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 #ifndef ART_COMPILER_OPTIMIZING_CODE_GENERATOR_ARM64_H_
18 #define ART_COMPILER_OPTIMIZING_CODE_GENERATOR_ARM64_H_
19
20 #include "base/bit_field.h"
21 #include "base/macros.h"
22 #include "class_root.h"
23 #include "code_generator.h"
24 #include "common_arm64.h"
25 #include "dex/dex_file_types.h"
26 #include "dex/string_reference.h"
27 #include "dex/type_reference.h"
28 #include "driver/compiler_options.h"
29 #include "jit_patches_arm64.h"
30 #include "nodes.h"
31 #include "parallel_move_resolver.h"
32 #include "utils/arm64/assembler_arm64.h"
33
34 // TODO(VIXL): Make VIXL compile with -Wshadow.
35 #pragma GCC diagnostic push
36 #pragma GCC diagnostic ignored "-Wshadow"
37 #include "aarch64/disasm-aarch64.h"
38 #include "aarch64/macro-assembler-aarch64.h"
39 #pragma GCC diagnostic pop
40
41 namespace art HIDDEN {
42
43 namespace linker {
44 class Arm64RelativePatcherTest;
45 } // namespace linker
46
47 namespace arm64 {
48
49 class CodeGeneratorARM64;
50
51 // Use a local definition to prevent copying mistakes.
52 static constexpr size_t kArm64WordSize = static_cast<size_t>(kArm64PointerSize);
53
54 // This constant is used as an approximate margin when emission of veneer and literal pools
55 // must be blocked.
56 static constexpr int kMaxMacroInstructionSizeInBytes = 15 * vixl::aarch64::kInstructionSize;
57
58 static const vixl::aarch64::Register kParameterCoreRegisters[] = {
59 vixl::aarch64::x1,
60 vixl::aarch64::x2,
61 vixl::aarch64::x3,
62 vixl::aarch64::x4,
63 vixl::aarch64::x5,
64 vixl::aarch64::x6,
65 vixl::aarch64::x7
66 };
67 static constexpr size_t kParameterCoreRegistersLength = arraysize(kParameterCoreRegisters);
68 static const vixl::aarch64::VRegister kParameterFPRegisters[] = {
69 vixl::aarch64::d0,
70 vixl::aarch64::d1,
71 vixl::aarch64::d2,
72 vixl::aarch64::d3,
73 vixl::aarch64::d4,
74 vixl::aarch64::d5,
75 vixl::aarch64::d6,
76 vixl::aarch64::d7
77 };
78 static constexpr size_t kParameterFPRegistersLength = arraysize(kParameterFPRegisters);
79
80 // Thread Register.
81 const vixl::aarch64::Register tr = vixl::aarch64::x19;
82 // Marking Register.
83 const vixl::aarch64::Register mr = vixl::aarch64::x20;
84 // Implicit suspend check register.
85 const vixl::aarch64::Register kImplicitSuspendCheckRegister = vixl::aarch64::x21;
86 // Method register on invoke.
87 static const vixl::aarch64::Register kArtMethodRegister = vixl::aarch64::x0;
88 const vixl::aarch64::CPURegList vixl_reserved_core_registers(vixl::aarch64::ip0,
89 vixl::aarch64::ip1);
90 const vixl::aarch64::CPURegList vixl_reserved_fp_registers(vixl::aarch64::d31);
91
92 const vixl::aarch64::CPURegList runtime_reserved_core_registers =
93 vixl::aarch64::CPURegList(
94 tr,
95 // Reserve X20 as Marking Register when emitting Baker read barriers.
96 // TODO: We don't need to reserve marking-register for userfaultfd GC. But
97 // that would require some work in the assembler code as the right GC is
98 // chosen at load-time and not compile time.
99 (kReserveMarkingRegister ? mr : vixl::aarch64::NoCPUReg),
100 kImplicitSuspendCheckRegister,
101 vixl::aarch64::lr);
102
103 // Some instructions have special requirements for a temporary, for example
104 // LoadClass/kBssEntry and LoadString/kBssEntry for Baker read barrier require
105 // temp that's not an R0 (to avoid an extra move) and Baker read barrier field
106 // loads with large offsets need a fixed register to limit the number of link-time
107 // thunks we generate. For these and similar cases, we want to reserve a specific
108 // register that's neither callee-save nor an argument register. We choose x15.
FixedTempLocation()109 inline Location FixedTempLocation() {
110 return Location::RegisterLocation(vixl::aarch64::x15.GetCode());
111 }
112
113 // Callee-save registers AAPCS64, without x19 (Thread Register) (nor
114 // x20 (Marking Register) when emitting Baker read barriers).
115 const vixl::aarch64::CPURegList callee_saved_core_registers(
116 vixl::aarch64::CPURegister::kRegister,
117 vixl::aarch64::kXRegSize,
118 (kReserveMarkingRegister ? vixl::aarch64::x21.GetCode() : vixl::aarch64::x20.GetCode()),
119 vixl::aarch64::x30.GetCode());
120 const vixl::aarch64::CPURegList callee_saved_fp_registers(vixl::aarch64::CPURegister::kVRegister,
121 vixl::aarch64::kDRegSize,
122 vixl::aarch64::d8.GetCode(),
123 vixl::aarch64::d15.GetCode());
124 Location ARM64ReturnLocation(DataType::Type return_type);
125
126 #define UNIMPLEMENTED_INTRINSIC_LIST_ARM64(V) \
127 V(IntegerRemainderUnsigned) \
128 V(LongRemainderUnsigned) \
129 V(StringStringIndexOf) \
130 V(StringStringIndexOfAfter) \
131 V(StringBufferAppend) \
132 V(StringBufferLength) \
133 V(StringBufferToString) \
134 V(StringBuilderAppendObject) \
135 V(StringBuilderAppendString) \
136 V(StringBuilderAppendCharSequence) \
137 V(StringBuilderAppendCharArray) \
138 V(StringBuilderAppendBoolean) \
139 V(StringBuilderAppendChar) \
140 V(StringBuilderAppendInt) \
141 V(StringBuilderAppendLong) \
142 V(StringBuilderAppendFloat) \
143 V(StringBuilderAppendDouble) \
144 V(StringBuilderLength) \
145 V(StringBuilderToString) \
146 V(SystemArrayCopyByte) \
147 V(SystemArrayCopyInt) \
148 /* 1.8 */ \
149 V(MethodHandleInvokeExact) \
150 V(MethodHandleInvoke)
151
152 class SlowPathCodeARM64 : public SlowPathCode {
153 public:
SlowPathCodeARM64(HInstruction * instruction)154 explicit SlowPathCodeARM64(HInstruction* instruction)
155 : SlowPathCode(instruction), entry_label_(), exit_label_() {}
156
GetEntryLabel()157 vixl::aarch64::Label* GetEntryLabel() { return &entry_label_; }
GetExitLabel()158 vixl::aarch64::Label* GetExitLabel() { return &exit_label_; }
159
160 void SaveLiveRegisters(CodeGenerator* codegen, LocationSummary* locations) override;
161 void RestoreLiveRegisters(CodeGenerator* codegen, LocationSummary* locations) override;
162
163 private:
164 vixl::aarch64::Label entry_label_;
165 vixl::aarch64::Label exit_label_;
166
167 DISALLOW_COPY_AND_ASSIGN(SlowPathCodeARM64);
168 };
169
170 class JumpTableARM64 : public DeletableArenaObject<kArenaAllocSwitchTable> {
171 public:
JumpTableARM64(HPackedSwitch * switch_instr)172 explicit JumpTableARM64(HPackedSwitch* switch_instr)
173 : switch_instr_(switch_instr), table_start_() {}
174
GetTableStartLabel()175 vixl::aarch64::Label* GetTableStartLabel() { return &table_start_; }
176
177 void EmitTable(CodeGeneratorARM64* codegen);
178
179 private:
180 HPackedSwitch* const switch_instr_;
181 vixl::aarch64::Label table_start_;
182
183 DISALLOW_COPY_AND_ASSIGN(JumpTableARM64);
184 };
185
186 static const vixl::aarch64::Register kRuntimeParameterCoreRegisters[] = {
187 vixl::aarch64::x0,
188 vixl::aarch64::x1,
189 vixl::aarch64::x2,
190 vixl::aarch64::x3,
191 vixl::aarch64::x4,
192 vixl::aarch64::x5,
193 vixl::aarch64::x6,
194 vixl::aarch64::x7
195 };
196 static constexpr size_t kRuntimeParameterCoreRegistersLength =
197 arraysize(kRuntimeParameterCoreRegisters);
198 static const vixl::aarch64::VRegister kRuntimeParameterFpuRegisters[] = {
199 vixl::aarch64::d0,
200 vixl::aarch64::d1,
201 vixl::aarch64::d2,
202 vixl::aarch64::d3,
203 vixl::aarch64::d4,
204 vixl::aarch64::d5,
205 vixl::aarch64::d6,
206 vixl::aarch64::d7
207 };
208 static constexpr size_t kRuntimeParameterFpuRegistersLength =
209 arraysize(kRuntimeParameterCoreRegisters);
210
211 class InvokeRuntimeCallingConvention : public CallingConvention<vixl::aarch64::Register,
212 vixl::aarch64::VRegister> {
213 public:
InvokeRuntimeCallingConvention()214 InvokeRuntimeCallingConvention()
215 : CallingConvention(kRuntimeParameterCoreRegisters,
216 kRuntimeParameterCoreRegistersLength,
217 kRuntimeParameterFpuRegisters,
218 kRuntimeParameterFpuRegistersLength,
219 kArm64PointerSize) {}
220
221 Location GetReturnLocation(DataType::Type return_type);
222
223 private:
224 DISALLOW_COPY_AND_ASSIGN(InvokeRuntimeCallingConvention);
225 };
226
227 class InvokeDexCallingConvention : public CallingConvention<vixl::aarch64::Register,
228 vixl::aarch64::VRegister> {
229 public:
InvokeDexCallingConvention()230 InvokeDexCallingConvention()
231 : CallingConvention(kParameterCoreRegisters,
232 kParameterCoreRegistersLength,
233 kParameterFPRegisters,
234 kParameterFPRegistersLength,
235 kArm64PointerSize) {}
236
GetReturnLocation(DataType::Type return_type)237 Location GetReturnLocation(DataType::Type return_type) const {
238 return ARM64ReturnLocation(return_type);
239 }
240
241
242 private:
243 DISALLOW_COPY_AND_ASSIGN(InvokeDexCallingConvention);
244 };
245
246 class InvokeDexCallingConventionVisitorARM64 : public InvokeDexCallingConventionVisitor {
247 public:
InvokeDexCallingConventionVisitorARM64()248 InvokeDexCallingConventionVisitorARM64() {}
~InvokeDexCallingConventionVisitorARM64()249 virtual ~InvokeDexCallingConventionVisitorARM64() {}
250
251 Location GetNextLocation(DataType::Type type) override;
GetReturnLocation(DataType::Type return_type)252 Location GetReturnLocation(DataType::Type return_type) const override {
253 return calling_convention.GetReturnLocation(return_type);
254 }
255 Location GetMethodLocation() const override;
256
257 private:
258 InvokeDexCallingConvention calling_convention;
259
260 DISALLOW_COPY_AND_ASSIGN(InvokeDexCallingConventionVisitorARM64);
261 };
262
263 class CriticalNativeCallingConventionVisitorARM64 : public InvokeDexCallingConventionVisitor {
264 public:
CriticalNativeCallingConventionVisitorARM64(bool for_register_allocation)265 explicit CriticalNativeCallingConventionVisitorARM64(bool for_register_allocation)
266 : for_register_allocation_(for_register_allocation) {}
267
~CriticalNativeCallingConventionVisitorARM64()268 virtual ~CriticalNativeCallingConventionVisitorARM64() {}
269
270 Location GetNextLocation(DataType::Type type) override;
271 Location GetReturnLocation(DataType::Type type) const override;
272 Location GetMethodLocation() const override;
273
GetStackOffset()274 size_t GetStackOffset() const { return stack_offset_; }
275
276 private:
277 // Register allocator does not support adjusting frame size, so we cannot provide final locations
278 // of stack arguments for register allocation. We ask the register allocator for any location and
279 // move these arguments to the right place after adjusting the SP when generating the call.
280 const bool for_register_allocation_;
281 size_t gpr_index_ = 0u;
282 size_t fpr_index_ = 0u;
283 size_t stack_offset_ = 0u;
284
285 DISALLOW_COPY_AND_ASSIGN(CriticalNativeCallingConventionVisitorARM64);
286 };
287
288 class FieldAccessCallingConventionARM64 : public FieldAccessCallingConvention {
289 public:
FieldAccessCallingConventionARM64()290 FieldAccessCallingConventionARM64() {}
291
GetObjectLocation()292 Location GetObjectLocation() const override {
293 return helpers::LocationFrom(vixl::aarch64::x1);
294 }
GetFieldIndexLocation()295 Location GetFieldIndexLocation() const override {
296 return helpers::LocationFrom(vixl::aarch64::x0);
297 }
GetReturnLocation(DataType::Type type)298 Location GetReturnLocation([[maybe_unused]] DataType::Type type) const override {
299 return helpers::LocationFrom(vixl::aarch64::x0);
300 }
GetSetValueLocation(DataType::Type type,bool is_instance)301 Location GetSetValueLocation([[maybe_unused]] DataType::Type type,
302 bool is_instance) const override {
303 return is_instance
304 ? helpers::LocationFrom(vixl::aarch64::x2)
305 : helpers::LocationFrom(vixl::aarch64::x1);
306 }
GetFpuLocation(DataType::Type type)307 Location GetFpuLocation([[maybe_unused]] DataType::Type type) const override {
308 return helpers::LocationFrom(vixl::aarch64::d0);
309 }
310
311 private:
312 DISALLOW_COPY_AND_ASSIGN(FieldAccessCallingConventionARM64);
313 };
314
315 class InstructionCodeGeneratorARM64 : public InstructionCodeGenerator {
316 public:
317 InstructionCodeGeneratorARM64(HGraph* graph, CodeGeneratorARM64* codegen);
318
319 #define DECLARE_VISIT_INSTRUCTION(name, super) \
320 void Visit##name(H##name* instr) override;
321
322 FOR_EACH_CONCRETE_INSTRUCTION_SCALAR_COMMON(DECLARE_VISIT_INSTRUCTION)
FOR_EACH_CONCRETE_INSTRUCTION_ARM64(DECLARE_VISIT_INSTRUCTION)323 FOR_EACH_CONCRETE_INSTRUCTION_ARM64(DECLARE_VISIT_INSTRUCTION)
324 FOR_EACH_CONCRETE_INSTRUCTION_SHARED(DECLARE_VISIT_INSTRUCTION)
325
326 #undef DECLARE_VISIT_INSTRUCTION
327
328 void VisitInstruction(HInstruction* instruction) override {
329 LOG(FATAL) << "Unreachable instruction " << instruction->DebugName()
330 << " (id " << instruction->GetId() << ")";
331 }
332
GetAssembler()333 Arm64Assembler* GetAssembler() const { return assembler_; }
GetVIXLAssembler()334 vixl::aarch64::MacroAssembler* GetVIXLAssembler() { return GetAssembler()->GetVIXLAssembler(); }
335
336 // SIMD helpers.
337 virtual Location AllocateSIMDScratchLocation(vixl::aarch64::UseScratchRegisterScope* scope) = 0;
338 virtual void FreeSIMDScratchLocation(Location loc,
339 vixl::aarch64::UseScratchRegisterScope* scope) = 0;
340 virtual void LoadSIMDRegFromStack(Location destination, Location source) = 0;
341 virtual void MoveSIMDRegToSIMDReg(Location destination, Location source) = 0;
342 virtual void MoveToSIMDStackSlot(Location destination, Location source) = 0;
343 virtual void SaveLiveRegistersHelper(LocationSummary* locations,
344 int64_t spill_offset) = 0;
345 virtual void RestoreLiveRegistersHelper(LocationSummary* locations,
346 int64_t spill_offset) = 0;
347
348 protected:
349 void GenerateClassInitializationCheck(SlowPathCodeARM64* slow_path,
350 vixl::aarch64::Register class_reg);
351 void GenerateBitstringTypeCheckCompare(HTypeCheckInstruction* check,
352 vixl::aarch64::Register temp);
353 void GenerateSuspendCheck(HSuspendCheck* instruction, HBasicBlock* successor);
354 void HandleBinaryOp(HBinaryOperation* instr);
355
356 void HandleFieldSet(HInstruction* instruction,
357 const FieldInfo& field_info,
358 bool value_can_be_null,
359 WriteBarrierKind write_barrier_kind);
360 void HandleFieldGet(HInstruction* instruction, const FieldInfo& field_info);
361 void HandleCondition(HCondition* instruction);
362
363 // Generate a heap reference load using one register `out`:
364 //
365 // out <- *(out + offset)
366 //
367 // while honoring heap poisoning and/or read barriers (if any).
368 //
369 // Location `maybe_temp` is used when generating a read barrier and
370 // shall be a register in that case; it may be an invalid location
371 // otherwise.
372 void GenerateReferenceLoadOneRegister(HInstruction* instruction,
373 Location out,
374 uint32_t offset,
375 Location maybe_temp,
376 ReadBarrierOption read_barrier_option);
377 // Generate a heap reference load using two different registers
378 // `out` and `obj`:
379 //
380 // out <- *(obj + offset)
381 //
382 // while honoring heap poisoning and/or read barriers (if any).
383 //
384 // Location `maybe_temp` is used when generating a Baker's (fast
385 // path) read barrier and shall be a register in that case; it may
386 // be an invalid location otherwise.
387 void GenerateReferenceLoadTwoRegisters(HInstruction* instruction,
388 Location out,
389 Location obj,
390 uint32_t offset,
391 Location maybe_temp,
392 ReadBarrierOption read_barrier_option);
393
394 // Generate a floating-point comparison.
395 void GenerateFcmp(HInstruction* instruction);
396
397 void HandleShift(HBinaryOperation* instr);
398 void GenerateTestAndBranch(HInstruction* instruction,
399 size_t condition_input_index,
400 vixl::aarch64::Label* true_target,
401 vixl::aarch64::Label* false_target);
402 void DivRemOneOrMinusOne(HBinaryOperation* instruction);
403 void DivRemByPowerOfTwo(HBinaryOperation* instruction);
404 void GenerateIncrementNegativeByOne(vixl::aarch64::Register out,
405 vixl::aarch64::Register in, bool use_cond_inc);
406 void GenerateResultRemWithAnyConstant(vixl::aarch64::Register out,
407 vixl::aarch64::Register dividend,
408 vixl::aarch64::Register quotient,
409 int64_t divisor,
410 // This function may acquire a scratch register.
411 vixl::aarch64::UseScratchRegisterScope* temps_scope);
412 void GenerateInt64UnsignedDivRemWithAnyPositiveConstant(HBinaryOperation* instruction);
413 void GenerateInt64DivRemWithAnyConstant(HBinaryOperation* instruction);
414 void GenerateInt32DivRemWithAnyConstant(HBinaryOperation* instruction);
415 void GenerateDivRemWithAnyConstant(HBinaryOperation* instruction, int64_t divisor);
416 void GenerateIntDiv(HDiv* instruction);
417 void GenerateIntDivForConstDenom(HDiv *instruction);
418 void GenerateIntDivForPower2Denom(HDiv *instruction);
419 void GenerateIntRem(HRem* instruction);
420 void GenerateIntRemForConstDenom(HRem *instruction);
421 void GenerateIntRemForPower2Denom(HRem *instruction);
422 void HandleGoto(HInstruction* got, HBasicBlock* successor);
423 void GenerateMethodEntryExitHook(HInstruction* instruction);
424
425 // Helpers to set up locations for vector memory operations. Returns the memory operand and,
426 // if used, sets the output parameter scratch to a temporary register used in this operand,
427 // so that the client can release it right after the memory operand use.
428 // Neon version.
429 vixl::aarch64::MemOperand VecNEONAddress(
430 HVecMemoryOperation* instruction,
431 // This function may acquire a scratch register.
432 vixl::aarch64::UseScratchRegisterScope* temps_scope,
433 size_t size,
434 bool is_string_char_at,
435 /*out*/ vixl::aarch64::Register* scratch);
436 // SVE version.
437 vixl::aarch64::SVEMemOperand VecSVEAddress(
438 HVecMemoryOperation* instruction,
439 // This function may acquire a scratch register.
440 vixl::aarch64::UseScratchRegisterScope* temps_scope,
441 size_t size,
442 bool is_string_char_at,
443 /*out*/ vixl::aarch64::Register* scratch);
444
445 Arm64Assembler* const assembler_;
446 CodeGeneratorARM64* const codegen_;
447
448 DISALLOW_COPY_AND_ASSIGN(InstructionCodeGeneratorARM64);
449 };
450
451 class LocationsBuilderARM64 : public HGraphVisitor {
452 public:
LocationsBuilderARM64(HGraph * graph,CodeGeneratorARM64 * codegen)453 LocationsBuilderARM64(HGraph* graph, CodeGeneratorARM64* codegen)
454 : HGraphVisitor(graph), codegen_(codegen) {}
455
456 #define DECLARE_VISIT_INSTRUCTION(name, super) \
457 void Visit##name(H##name* instr) override;
458
459 FOR_EACH_CONCRETE_INSTRUCTION_SCALAR_COMMON(DECLARE_VISIT_INSTRUCTION)
FOR_EACH_CONCRETE_INSTRUCTION_ARM64(DECLARE_VISIT_INSTRUCTION)460 FOR_EACH_CONCRETE_INSTRUCTION_ARM64(DECLARE_VISIT_INSTRUCTION)
461 FOR_EACH_CONCRETE_INSTRUCTION_SHARED(DECLARE_VISIT_INSTRUCTION)
462
463 #undef DECLARE_VISIT_INSTRUCTION
464
465 void VisitInstruction(HInstruction* instruction) override {
466 LOG(FATAL) << "Unreachable instruction " << instruction->DebugName()
467 << " (id " << instruction->GetId() << ")";
468 }
469
470 protected:
471 void HandleBinaryOp(HBinaryOperation* instr);
472 void HandleFieldSet(HInstruction* instruction);
473 void HandleFieldGet(HInstruction* instruction, const FieldInfo& field_info);
474 void HandleInvoke(HInvoke* instr);
475 void HandleCondition(HCondition* instruction);
476 void HandleShift(HBinaryOperation* instr);
477
478 CodeGeneratorARM64* const codegen_;
479 InvokeDexCallingConventionVisitorARM64 parameter_visitor_;
480
481 DISALLOW_COPY_AND_ASSIGN(LocationsBuilderARM64);
482 };
483
484 class InstructionCodeGeneratorARM64Neon : public InstructionCodeGeneratorARM64 {
485 public:
InstructionCodeGeneratorARM64Neon(HGraph * graph,CodeGeneratorARM64 * codegen)486 InstructionCodeGeneratorARM64Neon(HGraph* graph, CodeGeneratorARM64* codegen) :
487 InstructionCodeGeneratorARM64(graph, codegen) {}
488
489 #define DECLARE_VISIT_INSTRUCTION(name, super) \
490 void Visit##name(H##name* instr) override;
491
492 FOR_EACH_CONCRETE_INSTRUCTION_VECTOR_COMMON(DECLARE_VISIT_INSTRUCTION)
493
494 #undef DECLARE_VISIT_INSTRUCTION
495
496 Location AllocateSIMDScratchLocation(vixl::aarch64::UseScratchRegisterScope* scope) override;
497 void FreeSIMDScratchLocation(Location loc,
498 vixl::aarch64::UseScratchRegisterScope* scope) override;
499 void LoadSIMDRegFromStack(Location destination, Location source) override;
500 void MoveSIMDRegToSIMDReg(Location destination, Location source) override;
501 void MoveToSIMDStackSlot(Location destination, Location source) override;
502 void SaveLiveRegistersHelper(LocationSummary* locations, int64_t spill_offset) override;
503 void RestoreLiveRegistersHelper(LocationSummary* locations, int64_t spill_offset) override;
504 };
505
506 class LocationsBuilderARM64Neon : public LocationsBuilderARM64 {
507 public:
LocationsBuilderARM64Neon(HGraph * graph,CodeGeneratorARM64 * codegen)508 LocationsBuilderARM64Neon(HGraph* graph, CodeGeneratorARM64* codegen) :
509 LocationsBuilderARM64(graph, codegen) {}
510
511 #define DECLARE_VISIT_INSTRUCTION(name, super) \
512 void Visit##name(H##name* instr) override;
513
514 FOR_EACH_CONCRETE_INSTRUCTION_VECTOR_COMMON(DECLARE_VISIT_INSTRUCTION)
515
516 #undef DECLARE_VISIT_INSTRUCTION
517 };
518
519 class InstructionCodeGeneratorARM64Sve : public InstructionCodeGeneratorARM64 {
520 public:
InstructionCodeGeneratorARM64Sve(HGraph * graph,CodeGeneratorARM64 * codegen)521 InstructionCodeGeneratorARM64Sve(HGraph* graph, CodeGeneratorARM64* codegen) :
522 InstructionCodeGeneratorARM64(graph, codegen) {}
523
524 #define DECLARE_VISIT_INSTRUCTION(name, super) \
525 void Visit##name(H##name* instr) override;
526
527 FOR_EACH_CONCRETE_INSTRUCTION_VECTOR_COMMON(DECLARE_VISIT_INSTRUCTION)
528
529 #undef DECLARE_VISIT_INSTRUCTION
530
531 Location AllocateSIMDScratchLocation(vixl::aarch64::UseScratchRegisterScope* scope) override;
532 void FreeSIMDScratchLocation(Location loc,
533 vixl::aarch64::UseScratchRegisterScope* scope) override;
534 void LoadSIMDRegFromStack(Location destination, Location source) override;
535 void MoveSIMDRegToSIMDReg(Location destination, Location source) override;
536 void MoveToSIMDStackSlot(Location destination, Location source) override;
537 void SaveLiveRegistersHelper(LocationSummary* locations, int64_t spill_offset) override;
538 void RestoreLiveRegistersHelper(LocationSummary* locations, int64_t spill_offset) override;
539
540 private:
541 // Validate that instruction vector length and packed type are compliant with the SIMD
542 // register size (full SIMD register is used).
543 void ValidateVectorLength(HVecOperation* instr) const;
544
GetVecGoverningPReg(HVecOperation * instr)545 vixl::aarch64::PRegister GetVecGoverningPReg(HVecOperation* instr) {
546 return GetVecPredSetFixedOutPReg(instr->GetGoverningPredicate());
547 }
548
549 // Returns a fixed p-reg for predicate setting instruction.
550 //
551 // Currently we only support diamond CF loops for predicated vectorization; also we don't have
552 // register allocator support for vector predicates. Thus we use fixed P-regs for loop main,
553 // True and False predicates as a temporary solution.
554 //
555 // TODO: Support SIMD types and registers in ART.
GetVecPredSetFixedOutPReg(HVecPredSetOperation * instr)556 static vixl::aarch64::PRegister GetVecPredSetFixedOutPReg(HVecPredSetOperation* instr) {
557 if (instr->IsVecPredWhile() || instr->IsVecPredSetAll()) {
558 // VecPredWhile and VecPredSetAll live ranges never overlap due to the current vectorization
559 // scheme: the former only is live inside a vectorized loop and the later is never in a
560 // loop and never spans across loops.
561 return vixl::aarch64::p0;
562 } else if (instr->IsVecPredNot()) {
563 // This relies on the fact that we only use PredNot manually in the autovectorizer,
564 // so there is only one of them in each loop.
565 return vixl::aarch64::p1;
566 } else {
567 DCHECK(instr->IsVecCondition());
568 return vixl::aarch64::p2;
569 }
570 }
571 };
572
573 class LocationsBuilderARM64Sve : public LocationsBuilderARM64 {
574 public:
LocationsBuilderARM64Sve(HGraph * graph,CodeGeneratorARM64 * codegen)575 LocationsBuilderARM64Sve(HGraph* graph, CodeGeneratorARM64* codegen) :
576 LocationsBuilderARM64(graph, codegen) {}
577
578 #define DECLARE_VISIT_INSTRUCTION(name, super) \
579 void Visit##name(H##name* instr) override;
580
581 FOR_EACH_CONCRETE_INSTRUCTION_VECTOR_COMMON(DECLARE_VISIT_INSTRUCTION)
582
583 #undef DECLARE_VISIT_INSTRUCTION
584 };
585
586 class ParallelMoveResolverARM64 : public ParallelMoveResolverNoSwap {
587 public:
ParallelMoveResolverARM64(ArenaAllocator * allocator,CodeGeneratorARM64 * codegen)588 ParallelMoveResolverARM64(ArenaAllocator* allocator, CodeGeneratorARM64* codegen)
589 : ParallelMoveResolverNoSwap(allocator), codegen_(codegen), vixl_temps_() {}
590
591 protected:
592 void PrepareForEmitNativeCode() override;
593 void FinishEmitNativeCode() override;
594 Location AllocateScratchLocationFor(Location::Kind kind) override;
595 void FreeScratchLocation(Location loc) override;
596 void EmitMove(size_t index) override;
597
598 private:
599 Arm64Assembler* GetAssembler() const;
GetVIXLAssembler()600 vixl::aarch64::MacroAssembler* GetVIXLAssembler() const {
601 return GetAssembler()->GetVIXLAssembler();
602 }
603
604 CodeGeneratorARM64* const codegen_;
605 vixl::aarch64::UseScratchRegisterScope vixl_temps_;
606
607 DISALLOW_COPY_AND_ASSIGN(ParallelMoveResolverARM64);
608 };
609
610 class CodeGeneratorARM64 : public CodeGenerator {
611 public:
612 CodeGeneratorARM64(HGraph* graph,
613 const CompilerOptions& compiler_options,
614 OptimizingCompilerStats* stats = nullptr);
~CodeGeneratorARM64()615 virtual ~CodeGeneratorARM64() {}
616
617 void GenerateFrameEntry() override;
618 void GenerateFrameExit() override;
619
620 vixl::aarch64::CPURegList GetFramePreservedCoreRegisters() const;
621 vixl::aarch64::CPURegList GetFramePreservedFPRegisters() const;
622
623 void Bind(HBasicBlock* block) override;
624
GetLabelOf(HBasicBlock * block)625 vixl::aarch64::Label* GetLabelOf(HBasicBlock* block) {
626 block = FirstNonEmptyBlock(block);
627 return &(block_labels_[block->GetBlockId()]);
628 }
629
GetWordSize()630 size_t GetWordSize() const override {
631 return kArm64WordSize;
632 }
633
SupportsPredicatedSIMD()634 bool SupportsPredicatedSIMD() const override { return ShouldUseSVE(); }
635
GetSlowPathFPWidth()636 size_t GetSlowPathFPWidth() const override {
637 return GetGraph()->HasSIMD()
638 ? GetSIMDRegisterWidth()
639 : vixl::aarch64::kDRegSizeInBytes;
640 }
641
GetCalleePreservedFPWidth()642 size_t GetCalleePreservedFPWidth() const override {
643 return vixl::aarch64::kDRegSizeInBytes;
644 }
645
646 size_t GetSIMDRegisterWidth() const override;
647
GetAddressOf(HBasicBlock * block)648 uintptr_t GetAddressOf(HBasicBlock* block) override {
649 vixl::aarch64::Label* block_entry_label = GetLabelOf(block);
650 DCHECK(block_entry_label->IsBound());
651 return block_entry_label->GetLocation();
652 }
653
GetLocationBuilder()654 HGraphVisitor* GetLocationBuilder() override { return location_builder_; }
GetInstructionCodeGeneratorArm64()655 InstructionCodeGeneratorARM64* GetInstructionCodeGeneratorArm64() {
656 return instruction_visitor_;
657 }
GetInstructionVisitor()658 HGraphVisitor* GetInstructionVisitor() override { return GetInstructionCodeGeneratorArm64(); }
GetAssembler()659 Arm64Assembler* GetAssembler() override { return &assembler_; }
GetAssembler()660 const Arm64Assembler& GetAssembler() const override { return assembler_; }
GetVIXLAssembler()661 vixl::aarch64::MacroAssembler* GetVIXLAssembler() { return GetAssembler()->GetVIXLAssembler(); }
662
663 // Emit a write barrier if:
664 // A) emit_null_check is false
665 // B) emit_null_check is true, and value is not null.
666 void MaybeMarkGCCard(vixl::aarch64::Register object,
667 vixl::aarch64::Register value,
668 bool emit_null_check);
669
670 // Emit a write barrier unconditionally.
671 void MarkGCCard(vixl::aarch64::Register object);
672
673 // Crash if the card table is not valid. This check is only emitted for the CC GC. We assert
674 // `(!clean || !self->is_gc_marking)`, since the card table should not be set to clean when the CC
675 // GC is marking for eliminated write barriers.
676 void CheckGCCardIsValid(vixl::aarch64::Register object);
677
678 void GenerateMemoryBarrier(MemBarrierKind kind);
679
680 // Register allocation.
681
682 void SetupBlockedRegisters() const override;
683
684 size_t SaveCoreRegister(size_t stack_index, uint32_t reg_id) override;
685 size_t RestoreCoreRegister(size_t stack_index, uint32_t reg_id) override;
686 size_t SaveFloatingPointRegister(size_t stack_index, uint32_t reg_id) override;
687 size_t RestoreFloatingPointRegister(size_t stack_index, uint32_t reg_id) override;
688
689 // The number of registers that can be allocated. The register allocator may
690 // decide to reserve and not use a few of them.
691 // We do not consider registers sp, xzr, wzr. They are either not allocatable
692 // (xzr, wzr), or make for poor allocatable registers (sp alignment
693 // requirements, etc.). This also facilitates our task as all other registers
694 // can easily be mapped via to or from their type and index or code.
695 static const int kNumberOfAllocatableRegisters = vixl::aarch64::kNumberOfRegisters - 1;
696 static const int kNumberOfAllocatableFPRegisters = vixl::aarch64::kNumberOfVRegisters;
697 static constexpr int kNumberOfAllocatableRegisterPairs = 0;
698
699 void DumpCoreRegister(std::ostream& stream, int reg) const override;
700 void DumpFloatingPointRegister(std::ostream& stream, int reg) const override;
701
GetInstructionSet()702 InstructionSet GetInstructionSet() const override {
703 return InstructionSet::kArm64;
704 }
705
706 const Arm64InstructionSetFeatures& GetInstructionSetFeatures() const;
707
Initialize()708 void Initialize() override {
709 block_labels_.resize(GetGraph()->GetBlocks().size());
710 }
711
712 // We want to use the STP and LDP instructions to spill and restore registers for slow paths.
713 // These instructions can only encode offsets that are multiples of the register size accessed.
GetPreferredSlotsAlignment()714 uint32_t GetPreferredSlotsAlignment() const override { return vixl::aarch64::kXRegSizeInBytes; }
715
CreateJumpTable(HPackedSwitch * switch_instr)716 JumpTableARM64* CreateJumpTable(HPackedSwitch* switch_instr) {
717 jump_tables_.emplace_back(new (GetGraph()->GetAllocator()) JumpTableARM64(switch_instr));
718 return jump_tables_.back().get();
719 }
720
721 void Finalize() override;
722
723 // Code generation helpers.
724 void MoveConstant(vixl::aarch64::CPURegister destination, HConstant* constant);
725 void MoveConstant(Location destination, int32_t value) override;
726 void MoveLocation(Location dst, Location src, DataType::Type dst_type) override;
727 void AddLocationAsTemp(Location location, LocationSummary* locations) override;
728
729 void Load(DataType::Type type,
730 vixl::aarch64::CPURegister dst,
731 const vixl::aarch64::MemOperand& src);
732 void Store(DataType::Type type,
733 vixl::aarch64::CPURegister src,
734 const vixl::aarch64::MemOperand& dst);
735 void LoadAcquire(HInstruction* instruction,
736 DataType::Type type,
737 vixl::aarch64::CPURegister dst,
738 const vixl::aarch64::MemOperand& src,
739 bool needs_null_check);
740 void StoreRelease(HInstruction* instruction,
741 DataType::Type type,
742 vixl::aarch64::CPURegister src,
743 const vixl::aarch64::MemOperand& dst,
744 bool needs_null_check);
745
746 // Generate code to invoke a runtime entry point.
747 void InvokeRuntime(QuickEntrypointEnum entrypoint,
748 HInstruction* instruction,
749 uint32_t dex_pc,
750 SlowPathCode* slow_path = nullptr) override;
751
752 // Generate code to invoke a runtime entry point, but do not record
753 // PC-related information in a stack map.
754 void InvokeRuntimeWithoutRecordingPcInfo(int32_t entry_point_offset,
755 HInstruction* instruction,
756 SlowPathCode* slow_path);
757
GetMoveResolver()758 ParallelMoveResolverARM64* GetMoveResolver() override { return &move_resolver_; }
759
NeedsTwoRegisters(DataType::Type type)760 bool NeedsTwoRegisters([[maybe_unused]] DataType::Type type) const override { return false; }
761
762 // Check if the desired_string_load_kind is supported. If it is, return it,
763 // otherwise return a fall-back kind that should be used instead.
764 HLoadString::LoadKind GetSupportedLoadStringKind(
765 HLoadString::LoadKind desired_string_load_kind) override;
766
767 // Check if the desired_class_load_kind is supported. If it is, return it,
768 // otherwise return a fall-back kind that should be used instead.
769 HLoadClass::LoadKind GetSupportedLoadClassKind(
770 HLoadClass::LoadKind desired_class_load_kind) override;
771
772 // Check if the desired_dispatch_info is supported. If it is, return it,
773 // otherwise return a fall-back info that should be used instead.
774 HInvokeStaticOrDirect::DispatchInfo GetSupportedInvokeStaticOrDirectDispatch(
775 const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
776 ArtMethod* method) override;
777
778 void LoadMethod(MethodLoadKind load_kind, Location temp, HInvoke* invoke);
779 void GenerateStaticOrDirectCall(
780 HInvokeStaticOrDirect* invoke, Location temp, SlowPathCode* slow_path = nullptr) override;
781 void GenerateVirtualCall(
782 HInvokeVirtual* invoke, Location temp, SlowPathCode* slow_path = nullptr) override;
783
784 void MoveFromReturnRegister(Location trg, DataType::Type type) override;
785
786 // Add a new boot image intrinsic patch for an instruction and return the label
787 // to be bound before the instruction. The instruction will be either the
788 // ADRP (pass `adrp_label = null`) or the ADD (pass `adrp_label` pointing
789 // to the associated ADRP patch label).
790 vixl::aarch64::Label* NewBootImageIntrinsicPatch(uint32_t intrinsic_data,
791 vixl::aarch64::Label* adrp_label = nullptr);
792
793 // Add a new boot image relocation patch for an instruction and return the label
794 // to be bound before the instruction. The instruction will be either the
795 // ADRP (pass `adrp_label = null`) or the LDR (pass `adrp_label` pointing
796 // to the associated ADRP patch label).
797 vixl::aarch64::Label* NewBootImageRelRoPatch(uint32_t boot_image_offset,
798 vixl::aarch64::Label* adrp_label = nullptr);
799
800 // Add a new boot image method patch for an instruction and return the label
801 // to be bound before the instruction. The instruction will be either the
802 // ADRP (pass `adrp_label = null`) or the ADD (pass `adrp_label` pointing
803 // to the associated ADRP patch label).
804 vixl::aarch64::Label* NewBootImageMethodPatch(MethodReference target_method,
805 vixl::aarch64::Label* adrp_label = nullptr);
806
807 // Add a new .bss entry method patch for an instruction and return
808 // the label to be bound before the instruction. The instruction will be
809 // either the ADRP (pass `adrp_label = null`) or the LDR (pass `adrp_label`
810 // pointing to the associated ADRP patch label).
811 vixl::aarch64::Label* NewMethodBssEntryPatch(MethodReference target_method,
812 vixl::aarch64::Label* adrp_label = nullptr);
813
814 // Add a new boot image type patch for an instruction and return the label
815 // to be bound before the instruction. The instruction will be either the
816 // ADRP (pass `adrp_label = null`) or the ADD (pass `adrp_label` pointing
817 // to the associated ADRP patch label).
818 vixl::aarch64::Label* NewBootImageTypePatch(const DexFile& dex_file,
819 dex::TypeIndex type_index,
820 vixl::aarch64::Label* adrp_label = nullptr);
821
822 // Add a new app image type patch for an instruction and return the label
823 // to be bound before the instruction. The instruction will be either the
824 // ADRP (pass `adrp_label = null`) or the LDR (pass `adrp_label` pointing
825 // to the associated ADRP patch label).
826 vixl::aarch64::Label* NewAppImageTypePatch(const DexFile& dex_file,
827 dex::TypeIndex type_index,
828 vixl::aarch64::Label* adrp_label = nullptr);
829
830 // Add a new .bss entry type patch for an instruction and return the label
831 // to be bound before the instruction. The instruction will be either the
832 // ADRP (pass `adrp_label = null`) or the ADD (pass `adrp_label` pointing
833 // to the associated ADRP patch label).
834 vixl::aarch64::Label* NewBssEntryTypePatch(HLoadClass* load_class,
835 vixl::aarch64::Label* adrp_label = nullptr);
836
837 // Add a new boot image string patch for an instruction and return the label
838 // to be bound before the instruction. The instruction will be either the
839 // ADRP (pass `adrp_label = null`) or the ADD (pass `adrp_label` pointing
840 // to the associated ADRP patch label).
841 vixl::aarch64::Label* NewBootImageStringPatch(const DexFile& dex_file,
842 dex::StringIndex string_index,
843 vixl::aarch64::Label* adrp_label = nullptr);
844
845 // Add a new .bss entry string patch for an instruction and return the label
846 // to be bound before the instruction. The instruction will be either the
847 // ADRP (pass `adrp_label = null`) or the ADD (pass `adrp_label` pointing
848 // to the associated ADRP patch label).
849 vixl::aarch64::Label* NewStringBssEntryPatch(const DexFile& dex_file,
850 dex::StringIndex string_index,
851 vixl::aarch64::Label* adrp_label = nullptr);
852
853 // Add a new boot image JNI entrypoint patch for an instruction and return the label
854 // to be bound before the instruction. The instruction will be either the
855 // ADRP (pass `adrp_label = null`) or the LDR (pass `adrp_label` pointing
856 // to the associated ADRP patch label).
857 vixl::aarch64::Label* NewBootImageJniEntrypointPatch(MethodReference target_method,
858 vixl::aarch64::Label* adrp_label = nullptr);
859
860 // Emit the BL instruction for entrypoint thunk call and record the associated patch for AOT.
861 void EmitEntrypointThunkCall(ThreadOffset64 entrypoint_offset);
862
863 // Emit the CBNZ instruction for baker read barrier and record
864 // the associated patch for AOT or slow path for JIT.
865 void EmitBakerReadBarrierCbnz(uint32_t custom_data);
866
DeduplicateBootImageAddressLiteral(uint64_t address)867 vixl::aarch64::Literal<uint32_t>* DeduplicateBootImageAddressLiteral(uint64_t address) {
868 return jit_patches_.DeduplicateBootImageAddressLiteral(address);
869 }
DeduplicateJitStringLiteral(const DexFile & dex_file,dex::StringIndex string_index,Handle<mirror::String> handle)870 vixl::aarch64::Literal<uint32_t>* DeduplicateJitStringLiteral(const DexFile& dex_file,
871 dex::StringIndex string_index,
872 Handle<mirror::String> handle) {
873 return jit_patches_.DeduplicateJitStringLiteral(
874 dex_file, string_index, handle, GetCodeGenerationData());
875 }
DeduplicateJitClassLiteral(const DexFile & dex_file,dex::TypeIndex class_index,Handle<mirror::Class> handle)876 vixl::aarch64::Literal<uint32_t>* DeduplicateJitClassLiteral(const DexFile& dex_file,
877 dex::TypeIndex class_index,
878 Handle<mirror::Class> handle) {
879 return jit_patches_.DeduplicateJitClassLiteral(
880 dex_file, class_index, handle, GetCodeGenerationData());
881 }
882
883 void EmitAdrpPlaceholder(vixl::aarch64::Label* fixup_label, vixl::aarch64::Register reg);
884 void EmitAddPlaceholder(vixl::aarch64::Label* fixup_label,
885 vixl::aarch64::Register out,
886 vixl::aarch64::Register base);
887 void EmitLdrOffsetPlaceholder(vixl::aarch64::Label* fixup_label,
888 vixl::aarch64::Register out,
889 vixl::aarch64::Register base);
890
891 void LoadBootImageRelRoEntry(vixl::aarch64::Register reg, uint32_t boot_image_offset);
892 void LoadBootImageAddress(vixl::aarch64::Register reg, uint32_t boot_image_reference);
893 void LoadTypeForBootImageIntrinsic(vixl::aarch64::Register reg, TypeReference type_reference);
894 void LoadIntrinsicDeclaringClass(vixl::aarch64::Register reg, HInvoke* invoke);
895 void LoadClassRootForIntrinsic(vixl::aarch64::Register reg, ClassRoot class_root);
896
897 void EmitLinkerPatches(ArenaVector<linker::LinkerPatch>* linker_patches) override;
898 bool NeedsThunkCode(const linker::LinkerPatch& patch) const override;
899 void EmitThunkCode(const linker::LinkerPatch& patch,
900 /*out*/ ArenaVector<uint8_t>* code,
901 /*out*/ std::string* debug_name) override;
902
903 void EmitJitRootPatches(uint8_t* code, const uint8_t* roots_data) override;
904
905 // Generate a GC root reference load:
906 //
907 // root <- *(obj + offset)
908 //
909 // while honoring read barriers based on read_barrier_option.
910 void GenerateGcRootFieldLoad(HInstruction* instruction,
911 Location root,
912 vixl::aarch64::Register obj,
913 uint32_t offset,
914 vixl::aarch64::Label* fixup_label,
915 ReadBarrierOption read_barrier_option);
916 // Generate MOV for the `old_value` in intrinsic and mark it with Baker read barrier.
917 void GenerateIntrinsicMoveWithBakerReadBarrier(vixl::aarch64::Register marked_old_value,
918 vixl::aarch64::Register old_value);
919 // Fast path implementation of ReadBarrier::Barrier for a heap
920 // reference field load when Baker's read barriers are used.
921 // Overload suitable for Unsafe.getObject/-Volatile() intrinsic.
922 void GenerateFieldLoadWithBakerReadBarrier(HInstruction* instruction,
923 Location ref,
924 vixl::aarch64::Register obj,
925 const vixl::aarch64::MemOperand& src,
926 bool needs_null_check,
927 bool use_load_acquire);
928 // Fast path implementation of ReadBarrier::Barrier for a heap
929 // reference field load when Baker's read barriers are used.
930 void GenerateFieldLoadWithBakerReadBarrier(HInstruction* instruction,
931 Location ref,
932 vixl::aarch64::Register obj,
933 uint32_t offset,
934 Location maybe_temp,
935 bool needs_null_check,
936 bool use_load_acquire);
937 // Fast path implementation of ReadBarrier::Barrier for a heap
938 // reference array load when Baker's read barriers are used.
939 void GenerateArrayLoadWithBakerReadBarrier(HArrayGet* instruction,
940 Location ref,
941 vixl::aarch64::Register obj,
942 uint32_t data_offset,
943 Location index,
944 bool needs_null_check);
945
946 // Emit code checking the status of the Marking Register, and
947 // aborting the program if MR does not match the value stored in the
948 // art::Thread object. Code is only emitted in debug mode and if
949 // CompilerOptions::EmitRunTimeChecksInDebugMode returns true.
950 //
951 // Argument `code` is used to identify the different occurrences of
952 // MaybeGenerateMarkingRegisterCheck in the code generator, and is
953 // passed to the BRK instruction.
954 //
955 // If `temp_loc` is a valid location, it is expected to be a
956 // register and will be used as a temporary to generate code;
957 // otherwise, a temporary will be fetched from the core register
958 // scratch pool.
959 virtual void MaybeGenerateMarkingRegisterCheck(int code,
960 Location temp_loc = Location::NoLocation());
961
962 // Create slow path for a read barrier for a heap reference within `instruction`.
963 //
964 // This is a helper function for GenerateReadBarrierSlow() that has the same
965 // arguments. The creation and adding of the slow path is exposed for intrinsics
966 // that cannot use GenerateReadBarrierSlow() from their own slow paths.
967 SlowPathCodeARM64* AddReadBarrierSlowPath(HInstruction* instruction,
968 Location out,
969 Location ref,
970 Location obj,
971 uint32_t offset,
972 Location index);
973
974 // Generate a read barrier for a heap reference within `instruction`
975 // using a slow path.
976 //
977 // A read barrier for an object reference read from the heap is
978 // implemented as a call to the artReadBarrierSlow runtime entry
979 // point, which is passed the values in locations `ref`, `obj`, and
980 // `offset`:
981 //
982 // mirror::Object* artReadBarrierSlow(mirror::Object* ref,
983 // mirror::Object* obj,
984 // uint32_t offset);
985 //
986 // The `out` location contains the value returned by
987 // artReadBarrierSlow.
988 //
989 // When `index` is provided (i.e. for array accesses), the offset
990 // value passed to artReadBarrierSlow is adjusted to take `index`
991 // into account.
992 void GenerateReadBarrierSlow(HInstruction* instruction,
993 Location out,
994 Location ref,
995 Location obj,
996 uint32_t offset,
997 Location index = Location::NoLocation());
998
999 // If read barriers are enabled, generate a read barrier for a heap
1000 // reference using a slow path. If heap poisoning is enabled, also
1001 // unpoison the reference in `out`.
1002 void MaybeGenerateReadBarrierSlow(HInstruction* instruction,
1003 Location out,
1004 Location ref,
1005 Location obj,
1006 uint32_t offset,
1007 Location index = Location::NoLocation());
1008
1009 // Generate a read barrier for a GC root within `instruction` using
1010 // a slow path.
1011 //
1012 // A read barrier for an object reference GC root is implemented as
1013 // a call to the artReadBarrierForRootSlow runtime entry point,
1014 // which is passed the value in location `root`:
1015 //
1016 // mirror::Object* artReadBarrierForRootSlow(GcRoot<mirror::Object>* root);
1017 //
1018 // The `out` location contains the value returned by
1019 // artReadBarrierForRootSlow.
1020 void GenerateReadBarrierForRootSlow(HInstruction* instruction, Location out, Location root);
1021
1022 void IncreaseFrame(size_t adjustment) override;
1023 void DecreaseFrame(size_t adjustment) override;
1024
1025 void GenerateNop() override;
1026
1027 void GenerateImplicitNullCheck(HNullCheck* instruction) override;
1028 void GenerateExplicitNullCheck(HNullCheck* instruction) override;
1029
MaybeRecordImplicitNullCheck(HInstruction * instr)1030 void MaybeRecordImplicitNullCheck(HInstruction* instr) final {
1031 // The function must be only called within special scopes
1032 // (EmissionCheckScope, ExactAssemblyScope) which prevent generation of
1033 // veneer/literal pools by VIXL assembler.
1034 CHECK_EQ(GetVIXLAssembler()->ArePoolsBlocked(), true)
1035 << "The function must only be called within EmissionCheckScope or ExactAssemblyScope";
1036 CodeGenerator::MaybeRecordImplicitNullCheck(instr);
1037 }
1038
1039 void MaybeGenerateInlineCacheCheck(HInstruction* instruction, vixl::aarch64::Register klass);
1040 void MaybeIncrementHotness(HSuspendCheck* suspend_check, bool is_frame_entry);
1041
1042 bool CanUseImplicitSuspendCheck() const;
1043
1044 private:
1045 // Encoding of thunk type and data for link-time generated thunks for Baker read barriers.
1046
1047 enum class BakerReadBarrierKind : uint8_t {
1048 kField, // Field get or array get with constant offset (i.e. constant index).
1049 kAcquire, // Volatile field get.
1050 kArray, // Array get with index in register.
1051 kGcRoot, // GC root load.
1052 kLast = kGcRoot
1053 };
1054
1055 static constexpr uint32_t kBakerReadBarrierInvalidEncodedReg = /* sp/zr is invalid */ 31u;
1056
1057 static constexpr size_t kBitsForBakerReadBarrierKind =
1058 MinimumBitsToStore(static_cast<size_t>(BakerReadBarrierKind::kLast));
1059 static constexpr size_t kBakerReadBarrierBitsForRegister =
1060 MinimumBitsToStore(kBakerReadBarrierInvalidEncodedReg);
1061 using BakerReadBarrierKindField =
1062 BitField<BakerReadBarrierKind, 0, kBitsForBakerReadBarrierKind>;
1063 using BakerReadBarrierFirstRegField =
1064 BitField<uint32_t, kBitsForBakerReadBarrierKind, kBakerReadBarrierBitsForRegister>;
1065 using BakerReadBarrierSecondRegField =
1066 BitField<uint32_t,
1067 kBitsForBakerReadBarrierKind + kBakerReadBarrierBitsForRegister,
1068 kBakerReadBarrierBitsForRegister>;
1069
CheckValidReg(uint32_t reg)1070 static void CheckValidReg(uint32_t reg) {
1071 DCHECK(reg < vixl::aarch64::lr.GetCode() &&
1072 reg != vixl::aarch64::ip0.GetCode() &&
1073 reg != vixl::aarch64::ip1.GetCode()) << reg;
1074 }
1075
EncodeBakerReadBarrierFieldData(uint32_t base_reg,uint32_t holder_reg)1076 static inline uint32_t EncodeBakerReadBarrierFieldData(uint32_t base_reg, uint32_t holder_reg) {
1077 CheckValidReg(base_reg);
1078 CheckValidReg(holder_reg);
1079 return BakerReadBarrierKindField::Encode(BakerReadBarrierKind::kField) |
1080 BakerReadBarrierFirstRegField::Encode(base_reg) |
1081 BakerReadBarrierSecondRegField::Encode(holder_reg);
1082 }
1083
EncodeBakerReadBarrierAcquireData(uint32_t base_reg,uint32_t holder_reg)1084 static inline uint32_t EncodeBakerReadBarrierAcquireData(uint32_t base_reg, uint32_t holder_reg) {
1085 CheckValidReg(base_reg);
1086 CheckValidReg(holder_reg);
1087 DCHECK_NE(base_reg, holder_reg);
1088 return BakerReadBarrierKindField::Encode(BakerReadBarrierKind::kAcquire) |
1089 BakerReadBarrierFirstRegField::Encode(base_reg) |
1090 BakerReadBarrierSecondRegField::Encode(holder_reg);
1091 }
1092
EncodeBakerReadBarrierArrayData(uint32_t base_reg)1093 static inline uint32_t EncodeBakerReadBarrierArrayData(uint32_t base_reg) {
1094 CheckValidReg(base_reg);
1095 return BakerReadBarrierKindField::Encode(BakerReadBarrierKind::kArray) |
1096 BakerReadBarrierFirstRegField::Encode(base_reg) |
1097 BakerReadBarrierSecondRegField::Encode(kBakerReadBarrierInvalidEncodedReg);
1098 }
1099
EncodeBakerReadBarrierGcRootData(uint32_t root_reg)1100 static inline uint32_t EncodeBakerReadBarrierGcRootData(uint32_t root_reg) {
1101 CheckValidReg(root_reg);
1102 return BakerReadBarrierKindField::Encode(BakerReadBarrierKind::kGcRoot) |
1103 BakerReadBarrierFirstRegField::Encode(root_reg) |
1104 BakerReadBarrierSecondRegField::Encode(kBakerReadBarrierInvalidEncodedReg);
1105 }
1106
1107 void CompileBakerReadBarrierThunk(Arm64Assembler& assembler,
1108 uint32_t encoded_data,
1109 /*out*/ std::string* debug_name);
1110
1111 // The PcRelativePatchInfo is used for PC-relative addressing of methods/strings/types,
1112 // whether through .data.img.rel.ro, .bss, or directly in the boot image.
1113 struct PcRelativePatchInfo : PatchInfo<vixl::aarch64::Label> {
PcRelativePatchInfoPcRelativePatchInfo1114 PcRelativePatchInfo(const DexFile* dex_file, uint32_t off_or_idx)
1115 : PatchInfo<vixl::aarch64::Label>(dex_file, off_or_idx), pc_insn_label() { }
1116
1117 vixl::aarch64::Label* pc_insn_label;
1118 };
1119
1120 struct BakerReadBarrierPatchInfo {
BakerReadBarrierPatchInfoBakerReadBarrierPatchInfo1121 explicit BakerReadBarrierPatchInfo(uint32_t data) : label(), custom_data(data) { }
1122
1123 vixl::aarch64::Label label;
1124 uint32_t custom_data;
1125 };
1126
1127 vixl::aarch64::Label* NewPcRelativePatch(const DexFile* dex_file,
1128 uint32_t offset_or_index,
1129 vixl::aarch64::Label* adrp_label,
1130 ArenaDeque<PcRelativePatchInfo>* patches);
1131
1132 void EmitJumpTables();
1133
1134 template <linker::LinkerPatch (*Factory)(size_t, const DexFile*, uint32_t, uint32_t)>
1135 static void EmitPcRelativeLinkerPatches(const ArenaDeque<PcRelativePatchInfo>& infos,
1136 ArenaVector<linker::LinkerPatch>* linker_patches);
1137
1138 // Returns whether SVE features are supported and should be used.
1139 bool ShouldUseSVE() const;
1140
1141 // Labels for each block that will be compiled.
1142 // We use a deque so that the `vixl::aarch64::Label` objects do not move in memory.
1143 ArenaDeque<vixl::aarch64::Label> block_labels_; // Indexed by block id.
1144 vixl::aarch64::Label frame_entry_label_;
1145 ArenaVector<std::unique_ptr<JumpTableARM64>> jump_tables_;
1146
1147 LocationsBuilderARM64Neon location_builder_neon_;
1148 InstructionCodeGeneratorARM64Neon instruction_visitor_neon_;
1149 LocationsBuilderARM64Sve location_builder_sve_;
1150 InstructionCodeGeneratorARM64Sve instruction_visitor_sve_;
1151
1152 LocationsBuilderARM64* location_builder_;
1153 InstructionCodeGeneratorARM64* instruction_visitor_;
1154 ParallelMoveResolverARM64 move_resolver_;
1155 Arm64Assembler assembler_;
1156
1157 // PC-relative method patch info for kBootImageLinkTimePcRelative.
1158 ArenaDeque<PcRelativePatchInfo> boot_image_method_patches_;
1159 // PC-relative method patch info for kBssEntry.
1160 ArenaDeque<PcRelativePatchInfo> method_bss_entry_patches_;
1161 // PC-relative type patch info for kBootImageLinkTimePcRelative.
1162 ArenaDeque<PcRelativePatchInfo> boot_image_type_patches_;
1163 // PC-relative type patch info for kAppImageRelRo.
1164 ArenaDeque<PcRelativePatchInfo> app_image_type_patches_;
1165 // PC-relative type patch info for kBssEntry.
1166 ArenaDeque<PcRelativePatchInfo> type_bss_entry_patches_;
1167 // PC-relative public type patch info for kBssEntryPublic.
1168 ArenaDeque<PcRelativePatchInfo> public_type_bss_entry_patches_;
1169 // PC-relative package type patch info for kBssEntryPackage.
1170 ArenaDeque<PcRelativePatchInfo> package_type_bss_entry_patches_;
1171 // PC-relative String patch info for kBootImageLinkTimePcRelative.
1172 ArenaDeque<PcRelativePatchInfo> boot_image_string_patches_;
1173 // PC-relative String patch info for kBssEntry.
1174 ArenaDeque<PcRelativePatchInfo> string_bss_entry_patches_;
1175 // PC-relative method patch info for kBootImageLinkTimePcRelative+kCallCriticalNative.
1176 ArenaDeque<PcRelativePatchInfo> boot_image_jni_entrypoint_patches_;
1177 // PC-relative patch info for IntrinsicObjects for the boot image,
1178 // and for method/type/string patches for kBootImageRelRo otherwise.
1179 ArenaDeque<PcRelativePatchInfo> boot_image_other_patches_;
1180 // Patch info for calls to entrypoint dispatch thunks. Used for slow paths.
1181 ArenaDeque<PatchInfo<vixl::aarch64::Label>> call_entrypoint_patches_;
1182 // Baker read barrier patch info.
1183 ArenaDeque<BakerReadBarrierPatchInfo> baker_read_barrier_patches_;
1184
1185 JitPatchesARM64 jit_patches_;
1186
1187 // Baker read barrier slow paths, mapping custom data (uint32_t) to label.
1188 // Wrap the label to work around vixl::aarch64::Label being non-copyable
1189 // and non-moveable and as such unusable in ArenaSafeMap<>.
1190 struct LabelWrapper {
LabelWrapperLabelWrapper1191 LabelWrapper(const LabelWrapper& src)
1192 : label() {
1193 DCHECK(!src.label.IsLinked() && !src.label.IsBound());
1194 }
1195 LabelWrapper() = default;
1196 vixl::aarch64::Label label;
1197 };
1198 ArenaSafeMap<uint32_t, LabelWrapper> jit_baker_read_barrier_slow_paths_;
1199
1200 friend class linker::Arm64RelativePatcherTest;
1201 DISALLOW_COPY_AND_ASSIGN(CodeGeneratorARM64);
1202 };
1203
GetAssembler()1204 inline Arm64Assembler* ParallelMoveResolverARM64::GetAssembler() const {
1205 return codegen_->GetAssembler();
1206 }
1207
1208 } // namespace arm64
1209 } // namespace art
1210
1211 #endif // ART_COMPILER_OPTIMIZING_CODE_GENERATOR_ARM64_H_
1212