code_generator.cc revision 015c7e63604c038e866d7af3850c557403cddc8b 
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#include "code_generator.h"
18
19#include "code_generator_arm.h"
20#include "code_generator_arm64.h"
21#include "code_generator_x86.h"
22#include "code_generator_x86_64.h"
23#include "compiled_method.h"
24#include "dex/verified_method.h"
25#include "driver/dex_compilation_unit.h"
26#include "gc_map_builder.h"
27#include "leb128.h"
28#include "mapping_table.h"
29#include "mirror/array-inl.h"
30#include "mirror/object_array-inl.h"
31#include "mirror/object_reference.h"
32#include "ssa_liveness_analysis.h"
33#include "utils/assembler.h"
34#include "verifier/dex_gc_map.h"
35#include "vmap_table.h"
36
37namespace art {
38
39// Return whether a location is consistent with a type.
40static bool CheckType(Primitive::Type type, Location location) {
41  if (location.IsFpuRegister()
42      || (location.IsUnallocated() && (location.GetPolicy() == Location::kRequiresFpuRegister))) {
43    return (type == Primitive::kPrimFloat) || (type == Primitive::kPrimDouble);
44  } else if (location.IsRegister() ||
45             (location.IsUnallocated() && (location.GetPolicy() == Location::kRequiresRegister))) {
46    return Primitive::IsIntegralType(type) || (type == Primitive::kPrimNot);
47  } else if (location.IsRegisterPair()) {
48    return type == Primitive::kPrimLong;
49  } else if (location.IsFpuRegisterPair()) {
50    return type == Primitive::kPrimDouble;
51  } else if (location.IsStackSlot()) {
52    return (Primitive::IsIntegralType(type) && type != Primitive::kPrimLong)
53           || (type == Primitive::kPrimFloat)
54           || (type == Primitive::kPrimNot);
55  } else if (location.IsDoubleStackSlot()) {
56    return (type == Primitive::kPrimLong) || (type == Primitive::kPrimDouble);
57  } else if (location.IsConstant()) {
58    if (location.GetConstant()->IsIntConstant()) {
59      return Primitive::IsIntegralType(type) && (type != Primitive::kPrimLong);
60    } else if (location.GetConstant()->IsNullConstant()) {
61      return type == Primitive::kPrimNot;
62    } else if (location.GetConstant()->IsLongConstant()) {
63      return type == Primitive::kPrimLong;
64    } else if (location.GetConstant()->IsFloatConstant()) {
65      return type == Primitive::kPrimFloat;
66    } else {
67      return location.GetConstant()->IsDoubleConstant()
68          && (type == Primitive::kPrimDouble);
69    }
70  } else {
71    return location.IsInvalid() || (location.GetPolicy() == Location::kAny);
72  }
73}
74
75// Check that a location summary is consistent with an instruction.
76static bool CheckTypeConsistency(HInstruction* instruction) {
77  LocationSummary* locations = instruction->GetLocations();
78  if (locations == nullptr) {
79    return true;
80  }
81
82  if (locations->Out().IsUnallocated()
83      && (locations->Out().GetPolicy() == Location::kSameAsFirstInput)) {
84    DCHECK(CheckType(instruction->GetType(), locations->InAt(0)))
85        << instruction->GetType()
86        << " " << locations->InAt(0);
87  } else {
88    DCHECK(CheckType(instruction->GetType(), locations->Out()))
89        << instruction->GetType()
90        << " " << locations->Out();
91  }
92
93  for (size_t i = 0, e = instruction->InputCount(); i < e; ++i) {
94    DCHECK(CheckType(instruction->InputAt(i)->GetType(), locations->InAt(i)))
95      << instruction->InputAt(i)->GetType()
96      << " " << locations->InAt(i);
97  }
98
99  HEnvironment* environment = instruction->GetEnvironment();
100  for (size_t i = 0; i < instruction->EnvironmentSize(); ++i) {
101    if (environment->GetInstructionAt(i) != nullptr) {
102      Primitive::Type type = environment->GetInstructionAt(i)->GetType();
103      DCHECK(CheckType(type, environment->GetLocationAt(i)))
104        << type << " " << environment->GetLocationAt(i);
105    } else {
106      DCHECK(environment->GetLocationAt(i).IsInvalid())
107        << environment->GetLocationAt(i);
108    }
109  }
110  return true;
111}
112
113size_t CodeGenerator::GetCacheOffset(uint32_t index) {
114  return mirror::ObjectArray<mirror::Object>::OffsetOfElement(index).SizeValue();
115}
116
117size_t CodeGenerator::GetCachePointerOffset(uint32_t index) {
118  auto pointer_size = InstructionSetPointerSize(GetInstructionSet());
119  return mirror::Array::DataOffset(pointer_size).Uint32Value() + pointer_size * index;
120}
121
122void CodeGenerator::CompileBaseline(CodeAllocator* allocator, bool is_leaf) {
123  Initialize();
124  if (!is_leaf) {
125    MarkNotLeaf();
126  }
127  const bool is_64_bit = Is64BitInstructionSet(GetInstructionSet());
128  InitializeCodeGeneration(GetGraph()->GetNumberOfLocalVRegs()
129                             + GetGraph()->GetTemporariesVRegSlots()
130                             + 1 /* filler */,
131                           0, /* the baseline compiler does not have live registers at slow path */
132                           0, /* the baseline compiler does not have live registers at slow path */
133                           GetGraph()->GetMaximumNumberOfOutVRegs()
134                             + (is_64_bit ? 2 : 1) /* current method */,
135                           GetGraph()->GetBlocks());
136  CompileInternal(allocator, /* is_baseline */ true);
137}
138
139bool CodeGenerator::GoesToNextBlock(HBasicBlock* current, HBasicBlock* next) const {
140  DCHECK_EQ(block_order_->Get(current_block_index_), current);
141  return GetNextBlockToEmit() == FirstNonEmptyBlock(next);
142}
143
144HBasicBlock* CodeGenerator::GetNextBlockToEmit() const {
145  for (size_t i = current_block_index_ + 1; i < block_order_->Size(); ++i) {
146    HBasicBlock* block = block_order_->Get(i);
147    if (!block->IsSingleGoto()) {
148      return block;
149    }
150  }
151  return nullptr;
152}
153
154HBasicBlock* CodeGenerator::FirstNonEmptyBlock(HBasicBlock* block) const {
155  while (block->IsSingleGoto()) {
156    block = block->GetSuccessors().Get(0);
157  }
158  return block;
159}
160
161void CodeGenerator::CompileInternal(CodeAllocator* allocator, bool is_baseline) {
162  is_baseline_ = is_baseline;
163  HGraphVisitor* instruction_visitor = GetInstructionVisitor();
164  DCHECK_EQ(current_block_index_, 0u);
165  GenerateFrameEntry();
166  DCHECK_EQ(GetAssembler()->cfi().GetCurrentCFAOffset(), static_cast<int>(frame_size_));
167  for (size_t e = block_order_->Size(); current_block_index_ < e; ++current_block_index_) {
168    HBasicBlock* block = block_order_->Get(current_block_index_);
169    // Don't generate code for an empty block. Its predecessors will branch to its successor
170    // directly. Also, the label of that block will not be emitted, so this helps catch
171    // errors where we reference that label.
172    if (block->IsSingleGoto()) continue;
173    Bind(block);
174    for (HInstructionIterator it(block->GetInstructions()); !it.Done(); it.Advance()) {
175      HInstruction* current = it.Current();
176      if (is_baseline) {
177        InitLocationsBaseline(current);
178      }
179      DCHECK(CheckTypeConsistency(current));
180      current->Accept(instruction_visitor);
181    }
182  }
183
184  // Generate the slow paths.
185  for (size_t i = 0, e = slow_paths_.Size(); i < e; ++i) {
186    slow_paths_.Get(i)->EmitNativeCode(this);
187  }
188
189  // Finalize instructions in assember;
190  Finalize(allocator);
191}
192
193void CodeGenerator::CompileOptimized(CodeAllocator* allocator) {
194  // The register allocator already called `InitializeCodeGeneration`,
195  // where the frame size has been computed.
196  DCHECK(block_order_ != nullptr);
197  Initialize();
198  CompileInternal(allocator, /* is_baseline */ false);
199}
200
201void CodeGenerator::Finalize(CodeAllocator* allocator) {
202  size_t code_size = GetAssembler()->CodeSize();
203  uint8_t* buffer = allocator->Allocate(code_size);
204
205  MemoryRegion code(buffer, code_size);
206  GetAssembler()->FinalizeInstructions(code);
207}
208
209size_t CodeGenerator::FindFreeEntry(bool* array, size_t length) {
210  for (size_t i = 0; i < length; ++i) {
211    if (!array[i]) {
212      array[i] = true;
213      return i;
214    }
215  }
216  LOG(FATAL) << "Could not find a register in baseline register allocator";
217  UNREACHABLE();
218}
219
220size_t CodeGenerator::FindTwoFreeConsecutiveAlignedEntries(bool* array, size_t length) {
221  for (size_t i = 0; i < length - 1; i += 2) {
222    if (!array[i] && !array[i + 1]) {
223      array[i] = true;
224      array[i + 1] = true;
225      return i;
226    }
227  }
228  LOG(FATAL) << "Could not find a register in baseline register allocator";
229  UNREACHABLE();
230}
231
232void CodeGenerator::InitializeCodeGeneration(size_t number_of_spill_slots,
233                                             size_t maximum_number_of_live_core_registers,
234                                             size_t maximum_number_of_live_fp_registers,
235                                             size_t number_of_out_slots,
236                                             const GrowableArray<HBasicBlock*>& block_order) {
237  block_order_ = &block_order;
238  DCHECK(block_order_->Get(0) == GetGraph()->GetEntryBlock());
239  ComputeSpillMask();
240  first_register_slot_in_slow_path_ = (number_of_out_slots + number_of_spill_slots) * kVRegSize;
241
242  if (number_of_spill_slots == 0
243      && !HasAllocatedCalleeSaveRegisters()
244      && IsLeafMethod()
245      && !RequiresCurrentMethod()) {
246    DCHECK_EQ(maximum_number_of_live_core_registers, 0u);
247    DCHECK_EQ(maximum_number_of_live_fp_registers, 0u);
248    SetFrameSize(CallPushesPC() ? GetWordSize() : 0);
249  } else {
250    SetFrameSize(RoundUp(
251        number_of_spill_slots * kVRegSize
252        + number_of_out_slots * kVRegSize
253        + maximum_number_of_live_core_registers * GetWordSize()
254        + maximum_number_of_live_fp_registers * GetFloatingPointSpillSlotSize()
255        + FrameEntrySpillSize(),
256        kStackAlignment));
257  }
258}
259
260Location CodeGenerator::GetTemporaryLocation(HTemporary* temp) const {
261  uint16_t number_of_locals = GetGraph()->GetNumberOfLocalVRegs();
262  // The type of the previous instruction tells us if we need a single or double stack slot.
263  Primitive::Type type = temp->GetType();
264  int32_t temp_size = (type == Primitive::kPrimLong) || (type == Primitive::kPrimDouble) ? 2 : 1;
265  // Use the temporary region (right below the dex registers).
266  int32_t slot = GetFrameSize() - FrameEntrySpillSize()
267                                - kVRegSize  // filler
268                                - (number_of_locals * kVRegSize)
269                                - ((temp_size + temp->GetIndex()) * kVRegSize);
270  return temp_size == 2 ? Location::DoubleStackSlot(slot) : Location::StackSlot(slot);
271}
272
273int32_t CodeGenerator::GetStackSlot(HLocal* local) const {
274  uint16_t reg_number = local->GetRegNumber();
275  uint16_t number_of_locals = GetGraph()->GetNumberOfLocalVRegs();
276  if (reg_number >= number_of_locals) {
277    // Local is a parameter of the method. It is stored in the caller's frame.
278    // TODO: Share this logic with StackVisitor::GetVRegOffsetFromQuickCode.
279    return GetFrameSize() + InstructionSetPointerSize(GetInstructionSet())  // ART method
280                          + (reg_number - number_of_locals) * kVRegSize;
281  } else {
282    // Local is a temporary in this method. It is stored in this method's frame.
283    return GetFrameSize() - FrameEntrySpillSize()
284                          - kVRegSize  // filler.
285                          - (number_of_locals * kVRegSize)
286                          + (reg_number * kVRegSize);
287  }
288}
289
290void CodeGenerator::CreateCommonInvokeLocationSummary(
291    HInvoke* invoke, InvokeDexCallingConventionVisitor* visitor) {
292  ArenaAllocator* allocator = invoke->GetBlock()->GetGraph()->GetArena();
293  LocationSummary* locations = new (allocator) LocationSummary(invoke, LocationSummary::kCall);
294
295  for (size_t i = 0; i < invoke->GetNumberOfArguments(); i++) {
296    HInstruction* input = invoke->InputAt(i);
297    locations->SetInAt(i, visitor->GetNextLocation(input->GetType()));
298  }
299
300  locations->SetOut(visitor->GetReturnLocation(invoke->GetType()));
301
302  if (invoke->IsInvokeStaticOrDirect()) {
303    HInvokeStaticOrDirect* call = invoke->AsInvokeStaticOrDirect();
304    if (call->IsStringInit()) {
305      locations->AddTemp(visitor->GetMethodLocation());
306    } else if (call->IsRecursive()) {
307      locations->SetInAt(call->GetCurrentMethodInputIndex(), visitor->GetMethodLocation());
308    } else {
309      locations->AddTemp(visitor->GetMethodLocation());
310      locations->SetInAt(call->GetCurrentMethodInputIndex(), Location::RequiresRegister());
311    }
312  } else {
313    locations->AddTemp(visitor->GetMethodLocation());
314  }
315}
316
317void CodeGenerator::BlockIfInRegister(Location location, bool is_out) const {
318  // The DCHECKS below check that a register is not specified twice in
319  // the summary. The out location can overlap with an input, so we need
320  // to special case it.
321  if (location.IsRegister()) {
322    DCHECK(is_out || !blocked_core_registers_[location.reg()]);
323    blocked_core_registers_[location.reg()] = true;
324  } else if (location.IsFpuRegister()) {
325    DCHECK(is_out || !blocked_fpu_registers_[location.reg()]);
326    blocked_fpu_registers_[location.reg()] = true;
327  } else if (location.IsFpuRegisterPair()) {
328    DCHECK(is_out || !blocked_fpu_registers_[location.AsFpuRegisterPairLow<int>()]);
329    blocked_fpu_registers_[location.AsFpuRegisterPairLow<int>()] = true;
330    DCHECK(is_out || !blocked_fpu_registers_[location.AsFpuRegisterPairHigh<int>()]);
331    blocked_fpu_registers_[location.AsFpuRegisterPairHigh<int>()] = true;
332  } else if (location.IsRegisterPair()) {
333    DCHECK(is_out || !blocked_core_registers_[location.AsRegisterPairLow<int>()]);
334    blocked_core_registers_[location.AsRegisterPairLow<int>()] = true;
335    DCHECK(is_out || !blocked_core_registers_[location.AsRegisterPairHigh<int>()]);
336    blocked_core_registers_[location.AsRegisterPairHigh<int>()] = true;
337  }
338}
339
340void CodeGenerator::AllocateRegistersLocally(HInstruction* instruction) const {
341  LocationSummary* locations = instruction->GetLocations();
342  if (locations == nullptr) return;
343
344  for (size_t i = 0, e = GetNumberOfCoreRegisters(); i < e; ++i) {
345    blocked_core_registers_[i] = false;
346  }
347
348  for (size_t i = 0, e = GetNumberOfFloatingPointRegisters(); i < e; ++i) {
349    blocked_fpu_registers_[i] = false;
350  }
351
352  for (size_t i = 0, e = number_of_register_pairs_; i < e; ++i) {
353    blocked_register_pairs_[i] = false;
354  }
355
356  // Mark all fixed input, temp and output registers as used.
357  for (size_t i = 0, e = locations->GetInputCount(); i < e; ++i) {
358    BlockIfInRegister(locations->InAt(i));
359  }
360
361  for (size_t i = 0, e = locations->GetTempCount(); i < e; ++i) {
362    Location loc = locations->GetTemp(i);
363    BlockIfInRegister(loc);
364  }
365  Location result_location = locations->Out();
366  if (locations->OutputCanOverlapWithInputs()) {
367    BlockIfInRegister(result_location, /* is_out */ true);
368  }
369
370  SetupBlockedRegisters(/* is_baseline */ true);
371
372  // Allocate all unallocated input locations.
373  for (size_t i = 0, e = locations->GetInputCount(); i < e; ++i) {
374    Location loc = locations->InAt(i);
375    HInstruction* input = instruction->InputAt(i);
376    if (loc.IsUnallocated()) {
377      if ((loc.GetPolicy() == Location::kRequiresRegister)
378          || (loc.GetPolicy() == Location::kRequiresFpuRegister)) {
379        loc = AllocateFreeRegister(input->GetType());
380      } else {
381        DCHECK_EQ(loc.GetPolicy(), Location::kAny);
382        HLoadLocal* load = input->AsLoadLocal();
383        if (load != nullptr) {
384          loc = GetStackLocation(load);
385        } else {
386          loc = AllocateFreeRegister(input->GetType());
387        }
388      }
389      locations->SetInAt(i, loc);
390    }
391  }
392
393  // Allocate all unallocated temp locations.
394  for (size_t i = 0, e = locations->GetTempCount(); i < e; ++i) {
395    Location loc = locations->GetTemp(i);
396    if (loc.IsUnallocated()) {
397      switch (loc.GetPolicy()) {
398        case Location::kRequiresRegister:
399          // Allocate a core register (large enough to fit a 32-bit integer).
400          loc = AllocateFreeRegister(Primitive::kPrimInt);
401          break;
402
403        case Location::kRequiresFpuRegister:
404          // Allocate a core register (large enough to fit a 64-bit double).
405          loc = AllocateFreeRegister(Primitive::kPrimDouble);
406          break;
407
408        default:
409          LOG(FATAL) << "Unexpected policy for temporary location "
410                     << loc.GetPolicy();
411      }
412      locations->SetTempAt(i, loc);
413    }
414  }
415  if (result_location.IsUnallocated()) {
416    switch (result_location.GetPolicy()) {
417      case Location::kAny:
418      case Location::kRequiresRegister:
419      case Location::kRequiresFpuRegister:
420        result_location = AllocateFreeRegister(instruction->GetType());
421        break;
422      case Location::kSameAsFirstInput:
423        result_location = locations->InAt(0);
424        break;
425    }
426    locations->UpdateOut(result_location);
427  }
428}
429
430void CodeGenerator::InitLocationsBaseline(HInstruction* instruction) {
431  AllocateLocations(instruction);
432  if (instruction->GetLocations() == nullptr) {
433    if (instruction->IsTemporary()) {
434      HInstruction* previous = instruction->GetPrevious();
435      Location temp_location = GetTemporaryLocation(instruction->AsTemporary());
436      Move(previous, temp_location, instruction);
437    }
438    return;
439  }
440  AllocateRegistersLocally(instruction);
441  for (size_t i = 0, e = instruction->InputCount(); i < e; ++i) {
442    Location location = instruction->GetLocations()->InAt(i);
443    HInstruction* input = instruction->InputAt(i);
444    if (location.IsValid()) {
445      // Move the input to the desired location.
446      if (input->GetNext()->IsTemporary()) {
447        // If the input was stored in a temporary, use that temporary to
448        // perform the move.
449        Move(input->GetNext(), location, instruction);
450      } else {
451        Move(input, location, instruction);
452      }
453    }
454  }
455}
456
457void CodeGenerator::AllocateLocations(HInstruction* instruction) {
458  instruction->Accept(GetLocationBuilder());
459  DCHECK(CheckTypeConsistency(instruction));
460  LocationSummary* locations = instruction->GetLocations();
461  if (!instruction->IsSuspendCheckEntry()) {
462    if (locations != nullptr && locations->CanCall()) {
463      MarkNotLeaf();
464    }
465    if (instruction->NeedsCurrentMethod()) {
466      SetRequiresCurrentMethod();
467    }
468  }
469}
470
471CodeGenerator* CodeGenerator::Create(HGraph* graph,
472                                     InstructionSet instruction_set,
473                                     const InstructionSetFeatures& isa_features,
474                                     const CompilerOptions& compiler_options) {
475  switch (instruction_set) {
476    case kArm:
477    case kThumb2: {
478      return new arm::CodeGeneratorARM(graph,
479          *isa_features.AsArmInstructionSetFeatures(),
480          compiler_options);
481    }
482    case kArm64: {
483      return new arm64::CodeGeneratorARM64(graph,
484          *isa_features.AsArm64InstructionSetFeatures(),
485          compiler_options);
486    }
487    case kMips:
488      return nullptr;
489    case kX86: {
490      return new x86::CodeGeneratorX86(graph,
491           *isa_features.AsX86InstructionSetFeatures(),
492           compiler_options);
493    }
494    case kX86_64: {
495      return new x86_64::CodeGeneratorX86_64(graph,
496          *isa_features.AsX86_64InstructionSetFeatures(),
497          compiler_options);
498    }
499    default:
500      return nullptr;
501  }
502}
503
504void CodeGenerator::BuildNativeGCMap(
505    std::vector<uint8_t>* data, const DexCompilationUnit& dex_compilation_unit) const {
506  const std::vector<uint8_t>& gc_map_raw =
507      dex_compilation_unit.GetVerifiedMethod()->GetDexGcMap();
508  verifier::DexPcToReferenceMap dex_gc_map(&(gc_map_raw)[0]);
509
510  uint32_t max_native_offset = stack_map_stream_.ComputeMaxNativePcOffset();
511
512  size_t num_stack_maps = stack_map_stream_.GetNumberOfStackMaps();
513  GcMapBuilder builder(data, num_stack_maps, max_native_offset, dex_gc_map.RegWidth());
514  for (size_t i = 0; i != num_stack_maps; ++i) {
515    const StackMapStream::StackMapEntry& stack_map_entry = stack_map_stream_.GetStackMap(i);
516    uint32_t native_offset = stack_map_entry.native_pc_offset;
517    uint32_t dex_pc = stack_map_entry.dex_pc;
518    const uint8_t* references = dex_gc_map.FindBitMap(dex_pc, false);
519    CHECK(references != nullptr) << "Missing ref for dex pc 0x" << std::hex << dex_pc;
520    builder.AddEntry(native_offset, references);
521  }
522}
523
524void CodeGenerator::BuildSourceMap(DefaultSrcMap* src_map) const {
525  for (size_t i = 0, num = stack_map_stream_.GetNumberOfStackMaps(); i != num; ++i) {
526    const StackMapStream::StackMapEntry& stack_map_entry = stack_map_stream_.GetStackMap(i);
527    uint32_t pc2dex_offset = stack_map_entry.native_pc_offset;
528    int32_t pc2dex_dalvik_offset = stack_map_entry.dex_pc;
529    src_map->push_back(SrcMapElem({pc2dex_offset, pc2dex_dalvik_offset}));
530  }
531}
532
533void CodeGenerator::BuildMappingTable(std::vector<uint8_t>* data) const {
534  uint32_t pc2dex_data_size = 0u;
535  uint32_t pc2dex_entries = stack_map_stream_.GetNumberOfStackMaps();
536  uint32_t pc2dex_offset = 0u;
537  int32_t pc2dex_dalvik_offset = 0;
538  uint32_t dex2pc_data_size = 0u;
539  uint32_t dex2pc_entries = 0u;
540  uint32_t dex2pc_offset = 0u;
541  int32_t dex2pc_dalvik_offset = 0;
542
543  for (size_t i = 0; i < pc2dex_entries; i++) {
544    const StackMapStream::StackMapEntry& stack_map_entry = stack_map_stream_.GetStackMap(i);
545    pc2dex_data_size += UnsignedLeb128Size(stack_map_entry.native_pc_offset - pc2dex_offset);
546    pc2dex_data_size += SignedLeb128Size(stack_map_entry.dex_pc - pc2dex_dalvik_offset);
547    pc2dex_offset = stack_map_entry.native_pc_offset;
548    pc2dex_dalvik_offset = stack_map_entry.dex_pc;
549  }
550
551  // Walk over the blocks and find which ones correspond to catch block entries.
552  for (size_t i = 0; i < graph_->GetBlocks().Size(); ++i) {
553    HBasicBlock* block = graph_->GetBlocks().Get(i);
554    if (block->IsCatchBlock()) {
555      intptr_t native_pc = GetAddressOf(block);
556      ++dex2pc_entries;
557      dex2pc_data_size += UnsignedLeb128Size(native_pc - dex2pc_offset);
558      dex2pc_data_size += SignedLeb128Size(block->GetDexPc() - dex2pc_dalvik_offset);
559      dex2pc_offset = native_pc;
560      dex2pc_dalvik_offset = block->GetDexPc();
561    }
562  }
563
564  uint32_t total_entries = pc2dex_entries + dex2pc_entries;
565  uint32_t hdr_data_size = UnsignedLeb128Size(total_entries) + UnsignedLeb128Size(pc2dex_entries);
566  uint32_t data_size = hdr_data_size + pc2dex_data_size + dex2pc_data_size;
567  data->resize(data_size);
568
569  uint8_t* data_ptr = &(*data)[0];
570  uint8_t* write_pos = data_ptr;
571
572  write_pos = EncodeUnsignedLeb128(write_pos, total_entries);
573  write_pos = EncodeUnsignedLeb128(write_pos, pc2dex_entries);
574  DCHECK_EQ(static_cast<size_t>(write_pos - data_ptr), hdr_data_size);
575  uint8_t* write_pos2 = write_pos + pc2dex_data_size;
576
577  pc2dex_offset = 0u;
578  pc2dex_dalvik_offset = 0u;
579  dex2pc_offset = 0u;
580  dex2pc_dalvik_offset = 0u;
581
582  for (size_t i = 0; i < pc2dex_entries; i++) {
583    const StackMapStream::StackMapEntry& stack_map_entry = stack_map_stream_.GetStackMap(i);
584    DCHECK(pc2dex_offset <= stack_map_entry.native_pc_offset);
585    write_pos = EncodeUnsignedLeb128(write_pos, stack_map_entry.native_pc_offset - pc2dex_offset);
586    write_pos = EncodeSignedLeb128(write_pos, stack_map_entry.dex_pc - pc2dex_dalvik_offset);
587    pc2dex_offset = stack_map_entry.native_pc_offset;
588    pc2dex_dalvik_offset = stack_map_entry.dex_pc;
589  }
590
591  for (size_t i = 0; i < graph_->GetBlocks().Size(); ++i) {
592    HBasicBlock* block = graph_->GetBlocks().Get(i);
593    if (block->IsCatchBlock()) {
594      intptr_t native_pc = GetAddressOf(block);
595      write_pos2 = EncodeUnsignedLeb128(write_pos2, native_pc - dex2pc_offset);
596      write_pos2 = EncodeSignedLeb128(write_pos2, block->GetDexPc() - dex2pc_dalvik_offset);
597      dex2pc_offset = native_pc;
598      dex2pc_dalvik_offset = block->GetDexPc();
599    }
600  }
601
602
603  DCHECK_EQ(static_cast<size_t>(write_pos - data_ptr), hdr_data_size + pc2dex_data_size);
604  DCHECK_EQ(static_cast<size_t>(write_pos2 - data_ptr), data_size);
605
606  if (kIsDebugBuild) {
607    // Verify the encoded table holds the expected data.
608    MappingTable table(data_ptr);
609    CHECK_EQ(table.TotalSize(), total_entries);
610    CHECK_EQ(table.PcToDexSize(), pc2dex_entries);
611    auto it = table.PcToDexBegin();
612    auto it2 = table.DexToPcBegin();
613    for (size_t i = 0; i < pc2dex_entries; i++) {
614      const StackMapStream::StackMapEntry& stack_map_entry = stack_map_stream_.GetStackMap(i);
615      CHECK_EQ(stack_map_entry.native_pc_offset, it.NativePcOffset());
616      CHECK_EQ(stack_map_entry.dex_pc, it.DexPc());
617      ++it;
618    }
619    for (size_t i = 0; i < graph_->GetBlocks().Size(); ++i) {
620      HBasicBlock* block = graph_->GetBlocks().Get(i);
621      if (block->IsCatchBlock()) {
622        CHECK_EQ(GetAddressOf(block), it2.NativePcOffset());
623        CHECK_EQ(block->GetDexPc(), it2.DexPc());
624        ++it2;
625      }
626    }
627    CHECK(it == table.PcToDexEnd());
628    CHECK(it2 == table.DexToPcEnd());
629  }
630}
631
632void CodeGenerator::BuildVMapTable(std::vector<uint8_t>* data) const {
633  Leb128EncodingVector vmap_encoder;
634  // We currently don't use callee-saved registers.
635  size_t size = 0 + 1 /* marker */ + 0;
636  vmap_encoder.Reserve(size + 1u);  // All values are likely to be one byte in ULEB128 (<128).
637  vmap_encoder.PushBackUnsigned(size);
638  vmap_encoder.PushBackUnsigned(VmapTable::kAdjustedFpMarker);
639
640  *data = vmap_encoder.GetData();
641}
642
643void CodeGenerator::BuildStackMaps(std::vector<uint8_t>* data) {
644  uint32_t size = stack_map_stream_.PrepareForFillIn();
645  data->resize(size);
646  MemoryRegion region(data->data(), size);
647  stack_map_stream_.FillIn(region);
648}
649
650void CodeGenerator::RecordPcInfo(HInstruction* instruction,
651                                 uint32_t dex_pc,
652                                 SlowPathCode* slow_path) {
653  if (instruction != nullptr) {
654    // The code generated for some type conversions may call the
655    // runtime, thus normally requiring a subsequent call to this
656    // method.  However, the method verifier does not produce PC
657    // information for certain instructions, which are considered "atomic"
658    // (they cannot join a GC).
659    // Therefore we do not currently record PC information for such
660    // instructions.  As this may change later, we added this special
661    // case so that code generators may nevertheless call
662    // CodeGenerator::RecordPcInfo without triggering an error in
663    // CodeGenerator::BuildNativeGCMap ("Missing ref for dex pc 0x")
664    // thereafter.
665    if (instruction->IsTypeConversion()) {
666      return;
667    }
668    if (instruction->IsRem()) {
669      Primitive::Type type = instruction->AsRem()->GetResultType();
670      if ((type == Primitive::kPrimFloat) || (type == Primitive::kPrimDouble)) {
671        return;
672      }
673    }
674  }
675
676  uint32_t outer_dex_pc = dex_pc;
677  uint32_t outer_environment_size = 0;
678  uint32_t inlining_depth = 0;
679  if (instruction != nullptr) {
680    for (HEnvironment* environment = instruction->GetEnvironment();
681         environment != nullptr;
682         environment = environment->GetParent()) {
683      outer_dex_pc = environment->GetDexPc();
684      outer_environment_size = environment->Size();
685      if (environment != instruction->GetEnvironment()) {
686        inlining_depth++;
687      }
688    }
689  }
690
691  // Collect PC infos for the mapping table.
692  uint32_t native_pc = GetAssembler()->CodeSize();
693
694  if (instruction == nullptr) {
695    // For stack overflow checks.
696    stack_map_stream_.BeginStackMapEntry(outer_dex_pc, native_pc, 0, 0, 0, 0);
697    stack_map_stream_.EndStackMapEntry();
698    return;
699  }
700  LocationSummary* locations = instruction->GetLocations();
701
702  uint32_t register_mask = locations->GetRegisterMask();
703  if (locations->OnlyCallsOnSlowPath()) {
704    // In case of slow path, we currently set the location of caller-save registers
705    // to register (instead of their stack location when pushed before the slow-path
706    // call). Therefore register_mask contains both callee-save and caller-save
707    // registers that hold objects. We must remove the caller-save from the mask, since
708    // they will be overwritten by the callee.
709    register_mask &= core_callee_save_mask_;
710  }
711  // The register mask must be a subset of callee-save registers.
712  DCHECK_EQ(register_mask & core_callee_save_mask_, register_mask);
713  stack_map_stream_.BeginStackMapEntry(outer_dex_pc,
714                                       native_pc,
715                                       register_mask,
716                                       locations->GetStackMask(),
717                                       outer_environment_size,
718                                       inlining_depth);
719
720  EmitEnvironment(instruction->GetEnvironment(), slow_path);
721  stack_map_stream_.EndStackMapEntry();
722}
723
724void CodeGenerator::EmitEnvironment(HEnvironment* environment, SlowPathCode* slow_path) {
725  if (environment == nullptr) return;
726
727  if (environment->GetParent() != nullptr) {
728    // We emit the parent environment first.
729    EmitEnvironment(environment->GetParent(), slow_path);
730    stack_map_stream_.BeginInlineInfoEntry(environment->GetMethodIdx(),
731                                           environment->GetDexPc(),
732                                           environment->GetInvokeType(),
733                                           environment->Size());
734  }
735
736  // Walk over the environment, and record the location of dex registers.
737  for (size_t i = 0, environment_size = environment->Size(); i < environment_size; ++i) {
738    HInstruction* current = environment->GetInstructionAt(i);
739    if (current == nullptr) {
740      stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kNone, 0);
741      continue;
742    }
743
744    Location location = environment->GetLocationAt(i);
745    switch (location.GetKind()) {
746      case Location::kConstant: {
747        DCHECK_EQ(current, location.GetConstant());
748        if (current->IsLongConstant()) {
749          int64_t value = current->AsLongConstant()->GetValue();
750          stack_map_stream_.AddDexRegisterEntry(
751              DexRegisterLocation::Kind::kConstant, Low32Bits(value));
752          stack_map_stream_.AddDexRegisterEntry(
753              DexRegisterLocation::Kind::kConstant, High32Bits(value));
754          ++i;
755          DCHECK_LT(i, environment_size);
756        } else if (current->IsDoubleConstant()) {
757          int64_t value = bit_cast<int64_t, double>(current->AsDoubleConstant()->GetValue());
758          stack_map_stream_.AddDexRegisterEntry(
759              DexRegisterLocation::Kind::kConstant, Low32Bits(value));
760          stack_map_stream_.AddDexRegisterEntry(
761              DexRegisterLocation::Kind::kConstant, High32Bits(value));
762          ++i;
763          DCHECK_LT(i, environment_size);
764        } else if (current->IsIntConstant()) {
765          int32_t value = current->AsIntConstant()->GetValue();
766          stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kConstant, value);
767        } else if (current->IsNullConstant()) {
768          stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kConstant, 0);
769        } else {
770          DCHECK(current->IsFloatConstant()) << current->DebugName();
771          int32_t value = bit_cast<int32_t, float>(current->AsFloatConstant()->GetValue());
772          stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kConstant, value);
773        }
774        break;
775      }
776
777      case Location::kStackSlot: {
778        stack_map_stream_.AddDexRegisterEntry(
779            DexRegisterLocation::Kind::kInStack, location.GetStackIndex());
780        break;
781      }
782
783      case Location::kDoubleStackSlot: {
784        stack_map_stream_.AddDexRegisterEntry(
785            DexRegisterLocation::Kind::kInStack, location.GetStackIndex());
786        stack_map_stream_.AddDexRegisterEntry(
787            DexRegisterLocation::Kind::kInStack, location.GetHighStackIndex(kVRegSize));
788        ++i;
789        DCHECK_LT(i, environment_size);
790        break;
791      }
792
793      case Location::kRegister : {
794        int id = location.reg();
795        if (slow_path != nullptr && slow_path->IsCoreRegisterSaved(id)) {
796          uint32_t offset = slow_path->GetStackOffsetOfCoreRegister(id);
797          stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInStack, offset);
798          if (current->GetType() == Primitive::kPrimLong) {
799            stack_map_stream_.AddDexRegisterEntry(
800                DexRegisterLocation::Kind::kInStack, offset + kVRegSize);
801            ++i;
802            DCHECK_LT(i, environment_size);
803          }
804        } else {
805          stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInRegister, id);
806          if (current->GetType() == Primitive::kPrimLong) {
807            stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInRegister, id);
808            ++i;
809            DCHECK_LT(i, environment_size);
810          }
811        }
812        break;
813      }
814
815      case Location::kFpuRegister : {
816        int id = location.reg();
817        if (slow_path != nullptr && slow_path->IsFpuRegisterSaved(id)) {
818          uint32_t offset = slow_path->GetStackOffsetOfFpuRegister(id);
819          stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInStack, offset);
820          if (current->GetType() == Primitive::kPrimDouble) {
821            stack_map_stream_.AddDexRegisterEntry(
822                DexRegisterLocation::Kind::kInStack, offset + kVRegSize);
823            ++i;
824            DCHECK_LT(i, environment_size);
825          }
826        } else {
827          stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInFpuRegister, id);
828          if (current->GetType() == Primitive::kPrimDouble) {
829            stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInFpuRegister, id);
830            ++i;
831            DCHECK_LT(i, environment_size);
832          }
833        }
834        break;
835      }
836
837      case Location::kFpuRegisterPair : {
838        int low = location.low();
839        int high = location.high();
840        if (slow_path != nullptr && slow_path->IsFpuRegisterSaved(low)) {
841          uint32_t offset = slow_path->GetStackOffsetOfFpuRegister(low);
842          stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInStack, offset);
843        } else {
844          stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInFpuRegister, low);
845        }
846        if (slow_path != nullptr && slow_path->IsFpuRegisterSaved(high)) {
847          uint32_t offset = slow_path->GetStackOffsetOfFpuRegister(high);
848          stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInStack, offset);
849          ++i;
850        } else {
851          stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInFpuRegister, high);
852          ++i;
853        }
854        DCHECK_LT(i, environment_size);
855        break;
856      }
857
858      case Location::kRegisterPair : {
859        int low = location.low();
860        int high = location.high();
861        if (slow_path != nullptr && slow_path->IsCoreRegisterSaved(low)) {
862          uint32_t offset = slow_path->GetStackOffsetOfCoreRegister(low);
863          stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInStack, offset);
864        } else {
865          stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInRegister, low);
866        }
867        if (slow_path != nullptr && slow_path->IsCoreRegisterSaved(high)) {
868          uint32_t offset = slow_path->GetStackOffsetOfCoreRegister(high);
869          stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInStack, offset);
870        } else {
871          stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kInRegister, high);
872        }
873        ++i;
874        DCHECK_LT(i, environment_size);
875        break;
876      }
877
878      case Location::kInvalid: {
879        stack_map_stream_.AddDexRegisterEntry(DexRegisterLocation::Kind::kNone, 0);
880        break;
881      }
882
883      default:
884        LOG(FATAL) << "Unexpected kind " << location.GetKind();
885    }
886  }
887
888  if (environment->GetParent() != nullptr) {
889    stack_map_stream_.EndInlineInfoEntry();
890  }
891}
892
893bool CodeGenerator::CanMoveNullCheckToUser(HNullCheck* null_check) {
894  HInstruction* first_next_not_move = null_check->GetNextDisregardingMoves();
895
896  return (first_next_not_move != nullptr)
897      && first_next_not_move->CanDoImplicitNullCheckOn(null_check->InputAt(0));
898}
899
900void CodeGenerator::MaybeRecordImplicitNullCheck(HInstruction* instr) {
901  // If we are from a static path don't record the pc as we can't throw NPE.
902  // NB: having the checks here makes the code much less verbose in the arch
903  // specific code generators.
904  if (instr->IsStaticFieldSet() || instr->IsStaticFieldGet()) {
905    return;
906  }
907
908  if (!compiler_options_.GetImplicitNullChecks()) {
909    return;
910  }
911
912  if (!instr->CanDoImplicitNullCheckOn(instr->InputAt(0))) {
913    return;
914  }
915
916  // Find the first previous instruction which is not a move.
917  HInstruction* first_prev_not_move = instr->GetPreviousDisregardingMoves();
918
919  // If the instruction is a null check it means that `instr` is the first user
920  // and needs to record the pc.
921  if (first_prev_not_move != nullptr && first_prev_not_move->IsNullCheck()) {
922    HNullCheck* null_check = first_prev_not_move->AsNullCheck();
923    // TODO: The parallel moves modify the environment. Their changes need to be reverted
924    // otherwise the stack maps at the throw point will not be correct.
925    RecordPcInfo(null_check, null_check->GetDexPc());
926  }
927}
928
929void CodeGenerator::ClearSpillSlotsFromLoopPhisInStackMap(HSuspendCheck* suspend_check) const {
930  LocationSummary* locations = suspend_check->GetLocations();
931  HBasicBlock* block = suspend_check->GetBlock();
932  DCHECK(block->GetLoopInformation()->GetSuspendCheck() == suspend_check);
933  DCHECK(block->IsLoopHeader());
934
935  for (HInstructionIterator it(block->GetPhis()); !it.Done(); it.Advance()) {
936    HInstruction* current = it.Current();
937    LiveInterval* interval = current->GetLiveInterval();
938    // We only need to clear bits of loop phis containing objects and allocated in register.
939    // Loop phis allocated on stack already have the object in the stack.
940    if (current->GetType() == Primitive::kPrimNot
941        && interval->HasRegister()
942        && interval->HasSpillSlot()) {
943      locations->ClearStackBit(interval->GetSpillSlot() / kVRegSize);
944    }
945  }
946}
947
948void CodeGenerator::EmitParallelMoves(Location from1,
949                                      Location to1,
950                                      Primitive::Type type1,
951                                      Location from2,
952                                      Location to2,
953                                      Primitive::Type type2) {
954  HParallelMove parallel_move(GetGraph()->GetArena());
955  parallel_move.AddMove(from1, to1, type1, nullptr);
956  parallel_move.AddMove(from2, to2, type2, nullptr);
957  GetMoveResolver()->EmitNativeCode(&parallel_move);
958}
959
960void SlowPathCode::RecordPcInfo(CodeGenerator* codegen, HInstruction* instruction, uint32_t dex_pc) {
961  codegen->RecordPcInfo(instruction, dex_pc, this);
962}
963
964void SlowPathCode::SaveLiveRegisters(CodeGenerator* codegen, LocationSummary* locations) {
965  RegisterSet* register_set = locations->GetLiveRegisters();
966  size_t stack_offset = codegen->GetFirstRegisterSlotInSlowPath();
967  for (size_t i = 0, e = codegen->GetNumberOfCoreRegisters(); i < e; ++i) {
968    if (!codegen->IsCoreCalleeSaveRegister(i)) {
969      if (register_set->ContainsCoreRegister(i)) {
970        // If the register holds an object, update the stack mask.
971        if (locations->RegisterContainsObject(i)) {
972          locations->SetStackBit(stack_offset / kVRegSize);
973        }
974        DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize());
975        DCHECK_LT(i, kMaximumNumberOfExpectedRegisters);
976        saved_core_stack_offsets_[i] = stack_offset;
977        stack_offset += codegen->SaveCoreRegister(stack_offset, i);
978      }
979    }
980  }
981
982  for (size_t i = 0, e = codegen->GetNumberOfFloatingPointRegisters(); i < e; ++i) {
983    if (!codegen->IsFloatingPointCalleeSaveRegister(i)) {
984      if (register_set->ContainsFloatingPointRegister(i)) {
985        DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize());
986        DCHECK_LT(i, kMaximumNumberOfExpectedRegisters);
987        saved_fpu_stack_offsets_[i] = stack_offset;
988        stack_offset += codegen->SaveFloatingPointRegister(stack_offset, i);
989      }
990    }
991  }
992}
993
994void SlowPathCode::RestoreLiveRegisters(CodeGenerator* codegen, LocationSummary* locations) {
995  RegisterSet* register_set = locations->GetLiveRegisters();
996  size_t stack_offset = codegen->GetFirstRegisterSlotInSlowPath();
997  for (size_t i = 0, e = codegen->GetNumberOfCoreRegisters(); i < e; ++i) {
998    if (!codegen->IsCoreCalleeSaveRegister(i)) {
999      if (register_set->ContainsCoreRegister(i)) {
1000        DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize());
1001        stack_offset += codegen->RestoreCoreRegister(stack_offset, i);
1002      }
1003    }
1004  }
1005
1006  for (size_t i = 0, e = codegen->GetNumberOfFloatingPointRegisters(); i < e; ++i) {
1007    if (!codegen->IsFloatingPointCalleeSaveRegister(i)) {
1008      if (register_set->ContainsFloatingPointRegister(i)) {
1009        DCHECK_LT(stack_offset, codegen->GetFrameSize() - codegen->FrameEntrySpillSize());
1010        stack_offset += codegen->RestoreFloatingPointRegister(stack_offset, i);
1011      }
1012    }
1013  }
1014}
1015
1016}  // namespace art
1017