code_generator_arm64.cc revision 42249c3602c3d0243396ee3627ffb5906aa77c1e
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_arm64.h" 18 19#include "arch/arm64/instruction_set_features_arm64.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 "entrypoints/quick/quick_entrypoints_enum.h" 25#include "gc/accounting/card_table.h" 26#include "intrinsics.h" 27#include "intrinsics_arm64.h" 28#include "mirror/array-inl.h" 29#include "mirror/class-inl.h" 30#include "offsets.h" 31#include "thread.h" 32#include "utils/arm64/assembler_arm64.h" 33#include "utils/assembler.h" 34#include "utils/stack_checks.h" 35 36 37using namespace vixl; // NOLINT(build/namespaces) 38 39#ifdef __ 40#error "ARM64 Codegen VIXL macro-assembler macro already defined." 41#endif 42 43namespace art { 44 45template<class MirrorType> 46class GcRoot; 47 48namespace arm64 { 49 50using helpers::CPURegisterFrom; 51using helpers::DRegisterFrom; 52using helpers::FPRegisterFrom; 53using helpers::HeapOperand; 54using helpers::HeapOperandFrom; 55using helpers::InputCPURegisterAt; 56using helpers::InputFPRegisterAt; 57using helpers::InputRegisterAt; 58using helpers::InputOperandAt; 59using helpers::Int64ConstantFrom; 60using helpers::LocationFrom; 61using helpers::OperandFromMemOperand; 62using helpers::OutputCPURegister; 63using helpers::OutputFPRegister; 64using helpers::OutputRegister; 65using helpers::RegisterFrom; 66using helpers::StackOperandFrom; 67using helpers::VIXLRegCodeFromART; 68using helpers::WRegisterFrom; 69using helpers::XRegisterFrom; 70using helpers::ARM64EncodableConstantOrRegister; 71using helpers::ArtVixlRegCodeCoherentForRegSet; 72 73static constexpr int kCurrentMethodStackOffset = 0; 74// The compare/jump sequence will generate about (1.5 * num_entries + 3) instructions. While jump 75// table version generates 7 instructions and num_entries literals. Compare/jump sequence will 76// generates less code/data with a small num_entries. 77static constexpr uint32_t kPackedSwitchCompareJumpThreshold = 7; 78 79inline Condition ARM64Condition(IfCondition cond) { 80 switch (cond) { 81 case kCondEQ: return eq; 82 case kCondNE: return ne; 83 case kCondLT: return lt; 84 case kCondLE: return le; 85 case kCondGT: return gt; 86 case kCondGE: return ge; 87 case kCondB: return lo; 88 case kCondBE: return ls; 89 case kCondA: return hi; 90 case kCondAE: return hs; 91 } 92 LOG(FATAL) << "Unreachable"; 93 UNREACHABLE(); 94} 95 96Location ARM64ReturnLocation(Primitive::Type return_type) { 97 // Note that in practice, `LocationFrom(x0)` and `LocationFrom(w0)` create the 98 // same Location object, and so do `LocationFrom(d0)` and `LocationFrom(s0)`, 99 // but we use the exact registers for clarity. 100 if (return_type == Primitive::kPrimFloat) { 101 return LocationFrom(s0); 102 } else if (return_type == Primitive::kPrimDouble) { 103 return LocationFrom(d0); 104 } else if (return_type == Primitive::kPrimLong) { 105 return LocationFrom(x0); 106 } else if (return_type == Primitive::kPrimVoid) { 107 return Location::NoLocation(); 108 } else { 109 return LocationFrom(w0); 110 } 111} 112 113Location InvokeRuntimeCallingConvention::GetReturnLocation(Primitive::Type return_type) { 114 return ARM64ReturnLocation(return_type); 115} 116 117#define __ down_cast<CodeGeneratorARM64*>(codegen)->GetVIXLAssembler()-> 118#define QUICK_ENTRY_POINT(x) QUICK_ENTRYPOINT_OFFSET(kArm64WordSize, x).Int32Value() 119 120// Calculate memory accessing operand for save/restore live registers. 121static void SaveRestoreLiveRegistersHelper(CodeGenerator* codegen, 122 RegisterSet* register_set, 123 int64_t spill_offset, 124 bool is_save) { 125 DCHECK(ArtVixlRegCodeCoherentForRegSet(register_set->GetCoreRegisters(), 126 codegen->GetNumberOfCoreRegisters(), 127 register_set->GetFloatingPointRegisters(), 128 codegen->GetNumberOfFloatingPointRegisters())); 129 130 CPURegList core_list = CPURegList(CPURegister::kRegister, kXRegSize, 131 register_set->GetCoreRegisters() & (~callee_saved_core_registers.list())); 132 CPURegList fp_list = CPURegList(CPURegister::kFPRegister, kDRegSize, 133 register_set->GetFloatingPointRegisters() & (~callee_saved_fp_registers.list())); 134 135 MacroAssembler* masm = down_cast<CodeGeneratorARM64*>(codegen)->GetVIXLAssembler(); 136 UseScratchRegisterScope temps(masm); 137 138 Register base = masm->StackPointer(); 139 int64_t core_spill_size = core_list.TotalSizeInBytes(); 140 int64_t fp_spill_size = fp_list.TotalSizeInBytes(); 141 int64_t reg_size = kXRegSizeInBytes; 142 int64_t max_ls_pair_offset = spill_offset + core_spill_size + fp_spill_size - 2 * reg_size; 143 uint32_t ls_access_size = WhichPowerOf2(reg_size); 144 if (((core_list.Count() > 1) || (fp_list.Count() > 1)) && 145 !masm->IsImmLSPair(max_ls_pair_offset, ls_access_size)) { 146 // If the offset does not fit in the instruction's immediate field, use an alternate register 147 // to compute the base address(float point registers spill base address). 148 Register new_base = temps.AcquireSameSizeAs(base); 149 __ Add(new_base, base, Operand(spill_offset + core_spill_size)); 150 base = new_base; 151 spill_offset = -core_spill_size; 152 int64_t new_max_ls_pair_offset = fp_spill_size - 2 * reg_size; 153 DCHECK(masm->IsImmLSPair(spill_offset, ls_access_size)); 154 DCHECK(masm->IsImmLSPair(new_max_ls_pair_offset, ls_access_size)); 155 } 156 157 if (is_save) { 158 __ StoreCPURegList(core_list, MemOperand(base, spill_offset)); 159 __ StoreCPURegList(fp_list, MemOperand(base, spill_offset + core_spill_size)); 160 } else { 161 __ LoadCPURegList(core_list, MemOperand(base, spill_offset)); 162 __ LoadCPURegList(fp_list, MemOperand(base, spill_offset + core_spill_size)); 163 } 164} 165 166void SlowPathCodeARM64::SaveLiveRegisters(CodeGenerator* codegen, LocationSummary* locations) { 167 RegisterSet* register_set = locations->GetLiveRegisters(); 168 size_t stack_offset = codegen->GetFirstRegisterSlotInSlowPath(); 169 for (size_t i = 0, e = codegen->GetNumberOfCoreRegisters(); i < e; ++i) { 170 if (!codegen->IsCoreCalleeSaveRegister(i) && register_set->ContainsCoreRegister(i)) { 171 // If the register holds an object, update the stack mask. 172 if (locations->RegisterContainsObject(i)) { 173 locations->SetStackBit(stack_offset / kVRegSize); 174 } 175 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 176 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 177 saved_core_stack_offsets_[i] = stack_offset; 178 stack_offset += kXRegSizeInBytes; 179 } 180 } 181 182 for (size_t i = 0, e = codegen->GetNumberOfFloatingPointRegisters(); i < e; ++i) { 183 if (!codegen->IsFloatingPointCalleeSaveRegister(i) && 184 register_set->ContainsFloatingPointRegister(i)) { 185 DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize()); 186 DCHECK_LT(i, kMaximumNumberOfExpectedRegisters); 187 saved_fpu_stack_offsets_[i] = stack_offset; 188 stack_offset += kDRegSizeInBytes; 189 } 190 } 191 192 SaveRestoreLiveRegistersHelper(codegen, register_set, 193 codegen->GetFirstRegisterSlotInSlowPath(), true /* is_save */); 194} 195 196void SlowPathCodeARM64::RestoreLiveRegisters(CodeGenerator* codegen, LocationSummary* locations) { 197 RegisterSet* register_set = locations->GetLiveRegisters(); 198 SaveRestoreLiveRegistersHelper(codegen, register_set, 199 codegen->GetFirstRegisterSlotInSlowPath(), false /* is_save */); 200} 201 202class BoundsCheckSlowPathARM64 : public SlowPathCodeARM64 { 203 public: 204 explicit BoundsCheckSlowPathARM64(HBoundsCheck* instruction) : instruction_(instruction) {} 205 206 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 207 LocationSummary* locations = instruction_->GetLocations(); 208 CodeGeneratorARM64* arm64_codegen = down_cast<CodeGeneratorARM64*>(codegen); 209 210 __ Bind(GetEntryLabel()); 211 if (instruction_->CanThrowIntoCatchBlock()) { 212 // Live registers will be restored in the catch block if caught. 213 SaveLiveRegisters(codegen, instruction_->GetLocations()); 214 } 215 // We're moving two locations to locations that could overlap, so we need a parallel 216 // move resolver. 217 InvokeRuntimeCallingConvention calling_convention; 218 codegen->EmitParallelMoves( 219 locations->InAt(0), LocationFrom(calling_convention.GetRegisterAt(0)), Primitive::kPrimInt, 220 locations->InAt(1), LocationFrom(calling_convention.GetRegisterAt(1)), Primitive::kPrimInt); 221 arm64_codegen->InvokeRuntime( 222 QUICK_ENTRY_POINT(pThrowArrayBounds), instruction_, instruction_->GetDexPc(), this); 223 CheckEntrypointTypes<kQuickThrowArrayBounds, void, int32_t, int32_t>(); 224 } 225 226 bool IsFatal() const OVERRIDE { return true; } 227 228 const char* GetDescription() const OVERRIDE { return "BoundsCheckSlowPathARM64"; } 229 230 private: 231 HBoundsCheck* const instruction_; 232 233 DISALLOW_COPY_AND_ASSIGN(BoundsCheckSlowPathARM64); 234}; 235 236class DivZeroCheckSlowPathARM64 : public SlowPathCodeARM64 { 237 public: 238 explicit DivZeroCheckSlowPathARM64(HDivZeroCheck* instruction) : instruction_(instruction) {} 239 240 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 241 CodeGeneratorARM64* arm64_codegen = down_cast<CodeGeneratorARM64*>(codegen); 242 __ Bind(GetEntryLabel()); 243 if (instruction_->CanThrowIntoCatchBlock()) { 244 // Live registers will be restored in the catch block if caught. 245 SaveLiveRegisters(codegen, instruction_->GetLocations()); 246 } 247 arm64_codegen->InvokeRuntime( 248 QUICK_ENTRY_POINT(pThrowDivZero), instruction_, instruction_->GetDexPc(), this); 249 CheckEntrypointTypes<kQuickThrowDivZero, void, void>(); 250 } 251 252 bool IsFatal() const OVERRIDE { return true; } 253 254 const char* GetDescription() const OVERRIDE { return "DivZeroCheckSlowPathARM64"; } 255 256 private: 257 HDivZeroCheck* const instruction_; 258 DISALLOW_COPY_AND_ASSIGN(DivZeroCheckSlowPathARM64); 259}; 260 261class LoadClassSlowPathARM64 : public SlowPathCodeARM64 { 262 public: 263 LoadClassSlowPathARM64(HLoadClass* cls, 264 HInstruction* at, 265 uint32_t dex_pc, 266 bool do_clinit) 267 : cls_(cls), at_(at), dex_pc_(dex_pc), do_clinit_(do_clinit) { 268 DCHECK(at->IsLoadClass() || at->IsClinitCheck()); 269 } 270 271 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 272 LocationSummary* locations = at_->GetLocations(); 273 CodeGeneratorARM64* arm64_codegen = down_cast<CodeGeneratorARM64*>(codegen); 274 275 __ Bind(GetEntryLabel()); 276 SaveLiveRegisters(codegen, locations); 277 278 InvokeRuntimeCallingConvention calling_convention; 279 __ Mov(calling_convention.GetRegisterAt(0).W(), cls_->GetTypeIndex()); 280 int32_t entry_point_offset = do_clinit_ ? QUICK_ENTRY_POINT(pInitializeStaticStorage) 281 : QUICK_ENTRY_POINT(pInitializeType); 282 arm64_codegen->InvokeRuntime(entry_point_offset, at_, dex_pc_, this); 283 if (do_clinit_) { 284 CheckEntrypointTypes<kQuickInitializeStaticStorage, void*, uint32_t>(); 285 } else { 286 CheckEntrypointTypes<kQuickInitializeType, void*, uint32_t>(); 287 } 288 289 // Move the class to the desired location. 290 Location out = locations->Out(); 291 if (out.IsValid()) { 292 DCHECK(out.IsRegister() && !locations->GetLiveRegisters()->ContainsCoreRegister(out.reg())); 293 Primitive::Type type = at_->GetType(); 294 arm64_codegen->MoveLocation(out, calling_convention.GetReturnLocation(type), type); 295 } 296 297 RestoreLiveRegisters(codegen, locations); 298 __ B(GetExitLabel()); 299 } 300 301 const char* GetDescription() const OVERRIDE { return "LoadClassSlowPathARM64"; } 302 303 private: 304 // The class this slow path will load. 305 HLoadClass* const cls_; 306 307 // The instruction where this slow path is happening. 308 // (Might be the load class or an initialization check). 309 HInstruction* const at_; 310 311 // The dex PC of `at_`. 312 const uint32_t dex_pc_; 313 314 // Whether to initialize the class. 315 const bool do_clinit_; 316 317 DISALLOW_COPY_AND_ASSIGN(LoadClassSlowPathARM64); 318}; 319 320class LoadStringSlowPathARM64 : public SlowPathCodeARM64 { 321 public: 322 explicit LoadStringSlowPathARM64(HLoadString* instruction) : instruction_(instruction) {} 323 324 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 325 LocationSummary* locations = instruction_->GetLocations(); 326 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg())); 327 CodeGeneratorARM64* arm64_codegen = down_cast<CodeGeneratorARM64*>(codegen); 328 329 __ Bind(GetEntryLabel()); 330 SaveLiveRegisters(codegen, locations); 331 332 InvokeRuntimeCallingConvention calling_convention; 333 __ Mov(calling_convention.GetRegisterAt(0).W(), instruction_->GetStringIndex()); 334 arm64_codegen->InvokeRuntime( 335 QUICK_ENTRY_POINT(pResolveString), instruction_, instruction_->GetDexPc(), this); 336 CheckEntrypointTypes<kQuickResolveString, void*, uint32_t>(); 337 Primitive::Type type = instruction_->GetType(); 338 arm64_codegen->MoveLocation(locations->Out(), calling_convention.GetReturnLocation(type), type); 339 340 RestoreLiveRegisters(codegen, locations); 341 __ B(GetExitLabel()); 342 } 343 344 const char* GetDescription() const OVERRIDE { return "LoadStringSlowPathARM64"; } 345 346 private: 347 HLoadString* const instruction_; 348 349 DISALLOW_COPY_AND_ASSIGN(LoadStringSlowPathARM64); 350}; 351 352class NullCheckSlowPathARM64 : public SlowPathCodeARM64 { 353 public: 354 explicit NullCheckSlowPathARM64(HNullCheck* instr) : instruction_(instr) {} 355 356 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 357 CodeGeneratorARM64* arm64_codegen = down_cast<CodeGeneratorARM64*>(codegen); 358 __ Bind(GetEntryLabel()); 359 if (instruction_->CanThrowIntoCatchBlock()) { 360 // Live registers will be restored in the catch block if caught. 361 SaveLiveRegisters(codegen, instruction_->GetLocations()); 362 } 363 arm64_codegen->InvokeRuntime( 364 QUICK_ENTRY_POINT(pThrowNullPointer), instruction_, instruction_->GetDexPc(), this); 365 CheckEntrypointTypes<kQuickThrowNullPointer, void, void>(); 366 } 367 368 bool IsFatal() const OVERRIDE { return true; } 369 370 const char* GetDescription() const OVERRIDE { return "NullCheckSlowPathARM64"; } 371 372 private: 373 HNullCheck* const instruction_; 374 375 DISALLOW_COPY_AND_ASSIGN(NullCheckSlowPathARM64); 376}; 377 378class SuspendCheckSlowPathARM64 : public SlowPathCodeARM64 { 379 public: 380 SuspendCheckSlowPathARM64(HSuspendCheck* instruction, HBasicBlock* successor) 381 : instruction_(instruction), successor_(successor) {} 382 383 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 384 CodeGeneratorARM64* arm64_codegen = down_cast<CodeGeneratorARM64*>(codegen); 385 __ Bind(GetEntryLabel()); 386 SaveLiveRegisters(codegen, instruction_->GetLocations()); 387 arm64_codegen->InvokeRuntime( 388 QUICK_ENTRY_POINT(pTestSuspend), instruction_, instruction_->GetDexPc(), this); 389 CheckEntrypointTypes<kQuickTestSuspend, void, void>(); 390 RestoreLiveRegisters(codegen, instruction_->GetLocations()); 391 if (successor_ == nullptr) { 392 __ B(GetReturnLabel()); 393 } else { 394 __ B(arm64_codegen->GetLabelOf(successor_)); 395 } 396 } 397 398 vixl::Label* GetReturnLabel() { 399 DCHECK(successor_ == nullptr); 400 return &return_label_; 401 } 402 403 HBasicBlock* GetSuccessor() const { 404 return successor_; 405 } 406 407 const char* GetDescription() const OVERRIDE { return "SuspendCheckSlowPathARM64"; } 408 409 private: 410 HSuspendCheck* const instruction_; 411 // If not null, the block to branch to after the suspend check. 412 HBasicBlock* const successor_; 413 414 // If `successor_` is null, the label to branch to after the suspend check. 415 vixl::Label return_label_; 416 417 DISALLOW_COPY_AND_ASSIGN(SuspendCheckSlowPathARM64); 418}; 419 420class TypeCheckSlowPathARM64 : public SlowPathCodeARM64 { 421 public: 422 TypeCheckSlowPathARM64(HInstruction* instruction, bool is_fatal) 423 : instruction_(instruction), is_fatal_(is_fatal) {} 424 425 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 426 LocationSummary* locations = instruction_->GetLocations(); 427 Location class_to_check = locations->InAt(1); 428 Location object_class = instruction_->IsCheckCast() ? locations->GetTemp(0) 429 : locations->Out(); 430 DCHECK(instruction_->IsCheckCast() 431 || !locations->GetLiveRegisters()->ContainsCoreRegister(locations->Out().reg())); 432 CodeGeneratorARM64* arm64_codegen = down_cast<CodeGeneratorARM64*>(codegen); 433 uint32_t dex_pc = instruction_->GetDexPc(); 434 435 __ Bind(GetEntryLabel()); 436 437 if (!is_fatal_) { 438 SaveLiveRegisters(codegen, locations); 439 } 440 441 // We're moving two locations to locations that could overlap, so we need a parallel 442 // move resolver. 443 InvokeRuntimeCallingConvention calling_convention; 444 codegen->EmitParallelMoves( 445 class_to_check, LocationFrom(calling_convention.GetRegisterAt(0)), Primitive::kPrimNot, 446 object_class, LocationFrom(calling_convention.GetRegisterAt(1)), Primitive::kPrimNot); 447 448 if (instruction_->IsInstanceOf()) { 449 arm64_codegen->InvokeRuntime( 450 QUICK_ENTRY_POINT(pInstanceofNonTrivial), instruction_, dex_pc, this); 451 CheckEntrypointTypes<kQuickInstanceofNonTrivial, uint32_t, 452 const mirror::Class*, const mirror::Class*>(); 453 Primitive::Type ret_type = instruction_->GetType(); 454 Location ret_loc = calling_convention.GetReturnLocation(ret_type); 455 arm64_codegen->MoveLocation(locations->Out(), ret_loc, ret_type); 456 } else { 457 DCHECK(instruction_->IsCheckCast()); 458 arm64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pCheckCast), instruction_, dex_pc, this); 459 CheckEntrypointTypes<kQuickCheckCast, void, const mirror::Class*, const mirror::Class*>(); 460 } 461 462 if (!is_fatal_) { 463 RestoreLiveRegisters(codegen, locations); 464 __ B(GetExitLabel()); 465 } 466 } 467 468 const char* GetDescription() const OVERRIDE { return "TypeCheckSlowPathARM64"; } 469 bool IsFatal() const { return is_fatal_; } 470 471 private: 472 HInstruction* const instruction_; 473 const bool is_fatal_; 474 475 DISALLOW_COPY_AND_ASSIGN(TypeCheckSlowPathARM64); 476}; 477 478class DeoptimizationSlowPathARM64 : public SlowPathCodeARM64 { 479 public: 480 explicit DeoptimizationSlowPathARM64(HDeoptimize* instruction) 481 : instruction_(instruction) {} 482 483 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 484 CodeGeneratorARM64* arm64_codegen = down_cast<CodeGeneratorARM64*>(codegen); 485 __ Bind(GetEntryLabel()); 486 SaveLiveRegisters(codegen, instruction_->GetLocations()); 487 arm64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pDeoptimize), 488 instruction_, 489 instruction_->GetDexPc(), 490 this); 491 CheckEntrypointTypes<kQuickDeoptimize, void, void>(); 492 } 493 494 const char* GetDescription() const OVERRIDE { return "DeoptimizationSlowPathARM64"; } 495 496 private: 497 HDeoptimize* const instruction_; 498 DISALLOW_COPY_AND_ASSIGN(DeoptimizationSlowPathARM64); 499}; 500 501class ArraySetSlowPathARM64 : public SlowPathCodeARM64 { 502 public: 503 explicit ArraySetSlowPathARM64(HInstruction* instruction) : instruction_(instruction) {} 504 505 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 506 LocationSummary* locations = instruction_->GetLocations(); 507 __ Bind(GetEntryLabel()); 508 SaveLiveRegisters(codegen, locations); 509 510 InvokeRuntimeCallingConvention calling_convention; 511 HParallelMove parallel_move(codegen->GetGraph()->GetArena()); 512 parallel_move.AddMove( 513 locations->InAt(0), 514 LocationFrom(calling_convention.GetRegisterAt(0)), 515 Primitive::kPrimNot, 516 nullptr); 517 parallel_move.AddMove( 518 locations->InAt(1), 519 LocationFrom(calling_convention.GetRegisterAt(1)), 520 Primitive::kPrimInt, 521 nullptr); 522 parallel_move.AddMove( 523 locations->InAt(2), 524 LocationFrom(calling_convention.GetRegisterAt(2)), 525 Primitive::kPrimNot, 526 nullptr); 527 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 528 529 CodeGeneratorARM64* arm64_codegen = down_cast<CodeGeneratorARM64*>(codegen); 530 arm64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pAputObject), 531 instruction_, 532 instruction_->GetDexPc(), 533 this); 534 CheckEntrypointTypes<kQuickAputObject, void, mirror::Array*, int32_t, mirror::Object*>(); 535 RestoreLiveRegisters(codegen, locations); 536 __ B(GetExitLabel()); 537 } 538 539 const char* GetDescription() const OVERRIDE { return "ArraySetSlowPathARM64"; } 540 541 private: 542 HInstruction* const instruction_; 543 544 DISALLOW_COPY_AND_ASSIGN(ArraySetSlowPathARM64); 545}; 546 547void JumpTableARM64::EmitTable(CodeGeneratorARM64* codegen) { 548 uint32_t num_entries = switch_instr_->GetNumEntries(); 549 DCHECK_GE(num_entries, kPackedSwitchCompareJumpThreshold); 550 551 // We are about to use the assembler to place literals directly. Make sure we have enough 552 // underlying code buffer and we have generated the jump table with right size. 553 CodeBufferCheckScope scope(codegen->GetVIXLAssembler(), num_entries * sizeof(int32_t), 554 CodeBufferCheckScope::kCheck, CodeBufferCheckScope::kExactSize); 555 556 __ Bind(&table_start_); 557 const ArenaVector<HBasicBlock*>& successors = switch_instr_->GetBlock()->GetSuccessors(); 558 for (uint32_t i = 0; i < num_entries; i++) { 559 vixl::Label* target_label = codegen->GetLabelOf(successors[i]); 560 DCHECK(target_label->IsBound()); 561 ptrdiff_t jump_offset = target_label->location() - table_start_.location(); 562 DCHECK_GT(jump_offset, std::numeric_limits<int32_t>::min()); 563 DCHECK_LE(jump_offset, std::numeric_limits<int32_t>::max()); 564 Literal<int32_t> literal(jump_offset); 565 __ place(&literal); 566 } 567} 568 569// Slow path generating a read barrier for a heap reference. 570class ReadBarrierForHeapReferenceSlowPathARM64 : public SlowPathCodeARM64 { 571 public: 572 ReadBarrierForHeapReferenceSlowPathARM64(HInstruction* instruction, 573 Location out, 574 Location ref, 575 Location obj, 576 uint32_t offset, 577 Location index) 578 : instruction_(instruction), 579 out_(out), 580 ref_(ref), 581 obj_(obj), 582 offset_(offset), 583 index_(index) { 584 DCHECK(kEmitCompilerReadBarrier); 585 // If `obj` is equal to `out` or `ref`, it means the initial object 586 // has been overwritten by (or after) the heap object reference load 587 // to be instrumented, e.g.: 588 // 589 // __ Ldr(out, HeapOperand(out, class_offset); 590 // codegen_->GenerateReadBarrier(instruction, out_loc, out_loc, out_loc, offset); 591 // 592 // In that case, we have lost the information about the original 593 // object, and the emitted read barrier cannot work properly. 594 DCHECK(!obj.Equals(out)) << "obj=" << obj << " out=" << out; 595 DCHECK(!obj.Equals(ref)) << "obj=" << obj << " ref=" << ref; 596 } 597 598 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 599 CodeGeneratorARM64* arm64_codegen = down_cast<CodeGeneratorARM64*>(codegen); 600 LocationSummary* locations = instruction_->GetLocations(); 601 Primitive::Type type = Primitive::kPrimNot; 602 DCHECK(locations->CanCall()); 603 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(out_.reg())); 604 DCHECK(!instruction_->IsInvoke() || 605 (instruction_->IsInvokeStaticOrDirect() && 606 instruction_->GetLocations()->Intrinsified())); 607 608 __ Bind(GetEntryLabel()); 609 610 // Note: In the case of a HArrayGet instruction, when the base 611 // address is a HArm64IntermediateAddress instruction, it does not 612 // point to the array object itself, but to an offset within this 613 // object. However, the read barrier entry point needs the array 614 // object address to be passed as first argument. So we 615 // temporarily set back `obj_` to that address, and restore its 616 // initial value later. 617 if (instruction_->IsArrayGet() && 618 instruction_->AsArrayGet()->GetArray()->IsArm64IntermediateAddress()) { 619 if (kIsDebugBuild) { 620 HArm64IntermediateAddress* intermediate_address = 621 instruction_->AsArrayGet()->GetArray()->AsArm64IntermediateAddress(); 622 uint32_t intermediate_address_offset = 623 intermediate_address->GetOffset()->AsIntConstant()->GetValueAsUint64(); 624 DCHECK_EQ(intermediate_address_offset, offset_); 625 DCHECK_EQ(mirror::Array::DataOffset(Primitive::ComponentSize(type)).Uint32Value(), offset_); 626 } 627 Register obj_reg = RegisterFrom(obj_, Primitive::kPrimInt); 628 __ Sub(obj_reg, obj_reg, offset_); 629 } 630 631 SaveLiveRegisters(codegen, locations); 632 633 // We may have to change the index's value, but as `index_` is a 634 // constant member (like other "inputs" of this slow path), 635 // introduce a copy of it, `index`. 636 Location index = index_; 637 if (index_.IsValid()) { 638 // Handle `index_` for HArrayGet and intrinsic UnsafeGetObject. 639 if (instruction_->IsArrayGet()) { 640 // Compute the actual memory offset and store it in `index`. 641 Register index_reg = RegisterFrom(index_, Primitive::kPrimInt); 642 DCHECK(locations->GetLiveRegisters()->ContainsCoreRegister(index_.reg())); 643 if (codegen->IsCoreCalleeSaveRegister(index_.reg())) { 644 // We are about to change the value of `index_reg` (see the 645 // calls to vixl::MacroAssembler::Lsl and 646 // vixl::MacroAssembler::Mov below), but it has 647 // not been saved by the previous call to 648 // art::SlowPathCode::SaveLiveRegisters, as it is a 649 // callee-save register -- 650 // art::SlowPathCode::SaveLiveRegisters does not consider 651 // callee-save registers, as it has been designed with the 652 // assumption that callee-save registers are supposed to be 653 // handled by the called function. So, as a callee-save 654 // register, `index_reg` _would_ eventually be saved onto 655 // the stack, but it would be too late: we would have 656 // changed its value earlier. Therefore, we manually save 657 // it here into another freely available register, 658 // `free_reg`, chosen of course among the caller-save 659 // registers (as a callee-save `free_reg` register would 660 // exhibit the same problem). 661 // 662 // Note we could have requested a temporary register from 663 // the register allocator instead; but we prefer not to, as 664 // this is a slow path, and we know we can find a 665 // caller-save register that is available. 666 Register free_reg = FindAvailableCallerSaveRegister(codegen); 667 __ Mov(free_reg.W(), index_reg); 668 index_reg = free_reg; 669 index = LocationFrom(index_reg); 670 } else { 671 // The initial register stored in `index_` has already been 672 // saved in the call to art::SlowPathCode::SaveLiveRegisters 673 // (as it is not a callee-save register), so we can freely 674 // use it. 675 } 676 // Shifting the index value contained in `index_reg` by the scale 677 // factor (2) cannot overflow in practice, as the runtime is 678 // unable to allocate object arrays with a size larger than 679 // 2^26 - 1 (that is, 2^28 - 4 bytes). 680 __ Lsl(index_reg, index_reg, Primitive::ComponentSizeShift(type)); 681 static_assert( 682 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 683 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 684 __ Add(index_reg, index_reg, Operand(offset_)); 685 } else { 686 DCHECK(instruction_->IsInvoke()); 687 DCHECK(instruction_->GetLocations()->Intrinsified()); 688 DCHECK((instruction_->AsInvoke()->GetIntrinsic() == Intrinsics::kUnsafeGetObject) || 689 (instruction_->AsInvoke()->GetIntrinsic() == Intrinsics::kUnsafeGetObjectVolatile)) 690 << instruction_->AsInvoke()->GetIntrinsic(); 691 DCHECK_EQ(offset_, 0U); 692 DCHECK(index_.IsRegisterPair()); 693 // UnsafeGet's offset location is a register pair, the low 694 // part contains the correct offset. 695 index = index_.ToLow(); 696 } 697 } 698 699 // We're moving two or three locations to locations that could 700 // overlap, so we need a parallel move resolver. 701 InvokeRuntimeCallingConvention calling_convention; 702 HParallelMove parallel_move(codegen->GetGraph()->GetArena()); 703 parallel_move.AddMove(ref_, 704 LocationFrom(calling_convention.GetRegisterAt(0)), 705 type, 706 nullptr); 707 parallel_move.AddMove(obj_, 708 LocationFrom(calling_convention.GetRegisterAt(1)), 709 type, 710 nullptr); 711 if (index.IsValid()) { 712 parallel_move.AddMove(index, 713 LocationFrom(calling_convention.GetRegisterAt(2)), 714 Primitive::kPrimInt, 715 nullptr); 716 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 717 } else { 718 codegen->GetMoveResolver()->EmitNativeCode(¶llel_move); 719 arm64_codegen->MoveConstant(LocationFrom(calling_convention.GetRegisterAt(2)), offset_); 720 } 721 arm64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierSlow), 722 instruction_, 723 instruction_->GetDexPc(), 724 this); 725 CheckEntrypointTypes< 726 kQuickReadBarrierSlow, mirror::Object*, mirror::Object*, mirror::Object*, uint32_t>(); 727 arm64_codegen->MoveLocation(out_, calling_convention.GetReturnLocation(type), type); 728 729 RestoreLiveRegisters(codegen, locations); 730 731 // Restore the value of `obj_` when it corresponds to a 732 // HArm64IntermediateAddress instruction. 733 if (instruction_->IsArrayGet() && 734 instruction_->AsArrayGet()->GetArray()->IsArm64IntermediateAddress()) { 735 if (kIsDebugBuild) { 736 HArm64IntermediateAddress* intermediate_address = 737 instruction_->AsArrayGet()->GetArray()->AsArm64IntermediateAddress(); 738 uint32_t intermediate_address_offset = 739 intermediate_address->GetOffset()->AsIntConstant()->GetValueAsUint64(); 740 DCHECK_EQ(intermediate_address_offset, offset_); 741 DCHECK_EQ(mirror::Array::DataOffset(Primitive::ComponentSize(type)).Uint32Value(), offset_); 742 } 743 Register obj_reg = RegisterFrom(obj_, Primitive::kPrimInt); 744 __ Add(obj_reg, obj_reg, offset_); 745 } 746 747 __ B(GetExitLabel()); 748 } 749 750 const char* GetDescription() const OVERRIDE { return "ReadBarrierForHeapReferenceSlowPathARM64"; } 751 752 private: 753 Register FindAvailableCallerSaveRegister(CodeGenerator* codegen) { 754 size_t ref = static_cast<int>(XRegisterFrom(ref_).code()); 755 size_t obj = static_cast<int>(XRegisterFrom(obj_).code()); 756 for (size_t i = 0, e = codegen->GetNumberOfCoreRegisters(); i < e; ++i) { 757 if (i != ref && i != obj && !codegen->IsCoreCalleeSaveRegister(i)) { 758 return Register(VIXLRegCodeFromART(i), kXRegSize); 759 } 760 } 761 // We shall never fail to find a free caller-save register, as 762 // there are more than two core caller-save registers on ARM64 763 // (meaning it is possible to find one which is different from 764 // `ref` and `obj`). 765 DCHECK_GT(codegen->GetNumberOfCoreCallerSaveRegisters(), 2u); 766 LOG(FATAL) << "Could not find a free register"; 767 UNREACHABLE(); 768 } 769 770 HInstruction* const instruction_; 771 const Location out_; 772 const Location ref_; 773 const Location obj_; 774 const uint32_t offset_; 775 // An additional location containing an index to an array. 776 // Only used for HArrayGet and the UnsafeGetObject & 777 // UnsafeGetObjectVolatile intrinsics. 778 const Location index_; 779 780 DISALLOW_COPY_AND_ASSIGN(ReadBarrierForHeapReferenceSlowPathARM64); 781}; 782 783// Slow path generating a read barrier for a GC root. 784class ReadBarrierForRootSlowPathARM64 : public SlowPathCodeARM64 { 785 public: 786 ReadBarrierForRootSlowPathARM64(HInstruction* instruction, Location out, Location root) 787 : instruction_(instruction), out_(out), root_(root) {} 788 789 void EmitNativeCode(CodeGenerator* codegen) OVERRIDE { 790 LocationSummary* locations = instruction_->GetLocations(); 791 Primitive::Type type = Primitive::kPrimNot; 792 DCHECK(locations->CanCall()); 793 DCHECK(!locations->GetLiveRegisters()->ContainsCoreRegister(out_.reg())); 794 DCHECK(instruction_->IsLoadClass() || instruction_->IsLoadString()); 795 796 __ Bind(GetEntryLabel()); 797 SaveLiveRegisters(codegen, locations); 798 799 InvokeRuntimeCallingConvention calling_convention; 800 CodeGeneratorARM64* arm64_codegen = down_cast<CodeGeneratorARM64*>(codegen); 801 // The argument of the ReadBarrierForRootSlow is not a managed 802 // reference (`mirror::Object*`), but a `GcRoot<mirror::Object>*`; 803 // thus we need a 64-bit move here, and we cannot use 804 // 805 // arm64_codegen->MoveLocation( 806 // LocationFrom(calling_convention.GetRegisterAt(0)), 807 // root_, 808 // type); 809 // 810 // which would emit a 32-bit move, as `type` is a (32-bit wide) 811 // reference type (`Primitive::kPrimNot`). 812 __ Mov(calling_convention.GetRegisterAt(0), XRegisterFrom(out_)); 813 arm64_codegen->InvokeRuntime(QUICK_ENTRY_POINT(pReadBarrierForRootSlow), 814 instruction_, 815 instruction_->GetDexPc(), 816 this); 817 CheckEntrypointTypes<kQuickReadBarrierForRootSlow, mirror::Object*, GcRoot<mirror::Object>*>(); 818 arm64_codegen->MoveLocation(out_, calling_convention.GetReturnLocation(type), type); 819 820 RestoreLiveRegisters(codegen, locations); 821 __ B(GetExitLabel()); 822 } 823 824 const char* GetDescription() const OVERRIDE { return "ReadBarrierForRootSlowPathARM64"; } 825 826 private: 827 HInstruction* const instruction_; 828 const Location out_; 829 const Location root_; 830 831 DISALLOW_COPY_AND_ASSIGN(ReadBarrierForRootSlowPathARM64); 832}; 833 834#undef __ 835 836Location InvokeDexCallingConventionVisitorARM64::GetNextLocation(Primitive::Type type) { 837 Location next_location; 838 if (type == Primitive::kPrimVoid) { 839 LOG(FATAL) << "Unreachable type " << type; 840 } 841 842 if (Primitive::IsFloatingPointType(type) && 843 (float_index_ < calling_convention.GetNumberOfFpuRegisters())) { 844 next_location = LocationFrom(calling_convention.GetFpuRegisterAt(float_index_++)); 845 } else if (!Primitive::IsFloatingPointType(type) && 846 (gp_index_ < calling_convention.GetNumberOfRegisters())) { 847 next_location = LocationFrom(calling_convention.GetRegisterAt(gp_index_++)); 848 } else { 849 size_t stack_offset = calling_convention.GetStackOffsetOf(stack_index_); 850 next_location = Primitive::Is64BitType(type) ? Location::DoubleStackSlot(stack_offset) 851 : Location::StackSlot(stack_offset); 852 } 853 854 // Space on the stack is reserved for all arguments. 855 stack_index_ += Primitive::Is64BitType(type) ? 2 : 1; 856 return next_location; 857} 858 859Location InvokeDexCallingConventionVisitorARM64::GetMethodLocation() const { 860 return LocationFrom(kArtMethodRegister); 861} 862 863CodeGeneratorARM64::CodeGeneratorARM64(HGraph* graph, 864 const Arm64InstructionSetFeatures& isa_features, 865 const CompilerOptions& compiler_options, 866 OptimizingCompilerStats* stats) 867 : CodeGenerator(graph, 868 kNumberOfAllocatableRegisters, 869 kNumberOfAllocatableFPRegisters, 870 kNumberOfAllocatableRegisterPairs, 871 callee_saved_core_registers.list(), 872 callee_saved_fp_registers.list(), 873 compiler_options, 874 stats), 875 block_labels_(nullptr), 876 jump_tables_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 877 location_builder_(graph, this), 878 instruction_visitor_(graph, this), 879 move_resolver_(graph->GetArena(), this), 880 isa_features_(isa_features), 881 uint64_literals_(std::less<uint64_t>(), 882 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 883 method_patches_(MethodReferenceComparator(), 884 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 885 call_patches_(MethodReferenceComparator(), 886 graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 887 relative_call_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)), 888 pc_relative_dex_cache_patches_(graph->GetArena()->Adapter(kArenaAllocCodeGenerator)) { 889 // Save the link register (containing the return address) to mimic Quick. 890 AddAllocatedRegister(LocationFrom(lr)); 891} 892 893#define __ GetVIXLAssembler()-> 894 895void CodeGeneratorARM64::EmitJumpTables() { 896 for (auto jump_table : jump_tables_) { 897 jump_table->EmitTable(this); 898 } 899} 900 901void CodeGeneratorARM64::Finalize(CodeAllocator* allocator) { 902 EmitJumpTables(); 903 // Ensure we emit the literal pool. 904 __ FinalizeCode(); 905 906 CodeGenerator::Finalize(allocator); 907} 908 909void ParallelMoveResolverARM64::PrepareForEmitNativeCode() { 910 // Note: There are 6 kinds of moves: 911 // 1. constant -> GPR/FPR (non-cycle) 912 // 2. constant -> stack (non-cycle) 913 // 3. GPR/FPR -> GPR/FPR 914 // 4. GPR/FPR -> stack 915 // 5. stack -> GPR/FPR 916 // 6. stack -> stack (non-cycle) 917 // Case 1, 2 and 6 should never be included in a dependency cycle on ARM64. For case 3, 4, and 5 918 // VIXL uses at most 1 GPR. VIXL has 2 GPR and 1 FPR temps, and there should be no intersecting 919 // cycles on ARM64, so we always have 1 GPR and 1 FPR available VIXL temps to resolve the 920 // dependency. 921 vixl_temps_.Open(GetVIXLAssembler()); 922} 923 924void ParallelMoveResolverARM64::FinishEmitNativeCode() { 925 vixl_temps_.Close(); 926} 927 928Location ParallelMoveResolverARM64::AllocateScratchLocationFor(Location::Kind kind) { 929 DCHECK(kind == Location::kRegister || kind == Location::kFpuRegister || 930 kind == Location::kStackSlot || kind == Location::kDoubleStackSlot); 931 kind = (kind == Location::kFpuRegister) ? Location::kFpuRegister : Location::kRegister; 932 Location scratch = GetScratchLocation(kind); 933 if (!scratch.Equals(Location::NoLocation())) { 934 return scratch; 935 } 936 // Allocate from VIXL temp registers. 937 if (kind == Location::kRegister) { 938 scratch = LocationFrom(vixl_temps_.AcquireX()); 939 } else { 940 DCHECK(kind == Location::kFpuRegister); 941 scratch = LocationFrom(vixl_temps_.AcquireD()); 942 } 943 AddScratchLocation(scratch); 944 return scratch; 945} 946 947void ParallelMoveResolverARM64::FreeScratchLocation(Location loc) { 948 if (loc.IsRegister()) { 949 vixl_temps_.Release(XRegisterFrom(loc)); 950 } else { 951 DCHECK(loc.IsFpuRegister()); 952 vixl_temps_.Release(DRegisterFrom(loc)); 953 } 954 RemoveScratchLocation(loc); 955} 956 957void ParallelMoveResolverARM64::EmitMove(size_t index) { 958 MoveOperands* move = moves_[index]; 959 codegen_->MoveLocation(move->GetDestination(), move->GetSource(), Primitive::kPrimVoid); 960} 961 962void CodeGeneratorARM64::GenerateFrameEntry() { 963 MacroAssembler* masm = GetVIXLAssembler(); 964 BlockPoolsScope block_pools(masm); 965 __ Bind(&frame_entry_label_); 966 967 bool do_overflow_check = FrameNeedsStackCheck(GetFrameSize(), kArm64) || !IsLeafMethod(); 968 if (do_overflow_check) { 969 UseScratchRegisterScope temps(masm); 970 Register temp = temps.AcquireX(); 971 DCHECK(GetCompilerOptions().GetImplicitStackOverflowChecks()); 972 __ Sub(temp, sp, static_cast<int32_t>(GetStackOverflowReservedBytes(kArm64))); 973 __ Ldr(wzr, MemOperand(temp, 0)); 974 RecordPcInfo(nullptr, 0); 975 } 976 977 if (!HasEmptyFrame()) { 978 int frame_size = GetFrameSize(); 979 // Stack layout: 980 // sp[frame_size - 8] : lr. 981 // ... : other preserved core registers. 982 // ... : other preserved fp registers. 983 // ... : reserved frame space. 984 // sp[0] : current method. 985 __ Str(kArtMethodRegister, MemOperand(sp, -frame_size, PreIndex)); 986 GetAssembler()->cfi().AdjustCFAOffset(frame_size); 987 GetAssembler()->SpillRegisters(GetFramePreservedCoreRegisters(), 988 frame_size - GetCoreSpillSize()); 989 GetAssembler()->SpillRegisters(GetFramePreservedFPRegisters(), 990 frame_size - FrameEntrySpillSize()); 991 } 992} 993 994void CodeGeneratorARM64::GenerateFrameExit() { 995 BlockPoolsScope block_pools(GetVIXLAssembler()); 996 GetAssembler()->cfi().RememberState(); 997 if (!HasEmptyFrame()) { 998 int frame_size = GetFrameSize(); 999 GetAssembler()->UnspillRegisters(GetFramePreservedFPRegisters(), 1000 frame_size - FrameEntrySpillSize()); 1001 GetAssembler()->UnspillRegisters(GetFramePreservedCoreRegisters(), 1002 frame_size - GetCoreSpillSize()); 1003 __ Drop(frame_size); 1004 GetAssembler()->cfi().AdjustCFAOffset(-frame_size); 1005 } 1006 __ Ret(); 1007 GetAssembler()->cfi().RestoreState(); 1008 GetAssembler()->cfi().DefCFAOffset(GetFrameSize()); 1009} 1010 1011vixl::CPURegList CodeGeneratorARM64::GetFramePreservedCoreRegisters() const { 1012 DCHECK(ArtVixlRegCodeCoherentForRegSet(core_spill_mask_, GetNumberOfCoreRegisters(), 0, 0)); 1013 return vixl::CPURegList(vixl::CPURegister::kRegister, vixl::kXRegSize, 1014 core_spill_mask_); 1015} 1016 1017vixl::CPURegList CodeGeneratorARM64::GetFramePreservedFPRegisters() const { 1018 DCHECK(ArtVixlRegCodeCoherentForRegSet(0, 0, fpu_spill_mask_, 1019 GetNumberOfFloatingPointRegisters())); 1020 return vixl::CPURegList(vixl::CPURegister::kFPRegister, vixl::kDRegSize, 1021 fpu_spill_mask_); 1022} 1023 1024void CodeGeneratorARM64::Bind(HBasicBlock* block) { 1025 __ Bind(GetLabelOf(block)); 1026} 1027 1028void CodeGeneratorARM64::Move(HInstruction* instruction, 1029 Location location, 1030 HInstruction* move_for) { 1031 LocationSummary* locations = instruction->GetLocations(); 1032 Primitive::Type type = instruction->GetType(); 1033 DCHECK_NE(type, Primitive::kPrimVoid); 1034 1035 if (instruction->IsFakeString()) { 1036 // The fake string is an alias for null. 1037 DCHECK(IsBaseline()); 1038 instruction = locations->Out().GetConstant(); 1039 DCHECK(instruction->IsNullConstant()) << instruction->DebugName(); 1040 } 1041 1042 if (instruction->IsCurrentMethod()) { 1043 MoveLocation(location, 1044 Location::DoubleStackSlot(kCurrentMethodStackOffset), 1045 Primitive::kPrimVoid); 1046 } else if (locations != nullptr && locations->Out().Equals(location)) { 1047 return; 1048 } else if (instruction->IsIntConstant() 1049 || instruction->IsLongConstant() 1050 || instruction->IsNullConstant()) { 1051 int64_t value = GetInt64ValueOf(instruction->AsConstant()); 1052 if (location.IsRegister()) { 1053 Register dst = RegisterFrom(location, type); 1054 DCHECK(((instruction->IsIntConstant() || instruction->IsNullConstant()) && dst.Is32Bits()) || 1055 (instruction->IsLongConstant() && dst.Is64Bits())); 1056 __ Mov(dst, value); 1057 } else { 1058 DCHECK(location.IsStackSlot() || location.IsDoubleStackSlot()); 1059 UseScratchRegisterScope temps(GetVIXLAssembler()); 1060 Register temp = (instruction->IsIntConstant() || instruction->IsNullConstant()) 1061 ? temps.AcquireW() 1062 : temps.AcquireX(); 1063 __ Mov(temp, value); 1064 __ Str(temp, StackOperandFrom(location)); 1065 } 1066 } else if (instruction->IsTemporary()) { 1067 Location temp_location = GetTemporaryLocation(instruction->AsTemporary()); 1068 MoveLocation(location, temp_location, type); 1069 } else if (instruction->IsLoadLocal()) { 1070 uint32_t stack_slot = GetStackSlot(instruction->AsLoadLocal()->GetLocal()); 1071 if (Primitive::Is64BitType(type)) { 1072 MoveLocation(location, Location::DoubleStackSlot(stack_slot), type); 1073 } else { 1074 MoveLocation(location, Location::StackSlot(stack_slot), type); 1075 } 1076 1077 } else { 1078 DCHECK((instruction->GetNext() == move_for) || instruction->GetNext()->IsTemporary()); 1079 MoveLocation(location, locations->Out(), type); 1080 } 1081} 1082 1083void CodeGeneratorARM64::MoveConstant(Location location, int32_t value) { 1084 DCHECK(location.IsRegister()); 1085 __ Mov(RegisterFrom(location, Primitive::kPrimInt), value); 1086} 1087 1088void CodeGeneratorARM64::AddLocationAsTemp(Location location, LocationSummary* locations) { 1089 if (location.IsRegister()) { 1090 locations->AddTemp(location); 1091 } else { 1092 UNIMPLEMENTED(FATAL) << "AddLocationAsTemp not implemented for location " << location; 1093 } 1094} 1095 1096Location CodeGeneratorARM64::GetStackLocation(HLoadLocal* load) const { 1097 Primitive::Type type = load->GetType(); 1098 1099 switch (type) { 1100 case Primitive::kPrimNot: 1101 case Primitive::kPrimInt: 1102 case Primitive::kPrimFloat: 1103 return Location::StackSlot(GetStackSlot(load->GetLocal())); 1104 1105 case Primitive::kPrimLong: 1106 case Primitive::kPrimDouble: 1107 return Location::DoubleStackSlot(GetStackSlot(load->GetLocal())); 1108 1109 case Primitive::kPrimBoolean: 1110 case Primitive::kPrimByte: 1111 case Primitive::kPrimChar: 1112 case Primitive::kPrimShort: 1113 case Primitive::kPrimVoid: 1114 LOG(FATAL) << "Unexpected type " << type; 1115 } 1116 1117 LOG(FATAL) << "Unreachable"; 1118 return Location::NoLocation(); 1119} 1120 1121void CodeGeneratorARM64::MarkGCCard(Register object, Register value, bool value_can_be_null) { 1122 UseScratchRegisterScope temps(GetVIXLAssembler()); 1123 Register card = temps.AcquireX(); 1124 Register temp = temps.AcquireW(); // Index within the CardTable - 32bit. 1125 vixl::Label done; 1126 if (value_can_be_null) { 1127 __ Cbz(value, &done); 1128 } 1129 __ Ldr(card, MemOperand(tr, Thread::CardTableOffset<kArm64WordSize>().Int32Value())); 1130 __ Lsr(temp, object, gc::accounting::CardTable::kCardShift); 1131 __ Strb(card, MemOperand(card, temp.X())); 1132 if (value_can_be_null) { 1133 __ Bind(&done); 1134 } 1135} 1136 1137void CodeGeneratorARM64::SetupBlockedRegisters(bool is_baseline) const { 1138 // Blocked core registers: 1139 // lr : Runtime reserved. 1140 // tr : Runtime reserved. 1141 // xSuspend : Runtime reserved. TODO: Unblock this when the runtime stops using it. 1142 // ip1 : VIXL core temp. 1143 // ip0 : VIXL core temp. 1144 // 1145 // Blocked fp registers: 1146 // d31 : VIXL fp temp. 1147 CPURegList reserved_core_registers = vixl_reserved_core_registers; 1148 reserved_core_registers.Combine(runtime_reserved_core_registers); 1149 while (!reserved_core_registers.IsEmpty()) { 1150 blocked_core_registers_[reserved_core_registers.PopLowestIndex().code()] = true; 1151 } 1152 1153 CPURegList reserved_fp_registers = vixl_reserved_fp_registers; 1154 while (!reserved_fp_registers.IsEmpty()) { 1155 blocked_fpu_registers_[reserved_fp_registers.PopLowestIndex().code()] = true; 1156 } 1157 1158 if (is_baseline) { 1159 CPURegList reserved_core_baseline_registers = callee_saved_core_registers; 1160 while (!reserved_core_baseline_registers.IsEmpty()) { 1161 blocked_core_registers_[reserved_core_baseline_registers.PopLowestIndex().code()] = true; 1162 } 1163 } 1164 1165 if (is_baseline || GetGraph()->IsDebuggable()) { 1166 // Stubs do not save callee-save floating point registers. If the graph 1167 // is debuggable, we need to deal with these registers differently. For 1168 // now, just block them. 1169 CPURegList reserved_fp_baseline_registers = callee_saved_fp_registers; 1170 while (!reserved_fp_baseline_registers.IsEmpty()) { 1171 blocked_fpu_registers_[reserved_fp_baseline_registers.PopLowestIndex().code()] = true; 1172 } 1173 } 1174} 1175 1176Location CodeGeneratorARM64::AllocateFreeRegister(Primitive::Type type) const { 1177 if (type == Primitive::kPrimVoid) { 1178 LOG(FATAL) << "Unreachable type " << type; 1179 } 1180 1181 if (Primitive::IsFloatingPointType(type)) { 1182 ssize_t reg = FindFreeEntry(blocked_fpu_registers_, kNumberOfAllocatableFPRegisters); 1183 DCHECK_NE(reg, -1); 1184 return Location::FpuRegisterLocation(reg); 1185 } else { 1186 ssize_t reg = FindFreeEntry(blocked_core_registers_, kNumberOfAllocatableRegisters); 1187 DCHECK_NE(reg, -1); 1188 return Location::RegisterLocation(reg); 1189 } 1190} 1191 1192size_t CodeGeneratorARM64::SaveCoreRegister(size_t stack_index, uint32_t reg_id) { 1193 Register reg = Register(VIXLRegCodeFromART(reg_id), kXRegSize); 1194 __ Str(reg, MemOperand(sp, stack_index)); 1195 return kArm64WordSize; 1196} 1197 1198size_t CodeGeneratorARM64::RestoreCoreRegister(size_t stack_index, uint32_t reg_id) { 1199 Register reg = Register(VIXLRegCodeFromART(reg_id), kXRegSize); 1200 __ Ldr(reg, MemOperand(sp, stack_index)); 1201 return kArm64WordSize; 1202} 1203 1204size_t CodeGeneratorARM64::SaveFloatingPointRegister(size_t stack_index, uint32_t reg_id) { 1205 FPRegister reg = FPRegister(reg_id, kDRegSize); 1206 __ Str(reg, MemOperand(sp, stack_index)); 1207 return kArm64WordSize; 1208} 1209 1210size_t CodeGeneratorARM64::RestoreFloatingPointRegister(size_t stack_index, uint32_t reg_id) { 1211 FPRegister reg = FPRegister(reg_id, kDRegSize); 1212 __ Ldr(reg, MemOperand(sp, stack_index)); 1213 return kArm64WordSize; 1214} 1215 1216void CodeGeneratorARM64::DumpCoreRegister(std::ostream& stream, int reg) const { 1217 stream << XRegister(reg); 1218} 1219 1220void CodeGeneratorARM64::DumpFloatingPointRegister(std::ostream& stream, int reg) const { 1221 stream << DRegister(reg); 1222} 1223 1224void CodeGeneratorARM64::MoveConstant(CPURegister destination, HConstant* constant) { 1225 if (constant->IsIntConstant()) { 1226 __ Mov(Register(destination), constant->AsIntConstant()->GetValue()); 1227 } else if (constant->IsLongConstant()) { 1228 __ Mov(Register(destination), constant->AsLongConstant()->GetValue()); 1229 } else if (constant->IsNullConstant()) { 1230 __ Mov(Register(destination), 0); 1231 } else if (constant->IsFloatConstant()) { 1232 __ Fmov(FPRegister(destination), constant->AsFloatConstant()->GetValue()); 1233 } else { 1234 DCHECK(constant->IsDoubleConstant()); 1235 __ Fmov(FPRegister(destination), constant->AsDoubleConstant()->GetValue()); 1236 } 1237} 1238 1239 1240static bool CoherentConstantAndType(Location constant, Primitive::Type type) { 1241 DCHECK(constant.IsConstant()); 1242 HConstant* cst = constant.GetConstant(); 1243 return (cst->IsIntConstant() && type == Primitive::kPrimInt) || 1244 // Null is mapped to a core W register, which we associate with kPrimInt. 1245 (cst->IsNullConstant() && type == Primitive::kPrimInt) || 1246 (cst->IsLongConstant() && type == Primitive::kPrimLong) || 1247 (cst->IsFloatConstant() && type == Primitive::kPrimFloat) || 1248 (cst->IsDoubleConstant() && type == Primitive::kPrimDouble); 1249} 1250 1251void CodeGeneratorARM64::MoveLocation(Location destination, 1252 Location source, 1253 Primitive::Type dst_type) { 1254 if (source.Equals(destination)) { 1255 return; 1256 } 1257 1258 // A valid move can always be inferred from the destination and source 1259 // locations. When moving from and to a register, the argument type can be 1260 // used to generate 32bit instead of 64bit moves. In debug mode we also 1261 // checks the coherency of the locations and the type. 1262 bool unspecified_type = (dst_type == Primitive::kPrimVoid); 1263 1264 if (destination.IsRegister() || destination.IsFpuRegister()) { 1265 if (unspecified_type) { 1266 HConstant* src_cst = source.IsConstant() ? source.GetConstant() : nullptr; 1267 if (source.IsStackSlot() || 1268 (src_cst != nullptr && (src_cst->IsIntConstant() 1269 || src_cst->IsFloatConstant() 1270 || src_cst->IsNullConstant()))) { 1271 // For stack slots and 32bit constants, a 64bit type is appropriate. 1272 dst_type = destination.IsRegister() ? Primitive::kPrimInt : Primitive::kPrimFloat; 1273 } else { 1274 // If the source is a double stack slot or a 64bit constant, a 64bit 1275 // type is appropriate. Else the source is a register, and since the 1276 // type has not been specified, we chose a 64bit type to force a 64bit 1277 // move. 1278 dst_type = destination.IsRegister() ? Primitive::kPrimLong : Primitive::kPrimDouble; 1279 } 1280 } 1281 DCHECK((destination.IsFpuRegister() && Primitive::IsFloatingPointType(dst_type)) || 1282 (destination.IsRegister() && !Primitive::IsFloatingPointType(dst_type))); 1283 CPURegister dst = CPURegisterFrom(destination, dst_type); 1284 if (source.IsStackSlot() || source.IsDoubleStackSlot()) { 1285 DCHECK(dst.Is64Bits() == source.IsDoubleStackSlot()); 1286 __ Ldr(dst, StackOperandFrom(source)); 1287 } else if (source.IsConstant()) { 1288 DCHECK(CoherentConstantAndType(source, dst_type)); 1289 MoveConstant(dst, source.GetConstant()); 1290 } else if (source.IsRegister()) { 1291 if (destination.IsRegister()) { 1292 __ Mov(Register(dst), RegisterFrom(source, dst_type)); 1293 } else { 1294 DCHECK(destination.IsFpuRegister()); 1295 Primitive::Type source_type = Primitive::Is64BitType(dst_type) 1296 ? Primitive::kPrimLong 1297 : Primitive::kPrimInt; 1298 __ Fmov(FPRegisterFrom(destination, dst_type), RegisterFrom(source, source_type)); 1299 } 1300 } else { 1301 DCHECK(source.IsFpuRegister()); 1302 if (destination.IsRegister()) { 1303 Primitive::Type source_type = Primitive::Is64BitType(dst_type) 1304 ? Primitive::kPrimDouble 1305 : Primitive::kPrimFloat; 1306 __ Fmov(RegisterFrom(destination, dst_type), FPRegisterFrom(source, source_type)); 1307 } else { 1308 DCHECK(destination.IsFpuRegister()); 1309 __ Fmov(FPRegister(dst), FPRegisterFrom(source, dst_type)); 1310 } 1311 } 1312 } else { // The destination is not a register. It must be a stack slot. 1313 DCHECK(destination.IsStackSlot() || destination.IsDoubleStackSlot()); 1314 if (source.IsRegister() || source.IsFpuRegister()) { 1315 if (unspecified_type) { 1316 if (source.IsRegister()) { 1317 dst_type = destination.IsStackSlot() ? Primitive::kPrimInt : Primitive::kPrimLong; 1318 } else { 1319 dst_type = destination.IsStackSlot() ? Primitive::kPrimFloat : Primitive::kPrimDouble; 1320 } 1321 } 1322 DCHECK((destination.IsDoubleStackSlot() == Primitive::Is64BitType(dst_type)) && 1323 (source.IsFpuRegister() == Primitive::IsFloatingPointType(dst_type))); 1324 __ Str(CPURegisterFrom(source, dst_type), StackOperandFrom(destination)); 1325 } else if (source.IsConstant()) { 1326 DCHECK(unspecified_type || CoherentConstantAndType(source, dst_type)) 1327 << source << " " << dst_type; 1328 UseScratchRegisterScope temps(GetVIXLAssembler()); 1329 HConstant* src_cst = source.GetConstant(); 1330 CPURegister temp; 1331 if (src_cst->IsIntConstant() || src_cst->IsNullConstant()) { 1332 temp = temps.AcquireW(); 1333 } else if (src_cst->IsLongConstant()) { 1334 temp = temps.AcquireX(); 1335 } else if (src_cst->IsFloatConstant()) { 1336 temp = temps.AcquireS(); 1337 } else { 1338 DCHECK(src_cst->IsDoubleConstant()); 1339 temp = temps.AcquireD(); 1340 } 1341 MoveConstant(temp, src_cst); 1342 __ Str(temp, StackOperandFrom(destination)); 1343 } else { 1344 DCHECK(source.IsStackSlot() || source.IsDoubleStackSlot()); 1345 DCHECK(source.IsDoubleStackSlot() == destination.IsDoubleStackSlot()); 1346 UseScratchRegisterScope temps(GetVIXLAssembler()); 1347 // There is generally less pressure on FP registers. 1348 FPRegister temp = destination.IsDoubleStackSlot() ? temps.AcquireD() : temps.AcquireS(); 1349 __ Ldr(temp, StackOperandFrom(source)); 1350 __ Str(temp, StackOperandFrom(destination)); 1351 } 1352 } 1353} 1354 1355void CodeGeneratorARM64::Load(Primitive::Type type, 1356 CPURegister dst, 1357 const MemOperand& src) { 1358 switch (type) { 1359 case Primitive::kPrimBoolean: 1360 __ Ldrb(Register(dst), src); 1361 break; 1362 case Primitive::kPrimByte: 1363 __ Ldrsb(Register(dst), src); 1364 break; 1365 case Primitive::kPrimShort: 1366 __ Ldrsh(Register(dst), src); 1367 break; 1368 case Primitive::kPrimChar: 1369 __ Ldrh(Register(dst), src); 1370 break; 1371 case Primitive::kPrimInt: 1372 case Primitive::kPrimNot: 1373 case Primitive::kPrimLong: 1374 case Primitive::kPrimFloat: 1375 case Primitive::kPrimDouble: 1376 DCHECK_EQ(dst.Is64Bits(), Primitive::Is64BitType(type)); 1377 __ Ldr(dst, src); 1378 break; 1379 case Primitive::kPrimVoid: 1380 LOG(FATAL) << "Unreachable type " << type; 1381 } 1382} 1383 1384void CodeGeneratorARM64::LoadAcquire(HInstruction* instruction, 1385 CPURegister dst, 1386 const MemOperand& src) { 1387 MacroAssembler* masm = GetVIXLAssembler(); 1388 BlockPoolsScope block_pools(masm); 1389 UseScratchRegisterScope temps(masm); 1390 Register temp_base = temps.AcquireX(); 1391 Primitive::Type type = instruction->GetType(); 1392 1393 DCHECK(!src.IsPreIndex()); 1394 DCHECK(!src.IsPostIndex()); 1395 1396 // TODO(vixl): Let the MacroAssembler handle MemOperand. 1397 __ Add(temp_base, src.base(), OperandFromMemOperand(src)); 1398 MemOperand base = MemOperand(temp_base); 1399 switch (type) { 1400 case Primitive::kPrimBoolean: 1401 __ Ldarb(Register(dst), base); 1402 MaybeRecordImplicitNullCheck(instruction); 1403 break; 1404 case Primitive::kPrimByte: 1405 __ Ldarb(Register(dst), base); 1406 MaybeRecordImplicitNullCheck(instruction); 1407 __ Sbfx(Register(dst), Register(dst), 0, Primitive::ComponentSize(type) * kBitsPerByte); 1408 break; 1409 case Primitive::kPrimChar: 1410 __ Ldarh(Register(dst), base); 1411 MaybeRecordImplicitNullCheck(instruction); 1412 break; 1413 case Primitive::kPrimShort: 1414 __ Ldarh(Register(dst), base); 1415 MaybeRecordImplicitNullCheck(instruction); 1416 __ Sbfx(Register(dst), Register(dst), 0, Primitive::ComponentSize(type) * kBitsPerByte); 1417 break; 1418 case Primitive::kPrimInt: 1419 case Primitive::kPrimNot: 1420 case Primitive::kPrimLong: 1421 DCHECK_EQ(dst.Is64Bits(), Primitive::Is64BitType(type)); 1422 __ Ldar(Register(dst), base); 1423 MaybeRecordImplicitNullCheck(instruction); 1424 break; 1425 case Primitive::kPrimFloat: 1426 case Primitive::kPrimDouble: { 1427 DCHECK(dst.IsFPRegister()); 1428 DCHECK_EQ(dst.Is64Bits(), Primitive::Is64BitType(type)); 1429 1430 Register temp = dst.Is64Bits() ? temps.AcquireX() : temps.AcquireW(); 1431 __ Ldar(temp, base); 1432 MaybeRecordImplicitNullCheck(instruction); 1433 __ Fmov(FPRegister(dst), temp); 1434 break; 1435 } 1436 case Primitive::kPrimVoid: 1437 LOG(FATAL) << "Unreachable type " << type; 1438 } 1439} 1440 1441void CodeGeneratorARM64::Store(Primitive::Type type, 1442 CPURegister src, 1443 const MemOperand& dst) { 1444 switch (type) { 1445 case Primitive::kPrimBoolean: 1446 case Primitive::kPrimByte: 1447 __ Strb(Register(src), dst); 1448 break; 1449 case Primitive::kPrimChar: 1450 case Primitive::kPrimShort: 1451 __ Strh(Register(src), dst); 1452 break; 1453 case Primitive::kPrimInt: 1454 case Primitive::kPrimNot: 1455 case Primitive::kPrimLong: 1456 case Primitive::kPrimFloat: 1457 case Primitive::kPrimDouble: 1458 DCHECK_EQ(src.Is64Bits(), Primitive::Is64BitType(type)); 1459 __ Str(src, dst); 1460 break; 1461 case Primitive::kPrimVoid: 1462 LOG(FATAL) << "Unreachable type " << type; 1463 } 1464} 1465 1466void CodeGeneratorARM64::StoreRelease(Primitive::Type type, 1467 CPURegister src, 1468 const MemOperand& dst) { 1469 UseScratchRegisterScope temps(GetVIXLAssembler()); 1470 Register temp_base = temps.AcquireX(); 1471 1472 DCHECK(!dst.IsPreIndex()); 1473 DCHECK(!dst.IsPostIndex()); 1474 1475 // TODO(vixl): Let the MacroAssembler handle this. 1476 Operand op = OperandFromMemOperand(dst); 1477 __ Add(temp_base, dst.base(), op); 1478 MemOperand base = MemOperand(temp_base); 1479 switch (type) { 1480 case Primitive::kPrimBoolean: 1481 case Primitive::kPrimByte: 1482 __ Stlrb(Register(src), base); 1483 break; 1484 case Primitive::kPrimChar: 1485 case Primitive::kPrimShort: 1486 __ Stlrh(Register(src), base); 1487 break; 1488 case Primitive::kPrimInt: 1489 case Primitive::kPrimNot: 1490 case Primitive::kPrimLong: 1491 DCHECK_EQ(src.Is64Bits(), Primitive::Is64BitType(type)); 1492 __ Stlr(Register(src), base); 1493 break; 1494 case Primitive::kPrimFloat: 1495 case Primitive::kPrimDouble: { 1496 DCHECK(src.IsFPRegister()); 1497 DCHECK_EQ(src.Is64Bits(), Primitive::Is64BitType(type)); 1498 1499 Register temp = src.Is64Bits() ? temps.AcquireX() : temps.AcquireW(); 1500 __ Fmov(temp, FPRegister(src)); 1501 __ Stlr(temp, base); 1502 break; 1503 } 1504 case Primitive::kPrimVoid: 1505 LOG(FATAL) << "Unreachable type " << type; 1506 } 1507} 1508 1509void CodeGeneratorARM64::InvokeRuntime(QuickEntrypointEnum entrypoint, 1510 HInstruction* instruction, 1511 uint32_t dex_pc, 1512 SlowPathCode* slow_path) { 1513 InvokeRuntime(GetThreadOffset<kArm64WordSize>(entrypoint).Int32Value(), 1514 instruction, 1515 dex_pc, 1516 slow_path); 1517} 1518 1519void CodeGeneratorARM64::InvokeRuntime(int32_t entry_point_offset, 1520 HInstruction* instruction, 1521 uint32_t dex_pc, 1522 SlowPathCode* slow_path) { 1523 ValidateInvokeRuntime(instruction, slow_path); 1524 BlockPoolsScope block_pools(GetVIXLAssembler()); 1525 __ Ldr(lr, MemOperand(tr, entry_point_offset)); 1526 __ Blr(lr); 1527 RecordPcInfo(instruction, dex_pc, slow_path); 1528} 1529 1530void InstructionCodeGeneratorARM64::GenerateClassInitializationCheck(SlowPathCodeARM64* slow_path, 1531 vixl::Register class_reg) { 1532 UseScratchRegisterScope temps(GetVIXLAssembler()); 1533 Register temp = temps.AcquireW(); 1534 size_t status_offset = mirror::Class::StatusOffset().SizeValue(); 1535 bool use_acquire_release = codegen_->GetInstructionSetFeatures().PreferAcquireRelease(); 1536 1537 // Even if the initialized flag is set, we need to ensure consistent memory ordering. 1538 if (use_acquire_release) { 1539 // TODO(vixl): Let the MacroAssembler handle MemOperand. 1540 __ Add(temp, class_reg, status_offset); 1541 __ Ldar(temp, HeapOperand(temp)); 1542 __ Cmp(temp, mirror::Class::kStatusInitialized); 1543 __ B(lt, slow_path->GetEntryLabel()); 1544 } else { 1545 __ Ldr(temp, HeapOperand(class_reg, status_offset)); 1546 __ Cmp(temp, mirror::Class::kStatusInitialized); 1547 __ B(lt, slow_path->GetEntryLabel()); 1548 __ Dmb(InnerShareable, BarrierReads); 1549 } 1550 __ Bind(slow_path->GetExitLabel()); 1551} 1552 1553void InstructionCodeGeneratorARM64::GenerateMemoryBarrier(MemBarrierKind kind) { 1554 BarrierType type = BarrierAll; 1555 1556 switch (kind) { 1557 case MemBarrierKind::kAnyAny: 1558 case MemBarrierKind::kAnyStore: { 1559 type = BarrierAll; 1560 break; 1561 } 1562 case MemBarrierKind::kLoadAny: { 1563 type = BarrierReads; 1564 break; 1565 } 1566 case MemBarrierKind::kStoreStore: { 1567 type = BarrierWrites; 1568 break; 1569 } 1570 default: 1571 LOG(FATAL) << "Unexpected memory barrier " << kind; 1572 } 1573 __ Dmb(InnerShareable, type); 1574} 1575 1576void InstructionCodeGeneratorARM64::GenerateSuspendCheck(HSuspendCheck* instruction, 1577 HBasicBlock* successor) { 1578 SuspendCheckSlowPathARM64* slow_path = 1579 down_cast<SuspendCheckSlowPathARM64*>(instruction->GetSlowPath()); 1580 if (slow_path == nullptr) { 1581 slow_path = new (GetGraph()->GetArena()) SuspendCheckSlowPathARM64(instruction, successor); 1582 instruction->SetSlowPath(slow_path); 1583 codegen_->AddSlowPath(slow_path); 1584 if (successor != nullptr) { 1585 DCHECK(successor->IsLoopHeader()); 1586 codegen_->ClearSpillSlotsFromLoopPhisInStackMap(instruction); 1587 } 1588 } else { 1589 DCHECK_EQ(slow_path->GetSuccessor(), successor); 1590 } 1591 1592 UseScratchRegisterScope temps(codegen_->GetVIXLAssembler()); 1593 Register temp = temps.AcquireW(); 1594 1595 __ Ldrh(temp, MemOperand(tr, Thread::ThreadFlagsOffset<kArm64WordSize>().SizeValue())); 1596 if (successor == nullptr) { 1597 __ Cbnz(temp, slow_path->GetEntryLabel()); 1598 __ Bind(slow_path->GetReturnLabel()); 1599 } else { 1600 __ Cbz(temp, codegen_->GetLabelOf(successor)); 1601 __ B(slow_path->GetEntryLabel()); 1602 // slow_path will return to GetLabelOf(successor). 1603 } 1604} 1605 1606InstructionCodeGeneratorARM64::InstructionCodeGeneratorARM64(HGraph* graph, 1607 CodeGeneratorARM64* codegen) 1608 : InstructionCodeGenerator(graph, codegen), 1609 assembler_(codegen->GetAssembler()), 1610 codegen_(codegen) {} 1611 1612#define FOR_EACH_UNIMPLEMENTED_INSTRUCTION(M) \ 1613 /* No unimplemented IR. */ 1614 1615#define UNIMPLEMENTED_INSTRUCTION_BREAK_CODE(name) name##UnimplementedInstructionBreakCode 1616 1617enum UnimplementedInstructionBreakCode { 1618 // Using a base helps identify when we hit such breakpoints. 1619 UnimplementedInstructionBreakCodeBaseCode = 0x900, 1620#define ENUM_UNIMPLEMENTED_INSTRUCTION(name) UNIMPLEMENTED_INSTRUCTION_BREAK_CODE(name), 1621 FOR_EACH_UNIMPLEMENTED_INSTRUCTION(ENUM_UNIMPLEMENTED_INSTRUCTION) 1622#undef ENUM_UNIMPLEMENTED_INSTRUCTION 1623}; 1624 1625#define DEFINE_UNIMPLEMENTED_INSTRUCTION_VISITORS(name) \ 1626 void InstructionCodeGeneratorARM64::Visit##name(H##name* instr ATTRIBUTE_UNUSED) { \ 1627 __ Brk(UNIMPLEMENTED_INSTRUCTION_BREAK_CODE(name)); \ 1628 } \ 1629 void LocationsBuilderARM64::Visit##name(H##name* instr) { \ 1630 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instr); \ 1631 locations->SetOut(Location::Any()); \ 1632 } 1633 FOR_EACH_UNIMPLEMENTED_INSTRUCTION(DEFINE_UNIMPLEMENTED_INSTRUCTION_VISITORS) 1634#undef DEFINE_UNIMPLEMENTED_INSTRUCTION_VISITORS 1635 1636#undef UNIMPLEMENTED_INSTRUCTION_BREAK_CODE 1637#undef FOR_EACH_UNIMPLEMENTED_INSTRUCTION 1638 1639void LocationsBuilderARM64::HandleBinaryOp(HBinaryOperation* instr) { 1640 DCHECK_EQ(instr->InputCount(), 2U); 1641 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instr); 1642 Primitive::Type type = instr->GetResultType(); 1643 switch (type) { 1644 case Primitive::kPrimInt: 1645 case Primitive::kPrimLong: 1646 locations->SetInAt(0, Location::RequiresRegister()); 1647 locations->SetInAt(1, ARM64EncodableConstantOrRegister(instr->InputAt(1), instr)); 1648 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 1649 break; 1650 1651 case Primitive::kPrimFloat: 1652 case Primitive::kPrimDouble: 1653 locations->SetInAt(0, Location::RequiresFpuRegister()); 1654 locations->SetInAt(1, Location::RequiresFpuRegister()); 1655 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 1656 break; 1657 1658 default: 1659 LOG(FATAL) << "Unexpected " << instr->DebugName() << " type " << type; 1660 } 1661} 1662 1663void LocationsBuilderARM64::HandleFieldGet(HInstruction* instruction) { 1664 DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet()); 1665 1666 bool object_field_get_with_read_barrier = 1667 kEmitCompilerReadBarrier && (instruction->GetType() == Primitive::kPrimNot); 1668 LocationSummary* locations = 1669 new (GetGraph()->GetArena()) LocationSummary(instruction, 1670 object_field_get_with_read_barrier ? 1671 LocationSummary::kCallOnSlowPath : 1672 LocationSummary::kNoCall); 1673 locations->SetInAt(0, Location::RequiresRegister()); 1674 if (Primitive::IsFloatingPointType(instruction->GetType())) { 1675 locations->SetOut(Location::RequiresFpuRegister()); 1676 } else { 1677 // The output overlaps for an object field get when read barriers 1678 // are enabled: we do not want the load to overwrite the object's 1679 // location, as we need it to emit the read barrier. 1680 locations->SetOut( 1681 Location::RequiresRegister(), 1682 object_field_get_with_read_barrier ? Location::kOutputOverlap : Location::kNoOutputOverlap); 1683 } 1684} 1685 1686void InstructionCodeGeneratorARM64::HandleFieldGet(HInstruction* instruction, 1687 const FieldInfo& field_info) { 1688 DCHECK(instruction->IsInstanceFieldGet() || instruction->IsStaticFieldGet()); 1689 Primitive::Type field_type = field_info.GetFieldType(); 1690 BlockPoolsScope block_pools(GetVIXLAssembler()); 1691 1692 MemOperand field = HeapOperand(InputRegisterAt(instruction, 0), field_info.GetFieldOffset()); 1693 bool use_acquire_release = codegen_->GetInstructionSetFeatures().PreferAcquireRelease(); 1694 1695 if (field_info.IsVolatile()) { 1696 if (use_acquire_release) { 1697 // NB: LoadAcquire will record the pc info if needed. 1698 codegen_->LoadAcquire(instruction, OutputCPURegister(instruction), field); 1699 } else { 1700 codegen_->Load(field_type, OutputCPURegister(instruction), field); 1701 codegen_->MaybeRecordImplicitNullCheck(instruction); 1702 // For IRIW sequential consistency kLoadAny is not sufficient. 1703 GenerateMemoryBarrier(MemBarrierKind::kAnyAny); 1704 } 1705 } else { 1706 codegen_->Load(field_type, OutputCPURegister(instruction), field); 1707 codegen_->MaybeRecordImplicitNullCheck(instruction); 1708 } 1709 1710 if (field_type == Primitive::kPrimNot) { 1711 LocationSummary* locations = instruction->GetLocations(); 1712 Location base = locations->InAt(0); 1713 Location out = locations->Out(); 1714 uint32_t offset = field_info.GetFieldOffset().Uint32Value(); 1715 codegen_->MaybeGenerateReadBarrier(instruction, out, out, base, offset); 1716 } 1717} 1718 1719void LocationsBuilderARM64::HandleFieldSet(HInstruction* instruction) { 1720 LocationSummary* locations = 1721 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 1722 locations->SetInAt(0, Location::RequiresRegister()); 1723 if (Primitive::IsFloatingPointType(instruction->InputAt(1)->GetType())) { 1724 locations->SetInAt(1, Location::RequiresFpuRegister()); 1725 } else { 1726 locations->SetInAt(1, Location::RequiresRegister()); 1727 } 1728} 1729 1730void InstructionCodeGeneratorARM64::HandleFieldSet(HInstruction* instruction, 1731 const FieldInfo& field_info, 1732 bool value_can_be_null) { 1733 DCHECK(instruction->IsInstanceFieldSet() || instruction->IsStaticFieldSet()); 1734 BlockPoolsScope block_pools(GetVIXLAssembler()); 1735 1736 Register obj = InputRegisterAt(instruction, 0); 1737 CPURegister value = InputCPURegisterAt(instruction, 1); 1738 CPURegister source = value; 1739 Offset offset = field_info.GetFieldOffset(); 1740 Primitive::Type field_type = field_info.GetFieldType(); 1741 bool use_acquire_release = codegen_->GetInstructionSetFeatures().PreferAcquireRelease(); 1742 1743 { 1744 // We use a block to end the scratch scope before the write barrier, thus 1745 // freeing the temporary registers so they can be used in `MarkGCCard`. 1746 UseScratchRegisterScope temps(GetVIXLAssembler()); 1747 1748 if (kPoisonHeapReferences && field_type == Primitive::kPrimNot) { 1749 DCHECK(value.IsW()); 1750 Register temp = temps.AcquireW(); 1751 __ Mov(temp, value.W()); 1752 GetAssembler()->PoisonHeapReference(temp.W()); 1753 source = temp; 1754 } 1755 1756 if (field_info.IsVolatile()) { 1757 if (use_acquire_release) { 1758 codegen_->StoreRelease(field_type, source, HeapOperand(obj, offset)); 1759 codegen_->MaybeRecordImplicitNullCheck(instruction); 1760 } else { 1761 GenerateMemoryBarrier(MemBarrierKind::kAnyStore); 1762 codegen_->Store(field_type, source, HeapOperand(obj, offset)); 1763 codegen_->MaybeRecordImplicitNullCheck(instruction); 1764 GenerateMemoryBarrier(MemBarrierKind::kAnyAny); 1765 } 1766 } else { 1767 codegen_->Store(field_type, source, HeapOperand(obj, offset)); 1768 codegen_->MaybeRecordImplicitNullCheck(instruction); 1769 } 1770 } 1771 1772 if (CodeGenerator::StoreNeedsWriteBarrier(field_type, instruction->InputAt(1))) { 1773 codegen_->MarkGCCard(obj, Register(value), value_can_be_null); 1774 } 1775} 1776 1777void InstructionCodeGeneratorARM64::HandleBinaryOp(HBinaryOperation* instr) { 1778 Primitive::Type type = instr->GetType(); 1779 1780 switch (type) { 1781 case Primitive::kPrimInt: 1782 case Primitive::kPrimLong: { 1783 Register dst = OutputRegister(instr); 1784 Register lhs = InputRegisterAt(instr, 0); 1785 Operand rhs = InputOperandAt(instr, 1); 1786 if (instr->IsAdd()) { 1787 __ Add(dst, lhs, rhs); 1788 } else if (instr->IsAnd()) { 1789 __ And(dst, lhs, rhs); 1790 } else if (instr->IsOr()) { 1791 __ Orr(dst, lhs, rhs); 1792 } else if (instr->IsSub()) { 1793 __ Sub(dst, lhs, rhs); 1794 } else if (instr->IsRor()) { 1795 if (rhs.IsImmediate()) { 1796 uint32_t shift = rhs.immediate() & (lhs.SizeInBits() - 1); 1797 __ Ror(dst, lhs, shift); 1798 } else { 1799 // Ensure shift distance is in the same size register as the result. If 1800 // we are rotating a long and the shift comes in a w register originally, 1801 // we don't need to sxtw for use as an x since the shift distances are 1802 // all & reg_bits - 1. 1803 __ Ror(dst, lhs, RegisterFrom(instr->GetLocations()->InAt(1), type)); 1804 } 1805 } else { 1806 DCHECK(instr->IsXor()); 1807 __ Eor(dst, lhs, rhs); 1808 } 1809 break; 1810 } 1811 case Primitive::kPrimFloat: 1812 case Primitive::kPrimDouble: { 1813 FPRegister dst = OutputFPRegister(instr); 1814 FPRegister lhs = InputFPRegisterAt(instr, 0); 1815 FPRegister rhs = InputFPRegisterAt(instr, 1); 1816 if (instr->IsAdd()) { 1817 __ Fadd(dst, lhs, rhs); 1818 } else if (instr->IsSub()) { 1819 __ Fsub(dst, lhs, rhs); 1820 } else { 1821 LOG(FATAL) << "Unexpected floating-point binary operation"; 1822 } 1823 break; 1824 } 1825 default: 1826 LOG(FATAL) << "Unexpected binary operation type " << type; 1827 } 1828} 1829 1830void LocationsBuilderARM64::HandleShift(HBinaryOperation* instr) { 1831 DCHECK(instr->IsShl() || instr->IsShr() || instr->IsUShr()); 1832 1833 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instr); 1834 Primitive::Type type = instr->GetResultType(); 1835 switch (type) { 1836 case Primitive::kPrimInt: 1837 case Primitive::kPrimLong: { 1838 locations->SetInAt(0, Location::RequiresRegister()); 1839 locations->SetInAt(1, Location::RegisterOrConstant(instr->InputAt(1))); 1840 locations->SetOut(Location::RequiresRegister()); 1841 break; 1842 } 1843 default: 1844 LOG(FATAL) << "Unexpected shift type " << type; 1845 } 1846} 1847 1848void InstructionCodeGeneratorARM64::HandleShift(HBinaryOperation* instr) { 1849 DCHECK(instr->IsShl() || instr->IsShr() || instr->IsUShr()); 1850 1851 Primitive::Type type = instr->GetType(); 1852 switch (type) { 1853 case Primitive::kPrimInt: 1854 case Primitive::kPrimLong: { 1855 Register dst = OutputRegister(instr); 1856 Register lhs = InputRegisterAt(instr, 0); 1857 Operand rhs = InputOperandAt(instr, 1); 1858 if (rhs.IsImmediate()) { 1859 uint32_t shift_value = (type == Primitive::kPrimInt) 1860 ? static_cast<uint32_t>(rhs.immediate() & kMaxIntShiftValue) 1861 : static_cast<uint32_t>(rhs.immediate() & kMaxLongShiftValue); 1862 if (instr->IsShl()) { 1863 __ Lsl(dst, lhs, shift_value); 1864 } else if (instr->IsShr()) { 1865 __ Asr(dst, lhs, shift_value); 1866 } else { 1867 __ Lsr(dst, lhs, shift_value); 1868 } 1869 } else { 1870 Register rhs_reg = dst.IsX() ? rhs.reg().X() : rhs.reg().W(); 1871 1872 if (instr->IsShl()) { 1873 __ Lsl(dst, lhs, rhs_reg); 1874 } else if (instr->IsShr()) { 1875 __ Asr(dst, lhs, rhs_reg); 1876 } else { 1877 __ Lsr(dst, lhs, rhs_reg); 1878 } 1879 } 1880 break; 1881 } 1882 default: 1883 LOG(FATAL) << "Unexpected shift operation type " << type; 1884 } 1885} 1886 1887void LocationsBuilderARM64::VisitAdd(HAdd* instruction) { 1888 HandleBinaryOp(instruction); 1889} 1890 1891void InstructionCodeGeneratorARM64::VisitAdd(HAdd* instruction) { 1892 HandleBinaryOp(instruction); 1893} 1894 1895void LocationsBuilderARM64::VisitAnd(HAnd* instruction) { 1896 HandleBinaryOp(instruction); 1897} 1898 1899void InstructionCodeGeneratorARM64::VisitAnd(HAnd* instruction) { 1900 HandleBinaryOp(instruction); 1901} 1902 1903void LocationsBuilderARM64::VisitArm64DataProcWithShifterOp( 1904 HArm64DataProcWithShifterOp* instruction) { 1905 DCHECK(instruction->GetType() == Primitive::kPrimInt || 1906 instruction->GetType() == Primitive::kPrimLong); 1907 LocationSummary* locations = 1908 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 1909 if (instruction->GetInstrKind() == HInstruction::kNeg) { 1910 locations->SetInAt(0, Location::ConstantLocation(instruction->InputAt(0)->AsConstant())); 1911 } else { 1912 locations->SetInAt(0, Location::RequiresRegister()); 1913 } 1914 locations->SetInAt(1, Location::RequiresRegister()); 1915 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 1916} 1917 1918void InstructionCodeGeneratorARM64::VisitArm64DataProcWithShifterOp( 1919 HArm64DataProcWithShifterOp* instruction) { 1920 Primitive::Type type = instruction->GetType(); 1921 HInstruction::InstructionKind kind = instruction->GetInstrKind(); 1922 DCHECK(type == Primitive::kPrimInt || type == Primitive::kPrimLong); 1923 Register out = OutputRegister(instruction); 1924 Register left; 1925 if (kind != HInstruction::kNeg) { 1926 left = InputRegisterAt(instruction, 0); 1927 } 1928 // If this `HArm64DataProcWithShifterOp` was created by merging a type conversion as the 1929 // shifter operand operation, the IR generating `right_reg` (input to the type 1930 // conversion) can have a different type from the current instruction's type, 1931 // so we manually indicate the type. 1932 Register right_reg = RegisterFrom(instruction->GetLocations()->InAt(1), type); 1933 int64_t shift_amount = (type == Primitive::kPrimInt) 1934 ? static_cast<uint32_t>(instruction->GetShiftAmount() & kMaxIntShiftValue) 1935 : static_cast<uint32_t>(instruction->GetShiftAmount() & kMaxLongShiftValue); 1936 1937 Operand right_operand(0); 1938 1939 HArm64DataProcWithShifterOp::OpKind op_kind = instruction->GetOpKind(); 1940 if (HArm64DataProcWithShifterOp::IsExtensionOp(op_kind)) { 1941 right_operand = Operand(right_reg, helpers::ExtendFromOpKind(op_kind)); 1942 } else { 1943 right_operand = Operand(right_reg, helpers::ShiftFromOpKind(op_kind), shift_amount); 1944 } 1945 1946 // Logical binary operations do not support extension operations in the 1947 // operand. Note that VIXL would still manage if it was passed by generating 1948 // the extension as a separate instruction. 1949 // `HNeg` also does not support extension. See comments in `ShifterOperandSupportsExtension()`. 1950 DCHECK(!right_operand.IsExtendedRegister() || 1951 (kind != HInstruction::kAnd && kind != HInstruction::kOr && kind != HInstruction::kXor && 1952 kind != HInstruction::kNeg)); 1953 switch (kind) { 1954 case HInstruction::kAdd: 1955 __ Add(out, left, right_operand); 1956 break; 1957 case HInstruction::kAnd: 1958 __ And(out, left, right_operand); 1959 break; 1960 case HInstruction::kNeg: 1961 DCHECK(instruction->InputAt(0)->AsConstant()->IsZero()); 1962 __ Neg(out, right_operand); 1963 break; 1964 case HInstruction::kOr: 1965 __ Orr(out, left, right_operand); 1966 break; 1967 case HInstruction::kSub: 1968 __ Sub(out, left, right_operand); 1969 break; 1970 case HInstruction::kXor: 1971 __ Eor(out, left, right_operand); 1972 break; 1973 default: 1974 LOG(FATAL) << "Unexpected operation kind: " << kind; 1975 UNREACHABLE(); 1976 } 1977} 1978 1979void LocationsBuilderARM64::VisitArm64IntermediateAddress(HArm64IntermediateAddress* instruction) { 1980 LocationSummary* locations = 1981 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 1982 locations->SetInAt(0, Location::RequiresRegister()); 1983 locations->SetInAt(1, ARM64EncodableConstantOrRegister(instruction->GetOffset(), instruction)); 1984 locations->SetOut(Location::RequiresRegister()); 1985} 1986 1987void InstructionCodeGeneratorARM64::VisitArm64IntermediateAddress( 1988 HArm64IntermediateAddress* instruction) { 1989 __ Add(OutputRegister(instruction), 1990 InputRegisterAt(instruction, 0), 1991 Operand(InputOperandAt(instruction, 1))); 1992} 1993 1994void LocationsBuilderARM64::VisitArm64MultiplyAccumulate(HArm64MultiplyAccumulate* instr) { 1995 LocationSummary* locations = 1996 new (GetGraph()->GetArena()) LocationSummary(instr, LocationSummary::kNoCall); 1997 locations->SetInAt(HArm64MultiplyAccumulate::kInputAccumulatorIndex, 1998 Location::RequiresRegister()); 1999 locations->SetInAt(HArm64MultiplyAccumulate::kInputMulLeftIndex, Location::RequiresRegister()); 2000 locations->SetInAt(HArm64MultiplyAccumulate::kInputMulRightIndex, Location::RequiresRegister()); 2001 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2002} 2003 2004void InstructionCodeGeneratorARM64::VisitArm64MultiplyAccumulate(HArm64MultiplyAccumulate* instr) { 2005 Register res = OutputRegister(instr); 2006 Register accumulator = InputRegisterAt(instr, HArm64MultiplyAccumulate::kInputAccumulatorIndex); 2007 Register mul_left = InputRegisterAt(instr, HArm64MultiplyAccumulate::kInputMulLeftIndex); 2008 Register mul_right = InputRegisterAt(instr, HArm64MultiplyAccumulate::kInputMulRightIndex); 2009 2010 // Avoid emitting code that could trigger Cortex A53's erratum 835769. 2011 // This fixup should be carried out for all multiply-accumulate instructions: 2012 // madd, msub, smaddl, smsubl, umaddl and umsubl. 2013 if (instr->GetType() == Primitive::kPrimLong && 2014 codegen_->GetInstructionSetFeatures().NeedFixCortexA53_835769()) { 2015 MacroAssembler* masm = down_cast<CodeGeneratorARM64*>(codegen_)->GetVIXLAssembler(); 2016 vixl::Instruction* prev = masm->GetCursorAddress<vixl::Instruction*>() - vixl::kInstructionSize; 2017 if (prev->IsLoadOrStore()) { 2018 // Make sure we emit only exactly one nop. 2019 vixl::CodeBufferCheckScope scope(masm, 2020 vixl::kInstructionSize, 2021 vixl::CodeBufferCheckScope::kCheck, 2022 vixl::CodeBufferCheckScope::kExactSize); 2023 __ nop(); 2024 } 2025 } 2026 2027 if (instr->GetOpKind() == HInstruction::kAdd) { 2028 __ Madd(res, mul_left, mul_right, accumulator); 2029 } else { 2030 DCHECK(instr->GetOpKind() == HInstruction::kSub); 2031 __ Msub(res, mul_left, mul_right, accumulator); 2032 } 2033} 2034 2035void LocationsBuilderARM64::VisitArrayGet(HArrayGet* instruction) { 2036 bool object_array_get_with_read_barrier = 2037 kEmitCompilerReadBarrier && (instruction->GetType() == Primitive::kPrimNot); 2038 LocationSummary* locations = 2039 new (GetGraph()->GetArena()) LocationSummary(instruction, 2040 object_array_get_with_read_barrier ? 2041 LocationSummary::kCallOnSlowPath : 2042 LocationSummary::kNoCall); 2043 locations->SetInAt(0, Location::RequiresRegister()); 2044 locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1))); 2045 if (Primitive::IsFloatingPointType(instruction->GetType())) { 2046 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2047 } else { 2048 // The output overlaps in the case of an object array get with 2049 // read barriers enabled: we do not want the move to overwrite the 2050 // array's location, as we need it to emit the read barrier. 2051 locations->SetOut( 2052 Location::RequiresRegister(), 2053 object_array_get_with_read_barrier ? Location::kOutputOverlap : Location::kNoOutputOverlap); 2054 } 2055} 2056 2057void InstructionCodeGeneratorARM64::VisitArrayGet(HArrayGet* instruction) { 2058 Primitive::Type type = instruction->GetType(); 2059 Register obj = InputRegisterAt(instruction, 0); 2060 LocationSummary* locations = instruction->GetLocations(); 2061 Location index = locations->InAt(1); 2062 uint32_t offset = mirror::Array::DataOffset(Primitive::ComponentSize(type)).Uint32Value(); 2063 MemOperand source = HeapOperand(obj); 2064 CPURegister dest = OutputCPURegister(instruction); 2065 2066 MacroAssembler* masm = GetVIXLAssembler(); 2067 UseScratchRegisterScope temps(masm); 2068 // Block pools between `Load` and `MaybeRecordImplicitNullCheck`. 2069 BlockPoolsScope block_pools(masm); 2070 2071 if (index.IsConstant()) { 2072 offset += Int64ConstantFrom(index) << Primitive::ComponentSizeShift(type); 2073 source = HeapOperand(obj, offset); 2074 } else { 2075 Register temp = temps.AcquireSameSizeAs(obj); 2076 if (instruction->GetArray()->IsArm64IntermediateAddress()) { 2077 // We do not need to compute the intermediate address from the array: the 2078 // input instruction has done it already. See the comment in 2079 // `InstructionSimplifierArm64::TryExtractArrayAccessAddress()`. 2080 if (kIsDebugBuild) { 2081 HArm64IntermediateAddress* tmp = instruction->GetArray()->AsArm64IntermediateAddress(); 2082 DCHECK(tmp->GetOffset()->AsIntConstant()->GetValueAsUint64() == offset); 2083 } 2084 temp = obj; 2085 } else { 2086 __ Add(temp, obj, offset); 2087 } 2088 source = HeapOperand(temp, XRegisterFrom(index), LSL, Primitive::ComponentSizeShift(type)); 2089 } 2090 2091 codegen_->Load(type, dest, source); 2092 codegen_->MaybeRecordImplicitNullCheck(instruction); 2093 2094 if (type == Primitive::kPrimNot) { 2095 static_assert( 2096 sizeof(mirror::HeapReference<mirror::Object>) == sizeof(int32_t), 2097 "art::mirror::HeapReference<art::mirror::Object> and int32_t have different sizes."); 2098 Location obj_loc = locations->InAt(0); 2099 Location out = locations->Out(); 2100 if (index.IsConstant()) { 2101 codegen_->MaybeGenerateReadBarrier(instruction, out, out, obj_loc, offset); 2102 } else { 2103 // Note: when `obj_loc` is a HArm64IntermediateAddress, it does 2104 // not contain the base address of the array object, which is 2105 // needed by the read barrier entry point. So the read barrier 2106 // slow path will temporarily set back `obj_loc` to the right 2107 // address (see ReadBarrierForHeapReferenceSlowPathARM64::EmitNativeCode). 2108 codegen_->MaybeGenerateReadBarrier(instruction, out, out, obj_loc, offset, index); 2109 } 2110 } 2111} 2112 2113void LocationsBuilderARM64::VisitArrayLength(HArrayLength* instruction) { 2114 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction); 2115 locations->SetInAt(0, Location::RequiresRegister()); 2116 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2117} 2118 2119void InstructionCodeGeneratorARM64::VisitArrayLength(HArrayLength* instruction) { 2120 BlockPoolsScope block_pools(GetVIXLAssembler()); 2121 __ Ldr(OutputRegister(instruction), 2122 HeapOperand(InputRegisterAt(instruction, 0), mirror::Array::LengthOffset())); 2123 codegen_->MaybeRecordImplicitNullCheck(instruction); 2124} 2125 2126void LocationsBuilderARM64::VisitArraySet(HArraySet* instruction) { 2127 Primitive::Type value_type = instruction->GetComponentType(); 2128 2129 bool may_need_runtime_call_for_type_check = instruction->NeedsTypeCheck(); 2130 bool object_array_set_with_read_barrier = 2131 kEmitCompilerReadBarrier && (value_type == Primitive::kPrimNot); 2132 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary( 2133 instruction, 2134 (may_need_runtime_call_for_type_check || object_array_set_with_read_barrier) ? 2135 LocationSummary::kCallOnSlowPath : 2136 LocationSummary::kNoCall); 2137 locations->SetInAt(0, Location::RequiresRegister()); 2138 locations->SetInAt(1, Location::RegisterOrConstant(instruction->InputAt(1))); 2139 if (Primitive::IsFloatingPointType(value_type)) { 2140 locations->SetInAt(2, Location::RequiresFpuRegister()); 2141 } else { 2142 locations->SetInAt(2, Location::RequiresRegister()); 2143 } 2144} 2145 2146void InstructionCodeGeneratorARM64::VisitArraySet(HArraySet* instruction) { 2147 Primitive::Type value_type = instruction->GetComponentType(); 2148 LocationSummary* locations = instruction->GetLocations(); 2149 bool may_need_runtime_call_for_type_check = instruction->NeedsTypeCheck(); 2150 bool needs_write_barrier = 2151 CodeGenerator::StoreNeedsWriteBarrier(value_type, instruction->GetValue()); 2152 2153 Register array = InputRegisterAt(instruction, 0); 2154 CPURegister value = InputCPURegisterAt(instruction, 2); 2155 CPURegister source = value; 2156 Location index = locations->InAt(1); 2157 size_t offset = mirror::Array::DataOffset(Primitive::ComponentSize(value_type)).Uint32Value(); 2158 MemOperand destination = HeapOperand(array); 2159 MacroAssembler* masm = GetVIXLAssembler(); 2160 BlockPoolsScope block_pools(masm); 2161 2162 if (!needs_write_barrier) { 2163 DCHECK(!may_need_runtime_call_for_type_check); 2164 if (index.IsConstant()) { 2165 offset += Int64ConstantFrom(index) << Primitive::ComponentSizeShift(value_type); 2166 destination = HeapOperand(array, offset); 2167 } else { 2168 UseScratchRegisterScope temps(masm); 2169 Register temp = temps.AcquireSameSizeAs(array); 2170 if (instruction->GetArray()->IsArm64IntermediateAddress()) { 2171 // We do not need to compute the intermediate address from the array: the 2172 // input instruction has done it already. See the comment in 2173 // `InstructionSimplifierArm64::TryExtractArrayAccessAddress()`. 2174 if (kIsDebugBuild) { 2175 HArm64IntermediateAddress* tmp = instruction->GetArray()->AsArm64IntermediateAddress(); 2176 DCHECK(tmp->GetOffset()->AsIntConstant()->GetValueAsUint64() == offset); 2177 } 2178 temp = array; 2179 } else { 2180 __ Add(temp, array, offset); 2181 } 2182 destination = HeapOperand(temp, 2183 XRegisterFrom(index), 2184 LSL, 2185 Primitive::ComponentSizeShift(value_type)); 2186 } 2187 codegen_->Store(value_type, value, destination); 2188 codegen_->MaybeRecordImplicitNullCheck(instruction); 2189 } else { 2190 DCHECK(needs_write_barrier); 2191 DCHECK(!instruction->GetArray()->IsArm64IntermediateAddress()); 2192 vixl::Label done; 2193 SlowPathCodeARM64* slow_path = nullptr; 2194 { 2195 // We use a block to end the scratch scope before the write barrier, thus 2196 // freeing the temporary registers so they can be used in `MarkGCCard`. 2197 UseScratchRegisterScope temps(masm); 2198 Register temp = temps.AcquireSameSizeAs(array); 2199 if (index.IsConstant()) { 2200 offset += Int64ConstantFrom(index) << Primitive::ComponentSizeShift(value_type); 2201 destination = HeapOperand(array, offset); 2202 } else { 2203 destination = HeapOperand(temp, 2204 XRegisterFrom(index), 2205 LSL, 2206 Primitive::ComponentSizeShift(value_type)); 2207 } 2208 2209 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 2210 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 2211 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 2212 2213 if (may_need_runtime_call_for_type_check) { 2214 slow_path = new (GetGraph()->GetArena()) ArraySetSlowPathARM64(instruction); 2215 codegen_->AddSlowPath(slow_path); 2216 if (instruction->GetValueCanBeNull()) { 2217 vixl::Label non_zero; 2218 __ Cbnz(Register(value), &non_zero); 2219 if (!index.IsConstant()) { 2220 __ Add(temp, array, offset); 2221 } 2222 __ Str(wzr, destination); 2223 codegen_->MaybeRecordImplicitNullCheck(instruction); 2224 __ B(&done); 2225 __ Bind(&non_zero); 2226 } 2227 2228 if (kEmitCompilerReadBarrier) { 2229 // When read barriers are enabled, the type checking 2230 // instrumentation requires two read barriers: 2231 // 2232 // __ Mov(temp2, temp); 2233 // // /* HeapReference<Class> */ temp = temp->component_type_ 2234 // __ Ldr(temp, HeapOperand(temp, component_offset)); 2235 // codegen_->GenerateReadBarrier( 2236 // instruction, temp_loc, temp_loc, temp2_loc, component_offset); 2237 // 2238 // // /* HeapReference<Class> */ temp2 = value->klass_ 2239 // __ Ldr(temp2, HeapOperand(Register(value), class_offset)); 2240 // codegen_->GenerateReadBarrier( 2241 // instruction, temp2_loc, temp2_loc, value_loc, class_offset, temp_loc); 2242 // 2243 // __ Cmp(temp, temp2); 2244 // 2245 // However, the second read barrier may trash `temp`, as it 2246 // is a temporary register, and as such would not be saved 2247 // along with live registers before calling the runtime (nor 2248 // restored afterwards). So in this case, we bail out and 2249 // delegate the work to the array set slow path. 2250 // 2251 // TODO: Extend the register allocator to support a new 2252 // "(locally) live temp" location so as to avoid always 2253 // going into the slow path when read barriers are enabled. 2254 __ B(slow_path->GetEntryLabel()); 2255 } else { 2256 Register temp2 = temps.AcquireSameSizeAs(array); 2257 // /* HeapReference<Class> */ temp = array->klass_ 2258 __ Ldr(temp, HeapOperand(array, class_offset)); 2259 codegen_->MaybeRecordImplicitNullCheck(instruction); 2260 GetAssembler()->MaybeUnpoisonHeapReference(temp); 2261 2262 // /* HeapReference<Class> */ temp = temp->component_type_ 2263 __ Ldr(temp, HeapOperand(temp, component_offset)); 2264 // /* HeapReference<Class> */ temp2 = value->klass_ 2265 __ Ldr(temp2, HeapOperand(Register(value), class_offset)); 2266 // If heap poisoning is enabled, no need to unpoison `temp` 2267 // nor `temp2`, as we are comparing two poisoned references. 2268 __ Cmp(temp, temp2); 2269 2270 if (instruction->StaticTypeOfArrayIsObjectArray()) { 2271 vixl::Label do_put; 2272 __ B(eq, &do_put); 2273 // If heap poisoning is enabled, the `temp` reference has 2274 // not been unpoisoned yet; unpoison it now. 2275 GetAssembler()->MaybeUnpoisonHeapReference(temp); 2276 2277 // /* HeapReference<Class> */ temp = temp->super_class_ 2278 __ Ldr(temp, HeapOperand(temp, super_offset)); 2279 // If heap poisoning is enabled, no need to unpoison 2280 // `temp`, as we are comparing against null below. 2281 __ Cbnz(temp, slow_path->GetEntryLabel()); 2282 __ Bind(&do_put); 2283 } else { 2284 __ B(ne, slow_path->GetEntryLabel()); 2285 } 2286 temps.Release(temp2); 2287 } 2288 } 2289 2290 if (kPoisonHeapReferences) { 2291 Register temp2 = temps.AcquireSameSizeAs(array); 2292 DCHECK(value.IsW()); 2293 __ Mov(temp2, value.W()); 2294 GetAssembler()->PoisonHeapReference(temp2); 2295 source = temp2; 2296 } 2297 2298 if (!index.IsConstant()) { 2299 __ Add(temp, array, offset); 2300 } 2301 __ Str(source, destination); 2302 2303 if (!may_need_runtime_call_for_type_check) { 2304 codegen_->MaybeRecordImplicitNullCheck(instruction); 2305 } 2306 } 2307 2308 codegen_->MarkGCCard(array, value.W(), instruction->GetValueCanBeNull()); 2309 2310 if (done.IsLinked()) { 2311 __ Bind(&done); 2312 } 2313 2314 if (slow_path != nullptr) { 2315 __ Bind(slow_path->GetExitLabel()); 2316 } 2317 } 2318} 2319 2320void LocationsBuilderARM64::VisitBoundsCheck(HBoundsCheck* instruction) { 2321 LocationSummary::CallKind call_kind = instruction->CanThrowIntoCatchBlock() 2322 ? LocationSummary::kCallOnSlowPath 2323 : LocationSummary::kNoCall; 2324 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 2325 locations->SetInAt(0, Location::RequiresRegister()); 2326 locations->SetInAt(1, ARM64EncodableConstantOrRegister(instruction->InputAt(1), instruction)); 2327 if (instruction->HasUses()) { 2328 locations->SetOut(Location::SameAsFirstInput()); 2329 } 2330} 2331 2332void InstructionCodeGeneratorARM64::VisitBoundsCheck(HBoundsCheck* instruction) { 2333 BoundsCheckSlowPathARM64* slow_path = 2334 new (GetGraph()->GetArena()) BoundsCheckSlowPathARM64(instruction); 2335 codegen_->AddSlowPath(slow_path); 2336 2337 __ Cmp(InputRegisterAt(instruction, 0), InputOperandAt(instruction, 1)); 2338 __ B(slow_path->GetEntryLabel(), hs); 2339} 2340 2341void LocationsBuilderARM64::VisitClinitCheck(HClinitCheck* check) { 2342 LocationSummary* locations = 2343 new (GetGraph()->GetArena()) LocationSummary(check, LocationSummary::kCallOnSlowPath); 2344 locations->SetInAt(0, Location::RequiresRegister()); 2345 if (check->HasUses()) { 2346 locations->SetOut(Location::SameAsFirstInput()); 2347 } 2348} 2349 2350void InstructionCodeGeneratorARM64::VisitClinitCheck(HClinitCheck* check) { 2351 // We assume the class is not null. 2352 SlowPathCodeARM64* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathARM64( 2353 check->GetLoadClass(), check, check->GetDexPc(), true); 2354 codegen_->AddSlowPath(slow_path); 2355 GenerateClassInitializationCheck(slow_path, InputRegisterAt(check, 0)); 2356} 2357 2358static bool IsFloatingPointZeroConstant(HInstruction* instruction) { 2359 return (instruction->IsFloatConstant() && (instruction->AsFloatConstant()->GetValue() == 0.0f)) 2360 || (instruction->IsDoubleConstant() && (instruction->AsDoubleConstant()->GetValue() == 0.0)); 2361} 2362 2363void LocationsBuilderARM64::VisitCompare(HCompare* compare) { 2364 LocationSummary* locations = 2365 new (GetGraph()->GetArena()) LocationSummary(compare, LocationSummary::kNoCall); 2366 Primitive::Type in_type = compare->InputAt(0)->GetType(); 2367 switch (in_type) { 2368 case Primitive::kPrimLong: { 2369 locations->SetInAt(0, Location::RequiresRegister()); 2370 locations->SetInAt(1, ARM64EncodableConstantOrRegister(compare->InputAt(1), compare)); 2371 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2372 break; 2373 } 2374 case Primitive::kPrimFloat: 2375 case Primitive::kPrimDouble: { 2376 locations->SetInAt(0, Location::RequiresFpuRegister()); 2377 locations->SetInAt(1, 2378 IsFloatingPointZeroConstant(compare->InputAt(1)) 2379 ? Location::ConstantLocation(compare->InputAt(1)->AsConstant()) 2380 : Location::RequiresFpuRegister()); 2381 locations->SetOut(Location::RequiresRegister()); 2382 break; 2383 } 2384 default: 2385 LOG(FATAL) << "Unexpected type for compare operation " << in_type; 2386 } 2387} 2388 2389void InstructionCodeGeneratorARM64::VisitCompare(HCompare* compare) { 2390 Primitive::Type in_type = compare->InputAt(0)->GetType(); 2391 2392 // 0 if: left == right 2393 // 1 if: left > right 2394 // -1 if: left < right 2395 switch (in_type) { 2396 case Primitive::kPrimLong: { 2397 Register result = OutputRegister(compare); 2398 Register left = InputRegisterAt(compare, 0); 2399 Operand right = InputOperandAt(compare, 1); 2400 2401 __ Cmp(left, right); 2402 __ Cset(result, ne); 2403 __ Cneg(result, result, lt); 2404 break; 2405 } 2406 case Primitive::kPrimFloat: 2407 case Primitive::kPrimDouble: { 2408 Register result = OutputRegister(compare); 2409 FPRegister left = InputFPRegisterAt(compare, 0); 2410 if (compare->GetLocations()->InAt(1).IsConstant()) { 2411 DCHECK(IsFloatingPointZeroConstant(compare->GetLocations()->InAt(1).GetConstant())); 2412 // 0.0 is the only immediate that can be encoded directly in an FCMP instruction. 2413 __ Fcmp(left, 0.0); 2414 } else { 2415 __ Fcmp(left, InputFPRegisterAt(compare, 1)); 2416 } 2417 if (compare->IsGtBias()) { 2418 __ Cset(result, ne); 2419 } else { 2420 __ Csetm(result, ne); 2421 } 2422 __ Cneg(result, result, compare->IsGtBias() ? mi : gt); 2423 break; 2424 } 2425 default: 2426 LOG(FATAL) << "Unimplemented compare type " << in_type; 2427 } 2428} 2429 2430void LocationsBuilderARM64::HandleCondition(HCondition* instruction) { 2431 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction); 2432 2433 if (Primitive::IsFloatingPointType(instruction->InputAt(0)->GetType())) { 2434 locations->SetInAt(0, Location::RequiresFpuRegister()); 2435 locations->SetInAt(1, 2436 IsFloatingPointZeroConstant(instruction->InputAt(1)) 2437 ? Location::ConstantLocation(instruction->InputAt(1)->AsConstant()) 2438 : Location::RequiresFpuRegister()); 2439 } else { 2440 // Integer cases. 2441 locations->SetInAt(0, Location::RequiresRegister()); 2442 locations->SetInAt(1, ARM64EncodableConstantOrRegister(instruction->InputAt(1), instruction)); 2443 } 2444 2445 if (instruction->NeedsMaterialization()) { 2446 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2447 } 2448} 2449 2450void InstructionCodeGeneratorARM64::HandleCondition(HCondition* instruction) { 2451 if (!instruction->NeedsMaterialization()) { 2452 return; 2453 } 2454 2455 LocationSummary* locations = instruction->GetLocations(); 2456 Register res = RegisterFrom(locations->Out(), instruction->GetType()); 2457 IfCondition if_cond = instruction->GetCondition(); 2458 Condition arm64_cond = ARM64Condition(if_cond); 2459 2460 if (Primitive::IsFloatingPointType(instruction->InputAt(0)->GetType())) { 2461 FPRegister lhs = InputFPRegisterAt(instruction, 0); 2462 if (locations->InAt(1).IsConstant()) { 2463 DCHECK(IsFloatingPointZeroConstant(locations->InAt(1).GetConstant())); 2464 // 0.0 is the only immediate that can be encoded directly in an FCMP instruction. 2465 __ Fcmp(lhs, 0.0); 2466 } else { 2467 __ Fcmp(lhs, InputFPRegisterAt(instruction, 1)); 2468 } 2469 __ Cset(res, arm64_cond); 2470 if (instruction->IsFPConditionTrueIfNaN()) { 2471 // res = IsUnordered(arm64_cond) ? 1 : res <=> res = IsNotUnordered(arm64_cond) ? res : 1 2472 __ Csel(res, res, Operand(1), vc); // VC for "not unordered". 2473 } else if (instruction->IsFPConditionFalseIfNaN()) { 2474 // res = IsUnordered(arm64_cond) ? 0 : res <=> res = IsNotUnordered(arm64_cond) ? res : 0 2475 __ Csel(res, res, Operand(0), vc); // VC for "not unordered". 2476 } 2477 } else { 2478 // Integer cases. 2479 Register lhs = InputRegisterAt(instruction, 0); 2480 Operand rhs = InputOperandAt(instruction, 1); 2481 __ Cmp(lhs, rhs); 2482 __ Cset(res, arm64_cond); 2483 } 2484} 2485 2486#define FOR_EACH_CONDITION_INSTRUCTION(M) \ 2487 M(Equal) \ 2488 M(NotEqual) \ 2489 M(LessThan) \ 2490 M(LessThanOrEqual) \ 2491 M(GreaterThan) \ 2492 M(GreaterThanOrEqual) \ 2493 M(Below) \ 2494 M(BelowOrEqual) \ 2495 M(Above) \ 2496 M(AboveOrEqual) 2497#define DEFINE_CONDITION_VISITORS(Name) \ 2498void LocationsBuilderARM64::Visit##Name(H##Name* comp) { HandleCondition(comp); } \ 2499void InstructionCodeGeneratorARM64::Visit##Name(H##Name* comp) { HandleCondition(comp); } 2500FOR_EACH_CONDITION_INSTRUCTION(DEFINE_CONDITION_VISITORS) 2501#undef DEFINE_CONDITION_VISITORS 2502#undef FOR_EACH_CONDITION_INSTRUCTION 2503 2504void InstructionCodeGeneratorARM64::DivRemOneOrMinusOne(HBinaryOperation* instruction) { 2505 DCHECK(instruction->IsDiv() || instruction->IsRem()); 2506 2507 LocationSummary* locations = instruction->GetLocations(); 2508 Location second = locations->InAt(1); 2509 DCHECK(second.IsConstant()); 2510 2511 Register out = OutputRegister(instruction); 2512 Register dividend = InputRegisterAt(instruction, 0); 2513 int64_t imm = Int64FromConstant(second.GetConstant()); 2514 DCHECK(imm == 1 || imm == -1); 2515 2516 if (instruction->IsRem()) { 2517 __ Mov(out, 0); 2518 } else { 2519 if (imm == 1) { 2520 __ Mov(out, dividend); 2521 } else { 2522 __ Neg(out, dividend); 2523 } 2524 } 2525} 2526 2527void InstructionCodeGeneratorARM64::DivRemByPowerOfTwo(HBinaryOperation* instruction) { 2528 DCHECK(instruction->IsDiv() || instruction->IsRem()); 2529 2530 LocationSummary* locations = instruction->GetLocations(); 2531 Location second = locations->InAt(1); 2532 DCHECK(second.IsConstant()); 2533 2534 Register out = OutputRegister(instruction); 2535 Register dividend = InputRegisterAt(instruction, 0); 2536 int64_t imm = Int64FromConstant(second.GetConstant()); 2537 uint64_t abs_imm = static_cast<uint64_t>(AbsOrMin(imm)); 2538 int ctz_imm = CTZ(abs_imm); 2539 2540 UseScratchRegisterScope temps(GetVIXLAssembler()); 2541 Register temp = temps.AcquireSameSizeAs(out); 2542 2543 if (instruction->IsDiv()) { 2544 __ Add(temp, dividend, abs_imm - 1); 2545 __ Cmp(dividend, 0); 2546 __ Csel(out, temp, dividend, lt); 2547 if (imm > 0) { 2548 __ Asr(out, out, ctz_imm); 2549 } else { 2550 __ Neg(out, Operand(out, ASR, ctz_imm)); 2551 } 2552 } else { 2553 int bits = instruction->GetResultType() == Primitive::kPrimInt ? 32 : 64; 2554 __ Asr(temp, dividend, bits - 1); 2555 __ Lsr(temp, temp, bits - ctz_imm); 2556 __ Add(out, dividend, temp); 2557 __ And(out, out, abs_imm - 1); 2558 __ Sub(out, out, temp); 2559 } 2560} 2561 2562void InstructionCodeGeneratorARM64::GenerateDivRemWithAnyConstant(HBinaryOperation* instruction) { 2563 DCHECK(instruction->IsDiv() || instruction->IsRem()); 2564 2565 LocationSummary* locations = instruction->GetLocations(); 2566 Location second = locations->InAt(1); 2567 DCHECK(second.IsConstant()); 2568 2569 Register out = OutputRegister(instruction); 2570 Register dividend = InputRegisterAt(instruction, 0); 2571 int64_t imm = Int64FromConstant(second.GetConstant()); 2572 2573 Primitive::Type type = instruction->GetResultType(); 2574 DCHECK(type == Primitive::kPrimInt || type == Primitive::kPrimLong); 2575 2576 int64_t magic; 2577 int shift; 2578 CalculateMagicAndShiftForDivRem(imm, type == Primitive::kPrimLong /* is_long */, &magic, &shift); 2579 2580 UseScratchRegisterScope temps(GetVIXLAssembler()); 2581 Register temp = temps.AcquireSameSizeAs(out); 2582 2583 // temp = get_high(dividend * magic) 2584 __ Mov(temp, magic); 2585 if (type == Primitive::kPrimLong) { 2586 __ Smulh(temp, dividend, temp); 2587 } else { 2588 __ Smull(temp.X(), dividend, temp); 2589 __ Lsr(temp.X(), temp.X(), 32); 2590 } 2591 2592 if (imm > 0 && magic < 0) { 2593 __ Add(temp, temp, dividend); 2594 } else if (imm < 0 && magic > 0) { 2595 __ Sub(temp, temp, dividend); 2596 } 2597 2598 if (shift != 0) { 2599 __ Asr(temp, temp, shift); 2600 } 2601 2602 if (instruction->IsDiv()) { 2603 __ Sub(out, temp, Operand(temp, ASR, type == Primitive::kPrimLong ? 63 : 31)); 2604 } else { 2605 __ Sub(temp, temp, Operand(temp, ASR, type == Primitive::kPrimLong ? 63 : 31)); 2606 // TODO: Strength reduction for msub. 2607 Register temp_imm = temps.AcquireSameSizeAs(out); 2608 __ Mov(temp_imm, imm); 2609 __ Msub(out, temp, temp_imm, dividend); 2610 } 2611} 2612 2613void InstructionCodeGeneratorARM64::GenerateDivRemIntegral(HBinaryOperation* instruction) { 2614 DCHECK(instruction->IsDiv() || instruction->IsRem()); 2615 Primitive::Type type = instruction->GetResultType(); 2616 DCHECK(type == Primitive::kPrimInt || Primitive::kPrimLong); 2617 2618 LocationSummary* locations = instruction->GetLocations(); 2619 Register out = OutputRegister(instruction); 2620 Location second = locations->InAt(1); 2621 2622 if (second.IsConstant()) { 2623 int64_t imm = Int64FromConstant(second.GetConstant()); 2624 2625 if (imm == 0) { 2626 // Do not generate anything. DivZeroCheck would prevent any code to be executed. 2627 } else if (imm == 1 || imm == -1) { 2628 DivRemOneOrMinusOne(instruction); 2629 } else if (IsPowerOfTwo(AbsOrMin(imm))) { 2630 DivRemByPowerOfTwo(instruction); 2631 } else { 2632 DCHECK(imm <= -2 || imm >= 2); 2633 GenerateDivRemWithAnyConstant(instruction); 2634 } 2635 } else { 2636 Register dividend = InputRegisterAt(instruction, 0); 2637 Register divisor = InputRegisterAt(instruction, 1); 2638 if (instruction->IsDiv()) { 2639 __ Sdiv(out, dividend, divisor); 2640 } else { 2641 UseScratchRegisterScope temps(GetVIXLAssembler()); 2642 Register temp = temps.AcquireSameSizeAs(out); 2643 __ Sdiv(temp, dividend, divisor); 2644 __ Msub(out, temp, divisor, dividend); 2645 } 2646 } 2647} 2648 2649void LocationsBuilderARM64::VisitDiv(HDiv* div) { 2650 LocationSummary* locations = 2651 new (GetGraph()->GetArena()) LocationSummary(div, LocationSummary::kNoCall); 2652 switch (div->GetResultType()) { 2653 case Primitive::kPrimInt: 2654 case Primitive::kPrimLong: 2655 locations->SetInAt(0, Location::RequiresRegister()); 2656 locations->SetInAt(1, Location::RegisterOrConstant(div->InputAt(1))); 2657 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 2658 break; 2659 2660 case Primitive::kPrimFloat: 2661 case Primitive::kPrimDouble: 2662 locations->SetInAt(0, Location::RequiresFpuRegister()); 2663 locations->SetInAt(1, Location::RequiresFpuRegister()); 2664 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 2665 break; 2666 2667 default: 2668 LOG(FATAL) << "Unexpected div type " << div->GetResultType(); 2669 } 2670} 2671 2672void InstructionCodeGeneratorARM64::VisitDiv(HDiv* div) { 2673 Primitive::Type type = div->GetResultType(); 2674 switch (type) { 2675 case Primitive::kPrimInt: 2676 case Primitive::kPrimLong: 2677 GenerateDivRemIntegral(div); 2678 break; 2679 2680 case Primitive::kPrimFloat: 2681 case Primitive::kPrimDouble: 2682 __ Fdiv(OutputFPRegister(div), InputFPRegisterAt(div, 0), InputFPRegisterAt(div, 1)); 2683 break; 2684 2685 default: 2686 LOG(FATAL) << "Unexpected div type " << type; 2687 } 2688} 2689 2690void LocationsBuilderARM64::VisitDivZeroCheck(HDivZeroCheck* instruction) { 2691 LocationSummary::CallKind call_kind = instruction->CanThrowIntoCatchBlock() 2692 ? LocationSummary::kCallOnSlowPath 2693 : LocationSummary::kNoCall; 2694 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 2695 locations->SetInAt(0, Location::RegisterOrConstant(instruction->InputAt(0))); 2696 if (instruction->HasUses()) { 2697 locations->SetOut(Location::SameAsFirstInput()); 2698 } 2699} 2700 2701void InstructionCodeGeneratorARM64::VisitDivZeroCheck(HDivZeroCheck* instruction) { 2702 SlowPathCodeARM64* slow_path = 2703 new (GetGraph()->GetArena()) DivZeroCheckSlowPathARM64(instruction); 2704 codegen_->AddSlowPath(slow_path); 2705 Location value = instruction->GetLocations()->InAt(0); 2706 2707 Primitive::Type type = instruction->GetType(); 2708 2709 if ((type == Primitive::kPrimBoolean) || !Primitive::IsIntegralType(type)) { 2710 LOG(FATAL) << "Unexpected type " << type << " for DivZeroCheck."; 2711 return; 2712 } 2713 2714 if (value.IsConstant()) { 2715 int64_t divisor = Int64ConstantFrom(value); 2716 if (divisor == 0) { 2717 __ B(slow_path->GetEntryLabel()); 2718 } else { 2719 // A division by a non-null constant is valid. We don't need to perform 2720 // any check, so simply fall through. 2721 } 2722 } else { 2723 __ Cbz(InputRegisterAt(instruction, 0), slow_path->GetEntryLabel()); 2724 } 2725} 2726 2727void LocationsBuilderARM64::VisitDoubleConstant(HDoubleConstant* constant) { 2728 LocationSummary* locations = 2729 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 2730 locations->SetOut(Location::ConstantLocation(constant)); 2731} 2732 2733void InstructionCodeGeneratorARM64::VisitDoubleConstant( 2734 HDoubleConstant* constant ATTRIBUTE_UNUSED) { 2735 // Will be generated at use site. 2736} 2737 2738void LocationsBuilderARM64::VisitExit(HExit* exit) { 2739 exit->SetLocations(nullptr); 2740} 2741 2742void InstructionCodeGeneratorARM64::VisitExit(HExit* exit ATTRIBUTE_UNUSED) { 2743} 2744 2745void LocationsBuilderARM64::VisitFloatConstant(HFloatConstant* constant) { 2746 LocationSummary* locations = 2747 new (GetGraph()->GetArena()) LocationSummary(constant, LocationSummary::kNoCall); 2748 locations->SetOut(Location::ConstantLocation(constant)); 2749} 2750 2751void InstructionCodeGeneratorARM64::VisitFloatConstant(HFloatConstant* constant ATTRIBUTE_UNUSED) { 2752 // Will be generated at use site. 2753} 2754 2755void InstructionCodeGeneratorARM64::HandleGoto(HInstruction* got, HBasicBlock* successor) { 2756 DCHECK(!successor->IsExitBlock()); 2757 HBasicBlock* block = got->GetBlock(); 2758 HInstruction* previous = got->GetPrevious(); 2759 HLoopInformation* info = block->GetLoopInformation(); 2760 2761 if (info != nullptr && info->IsBackEdge(*block) && info->HasSuspendCheck()) { 2762 codegen_->ClearSpillSlotsFromLoopPhisInStackMap(info->GetSuspendCheck()); 2763 GenerateSuspendCheck(info->GetSuspendCheck(), successor); 2764 return; 2765 } 2766 if (block->IsEntryBlock() && (previous != nullptr) && previous->IsSuspendCheck()) { 2767 GenerateSuspendCheck(previous->AsSuspendCheck(), nullptr); 2768 } 2769 if (!codegen_->GoesToNextBlock(block, successor)) { 2770 __ B(codegen_->GetLabelOf(successor)); 2771 } 2772} 2773 2774void LocationsBuilderARM64::VisitGoto(HGoto* got) { 2775 got->SetLocations(nullptr); 2776} 2777 2778void InstructionCodeGeneratorARM64::VisitGoto(HGoto* got) { 2779 HandleGoto(got, got->GetSuccessor()); 2780} 2781 2782void LocationsBuilderARM64::VisitTryBoundary(HTryBoundary* try_boundary) { 2783 try_boundary->SetLocations(nullptr); 2784} 2785 2786void InstructionCodeGeneratorARM64::VisitTryBoundary(HTryBoundary* try_boundary) { 2787 HBasicBlock* successor = try_boundary->GetNormalFlowSuccessor(); 2788 if (!successor->IsExitBlock()) { 2789 HandleGoto(try_boundary, successor); 2790 } 2791} 2792 2793void InstructionCodeGeneratorARM64::GenerateTestAndBranch(HInstruction* instruction, 2794 size_t condition_input_index, 2795 vixl::Label* true_target, 2796 vixl::Label* false_target) { 2797 // FP branching requires both targets to be explicit. If either of the targets 2798 // is nullptr (fallthrough) use and bind `fallthrough_target` instead. 2799 vixl::Label fallthrough_target; 2800 HInstruction* cond = instruction->InputAt(condition_input_index); 2801 2802 if (true_target == nullptr && false_target == nullptr) { 2803 // Nothing to do. The code always falls through. 2804 return; 2805 } else if (cond->IsIntConstant()) { 2806 // Constant condition, statically compared against 1. 2807 if (cond->AsIntConstant()->IsOne()) { 2808 if (true_target != nullptr) { 2809 __ B(true_target); 2810 } 2811 } else { 2812 DCHECK(cond->AsIntConstant()->IsZero()); 2813 if (false_target != nullptr) { 2814 __ B(false_target); 2815 } 2816 } 2817 return; 2818 } 2819 2820 // The following code generates these patterns: 2821 // (1) true_target == nullptr && false_target != nullptr 2822 // - opposite condition true => branch to false_target 2823 // (2) true_target != nullptr && false_target == nullptr 2824 // - condition true => branch to true_target 2825 // (3) true_target != nullptr && false_target != nullptr 2826 // - condition true => branch to true_target 2827 // - branch to false_target 2828 if (IsBooleanValueOrMaterializedCondition(cond)) { 2829 // The condition instruction has been materialized, compare the output to 0. 2830 Location cond_val = instruction->GetLocations()->InAt(condition_input_index); 2831 DCHECK(cond_val.IsRegister()); 2832 if (true_target == nullptr) { 2833 __ Cbz(InputRegisterAt(instruction, condition_input_index), false_target); 2834 } else { 2835 __ Cbnz(InputRegisterAt(instruction, condition_input_index), true_target); 2836 } 2837 } else { 2838 // The condition instruction has not been materialized, use its inputs as 2839 // the comparison and its condition as the branch condition. 2840 HCondition* condition = cond->AsCondition(); 2841 2842 Primitive::Type type = condition->InputAt(0)->GetType(); 2843 if (Primitive::IsFloatingPointType(type)) { 2844 FPRegister lhs = InputFPRegisterAt(condition, 0); 2845 if (condition->GetLocations()->InAt(1).IsConstant()) { 2846 DCHECK(IsFloatingPointZeroConstant(condition->GetLocations()->InAt(1).GetConstant())); 2847 // 0.0 is the only immediate that can be encoded directly in an FCMP instruction. 2848 __ Fcmp(lhs, 0.0); 2849 } else { 2850 __ Fcmp(lhs, InputFPRegisterAt(condition, 1)); 2851 } 2852 if (condition->IsFPConditionTrueIfNaN()) { 2853 __ B(vs, true_target == nullptr ? &fallthrough_target : true_target); 2854 } else if (condition->IsFPConditionFalseIfNaN()) { 2855 __ B(vs, false_target == nullptr ? &fallthrough_target : false_target); 2856 } 2857 if (true_target == nullptr) { 2858 __ B(ARM64Condition(condition->GetOppositeCondition()), false_target); 2859 } else { 2860 __ B(ARM64Condition(condition->GetCondition()), true_target); 2861 } 2862 } else { 2863 // Integer cases. 2864 Register lhs = InputRegisterAt(condition, 0); 2865 Operand rhs = InputOperandAt(condition, 1); 2866 2867 Condition arm64_cond; 2868 vixl::Label* non_fallthrough_target; 2869 if (true_target == nullptr) { 2870 arm64_cond = ARM64Condition(condition->GetOppositeCondition()); 2871 non_fallthrough_target = false_target; 2872 } else { 2873 arm64_cond = ARM64Condition(condition->GetCondition()); 2874 non_fallthrough_target = true_target; 2875 } 2876 2877 if ((arm64_cond != gt && arm64_cond != le) && rhs.IsImmediate() && (rhs.immediate() == 0)) { 2878 switch (arm64_cond) { 2879 case eq: 2880 __ Cbz(lhs, non_fallthrough_target); 2881 break; 2882 case ne: 2883 __ Cbnz(lhs, non_fallthrough_target); 2884 break; 2885 case lt: 2886 // Test the sign bit and branch accordingly. 2887 __ Tbnz(lhs, (lhs.IsX() ? kXRegSize : kWRegSize) - 1, non_fallthrough_target); 2888 break; 2889 case ge: 2890 // Test the sign bit and branch accordingly. 2891 __ Tbz(lhs, (lhs.IsX() ? kXRegSize : kWRegSize) - 1, non_fallthrough_target); 2892 break; 2893 default: 2894 // Without the `static_cast` the compiler throws an error for 2895 // `-Werror=sign-promo`. 2896 LOG(FATAL) << "Unexpected condition: " << static_cast<int>(arm64_cond); 2897 } 2898 } else { 2899 __ Cmp(lhs, rhs); 2900 __ B(arm64_cond, non_fallthrough_target); 2901 } 2902 } 2903 } 2904 2905 // If neither branch falls through (case 3), the conditional branch to `true_target` 2906 // was already emitted (case 2) and we need to emit a jump to `false_target`. 2907 if (true_target != nullptr && false_target != nullptr) { 2908 __ B(false_target); 2909 } 2910 2911 if (fallthrough_target.IsLinked()) { 2912 __ Bind(&fallthrough_target); 2913 } 2914} 2915 2916void LocationsBuilderARM64::VisitIf(HIf* if_instr) { 2917 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(if_instr); 2918 if (IsBooleanValueOrMaterializedCondition(if_instr->InputAt(0))) { 2919 locations->SetInAt(0, Location::RequiresRegister()); 2920 } 2921} 2922 2923void InstructionCodeGeneratorARM64::VisitIf(HIf* if_instr) { 2924 HBasicBlock* true_successor = if_instr->IfTrueSuccessor(); 2925 HBasicBlock* false_successor = if_instr->IfFalseSuccessor(); 2926 vixl::Label* true_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), true_successor) ? 2927 nullptr : codegen_->GetLabelOf(true_successor); 2928 vixl::Label* false_target = codegen_->GoesToNextBlock(if_instr->GetBlock(), false_successor) ? 2929 nullptr : codegen_->GetLabelOf(false_successor); 2930 GenerateTestAndBranch(if_instr, /* condition_input_index */ 0, true_target, false_target); 2931} 2932 2933void LocationsBuilderARM64::VisitDeoptimize(HDeoptimize* deoptimize) { 2934 LocationSummary* locations = new (GetGraph()->GetArena()) 2935 LocationSummary(deoptimize, LocationSummary::kCallOnSlowPath); 2936 if (IsBooleanValueOrMaterializedCondition(deoptimize->InputAt(0))) { 2937 locations->SetInAt(0, Location::RequiresRegister()); 2938 } 2939} 2940 2941void InstructionCodeGeneratorARM64::VisitDeoptimize(HDeoptimize* deoptimize) { 2942 SlowPathCodeARM64* slow_path = 2943 deopt_slow_paths_.NewSlowPath<DeoptimizationSlowPathARM64>(deoptimize); 2944 GenerateTestAndBranch(deoptimize, 2945 /* condition_input_index */ 0, 2946 slow_path->GetEntryLabel(), 2947 /* false_target */ nullptr); 2948} 2949 2950void LocationsBuilderARM64::VisitNativeDebugInfo(HNativeDebugInfo* info) { 2951 new (GetGraph()->GetArena()) LocationSummary(info); 2952} 2953 2954void InstructionCodeGeneratorARM64::VisitNativeDebugInfo(HNativeDebugInfo* info) { 2955 if (codegen_->HasStackMapAtCurrentPc()) { 2956 // Ensure that we do not collide with the stack map of the previous instruction. 2957 __ Nop(); 2958 } 2959 codegen_->RecordPcInfo(info, info->GetDexPc()); 2960} 2961 2962void LocationsBuilderARM64::VisitInstanceFieldGet(HInstanceFieldGet* instruction) { 2963 HandleFieldGet(instruction); 2964} 2965 2966void InstructionCodeGeneratorARM64::VisitInstanceFieldGet(HInstanceFieldGet* instruction) { 2967 HandleFieldGet(instruction, instruction->GetFieldInfo()); 2968} 2969 2970void LocationsBuilderARM64::VisitInstanceFieldSet(HInstanceFieldSet* instruction) { 2971 HandleFieldSet(instruction); 2972} 2973 2974void InstructionCodeGeneratorARM64::VisitInstanceFieldSet(HInstanceFieldSet* instruction) { 2975 HandleFieldSet(instruction, instruction->GetFieldInfo(), instruction->GetValueCanBeNull()); 2976} 2977 2978void LocationsBuilderARM64::VisitInstanceOf(HInstanceOf* instruction) { 2979 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 2980 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 2981 switch (type_check_kind) { 2982 case TypeCheckKind::kExactCheck: 2983 case TypeCheckKind::kAbstractClassCheck: 2984 case TypeCheckKind::kClassHierarchyCheck: 2985 case TypeCheckKind::kArrayObjectCheck: 2986 call_kind = 2987 kEmitCompilerReadBarrier ? LocationSummary::kCallOnSlowPath : LocationSummary::kNoCall; 2988 break; 2989 case TypeCheckKind::kArrayCheck: 2990 case TypeCheckKind::kUnresolvedCheck: 2991 case TypeCheckKind::kInterfaceCheck: 2992 call_kind = LocationSummary::kCallOnSlowPath; 2993 break; 2994 } 2995 2996 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 2997 locations->SetInAt(0, Location::RequiresRegister()); 2998 locations->SetInAt(1, Location::RequiresRegister()); 2999 // The "out" register is used as a temporary, so it overlaps with the inputs. 3000 // Note that TypeCheckSlowPathARM64 uses this register too. 3001 locations->SetOut(Location::RequiresRegister(), Location::kOutputOverlap); 3002 // When read barriers are enabled, we need a temporary register for 3003 // some cases. 3004 if (kEmitCompilerReadBarrier && 3005 (type_check_kind == TypeCheckKind::kAbstractClassCheck || 3006 type_check_kind == TypeCheckKind::kClassHierarchyCheck || 3007 type_check_kind == TypeCheckKind::kArrayObjectCheck)) { 3008 locations->AddTemp(Location::RequiresRegister()); 3009 } 3010} 3011 3012void InstructionCodeGeneratorARM64::VisitInstanceOf(HInstanceOf* instruction) { 3013 LocationSummary* locations = instruction->GetLocations(); 3014 Location obj_loc = locations->InAt(0); 3015 Register obj = InputRegisterAt(instruction, 0); 3016 Register cls = InputRegisterAt(instruction, 1); 3017 Location out_loc = locations->Out(); 3018 Register out = OutputRegister(instruction); 3019 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 3020 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 3021 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 3022 uint32_t primitive_offset = mirror::Class::PrimitiveTypeOffset().Int32Value(); 3023 3024 vixl::Label done, zero; 3025 SlowPathCodeARM64* slow_path = nullptr; 3026 3027 // Return 0 if `obj` is null. 3028 // Avoid null check if we know `obj` is not null. 3029 if (instruction->MustDoNullCheck()) { 3030 __ Cbz(obj, &zero); 3031 } 3032 3033 // /* HeapReference<Class> */ out = obj->klass_ 3034 __ Ldr(out, HeapOperand(obj.W(), class_offset)); 3035 codegen_->MaybeGenerateReadBarrier(instruction, out_loc, out_loc, obj_loc, class_offset); 3036 3037 switch (instruction->GetTypeCheckKind()) { 3038 case TypeCheckKind::kExactCheck: { 3039 __ Cmp(out, cls); 3040 __ Cset(out, eq); 3041 if (zero.IsLinked()) { 3042 __ B(&done); 3043 } 3044 break; 3045 } 3046 3047 case TypeCheckKind::kAbstractClassCheck: { 3048 // If the class is abstract, we eagerly fetch the super class of the 3049 // object to avoid doing a comparison we know will fail. 3050 vixl::Label loop, success; 3051 __ Bind(&loop); 3052 Location temp_loc = kEmitCompilerReadBarrier ? locations->GetTemp(0) : Location::NoLocation(); 3053 if (kEmitCompilerReadBarrier) { 3054 // Save the value of `out` into `temp` before overwriting it 3055 // in the following move operation, as we will need it for the 3056 // read barrier below. 3057 Register temp = WRegisterFrom(temp_loc); 3058 __ Mov(temp, out); 3059 } 3060 // /* HeapReference<Class> */ out = out->super_class_ 3061 __ Ldr(out, HeapOperand(out, super_offset)); 3062 codegen_->MaybeGenerateReadBarrier(instruction, out_loc, out_loc, temp_loc, super_offset); 3063 // If `out` is null, we use it for the result, and jump to `done`. 3064 __ Cbz(out, &done); 3065 __ Cmp(out, cls); 3066 __ B(ne, &loop); 3067 __ Mov(out, 1); 3068 if (zero.IsLinked()) { 3069 __ B(&done); 3070 } 3071 break; 3072 } 3073 3074 case TypeCheckKind::kClassHierarchyCheck: { 3075 // Walk over the class hierarchy to find a match. 3076 vixl::Label loop, success; 3077 __ Bind(&loop); 3078 __ Cmp(out, cls); 3079 __ B(eq, &success); 3080 Location temp_loc = kEmitCompilerReadBarrier ? locations->GetTemp(0) : Location::NoLocation(); 3081 if (kEmitCompilerReadBarrier) { 3082 // Save the value of `out` into `temp` before overwriting it 3083 // in the following move operation, as we will need it for the 3084 // read barrier below. 3085 Register temp = WRegisterFrom(temp_loc); 3086 __ Mov(temp, out); 3087 } 3088 // /* HeapReference<Class> */ out = out->super_class_ 3089 __ Ldr(out, HeapOperand(out, super_offset)); 3090 codegen_->MaybeGenerateReadBarrier(instruction, out_loc, out_loc, temp_loc, super_offset); 3091 __ Cbnz(out, &loop); 3092 // If `out` is null, we use it for the result, and jump to `done`. 3093 __ B(&done); 3094 __ Bind(&success); 3095 __ Mov(out, 1); 3096 if (zero.IsLinked()) { 3097 __ B(&done); 3098 } 3099 break; 3100 } 3101 3102 case TypeCheckKind::kArrayObjectCheck: { 3103 // Do an exact check. 3104 vixl::Label exact_check; 3105 __ Cmp(out, cls); 3106 __ B(eq, &exact_check); 3107 // Otherwise, we need to check that the object's class is a non-primitive array. 3108 Location temp_loc = kEmitCompilerReadBarrier ? locations->GetTemp(0) : Location::NoLocation(); 3109 if (kEmitCompilerReadBarrier) { 3110 // Save the value of `out` into `temp` before overwriting it 3111 // in the following move operation, as we will need it for the 3112 // read barrier below. 3113 Register temp = WRegisterFrom(temp_loc); 3114 __ Mov(temp, out); 3115 } 3116 // /* HeapReference<Class> */ out = out->component_type_ 3117 __ Ldr(out, HeapOperand(out, component_offset)); 3118 codegen_->MaybeGenerateReadBarrier(instruction, out_loc, out_loc, temp_loc, component_offset); 3119 // If `out` is null, we use it for the result, and jump to `done`. 3120 __ Cbz(out, &done); 3121 __ Ldrh(out, HeapOperand(out, primitive_offset)); 3122 static_assert(Primitive::kPrimNot == 0, "Expected 0 for kPrimNot"); 3123 __ Cbnz(out, &zero); 3124 __ Bind(&exact_check); 3125 __ Mov(out, 1); 3126 __ B(&done); 3127 break; 3128 } 3129 3130 case TypeCheckKind::kArrayCheck: { 3131 __ Cmp(out, cls); 3132 DCHECK(locations->OnlyCallsOnSlowPath()); 3133 slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathARM64(instruction, 3134 /* is_fatal */ false); 3135 codegen_->AddSlowPath(slow_path); 3136 __ B(ne, slow_path->GetEntryLabel()); 3137 __ Mov(out, 1); 3138 if (zero.IsLinked()) { 3139 __ B(&done); 3140 } 3141 break; 3142 } 3143 3144 case TypeCheckKind::kUnresolvedCheck: 3145 case TypeCheckKind::kInterfaceCheck: { 3146 // Note that we indeed only call on slow path, but we always go 3147 // into the slow path for the unresolved and interface check 3148 // cases. 3149 // 3150 // We cannot directly call the InstanceofNonTrivial runtime 3151 // entry point without resorting to a type checking slow path 3152 // here (i.e. by calling InvokeRuntime directly), as it would 3153 // require to assign fixed registers for the inputs of this 3154 // HInstanceOf instruction (following the runtime calling 3155 // convention), which might be cluttered by the potential first 3156 // read barrier emission at the beginning of this method. 3157 DCHECK(locations->OnlyCallsOnSlowPath()); 3158 slow_path = new (GetGraph()->GetArena()) TypeCheckSlowPathARM64(instruction, 3159 /* is_fatal */ false); 3160 codegen_->AddSlowPath(slow_path); 3161 __ B(slow_path->GetEntryLabel()); 3162 if (zero.IsLinked()) { 3163 __ B(&done); 3164 } 3165 break; 3166 } 3167 } 3168 3169 if (zero.IsLinked()) { 3170 __ Bind(&zero); 3171 __ Mov(out, 0); 3172 } 3173 3174 if (done.IsLinked()) { 3175 __ Bind(&done); 3176 } 3177 3178 if (slow_path != nullptr) { 3179 __ Bind(slow_path->GetExitLabel()); 3180 } 3181} 3182 3183void LocationsBuilderARM64::VisitCheckCast(HCheckCast* instruction) { 3184 LocationSummary::CallKind call_kind = LocationSummary::kNoCall; 3185 bool throws_into_catch = instruction->CanThrowIntoCatchBlock(); 3186 3187 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 3188 switch (type_check_kind) { 3189 case TypeCheckKind::kExactCheck: 3190 case TypeCheckKind::kAbstractClassCheck: 3191 case TypeCheckKind::kClassHierarchyCheck: 3192 case TypeCheckKind::kArrayObjectCheck: 3193 call_kind = (throws_into_catch || kEmitCompilerReadBarrier) ? 3194 LocationSummary::kCallOnSlowPath : 3195 LocationSummary::kNoCall; // In fact, call on a fatal (non-returning) slow path. 3196 break; 3197 case TypeCheckKind::kArrayCheck: 3198 case TypeCheckKind::kUnresolvedCheck: 3199 case TypeCheckKind::kInterfaceCheck: 3200 call_kind = LocationSummary::kCallOnSlowPath; 3201 break; 3202 } 3203 3204 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 3205 locations->SetInAt(0, Location::RequiresRegister()); 3206 locations->SetInAt(1, Location::RequiresRegister()); 3207 // Note that TypeCheckSlowPathARM64 uses this "temp" register too. 3208 locations->AddTemp(Location::RequiresRegister()); 3209 locations->AddTemp(Location::RequiresRegister()); 3210 // When read barriers are enabled, we need an additional temporary 3211 // register for some cases. 3212 if (kEmitCompilerReadBarrier && 3213 (type_check_kind == TypeCheckKind::kAbstractClassCheck || 3214 type_check_kind == TypeCheckKind::kClassHierarchyCheck || 3215 type_check_kind == TypeCheckKind::kArrayObjectCheck)) { 3216 locations->AddTemp(Location::RequiresRegister()); 3217 } 3218} 3219 3220void InstructionCodeGeneratorARM64::VisitCheckCast(HCheckCast* instruction) { 3221 LocationSummary* locations = instruction->GetLocations(); 3222 Location obj_loc = locations->InAt(0); 3223 Register obj = InputRegisterAt(instruction, 0); 3224 Register cls = InputRegisterAt(instruction, 1); 3225 Location temp_loc = locations->GetTemp(0); 3226 Register temp = WRegisterFrom(temp_loc); 3227 uint32_t class_offset = mirror::Object::ClassOffset().Int32Value(); 3228 uint32_t super_offset = mirror::Class::SuperClassOffset().Int32Value(); 3229 uint32_t component_offset = mirror::Class::ComponentTypeOffset().Int32Value(); 3230 uint32_t primitive_offset = mirror::Class::PrimitiveTypeOffset().Int32Value(); 3231 3232 TypeCheckKind type_check_kind = instruction->GetTypeCheckKind(); 3233 bool is_type_check_slow_path_fatal = 3234 (type_check_kind == TypeCheckKind::kExactCheck || 3235 type_check_kind == TypeCheckKind::kAbstractClassCheck || 3236 type_check_kind == TypeCheckKind::kClassHierarchyCheck || 3237 type_check_kind == TypeCheckKind::kArrayObjectCheck) && 3238 !instruction->CanThrowIntoCatchBlock(); 3239 SlowPathCodeARM64* type_check_slow_path = 3240 new (GetGraph()->GetArena()) TypeCheckSlowPathARM64(instruction, 3241 is_type_check_slow_path_fatal); 3242 codegen_->AddSlowPath(type_check_slow_path); 3243 3244 vixl::Label done; 3245 // Avoid null check if we know obj is not null. 3246 if (instruction->MustDoNullCheck()) { 3247 __ Cbz(obj, &done); 3248 } 3249 3250 // /* HeapReference<Class> */ temp = obj->klass_ 3251 __ Ldr(temp, HeapOperand(obj, class_offset)); 3252 codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset); 3253 3254 switch (type_check_kind) { 3255 case TypeCheckKind::kExactCheck: 3256 case TypeCheckKind::kArrayCheck: { 3257 __ Cmp(temp, cls); 3258 // Jump to slow path for throwing the exception or doing a 3259 // more involved array check. 3260 __ B(ne, type_check_slow_path->GetEntryLabel()); 3261 break; 3262 } 3263 3264 case TypeCheckKind::kAbstractClassCheck: { 3265 // If the class is abstract, we eagerly fetch the super class of the 3266 // object to avoid doing a comparison we know will fail. 3267 vixl::Label loop, compare_classes; 3268 __ Bind(&loop); 3269 Location temp2_loc = 3270 kEmitCompilerReadBarrier ? locations->GetTemp(1) : Location::NoLocation(); 3271 if (kEmitCompilerReadBarrier) { 3272 // Save the value of `temp` into `temp2` before overwriting it 3273 // in the following move operation, as we will need it for the 3274 // read barrier below. 3275 Register temp2 = WRegisterFrom(temp2_loc); 3276 __ Mov(temp2, temp); 3277 } 3278 // /* HeapReference<Class> */ temp = temp->super_class_ 3279 __ Ldr(temp, HeapOperand(temp, super_offset)); 3280 codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, temp2_loc, super_offset); 3281 3282 // If the class reference currently in `temp` is not null, jump 3283 // to the `compare_classes` label to compare it with the checked 3284 // class. 3285 __ Cbnz(temp, &compare_classes); 3286 // Otherwise, jump to the slow path to throw the exception. 3287 // 3288 // But before, move back the object's class into `temp` before 3289 // going into the slow path, as it has been overwritten in the 3290 // meantime. 3291 // /* HeapReference<Class> */ temp = obj->klass_ 3292 __ Ldr(temp, HeapOperand(obj, class_offset)); 3293 codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset); 3294 __ B(type_check_slow_path->GetEntryLabel()); 3295 3296 __ Bind(&compare_classes); 3297 __ Cmp(temp, cls); 3298 __ B(ne, &loop); 3299 break; 3300 } 3301 3302 case TypeCheckKind::kClassHierarchyCheck: { 3303 // Walk over the class hierarchy to find a match. 3304 vixl::Label loop; 3305 __ Bind(&loop); 3306 __ Cmp(temp, cls); 3307 __ B(eq, &done); 3308 3309 Location temp2_loc = 3310 kEmitCompilerReadBarrier ? locations->GetTemp(1) : Location::NoLocation(); 3311 if (kEmitCompilerReadBarrier) { 3312 // Save the value of `temp` into `temp2` before overwriting it 3313 // in the following move operation, as we will need it for the 3314 // read barrier below. 3315 Register temp2 = WRegisterFrom(temp2_loc); 3316 __ Mov(temp2, temp); 3317 } 3318 // /* HeapReference<Class> */ temp = temp->super_class_ 3319 __ Ldr(temp, HeapOperand(temp, super_offset)); 3320 codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, temp2_loc, super_offset); 3321 3322 // If the class reference currently in `temp` is not null, jump 3323 // back at the beginning of the loop. 3324 __ Cbnz(temp, &loop); 3325 // Otherwise, jump to the slow path to throw the exception. 3326 // 3327 // But before, move back the object's class into `temp` before 3328 // going into the slow path, as it has been overwritten in the 3329 // meantime. 3330 // /* HeapReference<Class> */ temp = obj->klass_ 3331 __ Ldr(temp, HeapOperand(obj, class_offset)); 3332 codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset); 3333 __ B(type_check_slow_path->GetEntryLabel()); 3334 break; 3335 } 3336 3337 case TypeCheckKind::kArrayObjectCheck: { 3338 // Do an exact check. 3339 vixl::Label check_non_primitive_component_type; 3340 __ Cmp(temp, cls); 3341 __ B(eq, &done); 3342 3343 // Otherwise, we need to check that the object's class is a non-primitive array. 3344 Location temp2_loc = 3345 kEmitCompilerReadBarrier ? locations->GetTemp(1) : Location::NoLocation(); 3346 if (kEmitCompilerReadBarrier) { 3347 // Save the value of `temp` into `temp2` before overwriting it 3348 // in the following move operation, as we will need it for the 3349 // read barrier below. 3350 Register temp2 = WRegisterFrom(temp2_loc); 3351 __ Mov(temp2, temp); 3352 } 3353 // /* HeapReference<Class> */ temp = temp->component_type_ 3354 __ Ldr(temp, HeapOperand(temp, component_offset)); 3355 codegen_->MaybeGenerateReadBarrier( 3356 instruction, temp_loc, temp_loc, temp2_loc, component_offset); 3357 3358 // If the component type is not null (i.e. the object is indeed 3359 // an array), jump to label `check_non_primitive_component_type` 3360 // to further check that this component type is not a primitive 3361 // type. 3362 __ Cbnz(temp, &check_non_primitive_component_type); 3363 // Otherwise, jump to the slow path to throw the exception. 3364 // 3365 // But before, move back the object's class into `temp` before 3366 // going into the slow path, as it has been overwritten in the 3367 // meantime. 3368 // /* HeapReference<Class> */ temp = obj->klass_ 3369 __ Ldr(temp, HeapOperand(obj, class_offset)); 3370 codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset); 3371 __ B(type_check_slow_path->GetEntryLabel()); 3372 3373 __ Bind(&check_non_primitive_component_type); 3374 __ Ldrh(temp, HeapOperand(temp, primitive_offset)); 3375 static_assert(Primitive::kPrimNot == 0, "Expected 0 for kPrimNot"); 3376 __ Cbz(temp, &done); 3377 // Same comment as above regarding `temp` and the slow path. 3378 // /* HeapReference<Class> */ temp = obj->klass_ 3379 __ Ldr(temp, HeapOperand(obj, class_offset)); 3380 codegen_->MaybeGenerateReadBarrier(instruction, temp_loc, temp_loc, obj_loc, class_offset); 3381 __ B(type_check_slow_path->GetEntryLabel()); 3382 break; 3383 } 3384 3385 case TypeCheckKind::kUnresolvedCheck: 3386 case TypeCheckKind::kInterfaceCheck: 3387 // We always go into the type check slow path for the unresolved 3388 // and interface check cases. 3389 // 3390 // We cannot directly call the CheckCast runtime entry point 3391 // without resorting to a type checking slow path here (i.e. by 3392 // calling InvokeRuntime directly), as it would require to 3393 // assign fixed registers for the inputs of this HInstanceOf 3394 // instruction (following the runtime calling convention), which 3395 // might be cluttered by the potential first read barrier 3396 // emission at the beginning of this method. 3397 __ B(type_check_slow_path->GetEntryLabel()); 3398 break; 3399 } 3400 __ Bind(&done); 3401 3402 __ Bind(type_check_slow_path->GetExitLabel()); 3403} 3404 3405void LocationsBuilderARM64::VisitIntConstant(HIntConstant* constant) { 3406 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(constant); 3407 locations->SetOut(Location::ConstantLocation(constant)); 3408} 3409 3410void InstructionCodeGeneratorARM64::VisitIntConstant(HIntConstant* constant ATTRIBUTE_UNUSED) { 3411 // Will be generated at use site. 3412} 3413 3414void LocationsBuilderARM64::VisitNullConstant(HNullConstant* constant) { 3415 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(constant); 3416 locations->SetOut(Location::ConstantLocation(constant)); 3417} 3418 3419void InstructionCodeGeneratorARM64::VisitNullConstant(HNullConstant* constant ATTRIBUTE_UNUSED) { 3420 // Will be generated at use site. 3421} 3422 3423void LocationsBuilderARM64::VisitInvokeUnresolved(HInvokeUnresolved* invoke) { 3424 // The trampoline uses the same calling convention as dex calling conventions, 3425 // except instead of loading arg0/r0 with the target Method*, arg0/r0 will contain 3426 // the method_idx. 3427 HandleInvoke(invoke); 3428} 3429 3430void InstructionCodeGeneratorARM64::VisitInvokeUnresolved(HInvokeUnresolved* invoke) { 3431 codegen_->GenerateInvokeUnresolvedRuntimeCall(invoke); 3432} 3433 3434void LocationsBuilderARM64::HandleInvoke(HInvoke* invoke) { 3435 InvokeDexCallingConventionVisitorARM64 calling_convention_visitor; 3436 CodeGenerator::CreateCommonInvokeLocationSummary(invoke, &calling_convention_visitor); 3437} 3438 3439void LocationsBuilderARM64::VisitInvokeInterface(HInvokeInterface* invoke) { 3440 HandleInvoke(invoke); 3441} 3442 3443void InstructionCodeGeneratorARM64::VisitInvokeInterface(HInvokeInterface* invoke) { 3444 // TODO: b/18116999, our IMTs can miss an IncompatibleClassChangeError. 3445 LocationSummary* locations = invoke->GetLocations(); 3446 Register temp = XRegisterFrom(locations->GetTemp(0)); 3447 uint32_t method_offset = mirror::Class::EmbeddedImTableEntryOffset( 3448 invoke->GetImtIndex() % mirror::Class::kImtSize, kArm64PointerSize).Uint32Value(); 3449 Location receiver = locations->InAt(0); 3450 Offset class_offset = mirror::Object::ClassOffset(); 3451 Offset entry_point = ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArm64WordSize); 3452 3453 // The register ip1 is required to be used for the hidden argument in 3454 // art_quick_imt_conflict_trampoline, so prevent VIXL from using it. 3455 MacroAssembler* masm = GetVIXLAssembler(); 3456 UseScratchRegisterScope scratch_scope(masm); 3457 BlockPoolsScope block_pools(masm); 3458 scratch_scope.Exclude(ip1); 3459 __ Mov(ip1, invoke->GetDexMethodIndex()); 3460 3461 if (receiver.IsStackSlot()) { 3462 __ Ldr(temp.W(), StackOperandFrom(receiver)); 3463 // /* HeapReference<Class> */ temp = temp->klass_ 3464 __ Ldr(temp.W(), HeapOperand(temp.W(), class_offset)); 3465 } else { 3466 // /* HeapReference<Class> */ temp = receiver->klass_ 3467 __ Ldr(temp.W(), HeapOperandFrom(receiver, class_offset)); 3468 } 3469 codegen_->MaybeRecordImplicitNullCheck(invoke); 3470 // Instead of simply (possibly) unpoisoning `temp` here, we should 3471 // emit a read barrier for the previous class reference load. 3472 // However this is not required in practice, as this is an 3473 // intermediate/temporary reference and because the current 3474 // concurrent copying collector keeps the from-space memory 3475 // intact/accessible until the end of the marking phase (the 3476 // concurrent copying collector may not in the future). 3477 GetAssembler()->MaybeUnpoisonHeapReference(temp.W()); 3478 // temp = temp->GetImtEntryAt(method_offset); 3479 __ Ldr(temp, MemOperand(temp, method_offset)); 3480 // lr = temp->GetEntryPoint(); 3481 __ Ldr(lr, MemOperand(temp, entry_point.Int32Value())); 3482 // lr(); 3483 __ Blr(lr); 3484 DCHECK(!codegen_->IsLeafMethod()); 3485 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 3486} 3487 3488void LocationsBuilderARM64::VisitInvokeVirtual(HInvokeVirtual* invoke) { 3489 IntrinsicLocationsBuilderARM64 intrinsic(GetGraph()->GetArena()); 3490 if (intrinsic.TryDispatch(invoke)) { 3491 return; 3492 } 3493 3494 HandleInvoke(invoke); 3495} 3496 3497void LocationsBuilderARM64::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) { 3498 // When we do not run baseline, explicit clinit checks triggered by static 3499 // invokes must have been pruned by art::PrepareForRegisterAllocation. 3500 DCHECK(codegen_->IsBaseline() || !invoke->IsStaticWithExplicitClinitCheck()); 3501 3502 IntrinsicLocationsBuilderARM64 intrinsic(GetGraph()->GetArena()); 3503 if (intrinsic.TryDispatch(invoke)) { 3504 return; 3505 } 3506 3507 HandleInvoke(invoke); 3508} 3509 3510static bool TryGenerateIntrinsicCode(HInvoke* invoke, CodeGeneratorARM64* codegen) { 3511 if (invoke->GetLocations()->Intrinsified()) { 3512 IntrinsicCodeGeneratorARM64 intrinsic(codegen); 3513 intrinsic.Dispatch(invoke); 3514 return true; 3515 } 3516 return false; 3517} 3518 3519HInvokeStaticOrDirect::DispatchInfo CodeGeneratorARM64::GetSupportedInvokeStaticOrDirectDispatch( 3520 const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info, 3521 MethodReference target_method ATTRIBUTE_UNUSED) { 3522 // On arm64 we support all dispatch types. 3523 return desired_dispatch_info; 3524} 3525 3526void CodeGeneratorARM64::GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) { 3527 // For better instruction scheduling we load the direct code pointer before the method pointer. 3528 bool direct_code_loaded = false; 3529 switch (invoke->GetCodePtrLocation()) { 3530 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup: 3531 // LR = code address from literal pool with link-time patch. 3532 __ Ldr(lr, DeduplicateMethodCodeLiteral(invoke->GetTargetMethod())); 3533 direct_code_loaded = true; 3534 break; 3535 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect: 3536 // LR = invoke->GetDirectCodePtr(); 3537 __ Ldr(lr, DeduplicateUint64Literal(invoke->GetDirectCodePtr())); 3538 direct_code_loaded = true; 3539 break; 3540 default: 3541 break; 3542 } 3543 3544 // Make sure that ArtMethod* is passed in kArtMethodRegister as per the calling convention. 3545 Location callee_method = temp; // For all kinds except kRecursive, callee will be in temp. 3546 switch (invoke->GetMethodLoadKind()) { 3547 case HInvokeStaticOrDirect::MethodLoadKind::kStringInit: 3548 // temp = thread->string_init_entrypoint 3549 __ Ldr(XRegisterFrom(temp), MemOperand(tr, invoke->GetStringInitOffset())); 3550 break; 3551 case HInvokeStaticOrDirect::MethodLoadKind::kRecursive: 3552 callee_method = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 3553 break; 3554 case HInvokeStaticOrDirect::MethodLoadKind::kDirectAddress: 3555 // Load method address from literal pool. 3556 __ Ldr(XRegisterFrom(temp), DeduplicateUint64Literal(invoke->GetMethodAddress())); 3557 break; 3558 case HInvokeStaticOrDirect::MethodLoadKind::kDirectAddressWithFixup: 3559 // Load method address from literal pool with a link-time patch. 3560 __ Ldr(XRegisterFrom(temp), 3561 DeduplicateMethodAddressLiteral(invoke->GetTargetMethod())); 3562 break; 3563 case HInvokeStaticOrDirect::MethodLoadKind::kDexCachePcRelative: { 3564 // Add ADRP with its PC-relative DexCache access patch. 3565 pc_relative_dex_cache_patches_.emplace_back(*invoke->GetTargetMethod().dex_file, 3566 invoke->GetDexCacheArrayOffset()); 3567 vixl::Label* pc_insn_label = &pc_relative_dex_cache_patches_.back().label; 3568 { 3569 vixl::SingleEmissionCheckScope guard(GetVIXLAssembler()); 3570 __ Bind(pc_insn_label); 3571 __ adrp(XRegisterFrom(temp), 0); 3572 } 3573 pc_relative_dex_cache_patches_.back().pc_insn_label = pc_insn_label; 3574 // Add LDR with its PC-relative DexCache access patch. 3575 pc_relative_dex_cache_patches_.emplace_back(*invoke->GetTargetMethod().dex_file, 3576 invoke->GetDexCacheArrayOffset()); 3577 { 3578 vixl::SingleEmissionCheckScope guard(GetVIXLAssembler()); 3579 __ Bind(&pc_relative_dex_cache_patches_.back().label); 3580 __ ldr(XRegisterFrom(temp), MemOperand(XRegisterFrom(temp), 0)); 3581 pc_relative_dex_cache_patches_.back().pc_insn_label = pc_insn_label; 3582 } 3583 break; 3584 } 3585 case HInvokeStaticOrDirect::MethodLoadKind::kDexCacheViaMethod: { 3586 Location current_method = invoke->GetLocations()->InAt(invoke->GetSpecialInputIndex()); 3587 Register reg = XRegisterFrom(temp); 3588 Register method_reg; 3589 if (current_method.IsRegister()) { 3590 method_reg = XRegisterFrom(current_method); 3591 } else { 3592 DCHECK(invoke->GetLocations()->Intrinsified()); 3593 DCHECK(!current_method.IsValid()); 3594 method_reg = reg; 3595 __ Ldr(reg.X(), MemOperand(sp, kCurrentMethodStackOffset)); 3596 } 3597 3598 // /* ArtMethod*[] */ temp = temp.ptr_sized_fields_->dex_cache_resolved_methods_; 3599 __ Ldr(reg.X(), 3600 MemOperand(method_reg.X(), 3601 ArtMethod::DexCacheResolvedMethodsOffset(kArm64WordSize).Int32Value())); 3602 // temp = temp[index_in_cache]; 3603 uint32_t index_in_cache = invoke->GetTargetMethod().dex_method_index; 3604 __ Ldr(reg.X(), MemOperand(reg.X(), GetCachePointerOffset(index_in_cache))); 3605 break; 3606 } 3607 } 3608 3609 switch (invoke->GetCodePtrLocation()) { 3610 case HInvokeStaticOrDirect::CodePtrLocation::kCallSelf: 3611 __ Bl(&frame_entry_label_); 3612 break; 3613 case HInvokeStaticOrDirect::CodePtrLocation::kCallPCRelative: { 3614 relative_call_patches_.emplace_back(invoke->GetTargetMethod()); 3615 vixl::Label* label = &relative_call_patches_.back().label; 3616 vixl::SingleEmissionCheckScope guard(GetVIXLAssembler()); 3617 __ Bind(label); 3618 __ bl(0); // Branch and link to itself. This will be overriden at link time. 3619 break; 3620 } 3621 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirectWithFixup: 3622 case HInvokeStaticOrDirect::CodePtrLocation::kCallDirect: 3623 // LR prepared above for better instruction scheduling. 3624 DCHECK(direct_code_loaded); 3625 // lr() 3626 __ Blr(lr); 3627 break; 3628 case HInvokeStaticOrDirect::CodePtrLocation::kCallArtMethod: 3629 // LR = callee_method->entry_point_from_quick_compiled_code_; 3630 __ Ldr(lr, MemOperand( 3631 XRegisterFrom(callee_method), 3632 ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArm64WordSize).Int32Value())); 3633 // lr() 3634 __ Blr(lr); 3635 break; 3636 } 3637 3638 DCHECK(!IsLeafMethod()); 3639} 3640 3641void CodeGeneratorARM64::GenerateVirtualCall(HInvokeVirtual* invoke, Location temp_in) { 3642 // Use the calling convention instead of the location of the receiver, as 3643 // intrinsics may have put the receiver in a different register. In the intrinsics 3644 // slow path, the arguments have been moved to the right place, so here we are 3645 // guaranteed that the receiver is the first register of the calling convention. 3646 InvokeDexCallingConvention calling_convention; 3647 Register receiver = calling_convention.GetRegisterAt(0); 3648 Register temp = XRegisterFrom(temp_in); 3649 size_t method_offset = mirror::Class::EmbeddedVTableEntryOffset( 3650 invoke->GetVTableIndex(), kArm64PointerSize).SizeValue(); 3651 Offset class_offset = mirror::Object::ClassOffset(); 3652 Offset entry_point = ArtMethod::EntryPointFromQuickCompiledCodeOffset(kArm64WordSize); 3653 3654 BlockPoolsScope block_pools(GetVIXLAssembler()); 3655 3656 DCHECK(receiver.IsRegister()); 3657 // /* HeapReference<Class> */ temp = receiver->klass_ 3658 __ Ldr(temp.W(), HeapOperandFrom(LocationFrom(receiver), class_offset)); 3659 MaybeRecordImplicitNullCheck(invoke); 3660 // Instead of simply (possibly) unpoisoning `temp` here, we should 3661 // emit a read barrier for the previous class reference load. 3662 // intermediate/temporary reference and because the current 3663 // concurrent copying collector keeps the from-space memory 3664 // intact/accessible until the end of the marking phase (the 3665 // concurrent copying collector may not in the future). 3666 GetAssembler()->MaybeUnpoisonHeapReference(temp.W()); 3667 // temp = temp->GetMethodAt(method_offset); 3668 __ Ldr(temp, MemOperand(temp, method_offset)); 3669 // lr = temp->GetEntryPoint(); 3670 __ Ldr(lr, MemOperand(temp, entry_point.SizeValue())); 3671 // lr(); 3672 __ Blr(lr); 3673} 3674 3675void CodeGeneratorARM64::EmitLinkerPatches(ArenaVector<LinkerPatch>* linker_patches) { 3676 DCHECK(linker_patches->empty()); 3677 size_t size = 3678 method_patches_.size() + 3679 call_patches_.size() + 3680 relative_call_patches_.size() + 3681 pc_relative_dex_cache_patches_.size(); 3682 linker_patches->reserve(size); 3683 for (const auto& entry : method_patches_) { 3684 const MethodReference& target_method = entry.first; 3685 vixl::Literal<uint64_t>* literal = entry.second; 3686 linker_patches->push_back(LinkerPatch::MethodPatch(literal->offset(), 3687 target_method.dex_file, 3688 target_method.dex_method_index)); 3689 } 3690 for (const auto& entry : call_patches_) { 3691 const MethodReference& target_method = entry.first; 3692 vixl::Literal<uint64_t>* literal = entry.second; 3693 linker_patches->push_back(LinkerPatch::CodePatch(literal->offset(), 3694 target_method.dex_file, 3695 target_method.dex_method_index)); 3696 } 3697 for (const MethodPatchInfo<vixl::Label>& info : relative_call_patches_) { 3698 linker_patches->push_back(LinkerPatch::RelativeCodePatch(info.label.location(), 3699 info.target_method.dex_file, 3700 info.target_method.dex_method_index)); 3701 } 3702 for (const PcRelativeDexCacheAccessInfo& info : pc_relative_dex_cache_patches_) { 3703 linker_patches->push_back(LinkerPatch::DexCacheArrayPatch(info.label.location(), 3704 &info.target_dex_file, 3705 info.pc_insn_label->location(), 3706 info.element_offset)); 3707 } 3708} 3709 3710vixl::Literal<uint64_t>* CodeGeneratorARM64::DeduplicateUint64Literal(uint64_t value) { 3711 // Look up the literal for value. 3712 auto lb = uint64_literals_.lower_bound(value); 3713 if (lb != uint64_literals_.end() && !uint64_literals_.key_comp()(value, lb->first)) { 3714 return lb->second; 3715 } 3716 // We don't have a literal for this value, insert a new one. 3717 vixl::Literal<uint64_t>* literal = __ CreateLiteralDestroyedWithPool<uint64_t>(value); 3718 uint64_literals_.PutBefore(lb, value, literal); 3719 return literal; 3720} 3721 3722vixl::Literal<uint64_t>* CodeGeneratorARM64::DeduplicateMethodLiteral( 3723 MethodReference target_method, 3724 MethodToLiteralMap* map) { 3725 // Look up the literal for target_method. 3726 auto lb = map->lower_bound(target_method); 3727 if (lb != map->end() && !map->key_comp()(target_method, lb->first)) { 3728 return lb->second; 3729 } 3730 // We don't have a literal for this method yet, insert a new one. 3731 vixl::Literal<uint64_t>* literal = __ CreateLiteralDestroyedWithPool<uint64_t>(0u); 3732 map->PutBefore(lb, target_method, literal); 3733 return literal; 3734} 3735 3736vixl::Literal<uint64_t>* CodeGeneratorARM64::DeduplicateMethodAddressLiteral( 3737 MethodReference target_method) { 3738 return DeduplicateMethodLiteral(target_method, &method_patches_); 3739} 3740 3741vixl::Literal<uint64_t>* CodeGeneratorARM64::DeduplicateMethodCodeLiteral( 3742 MethodReference target_method) { 3743 return DeduplicateMethodLiteral(target_method, &call_patches_); 3744} 3745 3746 3747void InstructionCodeGeneratorARM64::VisitInvokeStaticOrDirect(HInvokeStaticOrDirect* invoke) { 3748 // When we do not run baseline, explicit clinit checks triggered by static 3749 // invokes must have been pruned by art::PrepareForRegisterAllocation. 3750 DCHECK(codegen_->IsBaseline() || !invoke->IsStaticWithExplicitClinitCheck()); 3751 3752 if (TryGenerateIntrinsicCode(invoke, codegen_)) { 3753 return; 3754 } 3755 3756 BlockPoolsScope block_pools(GetVIXLAssembler()); 3757 LocationSummary* locations = invoke->GetLocations(); 3758 codegen_->GenerateStaticOrDirectCall( 3759 invoke, locations->HasTemps() ? locations->GetTemp(0) : Location::NoLocation()); 3760 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 3761} 3762 3763void InstructionCodeGeneratorARM64::VisitInvokeVirtual(HInvokeVirtual* invoke) { 3764 if (TryGenerateIntrinsicCode(invoke, codegen_)) { 3765 return; 3766 } 3767 3768 codegen_->GenerateVirtualCall(invoke, invoke->GetLocations()->GetTemp(0)); 3769 DCHECK(!codegen_->IsLeafMethod()); 3770 codegen_->RecordPcInfo(invoke, invoke->GetDexPc()); 3771} 3772 3773void LocationsBuilderARM64::VisitLoadClass(HLoadClass* cls) { 3774 InvokeRuntimeCallingConvention calling_convention; 3775 CodeGenerator::CreateLoadClassLocationSummary( 3776 cls, 3777 LocationFrom(calling_convention.GetRegisterAt(0)), 3778 LocationFrom(vixl::x0), 3779 /* code_generator_supports_read_barrier */ true); 3780} 3781 3782void InstructionCodeGeneratorARM64::VisitLoadClass(HLoadClass* cls) { 3783 if (cls->NeedsAccessCheck()) { 3784 codegen_->MoveConstant(cls->GetLocations()->GetTemp(0), cls->GetTypeIndex()); 3785 codegen_->InvokeRuntime(QUICK_ENTRY_POINT(pInitializeTypeAndVerifyAccess), 3786 cls, 3787 cls->GetDexPc(), 3788 nullptr); 3789 CheckEntrypointTypes<kQuickInitializeTypeAndVerifyAccess, void*, uint32_t>(); 3790 return; 3791 } 3792 3793 Location out_loc = cls->GetLocations()->Out(); 3794 Register out = OutputRegister(cls); 3795 Register current_method = InputRegisterAt(cls, 0); 3796 if (cls->IsReferrersClass()) { 3797 DCHECK(!cls->CanCallRuntime()); 3798 DCHECK(!cls->MustGenerateClinitCheck()); 3799 uint32_t declaring_class_offset = ArtMethod::DeclaringClassOffset().Int32Value(); 3800 if (kEmitCompilerReadBarrier) { 3801 // /* GcRoot<mirror::Class>* */ out = &(current_method->declaring_class_) 3802 __ Add(out.X(), current_method.X(), declaring_class_offset); 3803 // /* mirror::Class* */ out = out->Read() 3804 codegen_->GenerateReadBarrierForRoot(cls, out_loc, out_loc); 3805 } else { 3806 // /* GcRoot<mirror::Class> */ out = current_method->declaring_class_ 3807 __ Ldr(out, MemOperand(current_method, declaring_class_offset)); 3808 } 3809 } else { 3810 MemberOffset resolved_types_offset = ArtMethod::DexCacheResolvedTypesOffset(kArm64PointerSize); 3811 // /* GcRoot<mirror::Class>[] */ out = 3812 // current_method.ptr_sized_fields_->dex_cache_resolved_types_ 3813 __ Ldr(out.X(), MemOperand(current_method, resolved_types_offset.Int32Value())); 3814 3815 size_t cache_offset = CodeGenerator::GetCacheOffset(cls->GetTypeIndex()); 3816 if (kEmitCompilerReadBarrier) { 3817 // /* GcRoot<mirror::Class>* */ out = &out[type_index] 3818 __ Add(out.X(), out.X(), cache_offset); 3819 // /* mirror::Class* */ out = out->Read() 3820 codegen_->GenerateReadBarrierForRoot(cls, out_loc, out_loc); 3821 } else { 3822 // /* GcRoot<mirror::Class> */ out = out[type_index] 3823 __ Ldr(out, MemOperand(out.X(), cache_offset)); 3824 } 3825 3826 if (!cls->IsInDexCache() || cls->MustGenerateClinitCheck()) { 3827 DCHECK(cls->CanCallRuntime()); 3828 SlowPathCodeARM64* slow_path = new (GetGraph()->GetArena()) LoadClassSlowPathARM64( 3829 cls, cls, cls->GetDexPc(), cls->MustGenerateClinitCheck()); 3830 codegen_->AddSlowPath(slow_path); 3831 if (!cls->IsInDexCache()) { 3832 __ Cbz(out, slow_path->GetEntryLabel()); 3833 } 3834 if (cls->MustGenerateClinitCheck()) { 3835 GenerateClassInitializationCheck(slow_path, out); 3836 } else { 3837 __ Bind(slow_path->GetExitLabel()); 3838 } 3839 } 3840 } 3841} 3842 3843static MemOperand GetExceptionTlsAddress() { 3844 return MemOperand(tr, Thread::ExceptionOffset<kArm64WordSize>().Int32Value()); 3845} 3846 3847void LocationsBuilderARM64::VisitLoadException(HLoadException* load) { 3848 LocationSummary* locations = 3849 new (GetGraph()->GetArena()) LocationSummary(load, LocationSummary::kNoCall); 3850 locations->SetOut(Location::RequiresRegister()); 3851} 3852 3853void InstructionCodeGeneratorARM64::VisitLoadException(HLoadException* instruction) { 3854 __ Ldr(OutputRegister(instruction), GetExceptionTlsAddress()); 3855} 3856 3857void LocationsBuilderARM64::VisitClearException(HClearException* clear) { 3858 new (GetGraph()->GetArena()) LocationSummary(clear, LocationSummary::kNoCall); 3859} 3860 3861void InstructionCodeGeneratorARM64::VisitClearException(HClearException* clear ATTRIBUTE_UNUSED) { 3862 __ Str(wzr, GetExceptionTlsAddress()); 3863} 3864 3865void LocationsBuilderARM64::VisitLoadLocal(HLoadLocal* load) { 3866 load->SetLocations(nullptr); 3867} 3868 3869void InstructionCodeGeneratorARM64::VisitLoadLocal(HLoadLocal* load ATTRIBUTE_UNUSED) { 3870 // Nothing to do, this is driven by the code generator. 3871} 3872 3873void LocationsBuilderARM64::VisitLoadString(HLoadString* load) { 3874 LocationSummary::CallKind call_kind = (!load->IsInDexCache() || kEmitCompilerReadBarrier) 3875 ? LocationSummary::kCallOnSlowPath 3876 : LocationSummary::kNoCall; 3877 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(load, call_kind); 3878 locations->SetInAt(0, Location::RequiresRegister()); 3879 locations->SetOut(Location::RequiresRegister()); 3880} 3881 3882void InstructionCodeGeneratorARM64::VisitLoadString(HLoadString* load) { 3883 Location out_loc = load->GetLocations()->Out(); 3884 Register out = OutputRegister(load); 3885 Register current_method = InputRegisterAt(load, 0); 3886 3887 uint32_t declaring_class_offset = ArtMethod::DeclaringClassOffset().Int32Value(); 3888 if (kEmitCompilerReadBarrier) { 3889 // /* GcRoot<mirror::Class>* */ out = &(current_method->declaring_class_) 3890 __ Add(out.X(), current_method.X(), declaring_class_offset); 3891 // /* mirror::Class* */ out = out->Read() 3892 codegen_->GenerateReadBarrierForRoot(load, out_loc, out_loc); 3893 } else { 3894 // /* GcRoot<mirror::Class> */ out = current_method->declaring_class_ 3895 __ Ldr(out, MemOperand(current_method, declaring_class_offset)); 3896 } 3897 3898 // /* GcRoot<mirror::String>[] */ out = out->dex_cache_strings_ 3899 __ Ldr(out.X(), HeapOperand(out, mirror::Class::DexCacheStringsOffset().Uint32Value())); 3900 3901 size_t cache_offset = CodeGenerator::GetCacheOffset(load->GetStringIndex()); 3902 if (kEmitCompilerReadBarrier) { 3903 // /* GcRoot<mirror::String>* */ out = &out[string_index] 3904 __ Add(out.X(), out.X(), cache_offset); 3905 // /* mirror::String* */ out = out->Read() 3906 codegen_->GenerateReadBarrierForRoot(load, out_loc, out_loc); 3907 } else { 3908 // /* GcRoot<mirror::String> */ out = out[string_index] 3909 __ Ldr(out, MemOperand(out.X(), cache_offset)); 3910 } 3911 3912 if (!load->IsInDexCache()) { 3913 SlowPathCodeARM64* slow_path = new (GetGraph()->GetArena()) LoadStringSlowPathARM64(load); 3914 codegen_->AddSlowPath(slow_path); 3915 __ Cbz(out, slow_path->GetEntryLabel()); 3916 __ Bind(slow_path->GetExitLabel()); 3917 } 3918} 3919 3920void LocationsBuilderARM64::VisitLocal(HLocal* local) { 3921 local->SetLocations(nullptr); 3922} 3923 3924void InstructionCodeGeneratorARM64::VisitLocal(HLocal* local) { 3925 DCHECK_EQ(local->GetBlock(), GetGraph()->GetEntryBlock()); 3926} 3927 3928void LocationsBuilderARM64::VisitLongConstant(HLongConstant* constant) { 3929 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(constant); 3930 locations->SetOut(Location::ConstantLocation(constant)); 3931} 3932 3933void InstructionCodeGeneratorARM64::VisitLongConstant(HLongConstant* constant ATTRIBUTE_UNUSED) { 3934 // Will be generated at use site. 3935} 3936 3937void LocationsBuilderARM64::VisitMonitorOperation(HMonitorOperation* instruction) { 3938 LocationSummary* locations = 3939 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall); 3940 InvokeRuntimeCallingConvention calling_convention; 3941 locations->SetInAt(0, LocationFrom(calling_convention.GetRegisterAt(0))); 3942} 3943 3944void InstructionCodeGeneratorARM64::VisitMonitorOperation(HMonitorOperation* instruction) { 3945 codegen_->InvokeRuntime(instruction->IsEnter() 3946 ? QUICK_ENTRY_POINT(pLockObject) : QUICK_ENTRY_POINT(pUnlockObject), 3947 instruction, 3948 instruction->GetDexPc(), 3949 nullptr); 3950 if (instruction->IsEnter()) { 3951 CheckEntrypointTypes<kQuickLockObject, void, mirror::Object*>(); 3952 } else { 3953 CheckEntrypointTypes<kQuickUnlockObject, void, mirror::Object*>(); 3954 } 3955} 3956 3957void LocationsBuilderARM64::VisitMul(HMul* mul) { 3958 LocationSummary* locations = 3959 new (GetGraph()->GetArena()) LocationSummary(mul, LocationSummary::kNoCall); 3960 switch (mul->GetResultType()) { 3961 case Primitive::kPrimInt: 3962 case Primitive::kPrimLong: 3963 locations->SetInAt(0, Location::RequiresRegister()); 3964 locations->SetInAt(1, Location::RequiresRegister()); 3965 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 3966 break; 3967 3968 case Primitive::kPrimFloat: 3969 case Primitive::kPrimDouble: 3970 locations->SetInAt(0, Location::RequiresFpuRegister()); 3971 locations->SetInAt(1, Location::RequiresFpuRegister()); 3972 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 3973 break; 3974 3975 default: 3976 LOG(FATAL) << "Unexpected mul type " << mul->GetResultType(); 3977 } 3978} 3979 3980void InstructionCodeGeneratorARM64::VisitMul(HMul* mul) { 3981 switch (mul->GetResultType()) { 3982 case Primitive::kPrimInt: 3983 case Primitive::kPrimLong: 3984 __ Mul(OutputRegister(mul), InputRegisterAt(mul, 0), InputRegisterAt(mul, 1)); 3985 break; 3986 3987 case Primitive::kPrimFloat: 3988 case Primitive::kPrimDouble: 3989 __ Fmul(OutputFPRegister(mul), InputFPRegisterAt(mul, 0), InputFPRegisterAt(mul, 1)); 3990 break; 3991 3992 default: 3993 LOG(FATAL) << "Unexpected mul type " << mul->GetResultType(); 3994 } 3995} 3996 3997void LocationsBuilderARM64::VisitNeg(HNeg* neg) { 3998 LocationSummary* locations = 3999 new (GetGraph()->GetArena()) LocationSummary(neg, LocationSummary::kNoCall); 4000 switch (neg->GetResultType()) { 4001 case Primitive::kPrimInt: 4002 case Primitive::kPrimLong: 4003 locations->SetInAt(0, ARM64EncodableConstantOrRegister(neg->InputAt(0), neg)); 4004 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 4005 break; 4006 4007 case Primitive::kPrimFloat: 4008 case Primitive::kPrimDouble: 4009 locations->SetInAt(0, Location::RequiresFpuRegister()); 4010 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 4011 break; 4012 4013 default: 4014 LOG(FATAL) << "Unexpected neg type " << neg->GetResultType(); 4015 } 4016} 4017 4018void InstructionCodeGeneratorARM64::VisitNeg(HNeg* neg) { 4019 switch (neg->GetResultType()) { 4020 case Primitive::kPrimInt: 4021 case Primitive::kPrimLong: 4022 __ Neg(OutputRegister(neg), InputOperandAt(neg, 0)); 4023 break; 4024 4025 case Primitive::kPrimFloat: 4026 case Primitive::kPrimDouble: 4027 __ Fneg(OutputFPRegister(neg), InputFPRegisterAt(neg, 0)); 4028 break; 4029 4030 default: 4031 LOG(FATAL) << "Unexpected neg type " << neg->GetResultType(); 4032 } 4033} 4034 4035void LocationsBuilderARM64::VisitNewArray(HNewArray* instruction) { 4036 LocationSummary* locations = 4037 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall); 4038 InvokeRuntimeCallingConvention calling_convention; 4039 locations->AddTemp(LocationFrom(calling_convention.GetRegisterAt(0))); 4040 locations->SetOut(LocationFrom(x0)); 4041 locations->SetInAt(0, LocationFrom(calling_convention.GetRegisterAt(1))); 4042 locations->SetInAt(1, LocationFrom(calling_convention.GetRegisterAt(2))); 4043} 4044 4045void InstructionCodeGeneratorARM64::VisitNewArray(HNewArray* instruction) { 4046 LocationSummary* locations = instruction->GetLocations(); 4047 InvokeRuntimeCallingConvention calling_convention; 4048 Register type_index = RegisterFrom(locations->GetTemp(0), Primitive::kPrimInt); 4049 DCHECK(type_index.Is(w0)); 4050 __ Mov(type_index, instruction->GetTypeIndex()); 4051 // Note: if heap poisoning is enabled, the entry point takes cares 4052 // of poisoning the reference. 4053 codegen_->InvokeRuntime(instruction->GetEntrypoint(), 4054 instruction, 4055 instruction->GetDexPc(), 4056 nullptr); 4057 CheckEntrypointTypes<kQuickAllocArrayWithAccessCheck, void*, uint32_t, int32_t, ArtMethod*>(); 4058} 4059 4060void LocationsBuilderARM64::VisitNewInstance(HNewInstance* instruction) { 4061 LocationSummary* locations = 4062 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall); 4063 InvokeRuntimeCallingConvention calling_convention; 4064 locations->SetInAt(0, LocationFrom(calling_convention.GetRegisterAt(0))); 4065 locations->SetInAt(1, LocationFrom(calling_convention.GetRegisterAt(1))); 4066 locations->SetOut(calling_convention.GetReturnLocation(Primitive::kPrimNot)); 4067} 4068 4069void InstructionCodeGeneratorARM64::VisitNewInstance(HNewInstance* instruction) { 4070 // Note: if heap poisoning is enabled, the entry point takes cares 4071 // of poisoning the reference. 4072 codegen_->InvokeRuntime(instruction->GetEntrypoint(), 4073 instruction, 4074 instruction->GetDexPc(), 4075 nullptr); 4076 CheckEntrypointTypes<kQuickAllocObjectWithAccessCheck, void*, uint32_t, ArtMethod*>(); 4077} 4078 4079void LocationsBuilderARM64::VisitNot(HNot* instruction) { 4080 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction); 4081 locations->SetInAt(0, Location::RequiresRegister()); 4082 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 4083} 4084 4085void InstructionCodeGeneratorARM64::VisitNot(HNot* instruction) { 4086 switch (instruction->GetResultType()) { 4087 case Primitive::kPrimInt: 4088 case Primitive::kPrimLong: 4089 __ Mvn(OutputRegister(instruction), InputOperandAt(instruction, 0)); 4090 break; 4091 4092 default: 4093 LOG(FATAL) << "Unexpected type for not operation " << instruction->GetResultType(); 4094 } 4095} 4096 4097void LocationsBuilderARM64::VisitBooleanNot(HBooleanNot* instruction) { 4098 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction); 4099 locations->SetInAt(0, Location::RequiresRegister()); 4100 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 4101} 4102 4103void InstructionCodeGeneratorARM64::VisitBooleanNot(HBooleanNot* instruction) { 4104 __ Eor(OutputRegister(instruction), InputRegisterAt(instruction, 0), vixl::Operand(1)); 4105} 4106 4107void LocationsBuilderARM64::VisitNullCheck(HNullCheck* instruction) { 4108 LocationSummary::CallKind call_kind = instruction->CanThrowIntoCatchBlock() 4109 ? LocationSummary::kCallOnSlowPath 4110 : LocationSummary::kNoCall; 4111 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction, call_kind); 4112 locations->SetInAt(0, Location::RequiresRegister()); 4113 if (instruction->HasUses()) { 4114 locations->SetOut(Location::SameAsFirstInput()); 4115 } 4116} 4117 4118void InstructionCodeGeneratorARM64::GenerateImplicitNullCheck(HNullCheck* instruction) { 4119 if (codegen_->CanMoveNullCheckToUser(instruction)) { 4120 return; 4121 } 4122 4123 BlockPoolsScope block_pools(GetVIXLAssembler()); 4124 Location obj = instruction->GetLocations()->InAt(0); 4125 __ Ldr(wzr, HeapOperandFrom(obj, Offset(0))); 4126 codegen_->RecordPcInfo(instruction, instruction->GetDexPc()); 4127} 4128 4129void InstructionCodeGeneratorARM64::GenerateExplicitNullCheck(HNullCheck* instruction) { 4130 SlowPathCodeARM64* slow_path = new (GetGraph()->GetArena()) NullCheckSlowPathARM64(instruction); 4131 codegen_->AddSlowPath(slow_path); 4132 4133 LocationSummary* locations = instruction->GetLocations(); 4134 Location obj = locations->InAt(0); 4135 4136 __ Cbz(RegisterFrom(obj, instruction->InputAt(0)->GetType()), slow_path->GetEntryLabel()); 4137} 4138 4139void InstructionCodeGeneratorARM64::VisitNullCheck(HNullCheck* instruction) { 4140 if (codegen_->IsImplicitNullCheckAllowed(instruction)) { 4141 GenerateImplicitNullCheck(instruction); 4142 } else { 4143 GenerateExplicitNullCheck(instruction); 4144 } 4145} 4146 4147void LocationsBuilderARM64::VisitOr(HOr* instruction) { 4148 HandleBinaryOp(instruction); 4149} 4150 4151void InstructionCodeGeneratorARM64::VisitOr(HOr* instruction) { 4152 HandleBinaryOp(instruction); 4153} 4154 4155void LocationsBuilderARM64::VisitParallelMove(HParallelMove* instruction ATTRIBUTE_UNUSED) { 4156 LOG(FATAL) << "Unreachable"; 4157} 4158 4159void InstructionCodeGeneratorARM64::VisitParallelMove(HParallelMove* instruction) { 4160 codegen_->GetMoveResolver()->EmitNativeCode(instruction); 4161} 4162 4163void LocationsBuilderARM64::VisitParameterValue(HParameterValue* instruction) { 4164 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction); 4165 Location location = parameter_visitor_.GetNextLocation(instruction->GetType()); 4166 if (location.IsStackSlot()) { 4167 location = Location::StackSlot(location.GetStackIndex() + codegen_->GetFrameSize()); 4168 } else if (location.IsDoubleStackSlot()) { 4169 location = Location::DoubleStackSlot(location.GetStackIndex() + codegen_->GetFrameSize()); 4170 } 4171 locations->SetOut(location); 4172} 4173 4174void InstructionCodeGeneratorARM64::VisitParameterValue( 4175 HParameterValue* instruction ATTRIBUTE_UNUSED) { 4176 // Nothing to do, the parameter is already at its location. 4177} 4178 4179void LocationsBuilderARM64::VisitCurrentMethod(HCurrentMethod* instruction) { 4180 LocationSummary* locations = 4181 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 4182 locations->SetOut(LocationFrom(kArtMethodRegister)); 4183} 4184 4185void InstructionCodeGeneratorARM64::VisitCurrentMethod( 4186 HCurrentMethod* instruction ATTRIBUTE_UNUSED) { 4187 // Nothing to do, the method is already at its location. 4188} 4189 4190void LocationsBuilderARM64::VisitPhi(HPhi* instruction) { 4191 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction); 4192 for (size_t i = 0, e = instruction->InputCount(); i < e; ++i) { 4193 locations->SetInAt(i, Location::Any()); 4194 } 4195 locations->SetOut(Location::Any()); 4196} 4197 4198void InstructionCodeGeneratorARM64::VisitPhi(HPhi* instruction ATTRIBUTE_UNUSED) { 4199 LOG(FATAL) << "Unreachable"; 4200} 4201 4202void LocationsBuilderARM64::VisitRem(HRem* rem) { 4203 Primitive::Type type = rem->GetResultType(); 4204 LocationSummary::CallKind call_kind = 4205 Primitive::IsFloatingPointType(type) ? LocationSummary::kCall : LocationSummary::kNoCall; 4206 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(rem, call_kind); 4207 4208 switch (type) { 4209 case Primitive::kPrimInt: 4210 case Primitive::kPrimLong: 4211 locations->SetInAt(0, Location::RequiresRegister()); 4212 locations->SetInAt(1, Location::RegisterOrConstant(rem->InputAt(1))); 4213 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 4214 break; 4215 4216 case Primitive::kPrimFloat: 4217 case Primitive::kPrimDouble: { 4218 InvokeRuntimeCallingConvention calling_convention; 4219 locations->SetInAt(0, LocationFrom(calling_convention.GetFpuRegisterAt(0))); 4220 locations->SetInAt(1, LocationFrom(calling_convention.GetFpuRegisterAt(1))); 4221 locations->SetOut(calling_convention.GetReturnLocation(type)); 4222 4223 break; 4224 } 4225 4226 default: 4227 LOG(FATAL) << "Unexpected rem type " << type; 4228 } 4229} 4230 4231void InstructionCodeGeneratorARM64::VisitRem(HRem* rem) { 4232 Primitive::Type type = rem->GetResultType(); 4233 4234 switch (type) { 4235 case Primitive::kPrimInt: 4236 case Primitive::kPrimLong: { 4237 GenerateDivRemIntegral(rem); 4238 break; 4239 } 4240 4241 case Primitive::kPrimFloat: 4242 case Primitive::kPrimDouble: { 4243 int32_t entry_offset = (type == Primitive::kPrimFloat) ? QUICK_ENTRY_POINT(pFmodf) 4244 : QUICK_ENTRY_POINT(pFmod); 4245 codegen_->InvokeRuntime(entry_offset, rem, rem->GetDexPc(), nullptr); 4246 if (type == Primitive::kPrimFloat) { 4247 CheckEntrypointTypes<kQuickFmodf, float, float, float>(); 4248 } else { 4249 CheckEntrypointTypes<kQuickFmod, double, double, double>(); 4250 } 4251 break; 4252 } 4253 4254 default: 4255 LOG(FATAL) << "Unexpected rem type " << type; 4256 UNREACHABLE(); 4257 } 4258} 4259 4260void LocationsBuilderARM64::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) { 4261 memory_barrier->SetLocations(nullptr); 4262} 4263 4264void InstructionCodeGeneratorARM64::VisitMemoryBarrier(HMemoryBarrier* memory_barrier) { 4265 GenerateMemoryBarrier(memory_barrier->GetBarrierKind()); 4266} 4267 4268void LocationsBuilderARM64::VisitReturn(HReturn* instruction) { 4269 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(instruction); 4270 Primitive::Type return_type = instruction->InputAt(0)->GetType(); 4271 locations->SetInAt(0, ARM64ReturnLocation(return_type)); 4272} 4273 4274void InstructionCodeGeneratorARM64::VisitReturn(HReturn* instruction ATTRIBUTE_UNUSED) { 4275 codegen_->GenerateFrameExit(); 4276} 4277 4278void LocationsBuilderARM64::VisitReturnVoid(HReturnVoid* instruction) { 4279 instruction->SetLocations(nullptr); 4280} 4281 4282void InstructionCodeGeneratorARM64::VisitReturnVoid(HReturnVoid* instruction ATTRIBUTE_UNUSED) { 4283 codegen_->GenerateFrameExit(); 4284} 4285 4286void LocationsBuilderARM64::VisitRor(HRor* ror) { 4287 HandleBinaryOp(ror); 4288} 4289 4290void InstructionCodeGeneratorARM64::VisitRor(HRor* ror) { 4291 HandleBinaryOp(ror); 4292} 4293 4294void LocationsBuilderARM64::VisitShl(HShl* shl) { 4295 HandleShift(shl); 4296} 4297 4298void InstructionCodeGeneratorARM64::VisitShl(HShl* shl) { 4299 HandleShift(shl); 4300} 4301 4302void LocationsBuilderARM64::VisitShr(HShr* shr) { 4303 HandleShift(shr); 4304} 4305 4306void InstructionCodeGeneratorARM64::VisitShr(HShr* shr) { 4307 HandleShift(shr); 4308} 4309 4310void LocationsBuilderARM64::VisitStoreLocal(HStoreLocal* store) { 4311 LocationSummary* locations = new (GetGraph()->GetArena()) LocationSummary(store); 4312 Primitive::Type field_type = store->InputAt(1)->GetType(); 4313 switch (field_type) { 4314 case Primitive::kPrimNot: 4315 case Primitive::kPrimBoolean: 4316 case Primitive::kPrimByte: 4317 case Primitive::kPrimChar: 4318 case Primitive::kPrimShort: 4319 case Primitive::kPrimInt: 4320 case Primitive::kPrimFloat: 4321 locations->SetInAt(1, Location::StackSlot(codegen_->GetStackSlot(store->GetLocal()))); 4322 break; 4323 4324 case Primitive::kPrimLong: 4325 case Primitive::kPrimDouble: 4326 locations->SetInAt(1, Location::DoubleStackSlot(codegen_->GetStackSlot(store->GetLocal()))); 4327 break; 4328 4329 default: 4330 LOG(FATAL) << "Unimplemented local type " << field_type; 4331 UNREACHABLE(); 4332 } 4333} 4334 4335void InstructionCodeGeneratorARM64::VisitStoreLocal(HStoreLocal* store ATTRIBUTE_UNUSED) { 4336} 4337 4338void LocationsBuilderARM64::VisitSub(HSub* instruction) { 4339 HandleBinaryOp(instruction); 4340} 4341 4342void InstructionCodeGeneratorARM64::VisitSub(HSub* instruction) { 4343 HandleBinaryOp(instruction); 4344} 4345 4346void LocationsBuilderARM64::VisitStaticFieldGet(HStaticFieldGet* instruction) { 4347 HandleFieldGet(instruction); 4348} 4349 4350void InstructionCodeGeneratorARM64::VisitStaticFieldGet(HStaticFieldGet* instruction) { 4351 HandleFieldGet(instruction, instruction->GetFieldInfo()); 4352} 4353 4354void LocationsBuilderARM64::VisitStaticFieldSet(HStaticFieldSet* instruction) { 4355 HandleFieldSet(instruction); 4356} 4357 4358void InstructionCodeGeneratorARM64::VisitStaticFieldSet(HStaticFieldSet* instruction) { 4359 HandleFieldSet(instruction, instruction->GetFieldInfo(), instruction->GetValueCanBeNull()); 4360} 4361 4362void LocationsBuilderARM64::VisitUnresolvedInstanceFieldGet( 4363 HUnresolvedInstanceFieldGet* instruction) { 4364 FieldAccessCallingConventionARM64 calling_convention; 4365 codegen_->CreateUnresolvedFieldLocationSummary( 4366 instruction, instruction->GetFieldType(), calling_convention); 4367} 4368 4369void InstructionCodeGeneratorARM64::VisitUnresolvedInstanceFieldGet( 4370 HUnresolvedInstanceFieldGet* instruction) { 4371 FieldAccessCallingConventionARM64 calling_convention; 4372 codegen_->GenerateUnresolvedFieldAccess(instruction, 4373 instruction->GetFieldType(), 4374 instruction->GetFieldIndex(), 4375 instruction->GetDexPc(), 4376 calling_convention); 4377} 4378 4379void LocationsBuilderARM64::VisitUnresolvedInstanceFieldSet( 4380 HUnresolvedInstanceFieldSet* instruction) { 4381 FieldAccessCallingConventionARM64 calling_convention; 4382 codegen_->CreateUnresolvedFieldLocationSummary( 4383 instruction, instruction->GetFieldType(), calling_convention); 4384} 4385 4386void InstructionCodeGeneratorARM64::VisitUnresolvedInstanceFieldSet( 4387 HUnresolvedInstanceFieldSet* instruction) { 4388 FieldAccessCallingConventionARM64 calling_convention; 4389 codegen_->GenerateUnresolvedFieldAccess(instruction, 4390 instruction->GetFieldType(), 4391 instruction->GetFieldIndex(), 4392 instruction->GetDexPc(), 4393 calling_convention); 4394} 4395 4396void LocationsBuilderARM64::VisitUnresolvedStaticFieldGet( 4397 HUnresolvedStaticFieldGet* instruction) { 4398 FieldAccessCallingConventionARM64 calling_convention; 4399 codegen_->CreateUnresolvedFieldLocationSummary( 4400 instruction, instruction->GetFieldType(), calling_convention); 4401} 4402 4403void InstructionCodeGeneratorARM64::VisitUnresolvedStaticFieldGet( 4404 HUnresolvedStaticFieldGet* instruction) { 4405 FieldAccessCallingConventionARM64 calling_convention; 4406 codegen_->GenerateUnresolvedFieldAccess(instruction, 4407 instruction->GetFieldType(), 4408 instruction->GetFieldIndex(), 4409 instruction->GetDexPc(), 4410 calling_convention); 4411} 4412 4413void LocationsBuilderARM64::VisitUnresolvedStaticFieldSet( 4414 HUnresolvedStaticFieldSet* instruction) { 4415 FieldAccessCallingConventionARM64 calling_convention; 4416 codegen_->CreateUnresolvedFieldLocationSummary( 4417 instruction, instruction->GetFieldType(), calling_convention); 4418} 4419 4420void InstructionCodeGeneratorARM64::VisitUnresolvedStaticFieldSet( 4421 HUnresolvedStaticFieldSet* instruction) { 4422 FieldAccessCallingConventionARM64 calling_convention; 4423 codegen_->GenerateUnresolvedFieldAccess(instruction, 4424 instruction->GetFieldType(), 4425 instruction->GetFieldIndex(), 4426 instruction->GetDexPc(), 4427 calling_convention); 4428} 4429 4430void LocationsBuilderARM64::VisitSuspendCheck(HSuspendCheck* instruction) { 4431 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCallOnSlowPath); 4432} 4433 4434void InstructionCodeGeneratorARM64::VisitSuspendCheck(HSuspendCheck* instruction) { 4435 HBasicBlock* block = instruction->GetBlock(); 4436 if (block->GetLoopInformation() != nullptr) { 4437 DCHECK(block->GetLoopInformation()->GetSuspendCheck() == instruction); 4438 // The back edge will generate the suspend check. 4439 return; 4440 } 4441 if (block->IsEntryBlock() && instruction->GetNext()->IsGoto()) { 4442 // The goto will generate the suspend check. 4443 return; 4444 } 4445 GenerateSuspendCheck(instruction, nullptr); 4446} 4447 4448void LocationsBuilderARM64::VisitTemporary(HTemporary* temp) { 4449 temp->SetLocations(nullptr); 4450} 4451 4452void InstructionCodeGeneratorARM64::VisitTemporary(HTemporary* temp ATTRIBUTE_UNUSED) { 4453 // Nothing to do, this is driven by the code generator. 4454} 4455 4456void LocationsBuilderARM64::VisitThrow(HThrow* instruction) { 4457 LocationSummary* locations = 4458 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kCall); 4459 InvokeRuntimeCallingConvention calling_convention; 4460 locations->SetInAt(0, LocationFrom(calling_convention.GetRegisterAt(0))); 4461} 4462 4463void InstructionCodeGeneratorARM64::VisitThrow(HThrow* instruction) { 4464 codegen_->InvokeRuntime( 4465 QUICK_ENTRY_POINT(pDeliverException), instruction, instruction->GetDexPc(), nullptr); 4466 CheckEntrypointTypes<kQuickDeliverException, void, mirror::Object*>(); 4467} 4468 4469void LocationsBuilderARM64::VisitTypeConversion(HTypeConversion* conversion) { 4470 LocationSummary* locations = 4471 new (GetGraph()->GetArena()) LocationSummary(conversion, LocationSummary::kNoCall); 4472 Primitive::Type input_type = conversion->GetInputType(); 4473 Primitive::Type result_type = conversion->GetResultType(); 4474 DCHECK_NE(input_type, result_type); 4475 if ((input_type == Primitive::kPrimNot) || (input_type == Primitive::kPrimVoid) || 4476 (result_type == Primitive::kPrimNot) || (result_type == Primitive::kPrimVoid)) { 4477 LOG(FATAL) << "Unexpected type conversion from " << input_type << " to " << result_type; 4478 } 4479 4480 if (Primitive::IsFloatingPointType(input_type)) { 4481 locations->SetInAt(0, Location::RequiresFpuRegister()); 4482 } else { 4483 locations->SetInAt(0, Location::RequiresRegister()); 4484 } 4485 4486 if (Primitive::IsFloatingPointType(result_type)) { 4487 locations->SetOut(Location::RequiresFpuRegister(), Location::kNoOutputOverlap); 4488 } else { 4489 locations->SetOut(Location::RequiresRegister(), Location::kNoOutputOverlap); 4490 } 4491} 4492 4493void InstructionCodeGeneratorARM64::VisitTypeConversion(HTypeConversion* conversion) { 4494 Primitive::Type result_type = conversion->GetResultType(); 4495 Primitive::Type input_type = conversion->GetInputType(); 4496 4497 DCHECK_NE(input_type, result_type); 4498 4499 if (Primitive::IsIntegralType(result_type) && Primitive::IsIntegralType(input_type)) { 4500 int result_size = Primitive::ComponentSize(result_type); 4501 int input_size = Primitive::ComponentSize(input_type); 4502 int min_size = std::min(result_size, input_size); 4503 Register output = OutputRegister(conversion); 4504 Register source = InputRegisterAt(conversion, 0); 4505 if (result_type == Primitive::kPrimInt && input_type == Primitive::kPrimLong) { 4506 // 'int' values are used directly as W registers, discarding the top 4507 // bits, so we don't need to sign-extend and can just perform a move. 4508 // We do not pass the `kDiscardForSameWReg` argument to force clearing the 4509 // top 32 bits of the target register. We theoretically could leave those 4510 // bits unchanged, but we would have to make sure that no code uses a 4511 // 32bit input value as a 64bit value assuming that the top 32 bits are 4512 // zero. 4513 __ Mov(output.W(), source.W()); 4514 } else if (result_type == Primitive::kPrimChar || 4515 (input_type == Primitive::kPrimChar && input_size < result_size)) { 4516 __ Ubfx(output, 4517 output.IsX() ? source.X() : source.W(), 4518 0, Primitive::ComponentSize(Primitive::kPrimChar) * kBitsPerByte); 4519 } else { 4520 __ Sbfx(output, output.IsX() ? source.X() : source.W(), 0, min_size * kBitsPerByte); 4521 } 4522 } else if (Primitive::IsFloatingPointType(result_type) && Primitive::IsIntegralType(input_type)) { 4523 __ Scvtf(OutputFPRegister(conversion), InputRegisterAt(conversion, 0)); 4524 } else if (Primitive::IsIntegralType(result_type) && Primitive::IsFloatingPointType(input_type)) { 4525 CHECK(result_type == Primitive::kPrimInt || result_type == Primitive::kPrimLong); 4526 __ Fcvtzs(OutputRegister(conversion), InputFPRegisterAt(conversion, 0)); 4527 } else if (Primitive::IsFloatingPointType(result_type) && 4528 Primitive::IsFloatingPointType(input_type)) { 4529 __ Fcvt(OutputFPRegister(conversion), InputFPRegisterAt(conversion, 0)); 4530 } else { 4531 LOG(FATAL) << "Unexpected or unimplemented type conversion from " << input_type 4532 << " to " << result_type; 4533 } 4534} 4535 4536void LocationsBuilderARM64::VisitUShr(HUShr* ushr) { 4537 HandleShift(ushr); 4538} 4539 4540void InstructionCodeGeneratorARM64::VisitUShr(HUShr* ushr) { 4541 HandleShift(ushr); 4542} 4543 4544void LocationsBuilderARM64::VisitXor(HXor* instruction) { 4545 HandleBinaryOp(instruction); 4546} 4547 4548void InstructionCodeGeneratorARM64::VisitXor(HXor* instruction) { 4549 HandleBinaryOp(instruction); 4550} 4551 4552void LocationsBuilderARM64::VisitBoundType(HBoundType* instruction ATTRIBUTE_UNUSED) { 4553 // Nothing to do, this should be removed during prepare for register allocator. 4554 LOG(FATAL) << "Unreachable"; 4555} 4556 4557void InstructionCodeGeneratorARM64::VisitBoundType(HBoundType* instruction ATTRIBUTE_UNUSED) { 4558 // Nothing to do, this should be removed during prepare for register allocator. 4559 LOG(FATAL) << "Unreachable"; 4560} 4561 4562void LocationsBuilderARM64::VisitFakeString(HFakeString* instruction) { 4563 DCHECK(codegen_->IsBaseline()); 4564 LocationSummary* locations = 4565 new (GetGraph()->GetArena()) LocationSummary(instruction, LocationSummary::kNoCall); 4566 locations->SetOut(Location::ConstantLocation(GetGraph()->GetNullConstant())); 4567} 4568 4569void InstructionCodeGeneratorARM64::VisitFakeString(HFakeString* instruction ATTRIBUTE_UNUSED) { 4570 DCHECK(codegen_->IsBaseline()); 4571 // Will be generated at use site. 4572} 4573 4574// Simple implementation of packed switch - generate cascaded compare/jumps. 4575void LocationsBuilderARM64::VisitPackedSwitch(HPackedSwitch* switch_instr) { 4576 LocationSummary* locations = 4577 new (GetGraph()->GetArena()) LocationSummary(switch_instr, LocationSummary::kNoCall); 4578 locations->SetInAt(0, Location::RequiresRegister()); 4579} 4580 4581void InstructionCodeGeneratorARM64::VisitPackedSwitch(HPackedSwitch* switch_instr) { 4582 int32_t lower_bound = switch_instr->GetStartValue(); 4583 uint32_t num_entries = switch_instr->GetNumEntries(); 4584 Register value_reg = InputRegisterAt(switch_instr, 0); 4585 HBasicBlock* default_block = switch_instr->GetDefaultBlock(); 4586 4587 // Roughly set 16 as max average assemblies generated per HIR in a graph. 4588 static constexpr int32_t kMaxExpectedSizePerHInstruction = 16 * vixl::kInstructionSize; 4589 // ADR has a limited range(+/-1MB), so we set a threshold for the number of HIRs in the graph to 4590 // make sure we don't emit it if the target may run out of range. 4591 // TODO: Instead of emitting all jump tables at the end of the code, we could keep track of ADR 4592 // ranges and emit the tables only as required. 4593 static constexpr int32_t kJumpTableInstructionThreshold = 1* MB / kMaxExpectedSizePerHInstruction; 4594 4595 if (num_entries <= kPackedSwitchCompareJumpThreshold || 4596 // Current instruction id is an upper bound of the number of HIRs in the graph. 4597 GetGraph()->GetCurrentInstructionId() > kJumpTableInstructionThreshold) { 4598 // Create a series of compare/jumps. 4599 UseScratchRegisterScope temps(codegen_->GetVIXLAssembler()); 4600 Register temp = temps.AcquireW(); 4601 __ Subs(temp, value_reg, Operand(lower_bound)); 4602 4603 const ArenaVector<HBasicBlock*>& successors = switch_instr->GetBlock()->GetSuccessors(); 4604 // Jump to successors[0] if value == lower_bound. 4605 __ B(eq, codegen_->GetLabelOf(successors[0])); 4606 int32_t last_index = 0; 4607 for (; num_entries - last_index > 2; last_index += 2) { 4608 __ Subs(temp, temp, Operand(2)); 4609 // Jump to successors[last_index + 1] if value < case_value[last_index + 2]. 4610 __ B(lo, codegen_->GetLabelOf(successors[last_index + 1])); 4611 // Jump to successors[last_index + 2] if value == case_value[last_index + 2]. 4612 __ B(eq, codegen_->GetLabelOf(successors[last_index + 2])); 4613 } 4614 if (num_entries - last_index == 2) { 4615 // The last missing case_value. 4616 __ Cmp(temp, Operand(1)); 4617 __ B(eq, codegen_->GetLabelOf(successors[last_index + 1])); 4618 } 4619 4620 // And the default for any other value. 4621 if (!codegen_->GoesToNextBlock(switch_instr->GetBlock(), default_block)) { 4622 __ B(codegen_->GetLabelOf(default_block)); 4623 } 4624 } else { 4625 JumpTableARM64* jump_table = new (GetGraph()->GetArena()) JumpTableARM64(switch_instr); 4626 codegen_->AddJumpTable(jump_table); 4627 4628 UseScratchRegisterScope temps(codegen_->GetVIXLAssembler()); 4629 4630 // Below instructions should use at most one blocked register. Since there are two blocked 4631 // registers, we are free to block one. 4632 Register temp_w = temps.AcquireW(); 4633 Register index; 4634 // Remove the bias. 4635 if (lower_bound != 0) { 4636 index = temp_w; 4637 __ Sub(index, value_reg, Operand(lower_bound)); 4638 } else { 4639 index = value_reg; 4640 } 4641 4642 // Jump to default block if index is out of the range. 4643 __ Cmp(index, Operand(num_entries)); 4644 __ B(hs, codegen_->GetLabelOf(default_block)); 4645 4646 // In current VIXL implementation, it won't require any blocked registers to encode the 4647 // immediate value for Adr. So we are free to use both VIXL blocked registers to reduce the 4648 // register pressure. 4649 Register table_base = temps.AcquireX(); 4650 // Load jump offset from the table. 4651 __ Adr(table_base, jump_table->GetTableStartLabel()); 4652 Register jump_offset = temp_w; 4653 __ Ldr(jump_offset, MemOperand(table_base, index, UXTW, 2)); 4654 4655 // Jump to target block by branching to table_base(pc related) + offset. 4656 Register target_address = table_base; 4657 __ Add(target_address, table_base, Operand(jump_offset, SXTW)); 4658 __ Br(target_address); 4659 } 4660} 4661 4662void CodeGeneratorARM64::GenerateReadBarrier(HInstruction* instruction, 4663 Location out, 4664 Location ref, 4665 Location obj, 4666 uint32_t offset, 4667 Location index) { 4668 DCHECK(kEmitCompilerReadBarrier); 4669 4670 // If heap poisoning is enabled, the unpoisoning of the loaded 4671 // reference will be carried out by the runtime within the slow 4672 // path. 4673 // 4674 // Note that `ref` currently does not get unpoisoned (when heap 4675 // poisoning is enabled), which is alright as the `ref` argument is 4676 // not used by the artReadBarrierSlow entry point. 4677 // 4678 // TODO: Unpoison `ref` when it is used by artReadBarrierSlow. 4679 SlowPathCodeARM64* slow_path = new (GetGraph()->GetArena()) 4680 ReadBarrierForHeapReferenceSlowPathARM64(instruction, out, ref, obj, offset, index); 4681 AddSlowPath(slow_path); 4682 4683 // TODO: When read barrier has a fast path, add it here. 4684 /* Currently the read barrier call is inserted after the original load. 4685 * However, if we have a fast path, we need to perform the load of obj.LockWord *before* the 4686 * original load. This load-load ordering is required by the read barrier. 4687 * The fast path/slow path (for Baker's algorithm) should look like: 4688 * 4689 * bool isGray = obj.LockWord & kReadBarrierMask; 4690 * lfence; // load fence or artificial data dependence to prevent load-load reordering 4691 * ref = obj.field; // this is the original load 4692 * if (isGray) { 4693 * ref = Mark(ref); // ideally the slow path just does Mark(ref) 4694 * } 4695 */ 4696 4697 __ B(slow_path->GetEntryLabel()); 4698 __ Bind(slow_path->GetExitLabel()); 4699} 4700 4701void CodeGeneratorARM64::MaybeGenerateReadBarrier(HInstruction* instruction, 4702 Location out, 4703 Location ref, 4704 Location obj, 4705 uint32_t offset, 4706 Location index) { 4707 if (kEmitCompilerReadBarrier) { 4708 // If heap poisoning is enabled, unpoisoning will be taken care of 4709 // by the runtime within the slow path. 4710 GenerateReadBarrier(instruction, out, ref, obj, offset, index); 4711 } else if (kPoisonHeapReferences) { 4712 GetAssembler()->UnpoisonHeapReference(WRegisterFrom(out)); 4713 } 4714} 4715 4716void CodeGeneratorARM64::GenerateReadBarrierForRoot(HInstruction* instruction, 4717 Location out, 4718 Location root) { 4719 DCHECK(kEmitCompilerReadBarrier); 4720 4721 // Note that GC roots are not affected by heap poisoning, so we do 4722 // not need to do anything special for this here. 4723 SlowPathCodeARM64* slow_path = 4724 new (GetGraph()->GetArena()) ReadBarrierForRootSlowPathARM64(instruction, out, root); 4725 AddSlowPath(slow_path); 4726 4727 // TODO: Implement a fast path for ReadBarrierForRoot, performing 4728 // the following operation (for Baker's algorithm): 4729 // 4730 // if (thread.tls32_.is_gc_marking) { 4731 // root = Mark(root); 4732 // } 4733 4734 __ B(slow_path->GetEntryLabel()); 4735 __ Bind(slow_path->GetExitLabel()); 4736} 4737 4738#undef __ 4739#undef QUICK_ENTRY_POINT 4740 4741} // namespace arm64 4742} // namespace art 4743