()) { 5457 // The goto will generate the suspend check. 5458 return; 5459 } 5460 GenerateSuspendCheck(instruction, nullptr); 5461} 5462 5463void InstructionCodeGeneratorARM::GenerateSuspendCheck(HSuspendCheck* instruction, 5464 HBasicBlock* successor) { 5465 SuspendCheckSlowPathARM* slow_path = 5466 down_cast<SuspendCheckSlowPathARM*>(instruction->GetSlowPath()); 5467 if (slow_path == nullptr) { 5468 slow_path = new (GetGraph()->GetArena()) SuspendCheckSlowPathARM(instruction, successor); 5469 instruction->SetSlowPath(slow_path); 5470 codegen_->AddSlowPath(slow_path); 5471 if (successor != nullptr) { 5472 DCHECK(successor->IsLoopHeader()); 5473 codegen_->ClearSpillSlotsFromLoopPhisInStackMap(instruction); 5474 } 5475 } else { 5476 DCHECK_EQ(slow_path->GetSuccessor(), successor); 5477 } 5478 5479 __ LoadFromOffset( 5480 kLoadUnsignedHalfword, IP, TR, Thread::ThreadFlagsOffset<kArmPointerSize>().Int32Value()); 5481 if (successor == nullptr) { 5482 __ CompareAndBranchIfNonZero(IP, slow_path->GetEntryLabel()); 5483 __ Bind(slow_path->GetReturnLabel()); 5484 } else { 5485 __ CompareAndBranchIfZero(IP, codegen_->GetLabelOf(successor)); 5486 __ b(slow_path->GetEntryLabel()); 5487 } 5488} 5489 5490ArmAssembler* ParallelMoveResolverARM::GetAssembler() const { 5491 return codegen_->GetAssembler(); 5492} 5493 5494void ParallelMoveResolverARM::EmitMove(size_t index) { 5495 MoveOperands* move = moves_[index]; 5496 Location source = move->GetSource(); 5497 Location destination = move->GetDestination(); 5498 5499 if (source.IsRegister()) { 5500 if (destination.IsRegister()) { 5501 __ Mov(destination.AsRegister<Register>(), source.AsRegister<Register>()); 5502 } else if (destination.IsFpuRegister()) { 5503 __ vmovsr(destination.AsFpuRegister<SRegister>(), source.AsRegister<Register>()); 5504 } else { 5505 DCHECK(destination.IsStackSlot()); 5506 __ StoreToOffset(kStoreWord, source.AsRegister<Register>(), 5507 SP, destination.GetStackIndex()); 5508 } 5509 } else if (source.IsStackSlot()) { 5510 if (destination.IsRegister()) { 5511 __ LoadFromOffset(kLoadWord, destination.AsRegister<Register>(), 5512 SP, source.GetStackIndex()); 5513 } else if (destination.IsFpuRegister()) { 5514 __ LoadSFromOffset(destination.AsFpuRegister<SRegister>(), SP, source.GetStackIndex()); 5515 } else { 5516 DCHECK(destination.IsStackSlot()); 5517 __ LoadFromOffset(kLoadWord, IP, SP, source.GetStackIndex()); 5518 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 5519 } 5520 } else if (source.IsFpuRegister()) { 5521 if (destination.IsRegister()) { 5522 __ vmovrs(destination.AsRegister<Register>(), source.AsFpuRegister<SRegister>()); 5523 } else if (destination.IsFpuRegister()) { 5524 __ vmovs(destination.AsFpuRegister<SRegister>(), source.AsFpuRegister<SRegister>()); 5525 } else { 5526 DCHECK(destination.IsStackSlot()); 5527 __ StoreSToOffset(source.AsFpuRegister<SRegister>(), SP, destination.GetStackIndex()); 5528 } 5529 } else if (source.IsDoubleStackSlot()) { 5530 if (destination.IsDoubleStackSlot()) { 5531 __ LoadDFromOffset(DTMP, SP, source.GetStackIndex()); 5532 __ StoreDToOffset(DTMP, SP, destination.GetStackIndex()); 5533 } else if (destination.IsRegisterPair()) { 5534 DCHECK(ExpectedPairLayout(destination)); 5535 __ LoadFromOffset( 5536 kLoadWordPair, destination.AsRegisterPairLow<Register>(), SP, source.GetStackIndex()); 5537 } else { 5538 DCHECK(destination.IsFpuRegisterPair()) << destination; 5539 __ LoadDFromOffset(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 5540 SP, 5541 source.GetStackIndex()); 5542 } 5543 } else if (source.IsRegisterPair()) { 5544 if (destination.IsRegisterPair()) { 5545 __ Mov(destination.AsRegisterPairLow<Register>(), source.AsRegisterPairLow<Register>()); 5546 __ Mov(destination.AsRegisterPairHigh<Register>(), source.AsRegisterPairHigh<Register>()); 5547 } else if (destination.IsFpuRegisterPair()) { 5548 __ vmovdrr(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 5549 source.AsRegisterPairLow<Register>(), 5550 source.AsRegisterPairHigh<Register>()); 5551 } else { 5552 DCHECK(destination.IsDoubleStackSlot()) << destination; 5553 DCHECK(ExpectedPairLayout(source)); 5554 __ StoreToOffset( 5555 kStoreWordPair, source.AsRegisterPairLow<Register>(), SP, destination.GetStackIndex()); 5556 } 5557 } else if (source.IsFpuRegisterPair()) { 5558 if (destination.IsRegisterPair()) { 5559 __ vmovrrd(destination.AsRegisterPairLow<Register>(), 5560 destination.AsRegisterPairHigh<Register>(), 5561 FromLowSToD(source.AsFpuRegisterPairLow<SRegister>())); 5562 } else if (destination.IsFpuRegisterPair()) { 5563 __ vmovd(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), 5564 FromLowSToD(source.AsFpuRegisterPairLow<SRegister>())); 5565 } else { 5566 DCHECK(destination.IsDoubleStackSlot()) << destination; 5567 __ StoreDToOffset(FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()), 5568 SP, 5569 destination.GetStackIndex()); 5570 } 5571 } else { 5572 DCHECK(source.IsConstant()) << source; 5573 HConstant* constant = source.GetConstant(); 5574 if (constant->IsIntConstant() || constant->IsNullConstant()) { 5575 int32_t value = CodeGenerator::GetInt32ValueOf(constant); 5576 if (destination.IsRegister()) { 5577 __ LoadImmediate(destination.AsRegister<Register>(), value); 5578 } else { 5579 DCHECK(destination.IsStackSlot()); 5580 __ LoadImmediate(IP, value); 5581 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 5582 } 5583 } else if (constant->IsLongConstant()) { 5584 int64_t value = constant->AsLongConstant()->GetValue(); 5585 if (destination.IsRegisterPair()) { 5586 __ LoadImmediate(destination.AsRegisterPairLow<Register>(), Low32Bits(value)); 5587 __ LoadImmediate(destination.AsRegisterPairHigh<Register>(), High32Bits(value)); 5588 } else { 5589 DCHECK(destination.IsDoubleStackSlot()) << destination; 5590 __ LoadImmediate(IP, Low32Bits(value)); 5591 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 5592 __ LoadImmediate(IP, High32Bits(value)); 5593 __ StoreToOffset(kStoreWord, IP, SP, destination.GetHighStackIndex(kArmWordSize)); 5594 } 5595 } else if (constant->IsDoubleConstant()) { 5596 double value = constant->AsDoubleConstant()->GetValue(); 5597 if (destination.IsFpuRegisterPair()) { 5598 __ LoadDImmediate(FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()), value); 5599 } else { 5600 DCHECK(destination.IsDoubleStackSlot()) << destination; 5601 uint64_t int_value = bit_cast<uint64_t, double>(value); 5602 __ LoadImmediate(IP, Low32Bits(int_value)); 5603 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 5604 __ LoadImmediate(IP, High32Bits(int_value)); 5605 __ StoreToOffset(kStoreWord, IP, SP, destination.GetHighStackIndex(kArmWordSize)); 5606 } 5607 } else { 5608 DCHECK(constant->IsFloatConstant()) << constant->DebugName(); 5609 float value = constant->AsFloatConstant()->GetValue(); 5610 if (destination.IsFpuRegister()) { 5611 __ LoadSImmediate(destination.AsFpuRegister<SRegister>(), value); 5612 } else { 5613 DCHECK(destination.IsStackSlot()); 5614 __ LoadImmediate(IP, bit_cast<int32_t, float>(value)); 5615 __ StoreToOffset(kStoreWord, IP, SP, destination.GetStackIndex()); 5616 } 5617 } 5618 } 5619} 5620 5621void ParallelMoveResolverARM::Exchange(Register reg, int mem) { 5622 __ Mov(IP, reg); 5623 __ LoadFromOffset(kLoadWord, reg, SP, mem); 5624 __ StoreToOffset(kStoreWord, IP, SP, mem); 5625} 5626 5627void ParallelMoveResolverARM::Exchange(int mem1, int mem2) { 5628 ScratchRegisterScope ensure_scratch(this, IP, R0, codegen_->GetNumberOfCoreRegisters()); 5629 int stack_offset = ensure_scratch.IsSpilled() ? kArmWordSize : 0; 5630 __ LoadFromOffset(kLoadWord, static_cast<Register>(ensure_scratch.GetRegister()), 5631 SP, mem1 + stack_offset); 5632 __ LoadFromOffset(kLoadWord, IP, SP, mem2 + stack_offset); 5633 __ StoreToOffset(kStoreWord, static_cast<Register>(ensure_scratch.GetRegister()), 5634 SP, mem2 + stack_offset); 5635 __ StoreToOffset(kStoreWord, IP, SP, mem1 + stack_offset); 5636} 5637 5638void ParallelMoveResolverARM::EmitSwap(size_t index) { 5639 MoveOperands* move = moves_[index]; 5640 Location source = move->GetSource(); 5641 Location destination = move->GetDestination(); 5642 5643 if (source.IsRegister() && destination.IsRegister()) { 5644 DCHECK_NE(source.AsRegister<Register>(), IP); 5645 DCHECK_NE(destination.AsRegister<Register>(), IP); 5646 __ Mov(IP, source.AsRegister<Register>()); 5647 __ Mov(source.AsRegister<Register>(), destination.AsRegister<Register>()); 5648 __ Mov(destination.AsRegister<Register>(), IP); 5649 } else if (source.IsRegister() && destination.IsStackSlot()) { 5650 Exchange(source.AsRegister<Register>(), destination.GetStackIndex()); 5651 } else if (source.IsStackSlot() && destination.IsRegister()) { 5652 Exchange(destination.AsRegister<Register>(), source.GetStackIndex()); 5653 } else if (source.IsStackSlot() && destination.IsStackSlot()) { 5654 Exchange(source.GetStackIndex(), destination.GetStackIndex()); 5655 } else if (source.IsFpuRegister() && destination.IsFpuRegister()) { 5656 __ vmovrs(IP, source.AsFpuRegister<SRegister>()); 5657 __ vmovs(source.AsFpuRegister<SRegister>(), destination.AsFpuRegister<SRegister>()); 5658 __ vmovsr(destination.AsFpuRegister<SRegister>(), IP); 5659 } else if (source.IsRegisterPair() && destination.IsRegisterPair()) { 5660 __ vmovdrr(DTMP, source.AsRegisterPairLow<Register>(), source.AsRegisterPairHigh<Register>()); 5661 __ Mov(source.AsRegisterPairLow<Register>(), destination.AsRegisterPairLow<Register>()); 5662 __ Mov(source.AsRegisterPairHigh<Register>(), destination.AsRegisterPairHigh<Register>()); 5663 __ vmovrrd(destination.AsRegisterPairLow<Register>(), 5664 destination.AsRegisterPairHigh<Register>(), 5665 DTMP); 5666 } else if (source.IsRegisterPair() || destination.IsRegisterPair()) { 5667 Register low_reg = source.IsRegisterPair() 5668 ? source.AsRegisterPairLow<Register>() 5669 : destination.AsRegisterPairLow<Register>(); 5670 int mem = source.IsRegisterPair() 5671 ? destination.GetStackIndex() 5672 : source.GetStackIndex(); 5673 DCHECK(ExpectedPairLayout(source.IsRegisterPair() ? source : destination)); 5674 __ vmovdrr(DTMP, low_reg, static_cast<Register>(low_reg + 1)); 5675 __ LoadFromOffset(kLoadWordPair, low_reg, SP, mem); 5676 __ StoreDToOffset(DTMP, SP, mem); 5677 } else if (source.IsFpuRegisterPair() && destination.IsFpuRegisterPair()) { 5678 DRegister first = FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()); 5679 DRegister second = FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()); 5680 __ vmovd(DTMP, first); 5681 __ vmovd(first, second); 5682 __ vmovd(second, DTMP); 5683 } else if (source.IsFpuRegisterPair() || destination.IsFpuRegisterPair()) { 5684 DRegister reg = source.IsFpuRegisterPair() 5685 ? FromLowSToD(source.AsFpuRegisterPairLow<SRegister>()) 5686 : FromLowSToD(destination.AsFpuRegisterPairLow<SRegister>()); 5687 int mem = source.IsFpuRegisterPair() 5688 ? destination.GetStackIndex() 5689 : source.GetStackIndex(); 5690 __ vmovd(DTMP, reg); 5691 __ LoadDFromOffset(reg, SP, mem); 5692 __ StoreDToOffset(DTMP, SP, mem); 5693 } else if (source.IsFpuRegister() || destination.IsFpuRegister()) { 5694 SRegister reg = source.IsFpuRegister() ? source.AsFpuRegister<SRegister>() 5695 : destination.AsFpuRegister<SRegister>(); 5696 int mem = source.IsFpuRegister() 5697 ? destination.GetStackIndex() 5698 : source.GetStackIndex(); 5699 5700 __ vmovrs(IP, reg); 5701 __ LoadSFromOffset(reg, SP, mem); 5702 __ StoreToOffset(kStoreWord, IP, SP, mem); 5703 } else if (source.IsDoubleStackSlot() && destination.IsDoubleStackSlot()) { 5704 Exchange(source.GetStackIndex(), destination.GetStackIndex()); 5705 Exchange(source.GetHighStackIndex(kArmWordSize), destination.GetHighStackIndex(kArmWordSize)); 5706 } else { 5707 LOG(FATAL) << "Unimplemented" << source << " <-> " << destination; 5708 } 5709} 5710 5711void ParallelMoveResolverARM::SpillScratch(int reg) { 5712 __ Push(static_cast<Register>(reg)); 5713} 5714 5715void ParallelMoveResolverARM::RestoreScratch(int reg) { 5716 __ Pop(static_cast<Register>(reg)); 5717} 5718 5719HLoadClass::LoadKind CodeGeneratorARM::GetSupportedLoadClassKind( 5720 HLoadClass::LoadKind desired_class_load_kind) { 5721 switch (desired_class_load_kind) { 5722 case HLoadClass::LoadKind::kReferrersClass: 5723 break; 5724 case HLoadClass::LoadKind::kBootImageLinkTimeAddress: 5725 DCHECK(!GetCompilerOptions().GetCompilePic()); 5726 break; 5727 case HLoadClass::LoadKind::kBootImageLinkTimePcRelative: 5728 DCHECK(GetCompilerOptions().GetCompilePic()); 5729 break; 5730 case HLoadClass::LoadKind::kBootImageAddress: 5731 break; 5732 case HLoadClass::LoadKind::kBssEntry: 5733 DCHECK(!Runtime::Current()->UseJitCompilation()); 5734 break; 5735 case HLoadClass::LoadKind::kJitTableAddress: 5736 DCHECK(Runtime::Current()->UseJitCompilation()); 5737 break; 5738 case HLoadClass::LoadKind::kDexCacheViaMethod: 5739 break; 5740 } 5741 return desired_class_load_kind; 5742} 5743 5744void LocationsBuilderARM::VisitLoadClass(HLoadClass* cls) { 5745 HLoadClass::LoadKind load_kind = cls->GetLoadKind(); 5746 if (load_kind == HLoadClass::LoadKind::kDexCacheViaMethod) { 5747 InvokeRuntimeCallingConvention calling_convention; 5748 CodeGenerator::CreateLoadClassRuntimeCallLocationSummary( 5749 cls, 5750 Location::RegisterLocation(calling_convention.GetRegisterAt(0)), 5751 Location::RegisterLocation(R0)); 5752 return; 5753 } 5754 DCHECK(!cls->NeedsAccessCheck()); 5755 5756 const bool requires_read_barrier = kEmitCompilerReadBarrier && !cls->IsInBootImage(); 5757 LocationSummary::CallKind call_kind = (cls->NeedsEnvironment() || requires_read_barrier) 5758 ? LocationSummary::kCallOnSlowPath 5759 : LocationSummary::kNoCall; 5760 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(cls, call_kind); 5761 if (kUseBakerReadBarrier && requires_read_barrier && !cls->NeedsEnvironment()) { 5762 locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty()); // No caller-save registers. 5763 } 5764 5765 if (load_kind == HLoadClass::LoadKind::kReferrersClass) { 5766 locations->SetInAt(0, Location::RequiresRegister()); 5767 } 5768 locations->SetOut(Location::RequiresRegister()); 5769} 5770 5771// NO_THREAD_SAFETY_ANALYSIS as we manipulate handles whose internal object we know does not 5772// move. 5773void InstructionCodeGeneratorARM::VisitLoadClass(HLoadClass* cls) NO_THREAD_SAFETY_ANALYSIS { 5774 HLoadClass::LoadKind load_kind = cls->GetLoadKind(); 5775 if (load_kind == HLoadClass::LoadKind::kDexCacheViaMethod) { 5776 codegen_->GenerateLoadClassRuntimeCall(cls); 5777 return; 5778 } 5779 DCHECK(!cls->NeedsAccessCheck()); 5780 5781 LocationSummary* locations = cls->GetLocations(); 5782 Location out_loc = locations->Out(); 5783 Register out = out_loc.AsRegister<Register>(); 5784 5785 const ReadBarrierOption read_barrier_option = cls->IsInBootImage() 5786 ? kWithoutReadBarrier 5787 : kCompilerReadBarrierOption; 5788 bool generate_null_check = false; 5789 switch (load_kind) { 5790 case HLoadClass::LoadKind::kReferrersClass: { 5791 DCHECK(!cls->CanCallRuntime()); 5792 DCHECK(!cls->MustGenerateClinitCheck()); 5793 // /* GcRoot<mirror::Class> */ out = current_method->declaring_class_ 5794 Register current_method = locations->InAt(0).AsRegister<Register>(); 5795 GenerateGcRootFieldLoad(cls, 5796 out_loc, 5797 current_method, 5798 ArtMethod::DeclaringClassOffset().Int32Value(), 5799 read_barrier_option); 5800 break; 5801 } 5802 case HLoadClass::LoadKind::kBootImageLinkTimeAddress: { 5803 DCHECK(codegen_->GetCompilerOptions().IsBootImage()); 5804 DCHECK_EQ(read_barrier_option, kWithoutReadBarrier); 5805 __ LoadLiteral(out, codegen_->DeduplicateBootImageTypeLiteral(cls->GetDexFile(), 5806 cls->GetTypeIndex())); 5807 break; 5808 } 5809 case HLoadClass::LoadKind::kBootImageLinkTimePcRelative: { 5810 DCHECK(codegen_->GetCompilerOptions().IsBootImage()); 5811 DCHECK_EQ(read_barrier_option, kWithoutReadBarrier); 5812 CodeGeneratorARM::PcRelativePatchInfo* labels = 5813 codegen_->NewPcRelativeTypePatch(cls->GetDexFile(), cls->GetTypeIndex()); 5814 __ BindTrackedLabel(&labels->movw_label); 5815 __ movw(out, /* placeholder */ 0u); 5816 __ BindTrackedLabel(&labels->movt_label); 5817 __ movt(out, /* placeholder */ 0u); 5818 __ BindTrackedLabel(&labels->add_pc_label); 5819 __ add(out, out, ShifterOperand(PC)); 5820 break; 5821 } 5822 case HLoadClass::LoadKind::kBootImageAddress: { 5823 DCHECK_EQ(read_barrier_option, kWithoutReadBarrier); 5824 uint32_t address = dchecked_integral_cast<uint32_t>( 5825 reinterpret_cast<uintptr_t>(cls->GetClass().Get())); 5826 DCHECK_NE(address, 0u); 5827 __ LoadLiteral(out, codegen_->DeduplicateBootImageAddressLiteral(address)); 5828 break; 5829 } 5830 case HLoadClass::LoadKind::kBssEntry: { 5831 CodeGeneratorARM::PcRelativePatchInfo* labels = 5832 codegen_->NewTypeBssEntryPatch(cls->GetDexFile(), cls->GetTypeIndex()); 5833 __ BindTrackedLabel(&labels->movw_label); 5834 __ movw(out, /* placeholder */ 0u); 5835 __ BindTrackedLabel(&labels->movt_label); 5836 __ movt(out, /* placeholder */ 0u); 5837 __ BindTrackedLabel(&labels->add_pc_label); 5838 __ add(out, out, ShifterOperand(PC)); 5839 GenerateGcRootFieldLoad(cls, out_loc, out, 0, kCompilerReadBarrierOption); 5840 generate_null_check = true; 5841 break; 5842 } 5843 case HLoadClass::LoadKind::kJitTableAddress: { 5844 __ LoadLiteral(out, codegen_->DeduplicateJitClassLiteral(cls->GetDexFile(), 5845 cls->GetTypeIndex(), 5846 cls->GetClass())); 5847 // /* GcRoot<mirror::Class> */ out = *out 5848 GenerateGcRootFieldLoad(cls, out_loc, out, /* offset */ 0, kCompilerReadBarrierOption); 5849 break; 5850 } 5851 case HLoadClass::LoadKind::kDexCacheViaMethod: 5852 LOG(FATAL) << "UNREACHABLE"; 5853 UNREACHABLE(); 5854 } 5855 5856 if (generate_null_check || cls->MustGenerateClinitCheck()) { 5857 DCHECK(cls->CanCallRuntime()); 5858 SlowPathCodeARM* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathARM( 5859 cls, cls, cls->GetDexPc(), cls->MustGenerateClinitCheck()); 5860 codegen_->AddSlowPath(slow_path); 5861 if (generate_null_check) { 5862 __ CompareAndBranchIfZero(out, slow_path->GetEntryLabel()); 5863 } 5864 if (cls->MustGenerateClinitCheck()) { 5865 GenerateClassInitializationCheck(slow_path, out); 5866 } else { 5867 __ Bind(slow_path->GetExitLabel()); 5868 } 5869 } 5870} 5871 5872void LocationsBuilderARM::VisitClinitCheck(HClinitCheck* check) { 5873 LocationSummary* locations = 5874 new (GetGraph()->GetArena()) LocationSummary(check, LocationSummary::kCallOnSlowPath); 5875 locations->SetInAt(0, Location::RequiresRegister()); 5876 if (check->HasUses()) { 5877 locations->SetOut(Location::SameAsFirstInput()); 5878 } 5879} 5880 5881void InstructionCodeGeneratorARM::VisitClinitCheck(HClinitCheck* check) { 5882 // We assume the class is not null. 5883 SlowPathCodeARM* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathARM( 5884 check->GetLoadClass(), check, check->GetDexPc(), true); 5885 codegen_->AddSlowPath(slow_path); 5886 GenerateClassInitializationCheck(slow_path, 5887 check->GetLocations()->InAt(0).AsRegister<Register>()); 5888} 5889 5890void InstructionCodeGeneratorARM::GenerateClassInitializationCheck( 5891 SlowPathCodeARM* slow_path, Register class_reg) { 5892 __ LoadFromOffset(kLoadWord, IP, class_reg, mirror::Class::StatusOffset().Int32Value()); 5893 __ cmp(IP, ShifterOperand(mirror::Class::kStatusInitialized)); 5894 __ b(slow_path->GetEntryLabel(), LT); 5895 // Even if the initialized flag is set, we may be in a situation where caches are not synced 5896 // properly. Therefore, we do a memory fence. 5897 __ dmb(ISH); 5898 __ Bind(slow_path->GetExitLabel()); 5899} 5900 5901HLoadString::LoadKind CodeGeneratorARM::GetSupportedLoadStringKind( 5902 HLoadString::LoadKind desired_string_load_kind) { 5903 switch (desired_string_load_kind) { 5904 case HLoadString::LoadKind::kBootImageLinkTimeAddress: 5905 DCHECK(!GetCompilerOptions().GetCompilePic()); 5906 break; 5907 case HLoadString::LoadKind::kBootImageLinkTimePcRelative: 5908 DCHECK(GetCompilerOptions().GetCompilePic()); 5909 break; 5910 case HLoadString::LoadKind::kBootImageAddress: 5911 break; 5912 case HLoadString::LoadKind::kBssEntry: 5913 DCHECK(!Runtime::Current()->UseJitCompilation()); 5914 break; 5915 case HLoadString::LoadKind::kJitTableAddress: 5916 DCHECK(Runtime::Current()->UseJitCompilation()); 5917 break; 5918 case HLoadString::LoadKind::kDexCacheViaMethod: 5919 break; 5920 } 5921 return desired_string_load_kind; 5922} 5923 5924void LocationsBuilderARM::VisitLoadString(HLoadString* load) { 5925 LocationSummary::CallKind call_kind = CodeGenerator::GetLoadStringCallKind(load); 5926 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(load, call_kind); 5927 HLoadString::LoadKind load_kind = load->GetLoadKind(); 5928 if (load_kind == HLoadString::LoadKind::kDexCacheViaMethod) { 5929 locations->SetOut(Location::RegisterLocation(R0)); 5930 } else { 5931 locations->SetOut(Location::RequiresRegister()); 5932 if (load_kind == HLoadString::LoadKind::kBssEntry) { 5933 if (!kUseReadBarrier || kUseBakerReadBarrier) { 5934 // Rely on the pResolveString and/or marking to save everything, including temps. 5935 // Note that IP may theoretically be clobbered by saving/restoring the live register 5936 // (only one thanks to the custom calling convention), so we request a different temp. 5937 locations->AddTemp(Location::RequiresRegister()); 5938 RegisterSet caller_saves = RegisterSet::Empty(); 5939 InvokeRuntimeCallingConvention calling_convention; 5940 caller_saves.Add(Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 5941 // TODO: Add GetReturnLocation() to the calling convention so that we can DCHECK() 5942 // that the the kPrimNot result register is the same as the first argument register. 5943 locations->SetCustomSlowPathCallerSaves(caller_saves); 5944 } else { 5945 // For non-Baker read barrier we have a temp-clobbering call. 5946 } 5947 } 5948 } 5949} 5950 5951// NO_THREAD_SAFETY_ANALYSIS as we manipulate handles whose internal object we know does not 5952// move. 5953void InstructionCodeGeneratorARM::VisitLoadString(HLoadString* load) NO_THREAD_SAFETY_ANALYSIS { 5954 LocationSummary* locations = load->GetLocations(); 5955 Location out_loc = locations->Out(); 5956 Register out = out_loc.AsRegister<Register>(); 5957 HLoadString::LoadKind load_kind = load->GetLoadKind(); 5958 5959 switch (load_kind) { 5960 case HLoadString::LoadKind::kBootImageLinkTimeAddress: { 5961 DCHECK(codegen_->GetCompilerOptions().IsBootImage()); 5962 __ LoadLiteral(out, codegen_->DeduplicateBootImageStringLiteral(load->GetDexFile(), 5963 load->GetStringIndex())); 5964 return; // No dex cache slow path. 5965 } 5966 case HLoadString::LoadKind::kBootImageLinkTimePcRelative: { 5967 DCHECK(codegen_->GetCompilerOptions().IsBootImage()); 5968 CodeGeneratorARM::PcRelativePatchInfo* labels = 5969 codegen_->NewPcRelativeStringPatch(load->GetDexFile(), load->GetStringIndex()); 5970 __ BindTrackedLabel(&labels->movw_label); 5971 __ movw(out, /* placeholder */ 0u); 5972 __ BindTrackedLabel(&labels->movt_label); 5973 __ movt(out, /* placeholder */ 0u); 5974 __ BindTrackedLabel(&labels->add_pc_label); 5975 __ add(out, out, ShifterOperand(PC)); 5976 return; // No dex cache slow path. 5977 } 5978 case HLoadString::LoadKind::kBootImageAddress: { 5979 uint32_t address = dchecked_integral_cast<uint32_t>( 5980 reinterpret_cast<uintptr_t>(load->GetString().Get())); 5981 DCHECK_NE(address, 0u); 5982 __ LoadLiteral(out, codegen_->DeduplicateBootImageAddressLiteral(address)); 5983 return; // No dex cache slow path. 5984 } 5985 case HLoadString::LoadKind::kBssEntry: { 5986 DCHECK(!codegen_->GetCompilerOptions().IsBootImage()); 5987 Register temp = locations->GetTemp(0).AsRegister<Register>(); 5988 CodeGeneratorARM::PcRelativePatchInfo* labels = 5989 codegen_->NewPcRelativeStringPatch(load->GetDexFile(), load->GetStringIndex()); 5990 __ BindTrackedLabel(&labels->movw_label); 5991 __ movw(temp, /* placeholder */ 0u); 5992 __ BindTrackedLabel(&labels->movt_label); 5993 __ movt(temp, /* placeholder */ 0u); 5994 __ BindTrackedLabel(&labels->add_pc_label); 5995 __ add(temp, temp, ShifterOperand(PC)); 5996 GenerateGcRootFieldLoad(load, out_loc, temp, /* offset */ 0, kCompilerReadBarrierOption); 5997 SlowPathCode* slow_path = new (GetGraph()->GetArena()) LoadStringSlowPathARM(load); 5998 codegen_->AddSlowPath(slow_path); 5999 __ CompareAndBranchIfZero(out, slow_path->GetEntryLabel()); 6000 __ Bind(slow_path->GetExitLabel()); 6001 return; 6002 } 6003 case HLoadString::LoadKind::kJitTableAddress: { 6004 __ LoadLiteral(out, codegen_->DeduplicateJitStringLiteral(load->GetDexFile(), 6005 load->GetStringIndex(), 6006 load->GetString())); 6007 // /* GcRoot<mirror::String> */ out = *out 6008 GenerateGcRootFieldLoad(load, out_loc, out, /* offset */ 0, kCompilerReadBarrierOption); 6009 return; 6010 } 6011 default: 6012 break; 6013 } 6014 6015 // TODO: Consider re-adding the compiler code to do string dex cache lookup again. 6016 DCHECK(load_kind == HLoadString::LoadKind::kDexCacheViaMethod); 6017 InvokeRuntimeCallingConvention calling_convention; 6018 DCHECK_EQ(calling_convention.GetRegisterAt(0), out); 6019 __ LoadImmediate(calling_convention.GetRegisterAt(0), load->GetStringIndex().index_); 6020 codegen_->InvokeRuntime(kQuickResolveString, load, load->GetDexPc()); 6021 CheckEntrypointTypes<kQuickResolveString, void*, uint32_t>(); 6022} 6023 6024static int32_t GetExceptionTlsOffset() { 6025 return Thread::ExceptionOffset<kArmPointerSize>().Int32Value(); 6026} 6027 6028void LocationsBuilderARM::VisitLoadException(HLoadException* load) { 6029 LocationSummary* locations = 6030 new (GetGraph()->GetArena()) LocationSummary(load, LocationSummary::kNoCall); 6031 locations->SetOut(Location::RequiresRegister()); 6032} 6033 6034void InstructionCodeGeneratorARM::VisitLoadException(HLoadException* load) { 6035 Register out = load->GetLocations()->Out().AsRegister<Register>(); 6036 __ LoadFromOffset(kLoadWord, out, TR, GetExceptionTlsOffset()); 6037} 6038 6039void LocationsBuilderARM::VisitClearException(HClearException* clear) { 6040 new (GetGraph()->GetArena()) LocationSummary(clear, LocationSummary::kNoCall); 6041} 6042 6043void InstructionCodeGeneratorARM::VisitClearException(HClearException* clear ATTRIBUTE_UNUSED) { 6044 __ LoadImmediate(IP, 0); 6045 __ StoreToOffset(kStoreWord, IP, TR, GetExceptionTlsOffset()); 6046} 6047 6048void LocationsBuilderARM::VisitThrow(HThrow* instruction) { 6049 LocationSummary* locations = 6050 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnMainOnly); 6051 InvokeRuntimeCallingConvention calling_convention; 6052 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 6053} 6054 6055void InstructionCodeGeneratorARM::VisitThrow(HThrow* instruction) { 6056 codegen_->InvokeRuntime(kQuickDeliverException, instruction, instruction->GetDexPc()); 6057 CheckEntrypointTypes<kQuickDeliverException, void, mirror::Object*>(); 6058} 6059 6060// Temp is used for read barrier. 6061static size_t NumberOfInstanceOfTemps(TypeCheckKind type_check_kind) { 6062 if (kEmitCompilerReadBarrier && 6063 (kUseBakerReadBarrier || 6064 type_check_kind == TypeCheckKind::kAbstractClassCheck || 6065 type_check_kind == TypeCheckKind::kClassHierarchyCheck || 6066 type_check_kind == TypeCheckKind::kArrayObjectCheck)) { 6067 return 1; 6068 } 6069 return 0; 6070} 6071 6072// Interface case has 3 temps, one for holding the number of interfaces, one for the current 6073// interface pointer, one for loading the current interface. 6074// The other checks have one temp for loading the object's class. 6075static size_t NumberOfCheckCastTemps(TypeCheckKind type_check_kind) { 6076 if (type_check_kind == TypeCheckKind::kInterfaceCheck) { 6077 return 3; 6078 } 6079 return 1 + NumberOfInstanceOfTemps(type_check_kind); 6080} 6081 6082void LocationsBuilderARM::VisitInstanceOf(HInstanceOf* instruction) { 6083 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 6084 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 6085 bool baker_read_barrier_slow_path = false; 6086 switch (type_check_kind) { 6087 case TypeCheckKind::kExactCheck: 6088 case TypeCheckKind::kAbstractClassCheck: 6089 case TypeCheckKind::kClassHierarchyCheck: 6090 case TypeCheckKind::kArrayObjectCheck: 6091 call_kind = 6092 kEmitCompilerReadBarrier ? LocationSummary::kCallOnSlowPath : LocationSummary::kNoCall; 6093 baker_read_barrier_slow_path = kUseBakerReadBarrier; 6094 break; 6095 case TypeCheckKind::kArrayCheck: 6096 case TypeCheckKind::kUnresolvedCheck: 6097 case TypeCheckKind::kInterfaceCheck: 6098 call_kind = LocationSummary::kCallOnSlowPath; 6099 break; 6100 } 6101 6102 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 6103 if (baker_read_barrier_slow_path) { 6104 locations->SetCustomSlowPathCallerSaves(RegisterSet::Empty()); // No caller-save registers. 6105 } 6106 locations->SetInAt(0, Location::RequiresRegister()); 6107 locations->SetInAt(1, Location::RequiresRegister()); 6108 // The "out" register is used as a temporary, so it overlaps with the inputs. 6109 // Note that TypeCheckSlowPathARM uses this register too. 6110 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 6111 locations->AddRegisterTemps(NumberOfInstanceOfTemps(type_check_kind)); 6112} 6113 6114void InstructionCodeGeneratorARM::VisitInstanceOf(HInstanceOf* instruction) { 6115 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 6116 LocationSummary* locations = instruction->GetLocations(); 6117 Location obj_loc = locations->InAt(0); 6118 Register obj = obj_loc.AsRegister<Register>(); 6119 Register cls = locations->InAt(1).AsRegister<Register>(); 6120 Location out_loc = locations->Out(); 6121 Register out = out_loc.AsRegister<Register>(); 6122 const size_t num_temps = NumberOfInstanceOfTemps(type_check_kind); 6123 DCHECK_LE(num_temps, 1u); 6124 Location maybe_temp_loc = (num_temps >= 1) ? locations->GetTemp(0) : Location::NoLocation(); 6125 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 6126 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 6127 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 6128 uint32_t primitive_offset = mirror::Class::PrimitiveTypeOffset().Int32Value(); 6129 Label done, zero; 6130 SlowPathCodeARM* slow_path = nullptr; 6131 6132 // Return 0 if `obj` is null. 6133 // avoid null check if we know obj is not null. 6134 if (instruction->MustDoNullCheck()) { 6135 __ CompareAndBranchIfZero(obj, &zero); 6136 } 6137 6138 switch (type_check_kind) { 6139 case TypeCheckKind::kExactCheck: { 6140 // /* HeapReference<Class> */ out = obj->klass_ 6141 GenerateReferenceLoadTwoRegisters(instruction, 6142 out_loc, 6143 obj_loc, 6144 class_offset, 6145 maybe_temp_loc, 6146 kCompilerReadBarrierOption); 6147 __ cmp(out, ShifterOperand(cls)); 6148 // Classes must be equal for the instanceof to succeed. 6149 __ b(&zero, NE); 6150 __ LoadImmediate(out, 1); 6151 __ b(&done); 6152 break; 6153 } 6154 6155 case TypeCheckKind::kAbstractClassCheck: { 6156 // /* HeapReference<Class> */ out = obj->klass_ 6157 GenerateReferenceLoadTwoRegisters(instruction, 6158 out_loc, 6159 obj_loc, 6160 class_offset, 6161 maybe_temp_loc, 6162 kCompilerReadBarrierOption); 6163 // If the class is abstract, we eagerly fetch the super class of the 6164 // object to avoid doing a comparison we know will fail. 6165 Label loop; 6166 __ Bind(&loop); 6167 // /* HeapReference<Class> */ out = out->super_class_ 6168 GenerateReferenceLoadOneRegister(instruction, 6169 out_loc, 6170 super_offset, 6171 maybe_temp_loc, 6172 kCompilerReadBarrierOption); 6173 // If `out` is null, we use it for the result, and jump to `done`. 6174 __ CompareAndBranchIfZero(out, &done); 6175 __ cmp(out, ShifterOperand(cls)); 6176 __ b(&loop, NE); 6177 __ LoadImmediate(out, 1); 6178 if (zero.IsLinked()) { 6179 __ b(&done); 6180 } 6181 break; 6182 } 6183 6184 case TypeCheckKind::kClassHierarchyCheck: { 6185 // /* HeapReference<Class> */ out = obj->klass_ 6186 GenerateReferenceLoadTwoRegisters(instruction, 6187 out_loc, 6188 obj_loc, 6189 class_offset, 6190 maybe_temp_loc, 6191 kCompilerReadBarrierOption); 6192 // Walk over the class hierarchy to find a match. 6193 Label loop, success; 6194 __ Bind(&loop); 6195 __ cmp(out, ShifterOperand(cls)); 6196 __ b(&success, EQ); 6197 // /* HeapReference<Class> */ out = out->super_class_ 6198 GenerateReferenceLoadOneRegister(instruction, 6199 out_loc, 6200 super_offset, 6201 maybe_temp_loc, 6202 kCompilerReadBarrierOption); 6203 __ CompareAndBranchIfNonZero(out, &loop); 6204 // If `out` is null, we use it for the result, and jump to `done`. 6205 __ b(&done); 6206 __ Bind(&success); 6207 __ LoadImmediate(out, 1); 6208 if (zero.IsLinked()) { 6209 __ b(&done); 6210 } 6211 break; 6212 } 6213 6214 case TypeCheckKind::kArrayObjectCheck: { 6215 // /* HeapReference<Class> */ out = obj->klass_ 6216 GenerateReferenceLoadTwoRegisters(instruction, 6217 out_loc, 6218 obj_loc, 6219 class_offset, 6220 maybe_temp_loc, 6221 kCompilerReadBarrierOption); 6222 // Do an exact check. 6223 Label exact_check; 6224 __ cmp(out, ShifterOperand(cls)); 6225 __ b(&exact_check, EQ); 6226 // Otherwise, we need to check that the object's class is a non-primitive array. 6227 // /* HeapReference<Class> */ out = out->component_type_ 6228 GenerateReferenceLoadOneRegister(instruction, 6229 out_loc, 6230 component_offset, 6231 maybe_temp_loc, 6232 kCompilerReadBarrierOption); 6233 // If `out` is null, we use it for the result, and jump to `done`. 6234 __ CompareAndBranchIfZero(out, &done); 6235 __ LoadFromOffset(kLoadUnsignedHalfword, out, out, primitive_offset); 6236 static_assert(Primitive::kPrimNot == 0, "Expected 0 for kPrimNot"); 6237 __ CompareAndBranchIfNonZero(out, &zero); 6238 __ Bind(&exact_check); 6239 __ LoadImmediate(out, 1); 6240 __ b(&done); 6241 break; 6242 } 6243 6244 case TypeCheckKind::kArrayCheck: { 6245 // No read barrier since the slow path will retry upon failure. 6246 // /* HeapReference<Class> */ out = obj->klass_ 6247 GenerateReferenceLoadTwoRegisters(instruction, 6248 out_loc, 6249 obj_loc, 6250 class_offset, 6251 maybe_temp_loc, 6252 kWithoutReadBarrier); 6253 __ cmp(out, ShifterOperand(cls)); 6254 DCHECK(locations->OnlyCallsOnSlowPath()); 6255 slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathARM(instruction, 6256 /* is_fatal */ false); 6257 codegen_->AddSlowPath(slow_path); 6258 __ b(slow_path->GetEntryLabel(), NE); 6259 __ LoadImmediate(out, 1); 6260 if (zero.IsLinked()) { 6261 __ b(&done); 6262 } 6263 break; 6264 } 6265 6266 case TypeCheckKind::kUnresolvedCheck: 6267 case TypeCheckKind::kInterfaceCheck: { 6268 // Note that we indeed only call on slow path, but we always go 6269 // into the slow path for the unresolved and interface check 6270 // cases. 6271 // 6272 // We cannot directly call the InstanceofNonTrivial runtime 6273 // entry point without resorting to a type checking slow path 6274 // here (i.e. by calling InvokeRuntime directly), as it would 6275 // require to assign fixed registers for the inputs of this 6276 // HInstanceOf instruction (following the runtime calling 6277 // convention), which might be cluttered by the potential first 6278 // read barrier emission at the beginning of this method. 6279 // 6280 // TODO: Introduce a new runtime entry point taking the object 6281 // to test (instead of its class) as argument, and let it deal 6282 // with the read barrier issues. This will let us refactor this 6283 // case of the `switch` code as it was previously (with a direct 6284 // call to the runtime not using a type checking slow path). 6285 // This should also be beneficial for the other cases above. 6286 DCHECK(locations->OnlyCallsOnSlowPath()); 6287 slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathARM(instruction, 6288 /* is_fatal */ false); 6289 codegen_->AddSlowPath(slow_path); 6290 __ b(slow_path->GetEntryLabel()); 6291 if (zero.IsLinked()) { 6292 __ b(&done); 6293 } 6294 break; 6295 } 6296 } 6297 6298 if (zero.IsLinked()) { 6299 __ Bind(&zero); 6300 __ LoadImmediate(out, 0); 6301 } 6302 6303 if (done.IsLinked()) { 6304 __ Bind(&done); 6305 } 6306 6307 if (slow_path != nullptr) { 6308 __ Bind(slow_path->GetExitLabel()); 6309 } 6310} 6311 6312void LocationsBuilderARM::VisitCheckCast(HCheckCast* instruction) { 6313 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 6314 bool throws_into_catch = instruction->CanThrowIntoCatchBlock(); 6315 6316 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 6317 switch (type_check_kind) { 6318 case TypeCheckKind::kExactCheck: 6319 case TypeCheckKind::kAbstractClassCheck: 6320 case TypeCheckKind::kClassHierarchyCheck: 6321 case TypeCheckKind::kArrayObjectCheck: 6322 call_kind = (throws_into_catch || kEmitCompilerReadBarrier) ? 6323 LocationSummary::kCallOnSlowPath : 6324 LocationSummary::kNoCall; // In fact, call on a fatal (non-returning) slow path. 6325 break; 6326 case TypeCheckKind::kArrayCheck: 6327 case TypeCheckKind::kUnresolvedCheck: 6328 case TypeCheckKind::kInterfaceCheck: 6329 call_kind = LocationSummary::kCallOnSlowPath; 6330 break; 6331 } 6332 6333 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 6334 locations->SetInAt(0, Location::RequiresRegister()); 6335 locations->SetInAt(1, Location::RequiresRegister()); 6336 locations->AddRegisterTemps(NumberOfCheckCastTemps(type_check_kind)); 6337} 6338 6339void InstructionCodeGeneratorARM::VisitCheckCast(HCheckCast* instruction) { 6340 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 6341 LocationSummary* locations = instruction->GetLocations(); 6342 Location obj_loc = locations->InAt(0); 6343 Register obj = obj_loc.AsRegister<Register>(); 6344 Register cls = locations->InAt(1).AsRegister<Register>(); 6345 Location temp_loc = locations->GetTemp(0); 6346 Register temp = temp_loc.AsRegister<Register>(); 6347 const size_t num_temps = NumberOfCheckCastTemps(type_check_kind); 6348 DCHECK_LE(num_temps, 3u); 6349 Location maybe_temp2_loc = (num_temps >= 2) ? locations->GetTemp(1) : Location::NoLocation(); 6350 Location maybe_temp3_loc = (num_temps >= 3) ? locations->GetTemp(2) : Location::NoLocation(); 6351 const uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 6352 const uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 6353 const uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 6354 const uint32_t primitive_offset = mirror::Class::PrimitiveTypeOffset().Int32Value(); 6355 const uint32_t iftable_offset = mirror::Class::IfTableOffset().Uint32Value(); 6356 const uint32_t array_length_offset = mirror::Array::LengthOffset().Uint32Value(); 6357 const uint32_t object_array_data_offset = 6358 mirror::Array::DataOffset(kHeapReferenceSize).Uint32Value(); 6359 6360 // Always false for read barriers since we may need to go to the entrypoint for non-fatal cases 6361 // from false negatives. The false negatives may come from avoiding read barriers below. Avoiding 6362 // read barriers is done for performance and code size reasons. 6363 bool is_type_check_slow_path_fatal = false; 6364 if (!kEmitCompilerReadBarrier) { 6365 is_type_check_slow_path_fatal = 6366 (type_check_kind == TypeCheckKind::kExactCheck || 6367 type_check_kind == TypeCheckKind::kAbstractClassCheck || 6368 type_check_kind == TypeCheckKind::kClassHierarchyCheck || 6369 type_check_kind == TypeCheckKind::kArrayObjectCheck) && 6370 !instruction->CanThrowIntoCatchBlock(); 6371 } 6372 SlowPathCodeARM* type_check_slow_path = 6373 new (GetGraph()->GetArena()) TypeCheckSlowPathARM(instruction, 6374 is_type_check_slow_path_fatal); 6375 codegen_->AddSlowPath(type_check_slow_path); 6376 6377 Label done; 6378 // Avoid null check if we know obj is not null. 6379 if (instruction->MustDoNullCheck()) { 6380 __ CompareAndBranchIfZero(obj, &done); 6381 } 6382 6383 switch (type_check_kind) { 6384 case TypeCheckKind::kExactCheck: 6385 case TypeCheckKind::kArrayCheck: { 6386 // /* HeapReference<Class> */ temp = obj->klass_ 6387 GenerateReferenceLoadTwoRegisters(instruction, 6388 temp_loc, 6389 obj_loc, 6390 class_offset, 6391 maybe_temp2_loc, 6392 kWithoutReadBarrier); 6393 6394 __ cmp(temp, ShifterOperand(cls)); 6395 // Jump to slow path for throwing the exception or doing a 6396 // more involved array check. 6397 __ b(type_check_slow_path->GetEntryLabel(), NE); 6398 break; 6399 } 6400 6401 case TypeCheckKind::kAbstractClassCheck: { 6402 // /* HeapReference<Class> */ temp = obj->klass_ 6403 GenerateReferenceLoadTwoRegisters(instruction, 6404 temp_loc, 6405 obj_loc, 6406 class_offset, 6407 maybe_temp2_loc, 6408 kWithoutReadBarrier); 6409 6410 // If the class is abstract, we eagerly fetch the super class of the 6411 // object to avoid doing a comparison we know will fail. 6412 Label loop; 6413 __ Bind(&loop); 6414 // /* HeapReference<Class> */ temp = temp->super_class_ 6415 GenerateReferenceLoadOneRegister(instruction, 6416 temp_loc, 6417 super_offset, 6418 maybe_temp2_loc, 6419 kWithoutReadBarrier); 6420 6421 // If the class reference currently in `temp` is null, jump to the slow path to throw the 6422 // exception. 6423 __ CompareAndBranchIfZero(temp, type_check_slow_path->GetEntryLabel()); 6424 6425 // Otherwise, compare the classes. 6426 __ cmp(temp, ShifterOperand(cls)); 6427 __ b(&loop, NE); 6428 break; 6429 } 6430 6431 case TypeCheckKind::kClassHierarchyCheck: { 6432 // /* HeapReference<Class> */ temp = obj->klass_ 6433 GenerateReferenceLoadTwoRegisters(instruction, 6434 temp_loc, 6435 obj_loc, 6436 class_offset, 6437 maybe_temp2_loc, 6438 kWithoutReadBarrier); 6439 6440 // Walk over the class hierarchy to find a match. 6441 Label loop; 6442 __ Bind(&loop); 6443 __ cmp(temp, ShifterOperand(cls)); 6444 __ b(&done, EQ); 6445 6446 // /* HeapReference<Class> */ temp = temp->super_class_ 6447 GenerateReferenceLoadOneRegister(instruction, 6448 temp_loc, 6449 super_offset, 6450 maybe_temp2_loc, 6451 kWithoutReadBarrier); 6452 6453 // If the class reference currently in `temp` is null, jump to the slow path to throw the 6454 // exception. 6455 __ CompareAndBranchIfZero(temp, type_check_slow_path->GetEntryLabel()); 6456 // Otherwise, jump to the beginning of the loop. 6457 __ b(&loop); 6458 break; 6459 } 6460 6461 case TypeCheckKind::kArrayObjectCheck: { 6462 // /* HeapReference<Class> */ temp = obj->klass_ 6463 GenerateReferenceLoadTwoRegisters(instruction, 6464 temp_loc, 6465 obj_loc, 6466 class_offset, 6467 maybe_temp2_loc, 6468 kWithoutReadBarrier); 6469 6470 // Do an exact check. 6471 __ cmp(temp, ShifterOperand(cls)); 6472 __ b(&done, EQ); 6473 6474 // Otherwise, we need to check that the object's class is a non-primitive array. 6475 // /* HeapReference<Class> */ temp = temp->component_type_ 6476 GenerateReferenceLoadOneRegister(instruction, 6477 temp_loc, 6478 component_offset, 6479 maybe_temp2_loc, 6480 kWithoutReadBarrier); 6481 // If the component type is null, jump to the slow path to throw the exception. 6482 __ CompareAndBranchIfZero(temp, type_check_slow_path->GetEntryLabel()); 6483 // Otherwise,the object is indeed an array, jump to label `check_non_primitive_component_type` 6484 // to further check that this component type is not a primitive type. 6485 __ LoadFromOffset(kLoadUnsignedHalfword, temp, temp, primitive_offset); 6486 static_assert(Primitive::kPrimNot == 0, "Expected 0 for art::Primitive::kPrimNot"); 6487 __ CompareAndBranchIfNonZero(temp, type_check_slow_path->GetEntryLabel()); 6488 break; 6489 } 6490 6491 case TypeCheckKind::kUnresolvedCheck: 6492 // We always go into the type check slow path for the unresolved check case. 6493 // We cannot directly call the CheckCast runtime entry point 6494 // without resorting to a type checking slow path here (i.e. by 6495 // calling InvokeRuntime directly), as it would require to 6496 // assign fixed registers for the inputs of this HInstanceOf 6497 // instruction (following the runtime calling convention), which 6498 // might be cluttered by the potential first read barrier 6499 // emission at the beginning of this method. 6500 6501 __ b(type_check_slow_path->GetEntryLabel()); 6502 break; 6503 6504 case TypeCheckKind::kInterfaceCheck: { 6505 // Avoid read barriers to improve performance of the fast path. We can not get false 6506 // positives by doing this. 6507 // /* HeapReference<Class> */ temp = obj->klass_ 6508 GenerateReferenceLoadTwoRegisters(instruction, 6509 temp_loc, 6510 obj_loc, 6511 class_offset, 6512 maybe_temp2_loc, 6513 kWithoutReadBarrier); 6514 6515 // /* HeapReference<Class> */ temp = temp->iftable_ 6516 GenerateReferenceLoadTwoRegisters(instruction, 6517 temp_loc, 6518 temp_loc, 6519 iftable_offset, 6520 maybe_temp2_loc, 6521 kWithoutReadBarrier); 6522 // Iftable is never null. 6523 __ ldr(maybe_temp2_loc.AsRegister<Register>(), Address(temp, array_length_offset)); 6524 // Loop through the iftable and check if any class matches. 6525 Label start_loop; 6526 __ Bind(&start_loop); 6527 __ CompareAndBranchIfZero(maybe_temp2_loc.AsRegister<Register>(), 6528 type_check_slow_path->GetEntryLabel()); 6529 __ ldr(maybe_temp3_loc.AsRegister<Register>(), Address(temp, object_array_data_offset)); 6530 __ MaybeUnpoisonHeapReference(maybe_temp3_loc.AsRegister<Register>()); 6531 // Go to next interface. 6532 __ add(temp, temp, ShifterOperand(2 * kHeapReferenceSize)); 6533 __ sub(maybe_temp2_loc.AsRegister<Register>(), 6534 maybe_temp2_loc.AsRegister<Register>(), 6535 ShifterOperand(2)); 6536 // Compare the classes and continue the loop if they do not match. 6537 __ cmp(cls, ShifterOperand(maybe_temp3_loc.AsRegister<Register>())); 6538 __ b(&start_loop, NE); 6539 break; 6540 } 6541 } 6542 __ Bind(&done); 6543 6544 __ Bind(type_check_slow_path->GetExitLabel()); 6545} 6546 6547void LocationsBuilderARM::VisitMonitorOperation(HMonitorOperation* instruction) { 6548 LocationSummary* locations = 6549 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnMainOnly); 6550 InvokeRuntimeCallingConvention calling_convention; 6551 locations->SetInAt(0, Location::RegisterLocation(calling_convention.GetRegisterAt(0))); 6552} 6553 6554void InstructionCodeGeneratorARM::VisitMonitorOperation(HMonitorOperation* instruction) { 6555 codegen_->InvokeRuntime(instruction->IsEnter() ? kQuickLockObject : kQuickUnlockObject, 6556 instruction, 6557 instruction->GetDexPc()); 6558 if (instruction->IsEnter()) { 6559 CheckEntrypointTypes<kQuickLockObject, void, mirror::Object*>(); 6560 } else { 6561 CheckEntrypointTypes<kQuickUnlockObject, void, mirror::Object*>(); 6562 } 6563} 6564 6565void LocationsBuilderARM::VisitAnd(HAnd* instruction) { HandleBitwiseOperation(instruction, AND); } 6566void LocationsBuilderARM::VisitOr(HOr* instruction) { HandleBitwiseOperation(instruction, ORR); } 6567void LocationsBuilderARM::VisitXor(HXor* instruction) { HandleBitwiseOperation(instruction, EOR); } 6568 6569void LocationsBuilderARM::HandleBitwiseOperation(HBinaryOperation* instruction, Opcode opcode) { 6570 LocationSummary* locations = 6571 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 6572 DCHECK(instruction->GetResultType() == Primitive::kPrimInt 6573 || instruction->GetResultType() == Primitive::kPrimLong); 6574 // Note: GVN reorders commutative operations to have the constant on the right hand side. 6575 locations->SetInAt(0, Location::RequiresRegister()); 6576 locations->SetInAt(1, ArmEncodableConstantOrRegister(instruction->InputAt(1), opcode)); 6577 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 6578} 6579 6580void InstructionCodeGeneratorARM::VisitAnd(HAnd* instruction) { 6581 HandleBitwiseOperation(instruction); 6582} 6583 6584void InstructionCodeGeneratorARM::VisitOr(HOr* instruction) { 6585 HandleBitwiseOperation(instruction); 6586} 6587 6588void InstructionCodeGeneratorARM::VisitXor(HXor* instruction) { 6589 HandleBitwiseOperation(instruction); 6590} 6591 6592 6593void LocationsBuilderARM::VisitBitwiseNegatedRight(HBitwiseNegatedRight* instruction) { 6594 LocationSummary* locations = 6595 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 6596 DCHECK(instruction->GetResultType() == Primitive::kPrimInt 6597 || instruction->GetResultType() == Primitive::kPrimLong); 6598 6599 locations->SetInAt(0, Location::RequiresRegister()); 6600 locations->SetInAt(1, Location::RequiresRegister()); 6601 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 6602} 6603 6604void InstructionCodeGeneratorARM::VisitBitwiseNegatedRight(HBitwiseNegatedRight* instruction) { 6605 LocationSummary* locations = instruction->GetLocations(); 6606 Location first = locations->InAt(0); 6607 Location second = locations->InAt(1); 6608 Location out = locations->Out(); 6609 6610 if (instruction->GetResultType() == Primitive::kPrimInt) { 6611 Register first_reg = first.AsRegister<Register>(); 6612 ShifterOperand second_reg(second.AsRegister<Register>()); 6613 Register out_reg = out.AsRegister<Register>(); 6614 6615 switch (instruction->GetOpKind()) { 6616 case HInstruction::kAnd: 6617 __ bic(out_reg, first_reg, second_reg); 6618 break; 6619 case HInstruction::kOr: 6620 __ orn(out_reg, first_reg, second_reg); 6621 break; 6622 // There is no EON on arm. 6623 case HInstruction::kXor: 6624 default: 6625 LOG(FATAL) << "Unexpected instruction " << instruction->DebugName(); 6626 UNREACHABLE(); 6627 } 6628 return; 6629 6630 } else { 6631 DCHECK_EQ(instruction->GetResultType(), Primitive::kPrimLong); 6632 Register first_low = first.AsRegisterPairLow<Register>(); 6633 Register first_high = first.AsRegisterPairHigh<Register>(); 6634 ShifterOperand second_low(second.AsRegisterPairLow<Register>()); 6635 ShifterOperand second_high(second.AsRegisterPairHigh<Register>()); 6636 Register out_low = out.AsRegisterPairLow<Register>(); 6637 Register out_high = out.AsRegisterPairHigh<Register>(); 6638 6639 switch (instruction->GetOpKind()) { 6640 case HInstruction::kAnd: 6641 __ bic(out_low, first_low, second_low); 6642 __ bic(out_high, first_high, second_high); 6643 break; 6644 case HInstruction::kOr: 6645 __ orn(out_low, first_low, second_low); 6646 __ orn(out_high, first_high, second_high); 6647 break; 6648 // There is no EON on arm. 6649 case HInstruction::kXor: 6650 default: 6651 LOG(FATAL) << "Unexpected instruction " << instruction->DebugName(); 6652 UNREACHABLE(); 6653 } 6654 } 6655} 6656 6657void InstructionCodeGeneratorARM::GenerateAndConst(Register out, Register first, uint32_t value) { 6658 // Optimize special cases for individual halfs of `and-long` (`and` is simplified earlier). 6659 if (value == 0xffffffffu) { 6660 if (out != first) { 6661 __ mov(out, ShifterOperand(first)); 6662 } 6663 return; 6664 } 6665 if (value == 0u) { 6666 __ mov(out, ShifterOperand(0)); 6667 return; 6668 } 6669 ShifterOperand so; 6670 if (__ ShifterOperandCanHold(kNoRegister, kNoRegister, AND, value, &so)) { 6671 __ and_(out, first, so); 6672 } else { 6673 DCHECK(__ ShifterOperandCanHold(kNoRegister, kNoRegister, BIC, ~value, &so)); 6674 __ bic(out, first, ShifterOperand(~value)); 6675 } 6676} 6677 6678void InstructionCodeGeneratorARM::GenerateOrrConst(Register out, Register first, uint32_t value) { 6679 // Optimize special cases for individual halfs of `or-long` (`or` is simplified earlier). 6680 if (value == 0u) { 6681 if (out != first) { 6682 __ mov(out, ShifterOperand(first)); 6683 } 6684 return; 6685 } 6686 if (value == 0xffffffffu) { 6687 __ mvn(out, ShifterOperand(0)); 6688 return; 6689 } 6690 ShifterOperand so; 6691 if (__ ShifterOperandCanHold(kNoRegister, kNoRegister, ORR, value, &so)) { 6692 __ orr(out, first, so); 6693 } else { 6694 DCHECK(__ ShifterOperandCanHold(kNoRegister, kNoRegister, ORN, ~value, &so)); 6695 __ orn(out, first, ShifterOperand(~value)); 6696 } 6697} 6698 6699void InstructionCodeGeneratorARM::GenerateEorConst(Register out, Register first, uint32_t value) { 6700 // Optimize special case for individual halfs of `xor-long` (`xor` is simplified earlier). 6701 if (value == 0u) { 6702 if (out != first) { 6703 __ mov(out, ShifterOperand(first)); 6704 } 6705 return; 6706 } 6707 __ eor(out, first, ShifterOperand(value)); 6708} 6709 6710void InstructionCodeGeneratorARM::GenerateAddLongConst(Location out, 6711 Location first, 6712 uint64_t value) { 6713 Register out_low = out.AsRegisterPairLow<Register>(); 6714 Register out_high = out.AsRegisterPairHigh<Register>(); 6715 Register first_low = first.AsRegisterPairLow<Register>(); 6716 Register first_high = first.AsRegisterPairHigh<Register>(); 6717 uint32_t value_low = Low32Bits(value); 6718 uint32_t value_high = High32Bits(value); 6719 if (value_low == 0u) { 6720 if (out_low != first_low) { 6721 __ mov(out_low, ShifterOperand(first_low)); 6722 } 6723 __ AddConstant(out_high, first_high, value_high); 6724 return; 6725 } 6726 __ AddConstantSetFlags(out_low, first_low, value_low); 6727 ShifterOperand so; 6728 if (__ ShifterOperandCanHold(out_high, first_high, ADC, value_high, kCcDontCare, &so)) { 6729 __ adc(out_high, first_high, so); 6730 } else if (__ ShifterOperandCanHold(out_low, first_low, SBC, ~value_high, kCcDontCare, &so)) { 6731 __ sbc(out_high, first_high, so); 6732 } else { 6733 LOG(FATAL) << "Unexpected constant " << value_high; 6734 UNREACHABLE(); 6735 } 6736} 6737 6738void InstructionCodeGeneratorARM::HandleBitwiseOperation(HBinaryOperation* instruction) { 6739 LocationSummary* locations = instruction->GetLocations(); 6740 Location first = locations->InAt(0); 6741 Location second = locations->InAt(1); 6742 Location out = locations->Out(); 6743 6744 if (second.IsConstant()) { 6745 uint64_t value = static_cast<uint64_t>(Int64FromConstant(second.GetConstant())); 6746 uint32_t value_low = Low32Bits(value); 6747 if (instruction->GetResultType() == Primitive::kPrimInt) { 6748 Register first_reg = first.AsRegister<Register>(); 6749 Register out_reg = out.AsRegister<Register>(); 6750 if (instruction->IsAnd()) { 6751 GenerateAndConst(out_reg, first_reg, value_low); 6752 } else if (instruction->IsOr()) { 6753 GenerateOrrConst(out_reg, first_reg, value_low); 6754 } else { 6755 DCHECK(instruction->IsXor()); 6756 GenerateEorConst(out_reg, first_reg, value_low); 6757 } 6758 } else { 6759 DCHECK_EQ(instruction->GetResultType(), Primitive::kPrimLong); 6760 uint32_t value_high = High32Bits(value); 6761 Register first_low = first.AsRegisterPairLow<Register>(); 6762 Register first_high = first.AsRegisterPairHigh<Register>(); 6763 Register out_low = out.AsRegisterPairLow<Register>(); 6764 Register out_high = out.AsRegisterPairHigh<Register>(); 6765 if (instruction->IsAnd()) { 6766 GenerateAndConst(out_low, first_low, value_low); 6767 GenerateAndConst(out_high, first_high, value_high); 6768 } else if (instruction->IsOr()) { 6769 GenerateOrrConst(out_low, first_low, value_low); 6770 GenerateOrrConst(out_high, first_high, value_high); 6771 } else { 6772 DCHECK(instruction->IsXor()); 6773 GenerateEorConst(out_low, first_low, value_low); 6774 GenerateEorConst(out_high, first_high, value_high); 6775 } 6776 } 6777 return; 6778 } 6779 6780 if (instruction->GetResultType() == Primitive::kPrimInt) { 6781 Register first_reg = first.AsRegister<Register>(); 6782 ShifterOperand second_reg(second.AsRegister<Register>()); 6783 Register out_reg = out.AsRegister<Register>(); 6784 if (instruction->IsAnd()) { 6785 __ and_(out_reg, first_reg, second_reg); 6786 } else if (instruction->IsOr()) { 6787 __ orr(out_reg, first_reg, second_reg); 6788 } else { 6789 DCHECK(instruction->IsXor()); 6790 __ eor(out_reg, first_reg, second_reg); 6791 } 6792 } else { 6793 DCHECK_EQ(instruction->GetResultType(), Primitive::kPrimLong); 6794 Register first_low = first.AsRegisterPairLow<Register>(); 6795 Register first_high = first.AsRegisterPairHigh<Register>(); 6796 ShifterOperand second_low(second.AsRegisterPairLow<Register>()); 6797 ShifterOperand second_high(second.AsRegisterPairHigh<Register>()); 6798 Register out_low = out.AsRegisterPairLow<Register>(); 6799 Register out_high = out.AsRegisterPairHigh<Register>(); 6800 if (instruction->IsAnd()) { 6801 __ and_(out_low, first_low, second_low); 6802 __ and_(out_high, first_high, second_high); 6803 } else if (instruction->IsOr()) { 6804 __ orr(out_low, first_low, second_low); 6805 __ orr(out_high, first_high, second_high); 6806 } else { 6807 DCHECK(instruction->IsXor()); 6808 __ eor(out_low, first_low, second_low); 6809 __ eor(out_high, first_high, second_high); 6810 } 6811 } 6812} 6813 6814void InstructionCodeGeneratorARM::GenerateReferenceLoadOneRegister( 6815 HInstruction* instruction, 6816 Location out, 6817 uint32_t offset, 6818 Location maybe_temp, 6819 ReadBarrierOption read_barrier_option) { 6820 Register out_reg = out.AsRegister<Register>(); 6821 if (read_barrier_option == kWithReadBarrier) { 6822 CHECK(kEmitCompilerReadBarrier); 6823 DCHECK(maybe_temp.IsRegister()) << maybe_temp; 6824 if (kUseBakerReadBarrier) { 6825 // Load with fast path based Baker's read barrier. 6826 // /* HeapReference<Object> */ out = *(out + offset) 6827 codegen_->GenerateFieldLoadWithBakerReadBarrier( 6828 instruction, out, out_reg, offset, maybe_temp, /* needs_null_check */ false); 6829 } else { 6830 // Load with slow path based read barrier. 6831 // Save the value of `out` into `maybe_temp` before overwriting it 6832 // in the following move operation, as we will need it for the 6833 // read barrier below. 6834 __ Mov(maybe_temp.AsRegister<Register>(), out_reg); 6835 // /* HeapReference<Object> */ out = *(out + offset) 6836 __ LoadFromOffset(kLoadWord, out_reg, out_reg, offset); 6837 codegen_->GenerateReadBarrierSlow(instruction, out, out, maybe_temp, offset); 6838 } 6839 } else { 6840 // Plain load with no read barrier. 6841 // /* HeapReference<Object> */ out = *(out + offset) 6842 __ LoadFromOffset(kLoadWord, out_reg, out_reg, offset); 6843 __ MaybeUnpoisonHeapReference(out_reg); 6844 } 6845} 6846 6847void InstructionCodeGeneratorARM::GenerateReferenceLoadTwoRegisters( 6848 HInstruction* instruction, 6849 Location out, 6850 Location obj, 6851 uint32_t offset, 6852 Location maybe_temp, 6853 ReadBarrierOption read_barrier_option) { 6854 Register out_reg = out.AsRegister<Register>(); 6855 Register obj_reg = obj.AsRegister<Register>(); 6856 if (read_barrier_option == kWithReadBarrier) { 6857 CHECK(kEmitCompilerReadBarrier); 6858 if (kUseBakerReadBarrier) { 6859 DCHECK(maybe_temp.IsRegister()) << maybe_temp; 6860 // Load with fast path based Baker's read barrier. 6861 // /* HeapReference<Object> */ out = *(obj + offset) 6862 codegen_->GenerateFieldLoadWithBakerReadBarrier( 6863 instruction, out, obj_reg, offset, maybe_temp, /* needs_null_check */ false); 6864 } else { 6865 // Load with slow path based read barrier. 6866 // /* HeapReference<Object> */ out = *(obj + offset) 6867 __ LoadFromOffset(kLoadWord, out_reg, obj_reg, offset); 6868 codegen_->GenerateReadBarrierSlow(instruction, out, out, obj, offset); 6869 } 6870 } else { 6871 // Plain load with no read barrier. 6872 // /* HeapReference<Object> */ out = *(obj + offset) 6873 __ LoadFromOffset(kLoadWord, out_reg, obj_reg, offset); 6874 __ MaybeUnpoisonHeapReference(out_reg); 6875 } 6876} 6877 6878void InstructionCodeGeneratorARM::GenerateGcRootFieldLoad(HInstruction* instruction, 6879 Location root, 6880 Register obj, 6881 uint32_t offset, 6882 ReadBarrierOption read_barrier_option) { 6883 Register root_reg = root.AsRegister<Register>(); 6884 if (read_barrier_option == kWithReadBarrier) { 6885 DCHECK(kEmitCompilerReadBarrier); 6886 if (kUseBakerReadBarrier) { 6887 // Fast path implementation of art::ReadBarrier::BarrierForRoot when 6888 // Baker's read barrier are used: 6889 // 6890 // root = obj.field; 6891 // temp = Thread::Current()->pReadBarrierMarkReg ## root.reg() 6892 // if (temp != null) { 6893 // root = temp(root) 6894 // } 6895 6896 // /* GcRoot<mirror::Object> */ root = *(obj + offset) 6897 __ LoadFromOffset(kLoadWord, root_reg, obj, offset); 6898 static_assert( 6899 sizeof(mirror::CompressedReference<mirror::Object>) == sizeof(GcRoot<mirror::Object>), 6900 "art::mirror::CompressedReference<mirror::Object> and art::GcRoot<mirror::Object> " 6901 "have different sizes."); 6902 static_assert(sizeof(mirror::CompressedReference<mirror::Object>) == sizeof(int32_t), 6903 "art::mirror::CompressedReference<mirror::Object> and int32_t " 6904 "have different sizes."); 6905 6906 // Slow path marking the GC root `root`. 6907 Location temp = Location::RegisterLocation(LR); 6908 SlowPathCodeARM* slow_path = 6909 new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM( 6910 instruction, 6911 root, 6912 /*entrypoint*/ temp); 6913 codegen_->AddSlowPath(slow_path); 6914 6915 // temp = Thread::Current()->pReadBarrierMarkReg ## root.reg() 6916 const int32_t entry_point_offset = 6917 CodeGenerator::GetReadBarrierMarkEntryPointsOffset<kArmPointerSize>(root.reg()); 6918 // Loading the entrypoint does not require a load acquire since it is only changed when 6919 // threads are suspended or running a checkpoint. 6920 __ LoadFromOffset(kLoadWord, temp.AsRegister<Register>(), TR, entry_point_offset); 6921 // The entrypoint is null when the GC is not marking, this prevents one load compared to 6922 // checking GetIsGcMarking. 6923 __ CompareAndBranchIfNonZero(temp.AsRegister<Register>(), slow_path->GetEntryLabel()); 6924 __ Bind(slow_path->GetExitLabel()); 6925 } else { 6926 // GC root loaded through a slow path for read barriers other 6927 // than Baker's. 6928 // /* GcRoot<mirror::Object>* */ root = obj + offset 6929 __ AddConstant(root_reg, obj, offset); 6930 // /* mirror::Object* */ root = root->Read() 6931 codegen_->GenerateReadBarrierForRootSlow(instruction, root, root); 6932 } 6933 } else { 6934 // Plain GC root load with no read barrier. 6935 // /* GcRoot<mirror::Object> */ root = *(obj + offset) 6936 __ LoadFromOffset(kLoadWord, root_reg, obj, offset); 6937 // Note that GC roots are not affected by heap poisoning, thus we 6938 // do not have to unpoison `root_reg` here. 6939 } 6940} 6941 6942void CodeGeneratorARM::GenerateFieldLoadWithBakerReadBarrier(HInstruction* instruction, 6943 Location ref, 6944 Register obj, 6945 uint32_t offset, 6946 Location temp, 6947 bool needs_null_check) { 6948 DCHECK(kEmitCompilerReadBarrier); 6949 DCHECK(kUseBakerReadBarrier); 6950 6951 // /* HeapReference<Object> */ ref = *(obj + offset) 6952 Location no_index = Location::NoLocation(); 6953 ScaleFactor no_scale_factor = TIMES_1; 6954 GenerateReferenceLoadWithBakerReadBarrier( 6955 instruction, ref, obj, offset, no_index, no_scale_factor, temp, needs_null_check); 6956} 6957 6958void CodeGeneratorARM::GenerateArrayLoadWithBakerReadBarrier(HInstruction* instruction, 6959 Location ref, 6960 Register obj, 6961 uint32_t data_offset, 6962 Location index, 6963 Location temp, 6964 bool needs_null_check) { 6965 DCHECK(kEmitCompilerReadBarrier); 6966 DCHECK(kUseBakerReadBarrier); 6967 6968 static_assert( 6969 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 6970 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 6971 // /* HeapReference<Object> */ ref = 6972 // *(obj + data_offset + index * sizeof(HeapReference<Object>)) 6973 ScaleFactor scale_factor = TIMES_4; 6974 GenerateReferenceLoadWithBakerReadBarrier( 6975 instruction, ref, obj, data_offset, index, scale_factor, temp, needs_null_check); 6976} 6977 6978void CodeGeneratorARM::GenerateReferenceLoadWithBakerReadBarrier(HInstruction* instruction, 6979 Location ref, 6980 Register obj, 6981 uint32_t offset, 6982 Location index, 6983 ScaleFactor scale_factor, 6984 Location temp, 6985 bool needs_null_check, 6986 bool always_update_field, 6987 Register* temp2) { 6988 DCHECK(kEmitCompilerReadBarrier); 6989 DCHECK(kUseBakerReadBarrier); 6990 6991 // In slow path based read barriers, the read barrier call is 6992 // inserted after the original load. However, in fast path based 6993 // Baker's read barriers, we need to perform the load of 6994 // mirror::Object::monitor_ *before* the original reference load. 6995 // This load-load ordering is required by the read barrier. 6996 // The fast path/slow path (for Baker's algorithm) should look like: 6997 // 6998 // uint32_t rb_state = Lockword(obj->monitor_).ReadBarrierState(); 6999 // lfence; // Load fence or artificial data dependency to prevent load-load reordering 7000 // HeapReference<Object> ref = *src; // Original reference load. 7001 // bool is_gray = (rb_state == ReadBarrier::GrayState()); 7002 // if (is_gray) { 7003 // ref = ReadBarrier::Mark(ref); // Performed by runtime entrypoint slow path. 7004 // } 7005 // 7006 // Note: the original implementation in ReadBarrier::Barrier is 7007 // slightly more complex as it performs additional checks that we do 7008 // not do here for performance reasons. 7009 7010 Register ref_reg = ref.AsRegister<Register>(); 7011 Register temp_reg = temp.AsRegister<Register>(); 7012 uint32_t monitor_offset = mirror::Object::MonitorOffset().Int32Value(); 7013 7014 // /* int32_t */ monitor = obj->monitor_ 7015 __ LoadFromOffset(kLoadWord, temp_reg, obj, monitor_offset); 7016 if (needs_null_check) { 7017 MaybeRecordImplicitNullCheck(instruction); 7018 } 7019 // /* LockWord */ lock_word = LockWord(monitor) 7020 static_assert(sizeof(LockWord) == sizeof(int32_t), 7021 "art::LockWord and int32_t have different sizes."); 7022 7023 // Introduce a dependency on the lock_word including the rb_state, 7024 // which shall prevent load-load reordering without using 7025 // a memory barrier (which would be more expensive). 7026 // `obj` is unchanged by this operation, but its value now depends 7027 // on `temp_reg`. 7028 __ add(obj, obj, ShifterOperand(temp_reg, LSR, 32)); 7029 7030 // The actual reference load. 7031 if (index.IsValid()) { 7032 // Load types involving an "index": ArrayGet, 7033 // UnsafeGetObject/UnsafeGetObjectVolatile and UnsafeCASObject 7034 // intrinsics. 7035 // /* HeapReference<Object> */ ref = *(obj + offset + (index << scale_factor)) 7036 if (index.IsConstant()) { 7037 size_t computed_offset = 7038 (index.GetConstant()->AsIntConstant()->GetValue() << scale_factor) + offset; 7039 __ LoadFromOffset(kLoadWord, ref_reg, obj, computed_offset); 7040 } else { 7041 // Handle the special case of the 7042 // UnsafeGetObject/UnsafeGetObjectVolatile and UnsafeCASObject 7043 // intrinsics, which use a register pair as index ("long 7044 // offset"), of which only the low part contains data. 7045 Register index_reg = index.IsRegisterPair() 7046 ? index.AsRegisterPairLow<Register>() 7047 : index.AsRegister<Register>(); 7048 __ add(IP, obj, ShifterOperand(index_reg, LSL, scale_factor)); 7049 __ LoadFromOffset(kLoadWord, ref_reg, IP, offset); 7050 } 7051 } else { 7052 // /* HeapReference<Object> */ ref = *(obj + offset) 7053 __ LoadFromOffset(kLoadWord, ref_reg, obj, offset); 7054 } 7055 7056 // Object* ref = ref_addr->AsMirrorPtr() 7057 __ MaybeUnpoisonHeapReference(ref_reg); 7058 7059 // Slow path marking the object `ref` when it is gray. 7060 SlowPathCodeARM* slow_path; 7061 if (always_update_field) { 7062 DCHECK(temp2 != nullptr); 7063 // ReadBarrierMarkAndUpdateFieldSlowPathARM only supports address 7064 // of the form `obj + field_offset`, where `obj` is a register and 7065 // `field_offset` is a register pair (of which only the lower half 7066 // is used). Thus `offset` and `scale_factor` above are expected 7067 // to be null in this code path. 7068 DCHECK_EQ(offset, 0u); 7069 DCHECK_EQ(scale_factor, ScaleFactor::TIMES_1); 7070 slow_path = new (GetGraph()->GetArena()) ReadBarrierMarkAndUpdateFieldSlowPathARM( 7071 instruction, ref, obj, /* field_offset */ index, temp_reg, *temp2); 7072 } else { 7073 slow_path = new (GetGraph()->GetArena()) ReadBarrierMarkSlowPathARM(instruction, ref); 7074 } 7075 AddSlowPath(slow_path); 7076 7077 // if (rb_state == ReadBarrier::GrayState()) 7078 // ref = ReadBarrier::Mark(ref); 7079 // Given the numeric representation, it's enough to check the low bit of the 7080 // rb_state. We do that by shifting the bit out of the lock word with LSRS 7081 // which can be a 16-bit instruction unlike the TST immediate. 7082 static_assert(ReadBarrier::WhiteState() == 0, "Expecting white to have value 0"); 7083 static_assert(ReadBarrier::GrayState() == 1, "Expecting gray to have value 1"); 7084 __ Lsrs(temp_reg, temp_reg, LockWord::kReadBarrierStateShift + 1); 7085 __ b(slow_path->GetEntryLabel(), CS); // Carry flag is the last bit shifted out by LSRS. 7086 __ Bind(slow_path->GetExitLabel()); 7087} 7088 7089void CodeGeneratorARM::GenerateReadBarrierSlow(HInstruction* instruction, 7090 Location out, 7091 Location ref, 7092 Location obj, 7093 uint32_t offset, 7094 Location index) { 7095 DCHECK(kEmitCompilerReadBarrier); 7096 7097 // Insert a slow path based read barrier *after* the reference load. 7098 // 7099 // If heap poisoning is enabled, the unpoisoning of the loaded 7100 // reference will be carried out by the runtime within the slow 7101 // path. 7102 // 7103 // Note that `ref` currently does not get unpoisoned (when heap 7104 // poisoning is enabled), which is alright as the `ref` argument is 7105 // not used by the artReadBarrierSlow entry point. 7106 // 7107 // TODO: Unpoison `ref` when it is used by artReadBarrierSlow. 7108 SlowPathCodeARM* slow_path = new (GetGraph()->GetArena()) 7109 ReadBarrierForHeapReferenceSlowPathARM(instruction, out, ref, obj, offset, index); 7110 AddSlowPath(slow_path); 7111 7112 __ b(slow_path->GetEntryLabel()); 7113 __ Bind(slow_path->GetExitLabel()); 7114} 7115 7116void CodeGeneratorARM::MaybeGenerateReadBarrierSlow(HInstruction* instruction, 7117 Location out, 7118 Location ref, 7119 Location obj, 7120 uint32_t offset, 7121 Location index) { 7122 if (kEmitCompilerReadBarrier) { 7123 // Baker's read barriers shall be handled by the fast path 7124 // (CodeGeneratorARM::GenerateReferenceLoadWithBakerReadBarrier). 7125 DCHECK(!kUseBakerReadBarrier); 7126 // If heap poisoning is enabled, unpoisoning will be taken care of 7127 // by the runtime within the slow path. 7128 GenerateReadBarrierSlow(instruction, out, ref, obj, offset, index); 7129 } else if (kPoisonHeapReferences) { 7130 __ UnpoisonHeapReference(out.AsRegister<Register>()); 7131 } 7132} 7133 7134void CodeGeneratorARM::GenerateReadBarrierForRootSlow(HInstruction* instruction, 7135 Location out, 7136 Location root) { 7137 DCHECK(kEmitCompilerReadBarrier); 7138 7139 // Insert a slow path based read barrier *after* the GC root load. 7140 // 7141 // Note that GC roots are not affected by heap poisoning, so we do 7142 // not need to do anything special for this here. 7143 SlowPathCodeARM* slow_path = 7144 new (GetGraph()->GetArena()) ReadBarrierForRootSlowPathARM(instruction, out, root); 7145 AddSlowPath(slow_path); 7146 7147 __ b(slow_path->GetEntryLabel()); 7148 __ Bind(slow_path->GetExitLabel()); 7149} 7150 7151HInvokeStaticOrDirect::DispatchInfo CodeGeneratorARM::GetSupportedInvokeStaticOrDirectDispatch( 7152 const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info, 7153 HInvokeStaticOrDirect* invoke ATTRIBUTE_UNUSED) { 7154 return desired_dispatch_info; 7155} 7156 7157Register CodeGeneratorARM::GetInvokeStaticOrDirectExtraParameter(HInvokeStaticOrDirect* invoke, 7158 Register temp) { 7159 DCHECK_EQ(invoke->InputCount(), invoke->GetNumberOfArguments() + 1u); 7160 Location location = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 7161 if (!invoke->GetLocations()->Intrinsified()) { 7162 return location.AsRegister<Register>(); 7163 } 7164 // For intrinsics we allow any location, so it may be on the stack. 7165 if (!location.IsRegister()) { 7166 __ LoadFromOffset(kLoadWord, temp, SP, location.GetStackIndex()); 7167 return temp; 7168 } 7169 // For register locations, check if the register was saved. If so, get it from the stack. 7170 // Note: There is a chance that the register was saved but not overwritten, so we could 7171 // save one load. However, since this is just an intrinsic slow path we prefer this 7172 // simple and more robust approach rather that trying to determine if that's the case. 7173 SlowPathCode* slow_path = GetCurrentSlowPath(); 7174 DCHECK(slow_path != nullptr); // For intrinsified invokes the call is emitted on the slow path. 7175 if (slow_path->IsCoreRegisterSaved(location.AsRegister<Register>())) { 7176 int stack_offset = slow_path->GetStackOffsetOfCoreRegister(location.AsRegister<Register>()); 7177 __ LoadFromOffset(kLoadWord, temp, SP, stack_offset); 7178 return temp; 7179 } 7180 return location.AsRegister<Register>(); 7181} 7182 7183void CodeGeneratorARM::GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) { 7184 Location callee_method = temp; // For all kinds except kRecursive, callee will be in temp. 7185 switch (invoke->GetMethodLoadKind()) { 7186 case HInvokeStaticOrDirect::MethodLoadKind::kStringInit: { 7187 uint32_t offset = 7188 GetThreadOffset<kArmPointerSize>(invoke->GetStringInitEntryPoint()).Int32Value(); 7189 // temp = thread->string_init_entrypoint 7190 __ LoadFromOffset(kLoadWord, temp.AsRegister<Register>(), TR, offset); 7191 break; 7192 } 7193 case HInvokeStaticOrDirect::MethodLoadKind::kRecursive: 7194 callee_method = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 7195 break; 7196 case HInvokeStaticOrDirect::MethodLoadKind::kDirectAddress: 7197 __ LoadImmediate(temp.AsRegister<Register>(), invoke->GetMethodAddress()); 7198 break; 7199 case HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative: { 7200 HArmDexCacheArraysBase* base = 7201 invoke->InputAt(invoke->GetSpecialInputIndex())->AsArmDexCacheArraysBase(); 7202 Register base_reg = GetInvokeStaticOrDirectExtraParameter(invoke, 7203 temp.AsRegister<Register>()); 7204 int32_t offset = invoke->GetDexCacheArrayOffset() - base->GetElementOffset(); 7205 __ LoadFromOffset(kLoadWord, temp.AsRegister<Register>(), base_reg, offset); 7206 break; 7207 } 7208 case HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod: { 7209 Location current_method = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 7210 Register method_reg; 7211 Register reg = temp.AsRegister<Register>(); 7212 if (current_method.IsRegister()) { 7213 method_reg = current_method.AsRegister<Register>(); 7214 } else { 7215 DCHECK(invoke->GetLocations()->Intrinsified()); 7216 DCHECK(!current_method.IsValid()); 7217 method_reg = reg; 7218 __ LoadFromOffset(kLoadWord, reg, SP, kCurrentMethodStackOffset); 7219 } 7220 // /* ArtMethod*[] */ temp = temp.ptr_sized_fields_->dex_cache_resolved_methods_; 7221 __ LoadFromOffset(kLoadWord, 7222 reg, 7223 method_reg, 7224 ArtMethod::DexCacheResolvedMethodsOffset(kArmPointerSize).Int32Value()); 7225 // temp = temp[index_in_cache]; 7226 // Note: Don't use invoke->GetTargetMethod() as it may point to a different dex file. 7227 uint32_t index_in_cache = invoke->GetDexMethodIndex(); 7228 __ LoadFromOffset(kLoadWord, reg, reg, CodeGenerator::GetCachePointerOffset(index_in_cache)); 7229 break; 7230 } 7231 } 7232 7233 switch (invoke->GetCodePtrLocation()) { 7234 case HInvokeStaticOrDirect::CodePtrLocation::kCallSelf: 7235 __ bl(GetFrameEntryLabel()); 7236 break; 7237 case HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod: 7238 // LR = callee_method->entry_point_from_quick_compiled_code_ 7239 __ LoadFromOffset( 7240 kLoadWord, LR, callee_method.AsRegister<Register>(), 7241 ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArmPointerSize).Int32Value()); 7242 // LR() 7243 __ blx(LR); 7244 break; 7245 } 7246 7247 DCHECK(!IsLeafMethod()); 7248} 7249 7250void CodeGeneratorARM::GenerateVirtualCall(HInvokeVirtual* invoke, Location temp_location) { 7251 Register temp = temp_location.AsRegister<Register>(); 7252 uint32_t method_offset = mirror::Class::EmbeddedVTableEntryOffset( 7253 invoke->GetVTableIndex(), kArmPointerSize).Uint32Value(); 7254 7255 // Use the calling convention instead of the location of the receiver, as 7256 // intrinsics may have put the receiver in a different register. In the intrinsics 7257 // slow path, the arguments have been moved to the right place, so here we are 7258 // guaranteed that the receiver is the first register of the calling convention. 7259 InvokeDexCallingConvention calling_convention; 7260 Register receiver = calling_convention.GetRegisterAt(0); 7261 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 7262 // /* HeapReference<Class> */ temp = receiver->klass_ 7263 __ LoadFromOffset(kLoadWord, temp, receiver, class_offset); 7264 MaybeRecordImplicitNullCheck(invoke); 7265 // Instead of simply (possibly) unpoisoning `temp` here, we should 7266 // emit a read barrier for the previous class reference load. 7267 // However this is not required in practice, as this is an 7268 // intermediate/temporary reference and because the current 7269 // concurrent copying collector keeps the from-space memory 7270 // intact/accessible until the end of the marking phase (the 7271 // concurrent copying collector may not in the future). 7272 __ MaybeUnpoisonHeapReference(temp); 7273 // temp = temp->GetMethodAt(method_offset); 7274 uint32_t entry_point = ArtMethod::EntryPointFromQuickCompiledCodeOffset( 7275 kArmPointerSize).Int32Value(); 7276 __ LoadFromOffset(kLoadWord, temp, temp, method_offset); 7277 // LR = temp->GetEntryPoint(); 7278 __ LoadFromOffset(kLoadWord, LR, temp, entry_point); 7279 // LR(); 7280 __ blx(LR); 7281} 7282 7283CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewPcRelativeStringPatch( 7284 const DexFile& dex_file, dex::StringIndex string_index) { 7285 return NewPcRelativePatch(dex_file, string_index.index_, &pc_relative_string_patches_); 7286} 7287 7288CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewPcRelativeTypePatch( 7289 const DexFile& dex_file, dex::TypeIndex type_index) { 7290 return NewPcRelativePatch(dex_file, type_index.index_, &pc_relative_type_patches_); 7291} 7292 7293CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewTypeBssEntryPatch( 7294 const DexFile& dex_file, dex::TypeIndex type_index) { 7295 return NewPcRelativePatch(dex_file, type_index.index_, &type_bss_entry_patches_); 7296} 7297 7298CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewPcRelativeDexCacheArrayPatch( 7299 const DexFile& dex_file, uint32_t element_offset) { 7300 return NewPcRelativePatch(dex_file, element_offset, &pc_relative_dex_cache_patches_); 7301} 7302 7303CodeGeneratorARM::PcRelativePatchInfo* CodeGeneratorARM::NewPcRelativePatch( 7304 const DexFile& dex_file, uint32_t offset_or_index, ArenaDeque<PcRelativePatchInfo>* patches) { 7305 patches->emplace_back(dex_file, offset_or_index); 7306 return &patches->back(); 7307} 7308 7309Literal* CodeGeneratorARM::DeduplicateBootImageStringLiteral(const DexFile& dex_file, 7310 dex::StringIndex string_index) { 7311 return boot_image_string_patches_.GetOrCreate( 7312 StringReference(&dex_file, string_index), 7313 [this]() { return __ NewLiteral<uint32_t>(/* placeholder */ 0u); }); 7314} 7315 7316Literal* CodeGeneratorARM::DeduplicateBootImageTypeLiteral(const DexFile& dex_file, 7317 dex::TypeIndex type_index) { 7318 return boot_image_type_patches_.GetOrCreate( 7319 TypeReference(&dex_file, type_index), 7320 [this]() { return __ NewLiteral<uint32_t>(/* placeholder */ 0u); }); 7321} 7322 7323Literal* CodeGeneratorARM::DeduplicateBootImageAddressLiteral(uint32_t address) { 7324 bool needs_patch = GetCompilerOptions().GetIncludePatchInformation(); 7325 Uint32ToLiteralMap* map = needs_patch ? &boot_image_address_patches_ : &uint32_literals_; 7326 return DeduplicateUint32Literal(dchecked_integral_cast<uint32_t>(address), map); 7327} 7328 7329Literal* CodeGeneratorARM::DeduplicateJitStringLiteral(const DexFile& dex_file, 7330 dex::StringIndex string_index, 7331 Handle<mirror::String> handle) { 7332 jit_string_roots_.Overwrite(StringReference(&dex_file, string_index), 7333 reinterpret_cast64<uint64_t>(handle.GetReference())); 7334 return jit_string_patches_.GetOrCreate( 7335 StringReference(&dex_file, string_index), 7336 [this]() { return __ NewLiteral<uint32_t>(/* placeholder */ 0u); }); 7337} 7338 7339Literal* CodeGeneratorARM::DeduplicateJitClassLiteral(const DexFile& dex_file, 7340 dex::TypeIndex type_index, 7341 Handle<mirror::Class> handle) { 7342 jit_class_roots_.Overwrite(TypeReference(&dex_file, type_index), 7343 reinterpret_cast64<uint64_t>(handle.GetReference())); 7344 return jit_class_patches_.GetOrCreate( 7345 TypeReference(&dex_file, type_index), 7346 [this]() { return __ NewLiteral<uint32_t>(/* placeholder */ 0u); }); 7347} 7348 7349template <LinkerPatch (*Factory)(size_t, const DexFile*, uint32_t, uint32_t)> 7350inline void CodeGeneratorARM::EmitPcRelativeLinkerPatches( 7351 const ArenaDeque<PcRelativePatchInfo>& infos, 7352 ArenaVector<LinkerPatch>* linker_patches) { 7353 for (const PcRelativePatchInfo& info : infos) { 7354 const DexFile& dex_file = info.target_dex_file; 7355 size_t offset_or_index = info.offset_or_index; 7356 DCHECK(info.add_pc_label.IsBound()); 7357 uint32_t add_pc_offset = dchecked_integral_cast<uint32_t>(info.add_pc_label.Position()); 7358 // Add MOVW patch. 7359 DCHECK(info.movw_label.IsBound()); 7360 uint32_t movw_offset = dchecked_integral_cast<uint32_t>(info.movw_label.Position()); 7361 linker_patches->push_back(Factory(movw_offset, &dex_file, add_pc_offset, offset_or_index)); 7362 // Add MOVT patch. 7363 DCHECK(info.movt_label.IsBound()); 7364 uint32_t movt_offset = dchecked_integral_cast<uint32_t>(info.movt_label.Position()); 7365 linker_patches->push_back(Factory(movt_offset, &dex_file, add_pc_offset, offset_or_index)); 7366 } 7367} 7368 7369void CodeGeneratorARM::EmitLinkerPatches(ArenaVector<LinkerPatch>* linker_patches) { 7370 DCHECK(linker_patches->empty()); 7371 size_t size = 7372 /* MOVW+MOVT for each entry */ 2u * pc_relative_dex_cache_patches_.size() + 7373 boot_image_string_patches_.size() + 7374 /* MOVW+MOVT for each entry */ 2u * pc_relative_string_patches_.size() + 7375 boot_image_type_patches_.size() + 7376 /* MOVW+MOVT for each entry */ 2u * pc_relative_type_patches_.size() + 7377 /* MOVW+MOVT for each entry */ 2u * type_bss_entry_patches_.size() + 7378 boot_image_address_patches_.size(); 7379 linker_patches->reserve(size); 7380 EmitPcRelativeLinkerPatches<LinkerPatch::DexCacheArrayPatch>(pc_relative_dex_cache_patches_, 7381 linker_patches); 7382 for (const auto& entry : boot_image_string_patches_) { 7383 const StringReference& target_string = entry.first; 7384 Literal* literal = entry.second; 7385 DCHECK(literal->GetLabel()->IsBound()); 7386 uint32_t literal_offset = literal->GetLabel()->Position(); 7387 linker_patches->push_back(LinkerPatch::StringPatch(literal_offset, 7388 target_string.dex_file, 7389 target_string.string_index.index_)); 7390 } 7391 if (!GetCompilerOptions().IsBootImage()) { 7392 DCHECK(pc_relative_type_patches_.empty()); 7393 EmitPcRelativeLinkerPatches<LinkerPatch::StringBssEntryPatch>(pc_relative_string_patches_, 7394 linker_patches); 7395 } else { 7396 EmitPcRelativeLinkerPatches<LinkerPatch::RelativeTypePatch>(pc_relative_type_patches_, 7397 linker_patches); 7398 EmitPcRelativeLinkerPatches<LinkerPatch::RelativeStringPatch>(pc_relative_string_patches_, 7399 linker_patches); 7400 } 7401 EmitPcRelativeLinkerPatches<LinkerPatch::TypeBssEntryPatch>(type_bss_entry_patches_, 7402 linker_patches); 7403 for (const auto& entry : boot_image_type_patches_) { 7404 const TypeReference& target_type = entry.first; 7405 Literal* literal = entry.second; 7406 DCHECK(literal->GetLabel()->IsBound()); 7407 uint32_t literal_offset = literal->GetLabel()->Position(); 7408 linker_patches->push_back(LinkerPatch::TypePatch(literal_offset, 7409 target_type.dex_file, 7410 target_type.type_index.index_)); 7411 } 7412 for (const auto& entry : boot_image_address_patches_) { 7413 DCHECK(GetCompilerOptions().GetIncludePatchInformation()); 7414 Literal* literal = entry.second; 7415 DCHECK(literal->GetLabel()->IsBound()); 7416 uint32_t literal_offset = literal->GetLabel()->Position(); 7417 linker_patches->push_back(LinkerPatch::RecordPosition(literal_offset)); 7418 } 7419 DCHECK_EQ(size, linker_patches->size()); 7420} 7421 7422Literal* CodeGeneratorARM::DeduplicateUint32Literal(uint32_t value, Uint32ToLiteralMap* map) { 7423 return map->GetOrCreate( 7424 value, 7425 [this, value]() { return __ NewLiteral<uint32_t>(value); }); 7426} 7427 7428Literal* CodeGeneratorARM::DeduplicateMethodLiteral(MethodReference target_method, 7429 MethodToLiteralMap* map) { 7430 return map->GetOrCreate( 7431 target_method, 7432 [this]() { return __ NewLiteral<uint32_t>(/* placeholder */ 0u); }); 7433} 7434 7435void LocationsBuilderARM::VisitMultiplyAccumulate(HMultiplyAccumulate* instr) { 7436 LocationSummary* locations = 7437 new (GetGraph()->GetArena()) LocationSummary(instr, LocationSummary::kNoCall); 7438 locations->SetInAt(HMultiplyAccumulate::kInputAccumulatorIndex, 7439 Location::RequiresRegister()); 7440 locations->SetInAt(HMultiplyAccumulate::kInputMulLeftIndex, Location::RequiresRegister()); 7441 locations->SetInAt(HMultiplyAccumulate::kInputMulRightIndex, Location::RequiresRegister()); 7442 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 7443} 7444 7445void InstructionCodeGeneratorARM::VisitMultiplyAccumulate(HMultiplyAccumulate* instr) { 7446 LocationSummary* locations = instr->GetLocations(); 7447 Register res = locations->Out().AsRegister<Register>(); 7448 Register accumulator = 7449 locations->InAt(HMultiplyAccumulate::kInputAccumulatorIndex).AsRegister<Register>(); 7450 Register mul_left = 7451 locations->InAt(HMultiplyAccumulate::kInputMulLeftIndex).AsRegister<Register>(); 7452 Register mul_right = 7453 locations->InAt(HMultiplyAccumulate::kInputMulRightIndex).AsRegister<Register>(); 7454 7455 if (instr->GetOpKind() == HInstruction::kAdd) { 7456 __ mla(res, mul_left, mul_right, accumulator); 7457 } else { 7458 __ mls(res, mul_left, mul_right, accumulator); 7459 } 7460} 7461 7462void LocationsBuilderARM::VisitBoundType(HBoundType* instruction ATTRIBUTE_UNUSED) { 7463 // Nothing to do, this should be removed during prepare for register allocator. 7464 LOG(FATAL) << "Unreachable"; 7465} 7466 7467void InstructionCodeGeneratorARM::VisitBoundType(HBoundType* instruction ATTRIBUTE_UNUSED) { 7468 // Nothing to do, this should be removed during prepare for register allocator. 7469 LOG(FATAL) << "Unreachable"; 7470} 7471 7472// Simple implementation of packed switch - generate cascaded compare/jumps. 7473void LocationsBuilderARM::VisitPackedSwitch(HPackedSwitch* switch_instr) { 7474 LocationSummary* locations = 7475 new (GetGraph()->GetArena()) LocationSummary(switch_instr, LocationSummary::kNoCall); 7476 locations->SetInAt(0, Location::RequiresRegister()); 7477 if (switch_instr->GetNumEntries() > kPackedSwitchCompareJumpThreshold && 7478 codegen_->GetAssembler()->IsThumb()) { 7479 locations->AddTemp(Location::RequiresRegister()); // We need a temp for the table base. 7480 if (switch_instr->GetStartValue() != 0) { 7481 locations->AddTemp(Location::RequiresRegister()); // We need a temp for the bias. 7482 } 7483 } 7484} 7485 7486void InstructionCodeGeneratorARM::VisitPackedSwitch(HPackedSwitch* switch_instr) { 7487 int32_t lower_bound = switch_instr->GetStartValue(); 7488 uint32_t num_entries = switch_instr->GetNumEntries(); 7489 LocationSummary* locations = switch_instr->GetLocations(); 7490 Register value_reg = locations->InAt(0).AsRegister<Register>(); 7491 HBasicBlock* default_block = switch_instr->GetDefaultBlock(); 7492 7493 if (num_entries <= kPackedSwitchCompareJumpThreshold || !codegen_->GetAssembler()->IsThumb()) { 7494 // Create a series of compare/jumps. 7495 Register temp_reg = IP; 7496 // Note: It is fine for the below AddConstantSetFlags() using IP register to temporarily store 7497 // the immediate, because IP is used as the destination register. For the other 7498 // AddConstantSetFlags() and GenerateCompareWithImmediate(), the immediate values are constant, 7499 // and they can be encoded in the instruction without making use of IP register. 7500 __ AddConstantSetFlags(temp_reg, value_reg, -lower_bound); 7501 7502 const ArenaVector<HBasicBlock*>& successors = switch_instr->GetBlock()->GetSuccessors(); 7503 // Jump to successors[0] if value == lower_bound. 7504 __ b(codegen_->GetLabelOf(successors[0]), EQ); 7505 int32_t last_index = 0; 7506 for (; num_entries - last_index > 2; last_index += 2) { 7507 __ AddConstantSetFlags(temp_reg, temp_reg, -2); 7508 // Jump to successors[last_index + 1] if value < case_value[last_index + 2]. 7509 __ b(codegen_->GetLabelOf(successors[last_index + 1]), LO); 7510 // Jump to successors[last_index + 2] if value == case_value[last_index + 2]. 7511 __ b(codegen_->GetLabelOf(successors[last_index + 2]), EQ); 7512 } 7513 if (num_entries - last_index == 2) { 7514 // The last missing case_value. 7515 __ CmpConstant(temp_reg, 1); 7516 __ b(codegen_->GetLabelOf(successors[last_index + 1]), EQ); 7517 } 7518 7519 // And the default for any other value. 7520 if (!codegen_->GoesToNextBlock(switch_instr->GetBlock(), default_block)) { 7521 __ b(codegen_->GetLabelOf(default_block)); 7522 } 7523 } else { 7524 // Create a table lookup. 7525 Register temp_reg = locations->GetTemp(0).AsRegister<Register>(); 7526 7527 // Materialize a pointer to the switch table 7528 std::vector<Label*> labels(num_entries); 7529 const ArenaVector<HBasicBlock*>& successors = switch_instr->GetBlock()->GetSuccessors(); 7530 for (uint32_t i = 0; i < num_entries; i++) { 7531 labels[i] = codegen_->GetLabelOf(successors[i]); 7532 } 7533 JumpTable* table = __ CreateJumpTable(std::move(labels), temp_reg); 7534 7535 // Remove the bias. 7536 Register key_reg; 7537 if (lower_bound != 0) { 7538 key_reg = locations->GetTemp(1).AsRegister<Register>(); 7539 __ AddConstant(key_reg, value_reg, -lower_bound); 7540 } else { 7541 key_reg = value_reg; 7542 } 7543 7544 // Check whether the value is in the table, jump to default block if not. 7545 __ CmpConstant(key_reg, num_entries - 1); 7546 __ b(codegen_->GetLabelOf(default_block), Condition::HI); 7547 7548 // Load the displacement from the table. 7549 __ ldr(temp_reg, Address(temp_reg, key_reg, Shift::LSL, 2)); 7550 7551 // Dispatch is a direct add to the PC (for Thumb2). 7552 __ EmitJumpTableDispatch(table, temp_reg); 7553 } 7554} 7555 7556void LocationsBuilderARM::VisitArmDexCacheArraysBase(HArmDexCacheArraysBase* base) { 7557 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(base); 7558 locations->SetOut(Location::RequiresRegister()); 7559} 7560 7561void InstructionCodeGeneratorARM::VisitArmDexCacheArraysBase(HArmDexCacheArraysBase* base) { 7562 Register base_reg = base->GetLocations()->Out().AsRegister<Register>(); 7563 CodeGeneratorARM::PcRelativePatchInfo* labels = 7564 codegen_->NewPcRelativeDexCacheArrayPatch(base->GetDexFile(), base->GetElementOffset()); 7565 __ BindTrackedLabel(&labels->movw_label); 7566 __ movw(base_reg, /* placeholder */ 0u); 7567 __ BindTrackedLabel(&labels->movt_label); 7568 __ movt(base_reg, /* placeholder */ 0u); 7569 __ BindTrackedLabel(&labels->add_pc_label); 7570 __ add(base_reg, base_reg, ShifterOperand(PC)); 7571} 7572 7573void CodeGeneratorARM::MoveFromReturnRegister(Location trg, Primitive::Type type) { 7574 if (!trg.IsValid()) { 7575 DCHECK_EQ(type, Primitive::kPrimVoid); 7576 return; 7577 } 7578 7579 DCHECK_NE(type, Primitive::kPrimVoid); 7580 7581 Location return_loc = InvokeDexCallingConventionVisitorARM().GetReturnLocation(type); 7582 if (return_loc.Equals(trg)) { 7583 return; 7584 } 7585 7586 // TODO: Consider pairs in the parallel move resolver, then this could be nicely merged 7587 // with the last branch. 7588 if (type == Primitive::kPrimLong) { 7589 HParallelMove parallel_move(GetGraph()->GetArena()); 7590 parallel_move.AddMove(return_loc.ToLow(), trg.ToLow(), Primitive::kPrimInt, nullptr); 7591 parallel_move.AddMove(return_loc.ToHigh(), trg.ToHigh(), Primitive::kPrimInt, nullptr); 7592 GetMoveResolver()->EmitNativeCode(¶llel_move); 7593 } else if (type == Primitive::kPrimDouble) { 7594 HParallelMove parallel_move(GetGraph()->GetArena()); 7595 parallel_move.AddMove(return_loc.ToLow(), trg.ToLow(), Primitive::kPrimFloat, nullptr); 7596 parallel_move.AddMove(return_loc.ToHigh(), trg.ToHigh(), Primitive::kPrimFloat, nullptr); 7597 GetMoveResolver()->EmitNativeCode(¶llel_move); 7598 } else { 7599 // Let the parallel move resolver take care of all of this. 7600 HParallelMove parallel_move(GetGraph()->GetArena()); 7601 parallel_move.AddMove(return_loc, trg, type, nullptr); 7602 GetMoveResolver()->EmitNativeCode(¶llel_move); 7603 } 7604} 7605 7606void LocationsBuilderARM::VisitClassTableGet(HClassTableGet* instruction) { 7607 LocationSummary* locations = 7608 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 7609 locations->SetInAt(0, Location::RequiresRegister()); 7610 locations->SetOut(Location::RequiresRegister()); 7611} 7612 7613void InstructionCodeGeneratorARM::VisitClassTableGet(HClassTableGet* instruction) { 7614 LocationSummary* locations = instruction->GetLocations(); 7615 if (instruction->GetTableKind() == HClassTableGet::TableKind::kVTable) { 7616 uint32_t method_offset = mirror::Class::EmbeddedVTableEntryOffset( 7617 instruction->GetIndex(), kArmPointerSize).SizeValue(); 7618 __ LoadFromOffset(kLoadWord, 7619 locations->Out().AsRegister<Register>(), 7620 locations->InAt(0).AsRegister<Register>(), 7621 method_offset); 7622 } else { 7623 uint32_t method_offset = static_cast<uint32_t>(ImTable::OffsetOfElement( 7624 instruction->GetIndex(), kArmPointerSize)); 7625 __ LoadFromOffset(kLoadWord, 7626 locations->Out().AsRegister<Register>(), 7627 locations->InAt(0).AsRegister<Register>(), 7628 mirror::Class::ImtPtrOffset(kArmPointerSize).Uint32Value()); 7629 __ LoadFromOffset(kLoadWord, 7630 locations->Out().AsRegister<Register>(), 7631 locations->Out().AsRegister<Register>(), 7632 method_offset); 7633 } 7634} 7635 7636static void PatchJitRootUse(uint8_t* code, 7637 const uint8_t* roots_data, 7638 Literal* literal, 7639 uint64_t index_in_table) { 7640 DCHECK(literal->GetLabel()->IsBound()); 7641 uint32_t literal_offset = literal->GetLabel()->Position(); 7642 uintptr_t address = 7643 reinterpret_cast<uintptr_t>(roots_data) + index_in_table * sizeof(GcRoot<mirror::Object>); 7644 uint8_t* data = code + literal_offset; 7645 reinterpret_cast<uint32_t*>(data)[0] = dchecked_integral_cast<uint32_t>(address); 7646} 7647 7648void CodeGeneratorARM::EmitJitRootPatches(uint8_t* code, const uint8_t* roots_data) { 7649 for (const auto& entry : jit_string_patches_) { 7650 const auto& it = jit_string_roots_.find(entry.first); 7651 DCHECK(it != jit_string_roots_.end()); 7652 PatchJitRootUse(code, roots_data, entry.second, it->second); 7653 } 7654 for (const auto& entry : jit_class_patches_) { 7655 const auto& it = jit_class_roots_.find(entry.first); 7656 DCHECK(it != jit_class_roots_.end()); 7657 PatchJitRootUse(code, roots_data, entry.second, it->second); 7658 } 7659} 7660 7661#undef __ 7662#undef QUICK_ENTRY_POINT 7663 7664} // namespace arm 7665} // namespace art 7666