code_generator_x86_64.h revision e460d1df1f789c7c8bb97024a8efbd713ac175e9 
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
33static constexpr Register kParameterCoreRegisters[] = { RSI, RDX, RCX, R8, R9 };
34static constexpr FloatRegister kParameterFloatRegisters[] =
35    { XMM0, XMM1, XMM2, XMM3, XMM4, XMM5, XMM6, XMM7 };
36
37static constexpr size_t kParameterCoreRegistersLength = arraysize(kParameterCoreRegisters);
38static constexpr size_t kParameterFloatRegistersLength = arraysize(kParameterFloatRegisters);
39
40static constexpr Register kRuntimeParameterCoreRegisters[] = { RDI, RSI, RDX, RCX };
41static constexpr size_t kRuntimeParameterCoreRegistersLength =
42    arraysize(kRuntimeParameterCoreRegisters);
43static constexpr FloatRegister kRuntimeParameterFpuRegisters[] = { XMM0, XMM1 };
44static constexpr size_t kRuntimeParameterFpuRegistersLength =
45    arraysize(kRuntimeParameterFpuRegisters);
46
47class InvokeRuntimeCallingConvention : public CallingConvention<Register, FloatRegister> {
48 public:
49  InvokeRuntimeCallingConvention()
50      : CallingConvention(kRuntimeParameterCoreRegisters,
51                          kRuntimeParameterCoreRegistersLength,
52                          kRuntimeParameterFpuRegisters,
53                          kRuntimeParameterFpuRegistersLength,
54                          kX86_64PointerSize) {}
55
56 private:
57  DISALLOW_COPY_AND_ASSIGN(InvokeRuntimeCallingConvention);
58};
59
60class InvokeDexCallingConvention : public CallingConvention<Register, FloatRegister> {
61 public:
62  InvokeDexCallingConvention() : CallingConvention(
63      kParameterCoreRegisters,
64      kParameterCoreRegistersLength,
65      kParameterFloatRegisters,
66      kParameterFloatRegistersLength,
67      kX86_64PointerSize) {}
68
69 private:
70  DISALLOW_COPY_AND_ASSIGN(InvokeDexCallingConvention);
71};
72
73class FieldAccessCallingConventionX86_64 : public FieldAccessCallingConvention {
74 public:
75  FieldAccessCallingConventionX86_64() {}
76
77  Location GetObjectLocation() const OVERRIDE {
78    return Location::RegisterLocation(RSI);
79  }
80  Location GetFieldIndexLocation() const OVERRIDE {
81    return Location::RegisterLocation(RDI);
82  }
83  Location GetReturnLocation(Primitive::Type type ATTRIBUTE_UNUSED) const OVERRIDE {
84    return Location::RegisterLocation(RAX);
85  }
86  Location GetSetValueLocation(Primitive::Type type, bool is_instance) const OVERRIDE {
87    return Primitive::Is64BitType(type)
88        ? Location::RegisterLocation(RDX)
89        : (is_instance
90            ? Location::RegisterLocation(RDX)
91            : Location::RegisterLocation(RSI));
92  }
93  Location GetFpuLocation(Primitive::Type type ATTRIBUTE_UNUSED) const OVERRIDE {
94    return Location::FpuRegisterLocation(XMM0);
95  }
96
97 private:
98  DISALLOW_COPY_AND_ASSIGN(FieldAccessCallingConventionX86_64);
99};
100
101
102class InvokeDexCallingConventionVisitorX86_64 : public InvokeDexCallingConventionVisitor {
103 public:
104  InvokeDexCallingConventionVisitorX86_64() {}
105  virtual ~InvokeDexCallingConventionVisitorX86_64() {}
106
107  Location GetNextLocation(Primitive::Type type) OVERRIDE;
108  Location GetReturnLocation(Primitive::Type type) const OVERRIDE;
109  Location GetMethodLocation() const OVERRIDE;
110
111 private:
112  InvokeDexCallingConvention calling_convention;
113
114  DISALLOW_COPY_AND_ASSIGN(InvokeDexCallingConventionVisitorX86_64);
115};
116
117class CodeGeneratorX86_64;
118
119class ParallelMoveResolverX86_64 : public ParallelMoveResolverWithSwap {
120 public:
121  ParallelMoveResolverX86_64(ArenaAllocator* allocator, CodeGeneratorX86_64* codegen)
122      : ParallelMoveResolverWithSwap(allocator), codegen_(codegen) {}
123
124  void EmitMove(size_t index) OVERRIDE;
125  void EmitSwap(size_t index) OVERRIDE;
126  void SpillScratch(int reg) OVERRIDE;
127  void RestoreScratch(int reg) OVERRIDE;
128
129  X86_64Assembler* GetAssembler() const;
130
131 private:
132  void Exchange32(CpuRegister reg, int mem);
133  void Exchange32(XmmRegister reg, int mem);
134  void Exchange32(int mem1, int mem2);
135  void Exchange64(CpuRegister reg, int mem);
136  void Exchange64(XmmRegister reg, int mem);
137  void Exchange64(int mem1, int mem2);
138
139  CodeGeneratorX86_64* const codegen_;
140
141  DISALLOW_COPY_AND_ASSIGN(ParallelMoveResolverX86_64);
142};
143
144class LocationsBuilderX86_64 : public HGraphVisitor {
145 public:
146  LocationsBuilderX86_64(HGraph* graph, CodeGeneratorX86_64* codegen)
147      : HGraphVisitor(graph), codegen_(codegen) {}
148
149#define DECLARE_VISIT_INSTRUCTION(name, super)     \
150  void Visit##name(H##name* instr) OVERRIDE;
151
152  FOR_EACH_CONCRETE_INSTRUCTION_COMMON(DECLARE_VISIT_INSTRUCTION)
153  FOR_EACH_CONCRETE_INSTRUCTION_X86_64(DECLARE_VISIT_INSTRUCTION)
154
155#undef DECLARE_VISIT_INSTRUCTION
156
157  void VisitInstruction(HInstruction* instruction) OVERRIDE {
158    LOG(FATAL) << "Unreachable instruction " << instruction->DebugName()
159               << " (id " << instruction->GetId() << ")";
160  }
161
162 private:
163  void HandleInvoke(HInvoke* invoke);
164  void HandleBitwiseOperation(HBinaryOperation* operation);
165  void HandleShift(HBinaryOperation* operation);
166  void HandleFieldSet(HInstruction* instruction, const FieldInfo& field_info);
167  void HandleFieldGet(HInstruction* instruction);
168
169  CodeGeneratorX86_64* const codegen_;
170  InvokeDexCallingConventionVisitorX86_64 parameter_visitor_;
171
172  DISALLOW_COPY_AND_ASSIGN(LocationsBuilderX86_64);
173};
174
175class InstructionCodeGeneratorX86_64 : public HGraphVisitor {
176 public:
177  InstructionCodeGeneratorX86_64(HGraph* graph, CodeGeneratorX86_64* codegen);
178
179#define DECLARE_VISIT_INSTRUCTION(name, super)     \
180  void Visit##name(H##name* instr) OVERRIDE;
181
182  FOR_EACH_CONCRETE_INSTRUCTION_COMMON(DECLARE_VISIT_INSTRUCTION)
183  FOR_EACH_CONCRETE_INSTRUCTION_X86_64(DECLARE_VISIT_INSTRUCTION)
184
185#undef DECLARE_VISIT_INSTRUCTION
186
187  void VisitInstruction(HInstruction* instruction) OVERRIDE {
188    LOG(FATAL) << "Unreachable instruction " << instruction->DebugName()
189               << " (id " << instruction->GetId() << ")";
190  }
191
192  X86_64Assembler* GetAssembler() const { return assembler_; }
193
194 private:
195  // Generate code for the given suspend check. If not null, `successor`
196  // is the block to branch to if the suspend check is not needed, and after
197  // the suspend call.
198  void GenerateSuspendCheck(HSuspendCheck* instruction, HBasicBlock* successor);
199  void GenerateClassInitializationCheck(SlowPathCode* slow_path, CpuRegister class_reg);
200  void HandleBitwiseOperation(HBinaryOperation* operation);
201  void GenerateRemFP(HRem* rem);
202  void DivRemOneOrMinusOne(HBinaryOperation* instruction);
203  void DivByPowerOfTwo(HDiv* instruction);
204  void GenerateDivRemWithAnyConstant(HBinaryOperation* instruction);
205  void GenerateDivRemIntegral(HBinaryOperation* instruction);
206  void HandleShift(HBinaryOperation* operation);
207  void GenerateMemoryBarrier(MemBarrierKind kind);
208  void HandleFieldSet(HInstruction* instruction,
209                      const FieldInfo& field_info,
210                      bool value_can_be_null);
211  void HandleFieldGet(HInstruction* instruction, const FieldInfo& field_info);
212  void GenerateImplicitNullCheck(HNullCheck* instruction);
213  void GenerateExplicitNullCheck(HNullCheck* instruction);
214  void PushOntoFPStack(Location source, uint32_t temp_offset,
215                       uint32_t stack_adjustment, bool is_float);
216  void GenerateTestAndBranch(HInstruction* instruction,
217                             Label* true_target,
218                             Label* false_target,
219                             Label* always_true_target);
220  void GenerateCompareTestAndBranch(HIf* if_inst,
221                                    HCondition* condition,
222                                    Label* true_target,
223                                    Label* false_target,
224                                    Label* always_true_target);
225  void GenerateFPJumps(HCondition* cond, Label* true_label, Label* false_label);
226  void HandleGoto(HInstruction* got, HBasicBlock* successor);
227
228  X86_64Assembler* const assembler_;
229  CodeGeneratorX86_64* const codegen_;
230
231  DISALLOW_COPY_AND_ASSIGN(InstructionCodeGeneratorX86_64);
232};
233
234class CodeGeneratorX86_64 : public CodeGenerator {
235 public:
236  CodeGeneratorX86_64(HGraph* graph,
237                  const X86_64InstructionSetFeatures& isa_features,
238                  const CompilerOptions& compiler_options,
239                  OptimizingCompilerStats* stats = nullptr);
240  virtual ~CodeGeneratorX86_64() {}
241
242  void GenerateFrameEntry() OVERRIDE;
243  void GenerateFrameExit() OVERRIDE;
244  void Bind(HBasicBlock* block) OVERRIDE;
245  void Move(HInstruction* instruction, Location location, HInstruction* move_for) OVERRIDE;
246  void MoveConstant(Location destination, int32_t value) OVERRIDE;
247  void MoveLocation(Location dst, Location src, Primitive::Type dst_type) OVERRIDE;
248  void AddLocationAsTemp(Location location, LocationSummary* locations) OVERRIDE;
249
250  size_t SaveCoreRegister(size_t stack_index, uint32_t reg_id) OVERRIDE;
251  size_t RestoreCoreRegister(size_t stack_index, uint32_t reg_id) OVERRIDE;
252  size_t SaveFloatingPointRegister(size_t stack_index, uint32_t reg_id) OVERRIDE;
253  size_t RestoreFloatingPointRegister(size_t stack_index, uint32_t reg_id) OVERRIDE;
254
255  // Generate code to invoke a runtime entry point.
256  void InvokeRuntime(QuickEntrypointEnum entrypoint,
257                     HInstruction* instruction,
258                     uint32_t dex_pc,
259                     SlowPathCode* slow_path) OVERRIDE;
260
261  void InvokeRuntime(int32_t entry_point_offset,
262                     HInstruction* instruction,
263                     uint32_t dex_pc,
264                     SlowPathCode* slow_path);
265
266  size_t GetWordSize() const OVERRIDE {
267    return kX86_64WordSize;
268  }
269
270  size_t GetFloatingPointSpillSlotSize() const OVERRIDE {
271    return kX86_64WordSize;
272  }
273
274  HGraphVisitor* GetLocationBuilder() OVERRIDE {
275    return &location_builder_;
276  }
277
278  HGraphVisitor* GetInstructionVisitor() OVERRIDE {
279    return &instruction_visitor_;
280  }
281
282  X86_64Assembler* GetAssembler() OVERRIDE {
283    return &assembler_;
284  }
285
286  const X86_64Assembler& GetAssembler() const OVERRIDE {
287    return assembler_;
288  }
289
290  ParallelMoveResolverX86_64* GetMoveResolver() OVERRIDE {
291    return &move_resolver_;
292  }
293
294  uintptr_t GetAddressOf(HBasicBlock* block) const OVERRIDE {
295    return GetLabelOf(block)->Position();
296  }
297
298  Location GetStackLocation(HLoadLocal* load) const OVERRIDE;
299
300  void SetupBlockedRegisters(bool is_baseline) const OVERRIDE;
301  Location AllocateFreeRegister(Primitive::Type type) const OVERRIDE;
302  void DumpCoreRegister(std::ostream& stream, int reg) const OVERRIDE;
303  void DumpFloatingPointRegister(std::ostream& stream, int reg) const OVERRIDE;
304  void Finalize(CodeAllocator* allocator) OVERRIDE;
305
306  InstructionSet GetInstructionSet() const OVERRIDE {
307    return InstructionSet::kX86_64;
308  }
309
310  // Emit a write barrier.
311  void MarkGCCard(CpuRegister temp,
312                  CpuRegister card,
313                  CpuRegister object,
314                  CpuRegister value,
315                  bool value_can_be_null);
316
317  // Helper method to move a value between two locations.
318  void Move(Location destination, Location source);
319
320  Label* GetLabelOf(HBasicBlock* block) const {
321    return CommonGetLabelOf<Label>(block_labels_, block);
322  }
323
324  void Initialize() OVERRIDE {
325    block_labels_ = CommonInitializeLabels<Label>();
326  }
327
328  bool NeedsTwoRegisters(Primitive::Type type ATTRIBUTE_UNUSED) const OVERRIDE {
329    return false;
330  }
331
332  void GenerateStaticOrDirectCall(HInvokeStaticOrDirect* invoke, Location temp) OVERRIDE;
333  void GenerateVirtualCall(HInvokeVirtual* invoke, Location temp) OVERRIDE;
334
335  void MoveFromReturnRegister(Location trg, Primitive::Type type) OVERRIDE;
336
337  void EmitLinkerPatches(ArenaVector<LinkerPatch>* linker_patches) OVERRIDE;
338
339  const X86_64InstructionSetFeatures& GetInstructionSetFeatures() const {
340    return isa_features_;
341  }
342
343  int ConstantAreaStart() const {
344    return constant_area_start_;
345  }
346
347  Address LiteralDoubleAddress(double v);
348  Address LiteralFloatAddress(float v);
349  Address LiteralInt32Address(int32_t v);
350  Address LiteralInt64Address(int64_t v);
351
352  // Load a 64 bit value into a register in the most efficient manner.
353  void Load64BitValue(CpuRegister dest, int64_t value);
354
355  // Store a 64 bit value into a DoubleStackSlot in the most efficient manner.
356  void Store64BitValueToStack(Location dest, int64_t value);
357
358 private:
359  struct PcRelativeDexCacheAccessInfo {
360    PcRelativeDexCacheAccessInfo(const DexFile& dex_file, uint32_t element_off)
361        : target_dex_file(dex_file), element_offset(element_off), label() { }
362
363    const DexFile& target_dex_file;
364    uint32_t element_offset;
365    Label label;
366  };
367
368  // Labels for each block that will be compiled.
369  Label* block_labels_;  // Indexed by block id.
370  Label frame_entry_label_;
371  LocationsBuilderX86_64 location_builder_;
372  InstructionCodeGeneratorX86_64 instruction_visitor_;
373  ParallelMoveResolverX86_64 move_resolver_;
374  X86_64Assembler assembler_;
375  const X86_64InstructionSetFeatures& isa_features_;
376
377  // Offset to the start of the constant area in the assembled code.
378  // Used for fixups to the constant area.
379  int constant_area_start_;
380
381  // Method patch info. Using ArenaDeque<> which retains element addresses on push/emplace_back().
382  ArenaDeque<MethodPatchInfo<Label>> method_patches_;
383  ArenaDeque<MethodPatchInfo<Label>> relative_call_patches_;
384  // PC-relative DexCache access info.
385  ArenaDeque<PcRelativeDexCacheAccessInfo> pc_rel_dex_cache_patches_;
386
387  // When we don't know the proper offset for the value, we use kDummy32BitOffset.
388  // We will fix this up in the linker later to have the right value.
389  static constexpr int32_t kDummy32BitOffset = 256;
390
391  DISALLOW_COPY_AND_ASSIGN(CodeGeneratorX86_64);
392};
393
394}  // namespace x86_64
395}  // namespace art
396
397#endif  // ART_COMPILER_OPTIMIZING_CODE_GENERATOR_X86_64_H_
398