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_X86_64_H_
18 #define ART_COMPILER_OPTIMIZING_CODE_GENERATOR_X86_64_H_
19 
20 #include "arch/x86_64/instruction_set_features_x86_64.h"
21 #include "code_generator.h"
22 #include "driver/compiler_options.h"
23 #include "nodes.h"
24 #include "parallel_move_resolver.h"
25 #include "utils/x86_64/assembler_x86_64.h"
26 
27 namespace art {
28 namespace x86_64 {
29 
30 // Use a local definition to prevent copying mistakes.
31 static constexpr size_t kX86_64WordSize = static_cast<size_t>(kX86_64PointerSize);
32 
33 // Some x86_64 instructions require a register to be available as temp.
34 static constexpr Register TMP = R11;
35 
36 static constexpr Register kParameterCoreRegisters[] = { RSI, RDX, RCX, R8, R9 };
37 static constexpr FloatRegister kParameterFloatRegisters[] =
38     { XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7 };
39 
40 static constexpr size_t kParameterCoreRegistersLength = arraysize(kParameterCoreRegisters);
41 static constexpr size_t kParameterFloatRegistersLength = arraysize(kParameterFloatRegisters);
42 
43 static constexpr Register kRuntimeParameterCoreRegisters[] = { RDI, RSI, RDX, RCX };
44 static constexpr size_t kRuntimeParameterCoreRegistersLength =
45     arraysize(kRuntimeParameterCoreRegisters);
46 static constexpr FloatRegister kRuntimeParameterFpuRegisters[] = { XMM0, XMM1 };
47 static constexpr size_t kRuntimeParameterFpuRegistersLength =
48     arraysize(kRuntimeParameterFpuRegisters);
49 
50 // These XMM registers are non-volatile in ART ABI, but volatile in native ABI.
51 // If the ART ABI changes, this list must be updated.  It is used to ensure that
52 // these are not clobbered by any direct call to native code (such as math intrinsics).
53 static constexpr FloatRegister non_volatile_xmm_regs[] = { XMM12, XMM13, XMM14, XMM15 };
54 
55 
56 class InvokeRuntimeCallingConvention : public CallingConvention<Register, FloatRegister> {
57  public:
InvokeRuntimeCallingConvention()58   InvokeRuntimeCallingConvention()
59       : CallingConvention(kRuntimeParameterCoreRegisters,
60                           kRuntimeParameterCoreRegistersLength,
61                           kRuntimeParameterFpuRegisters,
62                           kRuntimeParameterFpuRegistersLength,
63                           kX86_64PointerSize) {}
64 
65  private:
66   DISALLOW_COPY_AND_ASSIGN(InvokeRuntimeCallingConvention);
67 };
68 
69 class InvokeDexCallingConvention : public CallingConvention<Register, FloatRegister> {
70  public:
InvokeDexCallingConvention()71   InvokeDexCallingConvention() : CallingConvention(
72       kParameterCoreRegisters,
73       kParameterCoreRegistersLength,
74       kParameterFloatRegisters,
75       kParameterFloatRegistersLength,
76       kX86_64PointerSize) {}
77 
78  private:
79   DISALLOW_COPY_AND_ASSIGN(InvokeDexCallingConvention);
80 };
81 
82 class FieldAccessCallingConventionX86_64 : public FieldAccessCallingConvention {
83  public:
FieldAccessCallingConventionX86_64()84   FieldAccessCallingConventionX86_64() {}
85 
GetObjectLocation()86   Location GetObjectLocation() const override {
87     return Location::RegisterLocation(RSI);
88   }
GetFieldIndexLocation()89   Location GetFieldIndexLocation() const override {
90     return Location::RegisterLocation(RDI);
91   }
GetReturnLocation(DataType::Type type ATTRIBUTE_UNUSED)92   Location GetReturnLocation(DataType::Type type ATTRIBUTE_UNUSED) const override {
93     return Location::RegisterLocation(RAX);
94   }
GetSetValueLocation(DataType::Type type ATTRIBUTE_UNUSED,bool is_instance)95   Location GetSetValueLocation(DataType::Type type ATTRIBUTE_UNUSED, bool is_instance)
96       const override {
97     return is_instance
98         ? Location::RegisterLocation(RDX)
99         : Location::RegisterLocation(RSI);
100   }
GetFpuLocation(DataType::Type type ATTRIBUTE_UNUSED)101   Location GetFpuLocation(DataType::Type type ATTRIBUTE_UNUSED) const override {
102     return Location::FpuRegisterLocation(XMM0);
103   }
104 
105  private:
106   DISALLOW_COPY_AND_ASSIGN(FieldAccessCallingConventionX86_64);
107 };
108 
109 
110 class InvokeDexCallingConventionVisitorX86_64 : public InvokeDexCallingConventionVisitor {
111  public:
InvokeDexCallingConventionVisitorX86_64()112   InvokeDexCallingConventionVisitorX86_64() {}
~InvokeDexCallingConventionVisitorX86_64()113   virtual ~InvokeDexCallingConventionVisitorX86_64() {}
114 
115   Location GetNextLocation(DataType::Type type) override;
116   Location GetReturnLocation(DataType::Type type) const override;
117   Location GetMethodLocation() const override;
118 
119  private:
120   InvokeDexCallingConvention calling_convention;
121 
122   DISALLOW_COPY_AND_ASSIGN(InvokeDexCallingConventionVisitorX86_64);
123 };
124 
125 class CodeGeneratorX86_64;
126 
127 class ParallelMoveResolverX86_64 : public ParallelMoveResolverWithSwap {
128  public:
ParallelMoveResolverX86_64(ArenaAllocator * allocator,CodeGeneratorX86_64 * codegen)129   ParallelMoveResolverX86_64(ArenaAllocator* allocator, CodeGeneratorX86_64* codegen)
130       : ParallelMoveResolverWithSwap(allocator), codegen_(codegen) {}
131 
132   void EmitMove(size_t index) override;
133   void EmitSwap(size_t index) override;
134   void SpillScratch(int reg) override;
135   void RestoreScratch(int reg) override;
136 
137   X86_64Assembler* GetAssembler() const;
138 
139  private:
140   void Exchange32(CpuRegister reg, int mem);
141   void Exchange32(XmmRegister reg, int mem);
142   void Exchange64(CpuRegister reg1, CpuRegister reg2);
143   void Exchange64(CpuRegister reg, int mem);
144   void Exchange64(XmmRegister reg, int mem);
145   void Exchange128(XmmRegister reg, int mem);
146   void ExchangeMemory32(int mem1, int mem2);
147   void ExchangeMemory64(int mem1, int mem2, int num_of_qwords);
148 
149   CodeGeneratorX86_64* const codegen_;
150 
151   DISALLOW_COPY_AND_ASSIGN(ParallelMoveResolverX86_64);
152 };
153 
154 class LocationsBuilderX86_64 : public HGraphVisitor {
155  public:
LocationsBuilderX86_64(HGraph * graph,CodeGeneratorX86_64 * codegen)156   LocationsBuilderX86_64(HGraph* graph, CodeGeneratorX86_64* codegen)
157       : HGraphVisitor(graph), codegen_(codegen) {}
158 
159 #define DECLARE_VISIT_INSTRUCTION(name, super)     \
160   void Visit##name(H##name* instr) override;
161 
162   FOR_EACH_CONCRETE_INSTRUCTION_COMMON(DECLARE_VISIT_INSTRUCTION)
FOR_EACH_CONCRETE_INSTRUCTION_X86_64(DECLARE_VISIT_INSTRUCTION)163   FOR_EACH_CONCRETE_INSTRUCTION_X86_64(DECLARE_VISIT_INSTRUCTION)
164   FOR_EACH_CONCRETE_INSTRUCTION_X86_COMMON(DECLARE_VISIT_INSTRUCTION)
165 
166 #undef DECLARE_VISIT_INSTRUCTION
167 
168   void VisitInstruction(HInstruction* instruction) override {
169     LOG(FATAL) << "Unreachable instruction " << instruction->DebugName()
170                << " (id " << instruction->GetId() << ")";
171   }
172 
173  private:
174   void HandleInvoke(HInvoke* invoke);
175   void HandleBitwiseOperation(HBinaryOperation* operation);
176   void HandleCondition(HCondition* condition);
177   void HandleShift(HBinaryOperation* operation);
178   void HandleFieldSet(HInstruction* instruction, const FieldInfo& field_info);
179   void HandleFieldGet(HInstruction* instruction);
180 
181   CodeGeneratorX86_64* const codegen_;
182   InvokeDexCallingConventionVisitorX86_64 parameter_visitor_;
183 
184   DISALLOW_COPY_AND_ASSIGN(LocationsBuilderX86_64);
185 };
186 
187 class InstructionCodeGeneratorX86_64 : public InstructionCodeGenerator {
188  public:
189   InstructionCodeGeneratorX86_64(HGraph* graph, CodeGeneratorX86_64* codegen);
190 
191 #define DECLARE_VISIT_INSTRUCTION(name, super)     \
192   void Visit##name(H##name* instr) override;
193 
194   FOR_EACH_CONCRETE_INSTRUCTION_COMMON(DECLARE_VISIT_INSTRUCTION)
FOR_EACH_CONCRETE_INSTRUCTION_X86_64(DECLARE_VISIT_INSTRUCTION)195   FOR_EACH_CONCRETE_INSTRUCTION_X86_64(DECLARE_VISIT_INSTRUCTION)
196   FOR_EACH_CONCRETE_INSTRUCTION_X86_COMMON(DECLARE_VISIT_INSTRUCTION)
197 
198 #undef DECLARE_VISIT_INSTRUCTION
199 
200   void VisitInstruction(HInstruction* instruction) override {
201     LOG(FATAL) << "Unreachable instruction " << instruction->DebugName()
202                << " (id " << instruction->GetId() << ")";
203   }
204 
GetAssembler()205   X86_64Assembler* GetAssembler() const { return assembler_; }
206 
207  private:
208   // Generate code for the given suspend check. If not null, `successor`
209   // is the block to branch to if the suspend check is not needed, and after
210   // the suspend call.
211   void GenerateSuspendCheck(HSuspendCheck* instruction, HBasicBlock* successor);
212   void GenerateClassInitializationCheck(SlowPathCode* slow_path, CpuRegister class_reg);
213   void GenerateBitstringTypeCheckCompare(HTypeCheckInstruction* check, CpuRegister temp);
214   void HandleBitwiseOperation(HBinaryOperation* operation);
215   void GenerateRemFP(HRem* rem);
216   void DivRemOneOrMinusOne(HBinaryOperation* instruction);
217   void DivByPowerOfTwo(HDiv* instruction);
218   void RemByPowerOfTwo(HRem* instruction);
219   void GenerateDivRemWithAnyConstant(HBinaryOperation* instruction);
220   void GenerateDivRemIntegral(HBinaryOperation* instruction);
221   void HandleCondition(HCondition* condition);
222   void HandleShift(HBinaryOperation* operation);
223 
224   void HandleFieldSet(HInstruction* instruction,
225                       const FieldInfo& field_info,
226                       bool value_can_be_null);
227   void HandleFieldGet(HInstruction* instruction, const FieldInfo& field_info);
228 
229   void GenerateMinMaxInt(LocationSummary* locations, bool is_min, DataType::Type type);
230   void GenerateMinMaxFP(LocationSummary* locations, bool is_min, DataType::Type type);
231   void GenerateMinMax(HBinaryOperation* minmax, bool is_min);
232 
233   // Generate a heap reference load using one register `out`:
234   //
235   //   out <- *(out + offset)
236   //
237   // while honoring heap poisoning and/or read barriers (if any).
238   //
239   // Location `maybe_temp` is used when generating a read barrier and
240   // shall be a register in that case; it may be an invalid location
241   // otherwise.
242   void GenerateReferenceLoadOneRegister(HInstruction* instruction,
243                                         Location out,
244                                         uint32_t offset,
245                                         Location maybe_temp,
246                                         ReadBarrierOption read_barrier_option);
247   // Generate a heap reference load using two different registers
248   // `out` and `obj`:
249   //
250   //   out <- *(obj + offset)
251   //
252   // while honoring heap poisoning and/or read barriers (if any).
253   //
254   // Location `maybe_temp` is used when generating a Baker's (fast
255   // path) read barrier and shall be a register in that case; it may
256   // be an invalid location otherwise.
257   void GenerateReferenceLoadTwoRegisters(HInstruction* instruction,
258                                          Location out,
259                                          Location obj,
260                                          uint32_t offset,
261                                          ReadBarrierOption read_barrier_option);
262   // Generate a GC root reference load:
263   //
264   //   root <- *address
265   //
266   // while honoring read barriers based on read_barrier_option.
267   void GenerateGcRootFieldLoad(HInstruction* instruction,
268                                Location root,
269                                const Address& address,
270                                Label* fixup_label,
271                                ReadBarrierOption read_barrier_option);
272 
273   void PushOntoFPStack(Location source, uint32_t temp_offset,
274                        uint32_t stack_adjustment, bool is_float);
275   void GenerateCompareTest(HCondition* condition);
276   template<class LabelType>
277   void GenerateTestAndBranch(HInstruction* instruction,
278                              size_t condition_input_index,
279                              LabelType* true_target,
280                              LabelType* false_target);
281   template<class LabelType>
282   void GenerateCompareTestAndBranch(HCondition* condition,
283                                     LabelType* true_target,
284                                     LabelType* false_target);
285   template<class LabelType>
286   void GenerateFPJumps(HCondition* cond, LabelType* true_label, LabelType* false_label);
287 
288   void HandleGoto(HInstruction* got, HBasicBlock* successor);
289 
290   X86_64Assembler* const assembler_;
291   CodeGeneratorX86_64* const codegen_;
292 
293   DISALLOW_COPY_AND_ASSIGN(InstructionCodeGeneratorX86_64);
294 };
295 
296 // Class for fixups to jump tables.
297 class JumpTableRIPFixup;
298 
299 class CodeGeneratorX86_64 : public CodeGenerator {
300  public:
301   CodeGeneratorX86_64(HGraph* graph,
302                   const CompilerOptions& compiler_options,
303                   OptimizingCompilerStats* stats = nullptr);
~CodeGeneratorX86_64()304   virtual ~CodeGeneratorX86_64() {}
305 
306   void GenerateFrameEntry() override;
307   void GenerateFrameExit() override;
308   void Bind(HBasicBlock* block) override;
309   void MoveConstant(Location destination, int32_t value) override;
310   void MoveLocation(Location dst, Location src, DataType::Type dst_type) override;
311   void AddLocationAsTemp(Location location, LocationSummary* locations) override;
312 
313   size_t SaveCoreRegister(size_t stack_index, uint32_t reg_id) override;
314   size_t RestoreCoreRegister(size_t stack_index, uint32_t reg_id) override;
315   size_t SaveFloatingPointRegister(size_t stack_index, uint32_t reg_id) override;
316   size_t RestoreFloatingPointRegister(size_t stack_index, uint32_t reg_id) override;
317 
318   // Generate code to invoke a runtime entry point.
319   void InvokeRuntime(QuickEntrypointEnum entrypoint,
320                      HInstruction* instruction,
321                      uint32_t dex_pc,
322                      SlowPathCode* slow_path = nullptr) override;
323 
324   // Generate code to invoke a runtime entry point, but do not record
325   // PC-related information in a stack map.
326   void InvokeRuntimeWithoutRecordingPcInfo(int32_t entry_point_offset,
327                                            HInstruction* instruction,
328                                            SlowPathCode* slow_path);
329 
330   void GenerateInvokeRuntime(int32_t entry_point_offset);
331 
GetWordSize()332   size_t GetWordSize() const override {
333     return kX86_64WordSize;
334   }
335 
GetFloatingPointSpillSlotSize()336   size_t GetFloatingPointSpillSlotSize() const override {
337     return GetGraph()->HasSIMD()
338         ? 2 * kX86_64WordSize   // 16 bytes == 2 x86_64 words for each spill
339         : 1 * kX86_64WordSize;  //  8 bytes == 1 x86_64 words for each spill
340   }
341 
GetLocationBuilder()342   HGraphVisitor* GetLocationBuilder() override {
343     return &location_builder_;
344   }
345 
GetInstructionVisitor()346   HGraphVisitor* GetInstructionVisitor() override {
347     return &instruction_visitor_;
348   }
349 
GetAssembler()350   X86_64Assembler* GetAssembler() override {
351     return &assembler_;
352   }
353 
GetAssembler()354   const X86_64Assembler& GetAssembler() const override {
355     return assembler_;
356   }
357 
GetMoveResolver()358   ParallelMoveResolverX86_64* GetMoveResolver() override {
359     return &move_resolver_;
360   }
361 
GetAddressOf(HBasicBlock * block)362   uintptr_t GetAddressOf(HBasicBlock* block) override {
363     return GetLabelOf(block)->Position();
364   }
365 
366   void SetupBlockedRegisters() const override;
367   void DumpCoreRegister(std::ostream& stream, int reg) const override;
368   void DumpFloatingPointRegister(std::ostream& stream, int reg) const override;
369   void Finalize(CodeAllocator* allocator) override;
370 
GetInstructionSet()371   InstructionSet GetInstructionSet() const override {
372     return InstructionSet::kX86_64;
373   }
374 
375   const X86_64InstructionSetFeatures& GetInstructionSetFeatures() const;
376 
377   // Emit a write barrier.
378   void MarkGCCard(CpuRegister temp,
379                   CpuRegister card,
380                   CpuRegister object,
381                   CpuRegister value,
382                   bool value_can_be_null);
383 
384   void GenerateMemoryBarrier(MemBarrierKind kind);
385 
386   // Helper method to move a value between two locations.
387   void Move(Location destination, Location source);
388 
GetLabelOf(HBasicBlock * block)389   Label* GetLabelOf(HBasicBlock* block) const {
390     return CommonGetLabelOf<Label>(block_labels_, block);
391   }
392 
Initialize()393   void Initialize() override {
394     block_labels_ = CommonInitializeLabels<Label>();
395   }
396 
NeedsTwoRegisters(DataType::Type type ATTRIBUTE_UNUSED)397   bool NeedsTwoRegisters(DataType::Type type ATTRIBUTE_UNUSED) const override {
398     return false;
399   }
400 
401   // Check if the desired_string_load_kind is supported. If it is, return it,
402   // otherwise return a fall-back kind that should be used instead.
403   HLoadString::LoadKind GetSupportedLoadStringKind(
404       HLoadString::LoadKind desired_string_load_kind) override;
405 
406   // Check if the desired_class_load_kind is supported. If it is, return it,
407   // otherwise return a fall-back kind that should be used instead.
408   HLoadClass::LoadKind GetSupportedLoadClassKind(
409       HLoadClass::LoadKind desired_class_load_kind) override;
410 
411   // Check if the desired_dispatch_info is supported. If it is, return it,
412   // otherwise return a fall-back info that should be used instead.
413   HInvokeStaticOrDirect::DispatchInfo GetSupportedInvokeStaticOrDirectDispatch(
414       const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
415       ArtMethod* method) override;
416 
417   void GenerateStaticOrDirectCall(
418       HInvokeStaticOrDirect* invoke, Location temp, SlowPathCode* slow_path = nullptr) override;
419   void GenerateVirtualCall(
420       HInvokeVirtual* invoke, Location temp, SlowPathCode* slow_path = nullptr) override;
421 
422   void RecordBootImageIntrinsicPatch(uint32_t intrinsic_data);
423   void RecordBootImageRelRoPatch(uint32_t boot_image_offset);
424   void RecordBootImageMethodPatch(HInvokeStaticOrDirect* invoke);
425   void RecordMethodBssEntryPatch(HInvokeStaticOrDirect* invoke);
426   void RecordBootImageTypePatch(HLoadClass* load_class);
427   Label* NewTypeBssEntryPatch(HLoadClass* load_class);
428   void RecordBootImageStringPatch(HLoadString* load_string);
429   Label* NewStringBssEntryPatch(HLoadString* load_string);
430   Label* NewJitRootStringPatch(const DexFile& dex_file,
431                                dex::StringIndex string_index,
432                                Handle<mirror::String> handle);
433   Label* NewJitRootClassPatch(const DexFile& dex_file,
434                               dex::TypeIndex type_index,
435                               Handle<mirror::Class> handle);
436 
437   void LoadBootImageAddress(CpuRegister reg, uint32_t boot_image_reference);
438   void AllocateInstanceForIntrinsic(HInvokeStaticOrDirect* invoke, uint32_t boot_image_offset);
439 
440   void EmitLinkerPatches(ArenaVector<linker::LinkerPatch>* linker_patches) override;
441 
442   void PatchJitRootUse(uint8_t* code,
443                        const uint8_t* roots_data,
444                        const PatchInfo<Label>& info,
445                        uint64_t index_in_table) const;
446 
447   void EmitJitRootPatches(uint8_t* code, const uint8_t* roots_data) override;
448 
449   // Fast path implementation of ReadBarrier::Barrier for a heap
450   // reference field load when Baker's read barriers are used.
451   void GenerateFieldLoadWithBakerReadBarrier(HInstruction* instruction,
452                                              Location ref,
453                                              CpuRegister obj,
454                                              uint32_t offset,
455                                              bool needs_null_check);
456   // Fast path implementation of ReadBarrier::Barrier for a heap
457   // reference array load when Baker's read barriers are used.
458   void GenerateArrayLoadWithBakerReadBarrier(HInstruction* instruction,
459                                              Location ref,
460                                              CpuRegister obj,
461                                              uint32_t data_offset,
462                                              Location index,
463                                              bool needs_null_check);
464   // Factored implementation, used by GenerateFieldLoadWithBakerReadBarrier,
465   // GenerateArrayLoadWithBakerReadBarrier and some intrinsics.
466   //
467   // Load the object reference located at address `src`, held by
468   // object `obj`, into `ref`, and mark it if needed.  The base of
469   // address `src` must be `obj`.
470   //
471   // If `always_update_field` is true, the value of the reference is
472   // atomically updated in the holder (`obj`).  This operation
473   // requires two temporary registers, which must be provided as
474   // non-null pointers (`temp1` and `temp2`).
475   void GenerateReferenceLoadWithBakerReadBarrier(HInstruction* instruction,
476                                                  Location ref,
477                                                  CpuRegister obj,
478                                                  const Address& src,
479                                                  bool needs_null_check,
480                                                  bool always_update_field = false,
481                                                  CpuRegister* temp1 = nullptr,
482                                                  CpuRegister* temp2 = nullptr);
483 
484   // Generate a read barrier for a heap reference within `instruction`
485   // using a slow path.
486   //
487   // A read barrier for an object reference read from the heap is
488   // implemented as a call to the artReadBarrierSlow runtime entry
489   // point, which is passed the values in locations `ref`, `obj`, and
490   // `offset`:
491   //
492   //   mirror::Object* artReadBarrierSlow(mirror::Object* ref,
493   //                                      mirror::Object* obj,
494   //                                      uint32_t offset);
495   //
496   // The `out` location contains the value returned by
497   // artReadBarrierSlow.
498   //
499   // When `index` provided (i.e., when it is different from
500   // Location::NoLocation()), the offset value passed to
501   // artReadBarrierSlow is adjusted to take `index` into account.
502   void GenerateReadBarrierSlow(HInstruction* instruction,
503                                Location out,
504                                Location ref,
505                                Location obj,
506                                uint32_t offset,
507                                Location index = Location::NoLocation());
508 
509   // If read barriers are enabled, generate a read barrier for a heap
510   // reference using a slow path. If heap poisoning is enabled, also
511   // unpoison the reference in `out`.
512   void MaybeGenerateReadBarrierSlow(HInstruction* instruction,
513                                     Location out,
514                                     Location ref,
515                                     Location obj,
516                                     uint32_t offset,
517                                     Location index = Location::NoLocation());
518 
519   // Generate a read barrier for a GC root within `instruction` using
520   // a slow path.
521   //
522   // A read barrier for an object reference GC root is implemented as
523   // a call to the artReadBarrierForRootSlow runtime entry point,
524   // which is passed the value in location `root`:
525   //
526   //   mirror::Object* artReadBarrierForRootSlow(GcRoot<mirror::Object>* root);
527   //
528   // The `out` location contains the value returned by
529   // artReadBarrierForRootSlow.
530   void GenerateReadBarrierForRootSlow(HInstruction* instruction, Location out, Location root);
531 
ConstantAreaStart()532   int ConstantAreaStart() const {
533     return constant_area_start_;
534   }
535 
536   Address LiteralDoubleAddress(double v);
537   Address LiteralFloatAddress(float v);
538   Address LiteralInt32Address(int32_t v);
539   Address LiteralInt64Address(int64_t v);
540 
541   // Load a 32/64-bit value into a register in the most efficient manner.
542   void Load32BitValue(CpuRegister dest, int32_t value);
543   void Load64BitValue(CpuRegister dest, int64_t value);
544   void Load32BitValue(XmmRegister dest, int32_t value);
545   void Load64BitValue(XmmRegister dest, int64_t value);
546   void Load32BitValue(XmmRegister dest, float value);
547   void Load64BitValue(XmmRegister dest, double value);
548 
549   // Compare a register with a 32/64-bit value in the most efficient manner.
550   void Compare32BitValue(CpuRegister dest, int32_t value);
551   void Compare64BitValue(CpuRegister dest, int64_t value);
552 
553   // Compare int values. Supports register locations for `lhs`.
554   void GenerateIntCompare(Location lhs, Location rhs);
555   void GenerateIntCompare(CpuRegister lhs, Location rhs);
556 
557   // Compare long values. Supports only register locations for `lhs`.
558   void GenerateLongCompare(Location lhs, Location rhs);
559 
560   // Construct address for array access.
561   static Address ArrayAddress(CpuRegister obj,
562                               Location index,
563                               ScaleFactor scale,
564                               uint32_t data_offset);
565 
566   Address LiteralCaseTable(HPackedSwitch* switch_instr);
567 
568   // Store a 64 bit value into a DoubleStackSlot in the most efficient manner.
569   void Store64BitValueToStack(Location dest, int64_t value);
570 
571   void MoveFromReturnRegister(Location trg, DataType::Type type) override;
572 
573   // Assign a 64 bit constant to an address.
574   void MoveInt64ToAddress(const Address& addr_low,
575                           const Address& addr_high,
576                           int64_t v,
577                           HInstruction* instruction);
578 
579   // Ensure that prior stores complete to memory before subsequent loads.
580   // The locked add implementation will avoid serializing device memory, but will
581   // touch (but not change) the top of the stack.
582   // The 'non_temporal' parameter should be used to ensure ordering of non-temporal stores.
583   void MemoryFence(bool force_mfence = false) {
584     if (!force_mfence) {
585       assembler_.lock()->addl(Address(CpuRegister(RSP), 0), Immediate(0));
586     } else {
587       assembler_.mfence();
588     }
589   }
590 
591   void GenerateNop() override;
592   void GenerateImplicitNullCheck(HNullCheck* instruction) override;
593   void GenerateExplicitNullCheck(HNullCheck* instruction) override;
594 
595   // When we don't know the proper offset for the value, we use kDummy32BitOffset.
596   // We will fix this up in the linker later to have the right value.
597   static constexpr int32_t kDummy32BitOffset = 256;
598 
599  private:
600   template <linker::LinkerPatch (*Factory)(size_t, const DexFile*, uint32_t, uint32_t)>
601   static void EmitPcRelativeLinkerPatches(const ArenaDeque<PatchInfo<Label>>& infos,
602                                           ArenaVector<linker::LinkerPatch>* linker_patches);
603 
604   // Labels for each block that will be compiled.
605   Label* block_labels_;  // Indexed by block id.
606   Label frame_entry_label_;
607   LocationsBuilderX86_64 location_builder_;
608   InstructionCodeGeneratorX86_64 instruction_visitor_;
609   ParallelMoveResolverX86_64 move_resolver_;
610   X86_64Assembler assembler_;
611 
612   // Offset to the start of the constant area in the assembled code.
613   // Used for fixups to the constant area.
614   int constant_area_start_;
615 
616   // PC-relative method patch info for kBootImageLinkTimePcRelative/kBootImageRelRo.
617   // Also used for type/string patches for kBootImageRelRo (same linker patch as for methods).
618   ArenaDeque<PatchInfo<Label>> boot_image_method_patches_;
619   // PC-relative method patch info for kBssEntry.
620   ArenaDeque<PatchInfo<Label>> method_bss_entry_patches_;
621   // PC-relative type patch info for kBootImageLinkTimePcRelative.
622   ArenaDeque<PatchInfo<Label>> boot_image_type_patches_;
623   // PC-relative type patch info for kBssEntry.
624   ArenaDeque<PatchInfo<Label>> type_bss_entry_patches_;
625   // PC-relative String patch info for kBootImageLinkTimePcRelative.
626   ArenaDeque<PatchInfo<Label>> boot_image_string_patches_;
627   // PC-relative String patch info for kBssEntry.
628   ArenaDeque<PatchInfo<Label>> string_bss_entry_patches_;
629   // PC-relative patch info for IntrinsicObjects.
630   ArenaDeque<PatchInfo<Label>> boot_image_intrinsic_patches_;
631 
632   // Patches for string literals in JIT compiled code.
633   ArenaDeque<PatchInfo<Label>> jit_string_patches_;
634   // Patches for class literals in JIT compiled code.
635   ArenaDeque<PatchInfo<Label>> jit_class_patches_;
636 
637   // Fixups for jump tables need to be handled specially.
638   ArenaVector<JumpTableRIPFixup*> fixups_to_jump_tables_;
639 
640   DISALLOW_COPY_AND_ASSIGN(CodeGeneratorX86_64);
641 };
642 
643 }  // namespace x86_64
644 }  // namespace art
645 
646 #endif  // ART_COMPILER_OPTIMIZING_CODE_GENERATOR_X86_64_H_
647