• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
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(Primitive::Type type ATTRIBUTE_UNUSED)92   Location GetReturnLocation(Primitive::Type type ATTRIBUTE_UNUSED) const OVERRIDE {
93     return Location::RegisterLocation(RAX);
94   }
GetSetValueLocation(Primitive::Type type ATTRIBUTE_UNUSED,bool is_instance)95   Location GetSetValueLocation(Primitive::Type type ATTRIBUTE_UNUSED, bool is_instance)
96       const OVERRIDE {
97     return is_instance
98         ? Location::RegisterLocation(RDX)
99         : Location::RegisterLocation(RSI);
100   }
GetFpuLocation(Primitive::Type type ATTRIBUTE_UNUSED)101   Location GetFpuLocation(Primitive::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(Primitive::Type type) OVERRIDE;
116   Location GetReturnLocation(Primitive::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 Exchange32(int mem1, int mem2);
143   void Exchange64(CpuRegister reg1, CpuRegister reg2);
144   void Exchange64(CpuRegister reg, int mem);
145   void Exchange64(XmmRegister reg, int mem);
146   void Exchange64(int mem1, int mem2);
147 
148   CodeGeneratorX86_64* const codegen_;
149 
150   DISALLOW_COPY_AND_ASSIGN(ParallelMoveResolverX86_64);
151 };
152 
153 class LocationsBuilderX86_64 : public HGraphVisitor {
154  public:
LocationsBuilderX86_64(HGraph * graph,CodeGeneratorX86_64 * codegen)155   LocationsBuilderX86_64(HGraph* graph, CodeGeneratorX86_64* codegen)
156       : HGraphVisitor(graph), codegen_(codegen) {}
157 
158 #define DECLARE_VISIT_INSTRUCTION(name, super)     \
159   void Visit##name(H##name* instr) OVERRIDE;
160 
161   FOR_EACH_CONCRETE_INSTRUCTION_COMMON(DECLARE_VISIT_INSTRUCTION)
FOR_EACH_CONCRETE_INSTRUCTION_X86_64(DECLARE_VISIT_INSTRUCTION)162   FOR_EACH_CONCRETE_INSTRUCTION_X86_64(DECLARE_VISIT_INSTRUCTION)
163 
164 #undef DECLARE_VISIT_INSTRUCTION
165 
166   void VisitInstruction(HInstruction* instruction) OVERRIDE {
167     LOG(FATAL) << "Unreachable instruction " << instruction->DebugName()
168                << " (id " << instruction->GetId() << ")";
169   }
170 
171  private:
172   void HandleInvoke(HInvoke* invoke);
173   void HandleBitwiseOperation(HBinaryOperation* operation);
174   void HandleCondition(HCondition* condition);
175   void HandleShift(HBinaryOperation* operation);
176   void HandleFieldSet(HInstruction* instruction, const FieldInfo& field_info);
177   void HandleFieldGet(HInstruction* instruction);
178 
179   CodeGeneratorX86_64* const codegen_;
180   InvokeDexCallingConventionVisitorX86_64 parameter_visitor_;
181 
182   DISALLOW_COPY_AND_ASSIGN(LocationsBuilderX86_64);
183 };
184 
185 class InstructionCodeGeneratorX86_64 : public InstructionCodeGenerator {
186  public:
187   InstructionCodeGeneratorX86_64(HGraph* graph, CodeGeneratorX86_64* codegen);
188 
189 #define DECLARE_VISIT_INSTRUCTION(name, super)     \
190   void Visit##name(H##name* instr) OVERRIDE;
191 
192   FOR_EACH_CONCRETE_INSTRUCTION_COMMON(DECLARE_VISIT_INSTRUCTION)
FOR_EACH_CONCRETE_INSTRUCTION_X86_64(DECLARE_VISIT_INSTRUCTION)193   FOR_EACH_CONCRETE_INSTRUCTION_X86_64(DECLARE_VISIT_INSTRUCTION)
194 
195 #undef DECLARE_VISIT_INSTRUCTION
196 
197   void VisitInstruction(HInstruction* instruction) OVERRIDE {
198     LOG(FATAL) << "Unreachable instruction " << instruction->DebugName()
199                << " (id " << instruction->GetId() << ")";
200   }
201 
GetAssembler()202   X86_64Assembler* GetAssembler() const { return assembler_; }
203 
204  private:
205   // Generate code for the given suspend check. If not null, `successor`
206   // is the block to branch to if the suspend check is not needed, and after
207   // the suspend call.
208   void GenerateSuspendCheck(HSuspendCheck* instruction, HBasicBlock* successor);
209   void GenerateClassInitializationCheck(SlowPathCode* slow_path, CpuRegister class_reg);
210   void HandleBitwiseOperation(HBinaryOperation* operation);
211   void GenerateRemFP(HRem* rem);
212   void DivRemOneOrMinusOne(HBinaryOperation* instruction);
213   void DivByPowerOfTwo(HDiv* instruction);
214   void GenerateDivRemWithAnyConstant(HBinaryOperation* instruction);
215   void GenerateDivRemIntegral(HBinaryOperation* instruction);
216   void HandleCondition(HCondition* condition);
217   void HandleShift(HBinaryOperation* operation);
218 
219   void HandleFieldSet(HInstruction* instruction,
220                       const FieldInfo& field_info,
221                       bool value_can_be_null);
222   void HandleFieldGet(HInstruction* instruction, const FieldInfo& field_info);
223 
224   // Generate a heap reference load using one register `out`:
225   //
226   //   out <- *(out + offset)
227   //
228   // while honoring heap poisoning and/or read barriers (if any).
229   //
230   // Location `maybe_temp` is used when generating a read barrier and
231   // shall be a register in that case; it may be an invalid location
232   // otherwise.
233   void GenerateReferenceLoadOneRegister(HInstruction* instruction,
234                                         Location out,
235                                         uint32_t offset,
236                                         Location maybe_temp,
237                                         ReadBarrierOption read_barrier_option);
238   // Generate a heap reference load using two different registers
239   // `out` and `obj`:
240   //
241   //   out <- *(obj + offset)
242   //
243   // while honoring heap poisoning and/or read barriers (if any).
244   //
245   // Location `maybe_temp` is used when generating a Baker's (fast
246   // path) read barrier and shall be a register in that case; it may
247   // be an invalid location otherwise.
248   void GenerateReferenceLoadTwoRegisters(HInstruction* instruction,
249                                          Location out,
250                                          Location obj,
251                                          uint32_t offset,
252                                          ReadBarrierOption read_barrier_option);
253   // Generate a GC root reference load:
254   //
255   //   root <- *address
256   //
257   // while honoring read barriers based on read_barrier_option.
258   void GenerateGcRootFieldLoad(HInstruction* instruction,
259                                Location root,
260                                const Address& address,
261                                Label* fixup_label,
262                                ReadBarrierOption read_barrier_option);
263 
264   void PushOntoFPStack(Location source, uint32_t temp_offset,
265                        uint32_t stack_adjustment, bool is_float);
266   void GenerateCompareTest(HCondition* condition);
267   template<class LabelType>
268   void GenerateTestAndBranch(HInstruction* instruction,
269                              size_t condition_input_index,
270                              LabelType* true_target,
271                              LabelType* false_target);
272   template<class LabelType>
273   void GenerateCompareTestAndBranch(HCondition* condition,
274                                     LabelType* true_target,
275                                     LabelType* false_target);
276   template<class LabelType>
277   void GenerateFPJumps(HCondition* cond, LabelType* true_label, LabelType* false_label);
278 
279   void HandleGoto(HInstruction* got, HBasicBlock* successor);
280 
281   X86_64Assembler* const assembler_;
282   CodeGeneratorX86_64* const codegen_;
283 
284   DISALLOW_COPY_AND_ASSIGN(InstructionCodeGeneratorX86_64);
285 };
286 
287 // Class for fixups to jump tables.
288 class JumpTableRIPFixup;
289 
290 class CodeGeneratorX86_64 : public CodeGenerator {
291  public:
292   CodeGeneratorX86_64(HGraph* graph,
293                   const X86_64InstructionSetFeatures& isa_features,
294                   const CompilerOptions& compiler_options,
295                   OptimizingCompilerStats* stats = nullptr);
~CodeGeneratorX86_64()296   virtual ~CodeGeneratorX86_64() {}
297 
298   void GenerateFrameEntry() OVERRIDE;
299   void GenerateFrameExit() OVERRIDE;
300   void Bind(HBasicBlock* block) OVERRIDE;
301   void MoveConstant(Location destination, int32_t value) OVERRIDE;
302   void MoveLocation(Location dst, Location src, Primitive::Type dst_type) OVERRIDE;
303   void AddLocationAsTemp(Location location, LocationSummary* locations) OVERRIDE;
304 
305   size_t SaveCoreRegister(size_t stack_index, uint32_t reg_id) OVERRIDE;
306   size_t RestoreCoreRegister(size_t stack_index, uint32_t reg_id) OVERRIDE;
307   size_t SaveFloatingPointRegister(size_t stack_index, uint32_t reg_id) OVERRIDE;
308   size_t RestoreFloatingPointRegister(size_t stack_index, uint32_t reg_id) OVERRIDE;
309 
310   // Generate code to invoke a runtime entry point.
311   void InvokeRuntime(QuickEntrypointEnum entrypoint,
312                      HInstruction* instruction,
313                      uint32_t dex_pc,
314                      SlowPathCode* slow_path = nullptr) OVERRIDE;
315 
316   // Generate code to invoke a runtime entry point, but do not record
317   // PC-related information in a stack map.
318   void InvokeRuntimeWithoutRecordingPcInfo(int32_t entry_point_offset,
319                                            HInstruction* instruction,
320                                            SlowPathCode* slow_path);
321 
322   void GenerateInvokeRuntime(int32_t entry_point_offset);
323 
GetWordSize()324   size_t GetWordSize() const OVERRIDE {
325     return kX86_64WordSize;
326   }
327 
GetFloatingPointSpillSlotSize()328   size_t GetFloatingPointSpillSlotSize() const OVERRIDE {
329     return GetGraph()->HasSIMD()
330         ? 2 * kX86_64WordSize   // 16 bytes == 2 x86_64 words for each spill
331         : 1 * kX86_64WordSize;  //  8 bytes == 1 x86_64 words for each spill
332   }
333 
GetLocationBuilder()334   HGraphVisitor* GetLocationBuilder() OVERRIDE {
335     return &location_builder_;
336   }
337 
GetInstructionVisitor()338   HGraphVisitor* GetInstructionVisitor() OVERRIDE {
339     return &instruction_visitor_;
340   }
341 
GetAssembler()342   X86_64Assembler* GetAssembler() OVERRIDE {
343     return &assembler_;
344   }
345 
GetAssembler()346   const X86_64Assembler& GetAssembler() const OVERRIDE {
347     return assembler_;
348   }
349 
GetMoveResolver()350   ParallelMoveResolverX86_64* GetMoveResolver() OVERRIDE {
351     return &move_resolver_;
352   }
353 
GetAddressOf(HBasicBlock * block)354   uintptr_t GetAddressOf(HBasicBlock* block) OVERRIDE {
355     return GetLabelOf(block)->Position();
356   }
357 
358   void SetupBlockedRegisters() const OVERRIDE;
359   void DumpCoreRegister(std::ostream& stream, int reg) const OVERRIDE;
360   void DumpFloatingPointRegister(std::ostream& stream, int reg) const OVERRIDE;
361   void Finalize(CodeAllocator* allocator) OVERRIDE;
362 
GetInstructionSet()363   InstructionSet GetInstructionSet() const OVERRIDE {
364     return InstructionSet::kX86_64;
365   }
366 
367   // Emit a write barrier.
368   void MarkGCCard(CpuRegister temp,
369                   CpuRegister card,
370                   CpuRegister object,
371                   CpuRegister value,
372                   bool value_can_be_null);
373 
374   void GenerateMemoryBarrier(MemBarrierKind kind);
375 
376   // Helper method to move a value between two locations.
377   void Move(Location destination, Location source);
378 
GetLabelOf(HBasicBlock * block)379   Label* GetLabelOf(HBasicBlock* block) const {
380     return CommonGetLabelOf<Label>(block_labels_, block);
381   }
382 
Initialize()383   void Initialize() OVERRIDE {
384     block_labels_ = CommonInitializeLabels<Label>();
385   }
386 
NeedsTwoRegisters(Primitive::Type type ATTRIBUTE_UNUSED)387   bool NeedsTwoRegisters(Primitive::Type type ATTRIBUTE_UNUSED) const OVERRIDE {
388     return false;
389   }
390 
391   // Check if the desired_string_load_kind is supported. If it is, return it,
392   // otherwise return a fall-back kind that should be used instead.
393   HLoadString::LoadKind GetSupportedLoadStringKind(
394       HLoadString::LoadKind desired_string_load_kind) OVERRIDE;
395 
396   // Check if the desired_class_load_kind is supported. If it is, return it,
397   // otherwise return a fall-back kind that should be used instead.
398   HLoadClass::LoadKind GetSupportedLoadClassKind(
399       HLoadClass::LoadKind desired_class_load_kind) OVERRIDE;
400 
401   // Check if the desired_dispatch_info is supported. If it is, return it,
402   // otherwise return a fall-back info that should be used instead.
403   HInvokeStaticOrDirect::DispatchInfo GetSupportedInvokeStaticOrDirectDispatch(
404       const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
405       HInvokeStaticOrDirect* invoke) OVERRIDE;
406 
407   Location GenerateCalleeMethodStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp);
408   void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) OVERRIDE;
409   void GenerateVirtualCall(HInvokeVirtual* invoke, Location temp) OVERRIDE;
410 
411   void RecordBootStringPatch(HLoadString* load_string);
412   void RecordBootTypePatch(HLoadClass* load_class);
413   Label* NewTypeBssEntryPatch(HLoadClass* load_class);
414   Label* NewStringBssEntryPatch(HLoadString* load_string);
415   Label* NewPcRelativeDexCacheArrayPatch(const DexFile& dex_file, uint32_t element_offset);
416   Label* NewJitRootStringPatch(const DexFile& dex_file,
417                                dex::StringIndex dex_index,
418                                Handle<mirror::String> handle);
419   Label* NewJitRootClassPatch(const DexFile& dex_file,
420                               dex::TypeIndex dex_index,
421                               Handle<mirror::Class> handle);
422 
423   void MoveFromReturnRegister(Location trg, Primitive::Type type) OVERRIDE;
424 
425   void EmitLinkerPatches(ArenaVector<LinkerPatch>* linker_patches) OVERRIDE;
426 
427   void PatchJitRootUse(uint8_t* code,
428                        const uint8_t* roots_data,
429                        const PatchInfo<Label>& info,
430                        uint64_t index_in_table) const;
431 
432   void EmitJitRootPatches(uint8_t* code, const uint8_t* roots_data) OVERRIDE;
433 
GetInstructionSetFeatures()434   const X86_64InstructionSetFeatures& GetInstructionSetFeatures() const {
435     return isa_features_;
436   }
437 
438   // Fast path implementation of ReadBarrier::Barrier for a heap
439   // reference field load when Baker's read barriers are used.
440   void GenerateFieldLoadWithBakerReadBarrier(HInstruction* instruction,
441                                              Location ref,
442                                              CpuRegister obj,
443                                              uint32_t offset,
444                                              bool needs_null_check);
445   // Fast path implementation of ReadBarrier::Barrier for a heap
446   // reference array load when Baker's read barriers are used.
447   void GenerateArrayLoadWithBakerReadBarrier(HInstruction* instruction,
448                                              Location ref,
449                                              CpuRegister obj,
450                                              uint32_t data_offset,
451                                              Location index,
452                                              bool needs_null_check);
453   // Factored implementation, used by GenerateFieldLoadWithBakerReadBarrier,
454   // GenerateArrayLoadWithBakerReadBarrier and some intrinsics.
455   //
456   // Load the object reference located at address `src`, held by
457   // object `obj`, into `ref`, and mark it if needed.  The base of
458   // address `src` must be `obj`.
459   //
460   // If `always_update_field` is true, the value of the reference is
461   // atomically updated in the holder (`obj`).  This operation
462   // requires two temporary registers, which must be provided as
463   // non-null pointers (`temp1` and `temp2`).
464   void GenerateReferenceLoadWithBakerReadBarrier(HInstruction* instruction,
465                                                  Location ref,
466                                                  CpuRegister obj,
467                                                  const Address& src,
468                                                  bool needs_null_check,
469                                                  bool always_update_field = false,
470                                                  CpuRegister* temp1 = nullptr,
471                                                  CpuRegister* temp2 = nullptr);
472 
473   // Generate a read barrier for a heap reference within `instruction`
474   // using a slow path.
475   //
476   // A read barrier for an object reference read from the heap is
477   // implemented as a call to the artReadBarrierSlow runtime entry
478   // point, which is passed the values in locations `ref`, `obj`, and
479   // `offset`:
480   //
481   //   mirror::Object* artReadBarrierSlow(mirror::Object* ref,
482   //                                      mirror::Object* obj,
483   //                                      uint32_t offset);
484   //
485   // The `out` location contains the value returned by
486   // artReadBarrierSlow.
487   //
488   // When `index` provided (i.e., when it is different from
489   // Location::NoLocation()), the offset value passed to
490   // artReadBarrierSlow is adjusted to take `index` into account.
491   void GenerateReadBarrierSlow(HInstruction* instruction,
492                                Location out,
493                                Location ref,
494                                Location obj,
495                                uint32_t offset,
496                                Location index = Location::NoLocation());
497 
498   // If read barriers are enabled, generate a read barrier for a heap
499   // reference using a slow path. If heap poisoning is enabled, also
500   // unpoison the reference in `out`.
501   void MaybeGenerateReadBarrierSlow(HInstruction* instruction,
502                                     Location out,
503                                     Location ref,
504                                     Location obj,
505                                     uint32_t offset,
506                                     Location index = Location::NoLocation());
507 
508   // Generate a read barrier for a GC root within `instruction` using
509   // a slow path.
510   //
511   // A read barrier for an object reference GC root is implemented as
512   // a call to the artReadBarrierForRootSlow runtime entry point,
513   // which is passed the value in location `root`:
514   //
515   //   mirror::Object* artReadBarrierForRootSlow(GcRoot<mirror::Object>* root);
516   //
517   // The `out` location contains the value returned by
518   // artReadBarrierForRootSlow.
519   void GenerateReadBarrierForRootSlow(HInstruction* instruction, Location out, Location root);
520 
ConstantAreaStart()521   int ConstantAreaStart() const {
522     return constant_area_start_;
523   }
524 
525   Address LiteralDoubleAddress(double v);
526   Address LiteralFloatAddress(float v);
527   Address LiteralInt32Address(int32_t v);
528   Address LiteralInt64Address(int64_t v);
529 
530   // Load a 32/64-bit value into a register in the most efficient manner.
531   void Load32BitValue(CpuRegister dest, int32_t value);
532   void Load64BitValue(CpuRegister dest, int64_t value);
533   void Load32BitValue(XmmRegister dest, int32_t value);
534   void Load64BitValue(XmmRegister dest, int64_t value);
535   void Load32BitValue(XmmRegister dest, float value);
536   void Load64BitValue(XmmRegister dest, double value);
537 
538   // Compare a register with a 32/64-bit value in the most efficient manner.
539   void Compare32BitValue(CpuRegister dest, int32_t value);
540   void Compare64BitValue(CpuRegister dest, int64_t value);
541 
542   // Compare int values. Supports register locations for `lhs`.
543   void GenerateIntCompare(Location lhs, Location rhs);
544   void GenerateIntCompare(CpuRegister lhs, Location rhs);
545 
546   // Compare long values. Supports only register locations for `lhs`.
547   void GenerateLongCompare(Location lhs, Location rhs);
548 
549   // Construct address for array access.
550   static Address ArrayAddress(CpuRegister obj,
551                               Location index,
552                               ScaleFactor scale,
553                               uint32_t data_offset);
554 
555   Address LiteralCaseTable(HPackedSwitch* switch_instr);
556 
557   // Store a 64 bit value into a DoubleStackSlot in the most efficient manner.
558   void Store64BitValueToStack(Location dest, int64_t value);
559 
560   // Assign a 64 bit constant to an address.
561   void MoveInt64ToAddress(const Address& addr_low,
562                           const Address& addr_high,
563                           int64_t v,
564                           HInstruction* instruction);
565 
566   // Ensure that prior stores complete to memory before subsequent loads.
567   // The locked add implementation will avoid serializing device memory, but will
568   // touch (but not change) the top of the stack.
569   // The 'non_temporal' parameter should be used to ensure ordering of non-temporal stores.
570   void MemoryFence(bool force_mfence = false) {
571     if (!force_mfence) {
572       assembler_.lock()->addl(Address(CpuRegister(RSP), 0), Immediate(0));
573     } else {
574       assembler_.mfence();
575     }
576   }
577 
578   void GenerateNop() OVERRIDE;
579   void GenerateImplicitNullCheck(HNullCheck* instruction) OVERRIDE;
580   void GenerateExplicitNullCheck(HNullCheck* instruction) OVERRIDE;
581 
582   // When we don't know the proper offset for the value, we use kDummy32BitOffset.
583   // We will fix this up in the linker later to have the right value.
584   static constexpr int32_t kDummy32BitOffset = 256;
585 
586  private:
587   template <LinkerPatch (*Factory)(size_t, const DexFile*, uint32_t, uint32_t)>
588   static void EmitPcRelativeLinkerPatches(const ArenaDeque<PatchInfo<Label>>& infos,
589                                           ArenaVector<LinkerPatch>* linker_patches);
590 
591   // Labels for each block that will be compiled.
592   Label* block_labels_;  // Indexed by block id.
593   Label frame_entry_label_;
594   LocationsBuilderX86_64 location_builder_;
595   InstructionCodeGeneratorX86_64 instruction_visitor_;
596   ParallelMoveResolverX86_64 move_resolver_;
597   X86_64Assembler assembler_;
598   const X86_64InstructionSetFeatures& isa_features_;
599 
600   // Offset to the start of the constant area in the assembled code.
601   // Used for fixups to the constant area.
602   int constant_area_start_;
603 
604   // PC-relative DexCache access info.
605   ArenaDeque<PatchInfo<Label>> pc_relative_dex_cache_patches_;
606   // String patch locations; type depends on configuration (app .bss or boot image PIC).
607   ArenaDeque<PatchInfo<Label>> string_patches_;
608   // Type patch locations for boot image (always PIC).
609   ArenaDeque<PatchInfo<Label>> boot_image_type_patches_;
610   // Type patch locations for kBssEntry.
611   ArenaDeque<PatchInfo<Label>> type_bss_entry_patches_;
612 
613   // Fixups for jump tables need to be handled specially.
614   ArenaVector<JumpTableRIPFixup*> fixups_to_jump_tables_;
615 
616   // Patches for string literals in JIT compiled code.
617   ArenaDeque<PatchInfo<Label>> jit_string_patches_;
618 
619   // Patches for class literals in JIT compiled code.
620   ArenaDeque<PatchInfo<Label>> jit_class_patches_;
621 
622   DISALLOW_COPY_AND_ASSIGN(CodeGeneratorX86_64);
623 };
624 
625 }  // namespace x86_64
626 }  // namespace art
627 
628 #endif  // ART_COMPILER_OPTIMIZING_CODE_GENERATOR_X86_64_H_
629