code_generator_x86_64.h revision 0d5a281c671444bfa75d63caf1427a8c0e6e1177 
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 "code_generator.h"
21#include "dex/compiler_enums.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
27namespace art {
28namespace x86_64 {
29
30// Use a local definition to prevent copying mistakes.
31static constexpr size_t kX86_64WordSize = kX86_64PointerSize;
32
33// Some x86_64 instructions require a register to be available as temp.
34static constexpr Register TMP = R11;
35
36static constexpr Register kParameterCoreRegisters[] = { RSI, RDX, RCX, R8, R9 };
37static constexpr FloatRegister kParameterFloatRegisters[] =
38    { XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7 };
39
40static constexpr size_t kParameterCoreRegistersLength = arraysize(kParameterCoreRegisters);
41static constexpr size_t kParameterFloatRegistersLength = arraysize(kParameterFloatRegisters);
42
43static constexpr Register kRuntimeParameterCoreRegisters[] = { RDI, RSI, RDX, RCX };
44static constexpr size_t kRuntimeParameterCoreRegistersLength =
45    arraysize(kRuntimeParameterCoreRegisters);
46static constexpr FloatRegister kRuntimeParameterFpuRegisters[] = { XMM0, XMM1 };
47static constexpr size_t kRuntimeParameterFpuRegistersLength =
48    arraysize(kRuntimeParameterFpuRegisters);
49
50class InvokeRuntimeCallingConvention : public CallingConvention<Register, FloatRegister> {
51 public:
52  InvokeRuntimeCallingConvention()
53      : CallingConvention(kRuntimeParameterCoreRegisters,
54                          kRuntimeParameterCoreRegistersLength,
55                          kRuntimeParameterFpuRegisters,
56                          kRuntimeParameterFpuRegistersLength,
57                          kX86_64PointerSize) {}
58
59 private:
60  DISALLOW_COPY_AND_ASSIGN(InvokeRuntimeCallingConvention);
61};
62
63class InvokeDexCallingConvention : public CallingConvention<Register, FloatRegister> {
64 public:
65  InvokeDexCallingConvention() : CallingConvention(
66      kParameterCoreRegisters,
67      kParameterCoreRegistersLength,
68      kParameterFloatRegisters,
69      kParameterFloatRegistersLength,
70      kX86_64PointerSize) {}
71
72 private:
73  DISALLOW_COPY_AND_ASSIGN(InvokeDexCallingConvention);
74};
75
76class FieldAccessCallingConventionX86_64 : public FieldAccessCallingConvention {
77 public:
78  FieldAccessCallingConventionX86_64() {}
79
80  Location GetObjectLocation() const OVERRIDE {
81    return Location::RegisterLocation(RSI);
82  }
83  Location GetFieldIndexLocation() const OVERRIDE {
84    return Location::RegisterLocation(RDI);
85  }
86  Location GetReturnLocation(Primitive::Type type ATTRIBUTE_UNUSED) const OVERRIDE {
87    return Location::RegisterLocation(RAX);
88  }
89  Location GetSetValueLocation(Primitive::Type type, bool is_instance) const OVERRIDE {
90    return Primitive::Is64BitType(type)
91        ? Location::RegisterLocation(RDX)
92        : (is_instance
93            ? Location::RegisterLocation(RDX)
94            : Location::RegisterLocation(RSI));
95  }
96  Location GetFpuLocation(Primitive::Type type ATTRIBUTE_UNUSED) const OVERRIDE {
97    return Location::FpuRegisterLocation(XMM0);
98  }
99
100 private:
101  DISALLOW_COPY_AND_ASSIGN(FieldAccessCallingConventionX86_64);
102};
103
104
105class InvokeDexCallingConventionVisitorX86_64 : public InvokeDexCallingConventionVisitor {
106 public:
107  InvokeDexCallingConventionVisitorX86_64() {}
108  virtual ~InvokeDexCallingConventionVisitorX86_64() {}
109
110  Location GetNextLocation(Primitive::Type type) OVERRIDE;
111  Location GetReturnLocation(Primitive::Type type) const OVERRIDE;
112  Location GetMethodLocation() const OVERRIDE;
113
114 private:
115  InvokeDexCallingConvention calling_convention;
116
117  DISALLOW_COPY_AND_ASSIGN(InvokeDexCallingConventionVisitorX86_64);
118};
119
120class CodeGeneratorX86_64;
121
122class ParallelMoveResolverX86_64 : public ParallelMoveResolverWithSwap {
123 public:
124  ParallelMoveResolverX86_64(ArenaAllocator* allocator, CodeGeneratorX86_64* codegen)
125      : ParallelMoveResolverWithSwap(allocator), codegen_(codegen) {}
126
127  void EmitMove(size_t index) OVERRIDE;
128  void EmitSwap(size_t index) OVERRIDE;
129  void SpillScratch(int reg) OVERRIDE;
130  void RestoreScratch(int reg) OVERRIDE;
131
132  X86_64Assembler* GetAssembler() const;
133
134 private:
135  void Exchange32(CpuRegister reg, int mem);
136  void Exchange32(XmmRegister reg, int mem);
137  void Exchange32(int mem1, int mem2);
138  void Exchange64(CpuRegister reg, int mem);
139  void Exchange64(XmmRegister reg, int mem);
140  void Exchange64(int mem1, int mem2);
141
142  CodeGeneratorX86_64* const codegen_;
143
144  DISALLOW_COPY_AND_ASSIGN(ParallelMoveResolverX86_64);
145};
146
147class LocationsBuilderX86_64 : public HGraphVisitor {
148 public:
149  LocationsBuilderX86_64(HGraph* graph, CodeGeneratorX86_64* codegen)
150      : HGraphVisitor(graph), codegen_(codegen) {}
151
152#define DECLARE_VISIT_INSTRUCTION(name, super)     \
153  void Visit##name(H##name* instr) OVERRIDE;
154
155  FOR_EACH_CONCRETE_INSTRUCTION_COMMON(DECLARE_VISIT_INSTRUCTION)
156  FOR_EACH_CONCRETE_INSTRUCTION_X86_64(DECLARE_VISIT_INSTRUCTION)
157
158#undef DECLARE_VISIT_INSTRUCTION
159
160  void VisitInstruction(HInstruction* instruction) OVERRIDE {
161    LOG(FATAL) << "Unreachable instruction " << instruction->DebugName()
162               << " (id " << instruction->GetId() << ")";
163  }
164
165 private:
166  void HandleInvoke(HInvoke* invoke);
167  void HandleBitwiseOperation(HBinaryOperation* operation);
168  void HandleShift(HBinaryOperation* operation);
169  void HandleFieldSet(HInstruction* instruction, const FieldInfo& field_info);
170  void HandleFieldGet(HInstruction* instruction);
171
172  CodeGeneratorX86_64* const codegen_;
173  InvokeDexCallingConventionVisitorX86_64 parameter_visitor_;
174
175  DISALLOW_COPY_AND_ASSIGN(LocationsBuilderX86_64);
176};
177
178class InstructionCodeGeneratorX86_64 : public HGraphVisitor {
179 public:
180  InstructionCodeGeneratorX86_64(HGraph* graph, CodeGeneratorX86_64* codegen);
181
182#define DECLARE_VISIT_INSTRUCTION(name, super)     \
183  void Visit##name(H##name* instr) OVERRIDE;
184
185  FOR_EACH_CONCRETE_INSTRUCTION_COMMON(DECLARE_VISIT_INSTRUCTION)
186  FOR_EACH_CONCRETE_INSTRUCTION_X86_64(DECLARE_VISIT_INSTRUCTION)
187
188#undef DECLARE_VISIT_INSTRUCTION
189
190  void VisitInstruction(HInstruction* instruction) OVERRIDE {
191    LOG(FATAL) << "Unreachable instruction " << instruction->DebugName()
192               << " (id " << instruction->GetId() << ")";
193  }
194
195  X86_64Assembler* GetAssembler() const { return assembler_; }
196
197 private:
198  // Generate code for the given suspend check. If not null, `successor`
199  // is the block to branch to if the suspend check is not needed, and after
200  // the suspend call.
201  void GenerateSuspendCheck(HSuspendCheck* instruction, HBasicBlock* successor);
202  void GenerateClassInitializationCheck(SlowPathCode* slow_path, CpuRegister class_reg);
203  void HandleBitwiseOperation(HBinaryOperation* operation);
204  void GenerateRemFP(HRem* rem);
205  void DivRemOneOrMinusOne(HBinaryOperation* instruction);
206  void DivByPowerOfTwo(HDiv* instruction);
207  void GenerateDivRemWithAnyConstant(HBinaryOperation* instruction);
208  void GenerateDivRemIntegral(HBinaryOperation* instruction);
209  void HandleShift(HBinaryOperation* operation);
210  void GenerateMemoryBarrier(MemBarrierKind kind);
211  void HandleFieldSet(HInstruction* instruction,
212                      const FieldInfo& field_info,
213                      bool value_can_be_null);
214  void HandleFieldGet(HInstruction* instruction, const FieldInfo& field_info);
215  void GenerateImplicitNullCheck(HNullCheck* instruction);
216  void GenerateExplicitNullCheck(HNullCheck* instruction);
217  void PushOntoFPStack(Location source, uint32_t temp_offset,
218                       uint32_t stack_adjustment, bool is_float);
219  void GenerateTestAndBranch(HInstruction* instruction,
220                             Label* true_target,
221                             Label* false_target,
222                             Label* always_true_target);
223  void GenerateCompareTestAndBranch(HIf* if_inst,
224                                    HCondition* condition,
225                                    Label* true_target,
226                                    Label* false_target,
227                                    Label* always_true_target);
228  void GenerateFPJumps(HCondition* cond, Label* true_label, Label* false_label);
229  void HandleGoto(HInstruction* got, HBasicBlock* successor);
230
231  X86_64Assembler* const assembler_;
232  CodeGeneratorX86_64* const codegen_;
233
234  DISALLOW_COPY_AND_ASSIGN(InstructionCodeGeneratorX86_64);
235};
236
237// Class for fixups to jump tables.
238class JumpTableRIPFixup;
239
240class CodeGeneratorX86_64 : public CodeGenerator {
241 public:
242  CodeGeneratorX86_64(HGraph* graph,
243                  const X86_64InstructionSetFeatures& isa_features,
244                  const CompilerOptions& compiler_options,
245                  OptimizingCompilerStats* stats = nullptr);
246  virtual ~CodeGeneratorX86_64() {}
247
248  void GenerateFrameEntry() OVERRIDE;
249  void GenerateFrameExit() OVERRIDE;
250  void Bind(HBasicBlock* block) OVERRIDE;
251  void Move(HInstruction* instruction, Location location, HInstruction* move_for) OVERRIDE;
252  void MoveConstant(Location destination, int32_t value) OVERRIDE;
253  void MoveLocation(Location dst, Location src, Primitive::Type dst_type) OVERRIDE;
254  void AddLocationAsTemp(Location location, LocationSummary* locations) OVERRIDE;
255
256  size_t SaveCoreRegister(size_t stack_index, uint32_t reg_id) OVERRIDE;
257  size_t RestoreCoreRegister(size_t stack_index, uint32_t reg_id) OVERRIDE;
258  size_t SaveFloatingPointRegister(size_t stack_index, uint32_t reg_id) OVERRIDE;
259  size_t RestoreFloatingPointRegister(size_t stack_index, uint32_t reg_id) OVERRIDE;
260
261  // Generate code to invoke a runtime entry point.
262  void InvokeRuntime(QuickEntrypointEnum entrypoint,
263                     HInstruction* instruction,
264                     uint32_t dex_pc,
265                     SlowPathCode* slow_path) OVERRIDE;
266
267  void InvokeRuntime(int32_t entry_point_offset,
268                     HInstruction* instruction,
269                     uint32_t dex_pc,
270                     SlowPathCode* slow_path);
271
272  size_t GetWordSize() const OVERRIDE {
273    return kX86_64WordSize;
274  }
275
276  size_t GetFloatingPointSpillSlotSize() const OVERRIDE {
277    return kX86_64WordSize;
278  }
279
280  HGraphVisitor* GetLocationBuilder() OVERRIDE {
281    return &location_builder_;
282  }
283
284  HGraphVisitor* GetInstructionVisitor() OVERRIDE {
285    return &instruction_visitor_;
286  }
287
288  X86_64Assembler* GetAssembler() OVERRIDE {
289    return &assembler_;
290  }
291
292  const X86_64Assembler& GetAssembler() const OVERRIDE {
293    return assembler_;
294  }
295
296  ParallelMoveResolverX86_64* GetMoveResolver() OVERRIDE {
297    return &move_resolver_;
298  }
299
300  uintptr_t GetAddressOf(HBasicBlock* block) const OVERRIDE {
301    return GetLabelOf(block)->Position();
302  }
303
304  Location GetStackLocation(HLoadLocal* load) const OVERRIDE;
305
306  void SetupBlockedRegisters(bool is_baseline) const OVERRIDE;
307  Location AllocateFreeRegister(Primitive::Type type) const OVERRIDE;
308  void DumpCoreRegister(std::ostream& stream, int reg) const OVERRIDE;
309  void DumpFloatingPointRegister(std::ostream& stream, int reg) const OVERRIDE;
310  void Finalize(CodeAllocator* allocator) OVERRIDE;
311
312  InstructionSet GetInstructionSet() const OVERRIDE {
313    return InstructionSet::kX86_64;
314  }
315
316  // Emit a write barrier.
317  void MarkGCCard(CpuRegister temp,
318                  CpuRegister card,
319                  CpuRegister object,
320                  CpuRegister value,
321                  bool value_can_be_null);
322
323  // Helper method to move a value between two locations.
324  void Move(Location destination, Location source);
325
326  Label* GetLabelOf(HBasicBlock* block) const {
327    return CommonGetLabelOf<Label>(block_labels_, block);
328  }
329
330  void Initialize() OVERRIDE {
331    block_labels_ = CommonInitializeLabels<Label>();
332  }
333
334  bool NeedsTwoRegisters(Primitive::Type type ATTRIBUTE_UNUSED) const OVERRIDE {
335    return false;
336  }
337
338  // Check if the desired_dispatch_info is supported. If it is, return it,
339  // otherwise return a fall-back info that should be used instead.
340  HInvokeStaticOrDirect::DispatchInfo GetSupportedInvokeStaticOrDirectDispatch(
341      const HInvokeStaticOrDirect::DispatchInfo& desired_dispatch_info,
342      MethodReference target_method) OVERRIDE;
343
344  void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) OVERRIDE;
345  void GenerateVirtualCall(HInvokeVirtual* invoke, Location temp) OVERRIDE;
346
347  void MoveFromReturnRegister(Location trg, Primitive::Type type) OVERRIDE;
348
349  void EmitLinkerPatches(ArenaVector<LinkerPatch>* linker_patches) OVERRIDE;
350
351  const X86_64InstructionSetFeatures& GetInstructionSetFeatures() const {
352    return isa_features_;
353  }
354
355  // Generate a read barrier for a heap reference within `instruction`.
356  //
357  // A read barrier for an object reference read from the heap is
358  // implemented as a call to the artReadBarrierSlow runtime entry
359  // point, which is passed the values in locations `ref`, `obj`, and
360  // `offset`:
361  //
362  //   mirror::Object* artReadBarrierSlow(mirror::Object* ref,
363  //                                      mirror::Object* obj,
364  //                                      uint32_t offset);
365  //
366  // The `out` location contains the value returned by
367  // artReadBarrierSlow.
368  //
369  // When `index` provided (i.e., when it is different from
370  // Location::NoLocation()), the offset value passed to
371  // artReadBarrierSlow is adjusted to take `index` into account.
372  void GenerateReadBarrier(HInstruction* instruction,
373                           Location out,
374                           Location ref,
375                           Location obj,
376                           uint32_t offset,
377                           Location index = Location::NoLocation());
378
379  // If read barriers are enabled, generate a read barrier for a heap reference.
380  // If heap poisoning is enabled, also unpoison the reference in `out`.
381  void MaybeGenerateReadBarrier(HInstruction* instruction,
382                                Location out,
383                                Location ref,
384                                Location obj,
385                                uint32_t offset,
386                                Location index = Location::NoLocation());
387
388  // Generate a read barrier for a GC root within `instruction`.
389  //
390  // A read barrier for an object reference GC root is implemented as
391  // a call to the artReadBarrierForRootSlow runtime entry point,
392  // which is passed the value in location `root`:
393  //
394  //   mirror::Object* artReadBarrierForRootSlow(GcRoot<mirror::Object>* root);
395  //
396  // The `out` location contains the value returned by
397  // artReadBarrierForRootSlow.
398  void GenerateReadBarrierForRoot(HInstruction* instruction, Location out, Location root);
399
400  int ConstantAreaStart() const {
401    return constant_area_start_;
402  }
403
404  Address LiteralDoubleAddress(double v);
405  Address LiteralFloatAddress(float v);
406  Address LiteralInt32Address(int32_t v);
407  Address LiteralInt64Address(int64_t v);
408
409  // Load a 64 bit value into a register in the most efficient manner.
410  void Load64BitValue(CpuRegister dest, int64_t value);
411  Address LiteralCaseTable(HPackedSwitch* switch_instr);
412
413  // Store a 64 bit value into a DoubleStackSlot in the most efficient manner.
414  void Store64BitValueToStack(Location dest, int64_t value);
415
416  // Assign a 64 bit constant to an address.
417  void MoveInt64ToAddress(const Address& addr_low,
418                          const Address& addr_high,
419                          int64_t v,
420                          HInstruction* instruction);
421
422 private:
423  struct PcRelativeDexCacheAccessInfo {
424    PcRelativeDexCacheAccessInfo(const DexFile& dex_file, uint32_t element_off)
425        : target_dex_file(dex_file), element_offset(element_off), label() { }
426
427    const DexFile& target_dex_file;
428    uint32_t element_offset;
429    Label label;
430  };
431
432  // Labels for each block that will be compiled.
433  Label* block_labels_;  // Indexed by block id.
434  Label frame_entry_label_;
435  LocationsBuilderX86_64 location_builder_;
436  InstructionCodeGeneratorX86_64 instruction_visitor_;
437  ParallelMoveResolverX86_64 move_resolver_;
438  X86_64Assembler assembler_;
439  const X86_64InstructionSetFeatures& isa_features_;
440
441  // Offset to the start of the constant area in the assembled code.
442  // Used for fixups to the constant area.
443  int constant_area_start_;
444
445  // Method patch info. Using ArenaDeque<> which retains element addresses on push/emplace_back().
446  ArenaDeque<MethodPatchInfo<Label>> method_patches_;
447  ArenaDeque<MethodPatchInfo<Label>> relative_call_patches_;
448  // PC-relative DexCache access info.
449  ArenaDeque<PcRelativeDexCacheAccessInfo> pc_relative_dex_cache_patches_;
450
451  // When we don't know the proper offset for the value, we use kDummy32BitOffset.
452  // We will fix this up in the linker later to have the right value.
453  static constexpr int32_t kDummy32BitOffset = 256;
454
455  // Fixups for jump tables need to be handled specially.
456  ArenaVector<JumpTableRIPFixup*> fixups_to_jump_tables_;
457
458  DISALLOW_COPY_AND_ASSIGN(CodeGeneratorX86_64);
459};
460
461}  // namespace x86_64
462}  // namespace art
463
464#endif  // ART_COMPILER_OPTIMIZING_CODE_GENERATOR_X86_64_H_
465